diff --git a/.gitattributes b/.gitattributes index 3435ad956b0b19e6cd5ab52a1a995a94e39cdb45..118c229dee4e4533c53327de9ee3848cb3699f61 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5148,3 +5148,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Unraveling[[:space:]]Instance[[:space:]]Associations_[[:space:]]A[[:space:]]Closer[[:space:]]Look[[:space:]]for[[:space:]]Audio-Visual[[:space:]]Segmentation/bbcd7ad3-823f-4a3e-8587-9fa89b871675_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Unsegment[[:space:]]Anything[[:space:]]by[[:space:]]Simulating[[:space:]]Deformation/e8c8ca9c-763d-45fa-8f3b-d176b29e79d6_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Unsigned[[:space:]]Orthogonal[[:space:]]Distance[[:space:]]Fields_[[:space:]]An[[:space:]]Accurate[[:space:]]Neural[[:space:]]Implicit[[:space:]]Representation[[:space:]]for[[:space:]]Diverse[[:space:]]3D[[:space:]]Shapes/7478ca21-59b6-4d95-b955-47a5490cfeda_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]3D[[:space:]]Structure[[:space:]]Inference[[:space:]]from[[:space:]]Category-Specific[[:space:]]Image[[:space:]]Collections/9e81eb55-2017-47d1-ab50-718fc069c773_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Blind[[:space:]]Image[[:space:]]Deblurring[[:space:]]Based[[:space:]]on[[:space:]]Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Deep[[:space:]]Unrolling[[:space:]]Networks[[:space:]]for[[:space:]]Phase[[:space:]]Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Feature[[:space:]]Learning[[:space:]]with[[:space:]]Emergent[[:space:]]Data-Driven[[:space:]]Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Gaze[[:space:]]Representation[[:space:]]Learning[[:space:]]from[[:space:]]Multi-view[[:space:]]Face[[:space:]]Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Keypoints[[:space:]]from[[:space:]]Pretrained[[:space:]]Diffusion[[:space:]]Models/4645df0d-9e97-449d-a096-f3522e36e6da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Learning[[:space:]]of[[:space:]]Category-Level[[:space:]]3D[[:space:]]Pose[[:space:]]from[[:space:]]Object-Centric[[:space:]]Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Occupancy[[:space:]]Learning[[:space:]]from[[:space:]]Sparse[[:space:]]Point[[:space:]]Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Salient[[:space:]]Instance[[:space:]]Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]Through[[:space:]]Depth-Guided[[:space:]]Feature[[:space:]]Correlation[[:space:]]and[[:space:]]Sampling/490ac818-ae06-4da1-a305-2db07efc0513_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Template-assisted[[:space:]]Point[[:space:]]Cloud[[:space:]]Shape[[:space:]]Correspondence[[:space:]]Network/812a1d24-1f0f-400f-8c84-da6720a2a353_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Universal[[:space:]]Image[[:space:]]Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unsupervised[[:space:]]Video[[:space:]]Domain[[:space:]]Adaptation[[:space:]]with[[:space:]]Masked[[:space:]]Pre-Training[[:space:]]and[[:space:]]Collaborative[[:space:]]Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unveiling[[:space:]]Parts[[:space:]]Beyond[[:space:]]Objects_[[:space:]]Towards[[:space:]]Finer-Granularity[[:space:]]Referring[[:space:]]Expression[[:space:]]Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unveiling[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Audio-Visual[[:space:]]Early[[:space:]]Fusion[[:space:]]Transformers[[:space:]]with[[:space:]]Dense[[:space:]]Interactions[[:space:]]through[[:space:]]Masked[[:space:]]Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Unveiling[[:space:]]the[[:space:]]Unknown_[[:space:]]Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Unknown[[:space:]]to[[:space:]]Known[[:space:]]in[[:space:]]Open-Set[[:space:]]Source-Free[[:space:]]Domain[[:space:]]Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Upscale-A-Video_[[:space:]]Temporal-Consistent[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Real-World[[:space:]]Video[[:space:]]Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Using[[:space:]]Human[[:space:]]Feedback[[:space:]]to[[:space:]]Fine-tune[[:space:]]Diffusion[[:space:]]Models[[:space:]]without[[:space:]]Any[[:space:]]Reward[[:space:]]Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Utility-Fairness[[:space:]]Trade-Offs[[:space:]]and[[:space:]]How[[:space:]]to[[:space:]]Find[[:space:]]Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VA3_[[:space:]]Virtually[[:space:]]Assured[[:space:]]Amplification[[:space:]]Attack[[:space:]]on[[:space:]]Probabilistic[[:space:]]Copyright[[:space:]]Protection[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generative[[:space:]]Models/294458f2-3d11-4022-9469-5c90ec662dfc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VAREN_[[:space:]]Very[[:space:]]Accurate[[:space:]]and[[:space:]]Realistic[[:space:]]Equine[[:space:]]Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VBench_[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]Suite[[:space:]]for[[:space:]]Video[[:space:]]Generative[[:space:]]Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VCoder_[[:space:]]Versatile[[:space:]]Vision[[:space:]]Encoders[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VGGSfM_[[:space:]]Visual[[:space:]]Geometry[[:space:]]Grounded[[:space:]]Deep[[:space:]]Structure[[:space:]]From[[:space:]]Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VILA_[[:space:]]On[[:space:]]Pre-training[[:space:]]for[[:space:]]Visual[[:space:]]Language[[:space:]]Models/f648f3d5-06a0-470c-8e61-c14e811bec54_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VINECS_[[:space:]]Video-based[[:space:]]Neural[[:space:]]Character[[:space:]]Skinning/14057e08-ebbd-4725-9942-632446ff87e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VISTA-LLAMA_[[:space:]]Reducing[[:space:]]Hallucination[[:space:]]in[[:space:]]Video[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Equal[[:space:]]Distance[[:space:]]to[[:space:]]Visual[[:space:]]Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VLP_[[:space:]]Vision[[:space:]]Language[[:space:]]Planning[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VMC_[[:space:]]Video[[:space:]]Motion[[:space:]]Customization[[:space:]]using[[:space:]]Temporal[[:space:]]Attention[[:space:]]Adaption[[:space:]]for[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VMINer_[[:space:]]Versatile[[:space:]]Multi-view[[:space:]]Inverse[[:space:]]Rendering[[:space:]]with[[:space:]]Near-[[:space:]]and[[:space:]]Far-field[[:space:]]Light[[:space:]]Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VOODOO[[:space:]]3D_[[:space:]]Volumetric[[:space:]]Portrait[[:space:]]Disentanglement[[:space:]]For[[:space:]]One-Shot[[:space:]]3D[[:space:]]Head[[:space:]]Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VP3D_[[:space:]]Unleashing[[:space:]]2D[[:space:]]Visual[[:space:]]Prompt[[:space:]]for[[:space:]]Text-to-3D[[:space:]]Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VRP-SAM_[[:space:]]SAM[[:space:]]with[[:space:]]Visual[[:space:]]Reference[[:space:]]Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VRetouchEr_[[:space:]]Learning[[:space:]]Cross-frame[[:space:]]Feature[[:space:]]Interdependence[[:space:]]with[[:space:]]Imperfection[[:space:]]Flow[[:space:]]for[[:space:]]Face[[:space:]]Retouching[[:space:]]in[[:space:]]Videos/8a896d01-f26d-44df-a67e-710260f2b006_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VSCode_[[:space:]]General[[:space:]]Visual[[:space:]]Salient[[:space:]]and[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]2D[[:space:]]Prompt[[:space:]]Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VSRD_[[:space:]]Instance-Aware[[:space:]]Volumetric[[:space:]]Silhouette[[:space:]]Rendering[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]3D[[:space:]]Object[[:space:]]Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VS_[[:space:]]Reconstructing[[:space:]]Clothed[[:space:]]3D[[:space:]]Human[[:space:]]from[[:space:]]Single[[:space:]]Image[[:space:]]via[[:space:]]Vertex[[:space:]]Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VTQA_[[:space:]]Visual[[:space:]]Text[[:space:]]Question[[:space:]]Answering[[:space:]]via[[:space:]]Entity[[:space:]]Alignment[[:space:]]and[[:space:]]Cross-Media[[:space:]]Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VTimeLLM_[[:space:]]Empower[[:space:]]LLM[[:space:]]to[[:space:]]Grasp[[:space:]]Video[[:space:]]Moments/93551b6a-586f-4f31-83ad-4d86b76784d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/V__[[:space:]]Guided[[:space:]]Visual[[:space:]]Search[[:space:]]as[[:space:]]a[[:space:]]Core[[:space:]]Mechanism[[:space:]]in[[:space:]]Multimodal[[:space:]]LLMs/da9d620a-ef32-4c5e-b349-8d066a704f0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Validating[[:space:]]Privacy-Preserving[[:space:]]Face[[:space:]]Recognition[[:space:]]under[[:space:]]a[[:space:]]Minimum[[:space:]]Assumption/9139a567-fbb9-4834-bbbf-7d18ccf906f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Vanishing-Point-Guided[[:space:]]Video[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]of[[:space:]]Driving[[:space:]]Scenes/c95c8cd9-ea0a-498e-afad-6c04356cea08_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VastGaussian_[[:space:]]Vast[[:space:]]3D[[:space:]]Gaussians[[:space:]]for[[:space:]]Large[[:space:]]Scene[[:space:]]Reconstruction/dae866b8-bba9-4088-8662-ebfdb451138c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VecFusion_[[:space:]]Vector[[:space:]]Font[[:space:]]Generation[[:space:]]with[[:space:]]Diffusion/c5f5c319-ebd3-42c5-86d7-2ff36fd3c349_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Vector[[:space:]]Graphics[[:space:]]Generation[[:space:]]via[[:space:]]Mutually[[:space:]]Impulsed[[:space:]]Dual-domain[[:space:]]Diffusion/dad89f64-ff9a-4db0-b8ae-8201e4fb21a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Versatile[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation[[:space:]]Learned[[:space:]]from[[:space:]]Multi-Source[[:space:]]Datasets[[:space:]]via[[:space:]]Model[[:space:]]Self-Disambiguation/c3e7e066-94ab-4dd4-a693-b92631ccd737_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Versatile[[:space:]]Navigation[[:space:]]Under[[:space:]]Partial[[:space:]]Observability[[:space:]]via[[:space:]]Value-guided[[:space:]]Diffusion[[:space:]]Policy/dec738eb-2199-49f4-ba41-b8bdbd7e0fa4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViLa-MIL_[[:space:]]Dual-scale[[:space:]]Vision-Language[[:space:]]Multiple[[:space:]]Instance[[:space:]]Learning[[:space:]]for[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification/85372ebf-31af-4d52-a240-dcdd7001d47f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViP-LLaVA_[[:space:]]Making[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]Understand[[:space:]]Arbitrary[[:space:]]Visual[[:space:]]Prompts/4764e32b-b5e0-4cec-b885-ce28665cbc4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViT-CoMer_[[:space:]]Vision[[:space:]]Transformer[[:space:]]with[[:space:]]Convolutional[[:space:]]Multi-scale[[:space:]]Feature[[:space:]]Interaction[[:space:]]for[[:space:]]Dense[[:space:]]Predictions/2a3323b9-7324-4cd3-aae1-ea2eb71358b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViT-Lens_[[:space:]]Towards[[:space:]]Omni-modal[[:space:]]Representations/f3bd0731-24e0-43f0-9476-94f88d4b11b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViTamin_[[:space:]]Designing[[:space:]]Scalable[[:space:]]Vision[[:space:]]Models[[:space:]]in[[:space:]]the[[:space:]]Vision-Language[[:space:]]Era/ca2d983d-6b19-4d87-995a-d003078f34a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ViVid-1-to-3_[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis[[:space:]]with[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models/d4609e1c-e930-4647-bcd0-e97dd36ab5d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VicTR_[[:space:]]Video-conditioned[[:space:]]Text[[:space:]]Representations[[:space:]]for[[:space:]]Activity[[:space:]]Recognition/29323e50-5f81-4d07-b22d-e4ee0037aaef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VidLA_[[:space:]]Video-Language[[:space:]]Alignment[[:space:]]at[[:space:]]Scale/fc43318d-405d-4040-b280-5d6a6f3f19c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/VidToMe_[[:space:]]Video[[:space:]]Token[[:space:]]Merging[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Video[[:space:]]Editing/fb29b626-2268-4dba-853b-90322a5c4cdf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]via[[:space:]]Direct[[:space:]]Synthesis[[:space:]]with[[:space:]]the[[:space:]]Event-based[[:space:]]Reference/2d993f24-453b-4589-9f64-71049d897aff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Harmonization[[:space:]]with[[:space:]]Triplet[[:space:]]Spatio-Temporal[[:space:]]Variation[[:space:]]Patterns/5c610f72-b14a-4ec5-88f7-a1cf9a33dda3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Interpolation[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/fd333655-848a-4ba1-8c05-96c4cf9bfdad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Prediction[[:space:]]by[[:space:]]Modeling[[:space:]]Videos[[:space:]]as[[:space:]]Continuous[[:space:]]Multi-Dimensional[[:space:]]Processes/ccc42f26-9a12-4018-9b85-3006f577d2a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]ReCap_[[:space:]]Recursive[[:space:]]Captioning[[:space:]]of[[:space:]]Hour-Long[[:space:]]Videos/34833510-7e3f-46a3-8e63-d31ed9f4ff24_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Recognition[[:space:]]in[[:space:]]Portrait[[:space:]]Mode/1430817f-2fb2-4e8e-aaf4-dc2cf17e83e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video[[:space:]]Super-Resolution[[:space:]]Transformer[[:space:]]with[[:space:]]Masked[[:space:]]Inter&Intra-Frame[[:space:]]Attention/1e69759c-d8e8-49da-aff4-8c9d96153686_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Video-Based[[:space:]]Human[[:space:]]Pose[[:space:]]Regression[[:space:]]via[[:space:]]Decoupled[[:space:]]Space-Time[[:space:]]Aggregation/f3738d37-ef8c-4352-b1e1-4943413c7dec_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_content_list.json b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2246cb653417d7374475d66707576cef5c3590e9 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_content_list.json @@ -0,0 +1,1553 @@ +[ + { + "type": "text", + "text": "Unsupervised 3D Structure Inference from Category-Specific Image Collections", + "text_level": 1, + "bbox": [ + 84, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weikang Wang", + "bbox": [ + 241, + 180, + 367, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongliang Cao", + "bbox": [ + 419, + 181, + 542, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Florian Bernard", + "bbox": [ + 596, + 181, + 727, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Bonn, Germany", + "bbox": [ + 364, + 205, + 604, + 223 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6a1958682f39e314355aaa2084552487e22c9f5c40e7f1cdf28107a1b914b82.jpg", + "image_caption": [ + "Figure 1. Our method detects consistent and reasonable 3D structure for general object categories from a single 2D image. For training, our method only requires a collection of images from the same category without any additional information." + ], + "image_footnote": [], + "bbox": [ + 78, + 279, + 883, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 488, + 313, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding 3D object structure from image collections of general object categories remains a long-standing challenge in computer vision. Due to the high relevance of image keypoints (e.g. for graph matching, controlling generative models, scene understanding, etc.), in this work we specifically focus on inferring 3D structure in terms of sparse keypoints. Existing 3D keypoint inference approaches rely on strong priors, such as spatio-temporal consistency, multi-view images of the same object, 3D shape priors (e.g. templates, skeleton), or supervisory signals e.g. in the form of 2D keypoint annotations. In contrast, we propose the first unsupervised 3D keypoint inference approach that can be trained for general object categories solely from an inhomogeneous image collection (containing different instances of objects from the same category). Our experiments show that our method not only improves upon unsupervised 2D keypoint inference, but more importantly, it also produces reasonable 3D structure for various object categories, both qualitatively and quantitatively.", + "bbox": [ + 75, + 522, + 473, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding 3D object structure from 2D images is a fundamental problem in computer vision that has been studied", + "bbox": [ + 75, + 871, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for decades. A range of different 3D structure representations have been considered, including keypoints [14, 25, 59, 69], meshes [55, 56, 62], and voxels [35, 45, 47]. Among these, keypoint-based 3D structure representations have the strong advantage that they are simple and flexible, while at the same time being highly relevant for diverse downstream tasks. However, obtaining 3D keypoints from 2D images is hard due to unknown depth, pose variations, different appearance, and the lack of explicit geometric information. Due to these difficulties, prior works mostly focus on 2D keypoints detection instead of 3D [9, 12, 31, 36, 66].", + "bbox": [ + 496, + 489, + 893, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we investigate whether it is possible to directly obtain 3D keypoints without using any explicit supervision. Our work is motivated by the fact that 2D image keypoints are observations of points in 3D space that are projected onto the 2D image plane. By explicitly understanding and inferring 3D geometry we are able to improve upon the quality of detected keypoints. Specifically, we experimentally demonstrate that projecting our predicted 3D keypoints onto the image plane leads to more reliable 2D keypoints compared to existing 2D keypoint predictors. Moreover, our method can serve the increasing amount of tasks that rely on 3D keypoints, e.g. 3D structure-aware generative models [58], 3D avatar generation [34], or 3D reconstruction [55, 56]. Despite keypoint prediction being a well-studied problem in the 2D domain, 3D keypoint inference is under-explored and we thus specifically contribute", + "bbox": [ + 496, + 660, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10704", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "towards this area.", + "bbox": [ + 76, + 90, + 196, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing works on inferring 3D keypoints either use 3D point clouds or 2D images/videos from the same object instance as input data. While the former is conceptually a subset selection problem, in this work we solely focus on the more challenging latter case. Manually obtaining 3D keypoints in image data is much harder than annotating 2D keypoints. Thus, there are only very few datasets with 3D keypoints annotations, which in turn motivates unsupervised learning strategies that do not require expensive or even infeasible 3D annotations for training. Existing works that address 3D keypoint inference from image data use for example 2D keypoint annotations [38, 44], multiple views with known poses [11, 35, 48], spatial-temporal consistency in videos [9, 15], 2D/3D skeleton [55, 56] or templates [14]. In stark contrast, in this work we relax the assumptions about the availability of such supervisory signals and train a 3D keypoint predictor merely from category-specific image collections as input. Our method builds upon the assumption that object instances from the same category should have a similar 3D structure with moderate intra-class variations (e.g. faces, horses, birds). Table 1 summaries the differences among existing methods and Fig. 2 gives a visualization of some additional information within these methods and the comparison to our setting.", + "bbox": [ + 76, + 107, + 473, + 470 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/81e0bea5678df96e27ad1db1e5c7ed817d14febc41a76fe50fbacdd07b687111.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsUnsup.w/o PriorCat.-agnostic
Bulat et al. [3]X✓✓
Gou et al. [10]X✓✓
Zhao et al. [68]X✓✓
Tulyakov et al. [50]X✓✓
Pavlakos et al. [41]X✓X
Zhang et al. [66]X✓X
He et al. [14]XX✓
Mildenhall et al. [35]✓X✓
Wu et al. [54]✓X✓
Moniz et al. [37]✓XX
Supasorn et al. [48]✓X✓
Novotny et al. [38]✓X✓
Park et al. [40]✓X✓
Chen et al. [5]✓X✓
Reddy et al. [44]✓X✓
Zhou et al. [69]✓✓X
Dundar et al. [9]✓✓X
Xu et al. [59]✓✓X
Honari and Fua [15]✓X✓
Ours✓✓✓
", + "bbox": [ + 78, + 484, + 468, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. In comparison to existing 3D keypoints inference methods, our approach combines a unique set of desirable properties: it can be trained in an unsupervised manner without requiring any additional information; it does not require any shape prior and can thus be applied to different object categories.", + "bbox": [ + 75, + 825, + 468, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/31017baa371f382d7b690fc43418510f203ad267aa832ca5db33badeed8922a7.jpg", + "image_caption": [ + "Figure 2. (Top row from left to right) Existing methods require different additional information for 3D structure inference, e.g. multiple views (from [12]), 2D keypoint annotations (from [12]), and a symmetry prior (from [54]). (Bottom row) Our setting does not require such information but instead also infers a 3D shape space model that is learned in an unsupervised manner." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method enables to infer 3D keypoints from category-specific image collections without relying on additional prior information. To this end, we build upon the recent unsupervised 2D keypoint detection method by He et al. [12] and extend it by additionally learning a sparse 3D shape space model to represent 3D structure. To regularise the 3D structure, a self-supervised re-projection loss is introduced to couple 3D keypoints with corresponding 2D keypoints. By combining our 3D shape space model, re-projection loss, and additional regularisation loss terms, we demonstrate that our proposed method not only improves 2D keypoint accuracy, but also produces reasonable 3D structure. We summarise our main contributions as:", + "bbox": [ + 498, + 462, + 892, + 656 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time we obtain 3D keypoints from general category-specific image collections without relying on additional prior information.", + "- Our method obtains reasonable 3D keypoints based on our learnable 3D shape space model that is fused with a state-of-the-art unsupervised 2D keypoint detector.", + "- We demonstrate that our method both improves unsupervised 2D keypoint detection accuracy and obtains reasonable 3D shape on different benchmarks." + ], + "bbox": [ + 500, + 665, + 890, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 837, + 640, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following we summarize relevant 2D and 3D key-point prediction methods.", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 2D Keypoint Detection", + "text_level": 1, + "bbox": [ + 76, + 90, + 287, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised methods. Various domain-specific supervised 2D keypoint detection methods have been proposed that are trained on a large amount of labeled datasets, e.g. for faces [2, 57, 71], human bodies [1, 4, 18, 21, 32], or vehicles [43], etc. However, labeling keypoints for images is time-consuming, laborious and error-prone, especially when there are occlusions, extreme pose variations, or appearance changes. As a consequence, labeling errors will subsequently influence network training and downstream task performance.", + "bbox": [ + 75, + 113, + 468, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised methods. Due to these reasons, an increasing amount of attention is paid to unsupervised methods. One popular concept is multi-view geometric consistency. Thewlis et al. [49] detect keypoints by recovering transformations between one image and its deformed counterpart from pre-defined deformations. Without relying on pre-defined deformations, various video-based methods [9, 19, 20, 22, 24, 36, 46] infer keypoints based on a temporal coherence prior. Instead of using multiple views as input, Xu et al. [60] use unpaired input to disentangle appearance and poses by conducting a swapping-reconstruction strategy and encouraging reconstructed images to resemble the original. He et al. [13] and He et al. [11] use a generative adversarial training framework to constrain keypoint locations by interpreting them as latent codes for an image generator. However, collaboratively learning intermediate keypoints and complex adversarial models is unstable and does not scale. Starting from some initial point sets, Mallis et al. [31] filter keypoints by optimizing alternatively between clustering of keypoints features across a dataset, and training a keypoint detector with pseudo-labels offered by cluster centers. However, this method depends heavily on the initial point set. Zhang et al. [67] address single-image keypoint inference using an auto-encoder architecture to interpret keypoints as a latent representation, while reconstructing the input image from the concatenated keypoints and respective pixel feature descriptors. He et al. [12] use a similar idea by representing keypoints as latent features in an auto-encoder manner. However, they reconstruct the input from an edge map created from keypoints combined with random masking of appearance information. In this way, paired input data is not needed to disentangle appearance and poses, which makes this method applicable in wider scenarios.", + "bbox": [ + 75, + 265, + 472, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3D Keypoint Inference", + "text_level": 1, + "bbox": [ + 76, + 786, + 287, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to 2D keypoints detection, 3D keypoints detection is an under-explored topic, particularly in unsupervised settings without strong priors.", + "bbox": [ + 75, + 809, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised methods. Some earlier methods [3, 10, 68] use a two-stage strategy that infers 2D keypoints first, and then predicts depth for these keypoints. These methods", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not only need 3D keypoints or depth annotations for training, but are also sub-optimal due to their two-stage procedure. Tulyakov et al. [50] generalize a cascaded regression method to 3D for estimating 3D face landmarks in an end-to-end manner from images. Pavlakos et al. [41] generalize 2D heatmaps directly to 3D by inferring 3D keypoints, while Zhang et al. [66] improve this method via a joint voxel and coordinate regression to avoid the curse of dimensionality. However, these methods all needs 3D keypoints annotations, so that their scope of application is limited due to the absence of annotated 3D keypoints in many cases. Reddy et al. [44] propose a trifocal loss to learn (possibly occluded) 3D keypoints from multiple views by using 2D keypoints annotations. Closely related to 3D shape inference, sparse non-rigid structure-from-motion methods also take 2D keypoints as input to infer 3D structure of objects [8, 23, 38-40, 52, 64, 65]. Based on [12], He et al. [14] predict 3D keypoints using dozens of images with 2D annotations together with a skeleton template.", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised methods. In order to get rid of suboptimal two-stage approaches or 2D/3D keypoints annotations, some works aim at unsupervised 3D structure inference. Supasorn et al. [48] use paired views and known transformations between them as input to infer 3D keypoints. Chen et al. [5] use multiple views as input to infer 3D keypoints by averaging 2D keypoints and depth values detected in each view. Using a pair of views as input, Honari and Fua [15] propose a two-stream auto-encoder to encode 2D keypoints of each view, and then use triangulation to lift two sets of 2D keypoints to 3D. Various unsupervised 3D human pose estimation methods [25, 59, 69], either use multiple views or video as input, or they rely on some domain-specific losses as constraints.", + "bbox": [ + 496, + 378, + 892, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from all mentioned 3D keypoints inference settings, our method does not rely on 2D/3D keypoints annotations, 3D template/priors/skeletons, multiple views/video, or domain-specific 3D structure. Instead, for the first time we infer 3D structure only from category-specific collections of images for a broad range of object categories.", + "bbox": [ + 496, + 589, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 694, + 627, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method builds upon the recent AutoLink method [12] that addresses the task of unsupervised keypoint detection in 2D. In this section we give a brief overview of AutoLink.", + "bbox": [ + 496, + 719, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AutoLink comprises an auto-encoder with 2D keypoints as bottleneck latent embeddings. For training, it employs a reconstruction loss that compares the input image with a reconstructed image that is obtained based on edge maps and a randomly masked input image.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given an image $I \\in \\mathbb{R}^{W \\times H \\times 3}$ , the encoder takes $I$ as input and outputs $K$ heatmaps, denoted as $H_{1}, H_{2}, \\ldots, H_{K}$ , where $H_{i} \\in [0,1]^{W \\times H}$ . The $i$ -th heatmap represents the likelihood for each pixel to be the", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0ad3c2b3e0c004ec57e8380182823bf433d7ec03f3a63d3ad70b16ec9b91fd2f.jpg", + "image_caption": [ + "Figure 3. Overview of our pipeline. We sample images $I$ from a category-specific image collection, which are then piped into an encoder that predicts $K$ heatmaps. These heatmaps are used to extract a $K \\times 2$ keypoint matrix $P^{2D}$ . From $P^{2D}$ , an edge map $S$ is constructed, which is then, together with a randomly masked version of the input image $I_{\\mathrm{masked}}$ , fed into the decoder to reconstruct image $I'$ . Moreover, an MLP takes $P^{2D}$ as input and predicts with different heads, the coefficients $\\alpha$ , the 3D translation $T$ , and the 3D rotation $R$ . At the same time we optimize for a globally trainable mean shape $M$ and basis $B$ , which define a 3D shape space model used to represent our 3D keypoints, see Eq.(5). The main difference to AutoLink [12], which predicts 2D keypoints only, is that we infer 3D keypoints and a 3D shape space model." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "specified keypoint $i$ , where $K$ represents the total number of different keypoints. With that, the location of each 2D keypoint $P_{i}^{2\\mathrm{D}} = (x_{i},y_{i}), i = 1,2,\\ldots ,K$ can be represented as", + "bbox": [ + 75, + 395, + 468, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} ^ {\\mathrm {2 D}} = \\sum_ {p} \\frac {H _ {i} (p)}{\\sum_ {p ^ {\\prime}} H _ {i} \\left(p ^ {\\prime}\\right)} p, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 450, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p,p^{\\prime}\\in [-1,1]\\times [-1,1]$ are normalized pixel coordinates.", + "bbox": [ + 76, + 492, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each pair of 2D keypoints, $P_{i}^{2\\mathrm{D}}$ and $P_{j}^{2\\mathrm{D}}$ , $1 \\leq i,j \\leq K$ , the method constructs an edge map $S_{ij} \\in \\mathbb{R}^{H \\times W}$ defined on normalized pixel coordinates $p$ via", + "bbox": [ + 76, + 522, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i j} (p) = \\exp \\left(- d _ {i j} (p) / \\sigma^ {2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 577, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{ij}(p)$ is the $L_{2}$ distance of pixel $p$ to the line segment connecting $P_{i}^{2\\mathrm{D}}$ and $P_{j}^{2\\mathrm{D}}$ , and $\\sigma \\in \\mathbb{R}$ controls the thickness of the edge. The final edge map $S \\in \\mathbb{R}^{W \\times H}$ summarizing all edges between all keypoints is given by", + "bbox": [ + 76, + 602, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (p) = \\max _ {1 \\leq i, j \\leq K} \\omega_ {i j} S _ {i j} (p), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 672, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega_{ij}$ are learnable parameters that are shared by all images and optimized during training.", + "bbox": [ + 76, + 702, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the edge map $S$ and the (randomly masked) input image $I_{\\mathrm{masked}}$ are concatenated along the channel axis and fed into a decoder to reconstruct the input. The whole AutoLink pipeline is trained end-to-end only based on the reconstruction loss", + "bbox": [ + 76, + 732, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\| F (D (S \\odot I _ {\\text {m a s k e d}})) - F (I) \\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 816, + 468, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ is the decoder, $F$ is a pre-trained and fixed feature extractor (i.e. VGG network), $\\odot$ represents channel-wise concatenation, and $I_{\\mathrm{masked}}$ is the randomly masked original input image with $80\\%$ mask area.", + "bbox": [ + 76, + 839, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Unsupervised 3D Structure Inference", + "text_level": 1, + "bbox": [ + 498, + 393, + 834, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following we explain our method. To this end, we introduce our 3D shape space model for inferring 3D keypoint structure in Section 4.1. In section 4.2 we explain estimating the object pose. Section 4.3 introduces our unsupervised losses. Our overall pipeline is summarized in Fig. 3.", + "bbox": [ + 496, + 419, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Trainable 3D Shape Space Model", + "text_level": 1, + "bbox": [ + 498, + 501, + 790, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than working with high-dimensional and expensive 3D heatmaps, we infer 3D geometry by training a 2D keypoint predictor together with a 3D shape space model in an end-to-end manner, see Fig. 3. Our encoder first predicts 2D keypoint heatmaps from a given 2D image, which are then transformed into a 2D keypoint matrix using Eq. (1). Then a multi-layer perceptron (MLP) takes the 2D keypoint matrix as input to predict input-dependent coefficients $\\alpha$ of a 3D shape model. At the same time, the mean shape $M$ and the set of basis vectors $B$ of this 3D shape space model are optimized as global trainable parameters. Overall, this allows to obtain the 3D keypoints (in canonical pose) as", + "bbox": [ + 496, + 523, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {c a n o n i c a l}} ^ {\\mathrm {3 D}} = \\operatorname {m a t} (M + \\alpha B), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 710, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is the mean shape of size $1 \\times 3K$ , $B$ is the basis matrix of size $n \\times 3K$ and $\\alpha$ are the coefficients of size $1 \\times n$ , with $n$ denoting the number of basis functions. The operation $\\mathrm{mat}()$ reshapes a $1 \\times 3K$ row vector to a $K \\times 3$ matrix. By training these components simultaneously with our 2D heatmap predictor, we can ensure that during training our 2D heatmap predictor receives gradients that contain explicit information about 3D geometry.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We emphasize that opposed to most existing 3D shape space learning methods that require known keypoint locations, we simultaneously optimize for keypoint locations", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while using our trainable 3D shape space model as 3D geometry-aware regularizer. Unlike non-rigid structure-from-motion methods [8, 38-40] that utilize similar ideas, our approach does not require multiple images of the same object from different views. Instead, we use an inhomogeneous collection of category-specific images, where the 3D shape space model serves as flexible and adaptive shape prior that can adjust to the different geometries of objects (within the considered object category). As such, our approach is much more flexible, since it can handle single-view image collections as for example available in web collections.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 281, + 236, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the 3D keypoints $P_{\\mathrm{canonical}}^{\\mathrm{3D}}$ (in canonical pose) using Eq. (5), an input-dependent transformation matrix and scaling factor are estimated that transform $P_{\\mathrm{canonical}}^{\\mathrm{3D}}$ to the posed 3D keypoints $P^{3D}$ , such that the projection of $P^{3D}$ onto the 2D image plane coincides with the 2D structure of the object within the 2D image.", + "bbox": [ + 75, + 303, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-rigid object deformations are taken care of by the 3D shape model, so that we represent the object pose solely via a rigid body transformation matrix $(R,T)\\in \\mathrm{SE}(3)$ . With the 2D keypoint matrix $P^{2D}$ as input, an MLP computes the translation $T\\in \\mathbb{R}^{1\\times 3}$ and the rotation angles $\\Theta \\in \\mathbb{R}^{1\\times 3}$ , which are converted to a rotation matrix via the Rodrigues formula [30]:", + "bbox": [ + 75, + 395, + 470, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR = I _ {3} + \\frac {\\sin (\\theta)}{\\theta} [ \\Theta ] _ {\\times} + \\frac {1 - \\cos (\\theta)}{\\theta^ {2}} [ \\Theta ] _ {\\times} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 523, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta = \\| \\Theta \\|$ and the skew-symmetric operator $[\\cdot ]_{\\times}$ turns a vector in $\\mathbb{R}^{1\\times 3}$ into a skew-symmetric matrix. Finally the 3D keypoint matrix $P^{3D}$ is given as", + "bbox": [ + 75, + 561, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {3 \\mathrm {D}} = P _ {\\text {c a n o n i c a l}} ^ {3 \\mathrm {D}} R + \\mathbf {1} _ {K} T, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 618, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_K$ is a $K$ -dimensional vector of all ones.", + "bbox": [ + 76, + 647, + 398, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Unsupervised Losses", + "text_level": 1, + "bbox": [ + 76, + 671, + 272, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our re-projection loss penalizes the discrepancy between the projected 3D keypoints and the 2D heatmap keypoint predictions, i.e.", + "bbox": [ + 75, + 695, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p r o j}} = \\| P ^ {3 \\mathrm {D}} \\Pi - P ^ {2 \\mathrm {D}} \\| _ {F}, \\quad \\text {w i t h} \\quad \\Pi = \\left[ \\begin{array}{c c} s _ {1} & 0 \\\\ 0 & s _ {2} \\\\ 0 & 0 \\end{array} \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 750, + 468, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "being a simple orthographic projection (of $P^{3D}$ from Eq. (8)) combined with scaling $s = (s_1, s_2)$ , which is predicted by an MLP. This loss ensures consistent 2D keypoint predictions between different images, since the 3D shape structure is shared across the whole image collection via the 3D shape space model.", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, combining $\\mathcal{L}_{\\mathrm{rec}}$ and $\\mathcal{L}_{\\mathrm{proj}}$ alone will lead to degenerate solutions (i.e. all keypoints converge to a single location). In order to avoid this situation, we consider the repulsion loss", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e p}} = - \\sum_ {i = 1} ^ {K} \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} \\exp (- \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} / h), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 157, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ is a 2D keypoint (a row of $P^{2\\mathrm{D}}$ ), $\\mathcal{N}(p_i)$ is the nearest keypoint (other than $p_i$ ) of $p_i$ in $P^{2\\mathrm{D}}$ , and $h$ is a temperature parameter that controls the repulsion decrease rate. The effect of Eq. (9) is that neighboring keypoints do not move too close to each other.", + "bbox": [ + 496, + 212, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our total loss combines all terms and is defined as", + "bbox": [ + 517, + 287, + 852, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e p}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 310, + 890, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are weights of the individual terms.", + "bbox": [ + 496, + 334, + 823, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 361, + 633, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on various datasets and compare to existing keypoint prediction methods. Implementation details are provided in the supplementary document.", + "bbox": [ + 496, + 386, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 439, + 689, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on six datasets: CELEBA WILD [28], CUB-200-2011 [51], 300W-LP [71], AFLW2000-3D [71], HORSE [70] and AFHQ [6]. The CELEBA WILD dataset [28] contains celebrity faces in in-the-wild environments, along with 2D keypoint annotations for each image. Following the setting of He et al. [12], we remove the images whose face covers less than $30\\%$ of the area. The 300W-LP [71] dataset is a large synthetic face dataset containing 61,225 samples across large poses (1,786 from IBUG, 5,207 from AFW, 16,556 from LFPW and 37,676 from HELEN). The AFLW2000-3D dataset was introduced along with the 300W-LP dataset by the same data acquisition method, containing 2,000 facial images, each with 68 3D keypoints annotations. We train on the 300W-LP dataset and test on the AFLW2000-3D dataset to evaluate the quantitative 3D keypoint detection performance. We use subsets from the AFHQ dataset [6], namely CAT, DOG, TIGER, FOX, CHEETAH, WOLF, JAGUAR and LION, to get qualitative 3D structure inference results. CUB-200-2011 consists of 11,788 images of birds. Following the settings in [12, 29], we align all birds facing left, and remove seabirds. The HORSE dataset is extracted from the CycleGAN dataset [70] by removing images with multiple horses and aligning them to face left. CUB-200-2011, HORSE, AFHQ and CELEBA WILD are used to generate qualitative results. We quantitatively evaluate unsupervised 2D keypoint inference using the CELEBA WILD dataset and compare it with the recent state-of-the-art method by He et al. [12].", + "bbox": [ + 496, + 462, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/93cb6c61d32f561f61dfe16c14469edbb4f80ac17e3292d8fe053419ee327bbb.jpg", + "image_caption": [ + "Figure 4. Visualization of the 3D structure for different instances of the AFHQ dataset (CAT, DOG, FOX and TIGER subset in each row). Each 3D structure is shown in two forms: (i) 3D keypoints with linking edges between them (the edge thickness is proportional to the edge map coefficients $\\omega$ , which indicate their relative importance); (ii) Delaunay triangulation of the 3D keypoints rendered as shaded mesh. Different keypoints are shown in different colors to demonstrate consistency across different images (columns)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2b08cef81a9de29cca76b6e277f14809d743c8a15a112e845a5655263b8eadb.jpg", + "image_caption": [ + "Figure 5. Visualization of the 3D structure for different images from the HORSE and CUB-200-2011 datasets (rows)." + ], + "image_footnote": [], + "bbox": [ + 81, + 500, + 864, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. For unsupervised 2D keypoint inference, we follow the metrics used by He et al. [12], which is the normalized mean $L_{2}$ error (NME) of detected 2D keypoints aligned to the ground truth. For the 3D case, we generalize NME to 3D by calculating the normalized mean $L_{2}$ of detected 3D keypoints aligned to ground truth 3D keypoints. In both cases, the normalization is the ocular distances of each image.", + "bbox": [ + 75, + 731, + 472, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 729, + 684, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visualize the learned 3D structure on four subsets of AFHQ (Cat, Dog, Fox and Tiger; visualization of results of other subsets are in supplementary document), CUB-200-2011, and HORSE datasets in Fig. 4 and 5, respectively. From the visualizations we can directly recognize the 3D structure of each shown object category. For the four subsets of AFHQ dataset, despite these animal categories sharing many similarities, we can recognize the individual animal face characteristics from predicted 3D keypoints with", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10709", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "linking edges. Moreover, we find that the set of thick edges, i.e. the ones with large learned edge coefficient $\\omega$ , coincide with the most notable structures of each category, while negligible edges remain very thin. Finally, we demonstrate that keypoints are consistently detected among different instances of objects per category, which can be seen based on the consistency of the color of keypoints (both in 2D and 3D). More qualitative results and videos are included in the supplementary document and attached files.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting (cf. Table 1), as comparison we consider a simple baseline that first predicts 2D keypoints using AutoLink [12], and then lifts 2D keypoints to 3D by pinpointing learned 2D keypoints on the depth map (given by the pre-trained unsupervised monocular depth estimator MiDaS [42]). Fig. 6 shows a comparison on CELEBA WILD dataset.", + "bbox": [ + 75, + 231, + 472, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0981cd9de79f2f502192e66d557acf9c80fd59f00a18e2cf79b4899d710636a9.jpg", + "image_caption": [ + "Figure 6. Comparison of predicted 3D structures between our method (left) and AutoLink [12] combined with MiDaS [42] (right). Our method obtains more consistent 2D keypoint predictions and more realistic 3D structure." + ], + "image_footnote": [], + "bbox": [ + 76, + 373, + 467, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 736, + 272, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2D keypoint inference. First, we show that our method improves upon unsupervised 2D keypoint inference. Table 2 shows the NME values of our method and of various other unsupervised methods on the CELEBA WILD dataset. We also compare to the method by He et al. [12] for various choices of $K$ on the CELEBA WILD dataset to show the robustness of our method, as shown in Table 3. We can see that our method outperforms all competitors across all settings.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b4dcfd7d301c066b9664802ef902f27cb26a6d75692bf50e0e9bd330656950ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodK=8
DFF [7]31.30%*
SCOPS (w/o saliency) [17]22.11%†
SCOPS (w/saliency) [17]15.01%†
Liu et al. [27]12.26%†
Huang et al. [16]8.4%†
GANSeg [13]6.18%†
Thewlis et al. [49]31.30%*
Zhang et al. [67]40.82%*
LatentKeypointGAN [11]21.90%†
LatentKeypointGAN-tuned [11]5.63%†
Lorenz et al. [29]11.41%‡
IMM [19]8.74%‡
AutoLink [12]5.39%
Ours5.21%
", + "bbox": [ + 511, + 88, + 883, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/111c6b0b704dd0ee8be16722996fedc3cdb1c6b524012811f646efda75f79fa0.jpg", + "table_caption": [ + "Table 2. Normalized $L_{2}$ error (NME) for 2D keypoints inference of various unsupervised methods on CELEBA WILD datasets for $K = 8$ . (* reported from Collins et al. [7]; † reported from He et al. [12]; ‡ reported from Liu et al. [27])" + ], + "table_footnote": [], + "table_body": "
MethodK=8K=16K=24K=32
AutoLink [12]5.39%4.69%3.99%3.77%
Ours5.21%3.97%3.54%3.48%
", + "bbox": [ + 506, + 409, + 885, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d1c7fd5b453c2473f2e0b3682a8675f7dc925b99f9ea7f52d40f93ad375d14c8.jpg", + "table_caption": [ + "Table 3. Normalized $L_{2}$ error (NME) of AutoLink and our method for different numbers of keypoints using the CELEBA WILD dataset." + ], + "table_footnote": [], + "table_body": "
Supervised 3DDFAUnsupervised
AutoLink+Unsup3dAutoLink+MiDaSOurs
4.94%11.47%9.23%8.48%
", + "bbox": [ + 504, + 529, + 887, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Normalized $L_{2}$ error (NME) of our method, two unsupervised methods (AutoLink + MiDaS and AutoLink + Unsup3d) and one supervised method (3DDFA) for 3D keypoints inference (training on the 300W-LP dataset and testing on the AFLW2000-3D dataset).", + "bbox": [ + 498, + 590, + 890, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D keypoint inference. As stated in Sec. 5.2, there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting. So we define two unsupervised baselines by combining AutoLink [12] with MiDaS [42], and Autolink [12] with Unsup3d [54]. Moreover, we compare against the supervised method 3DDFA [71] to understand the gap between unsupervised and supervised methods. Table 4 summarises the NME results, in which our method achieves better performance compared to the two-stage baselines.", + "bbox": [ + 496, + 675, + 893, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 844, + 651, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we discuss our ablation studies. Results are shown in Fig. 7. The first and second row show the results", + "bbox": [ + 498, + 869, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10710", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with full training losses (i.e. Eq. (10)) with the number of keypoints being 32 and 16, respectively. We can see that with an increasing number of keypoints more detailed structure can be captured by our 3D keypoints. Furthermore, we remove $\\mathcal{L}_{\\mathrm{rep}}$ from our overall loss to train our framework (third row). The structure collapses into a degenerate keypoint configuration, which confirms the importance of $\\mathcal{L}_{\\mathrm{rep}}$ .", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A more detailed analysis on the behavior of $\\mathcal{L}_{\\mathrm{rep}}$ and on the influence of the number of basis functions $n$ of our shape model, are provided in the supplementary document.", + "bbox": [ + 75, + 198, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a695b4908032afebb8ae4149ae4dccf50bba929617161996c0adaef1275b137f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 261, + 467, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a47ce8d8b0192ae1201ffc992bcdad54407985e46cdfb9eec3728730d7a4472.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 335, + 465, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bca5f09139afcb2be47e22a310418022829def2841db0bd1641d921836046f4.jpg", + "image_caption": [ + "Figure 7. Ablation study on CELEBA WILD dataset. The first and second rows show results for a different number of keypoints (32 and 16). The third row shows results without repulsion loss $\\mathcal{L}_{\\mathrm{rep}}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 402, + 447, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 76, + 539, + 191, + 554 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2ba1e5d0f8a62ed02eea3d97d810c08cd08cfab4be9e549c9dc6e52525a3492.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 584, + 444, + 684 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e955484b80f02718e7796fa5f7ebde463e01e1a774206f5dcc0de8b42426761a.jpg", + "image_caption": [ + "Figure 8. Pose estimation visualization of two instances from CELEBA WILD dataset from two views." + ], + "image_footnote": [], + "bbox": [ + 91, + 702, + 442, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large pose estimation has always been an important yet challenging issue in many 3D shape inference settings", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[26, 33, 55, 56, 61, 63]. Recent works have found that when the pose variations falls into a small range, it can be learned directly by a network together with shape [53]. We observe similar results in our setting. For the CELEBA WILD, AFHQ (instances with moderate pose variations), CUB-200-2011 and HORSE (aligned facing direction) datasets, Fig. 4 and 5 have shown multiple instances in different poses with correct pose estimation. Fig. 8 gives more pose estimation results. However, for datasets consisting of instances with large pose variations, our method may fail to optimize for the correct poses. For example, our test on the HUMAN3.6M dataset [18] shows that, even though our model learns plausible human body structure, it fails to optimize for the correct pose of each instance, which also leads to incorrect instance-specific deformations, as shown in Fig. 9 – for such cases with strong articulation, domain-specific approaches that exploit domain knowledge (e.g. about the human kinematic skeleton structure) are at the time being more suitable. Yet, inferring such articulated objects is a highly relevant and interesting direction for future work.", + "bbox": [ + 496, + 90, + 890, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a865d2682ea8093bd3b5787e2fe64aa85dca0a3df57dd9242a0bede5b795d30.jpg", + "image_caption": [ + "Figure 9. Incorrect pose estimation visualization of HUMAN3.6M dataset. For severely articulated objects additional domain knowledge is currently needed (e.g. a human body kinematic skeleton)." + ], + "image_footnote": [], + "bbox": [ + 506, + 421, + 854, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, our method is the first unsupervised approach that can infer 3D keypoints from category-specific image collections without the need of any additional prior information. Conceptually, our main novelty lies in simultaneously learning 3D keypoints along with a 3D shape space model from an inhomogeneous collection of category-specific images (e.g. opposed to homogeneous collections of multiview images or videos, which show the same object from different views or in different poses). We experimentally demonstrate that our method not only improves upon 2D keypoint accuracy, but also infers plausible 3D structure for various object categories. We expect our work to be useful for diverse downstream tasks, for example related to 3D reconstruction or 3D shape-based generative models.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3", + "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 3", + "[3] Adrian Bulat, Tzimiropoulos, and Georgios. Two-stage convolutional part heatmap regression for the 1st 3d face alignment in the wild (3dfaw) challenge. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 616-624. Springer, 2016. 2, 3", + "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 3", + "[5] Boyuan Chen, Pieter Abbeel, and Deepak Pathak. Unsupervised learning of visual 3d keypoints for control. In International Conference on Machine Learning, pages 1539-1549. PMLR, 2021. 2, 3", + "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5", + "[7] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 7", + "[8] Hui Deng, Tong Zhang, Yuchao Dai, Jiawei Shi, Yiran Zhong, and Hongdong Li. Deep non-rigid structure-from-motion: A sequence-to-sequence translation perspective. arXiv preprint arXiv:2204.04730, 2022. 3, 5", + "[9] Aysegul Dundar, Kevin J Shih, Animesh Garg, Robert Pottorf, Anrew Tao, and Bryan Catanzaro. Unsupervised disentanglement of pose, appearance and background from images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3883-3894, 2021. 1, 2, 3", + "[10] Chao Gou, Yue Wu, Fei-Yue Wang, and Qiang Ji. Shape augmented regression for 3d face alignment. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 604-615. Springer, 2016. 2, 3", + "[11] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. arXiv preprint arXiv:2103.15812, 2021. 2, 3, 7", + "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. Advances in Neural Information Processing Systems, 35:36123-36141, 2022. 1, 2, 3, 4, 5, 6, 7", + "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1225-1235, 2022. 3, 7", + "[14] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21337-21348, 2023. 1, 2, 3", + "[15] Sina Honari and Pascal Fua. Unsupervised 3d keypoint estimation with multi-view geometry. arXiv preprint arXiv:2211.12829, 2022. 2, 3", + "[16] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662-8672, 2020. 7", + "[17] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 7", + "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3, 8", + "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3, 7", + "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3", + "[21] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 196-214. Springer, 2020. 3", + "[22] Yunji Kim, Seonghyeon Nam, In Cho, and Seon Joo Kim. Unsupervised keypoint learning for guiding class-conditional video prediction. Advances in neural information processing systems, 32, 2019. 3", + "[23] Chen Kong and Simon Lucey. Deep non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1558-1567, 2019. 3", + "[24] Tejas D Kulkarni, Ankush Gupta, Catalin Ionescu, Sebastian Borgeaud, Malcolm Reynolds, Andrew Zisserman, and Volodymyr Mnih. Unsupervised learning of object keypoints for perception and control. Advances in neural information processing systems, 32, 2019. 3", + "[25] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6152-6162, 2020. 1, 3" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 8", + "[27] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Unsupervised part segmentation through disentangling appearance and shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8355-8364, 2021. 7", + "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5", + "[29] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2019. 5, 7", + "[30] Y Ma, S Soatto, J Kosecka, and S Sastry. An invitation to 3d vision: From images to models. Springer verlag, 2003. 5", + "[31] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3", + "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3", + "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 8", + "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 1", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[36] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 1, 3", + "[37] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 2", + "[38] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression: A flexible alignment-based framework for non-rigid structure estimation. IEEE Transactions on Image Processing, 27(1):249-264, 2017.", + "[40] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression networks: Learning 3d structure of non-rigid objects from 2d annotations. In European Conference on Computer Vision, pages 1-18. Springer, 2020. 2, 3, 5", + "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2, 3", + "[42] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 7", + "[43] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 3", + "[44] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7326-7335, 2019. 2, 3", + "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1", + "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 3", + "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1", + "[48] Suwajanakorn Supasorn, Noah Snavely, Jonathan J Tompson, and Mohammad Norouzi. Discovery of latent 3d keypoints via end-to-end geometric reasoning. Advances in neural information processing systems, 31, 2018. 2, 3", + "[49] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3, 7", + "[50] Sergey Tulyakov, László A Jeni, Jeffrey F Cohn, and Nicu Sebe. Viewpoint-consistent 3d face alignment. IEEE transactions on pattern analysis and machine intelligence, 40(9): 2250-2264, 2017. 2, 3" + ], + "bbox": [ + 501, + 90, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10713", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[52] Chaoyang Wang, Chen-Hsuan Lin, and Simon Lucey. Deep nrsfm++: Towards unsupervised 2d-3d lifting in the wild. In 2020 International Conference on 3D Vision (3DV), pages 12–22. IEEE, 2020. 3", + "[53] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 8", + "[54] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2, 7", + "[55] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, pages 1-12, 2023. 1, 2, 8", + "[56] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 1, 2, 8", + "[57] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 3", + "[58] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 1", + "[59] Chenxin Xu, Siheng Chen, Maosen Li, and Ya Zhang. Invariant teacher and equivariant student for unsupervised 3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3013-3021, 2021. 1, 2, 3", + "[60] Yinghao Xu, Ceyuan Yang, Ziwei Liu, Bo Dai, and Bolei Zhou. Unsupervised landmark learning from unpaired data. arXiv preprint arXiv:2007.01053, 2020. 3", + "[61] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 8", + "[62] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 1", + "[63] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1323-1330. IEEE, 2021. 8" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Haitian Zeng, Yuchao Dai, Xin Yu, Xiaohan Wang, and Yi Yang. Pr-rrn: pairwise-regularized residual-recursive networks for non-rigid structure-from-motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5600–5609, 2021. 3", + "[65] Haitian Zeng, Xin Yu, Jiaxu Miao, and Yi Yang. Mhr-net: Multiple-hypothesis reconstruction of non-rigid shapes from 2d views. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 3", + "[66] Hongwen Zhang, Qi Li, and Zhenan Sun. Joint voxel and coordinate regression for accurate 3d facial landmark localization. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2202-2208. IEEE, 2018. 1, 2, 3", + "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3, 7", + "[68] Ruiqi Zhao, Yan Wang, C Fabian Benitez-Quiroz, Yaojie Liu, and Aleix M Martinez. Fast and precise face alignment and 3d shape reconstruction from a single 2d image. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 590-603. Springer, 2016. 2, 3", + "[69] Shihao Zhou, Mengxi Jiang, Qicong Wang, and Yunqi Lei. Towards locality similarity preserving to 3d human pose estimation. In Proceedings of the Asian Conference on Computer Vision, 2020. 1, 2, 3", + "[70] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5", + "[71] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 613 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10714", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_model.json b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ac9b5f155760ebfeb6578687f7001facf0795c5 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "Unsupervised 3D Structure Inference from Category-Specific Image Collections" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.181, + 0.368, + 0.201 + ], + "angle": 0, + "content": "Weikang Wang" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.182, + 0.544, + 0.199 + ], + "angle": 0, + "content": "Dongliang Cao" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.182, + 0.728, + 0.198 + ], + "angle": 0, + "content": "Florian Bernard" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.206, + 0.605, + 0.224 + ], + "angle": 0, + "content": "University of Bonn, Germany" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.28, + 0.884, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 1. Our method detects consistent and reasonable 3D structure for general object categories from a single 2D image. For training, our method only requires a collection of images from the same category without any additional information." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.489, + 0.314, + 0.505 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.474, + 0.812 + ], + "angle": 0, + "content": "Understanding 3D object structure from image collections of general object categories remains a long-standing challenge in computer vision. Due to the high relevance of image keypoints (e.g. for graph matching, controlling generative models, scene understanding, etc.), in this work we specifically focus on inferring 3D structure in terms of sparse keypoints. Existing 3D keypoint inference approaches rely on strong priors, such as spatio-temporal consistency, multi-view images of the same object, 3D shape priors (e.g. templates, skeleton), or supervisory signals e.g. in the form of 2D keypoint annotations. In contrast, we propose the first unsupervised 3D keypoint inference approach that can be trained for general object categories solely from an inhomogeneous image collection (containing different instances of objects from the same category). Our experiments show that our method not only improves upon unsupervised 2D keypoint inference, but more importantly, it also produces reasonable 3D structure for various object categories, both qualitatively and quantitatively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.872, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Understanding 3D object structure from 2D images is a fundamental problem in computer vision that has been studied" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.657 + ], + "angle": 0, + "content": "for decades. A range of different 3D structure representations have been considered, including keypoints [14, 25, 59, 69], meshes [55, 56, 62], and voxels [35, 45, 47]. Among these, keypoint-based 3D structure representations have the strong advantage that they are simple and flexible, while at the same time being highly relevant for diverse downstream tasks. However, obtaining 3D keypoints from 2D images is hard due to unknown depth, pose variations, different appearance, and the lack of explicit geometric information. Due to these difficulties, prior works mostly focus on 2D keypoints detection instead of 3D [9, 12, 31, 36, 66]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.896, + 0.903 + ], + "angle": 0, + "content": "In this work we investigate whether it is possible to directly obtain 3D keypoints without using any explicit supervision. Our work is motivated by the fact that 2D image keypoints are observations of points in 3D space that are projected onto the 2D image plane. By explicitly understanding and inferring 3D geometry we are able to improve upon the quality of detected keypoints. Specifically, we experimentally demonstrate that projecting our predicted 3D keypoints onto the image plane leads to more reliable 2D keypoints compared to existing 2D keypoint predictors. Moreover, our method can serve the increasing amount of tasks that rely on 3D keypoints, e.g. 3D structure-aware generative models [58], 3D avatar generation [34], or 3D reconstruction [55, 56]. Despite keypoint prediction being a well-studied problem in the 2D domain, 3D keypoint inference is under-explored and we thus specifically contribute" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.197, + 0.105 + ], + "angle": 0, + "content": "towards this area." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.108, + 0.474, + 0.471 + ], + "angle": 0, + "content": "Existing works on inferring 3D keypoints either use 3D point clouds or 2D images/videos from the same object instance as input data. While the former is conceptually a subset selection problem, in this work we solely focus on the more challenging latter case. Manually obtaining 3D keypoints in image data is much harder than annotating 2D keypoints. Thus, there are only very few datasets with 3D keypoints annotations, which in turn motivates unsupervised learning strategies that do not require expensive or even infeasible 3D annotations for training. Existing works that address 3D keypoint inference from image data use for example 2D keypoint annotations [38, 44], multiple views with known poses [11, 35, 48], spatial-temporal consistency in videos [9, 15], 2D/3D skeleton [55, 56] or templates [14]. In stark contrast, in this work we relax the assumptions about the availability of such supervisory signals and train a 3D keypoint predictor merely from category-specific image collections as input. Our method builds upon the assumption that object instances from the same category should have a similar 3D structure with moderate intra-class variations (e.g. faces, horses, birds). Table 1 summaries the differences among existing methods and Fig. 2 gives a visualization of some additional information within these methods and the comparison to our setting." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.485, + 0.47, + 0.817 + ], + "angle": 0, + "content": "
MethodsUnsup.w/o PriorCat.-agnostic
Bulat et al. [3]X✓✓
Gou et al. [10]X✓✓
Zhao et al. [68]X✓✓
Tulyakov et al. [50]X✓✓
Pavlakos et al. [41]X✓X
Zhang et al. [66]X✓X
He et al. [14]XX✓
Mildenhall et al. [35]✓X✓
Wu et al. [54]✓X✓
Moniz et al. [37]✓XX
Supasorn et al. [48]✓X✓
Novotny et al. [38]✓X✓
Park et al. [40]✓X✓
Chen et al. [5]✓X✓
Reddy et al. [44]✓X✓
Zhou et al. [69]✓✓X
Dundar et al. [9]✓✓X
Xu et al. [59]✓✓X
Honari and Fua [15]✓X✓
Ours✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.898 + ], + "angle": 0, + "content": "Table 1. In comparison to existing 3D keypoints inference methods, our approach combines a unique set of desirable properties: it can be trained in an unsupervised manner without requiring any additional information; it does not require any shape prior and can thus be applied to different object categories." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.358, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 2. (Top row from left to right) Existing methods require different additional information for 3D structure inference, e.g. multiple views (from [12]), 2D keypoint annotations (from [12]), and a symmetry prior (from [54]). (Bottom row) Our setting does not require such information but instead also infers a 3D shape space model that is learned in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.893, + 0.657 + ], + "angle": 0, + "content": "Our method enables to infer 3D keypoints from category-specific image collections without relying on additional prior information. To this end, we build upon the recent unsupervised 2D keypoint detection method by He et al. [12] and extend it by additionally learning a sparse 3D shape space model to represent 3D structure. To regularise the 3D structure, a self-supervised re-projection loss is introduced to couple 3D keypoints with corresponding 2D keypoints. By combining our 3D shape space model, re-projection loss, and additional regularisation loss terms, we demonstrate that our proposed method not only improves 2D keypoint accuracy, but also produces reasonable 3D structure. We summarise our main contributions as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.666, + 0.892, + 0.711 + ], + "angle": 0, + "content": "- For the first time we obtain 3D keypoints from general category-specific image collections without relying on additional prior information." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.712, + 0.892, + 0.756 + ], + "angle": 0, + "content": "- Our method obtains reasonable 3D keypoints based on our learnable 3D shape space model that is fused with a state-of-the-art unsupervised 2D keypoint detector." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.757, + 0.892, + 0.802 + ], + "angle": 0, + "content": "- We demonstrate that our method both improves unsupervised 2D keypoint detection accuracy and obtains reasonable 3D shape on different benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.666, + 0.892, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.838, + 0.642, + 0.853 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the following we summarize relevant 2D and 3D key-point prediction methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10705" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.289, + 0.108 + ], + "angle": 0, + "content": "2.1. 2D Keypoint Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Supervised methods. Various domain-specific supervised 2D keypoint detection methods have been proposed that are trained on a large amount of labeled datasets, e.g. for faces [2, 57, 71], human bodies [1, 4, 18, 21, 32], or vehicles [43], etc. However, labeling keypoints for images is time-consuming, laborious and error-prone, especially when there are occlusions, extreme pose variations, or appearance changes. As a consequence, labeling errors will subsequently influence network training and downstream task performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.473, + 0.78 + ], + "angle": 0, + "content": "Unsupervised methods. Due to these reasons, an increasing amount of attention is paid to unsupervised methods. One popular concept is multi-view geometric consistency. Thewlis et al. [49] detect keypoints by recovering transformations between one image and its deformed counterpart from pre-defined deformations. Without relying on pre-defined deformations, various video-based methods [9, 19, 20, 22, 24, 36, 46] infer keypoints based on a temporal coherence prior. Instead of using multiple views as input, Xu et al. [60] use unpaired input to disentangle appearance and poses by conducting a swapping-reconstruction strategy and encouraging reconstructed images to resemble the original. He et al. [13] and He et al. [11] use a generative adversarial training framework to constrain keypoint locations by interpreting them as latent codes for an image generator. However, collaboratively learning intermediate keypoints and complex adversarial models is unstable and does not scale. Starting from some initial point sets, Mallis et al. [31] filter keypoints by optimizing alternatively between clustering of keypoints features across a dataset, and training a keypoint detector with pseudo-labels offered by cluster centers. However, this method depends heavily on the initial point set. Zhang et al. [67] address single-image keypoint inference using an auto-encoder architecture to interpret keypoints as a latent representation, while reconstructing the input image from the concatenated keypoints and respective pixel feature descriptors. He et al. [12] use a similar idea by representing keypoints as latent features in an auto-encoder manner. However, they reconstruct the input from an edge map created from keypoints combined with random masking of appearance information. In this way, paired input data is not needed to disentangle appearance and poses, which makes this method applicable in wider scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.288, + 0.804 + ], + "angle": 0, + "content": "2.2.3D Keypoint Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Compared to 2D keypoints detection, 3D keypoints detection is an under-explored topic, particularly in unsupervised settings without strong priors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Supervised methods. Some earlier methods [3, 10, 68] use a two-stage strategy that infers 2D keypoints first, and then predicts depth for these keypoints. These methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.379 + ], + "angle": 0, + "content": "not only need 3D keypoints or depth annotations for training, but are also sub-optimal due to their two-stage procedure. Tulyakov et al. [50] generalize a cascaded regression method to 3D for estimating 3D face landmarks in an end-to-end manner from images. Pavlakos et al. [41] generalize 2D heatmaps directly to 3D by inferring 3D keypoints, while Zhang et al. [66] improve this method via a joint voxel and coordinate regression to avoid the curse of dimensionality. However, these methods all needs 3D keypoints annotations, so that their scope of application is limited due to the absence of annotated 3D keypoints in many cases. Reddy et al. [44] propose a trifocal loss to learn (possibly occluded) 3D keypoints from multiple views by using 2D keypoints annotations. Closely related to 3D shape inference, sparse non-rigid structure-from-motion methods also take 2D keypoints as input to infer 3D structure of objects [8, 23, 38-40, 52, 64, 65]. Based on [12], He et al. [14] predict 3D keypoints using dozens of images with 2D annotations together with a skeleton template." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.893, + 0.59 + ], + "angle": 0, + "content": "Unsupervised methods. In order to get rid of suboptimal two-stage approaches or 2D/3D keypoints annotations, some works aim at unsupervised 3D structure inference. Supasorn et al. [48] use paired views and known transformations between them as input to infer 3D keypoints. Chen et al. [5] use multiple views as input to infer 3D keypoints by averaging 2D keypoints and depth values detected in each view. Using a pair of views as input, Honari and Fua [15] propose a two-stream auto-encoder to encode 2D keypoints of each view, and then use triangulation to lift two sets of 2D keypoints to 3D. Various unsupervised 3D human pose estimation methods [25, 59, 69], either use multiple views or video as input, or they rely on some domain-specific losses as constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Different from all mentioned 3D keypoints inference settings, our method does not rely on 2D/3D keypoints annotations, 3D template/priors/skeletons, multiple views/video, or domain-specific 3D structure. Instead, for the first time we infer 3D structure only from category-specific collections of images for a broad range of object categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.628, + 0.712 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Our method builds upon the recent AutoLink method [12] that addresses the task of unsupervised keypoint detection in 2D. In this section we give a brief overview of AutoLink." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.84 + ], + "angle": 0, + "content": "AutoLink comprises an auto-encoder with 2D keypoints as bottleneck latent embeddings. For training, it employs a reconstruction loss that compares the input image with a reconstructed image that is obtained based on edge maps and a randomly masked input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Formally, given an image \\(I \\in \\mathbb{R}^{W \\times H \\times 3}\\), the encoder takes \\(I\\) as input and outputs \\(K\\) heatmaps, denoted as \\(H_{1}, H_{2}, \\ldots, H_{K}\\), where \\(H_{i} \\in [0,1]^{W \\times H}\\). The \\(i\\)-th heatmap represents the likelihood for each pixel to be the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10706" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.892, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Figure 3. Overview of our pipeline. We sample images \\(I\\) from a category-specific image collection, which are then piped into an encoder that predicts \\(K\\) heatmaps. These heatmaps are used to extract a \\(K \\times 2\\) keypoint matrix \\(P^{2D}\\). From \\(P^{2D}\\), an edge map \\(S\\) is constructed, which is then, together with a randomly masked version of the input image \\(I_{\\mathrm{masked}}\\), fed into the decoder to reconstruct image \\(I'\\). Moreover, an MLP takes \\(P^{2D}\\) as input and predicts with different heads, the coefficients \\(\\alpha\\), the 3D translation \\(T\\), and the 3D rotation \\(R\\). At the same time we optimize for a globally trainable mean shape \\(M\\) and basis \\(B\\), which define a 3D shape space model used to represent our 3D keypoints, see Eq.(5). The main difference to AutoLink [12], which predicts 2D keypoints only, is that we infer 3D keypoints and a 3D shape space model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.454 + ], + "angle": 0, + "content": "specified keypoint \\(i\\), where \\(K\\) represents the total number of different keypoints. With that, the location of each 2D keypoint \\(P_{i}^{2\\mathrm{D}} = (x_{i},y_{i}), i = 1,2,\\ldots ,K\\) can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.452, + 0.469, + 0.489 + ], + "angle": 0, + "content": "\\[\nP _ {i} ^ {\\mathrm {2 D}} = \\sum_ {p} \\frac {H _ {i} (p)}{\\sum_ {p ^ {\\prime}} H _ {i} \\left(p ^ {\\prime}\\right)} p, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.469, + 0.522 + ], + "angle": 0, + "content": "where \\(p,p^{\\prime}\\in [-1,1]\\times [-1,1]\\) are normalized pixel coordinates." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.523, + 0.469, + 0.571 + ], + "angle": 0, + "content": "For each pair of 2D keypoints, \\( P_{i}^{2\\mathrm{D}} \\) and \\( P_{j}^{2\\mathrm{D}} \\), \\( 1 \\leq i,j \\leq K \\), the method constructs an edge map \\( S_{ij} \\in \\mathbb{R}^{H \\times W} \\) defined on normalized pixel coordinates \\( p \\) via" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.578, + 0.469, + 0.596 + ], + "angle": 0, + "content": "\\[\nS _ {i j} (p) = \\exp \\left(- d _ {i j} (p) / \\sigma^ {2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.603, + 0.469, + 0.665 + ], + "angle": 0, + "content": "where \\(d_{ij}(p)\\) is the \\(L_{2}\\) distance of pixel \\(p\\) to the line segment connecting \\(P_{i}^{2\\mathrm{D}}\\) and \\(P_{j}^{2\\mathrm{D}}\\), and \\(\\sigma \\in \\mathbb{R}\\) controls the thickness of the edge. The final edge map \\(S \\in \\mathbb{R}^{W \\times H}\\) summarizing all edges between all keypoints is given by" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.673, + 0.469, + 0.697 + ], + "angle": 0, + "content": "\\[\nS (p) = \\max _ {1 \\leq i, j \\leq K} \\omega_ {i j} S _ {i j} (p), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.469, + 0.733 + ], + "angle": 0, + "content": "where \\(\\omega_{ij}\\) are learnable parameters that are shared by all images and optimized during training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.469, + 0.807 + ], + "angle": 0, + "content": "Finally, the edge map \\( S \\) and the (randomly masked) input image \\( I_{\\mathrm{masked}} \\) are concatenated along the channel axis and fed into a decoder to reconstruct the input. The whole AutoLink pipeline is trained end-to-end only based on the reconstruction loss" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.817, + 0.469, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\| F (D (S \\odot I _ {\\text {m a s k e d}})) - F (I) \\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\(D\\) is the decoder, \\(F\\) is a pre-trained and fixed feature extractor (i.e. VGG network), \\(\\odot\\) represents channel-wise concatenation, and \\(I_{\\mathrm{masked}}\\) is the randomly masked original input image with \\(80\\%\\) mask area." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.394, + 0.836, + 0.411 + ], + "angle": 0, + "content": "4. Unsupervised 3D Structure Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.495 + ], + "angle": 0, + "content": "In the following we explain our method. To this end, we introduce our 3D shape space model for inferring 3D keypoint structure in Section 4.1. In section 4.2 we explain estimating the object pose. Section 4.3 introduces our unsupervised losses. Our overall pipeline is summarized in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.502, + 0.792, + 0.518 + ], + "angle": 0, + "content": "4.1. Trainable 3D Shape Space Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Rather than working with high-dimensional and expensive 3D heatmaps, we infer 3D geometry by training a 2D keypoint predictor together with a 3D shape space model in an end-to-end manner, see Fig. 3. Our encoder first predicts 2D keypoint heatmaps from a given 2D image, which are then transformed into a 2D keypoint matrix using Eq. (1). Then a multi-layer perceptron (MLP) takes the 2D keypoint matrix as input to predict input-dependent coefficients \\(\\alpha\\) of a 3D shape model. At the same time, the mean shape \\(M\\) and the set of basis vectors \\(B\\) of this 3D shape space model are optimized as global trainable parameters. Overall, this allows to obtain the 3D keypoints (in canonical pose) as" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.712, + 0.891, + 0.73 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {c a n o n i c a l}} ^ {\\mathrm {3 D}} = \\operatorname {m a t} (M + \\alpha B), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.855 + ], + "angle": 0, + "content": "where \\( M \\) is the mean shape of size \\( 1 \\times 3K \\), \\( B \\) is the basis matrix of size \\( n \\times 3K \\) and \\( \\alpha \\) are the coefficients of size \\( 1 \\times n \\), with \\( n \\) denoting the number of basis functions. The operation \\( \\mathrm{mat}() \\) reshapes a \\( 1 \\times 3K \\) row vector to a \\( K \\times 3 \\) matrix. By training these components simultaneously with our 2D heatmap predictor, we can ensure that during training our 2D heatmap predictor receives gradients that contain explicit information about 3D geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We emphasize that opposed to most existing 3D shape space learning methods that require known keypoint locations, we simultaneously optimize for keypoint locations" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "while using our trainable 3D shape space model as 3D geometry-aware regularizer. Unlike non-rigid structure-from-motion methods [8, 38-40] that utilize similar ideas, our approach does not require multiple images of the same object from different views. Instead, we use an inhomogeneous collection of category-specific images, where the 3D shape space model serves as flexible and adaptive shape prior that can adjust to the different geometries of objects (within the considered object category). As such, our approach is much more flexible, since it can handle single-view image collections as for example available in web collections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.238, + 0.296 + ], + "angle": 0, + "content": "4.2. Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.395 + ], + "angle": 0, + "content": "After getting the 3D keypoints \\( P_{\\mathrm{canonical}}^{\\mathrm{3D}} \\) (in canonical pose) using Eq. (5), an input-dependent transformation matrix and scaling factor are estimated that transform \\( P_{\\mathrm{canonical}}^{\\mathrm{3D}} \\) to the posed 3D keypoints \\( P^{3D} \\), such that the projection of \\( P^{3D} \\) onto the 2D image plane coincides with the 2D structure of the object within the 2D image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.472, + 0.502 + ], + "angle": 0, + "content": "Non-rigid object deformations are taken care of by the 3D shape model, so that we represent the object pose solely via a rigid body transformation matrix \\((R,T)\\in \\mathrm{SE}(3)\\). With the 2D keypoint matrix \\(P^{2D}\\) as input, an MLP computes the translation \\(T\\in \\mathbb{R}^{1\\times 3}\\) and the rotation angles \\(\\Theta \\in \\mathbb{R}^{1\\times 3}\\), which are converted to a rotation matrix via the Rodrigues formula [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.524, + 0.47, + 0.556 + ], + "angle": 0, + "content": "\\[\nR = I _ {3} + \\frac {\\sin (\\theta)}{\\theta} [ \\Theta ] _ {\\times} + \\frac {1 - \\cos (\\theta)}{\\theta^ {2}} [ \\Theta ] _ {\\times} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.61 + ], + "angle": 0, + "content": "where \\(\\theta = \\| \\Theta \\|\\) and the skew-symmetric operator \\([\\cdot ]_{\\times}\\) turns a vector in \\(\\mathbb{R}^{1\\times 3}\\) into a skew-symmetric matrix. Finally the 3D keypoint matrix \\(P^{3D}\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.619, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\nP ^ {3 \\mathrm {D}} = P _ {\\text {c a n o n i c a l}} ^ {3 \\mathrm {D}} R + \\mathbf {1} _ {K} T, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.648, + 0.4, + 0.663 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_K\\) is a \\(K\\)-dimensional vector of all ones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.274, + 0.687 + ], + "angle": 0, + "content": "4.3. Unsupervised Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Our re-projection loss penalizes the discrepancy between the projected 3D keypoints and the 2D heatmap keypoint predictions, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.75, + 0.469, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p r o j}} = \\| P ^ {3 \\mathrm {D}} \\Pi - P ^ {2 \\mathrm {D}} \\| _ {F}, \\quad \\text {w i t h} \\quad \\Pi = \\left[ \\begin{array}{c c} s _ {1} & 0 \\\\ 0 & s _ {2} \\\\ 0 & 0 \\end{array} \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "being a simple orthographic projection (of \\( P^{3D} \\) from Eq. (8)) combined with scaling \\( s = (s_1, s_2) \\), which is predicted by an MLP. This loss ensures consistent 2D keypoint predictions between different images, since the 3D shape structure is shared across the whole image collection via the 3D shape space model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "However, combining \\(\\mathcal{L}_{\\mathrm{rec}}\\) and \\(\\mathcal{L}_{\\mathrm{proj}}\\) alone will lead to degenerate solutions (i.e. all keypoints converge to a single location). In order to avoid this situation, we consider the repulsion loss" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.159, + 0.892, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e p}} = - \\sum_ {i = 1} ^ {K} \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} \\exp (- \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} / h), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.287 + ], + "angle": 0, + "content": "where \\( p_i \\) is a 2D keypoint (a row of \\( P^{2\\mathrm{D}} \\)), \\( \\mathcal{N}(p_i) \\) is the nearest keypoint (other than \\( p_i \\)) of \\( p_i \\) in \\( P^{2\\mathrm{D}} \\), and \\( h \\) is a temperature parameter that controls the repulsion decrease rate. The effect of Eq. (9) is that neighboring keypoints do not move too close to each other." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.289, + 0.854, + 0.303 + ], + "angle": 0, + "content": "Our total loss combines all terms and is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.311, + 0.892, + 0.329 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e p}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.825, + 0.351 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are weights of the individual terms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.634, + 0.379 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.433 + ], + "angle": 0, + "content": "We evaluate our method on various datasets and compare to existing keypoint prediction methods. Implementation details are provided in the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.691, + 0.457 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on six datasets: CELEBA WILD [28], CUB-200-2011 [51], 300W-LP [71], AFLW2000-3D [71], HORSE [70] and AFHQ [6]. The CELEBA WILD dataset [28] contains celebrity faces in in-the-wild environments, along with 2D keypoint annotations for each image. Following the setting of He et al. [12], we remove the images whose face covers less than \\(30\\%\\) of the area. The 300W-LP [71] dataset is a large synthetic face dataset containing 61,225 samples across large poses (1,786 from IBUG, 5,207 from AFW, 16,556 from LFPW and 37,676 from HELEN). The AFLW2000-3D dataset was introduced along with the 300W-LP dataset by the same data acquisition method, containing 2,000 facial images, each with 68 3D keypoints annotations. We train on the 300W-LP dataset and test on the AFLW2000-3D dataset to evaluate the quantitative 3D keypoint detection performance. We use subsets from the AFHQ dataset [6], namely CAT, DOG, TIGER, FOX, CHEETAH, WOLF, JAGUAR and LION, to get qualitative 3D structure inference results. CUB-200-2011 consists of 11,788 images of birds. Following the settings in [12, 29], we align all birds facing left, and remove seabirds. The HORSE dataset is extracted from the CycleGAN dataset [70] by removing images with multiple horses and aligning them to face left. CUB-200-2011, HORSE, AFHQ and CELEBA WILD are used to generate qualitative results. We quantitatively evaluate unsupervised 2D keypoint inference using the CELEBA WILD dataset and compare it with the recent state-of-the-art method by He et al. [12]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10708" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.896, + 0.486 + ], + "angle": 0, + "content": "Figure 4. Visualization of the 3D structure for different instances of the AFHQ dataset (CAT, DOG, FOX and TIGER subset in each row). Each 3D structure is shown in two forms: (i) 3D keypoints with linking edges between them (the edge thickness is proportional to the edge map coefficients \\(\\omega\\), which indicate their relative importance); (ii) Delaunay triangulation of the 3D keypoints rendered as shaded mesh. Different keypoints are shown in different colors to demonstrate consistency across different images (columns)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.501, + 0.866, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.691, + 0.837, + 0.707 + ], + "angle": 0, + "content": "Figure 5. Visualization of the 3D structure for different images from the HORSE and CUB-200-2011 datasets (rows)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.473, + 0.853 + ], + "angle": 0, + "content": "Metrics. For unsupervised 2D keypoint inference, we follow the metrics used by He et al. [12], which is the normalized mean \\( L_{2} \\) error (NME) of detected 2D keypoints aligned to the ground truth. For the 3D case, we generalize NME to 3D by calculating the normalized mean \\( L_{2} \\) of detected 3D keypoints aligned to ground truth 3D keypoints. In both cases, the normalization is the ocular distances of each image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.731, + 0.685, + 0.748 + ], + "angle": 0, + "content": "5.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We visualize the learned 3D structure on four subsets of AFHQ (Cat, Dog, Fox and Tiger; visualization of results of other subsets are in supplementary document), CUB-200-2011, and HORSE datasets in Fig. 4 and 5, respectively. From the visualizations we can directly recognize the 3D structure of each shown object category. For the four subsets of AFHQ dataset, despite these animal categories sharing many similarities, we can recognize the individual animal face characteristics from predicted 3D keypoints with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10709" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "linking edges. Moreover, we find that the set of thick edges, i.e. the ones with large learned edge coefficient \\(\\omega\\), coincide with the most notable structures of each category, while negligible edges remain very thin. Finally, we demonstrate that keypoints are consistently detected among different instances of objects per category, which can be seen based on the consistency of the color of keypoints (both in 2D and 3D). More qualitative results and videos are included in the supplementary document and attached files." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.354 + ], + "angle": 0, + "content": "Since there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting (cf. Table 1), as comparison we consider a simple baseline that first predicts 2D keypoints using AutoLink [12], and then lifts 2D keypoints to 3D by pinpointing learned 2D keypoints on the depth map (given by the pre-trained unsupervised monocular depth estimator MiDaS [42]). Fig. 6 shows a comparison on CELEBA WILD dataset." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.374, + 0.468, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.637, + 0.472, + 0.693 + ], + "angle": 0, + "content": "Figure 6. Comparison of predicted 3D structures between our method (left) and AutoLink [12] combined with MiDaS [42] (right). Our method obtains more consistent 2D keypoint predictions and more realistic 3D structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.273, + 0.753 + ], + "angle": 0, + "content": "5.3. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "2D keypoint inference. First, we show that our method improves upon unsupervised 2D keypoint inference. Table 2 shows the NME values of our method and of various other unsupervised methods on the CELEBA WILD dataset. We also compare to the method by He et al. [12] for various choices of \\( K \\) on the CELEBA WILD dataset to show the robustness of our method, as shown in Table 3. We can see that our method outperforms all competitors across all settings." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.885, + 0.334 + ], + "angle": 0, + "content": "
MethodK=8
DFF [7]31.30%*
SCOPS (w/o saliency) [17]22.11%†
SCOPS (w/saliency) [17]15.01%†
Liu et al. [27]12.26%†
Huang et al. [16]8.4%†
GANSeg [13]6.18%†
Thewlis et al. [49]31.30%*
Zhang et al. [67]40.82%*
LatentKeypointGAN [11]21.90%†
LatentKeypointGAN-tuned [11]5.63%†
Lorenz et al. [29]11.41%‡
IMM [19]8.74%‡
AutoLink [12]5.39%
Ours5.21%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.343, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Table 2. Normalized \\( L_{2} \\) error (NME) for 2D keypoints inference of various unsupervised methods on CELEBA WILD datasets for \\( K = 8 \\). (* reported from Collins et al. [7]; † reported from He et al. [12]; ‡ reported from Liu et al. [27])" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.41, + 0.887, + 0.472 + ], + "angle": 0, + "content": "
MethodK=8K=16K=24K=32
AutoLink [12]5.39%4.69%3.99%3.77%
Ours5.21%3.97%3.54%3.48%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Table 3. Normalized \\( L_{2} \\) error (NME) of AutoLink and our method for different numbers of keypoints using the CELEBA WILD dataset." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.53, + 0.888, + 0.582 + ], + "angle": 0, + "content": "
Supervised 3DDFAUnsupervised
AutoLink+Unsup3dAutoLink+MiDaSOurs
4.94%11.47%9.23%8.48%
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Table 4. Normalized \\( L_{2} \\) error (NME) of our method, two unsupervised methods (AutoLink + MiDaS and AutoLink + Unsup3d) and one supervised method (3DDFA) for 3D keypoints inference (training on the 300W-LP dataset and testing on the AFLW2000-3D dataset)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.895, + 0.829 + ], + "angle": 0, + "content": "3D keypoint inference. As stated in Sec. 5.2, there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting. So we define two unsupervised baselines by combining AutoLink [12] with MiDaS [42], and Autolink [12] with Unsup3d [54]. Moreover, we compare against the supervised method 3DDFA [71] to understand the gap between unsupervised and supervised methods. Table 4 summarises the NME results, in which our method achieves better performance compared to the two-stage baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.652, + 0.862 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this section we discuss our ablation studies. Results are shown in Fig. 7. The first and second row show the results" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10710" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "with full training losses (i.e. Eq. (10)) with the number of keypoints being 32 and 16, respectively. We can see that with an increasing number of keypoints more detailed structure can be captured by our 3D keypoints. Furthermore, we remove \\(\\mathcal{L}_{\\mathrm{rep}}\\) from our overall loss to train our framework (third row). The structure collapses into a degenerate keypoint configuration, which confirms the importance of \\(\\mathcal{L}_{\\mathrm{rep}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.243 + ], + "angle": 0, + "content": "A more detailed analysis on the behavior of \\(\\mathcal{L}_{\\mathrm{rep}}\\) and on the influence of the number of basis functions \\(n\\) of our shape model, are provided in the supplementary document." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.262, + 0.468, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.337, + 0.466, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.403, + 0.449, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.468, + 0.508 + ], + "angle": 0, + "content": "Figure 7. Ablation study on CELEBA WILD dataset. The first and second rows show results for a different number of keypoints (32 and 16). The third row shows results without repulsion loss \\(\\mathcal{L}_{\\mathrm{rep}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.54, + 0.192, + 0.555 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.585, + 0.445, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.703, + 0.444, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.823, + 0.468, + 0.851 + ], + "angle": 0, + "content": "Figure 8. Pose estimation visualization of two instances from CELEBA WILD dataset from two views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Large pose estimation has always been an important yet challenging issue in many 3D shape inference settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[26, 33, 55, 56, 61, 63]. Recent works have found that when the pose variations falls into a small range, it can be learned directly by a network together with shape [53]. We observe similar results in our setting. For the CELEBA WILD, AFHQ (instances with moderate pose variations), CUB-200-2011 and HORSE (aligned facing direction) datasets, Fig. 4 and 5 have shown multiple instances in different poses with correct pose estimation. Fig. 8 gives more pose estimation results. However, for datasets consisting of instances with large pose variations, our method may fail to optimize for the correct poses. For example, our test on the HUMAN3.6M dataset [18] shows that, even though our model learns plausible human body structure, it fails to optimize for the correct pose of each instance, which also leads to incorrect instance-specific deformations, as shown in Fig. 9 – for such cases with strong articulation, domain-specific approaches that exploit domain knowledge (e.g. about the human kinematic skeleton structure) are at the time being more suitable. Yet, inferring such articulated objects is a highly relevant and interesting direction for future work." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.422, + 0.855, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.587, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Figure 9. Incorrect pose estimation visualization of HUMAN3.6M dataset. For severely articulated objects additional domain knowledge is currently needed (e.g. a human body kinematic skeleton)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.618, + 0.68 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall, our method is the first unsupervised approach that can infer 3D keypoints from category-specific image collections without the need of any additional prior information. Conceptually, our main novelty lies in simultaneously learning 3D keypoints along with a 3D shape space model from an inhomogeneous collection of category-specific images (e.g. opposed to homogeneous collections of multiview images or videos, which show the same object from different views or in different poses). We experimentally demonstrate that our method not only improves upon 2D keypoint accuracy, but also infers plausible 3D structure for various object categories. We expect our work to be useful for diverse downstream tasks, for example related to 3D reconstruction or 3D shape-based generative models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10711" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[3] Adrian Bulat, Tzimiropoulos, and Georgios. Two-stage convolutional part heatmap regression for the 1st 3d face alignment in the wild (3dfaw) challenge. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 616-624. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[5] Boyuan Chen, Pieter Abbeel, and Deepak Pathak. Unsupervised learning of visual 3d keypoints for control. In International Conference on Machine Learning, pages 1539-1549. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[7] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[8] Hui Deng, Tong Zhang, Yuchao Dai, Jiawei Shi, Yiran Zhong, and Hongdong Li. Deep non-rigid structure-from-motion: A sequence-to-sequence translation perspective. arXiv preprint arXiv:2204.04730, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[9] Aysegul Dundar, Kevin J Shih, Animesh Garg, Robert Pottorf, Anrew Tao, and Bryan Catanzaro. Unsupervised disentanglement of pose, appearance and background from images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3883-3894, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[10] Chao Gou, Yue Wu, Fei-Yue Wang, and Qiang Ji. Shape augmented regression for 3d face alignment. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 604-615. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. arXiv preprint arXiv:2103.15812, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. Advances in Neural Information Processing Systems, 35:36123-36141, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.134 + ], + "angle": 0, + "content": "generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1225-1235, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[14] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21337-21348, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Sina Honari and Pascal Fua. Unsupervised 3d keypoint estimation with multi-view geometry. arXiv preprint arXiv:2211.12829, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[16] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662-8672, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[17] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[21] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 196-214. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[22] Yunji Kim, Seonghyeon Nam, In Cho, and Seon Joo Kim. Unsupervised keypoint learning for guiding class-conditional video prediction. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[23] Chen Kong and Simon Lucey. Deep non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1558-1567, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[24] Tejas D Kulkarni, Ankush Gupta, Catalin Ionescu, Sebastian Borgeaud, Malcolm Reynolds, Andrew Zisserman, and Volodymyr Mnih. Unsupervised learning of object keypoints for perception and control. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6152-6162, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10712" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[27] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Unsupervised part segmentation through disentangling appearance and shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8355-8364, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[29] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[30] Y Ma, S Soatto, J Kosecka, and S Sastry. An invitation to 3d vision: From images to models. Springer verlag, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[31] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[36] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[37] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression: A flexible alignment-based framework for non-rigid structure estimation. IEEE Transactions on Image Processing, 27(1):249-264, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[40] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression networks: Learning 3d structure of non-rigid objects from 2d annotations. In European Conference on Computer Vision, pages 1-18. Springer, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.894, + 0.278 + ], + "angle": 0, + "content": "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[42] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.894, + 0.422 + ], + "angle": 0, + "content": "[43] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[44] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7326-7335, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.583, + 0.892, + 0.652 + ], + "angle": 0, + "content": "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[48] Suwajanakorn Supasorn, Noah Snavely, Jonathan J Tompson, and Mohammad Norouzi. Discovery of latent 3d keypoints via end-to-end geometric reasoning. Advances in neural information processing systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[49] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Sergey Tulyakov, László A Jeni, Jeffrey F Cohn, and Nicu Sebe. Viewpoint-consistent 3d face alignment. IEEE transactions on pattern analysis and machine intelligence, 40(9): 2250-2264, 2017. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10713" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[52] Chaoyang Wang, Chen-Hsuan Lin, and Simon Lucey. Deep nrsfm++: Towards unsupervised 2d-3d lifting in the wild. In 2020 International Conference on 3D Vision (3DV), pages 12–22. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[53] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.468, + 0.315 + ], + "angle": 0, + "content": "[54] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.468, + 0.371 + ], + "angle": 0, + "content": "[55] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, pages 1-12, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.372, + 0.468, + 0.44 + ], + "angle": 0, + "content": "[56] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.468, + 0.509 + ], + "angle": 0, + "content": "[57] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.468, + 0.566 + ], + "angle": 0, + "content": "[58] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.468, + 0.634 + ], + "angle": 0, + "content": "[59] Chenxin Xu, Siheng Chen, Maosen Li, and Ya Zhang. Invariant teacher and equivariant student for unsupervised 3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3013-3021, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.636, + 0.468, + 0.678 + ], + "angle": 0, + "content": "[60] Yinghao Xu, Ceyuan Yang, Ziwei Liu, Bo Dai, and Bolei Zhou. Unsupervised landmark learning from unpaired data. arXiv preprint arXiv:2007.01053, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[61] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.468, + 0.831 + ], + "angle": 0, + "content": "[62] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[63] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1323-1330. IEEE, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[64] Haitian Zeng, Yuchao Dai, Xin Yu, Xiaohan Wang, and Yi Yang. Pr-rrn: pairwise-regularized residual-recursive networks for non-rigid structure-from-motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5600–5609, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[65] Haitian Zeng, Xin Yu, Jiaxu Miao, and Yi Yang. Mhr-net: Multiple-hypothesis reconstruction of non-rigid shapes from 2d views. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[66] Hongwen Zhang, Qi Li, and Zhenan Sun. Joint voxel and coordinate regression for accurate 3d facial landmark localization. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2202-2208. IEEE, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[68] Ruiqi Zhao, Yan Wang, C Fabian Benitez-Quiroz, Yaojie Liu, and Aleix M Martinez. Fast and precise face alignment and 3d shape reconstruction from a single 2d image. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 590-603. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[69] Shihao Zhou, Mengxi Jiang, Qicong Wang, and Yunqi Lei. Towards locality similarity preserving to 3d human pose estimation. In Proceedings of the Asian Conference on Computer Vision, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.488, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[70] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[71] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10714" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_origin.pdf b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0633d91700db547670a03f80997f490ca7ebde8 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/9e81eb55-2017-47d1-ab50-718fc069c773_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4cb59ae604ebe07734df3b3af80fb40001c14d46dee96407715c05d8af94750 +size 8712684 diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/full.md b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f5dd90a95400128f8410a37bcbc5585f4a5352a8 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/full.md @@ -0,0 +1,324 @@ +# Unsupervised 3D Structure Inference from Category-Specific Image Collections + +Weikang Wang + +Dongliang Cao + +Florian Bernard + +University of Bonn, Germany + +![](images/f6a1958682f39e314355aaa2084552487e22c9f5c40e7f1cdf28107a1b914b82.jpg) +Figure 1. Our method detects consistent and reasonable 3D structure for general object categories from a single 2D image. For training, our method only requires a collection of images from the same category without any additional information. + +# Abstract + +Understanding 3D object structure from image collections of general object categories remains a long-standing challenge in computer vision. Due to the high relevance of image keypoints (e.g. for graph matching, controlling generative models, scene understanding, etc.), in this work we specifically focus on inferring 3D structure in terms of sparse keypoints. Existing 3D keypoint inference approaches rely on strong priors, such as spatio-temporal consistency, multi-view images of the same object, 3D shape priors (e.g. templates, skeleton), or supervisory signals e.g. in the form of 2D keypoint annotations. In contrast, we propose the first unsupervised 3D keypoint inference approach that can be trained for general object categories solely from an inhomogeneous image collection (containing different instances of objects from the same category). Our experiments show that our method not only improves upon unsupervised 2D keypoint inference, but more importantly, it also produces reasonable 3D structure for various object categories, both qualitatively and quantitatively. + +# 1. Introduction + +Understanding 3D object structure from 2D images is a fundamental problem in computer vision that has been studied + +for decades. A range of different 3D structure representations have been considered, including keypoints [14, 25, 59, 69], meshes [55, 56, 62], and voxels [35, 45, 47]. Among these, keypoint-based 3D structure representations have the strong advantage that they are simple and flexible, while at the same time being highly relevant for diverse downstream tasks. However, obtaining 3D keypoints from 2D images is hard due to unknown depth, pose variations, different appearance, and the lack of explicit geometric information. Due to these difficulties, prior works mostly focus on 2D keypoints detection instead of 3D [9, 12, 31, 36, 66]. + +In this work we investigate whether it is possible to directly obtain 3D keypoints without using any explicit supervision. Our work is motivated by the fact that 2D image keypoints are observations of points in 3D space that are projected onto the 2D image plane. By explicitly understanding and inferring 3D geometry we are able to improve upon the quality of detected keypoints. Specifically, we experimentally demonstrate that projecting our predicted 3D keypoints onto the image plane leads to more reliable 2D keypoints compared to existing 2D keypoint predictors. Moreover, our method can serve the increasing amount of tasks that rely on 3D keypoints, e.g. 3D structure-aware generative models [58], 3D avatar generation [34], or 3D reconstruction [55, 56]. Despite keypoint prediction being a well-studied problem in the 2D domain, 3D keypoint inference is under-explored and we thus specifically contribute + +towards this area. + +Existing works on inferring 3D keypoints either use 3D point clouds or 2D images/videos from the same object instance as input data. While the former is conceptually a subset selection problem, in this work we solely focus on the more challenging latter case. Manually obtaining 3D keypoints in image data is much harder than annotating 2D keypoints. Thus, there are only very few datasets with 3D keypoints annotations, which in turn motivates unsupervised learning strategies that do not require expensive or even infeasible 3D annotations for training. Existing works that address 3D keypoint inference from image data use for example 2D keypoint annotations [38, 44], multiple views with known poses [11, 35, 48], spatial-temporal consistency in videos [9, 15], 2D/3D skeleton [55, 56] or templates [14]. In stark contrast, in this work we relax the assumptions about the availability of such supervisory signals and train a 3D keypoint predictor merely from category-specific image collections as input. Our method builds upon the assumption that object instances from the same category should have a similar 3D structure with moderate intra-class variations (e.g. faces, horses, birds). Table 1 summaries the differences among existing methods and Fig. 2 gives a visualization of some additional information within these methods and the comparison to our setting. + +
MethodsUnsup.w/o PriorCat.-agnostic
Bulat et al. [3]X✓✓
Gou et al. [10]X✓✓
Zhao et al. [68]X✓✓
Tulyakov et al. [50]X✓✓
Pavlakos et al. [41]X✓X
Zhang et al. [66]X✓X
He et al. [14]XX✓
Mildenhall et al. [35]✓X✓
Wu et al. [54]✓X✓
Moniz et al. [37]✓XX
Supasorn et al. [48]✓X✓
Novotny et al. [38]✓X✓
Park et al. [40]✓X✓
Chen et al. [5]✓X✓
Reddy et al. [44]✓X✓
Zhou et al. [69]✓✓X
Dundar et al. [9]✓✓X
Xu et al. [59]✓✓X
Honari and Fua [15]✓X✓
Ours✓✓✓
+ +Table 1. In comparison to existing 3D keypoints inference methods, our approach combines a unique set of desirable properties: it can be trained in an unsupervised manner without requiring any additional information; it does not require any shape prior and can thus be applied to different object categories. + +![](images/31017baa371f382d7b690fc43418510f203ad267aa832ca5db33badeed8922a7.jpg) +Figure 2. (Top row from left to right) Existing methods require different additional information for 3D structure inference, e.g. multiple views (from [12]), 2D keypoint annotations (from [12]), and a symmetry prior (from [54]). (Bottom row) Our setting does not require such information but instead also infers a 3D shape space model that is learned in an unsupervised manner. + +Our method enables to infer 3D keypoints from category-specific image collections without relying on additional prior information. To this end, we build upon the recent unsupervised 2D keypoint detection method by He et al. [12] and extend it by additionally learning a sparse 3D shape space model to represent 3D structure. To regularise the 3D structure, a self-supervised re-projection loss is introduced to couple 3D keypoints with corresponding 2D keypoints. By combining our 3D shape space model, re-projection loss, and additional regularisation loss terms, we demonstrate that our proposed method not only improves 2D keypoint accuracy, but also produces reasonable 3D structure. We summarise our main contributions as: + +- For the first time we obtain 3D keypoints from general category-specific image collections without relying on additional prior information. +- Our method obtains reasonable 3D keypoints based on our learnable 3D shape space model that is fused with a state-of-the-art unsupervised 2D keypoint detector. +- We demonstrate that our method both improves unsupervised 2D keypoint detection accuracy and obtains reasonable 3D shape on different benchmarks. + +# 2. Related Work + +In the following we summarize relevant 2D and 3D key-point prediction methods. + +# 2.1. 2D Keypoint Detection + +Supervised methods. Various domain-specific supervised 2D keypoint detection methods have been proposed that are trained on a large amount of labeled datasets, e.g. for faces [2, 57, 71], human bodies [1, 4, 18, 21, 32], or vehicles [43], etc. However, labeling keypoints for images is time-consuming, laborious and error-prone, especially when there are occlusions, extreme pose variations, or appearance changes. As a consequence, labeling errors will subsequently influence network training and downstream task performance. + +Unsupervised methods. Due to these reasons, an increasing amount of attention is paid to unsupervised methods. One popular concept is multi-view geometric consistency. Thewlis et al. [49] detect keypoints by recovering transformations between one image and its deformed counterpart from pre-defined deformations. Without relying on pre-defined deformations, various video-based methods [9, 19, 20, 22, 24, 36, 46] infer keypoints based on a temporal coherence prior. Instead of using multiple views as input, Xu et al. [60] use unpaired input to disentangle appearance and poses by conducting a swapping-reconstruction strategy and encouraging reconstructed images to resemble the original. He et al. [13] and He et al. [11] use a generative adversarial training framework to constrain keypoint locations by interpreting them as latent codes for an image generator. However, collaboratively learning intermediate keypoints and complex adversarial models is unstable and does not scale. Starting from some initial point sets, Mallis et al. [31] filter keypoints by optimizing alternatively between clustering of keypoints features across a dataset, and training a keypoint detector with pseudo-labels offered by cluster centers. However, this method depends heavily on the initial point set. Zhang et al. [67] address single-image keypoint inference using an auto-encoder architecture to interpret keypoints as a latent representation, while reconstructing the input image from the concatenated keypoints and respective pixel feature descriptors. He et al. [12] use a similar idea by representing keypoints as latent features in an auto-encoder manner. However, they reconstruct the input from an edge map created from keypoints combined with random masking of appearance information. In this way, paired input data is not needed to disentangle appearance and poses, which makes this method applicable in wider scenarios. + +# 2.2.3D Keypoint Inference + +Compared to 2D keypoints detection, 3D keypoints detection is an under-explored topic, particularly in unsupervised settings without strong priors. + +Supervised methods. Some earlier methods [3, 10, 68] use a two-stage strategy that infers 2D keypoints first, and then predicts depth for these keypoints. These methods + +not only need 3D keypoints or depth annotations for training, but are also sub-optimal due to their two-stage procedure. Tulyakov et al. [50] generalize a cascaded regression method to 3D for estimating 3D face landmarks in an end-to-end manner from images. Pavlakos et al. [41] generalize 2D heatmaps directly to 3D by inferring 3D keypoints, while Zhang et al. [66] improve this method via a joint voxel and coordinate regression to avoid the curse of dimensionality. However, these methods all needs 3D keypoints annotations, so that their scope of application is limited due to the absence of annotated 3D keypoints in many cases. Reddy et al. [44] propose a trifocal loss to learn (possibly occluded) 3D keypoints from multiple views by using 2D keypoints annotations. Closely related to 3D shape inference, sparse non-rigid structure-from-motion methods also take 2D keypoints as input to infer 3D structure of objects [8, 23, 38-40, 52, 64, 65]. Based on [12], He et al. [14] predict 3D keypoints using dozens of images with 2D annotations together with a skeleton template. + +Unsupervised methods. In order to get rid of suboptimal two-stage approaches or 2D/3D keypoints annotations, some works aim at unsupervised 3D structure inference. Supasorn et al. [48] use paired views and known transformations between them as input to infer 3D keypoints. Chen et al. [5] use multiple views as input to infer 3D keypoints by averaging 2D keypoints and depth values detected in each view. Using a pair of views as input, Honari and Fua [15] propose a two-stream auto-encoder to encode 2D keypoints of each view, and then use triangulation to lift two sets of 2D keypoints to 3D. Various unsupervised 3D human pose estimation methods [25, 59, 69], either use multiple views or video as input, or they rely on some domain-specific losses as constraints. + +Different from all mentioned 3D keypoints inference settings, our method does not rely on 2D/3D keypoints annotations, 3D template/priors/skeletons, multiple views/video, or domain-specific 3D structure. Instead, for the first time we infer 3D structure only from category-specific collections of images for a broad range of object categories. + +# 3. Background + +Our method builds upon the recent AutoLink method [12] that addresses the task of unsupervised keypoint detection in 2D. In this section we give a brief overview of AutoLink. + +AutoLink comprises an auto-encoder with 2D keypoints as bottleneck latent embeddings. For training, it employs a reconstruction loss that compares the input image with a reconstructed image that is obtained based on edge maps and a randomly masked input image. + +Formally, given an image $I \in \mathbb{R}^{W \times H \times 3}$ , the encoder takes $I$ as input and outputs $K$ heatmaps, denoted as $H_{1}, H_{2}, \ldots, H_{K}$ , where $H_{i} \in [0,1]^{W \times H}$ . The $i$ -th heatmap represents the likelihood for each pixel to be the + +![](images/0ad3c2b3e0c004ec57e8380182823bf433d7ec03f3a63d3ad70b16ec9b91fd2f.jpg) +Figure 3. Overview of our pipeline. We sample images $I$ from a category-specific image collection, which are then piped into an encoder that predicts $K$ heatmaps. These heatmaps are used to extract a $K \times 2$ keypoint matrix $P^{2D}$ . From $P^{2D}$ , an edge map $S$ is constructed, which is then, together with a randomly masked version of the input image $I_{\mathrm{masked}}$ , fed into the decoder to reconstruct image $I'$ . Moreover, an MLP takes $P^{2D}$ as input and predicts with different heads, the coefficients $\alpha$ , the 3D translation $T$ , and the 3D rotation $R$ . At the same time we optimize for a globally trainable mean shape $M$ and basis $B$ , which define a 3D shape space model used to represent our 3D keypoints, see Eq.(5). The main difference to AutoLink [12], which predicts 2D keypoints only, is that we infer 3D keypoints and a 3D shape space model. + +specified keypoint $i$ , where $K$ represents the total number of different keypoints. With that, the location of each 2D keypoint $P_{i}^{2\mathrm{D}} = (x_{i},y_{i}), i = 1,2,\ldots ,K$ can be represented as + +$$ +P _ {i} ^ {\mathrm {2 D}} = \sum_ {p} \frac {H _ {i} (p)}{\sum_ {p ^ {\prime}} H _ {i} \left(p ^ {\prime}\right)} p, \tag {1} +$$ + +where $p,p^{\prime}\in [-1,1]\times [-1,1]$ are normalized pixel coordinates. + +For each pair of 2D keypoints, $P_{i}^{2\mathrm{D}}$ and $P_{j}^{2\mathrm{D}}$ , $1 \leq i,j \leq K$ , the method constructs an edge map $S_{ij} \in \mathbb{R}^{H \times W}$ defined on normalized pixel coordinates $p$ via + +$$ +S _ {i j} (p) = \exp \left(- d _ {i j} (p) / \sigma^ {2}\right), \tag {2} +$$ + +where $d_{ij}(p)$ is the $L_{2}$ distance of pixel $p$ to the line segment connecting $P_{i}^{2\mathrm{D}}$ and $P_{j}^{2\mathrm{D}}$ , and $\sigma \in \mathbb{R}$ controls the thickness of the edge. The final edge map $S \in \mathbb{R}^{W \times H}$ summarizing all edges between all keypoints is given by + +$$ +S (p) = \max _ {1 \leq i, j \leq K} \omega_ {i j} S _ {i j} (p), \tag {3} +$$ + +where $\omega_{ij}$ are learnable parameters that are shared by all images and optimized during training. + +Finally, the edge map $S$ and the (randomly masked) input image $I_{\mathrm{masked}}$ are concatenated along the channel axis and fed into a decoder to reconstruct the input. The whole AutoLink pipeline is trained end-to-end only based on the reconstruction loss + +$$ +\mathcal {L} _ {\text {r e c}} = \| F (D (S \odot I _ {\text {m a s k e d}})) - F (I) \| _ {2}, \tag {4} +$$ + +where $D$ is the decoder, $F$ is a pre-trained and fixed feature extractor (i.e. VGG network), $\odot$ represents channel-wise concatenation, and $I_{\mathrm{masked}}$ is the randomly masked original input image with $80\%$ mask area. + +# 4. Unsupervised 3D Structure Inference + +In the following we explain our method. To this end, we introduce our 3D shape space model for inferring 3D keypoint structure in Section 4.1. In section 4.2 we explain estimating the object pose. Section 4.3 introduces our unsupervised losses. Our overall pipeline is summarized in Fig. 3. + +# 4.1. Trainable 3D Shape Space Model + +Rather than working with high-dimensional and expensive 3D heatmaps, we infer 3D geometry by training a 2D keypoint predictor together with a 3D shape space model in an end-to-end manner, see Fig. 3. Our encoder first predicts 2D keypoint heatmaps from a given 2D image, which are then transformed into a 2D keypoint matrix using Eq. (1). Then a multi-layer perceptron (MLP) takes the 2D keypoint matrix as input to predict input-dependent coefficients $\alpha$ of a 3D shape model. At the same time, the mean shape $M$ and the set of basis vectors $B$ of this 3D shape space model are optimized as global trainable parameters. Overall, this allows to obtain the 3D keypoints (in canonical pose) as + +$$ +P _ {\text {c a n o n i c a l}} ^ {\mathrm {3 D}} = \operatorname {m a t} (M + \alpha B), \tag {5} +$$ + +where $M$ is the mean shape of size $1 \times 3K$ , $B$ is the basis matrix of size $n \times 3K$ and $\alpha$ are the coefficients of size $1 \times n$ , with $n$ denoting the number of basis functions. The operation $\mathrm{mat}()$ reshapes a $1 \times 3K$ row vector to a $K \times 3$ matrix. By training these components simultaneously with our 2D heatmap predictor, we can ensure that during training our 2D heatmap predictor receives gradients that contain explicit information about 3D geometry. + +We emphasize that opposed to most existing 3D shape space learning methods that require known keypoint locations, we simultaneously optimize for keypoint locations + +while using our trainable 3D shape space model as 3D geometry-aware regularizer. Unlike non-rigid structure-from-motion methods [8, 38-40] that utilize similar ideas, our approach does not require multiple images of the same object from different views. Instead, we use an inhomogeneous collection of category-specific images, where the 3D shape space model serves as flexible and adaptive shape prior that can adjust to the different geometries of objects (within the considered object category). As such, our approach is much more flexible, since it can handle single-view image collections as for example available in web collections. + +# 4.2. Pose Estimation + +After getting the 3D keypoints $P_{\mathrm{canonical}}^{\mathrm{3D}}$ (in canonical pose) using Eq. (5), an input-dependent transformation matrix and scaling factor are estimated that transform $P_{\mathrm{canonical}}^{\mathrm{3D}}$ to the posed 3D keypoints $P^{3D}$ , such that the projection of $P^{3D}$ onto the 2D image plane coincides with the 2D structure of the object within the 2D image. + +Non-rigid object deformations are taken care of by the 3D shape model, so that we represent the object pose solely via a rigid body transformation matrix $(R,T)\in \mathrm{SE}(3)$ . With the 2D keypoint matrix $P^{2D}$ as input, an MLP computes the translation $T\in \mathbb{R}^{1\times 3}$ and the rotation angles $\Theta \in \mathbb{R}^{1\times 3}$ , which are converted to a rotation matrix via the Rodrigues formula [30]: + +$$ +R = I _ {3} + \frac {\sin (\theta)}{\theta} [ \Theta ] _ {\times} + \frac {1 - \cos (\theta)}{\theta^ {2}} [ \Theta ] _ {\times} ^ {2}, \tag {6} +$$ + +where $\theta = \| \Theta \|$ and the skew-symmetric operator $[\cdot ]_{\times}$ turns a vector in $\mathbb{R}^{1\times 3}$ into a skew-symmetric matrix. Finally the 3D keypoint matrix $P^{3D}$ is given as + +$$ +P ^ {3 \mathrm {D}} = P _ {\text {c a n o n i c a l}} ^ {3 \mathrm {D}} R + \mathbf {1} _ {K} T, \tag {7} +$$ + +where $\mathbf{1}_K$ is a $K$ -dimensional vector of all ones. + +# 4.3. Unsupervised Losses + +Our re-projection loss penalizes the discrepancy between the projected 3D keypoints and the 2D heatmap keypoint predictions, i.e. + +$$ +\mathcal {L} _ {\mathrm {p r o j}} = \| P ^ {3 \mathrm {D}} \Pi - P ^ {2 \mathrm {D}} \| _ {F}, \quad \text {w i t h} \quad \Pi = \left[ \begin{array}{c c} s _ {1} & 0 \\ 0 & s _ {2} \\ 0 & 0 \end{array} \right] \tag {8} +$$ + +being a simple orthographic projection (of $P^{3D}$ from Eq. (8)) combined with scaling $s = (s_1, s_2)$ , which is predicted by an MLP. This loss ensures consistent 2D keypoint predictions between different images, since the 3D shape structure is shared across the whole image collection via the 3D shape space model. + +However, combining $\mathcal{L}_{\mathrm{rec}}$ and $\mathcal{L}_{\mathrm{proj}}$ alone will lead to degenerate solutions (i.e. all keypoints converge to a single location). In order to avoid this situation, we consider the repulsion loss + +$$ +\mathcal {L} _ {\text {r e p}} = - \sum_ {i = 1} ^ {K} \| p _ {i} - \mathcal {N} (p _ {i}) \| _ {2} \exp (- \| p _ {i} - \mathcal {N} (p _ {i}) \| _ {2} / h), \tag {9} +$$ + +where $p_i$ is a 2D keypoint (a row of $P^{2\mathrm{D}}$ ), $\mathcal{N}(p_i)$ is the nearest keypoint (other than $p_i$ ) of $p_i$ in $P^{2\mathrm{D}}$ , and $h$ is a temperature parameter that controls the repulsion decrease rate. The effect of Eq. (9) is that neighboring keypoints do not move too close to each other. + +Our total loss combines all terms and is defined as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \lambda_ {1} \mathcal {L} _ {\text {p r o j}} + \lambda_ {2} \mathcal {L} _ {\text {r e p}}, \tag {10} +$$ + +where $\lambda_1, \lambda_2$ are weights of the individual terms. + +# 5. Experiments + +We evaluate our method on various datasets and compare to existing keypoint prediction methods. Implementation details are provided in the supplementary document. + +# 5.1. Experimental Setup + +Datasets. We evaluate our method on six datasets: CELEBA WILD [28], CUB-200-2011 [51], 300W-LP [71], AFLW2000-3D [71], HORSE [70] and AFHQ [6]. The CELEBA WILD dataset [28] contains celebrity faces in in-the-wild environments, along with 2D keypoint annotations for each image. Following the setting of He et al. [12], we remove the images whose face covers less than $30\%$ of the area. The 300W-LP [71] dataset is a large synthetic face dataset containing 61,225 samples across large poses (1,786 from IBUG, 5,207 from AFW, 16,556 from LFPW and 37,676 from HELEN). The AFLW2000-3D dataset was introduced along with the 300W-LP dataset by the same data acquisition method, containing 2,000 facial images, each with 68 3D keypoints annotations. We train on the 300W-LP dataset and test on the AFLW2000-3D dataset to evaluate the quantitative 3D keypoint detection performance. We use subsets from the AFHQ dataset [6], namely CAT, DOG, TIGER, FOX, CHEETAH, WOLF, JAGUAR and LION, to get qualitative 3D structure inference results. CUB-200-2011 consists of 11,788 images of birds. Following the settings in [12, 29], we align all birds facing left, and remove seabirds. The HORSE dataset is extracted from the CycleGAN dataset [70] by removing images with multiple horses and aligning them to face left. CUB-200-2011, HORSE, AFHQ and CELEBA WILD are used to generate qualitative results. We quantitatively evaluate unsupervised 2D keypoint inference using the CELEBA WILD dataset and compare it with the recent state-of-the-art method by He et al. [12]. + +![](images/93cb6c61d32f561f61dfe16c14469edbb4f80ac17e3292d8fe053419ee327bbb.jpg) +Figure 4. Visualization of the 3D structure for different instances of the AFHQ dataset (CAT, DOG, FOX and TIGER subset in each row). Each 3D structure is shown in two forms: (i) 3D keypoints with linking edges between them (the edge thickness is proportional to the edge map coefficients $\omega$ , which indicate their relative importance); (ii) Delaunay triangulation of the 3D keypoints rendered as shaded mesh. Different keypoints are shown in different colors to demonstrate consistency across different images (columns). + +![](images/f2b08cef81a9de29cca76b6e277f14809d743c8a15a112e845a5655263b8eadb.jpg) +Figure 5. Visualization of the 3D structure for different images from the HORSE and CUB-200-2011 datasets (rows). + +Metrics. For unsupervised 2D keypoint inference, we follow the metrics used by He et al. [12], which is the normalized mean $L_{2}$ error (NME) of detected 2D keypoints aligned to the ground truth. For the 3D case, we generalize NME to 3D by calculating the normalized mean $L_{2}$ of detected 3D keypoints aligned to ground truth 3D keypoints. In both cases, the normalization is the ocular distances of each image. + +# 5.2. Qualitative Results + +We visualize the learned 3D structure on four subsets of AFHQ (Cat, Dog, Fox and Tiger; visualization of results of other subsets are in supplementary document), CUB-200-2011, and HORSE datasets in Fig. 4 and 5, respectively. From the visualizations we can directly recognize the 3D structure of each shown object category. For the four subsets of AFHQ dataset, despite these animal categories sharing many similarities, we can recognize the individual animal face characteristics from predicted 3D keypoints with + +linking edges. Moreover, we find that the set of thick edges, i.e. the ones with large learned edge coefficient $\omega$ , coincide with the most notable structures of each category, while negligible edges remain very thin. Finally, we demonstrate that keypoints are consistently detected among different instances of objects per category, which can be seen based on the consistency of the color of keypoints (both in 2D and 3D). More qualitative results and videos are included in the supplementary document and attached files. + +Since there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting (cf. Table 1), as comparison we consider a simple baseline that first predicts 2D keypoints using AutoLink [12], and then lifts 2D keypoints to 3D by pinpointing learned 2D keypoints on the depth map (given by the pre-trained unsupervised monocular depth estimator MiDaS [42]). Fig. 6 shows a comparison on CELEBA WILD dataset. + +![](images/0981cd9de79f2f502192e66d557acf9c80fd59f00a18e2cf79b4899d710636a9.jpg) +Figure 6. Comparison of predicted 3D structures between our method (left) and AutoLink [12] combined with MiDaS [42] (right). Our method obtains more consistent 2D keypoint predictions and more realistic 3D structure. + +# 5.3. Quantitative Results + +2D keypoint inference. First, we show that our method improves upon unsupervised 2D keypoint inference. Table 2 shows the NME values of our method and of various other unsupervised methods on the CELEBA WILD dataset. We also compare to the method by He et al. [12] for various choices of $K$ on the CELEBA WILD dataset to show the robustness of our method, as shown in Table 3. We can see that our method outperforms all competitors across all settings. + +
MethodK=8
DFF [7]31.30%*
SCOPS (w/o saliency) [17]22.11%†
SCOPS (w/saliency) [17]15.01%†
Liu et al. [27]12.26%†
Huang et al. [16]8.4%†
GANSeg [13]6.18%†
Thewlis et al. [49]31.30%*
Zhang et al. [67]40.82%*
LatentKeypointGAN [11]21.90%†
LatentKeypointGAN-tuned [11]5.63%†
Lorenz et al. [29]11.41%‡
IMM [19]8.74%‡
AutoLink [12]5.39%
Ours5.21%
+ +Table 2. Normalized $L_{2}$ error (NME) for 2D keypoints inference of various unsupervised methods on CELEBA WILD datasets for $K = 8$ . (* reported from Collins et al. [7]; † reported from He et al. [12]; ‡ reported from Liu et al. [27]) + +
MethodK=8K=16K=24K=32
AutoLink [12]5.39%4.69%3.99%3.77%
Ours5.21%3.97%3.54%3.48%
+ +Table 3. Normalized $L_{2}$ error (NME) of AutoLink and our method for different numbers of keypoints using the CELEBA WILD dataset. + +
Supervised 3DDFAUnsupervised
AutoLink+Unsup3dAutoLink+MiDaSOurs
4.94%11.47%9.23%8.48%
+ +Table 4. Normalized $L_{2}$ error (NME) of our method, two unsupervised methods (AutoLink + MiDaS and AutoLink + Unsup3d) and one supervised method (3DDFA) for 3D keypoints inference (training on the 300W-LP dataset and testing on the AFLW2000-3D dataset). + +3D keypoint inference. As stated in Sec. 5.2, there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting. So we define two unsupervised baselines by combining AutoLink [12] with MiDaS [42], and Autolink [12] with Unsup3d [54]. Moreover, we compare against the supervised method 3DDFA [71] to understand the gap between unsupervised and supervised methods. Table 4 summarises the NME results, in which our method achieves better performance compared to the two-stage baselines. + +# 6. Ablation Study + +In this section we discuss our ablation studies. Results are shown in Fig. 7. The first and second row show the results + +with full training losses (i.e. Eq. (10)) with the number of keypoints being 32 and 16, respectively. We can see that with an increasing number of keypoints more detailed structure can be captured by our 3D keypoints. Furthermore, we remove $\mathcal{L}_{\mathrm{rep}}$ from our overall loss to train our framework (third row). The structure collapses into a degenerate keypoint configuration, which confirms the importance of $\mathcal{L}_{\mathrm{rep}}$ . + +A more detailed analysis on the behavior of $\mathcal{L}_{\mathrm{rep}}$ and on the influence of the number of basis functions $n$ of our shape model, are provided in the supplementary document. + +![](images/a695b4908032afebb8ae4149ae4dccf50bba929617161996c0adaef1275b137f.jpg) + +![](images/3a47ce8d8b0192ae1201ffc992bcdad54407985e46cdfb9eec3728730d7a4472.jpg) + +![](images/0bca5f09139afcb2be47e22a310418022829def2841db0bd1641d921836046f4.jpg) +Figure 7. Ablation study on CELEBA WILD dataset. The first and second rows show results for a different number of keypoints (32 and 16). The third row shows results without repulsion loss $\mathcal{L}_{\mathrm{rep}}$ . + +# 7. Discussion + +![](images/b2ba1e5d0f8a62ed02eea3d97d810c08cd08cfab4be9e549c9dc6e52525a3492.jpg) + +![](images/e955484b80f02718e7796fa5f7ebde463e01e1a774206f5dcc0de8b42426761a.jpg) +Figure 8. Pose estimation visualization of two instances from CELEBA WILD dataset from two views. + +Large pose estimation has always been an important yet challenging issue in many 3D shape inference settings + +[26, 33, 55, 56, 61, 63]. Recent works have found that when the pose variations falls into a small range, it can be learned directly by a network together with shape [53]. We observe similar results in our setting. For the CELEBA WILD, AFHQ (instances with moderate pose variations), CUB-200-2011 and HORSE (aligned facing direction) datasets, Fig. 4 and 5 have shown multiple instances in different poses with correct pose estimation. Fig. 8 gives more pose estimation results. However, for datasets consisting of instances with large pose variations, our method may fail to optimize for the correct poses. For example, our test on the HUMAN3.6M dataset [18] shows that, even though our model learns plausible human body structure, it fails to optimize for the correct pose of each instance, which also leads to incorrect instance-specific deformations, as shown in Fig. 9 – for such cases with strong articulation, domain-specific approaches that exploit domain knowledge (e.g. about the human kinematic skeleton structure) are at the time being more suitable. Yet, inferring such articulated objects is a highly relevant and interesting direction for future work. + +![](images/8a865d2682ea8093bd3b5787e2fe64aa85dca0a3df57dd9242a0bede5b795d30.jpg) +Figure 9. Incorrect pose estimation visualization of HUMAN3.6M dataset. For severely articulated objects additional domain knowledge is currently needed (e.g. a human body kinematic skeleton). + +# 8. Conclusion + +Overall, our method is the first unsupervised approach that can infer 3D keypoints from category-specific image collections without the need of any additional prior information. Conceptually, our main novelty lies in simultaneously learning 3D keypoints along with a 3D shape space model from an inhomogeneous collection of category-specific images (e.g. opposed to homogeneous collections of multiview images or videos, which show the same object from different views or in different poses). We experimentally demonstrate that our method not only improves upon 2D keypoint accuracy, but also infers plausible 3D structure for various object categories. We expect our work to be useful for diverse downstream tasks, for example related to 3D reconstruction or 3D shape-based generative models. + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3 +[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 3 +[3] Adrian Bulat, Tzimiropoulos, and Georgios. Two-stage convolutional part heatmap regression for the 1st 3d face alignment in the wild (3dfaw) challenge. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 616-624. Springer, 2016. 2, 3 +[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 3 +[5] Boyuan Chen, Pieter Abbeel, and Deepak Pathak. Unsupervised learning of visual 3d keypoints for control. In International Conference on Machine Learning, pages 1539-1549. PMLR, 2021. 2, 3 +[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5 +[7] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 7 +[8] Hui Deng, Tong Zhang, Yuchao Dai, Jiawei Shi, Yiran Zhong, and Hongdong Li. Deep non-rigid structure-from-motion: A sequence-to-sequence translation perspective. arXiv preprint arXiv:2204.04730, 2022. 3, 5 +[9] Aysegul Dundar, Kevin J Shih, Animesh Garg, Robert Pottorf, Anrew Tao, and Bryan Catanzaro. Unsupervised disentanglement of pose, appearance and background from images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3883-3894, 2021. 1, 2, 3 +[10] Chao Gou, Yue Wu, Fei-Yue Wang, and Qiang Ji. Shape augmented regression for 3d face alignment. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 604-615. Springer, 2016. 2, 3 +[11] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. arXiv preprint arXiv:2103.15812, 2021. 2, 3, 7 +[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. Advances in Neural Information Processing Systems, 35:36123-36141, 2022. 1, 2, 3, 4, 5, 6, 7 +[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image + +generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1225-1235, 2022. 3, 7 +[14] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21337-21348, 2023. 1, 2, 3 +[15] Sina Honari and Pascal Fua. Unsupervised 3d keypoint estimation with multi-view geometry. arXiv preprint arXiv:2211.12829, 2022. 2, 3 +[16] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662-8672, 2020. 7 +[17] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 7 +[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3, 8 +[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3, 7 +[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3 +[21] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 196-214. Springer, 2020. 3 +[22] Yunji Kim, Seonghyeon Nam, In Cho, and Seon Joo Kim. Unsupervised keypoint learning for guiding class-conditional video prediction. Advances in neural information processing systems, 32, 2019. 3 +[23] Chen Kong and Simon Lucey. Deep non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1558-1567, 2019. 3 +[24] Tejas D Kulkarni, Ankush Gupta, Catalin Ionescu, Sebastian Borgeaud, Malcolm Reynolds, Andrew Zisserman, and Volodymyr Mnih. Unsupervised learning of object keypoints for perception and control. Advances in neural information processing systems, 32, 2019. 3 +[25] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6152-6162, 2020. 1, 3 + +[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 8 +[27] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Unsupervised part segmentation through disentangling appearance and shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8355-8364, 2021. 7 +[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5 +[29] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2019. 5, 7 +[30] Y Ma, S Soatto, J Kosecka, and S Sastry. An invitation to 3d vision: From images to models. Springer verlag, 2003. 5 +[31] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3 +[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3 +[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 8 +[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 1 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[36] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 1, 3 +[37] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 2 +[38] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 5 + +[39] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression: A flexible alignment-based framework for non-rigid structure estimation. IEEE Transactions on Image Processing, 27(1):249-264, 2017. +[40] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression networks: Learning 3d structure of non-rigid objects from 2d annotations. In European Conference on Computer Vision, pages 1-18. Springer, 2020. 2, 3, 5 +[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2, 3 +[42] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 7 +[43] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 3 +[44] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7326-7335, 2019. 2, 3 +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1 +[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 3 +[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1 +[48] Suwajanakorn Supasorn, Noah Snavely, Jonathan J Tompson, and Mohammad Norouzi. Discovery of latent 3d keypoints via end-to-end geometric reasoning. Advances in neural information processing systems, 31, 2018. 2, 3 +[49] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3, 7 +[50] Sergey Tulyakov, László A Jeni, Jeffrey F Cohn, and Nicu Sebe. Viewpoint-consistent 3d face alignment. IEEE transactions on pattern analysis and machine intelligence, 40(9): 2250-2264, 2017. 2, 3 + +[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[52] Chaoyang Wang, Chen-Hsuan Lin, and Simon Lucey. Deep nrsfm++: Towards unsupervised 2d-3d lifting in the wild. In 2020 International Conference on 3D Vision (3DV), pages 12–22. IEEE, 2020. 3 +[53] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 8 +[54] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2, 7 +[55] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, pages 1-12, 2023. 1, 2, 8 +[56] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 1, 2, 8 +[57] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 3 +[58] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 1 +[59] Chenxin Xu, Siheng Chen, Maosen Li, and Ya Zhang. Invariant teacher and equivariant student for unsupervised 3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3013-3021, 2021. 1, 2, 3 +[60] Yinghao Xu, Ceyuan Yang, Ziwei Liu, Bo Dai, and Bolei Zhou. Unsupervised landmark learning from unpaired data. arXiv preprint arXiv:2007.01053, 2020. 3 +[61] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 8 +[62] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 1 +[63] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1323-1330. IEEE, 2021. 8 + +[64] Haitian Zeng, Yuchao Dai, Xin Yu, Xiaohan Wang, and Yi Yang. Pr-rrn: pairwise-regularized residual-recursive networks for non-rigid structure-from-motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5600–5609, 2021. 3 +[65] Haitian Zeng, Xin Yu, Jiaxu Miao, and Yi Yang. Mhr-net: Multiple-hypothesis reconstruction of non-rigid shapes from 2d views. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 3 +[66] Hongwen Zhang, Qi Li, and Zhenan Sun. Joint voxel and coordinate regression for accurate 3d facial landmark localization. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2202-2208. IEEE, 2018. 1, 2, 3 +[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3, 7 +[68] Ruiqi Zhao, Yan Wang, C Fabian Benitez-Quiroz, Yaojie Liu, and Aleix M Martinez. Fast and precise face alignment and 3d shape reconstruction from a single 2d image. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 590-603. Springer, 2016. 2, 3 +[69] Shihao Zhou, Mengxi Jiang, Qicong Wang, and Yunqi Lei. Towards locality similarity preserving to 3d human pose estimation. In Proceedings of the Asian Conference on Computer Vision, 2020. 1, 2, 3 +[70] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5 +[71] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3, 5, 7 \ No newline at end of file diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/images.zip b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2b377c6ba05467c7bf854ba3ce9028750f73ef5 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e50c85ff74de724446d58b5fa0deeb9cae701196236388ccd118bcd00ad94bed +size 805394 diff --git a/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/layout.json b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85a4958869f131565342c0796e5419459fefe860 --- /dev/null +++ b/2024/Unsupervised 3D Structure Inference from Category-Specific Image Collections/layout.json @@ -0,0 +1,9123 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "type": "text", + "content": "Unsupervised 3D Structure Inference from Category-Specific Image Collections" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 143, + 225, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 143, + 225, + 159 + ], + "spans": [ + { + "bbox": [ + 148, + 143, + 225, + 159 + ], + "type": "text", + "content": "Weikang Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 144, + 332, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 144, + 332, + 157 + ], + "spans": [ + { + "bbox": [ + 257, + 144, + 332, + 157 + ], + "type": "text", + "content": "Dongliang Cao" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 365, + 144, + 445, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 144, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 365, + 144, + 445, + 156 + ], + "type": "text", + "content": "Florian Bernard" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 163, + 370, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 163, + 370, + 177 + ], + "spans": [ + { + "bbox": [ + 223, + 163, + 370, + 177 + ], + "type": "text", + "content": "University of Bonn, Germany" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 221, + 541, + 343 + ], + "blocks": [ + { + "bbox": [ + 48, + 221, + 541, + 343 + ], + "lines": [ + { + "bbox": [ + 48, + 221, + 541, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 541, + 343 + ], + "type": "image", + "image_path": "f6a1958682f39e314355aaa2084552487e22c9f5c40e7f1cdf28107a1b914b82.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 377 + ], + "type": "text", + "content": "Figure 1. Our method detects consistent and reasonable 3D structure for general object categories from a single 2D image. For training, our method only requires a collection of images from the same category without any additional information." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "spans": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 414, + 290, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 290, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 290, + 643 + ], + "type": "text", + "content": "Understanding 3D object structure from image collections of general object categories remains a long-standing challenge in computer vision. Due to the high relevance of image keypoints (e.g. for graph matching, controlling generative models, scene understanding, etc.), in this work we specifically focus on inferring 3D structure in terms of sparse keypoints. Existing 3D keypoint inference approaches rely on strong priors, such as spatio-temporal consistency, multi-view images of the same object, 3D shape priors (e.g. templates, skeleton), or supervisory signals e.g. in the form of 2D keypoint annotations. In contrast, we propose the first unsupervised 3D keypoint inference approach that can be trained for general object categories solely from an inhomogeneous image collection (containing different instances of objects from the same category). Our experiments show that our method not only improves upon unsupervised 2D keypoint inference, but more importantly, it also produces reasonable 3D structure for various object categories, both qualitatively and quantitatively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 690, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 690, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 690, + 288, + 715 + ], + "type": "text", + "content": "Understanding 3D object structure from 2D images is a fundamental problem in computer vision that has been studied" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": "for decades. A range of different 3D structure representations have been considered, including keypoints [14, 25, 59, 69], meshes [55, 56, 62], and voxels [35, 45, 47]. Among these, keypoint-based 3D structure representations have the strong advantage that they are simple and flexible, while at the same time being highly relevant for diverse downstream tasks. However, obtaining 3D keypoints from 2D images is hard due to unknown depth, pose variations, different appearance, and the lack of explicit geometric information. Due to these difficulties, prior works mostly focus on 2D keypoints detection instead of 3D [9, 12, 31, 36, 66]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 523, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 548, + 715 + ], + "type": "text", + "content": "In this work we investigate whether it is possible to directly obtain 3D keypoints without using any explicit supervision. Our work is motivated by the fact that 2D image keypoints are observations of points in 3D space that are projected onto the 2D image plane. By explicitly understanding and inferring 3D geometry we are able to improve upon the quality of detected keypoints. Specifically, we experimentally demonstrate that projecting our predicted 3D keypoints onto the image plane leads to more reliable 2D keypoints compared to existing 2D keypoint predictors. Moreover, our method can serve the increasing amount of tasks that rely on 3D keypoints, e.g. 3D structure-aware generative models [58], 3D avatar generation [34], or 3D reconstruction [55, 56]. Despite keypoint prediction being a well-studied problem in the 2D domain, 3D keypoint inference is under-explored and we thus specifically contribute" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10704" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 120, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 120, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 120, + 83 + ], + "type": "text", + "content": "towards this area." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 85, + 290, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 85, + 290, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 290, + 373 + ], + "type": "text", + "content": "Existing works on inferring 3D keypoints either use 3D point clouds or 2D images/videos from the same object instance as input data. While the former is conceptually a subset selection problem, in this work we solely focus on the more challenging latter case. Manually obtaining 3D keypoints in image data is much harder than annotating 2D keypoints. Thus, there are only very few datasets with 3D keypoints annotations, which in turn motivates unsupervised learning strategies that do not require expensive or even infeasible 3D annotations for training. Existing works that address 3D keypoint inference from image data use for example 2D keypoint annotations [38, 44], multiple views with known poses [11, 35, 48], spatial-temporal consistency in videos [9, 15], 2D/3D skeleton [55, 56] or templates [14]. In stark contrast, in this work we relax the assumptions about the availability of such supervisory signals and train a 3D keypoint predictor merely from category-specific image collections as input. Our method builds upon the assumption that object instances from the same category should have a similar 3D structure with moderate intra-class variations (e.g. faces, horses, birds). Table 1 summaries the differences among existing methods and Fig. 2 gives a visualization of some additional information within these methods and the comparison to our setting." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 384, + 287, + 647 + ], + "blocks": [ + { + "bbox": [ + 48, + 384, + 287, + 647 + ], + "lines": [ + { + "bbox": [ + 48, + 384, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 287, + 647 + ], + "type": "table", + "html": "
MethodsUnsup.w/o PriorCat.-agnostic
Bulat et al. [3]X✓✓
Gou et al. [10]X✓✓
Zhao et al. [68]X✓✓
Tulyakov et al. [50]X✓✓
Pavlakos et al. [41]X✓X
Zhang et al. [66]X✓X
He et al. [14]XX✓
Mildenhall et al. [35]✓X✓
Wu et al. [54]✓X✓
Moniz et al. [37]✓XX
Supasorn et al. [48]✓X✓
Novotny et al. [38]✓X✓
Park et al. [40]✓X✓
Chen et al. [5]✓X✓
Reddy et al. [44]✓X✓
Zhou et al. [69]✓✓X
Dundar et al. [9]✓✓X
Xu et al. [59]✓✓X
Honari and Fua [15]✓X✓
Ours✓✓✓
", + "image_path": "81e0bea5678df96e27ad1db1e5c7ed817d14febc41a76fe50fbacdd07b687111.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 654, + 287, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 711 + ], + "type": "text", + "content": "Table 1. In comparison to existing 3D keypoints inference methods, our approach combines a unique set of desirable properties: it can be trained in an unsupervised manner without requiring any additional information; it does not require any shape prior and can thus be applied to different object categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 273 + ], + "type": "image", + "image_path": "31017baa371f382d7b690fc43418510f203ad267aa832ca5db33badeed8922a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 283, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 283, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 546, + 350 + ], + "type": "text", + "content": "Figure 2. (Top row from left to right) Existing methods require different additional information for 3D structure inference, e.g. multiple views (from [12]), 2D keypoint annotations (from [12]), and a symmetry prior (from [54]). (Bottom row) Our setting does not require such information but instead also infers a 3D shape space model that is learned in an unsupervised manner." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 366, + 546, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 546, + 520 + ], + "type": "text", + "content": "Our method enables to infer 3D keypoints from category-specific image collections without relying on additional prior information. To this end, we build upon the recent unsupervised 2D keypoint detection method by He et al. [12] and extend it by additionally learning a sparse 3D shape space model to represent 3D structure. To regularise the 3D structure, a self-supervised re-projection loss is introduced to couple 3D keypoints with corresponding 2D keypoints. By combining our 3D shape space model, re-projection loss, and additional regularisation loss terms, we demonstrate that our proposed method not only improves 2D keypoint accuracy, but also produces reasonable 3D structure. We summarise our main contributions as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 527, + 545, + 635 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 527, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 545, + 563 + ], + "type": "text", + "content": "- For the first time we obtain 3D keypoints from general category-specific image collections without relying on additional prior information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 563, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 545, + 598 + ], + "type": "text", + "content": "- Our method obtains reasonable 3D keypoints based on our learnable 3D shape space model that is fused with a state-of-the-art unsupervised 2D keypoint detector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 599, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 545, + 635 + ], + "type": "text", + "content": "- We demonstrate that our method both improves unsupervised 2D keypoint detection accuracy and obtains reasonable 3D shape on different benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 663, + 392, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 392, + 675 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 392, + 675 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "In the following we summarize relevant 2D and 3D key-point prediction methods." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10705" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 176, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 176, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 176, + 85 + ], + "type": "text", + "content": "2.1. 2D Keypoint Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "content": "Supervised methods. Various domain-specific supervised 2D keypoint detection methods have been proposed that are trained on a large amount of labeled datasets, e.g. for faces [2, 57, 71], human bodies [1, 4, 18, 21, 32], or vehicles [43], etc. However, labeling keypoints for images is time-consuming, laborious and error-prone, especially when there are occlusions, extreme pose variations, or appearance changes. As a consequence, labeling errors will subsequently influence network training and downstream task performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 289, + 617 + ], + "type": "text", + "content": "Unsupervised methods. Due to these reasons, an increasing amount of attention is paid to unsupervised methods. One popular concept is multi-view geometric consistency. Thewlis et al. [49] detect keypoints by recovering transformations between one image and its deformed counterpart from pre-defined deformations. Without relying on pre-defined deformations, various video-based methods [9, 19, 20, 22, 24, 36, 46] infer keypoints based on a temporal coherence prior. Instead of using multiple views as input, Xu et al. [60] use unpaired input to disentangle appearance and poses by conducting a swapping-reconstruction strategy and encouraging reconstructed images to resemble the original. He et al. [13] and He et al. [11] use a generative adversarial training framework to constrain keypoint locations by interpreting them as latent codes for an image generator. However, collaboratively learning intermediate keypoints and complex adversarial models is unstable and does not scale. Starting from some initial point sets, Mallis et al. [31] filter keypoints by optimizing alternatively between clustering of keypoints features across a dataset, and training a keypoint detector with pseudo-labels offered by cluster centers. However, this method depends heavily on the initial point set. Zhang et al. [67] address single-image keypoint inference using an auto-encoder architecture to interpret keypoints as a latent representation, while reconstructing the input image from the concatenated keypoints and respective pixel feature descriptors. He et al. [12] use a similar idea by representing keypoints as latent features in an auto-encoder manner. However, they reconstruct the input from an edge map created from keypoints combined with random masking of appearance information. In this way, paired input data is not needed to disentangle appearance and poses, which makes this method applicable in wider scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 623, + 176, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 176, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 176, + 636 + ], + "type": "text", + "content": "2.2.3D Keypoint Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "Compared to 2D keypoints detection, 3D keypoints detection is an under-explored topic, particularly in unsupervised settings without strong priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Supervised methods. Some earlier methods [3, 10, 68] use a two-stage strategy that infers 2D keypoints first, and then predicts depth for these keypoints. These methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "not only need 3D keypoints or depth annotations for training, but are also sub-optimal due to their two-stage procedure. Tulyakov et al. [50] generalize a cascaded regression method to 3D for estimating 3D face landmarks in an end-to-end manner from images. Pavlakos et al. [41] generalize 2D heatmaps directly to 3D by inferring 3D keypoints, while Zhang et al. [66] improve this method via a joint voxel and coordinate regression to avoid the curse of dimensionality. However, these methods all needs 3D keypoints annotations, so that their scope of application is limited due to the absence of annotated 3D keypoints in many cases. Reddy et al. [44] propose a trifocal loss to learn (possibly occluded) 3D keypoints from multiple views by using 2D keypoints annotations. Closely related to 3D shape inference, sparse non-rigid structure-from-motion methods also take 2D keypoints as input to infer 3D structure of objects [8, 23, 38-40, 52, 64, 65]. Based on [12], He et al. [14] predict 3D keypoints using dozens of images with 2D annotations together with a skeleton template." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "type": "text", + "content": "Unsupervised methods. In order to get rid of suboptimal two-stage approaches or 2D/3D keypoints annotations, some works aim at unsupervised 3D structure inference. Supasorn et al. [48] use paired views and known transformations between them as input to infer 3D keypoints. Chen et al. [5] use multiple views as input to infer 3D keypoints by averaging 2D keypoints and depth values detected in each view. Using a pair of views as input, Honari and Fua [15] propose a two-stream auto-encoder to encode 2D keypoints of each view, and then use triangulation to lift two sets of 2D keypoints to 3D. Various unsupervised 3D human pose estimation methods [25, 59, 69], either use multiple views or video as input, or they rely on some domain-specific losses as constraints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 467, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 539 + ], + "type": "text", + "content": "Different from all mentioned 3D keypoints inference settings, our method does not rely on 2D/3D keypoints annotations, 3D template/priors/skeletons, multiple views/video, or domain-specific 3D structure. Instead, for the first time we infer 3D structure only from category-specific collections of images for a broad range of object categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 550, + 384, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 384, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 384, + 563 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "content": "Our method builds upon the recent AutoLink method [12] that addresses the task of unsupervised keypoint detection in 2D. In this section we give a brief overview of AutoLink." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "AutoLink comprises an auto-encoder with 2D keypoints as bottleneck latent embeddings. For training, it employs a reconstruction loss that compares the input image with a reconstructed image that is obtained based on edge maps and a randomly masked input image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Formally, given an image " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", the encoder takes " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " as input and outputs " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " heatmaps, denoted as " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{1}, H_{2}, \\ldots, H_{K}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "H_{i} \\in [0,1]^{W \\times H}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th heatmap represents the likelihood for each pixel to be the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10706" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 205 + ], + "type": "image", + "image_path": "0ad3c2b3e0c004ec57e8380182823bf433d7ec03f3a63d3ad70b16ec9b91fd2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": "Figure 3. Overview of our pipeline. We sample images " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " from a category-specific image collection, which are then piped into an encoder that predicts " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " heatmaps. These heatmaps are used to extract a " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "K \\times 2" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " keypoint matrix " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "P^{2D}" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ". From " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "P^{2D}" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ", an edge map " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " is constructed, which is then, together with a randomly masked version of the input image " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{masked}}" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ", fed into the decoder to reconstruct image " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ". Moreover, an MLP takes " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "P^{2D}" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " as input and predicts with different heads, the coefficients " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ", the 3D translation " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ", and the 3D rotation " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ". At the same time we optimize for a globally trainable mean shape " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": " and basis " + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 215, + 547, + 293 + ], + "type": "text", + "content": ", which define a 3D shape space model used to represent our 3D keypoints, see Eq.(5). The main difference to AutoLink [12], which predicts 2D keypoints only, is that we infer 3D keypoints and a 3D shape space model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "text", + "content": "specified keypoint " + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "text", + "content": " represents the total number of different keypoints. With that, the location of each 2D keypoint " + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "inline_equation", + "content": "P_{i}^{2\\mathrm{D}} = (x_{i},y_{i}), i = 1,2,\\ldots ,K" + }, + { + "bbox": [ + 46, + 313, + 287, + 359 + ], + "type": "text", + "content": " can be represented as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 357, + 287, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 357, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 114, + 357, + 287, + 387 + ], + "type": "interline_equation", + "content": "P _ {i} ^ {\\mathrm {2 D}} = \\sum_ {p} \\frac {H _ {i} (p)}{\\sum_ {p ^ {\\prime}} H _ {i} \\left(p ^ {\\prime}\\right)} p, \\tag {1}", + "image_path": "05a37dfad489270f2c2fc4c386bfc49893a02fb2bd777677ca2363821f2788f4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "inline_equation", + "content": "p,p^{\\prime}\\in [-1,1]\\times [-1,1]" + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "content": " are normalized pixel coordinates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": "For each pair of 2D keypoints, " + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "inline_equation", + "content": "P_{i}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "inline_equation", + "content": "P_{j}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "inline_equation", + "content": "1 \\leq i,j \\leq K" + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": ", the method constructs an edge map " + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "inline_equation", + "content": "S_{ij} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": " defined on normalized pixel coordinates " + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 414, + 287, + 452 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 457, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 457, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 108, + 457, + 287, + 472 + ], + "type": "interline_equation", + "content": "S _ {i j} (p) = \\exp \\left(- d _ {i j} (p) / \\sigma^ {2}\\right), \\tag {2}", + "image_path": "a9d03fe5398a1510f481e68046c482c29a09b97afbc360f2d17217fb9ab8e866.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "d_{ij}(p)" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " distance of pixel " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " to the line segment connecting " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "P_{i}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "P_{j}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " controls the thickness of the edge. The final edge map " + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 47, + 477, + 287, + 526 + ], + "type": "text", + "content": " summarizing all edges between all keypoints is given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 533, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 533, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 111, + 533, + 287, + 552 + ], + "type": "interline_equation", + "content": "S (p) = \\max _ {1 \\leq i, j \\leq K} \\omega_ {i j} S _ {i j} (p), \\tag {3}", + "image_path": "966d174bc753d16224670b6b805a7d81fcacdba894b011f97133451a7a251aec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 556, + 287, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 556, + 287, + 580 + ], + "type": "inline_equation", + "content": "\\omega_{ij}" + }, + { + "bbox": [ + 47, + 556, + 287, + 580 + ], + "type": "text", + "content": " are learnable parameters that are shared by all images and optimized during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "text", + "content": "Finally, the edge map " + }, + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "text", + "content": " and the (randomly masked) input image " + }, + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{masked}}" + }, + { + "bbox": [ + 47, + 580, + 287, + 639 + ], + "type": "text", + "content": " are concatenated along the channel axis and fed into a decoder to reconstruct the input. The whole AutoLink pipeline is trained end-to-end only based on the reconstruction loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 87, + 647, + 287, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 647, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 87, + 647, + 287, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\| F (D (S \\odot I _ {\\text {m a s k e d}})) - F (I) \\| _ {2}, \\tag {4}", + "image_path": "65c71b1a87219b8f77ad12292c2fc73b2e480c2b0d5f7e260f143b86c50b794e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": " is the decoder, " + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": " is a pre-trained and fixed feature extractor (i.e. VGG network), " + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": " represents channel-wise concatenation, and " + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{masked}}" + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": " is the randomly masked original input image with " + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": " mask area." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 312, + 511, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 511, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 511, + 325 + ], + "type": "text", + "content": "4. Unsupervised 3D Structure Inference" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "content": "In the following we explain our method. To this end, we introduce our 3D shape space model for inferring 3D keypoint structure in Section 4.1. In section 4.2 we explain estimating the object pose. Section 4.3 introduces our unsupervised losses. Our overall pipeline is summarized in Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 397, + 484, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 484, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 484, + 410 + ], + "type": "text", + "content": "4.1. Trainable 3D Shape Space Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "text", + "content": "Rather than working with high-dimensional and expensive 3D heatmaps, we infer 3D geometry by training a 2D keypoint predictor together with a 3D shape space model in an end-to-end manner, see Fig. 3. Our encoder first predicts 2D keypoint heatmaps from a given 2D image, which are then transformed into a 2D keypoint matrix using Eq. (1). Then a multi-layer perceptron (MLP) takes the 2D keypoint matrix as input to predict input-dependent coefficients " + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "text", + "content": " of a 3D shape model. At the same time, the mean shape " + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "text", + "content": " and the set of basis vectors " + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 415, + 545, + 559 + ], + "type": "text", + "content": " of this 3D shape space model are optimized as global trainable parameters. Overall, this allows to obtain the 3D keypoints (in canonical pose) as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 563, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 563, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 367, + 563, + 545, + 578 + ], + "type": "interline_equation", + "content": "P _ {\\text {c a n o n i c a l}} ^ {\\mathrm {3 D}} = \\operatorname {m a t} (M + \\alpha B), \\tag {5}", + "image_path": "1fc9ef0f4cdd2c044dd32e05924971d3685f83ac33a38c88c8eefad4e8bd2b18.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " is the mean shape of size " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "1 \\times 3K" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " is the basis matrix of size " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "n \\times 3K" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " are the coefficients of size " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "1 \\times n" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " denoting the number of basis functions. The operation " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{mat}()" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " reshapes a " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "1 \\times 3K" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " row vector to a " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "K \\times 3" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " matrix. By training these components simultaneously with our 2D heatmap predictor, we can ensure that during training our 2D heatmap predictor receives gradients that contain explicit information about 3D geometry." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We emphasize that opposed to most existing 3D shape space learning methods that require known keypoint locations, we simultaneously optimize for keypoint locations" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10707" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "while using our trainable 3D shape space model as 3D geometry-aware regularizer. Unlike non-rigid structure-from-motion methods [8, 38-40] that utilize similar ideas, our approach does not require multiple images of the same object from different views. Instead, we use an inhomogeneous collection of category-specific images, where the 3D shape space model serves as flexible and adaptive shape prior that can adjust to the different geometries of objects (within the considered object category). As such, our approach is much more flexible, since it can handle single-view image collections as for example available in web collections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 145, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 145, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 145, + 234 + ], + "type": "text", + "content": "4.2. Pose Estimation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": "After getting the 3D keypoints " + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{canonical}}^{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": " (in canonical pose) using Eq. (5), an input-dependent transformation matrix and scaling factor are estimated that transform " + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{canonical}}^{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": " to the posed 3D keypoints " + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "inline_equation", + "content": "P^{3D}" + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": ", such that the projection of " + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "inline_equation", + "content": "P^{3D}" + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": " onto the 2D image plane coincides with the 2D structure of the object within the 2D image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "content": "Non-rigid object deformations are taken care of by the 3D shape model, so that we represent the object pose solely via a rigid body transformation matrix " + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "inline_equation", + "content": "(R,T)\\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "content": ". With the 2D keypoint matrix " + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "inline_equation", + "content": "P^{2D}" + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "content": " as input, an MLP computes the translation " + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "inline_equation", + "content": "T\\in \\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "content": " and the rotation angles " + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 46, + 313, + 288, + 397 + ], + "type": "text", + "content": ", which are converted to a rotation matrix via the Rodrigues formula [30]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 415, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 78, + 415, + 287, + 440 + ], + "type": "interline_equation", + "content": "R = I _ {3} + \\frac {\\sin (\\theta)}{\\theta} [ \\Theta ] _ {\\times} + \\frac {1 - \\cos (\\theta)}{\\theta^ {2}} [ \\Theta ] _ {\\times} ^ {2}, \\tag {6}", + "image_path": "1abf39ba85d3021ee5a01c31070388ae620faf80df843c4417ade9b819aa1f4b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\theta = \\| \\Theta \\|" + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "content": " and the skew-symmetric operator " + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "inline_equation", + "content": "[\\cdot ]_{\\times}" + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "content": " turns a vector in " + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "content": " into a skew-symmetric matrix. Finally the 3D keypoint matrix " + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "inline_equation", + "content": "P^{3D}" + }, + { + "bbox": [ + 46, + 445, + 287, + 483 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 490, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 490, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 112, + 490, + 287, + 506 + ], + "type": "interline_equation", + "content": "P ^ {3 \\mathrm {D}} = P _ {\\text {c a n o n i c a l}} ^ {3 \\mathrm {D}} R + \\mathbf {1} _ {K} T, \\tag {7}", + "image_path": "f68a4e479b087bdac6a5b455c1e53abdf7f28ff8bbb890c39d36496d3898e3cc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_K" + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "text", + "content": "-dimensional vector of all ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 532, + 167, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 167, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 167, + 544 + ], + "type": "text", + "content": "4.3. Unsupervised Losses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 551, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 586 + ], + "type": "text", + "content": "Our re-projection loss penalizes the discrepancy between the projected 3D keypoints and the 2D heatmap keypoint predictions, i.e." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 594, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 287, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p r o j}} = \\| P ^ {3 \\mathrm {D}} \\Pi - P ^ {2 \\mathrm {D}} \\| _ {F}, \\quad \\text {w i t h} \\quad \\Pi = \\left[ \\begin{array}{c c} s _ {1} & 0 \\\\ 0 & s _ {2} \\\\ 0 & 0 \\end{array} \\right] \\tag {8}", + "image_path": "67e1b8182c820dec902e49763a54c7151ca08b2a459b8192474a9674c056fffa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "being a simple orthographic projection (of " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "P^{3D}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " from Eq. (8)) combined with scaling " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "s = (s_1, s_2)" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", which is predicted by an MLP. This loss ensures consistent 2D keypoint predictions between different images, since the 3D shape structure is shared across the whole image collection via the 3D shape space model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "However, combining " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " alone will lead to degenerate solutions (i.e. all keypoints converge to a single location). In order to avoid this situation, we consider the repulsion loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 125, + 545, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 125, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 545, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e p}} = - \\sum_ {i = 1} ^ {K} \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} \\exp (- \\| p _ {i} - \\mathcal {N} (p _ {i}) \\| _ {2} / h), \\tag {9}", + "image_path": "cb27a03cb027e277e55c88c8bcefd68839685f031ed212503614c62a6dc352dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": " is a 2D keypoint (a row of " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "P^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_i)" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": " is the nearest keypoint (other than " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": ") of " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "P^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": " is a temperature parameter that controls the repulsion decrease rate. The effect of Eq. (9) is that neighboring keypoints do not move too close to each other." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 228, + 522, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 522, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 522, + 239 + ], + "type": "text", + "content": "Our total loss combines all terms and is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 246, + 545, + 260 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 246, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 365, + 246, + 545, + 260 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e p}}, \\tag {10}", + "image_path": "0313a4f41e2a6080b9f74a94bb25cb172dac39b545ee2c859b16b2daec7c36a8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 265, + 504, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 504, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 265, + 504, + 277 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 304, + 265, + 504, + 277 + ], + "type": "text", + "content": " are weights of the individual terms." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 286, + 388, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 388, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 388, + 300 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "content": "We evaluate our method on various datasets and compare to existing keypoint prediction methods. Implementation details are provided in the supplementary document." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 348, + 422, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 422, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 422, + 361 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our method on six datasets: CELEBA WILD [28], CUB-200-2011 [51], 300W-LP [71], AFLW2000-3D [71], HORSE [70] and AFHQ [6]. The CELEBA WILD dataset [28] contains celebrity faces in in-the-wild environments, along with 2D keypoint annotations for each image. Following the setting of He et al. [12], we remove the images whose face covers less than " + }, + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 366, + 545, + 713 + ], + "type": "text", + "content": " of the area. The 300W-LP [71] dataset is a large synthetic face dataset containing 61,225 samples across large poses (1,786 from IBUG, 5,207 from AFW, 16,556 from LFPW and 37,676 from HELEN). The AFLW2000-3D dataset was introduced along with the 300W-LP dataset by the same data acquisition method, containing 2,000 facial images, each with 68 3D keypoints annotations. We train on the 300W-LP dataset and test on the AFLW2000-3D dataset to evaluate the quantitative 3D keypoint detection performance. We use subsets from the AFHQ dataset [6], namely CAT, DOG, TIGER, FOX, CHEETAH, WOLF, JAGUAR and LION, to get qualitative 3D structure inference results. CUB-200-2011 consists of 11,788 images of birds. Following the settings in [12, 29], we align all birds facing left, and remove seabirds. The HORSE dataset is extracted from the CycleGAN dataset [70] by removing images with multiple horses and aligning them to face left. CUB-200-2011, HORSE, AFHQ and CELEBA WILD are used to generate qualitative results. We quantitatively evaluate unsupervised 2D keypoint inference using the CELEBA WILD dataset and compare it with the recent state-of-the-art method by He et al. [12]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10708" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 329 + ], + "type": "image", + "image_path": "93cb6c61d32f561f61dfe16c14469edbb4f80ac17e3292d8fe053419ee327bbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 548, + 384 + ], + "type": "text", + "content": "Figure 4. Visualization of the 3D structure for different instances of the AFHQ dataset (CAT, DOG, FOX and TIGER subset in each row). Each 3D structure is shown in two forms: (i) 3D keypoints with linking edges between them (the edge thickness is proportional to the edge map coefficients " + }, + { + "bbox": [ + 46, + 338, + 548, + 384 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 338, + 548, + 384 + ], + "type": "text", + "content": ", which indicate their relative importance); (ii) Delaunay triangulation of the 3D keypoints rendered as shaded mesh. Different keypoints are shown in different colors to demonstrate consistency across different images (columns)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 396, + 529, + 544 + ], + "blocks": [ + { + "bbox": [ + 50, + 396, + 529, + 544 + ], + "lines": [ + { + "bbox": [ + 50, + 396, + 529, + 544 + ], + "spans": [ + { + "bbox": [ + 50, + 396, + 529, + 544 + ], + "type": "image", + "image_path": "f2b08cef81a9de29cca76b6e277f14809d743c8a15a112e845a5655263b8eadb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 547, + 512, + 559 + ], + "lines": [ + { + "bbox": [ + 82, + 547, + 512, + 559 + ], + "spans": [ + { + "bbox": [ + 82, + 547, + 512, + 559 + ], + "type": "text", + "content": "Figure 5. Visualization of the 3D structure for different images from the HORSE and CUB-200-2011 datasets (rows)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "text", + "content": "Metrics. For unsupervised 2D keypoint inference, we follow the metrics used by He et al. [12], which is the normalized mean " + }, + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "text", + "content": " error (NME) of detected 2D keypoints aligned to the ground truth. For the 3D case, we generalize NME to 3D by calculating the normalized mean " + }, + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 579, + 289, + 675 + ], + "type": "text", + "content": " of detected 3D keypoints aligned to ground truth 3D keypoints. In both cases, the normalization is the ocular distances of each image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 578, + 419, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 419, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 419, + 592 + ], + "type": "text", + "content": "5.2. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We visualize the learned 3D structure on four subsets of AFHQ (Cat, Dog, Fox and Tiger; visualization of results of other subsets are in supplementary document), CUB-200-2011, and HORSE datasets in Fig. 4 and 5, respectively. From the visualizations we can directly recognize the 3D structure of each shown object category. For the four subsets of AFHQ dataset, despite these animal categories sharing many similarities, we can recognize the individual animal face characteristics from predicted 3D keypoints with" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10709" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "linking edges. Moreover, we find that the set of thick edges, i.e. the ones with large learned edge coefficient " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", coincide with the most notable structures of each category, while negligible edges remain very thin. Finally, we demonstrate that keypoints are consistently detected among different instances of objects per category, which can be seen based on the consistency of the color of keypoints (both in 2D and 3D). More qualitative results and videos are included in the supplementary document and attached files." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 289, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 280 + ], + "type": "text", + "content": "Since there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting (cf. Table 1), as comparison we consider a simple baseline that first predicts 2D keypoints using AutoLink [12], and then lifts 2D keypoints to 3D by pinpointing learned 2D keypoints on the depth map (given by the pre-trained unsupervised monocular depth estimator MiDaS [42]). Fig. 6 shows a comparison on CELEBA WILD dataset." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 296, + 286, + 495 + ], + "blocks": [ + { + "bbox": [ + 47, + 296, + 286, + 495 + ], + "lines": [ + { + "bbox": [ + 47, + 296, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 286, + 495 + ], + "type": "image", + "image_path": "0981cd9de79f2f502192e66d557acf9c80fd59f00a18e2cf79b4899d710636a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 504, + 288, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 504, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 288, + 548 + ], + "type": "text", + "content": "Figure 6. Comparison of predicted 3D structures between our method (left) and AutoLink [12] combined with MiDaS [42] (right). Our method obtains more consistent 2D keypoint predictions and more realistic 3D structure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 583, + 167, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 167, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 167, + 596 + ], + "type": "text", + "content": "5.3. Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "2D keypoint inference. First, we show that our method improves upon unsupervised 2D keypoint inference. Table 2 shows the NME values of our method and of various other unsupervised methods on the CELEBA WILD dataset. We also compare to the method by He et al. [12] for various choices of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " on the CELEBA WILD dataset to show the robustness of our method, as shown in Table 3. We can see that our method outperforms all competitors across all settings." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 541, + 264 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 264 + ], + "type": "table", + "html": "
MethodK=8
DFF [7]31.30%*
SCOPS (w/o saliency) [17]22.11%†
SCOPS (w/saliency) [17]15.01%†
Liu et al. [27]12.26%†
Huang et al. [16]8.4%†
GANSeg [13]6.18%†
Thewlis et al. [49]31.30%*
Zhang et al. [67]40.82%*
LatentKeypointGAN [11]21.90%†
LatentKeypointGAN-tuned [11]5.63%†
Lorenz et al. [29]11.41%‡
IMM [19]8.74%‡
AutoLink [12]5.39%
Ours5.21%
", + "image_path": "b4dcfd7d301c066b9664802ef902f27cb26a6d75692bf50e0e9bd330656950ae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 324, + 542, + 373 + ], + "blocks": [ + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "type": "text", + "content": "Table 2. Normalized " + }, + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "type": "text", + "content": " error (NME) for 2D keypoints inference of various unsupervised methods on CELEBA WILD datasets for " + }, + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 305, + 271, + 547, + 312 + ], + "type": "text", + "content": ". (* reported from Collins et al. [7]; † reported from He et al. [12]; ‡ reported from Liu et al. [27])" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 324, + 542, + 373 + ], + "lines": [ + { + "bbox": [ + 310, + 324, + 542, + 373 + ], + "spans": [ + { + "bbox": [ + 310, + 324, + 542, + 373 + ], + "type": "table", + "html": "
MethodK=8K=16K=24K=32
AutoLink [12]5.39%4.69%3.99%3.77%
Ours5.21%3.97%3.54%3.48%
", + "image_path": "111c6b0b704dd0ee8be16722996fedc3cdb1c6b524012811f646efda75f79fa0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 419, + 543, + 460 + ], + "blocks": [ + { + "bbox": [ + 305, + 380, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 414 + ], + "type": "text", + "content": "Table 3. Normalized " + }, + { + "bbox": [ + 305, + 380, + 545, + 414 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 380, + 545, + 414 + ], + "type": "text", + "content": " error (NME) of AutoLink and our method for different numbers of keypoints using the CELEBA WILD dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 419, + 543, + 460 + ], + "lines": [ + { + "bbox": [ + 309, + 419, + 543, + 460 + ], + "spans": [ + { + "bbox": [ + 309, + 419, + 543, + 460 + ], + "type": "table", + "html": "
Supervised 3DDFAUnsupervised
AutoLink+Unsup3dAutoLink+MiDaSOurs
4.94%11.47%9.23%8.48%
", + "image_path": "d1c7fd5b453c2473f2e0b3682a8675f7dc925b99f9ea7f52d40f93ad375d14c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "type": "text", + "content": "Table 4. Normalized " + }, + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 468, + 545, + 523 + ], + "type": "text", + "content": " error (NME) of our method, two unsupervised methods (AutoLink + MiDaS and AutoLink + Unsup3d) and one supervised method (3DDFA) for 3D keypoints inference (training on the 300W-LP dataset and testing on the AFLW2000-3D dataset)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 535, + 547, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 547, + 656 + ], + "type": "text", + "content": "3D keypoint inference. As stated in Sec. 5.2, there does not exist any other 3D keypoint inference method that is applicable in our unsupervised setting. So we define two unsupervised baselines by combining AutoLink [12] with MiDaS [42], and Autolink [12] with Unsup3d [54]. Moreover, we compare against the supervised method 3DDFA [71] to understand the gap between unsupervised and supervised methods. Table 4 summarises the NME results, in which our method achieves better performance compared to the two-stage baselines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 669, + 399, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 399, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 399, + 682 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": "In this section we discuss our ablation studies. Results are shown in Fig. 7. The first and second row show the results" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10710" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "with full training losses (i.e. Eq. (10)) with the number of keypoints being 32 and 16, respectively. We can see that with an increasing number of keypoints more detailed structure can be captured by our 3D keypoints. Furthermore, we remove " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rep}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " from our overall loss to train our framework (third row). The structure collapses into a degenerate keypoint configuration, which confirms the importance of " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rep}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "text", + "content": "A more detailed analysis on the behavior of " + }, + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rep}}" + }, + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "text", + "content": " and on the influence of the number of basis functions " + }, + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 157, + 286, + 192 + ], + "type": "text", + "content": " of our shape model, are provided in the supplementary document." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 207, + 286, + 259 + ], + "blocks": [ + { + "bbox": [ + 51, + 207, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 51, + 207, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 51, + 207, + 286, + 259 + ], + "type": "image", + "image_path": "a695b4908032afebb8ae4149ae4dccf50bba929617161996c0adaef1275b137f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 266, + 285, + 314 + ], + "blocks": [ + { + "bbox": [ + 51, + 266, + 285, + 314 + ], + "lines": [ + { + "bbox": [ + 51, + 266, + 285, + 314 + ], + "spans": [ + { + "bbox": [ + 51, + 266, + 285, + 314 + ], + "type": "image", + "image_path": "3a47ce8d8b0192ae1201ffc992bcdad54407985e46cdfb9eec3728730d7a4472.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 319, + 274, + 357 + ], + "blocks": [ + { + "bbox": [ + 51, + 319, + 274, + 357 + ], + "lines": [ + { + "bbox": [ + 51, + 319, + 274, + 357 + ], + "spans": [ + { + "bbox": [ + 51, + 319, + 274, + 357 + ], + "type": "image", + "image_path": "0bca5f09139afcb2be47e22a310418022829def2841db0bd1641d921836046f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 286, + 402 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 286, + 402 + ], + "type": "text", + "content": "Figure 7. Ablation study on CELEBA WILD dataset. The first and second rows show results for a different number of keypoints (32 and 16). The third row shows results without repulsion loss " + }, + { + "bbox": [ + 46, + 369, + 286, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rep}}" + }, + { + "bbox": [ + 46, + 369, + 286, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 427, + 117, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 117, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 117, + 439 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 463, + 272, + 542 + ], + "blocks": [ + { + "bbox": [ + 56, + 463, + 272, + 542 + ], + "lines": [ + { + "bbox": [ + 56, + 463, + 272, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 272, + 542 + ], + "type": "image", + "image_path": "b2ba1e5d0f8a62ed02eea3d97d810c08cd08cfab4be9e549c9dc6e52525a3492.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 556, + 271, + 640 + ], + "blocks": [ + { + "bbox": [ + 56, + 556, + 271, + 640 + ], + "lines": [ + { + "bbox": [ + 56, + 556, + 271, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 271, + 640 + ], + "type": "image", + "image_path": "e955484b80f02718e7796fa5f7ebde463e01e1a774206f5dcc0de8b42426761a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 651, + 286, + 673 + ], + "lines": [ + { + "bbox": [ + 46, + 651, + 286, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 286, + 673 + ], + "type": "text", + "content": "Figure 8. Pose estimation visualization of two instances from CELEBA WILD dataset from two views." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Large pose estimation has always been an important yet challenging issue in many 3D shape inference settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 322 + ], + "type": "text", + "content": "[26, 33, 55, 56, 61, 63]. Recent works have found that when the pose variations falls into a small range, it can be learned directly by a network together with shape [53]. We observe similar results in our setting. For the CELEBA WILD, AFHQ (instances with moderate pose variations), CUB-200-2011 and HORSE (aligned facing direction) datasets, Fig. 4 and 5 have shown multiple instances in different poses with correct pose estimation. Fig. 8 gives more pose estimation results. However, for datasets consisting of instances with large pose variations, our method may fail to optimize for the correct poses. For example, our test on the HUMAN3.6M dataset [18] shows that, even though our model learns plausible human body structure, it fails to optimize for the correct pose of each instance, which also leads to incorrect instance-specific deformations, as shown in Fig. 9 – for such cases with strong articulation, domain-specific approaches that exploit domain knowledge (e.g. about the human kinematic skeleton structure) are at the time being more suitable. Yet, inferring such articulated objects is a highly relevant and interesting direction for future work." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 334, + 523, + 454 + ], + "blocks": [ + { + "bbox": [ + 310, + 334, + 523, + 454 + ], + "lines": [ + { + "bbox": [ + 310, + 334, + 523, + 454 + ], + "spans": [ + { + "bbox": [ + 310, + 334, + 523, + 454 + ], + "type": "image", + "image_path": "8a865d2682ea8093bd3b5787e2fe64aa85dca0a3df57dd9242a0bede5b795d30.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 464, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 498 + ], + "type": "text", + "content": "Figure 9. Incorrect pose estimation visualization of HUMAN3.6M dataset. For severely articulated objects additional domain knowledge is currently needed (e.g. a human body kinematic skeleton)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Overall, our method is the first unsupervised approach that can infer 3D keypoints from category-specific image collections without the need of any additional prior information. Conceptually, our main novelty lies in simultaneously learning 3D keypoints along with a 3D shape space model from an inhomogeneous collection of category-specific images (e.g. opposed to homogeneous collections of multiview images or videos, which show the same object from different views or in different poses). We experimentally demonstrate that our method not only improves upon 2D keypoint accuracy, but also infers plausible 3D structure for various object categories. We expect our work to be useful for diverse downstream tasks, for example related to 3D reconstruction or 3D shape-based generative models." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10711" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "text", + "content": "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "text", + "content": "[3] Adrian Bulat, Tzimiropoulos, and Georgios. Two-stage convolutional part heatmap regression for the 1st 3d face alignment in the wild (3dfaw) challenge. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 616-624. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "text", + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 356 + ], + "type": "text", + "content": "[5] Boyuan Chen, Pieter Abbeel, and Deepak Pathak. Unsupervised learning of visual 3d keypoints for control. In International Conference on Machine Learning, pages 1539-1549. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "text", + "content": "[7] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "text", + "content": "[8] Hui Deng, Tong Zhang, Yuchao Dai, Jiawei Shi, Yiran Zhong, and Hongdong Li. Deep non-rigid structure-from-motion: A sequence-to-sequence translation perspective. arXiv preprint arXiv:2204.04730, 2022. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 545 + ], + "type": "text", + "content": "[9] Aysegul Dundar, Kevin J Shih, Animesh Garg, Robert Pottorf, Anrew Tao, and Bryan Catanzaro. Unsupervised disentanglement of pose, appearance and background from images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3883-3894, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[10] Chao Gou, Yue Wu, Fei-Yue Wang, and Qiang Ji. Shape augmented regression for 3d face alignment. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 604-615. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[11] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. arXiv preprint arXiv:2103.15812, 2021. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. Advances in Neural Information Processing Systems, 35:36123-36141, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 106 + ], + "type": "text", + "content": "generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1225-1235, 2022. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "type": "text", + "content": "[14] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21337-21348, 2023. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[15] Sina Honari and Pascal Fua. Unsupervised 3d keypoint estimation with multi-view geometry. arXiv preprint arXiv:2211.12829, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[16] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8662-8672, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[17] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 295, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 348 + ], + "type": "text", + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "text", + "content": "[21] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 196-214. Springer, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[22] Yunji Kim, Seonghyeon Nam, In Cho, and Seon Joo Kim. Unsupervised keypoint learning for guiding class-conditional video prediction. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[23] Chen Kong and Simon Lucey. Deep non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1558-1567, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "type": "text", + "content": "[24] Tejas D Kulkarni, Ankush Gupta, Catalin Ionescu, Sebastian Borgeaud, Malcolm Reynolds, Andrew Zisserman, and Volodymyr Mnih. Unsupervised learning of object keypoints for perception and control. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[25] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6152-6162, 2020. 1, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10712" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[27] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Unsupervised part segmentation through disentangling appearance and shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8355-8364, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 215 + ], + "type": "text", + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 270 + ], + "type": "text", + "content": "[29] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2019. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "type": "text", + "content": "[30] Y Ma, S Soatto, J Kosecka, and S Sastry. An invitation to 3d vision: From images to models. Springer verlag, 2003. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 338 + ], + "type": "text", + "content": "[31] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "text", + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "text", + "content": "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "text", + "content": "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[36] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "text", + "content": "[37] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[38] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[39] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression: A flexible alignment-based framework for non-rigid structure estimation. IEEE Transactions on Image Processing, 27(1):249-264, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "text", + "content": "[40] Sungheon Park, Minsik Lee, and Nojun Kwak. Procrustean regression networks: Learning 3d structure of non-rigid objects from 2d annotations. In European Conference on Computer Vision, pages 1-18. Springer, 2020. 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 547, + 220 + ], + "type": "text", + "content": "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 222, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 277 + ], + "type": "text", + "content": "[42] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 279, + 547, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 547, + 334 + ], + "type": "text", + "content": "[43] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 335, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 392 + ], + "type": "text", + "content": "[44] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Occlusion-net: 2d/3d occluded keypoint localization using graph networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7326-7335, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 459 + ], + "type": "text", + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 461, + 545, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 461, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 307, + 461, + 545, + 516 + ], + "type": "text", + "content": "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 545, + 574 + ], + "type": "text", + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "text", + "content": "[48] Suwajanakorn Supasorn, Noah Snavely, Jonathan J Tompson, and Mohammad Norouzi. Discovery of latent 3d keypoints via end-to-end geometric reasoning. Advances in neural information processing systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 622, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 667 + ], + "type": "text", + "content": "[49] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[50] Sergey Tulyakov, László A Jeni, Jeffrey F Cohn, and Nicu Sebe. Viewpoint-consistent 3d face alignment. IEEE transactions on pattern analysis and machine intelligence, 40(9): 2250-2264, 2017. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10713" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 286, + 149 + ], + "type": "text", + "content": "[52] Chaoyang Wang, Chen-Hsuan Lin, and Simon Lucey. Deep nrsfm++: Towards unsupervised 2d-3d lifting in the wild. In 2020 International Conference on 3D Vision (3DV), pages 12–22. IEEE, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 286, + 194 + ], + "type": "text", + "content": "[53] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 286, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 286, + 249 + ], + "type": "text", + "content": "[54] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 286, + 293 + ], + "type": "text", + "content": "[55] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, pages 1-12, 2023. 1, 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 294, + 286, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 294, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 294, + 286, + 348 + ], + "type": "text", + "content": "[56] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "type": "text", + "content": "[57] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 404, + 286, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 286, + 448 + ], + "type": "text", + "content": "[58] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 286, + 502 + ], + "type": "text", + "content": "[59] Chenxin Xu, Siheng Chen, Maosen Li, and Ya Zhang. Invariant teacher and equivariant student for unsupervised 3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3013-3021, 2021. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 503, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 503, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 503, + 286, + 536 + ], + "type": "text", + "content": "[60] Yinghao Xu, Ceyuan Yang, Ziwei Liu, Bo Dai, and Bolei Zhou. Unsupervised landmark learning from unpaired data. arXiv preprint arXiv:2007.01053, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 537, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 286, + 591 + ], + "type": "text", + "content": "[61] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 592, + 286, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 286, + 658 + ], + "type": "text", + "content": "[62] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "type": "text", + "content": "[63] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1323-1330. IEEE, 2021. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 486 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[64] Haitian Zeng, Yuchao Dai, Xin Yu, Xiaohan Wang, and Yi Yang. Pr-rrn: pairwise-regularized residual-recursive networks for non-rigid structure-from-motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5600–5609, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[65] Haitian Zeng, Xin Yu, Jiaxu Miao, and Yi Yang. Mhr-net: Multiple-hypothesis reconstruction of non-rigid shapes from 2d views. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[66] Hongwen Zhang, Qi Li, and Zhenan Sun. Joint voxel and coordinate regression for accurate 3d facial landmark localization. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2202-2208. IEEE, 2018. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "text", + "content": "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "text", + "content": "[68] Ruiqi Zhao, Yan Wang, C Fabian Benitez-Quiroz, Yaojie Liu, and Aleix M Martinez. Fast and precise face alignment and 3d shape reconstruction from a single 2d image. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 590-603. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "type": "text", + "content": "[69] Shihao Zhou, Mengxi Jiang, Qicong Wang, and Yunqi Lei. Towards locality similarity preserving to 3d human pose estimation. In Proceedings of the Asian Conference on Computer Vision, 2020. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 386, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 386, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 308, + 386, + 545, + 439 + ], + "type": "text", + "content": "[70] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 441, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 486 + ], + "type": "text", + "content": "[71] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3, 5, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10714" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_content_list.json b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..425bbd1d894353795da75384a5776fe7964f3f77 --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_content_list.json @@ -0,0 +1,1530 @@ +[ + { + "type": "text", + "text": "Unsupervised Blind Image Deblurring Based on Self-Enhancement", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lufei Chen Xiangpeng Tian Shuhua Xiong Yinjie Lei Chao Ren* College of Electronics and Information Engineering, Sichuan University, China {chenlufei, tianxp}@stu.scu.edu.cn, {xiongh, yinjie, chaoren}@scu.edu.cn", + "bbox": [ + 151, + 180, + 823, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significant progress in image deblurring has been achieved by deep learning methods, especially the remarkable performance of supervised models on paired synthetic data. However, real-world quality degradation is more complex than synthetic datasets, and acquiring paired data in real-world scenarios poses significant challenges. To address these challenges, we propose a novel unsupervised image deblurring framework based on self-enhancement. The framework progressively generates improved pseudosharp and blurry image pairs without the need for real paired datasets, and the generated image pairs with higher qualities can be used to enhance the performance of the reconstructor. To ensure the generated blurry images are closer to the real blurry images, we propose a novel re-degradation principal component consistency loss, which enforces the principal components of the generated low-quality images to be similar to those of re-degraded images from the original sharp ones. Furthermore, we introduce the self-enhancement strategy that significantly improves deblurring performance without increasing the computational complexity of network during inference. Through extensive experiments on multiple real-world blurry datasets, we demonstrate the superiority of our approach over other state-of-the-art unsupervised methods.", + "bbox": [ + 75, + 301, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image deblurring is a classical problem in the field of computer vision, which aims to recover a clear image from its blurred version. The image deblurring tasks can be divided into blind and non-blind deblurring, where the blind image deblurring with unknown degradation is more challenging in general. Conventional model-based methods for blind image deblurring typically involve two main steps: estimating the blur kernel, and then reconstructing the sharp image from the blurred input [10, 26, 40, 42]. These methods' performance is largely constrained by the accuracy of blur kernel", + "bbox": [ + 75, + 722, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d984f4da6052005ac60be58ef7d25742297c3f10fd1fb34706ccec80470f4240.jpg", + "image_caption": [ + "Figure 1. Performance comparison of our proposed SEMGUD with other unsupervised methods [12, 23, 36, 39, 48, 50] on different datasets." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 893, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "estimation. For instance, in [27], the dark channel prior is used to estimate the blur kernel and reconstruct the sharp image. However, the blurry characteristics in real-world scenarios are quite complex, making it challenging to accurately estimate the optimal blur kernel. In addition, these methods often require complex iterative optimisation processes, which may lead to long inference time.", + "bbox": [ + 496, + 515, + 895, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) have been widely used in deblurring tasks, achieving significant success. The supervised methods [4-6, 15, 16, 19, 29, 34, 44-46] focus on training deep neural network models using a large number of paired sharp and blurry images. This enables the network to learn the mapping from blurry images to sharp images without the need for blur kernel estimation, achieving end-to-end reconstruction of the blurry and sharp images. For example, DeepDeblur [24] proposes a multi-scale CNN to implement a coarse-to-fine processing pipeline and directly restores sharp images. However, in the real world cases, for the supervised learning methods, collecting paired datasets from the real world is challenging, and manually synthesized datasets are difficult to simulate the complex real image degradation processes.", + "bbox": [ + 496, + 625, + 895, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to supervised deep learning methods, unsupervised deep learning methods [3, 25, 49, 50] for real world", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25691", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images typically achieve end-to-end image reconstruction without requiring real world paired sharp and blurry images during training. This allows unsupervised methods to more effectively handle complex real-world scenarios, especially when data collection cost is high, or the blur degradation is complex and challenging to model. Unfortunately, due to the high difficulty and challenges of unsupervised deblurring methods in training, there is still very little relevant research [12, 23, 36, 48] in this field. Specifically, because unsupervised methods lack the strong constraints provided by paired datasets, researchers tend to design more extensive deep models and complex inference processes to establish connections between inputs and outputs. This often leads to longer model inference times. Additionally, deploying large models in practical applications presents significant challenges. Therefore, how to synthesize high-quality pseudopaired datasets and how to improve the performance of unsupervised methods without increasing the complexity of the model have become key research issues.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we employ Generative Adversarial Network (GAN) [7] to learn the real-world image degradation distribution, addressing the issue of the lack of real-world paired data. We also introduce a novel re-degradation principal component consistency loss to more accurately synthesize blurred images. Considering that progressively update pseudo paired data can lead to higher performance [21], we propose a self-enhancement deblurring strategy within our unsupervised framework to further enhance the network's performance. This strategy significantly improves network deblurring performance without altering the existing network structure and without increasing inference computational complexity. The main contributions of this paper are as follows:", + "bbox": [ + 75, + 381, + 472, + 592 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel self-enhancement based unsupervised deblurring framework. This framework can progressively improve the generated pseudo-paired data and reconstruct without the need for real paired datasets, addressing the issues of insufficient paired data in real deblurring tasks, as well as the substantial complexity increase in the conventional geometric augmentation inference.", + "- We design a novel loss function called re-degradation principal component consistency ( $RPC^2$ ) loss. By introducing blur kernel prediction in principal component constraint, the $RPC^2$ loss enforces the principal components of the synthesized low-quality images to be similar to those of re-degraded images from the original sharp ones. It can decrease the effect of noise interfere in blurring images, and also make the synthetic image has more similar blurring degradation to the real data.", + "- Extensive experiments are conducted on typical real blurring datasets and the results verify the superior performance and generalization of our method over other existing unsupervised deblurring methods." + ], + "bbox": [ + 76, + 595, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Supervised Image Deblurring", + "text_level": 1, + "bbox": [ + 500, + 114, + 807, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, deep learning methods have achieved significant success in the field of image restoration [20-22, 24, 25, 28, 30, 34, 41]. For image deblurring, DL-MRF [35] proposes a CNN-based model to estimate the blur kernel and eliminate motion blur. In [2], a CNN is employed to compute an estimate of a clear image blurred by an unknown motion kernel. Thanks to the availability of some paired datasets, learning-based methods have gradually shifted their focus towards learning the mapping from given blurry images to the original sharp images, without explicitly estimating the blur kernel [11, 15, 16, 19, 29, 31, 44, 45]. MPRNet [45] introduces a novel multi-level progressive architecture to generate context-rich and spatially accurate output. DBGAN [47] designs an effective GAN-based model for simulating the synthesis of real-world blurry images. Stripformer [38] develops a transformer-based architecture by constructing horizontal and vertical labels to reweight image features. NAFNet [5] introduces a simple baseline network for image deblurring and denoising tasks. MRLPFNet [8] proposes a simple and effective multi-scale residual low-pass filter network that can better model both low and high frequency information. UFPNet [9] represents the motion blur kernel space in the latent space using normalized flows and designs a CNN to predict latent codes instead of motion blur kernels. Overall, supervised real image deblurring methods are dependent on real paired datasets, yet the scarcity of such data in real-world scenarios frequently limits their applicability.", + "bbox": [ + 496, + 138, + 895, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Deep Unsupervised Image Deblurring", + "text_level": 1, + "bbox": [ + 500, + 559, + 826, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to supervised methods, unsupervised real image deblurring methods [25, 39, 49, 50] do not involve real paired data during the training process. Consequently, unsupervised methods have weaker constraints between input and output, making it challenging to accurately learn the blur-to-clear mapping. Building on the basic GAN, UID-GAN [23] entangles the content and blur features of blurry images on a domain-specific dataset and then removes the blur from the blurry images. FCLGAN [48] proposes a lightweight and real-time unsupervised single-image blind deblurring baseline. USDF [12] introduces a multi-step self-supervised deblurring framework that iteratively decomposes and reassembles input images, exploiting the uncertainty of blur artifacts to generate a variety of pseudo-blurred and sharp image pairs. UADU [36] proposes an unsupervised semi-blind deblurring model that can effectively recover uniformly blurry image. However, unsupervised real deblurring methods have only received limited research attention in recent years. Existing methods overlook the potential in performance enhancement, and traditional model augmentation tends to increase inference computational complexity. Ad", + "bbox": [ + 496, + 583, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25692", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc8e8398487a43dcbb94198192aaa68c665d61ef37a6e72afcf760d96d261471.jpg", + "image_caption": [ + "Figure 2. Multi-Generator Unsupervised Deblurring (MGUD) framework. The whole framework employs four generators and discriminators and uses NAFNet [5] as reconstructor. The red arrows represent the backbone of MGUD, and the blue arrows, purple arrows, and green arrows respectively represent the different generator complementary constraint modules GECM-1, GECM-2, and GECM-3. DGIG Module denotes the degradation guidance information generation module." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ditionally, methods based on synthetic pseudo-real paired images don't address how to continuously improve the quality of synthetic images.", + "bbox": [ + 75, + 489, + 470, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 550, + 212, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unsupervised Deblurring Framework", + "text_level": 1, + "bbox": [ + 76, + 575, + 403, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide a detailed explanation of our proposed unsupervised framework in this subsection. Our approach aims to address the issue of insufficient paired data in real-world deblurring applications, and significantly improve network performance without altering the existing network structure and without increasing inference computational complexity.", + "bbox": [ + 75, + 599, + 470, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Overall Framework of Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 709, + 415, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In response to the lack of paired datasets in the real world, we propose the Multi-Generator Unsupervised Deblurring (MGUD) framework, as illustrated in Fig. 2. The framework employs ResNet-based generator with 6 residual blocks, PatchGAN [50] discriminator, and NAFNet [5] reconstructor. The detailed structures are presented in the supplementary materials.", + "bbox": [ + 75, + 733, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In generator, we effectively utilize unpaired sharp and blurry images to synthesize pseudo-paired datasets. Inspired by CycleGAN [50], we design a GAN structure that transitions from \"original sharp $\\rightarrow$ synthetic blur $\\rightarrow$ reconstructed", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sharp $\\rightarrow$ secondary synthetic blur\", as shown by the red and blue arrows in Fig. 2. The process $(x, \\mathcal{F}(y)) \\to x_{b1}$ transforms original sharp image $x$ to synthetic blurry image $x_{b1}$ according to degradation of blurry image $y$ , where “ $\\mathcal{F}(\\cdot)$ ” is the degradation guide information generation (DGIG) module, providing ample pseudo-paired data for deblurring. The process $x_{b1} \\to x_{rec}$ achieves deblurring by transforming $x_{b1}$ to reconstructed sharp image $x_{rec}$ . We incorporate a generator complementary constraint module (GECM) following the CycleGAN concept, $(x_{rec}, \\mathcal{F}(x_{b1})) \\to x_{b2}$ , facilitating the transition from reconstructed sharp image $x_{rec}$ to secondary synthetic blur image $x_{b2}$ , adding additional constraints and enhancing training stability.", + "bbox": [ + 496, + 489, + 893, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that the unsupervised deblurring framework's performance mainly relies on the blur image generator's capability, and also considering the challenges in GAN training, we further introduce two GECMs: $(y_{rec},\\mathcal{F}(y))\\to y_{b1}$ and $(y_{rec},\\mathcal{F}(x_{b1}))\\to y_{b2}$ , which substantially enhance the qualities of the synthetic images.", + "bbox": [ + 496, + 686, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Generation of Pseudo Paired Datasets and Reconstructor for Deblurring", + "text_level": 1, + "bbox": [ + 500, + 800, + 893, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to supervised methods, the training process of unsupervised methods struggle to accurately learn the mapping from blurry to sharp due to the lack of strong constraints from paired data. GAN is often preferred for unsupervised", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/82cdd8d7d5a86658278f0de09ea08e239e153d6bafa8870cf3eb004fceb285df.jpg", + "image_caption": [ + "Figure 3. The implementation principles of the proposed self-enhancement strategy and re-degradation principal component consistency loss function. $k$ denotes the blur kernel estimated from the real blurry images. $\\tilde{Rec}$ denotes using the fixed reconstructor trained in the last round at the input of DGIG to synthesize better pseudo-paired data." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "methods due to its superior data synthesis capability. Current mainstream unsupervised methods focus on developing more capable generators to simulate the image blurring process and generate synthetic images that closely resemble real blur. Theoretically, the closer the synthetic blur images are to actual blur, the better the reconstructor is expected to perform. The structure of our generator is depicted in Fig. 2, incorporating a sharp feature extraction module and 6 residual blocks.", + "bbox": [ + 75, + 405, + 470, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The synthesis of high-quality realistic blurred images is challenging due to varying image content influences. Considering many potential factors that could lead to undesired blurring artifacts, we initially process the blurred images $y$ using a DGIG module to obtain blurred features. The DGIG module employs a U-Net architecture, comprising a down-sampling layer followed by an upsampling layer. The $G$ learns the blurring characteristics of blurry images to guide the synthesis of sharp images towards realistic blurring. To ensure the generation of more realistic blurred images, we also train a discriminator $D$ to distinguish between synthesized and real-world blurred images, where the generator and discriminator learn collaboratively in an adversarial manner. In the MGUD framework, the adversarial loss $\\mathcal{L}_{adv1}$ of the backbone is constrained between the $y\\epsilon Y$ and $x_{b1}\\epsilon S$ , where $Y$ and $S$ denote the real blurred image and synthetic blurred image respectively. Considering that the pseudopaired images generated by the generator are crucial to the reconstructor's performance in this unsupervised framework, and accurate degradation is difficult to obtain, additional three adversarial losses are further introduced to enhance constraints as shown in Fig. 2:", + "bbox": [ + 75, + 542, + 472, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G A N} = \\mathcal {L} _ {a d v 1} + \\mathcal {L} _ {a d v 2} + \\mathcal {L} _ {a d v 3} + \\mathcal {L} _ {a d v 4} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 885, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the synthesized blurry image $x_{b1}$ , the deblurred image $x_{rec}$ is obtained through the reconstructor Rec. In our entire framework, we optimize the following loss function to train the reconstructor:", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R e c} = \\mathcal {L} _ {P S N R} (x, x _ {r e c}) + \\mathcal {L} _ {S S I M} (x, x _ {r e c}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 166, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{PSNR}(\\cdot)$ represents the PSNR loss used to constrain the peak signal-to-noise ratio of the image, and $\\mathcal{L}_{SSIM}(\\cdot)$ represents the SSIM loss used to constrain the structural information of the image.", + "bbox": [ + 498, + 196, + 893, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Re-Degradation Principal Component Consistency Loss", + "text_level": 1, + "bbox": [ + 498, + 276, + 893, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although adversarial losses can effectively improve the performance of the generator and discriminator, leading to more accurately synthesized realistic blurred images, the data synthesis process is highly susceptible to interference from the inherent information in blurred images, such as content and color. To mitigate this, we draw inspiration from [13] and design a novel loss function called re-degradation principal component consistency $(RPC^2)$ loss. By introducing blur kernel prediction in principal component constraint, this $RPC^2$ loss ensures the principal components of the synthesized low-quality images align with those of re-degraded images from the original sharp ones, which can decrease the impact of noise interfere in the blurring process. As shown in Fig. 3, we use the kernel estimation network to estimate the blur kernel of the blurry image $y$ , and then utilize this kernel to further re-blur the sharp image $x$ . Subsequently, we maintain the principal component consistency between $x$ and the synthesized blurred image $x_{b1}$ by introducing the L1 norm loss. The process is defined as follows:", + "bbox": [ + 496, + 315, + 893, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R P C ^ {2}} = \\sum_ {\\phi = 3, 5, 7} \\omega_ {\\phi} \\left\\| \\mathcal {G} _ {\\phi} (x \\otimes K (y)) - \\mathcal {G} _ {\\phi} \\left(x _ {b 1}\\right) \\right\\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 613, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K(\\cdot)$ denotes the blur kernel estimation process, $\\otimes$ denotes the blur operation, $\\mathcal{G}_{\\phi}(\\cdot)$ denotes a Gaussian filter operator with a kernel size of $\\phi$ , and $\\omega_{\\phi}$ represents the weight for level $\\phi$ .", + "bbox": [ + 496, + 657, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the $RPC^2$ loss's role is to guide $G$ in isolating the main degradation component (e.g., motion blur) from $y$ , derived from $K$ , with Gaussian filter acting to prevent $G$ from incorporating residual texture interference (like content or color from $y$ ) onto $x$ . Our test results in Section 4.3 show that without using the kernel estimation, the performance will significant decrease, verifying the necessity of the kernel estimation. In addition, we test three well-performed kernel estimation methods [1, 9, 20] and find that they achieve similar final performance. This is due to the introduction of multi-scale Gaussian low-pass filtering, the impact of noise and kernel estimation biases on the image is mitigated to", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25694", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e05c545e939cc3dd152d3ac67dae2211a1d7acfad508c69a9c0e07e59eb8a0bd.jpg", + "image_caption": [ + "Figure 4. Comparison of our proposed self-enhancement strategy with the conventional geometric augmentation strategy in terms of training expenses, testing expenses, and performance improvement respectively." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 331, + 260 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c05e55f9d459c976b6735499e4499f297c66ec2fbe9be28bef62e70cd0594e5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 620, + 260 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/273981147ef80c1f7e58524d328bb18efcc57eb937ad9bf2206e0aa98b913cbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 883, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "some extent, ensuring good consistency of the main content in the image.", + "bbox": [ + 75, + 329, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Inference Complexity Invariant Self-Enhancement Strategy", + "text_level": 1, + "bbox": [ + 76, + 369, + 468, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to enable the reconstructor to self-correct and enhance its performance without altering its structure or increasing the complexity of the original network, we propose the Self-Enhancement (SE) strategy. Fig. 3 illustrates the implementation scheme of one of the DGIGs integrating the SE strategy, and the other DGIG with the SE strategy is similar. By employing the SE strategy, a better reconstructor $(Rec)$ can be obtained, thereby learning more accurate degradation. This results in the generation of more realistic synthetic blurred-sharp image pairs and progressively improves the performance of the updated $Rec$ with higher quality synthetic samples. It is exciting to note that the performance of the $Rec$ using the SE strategy shows significant improvement over the one without the SE strategy. The implementation of the SE strategy involves several steps. First, we need to train an initial reconstructor $Rec_{1}$ . Then, as shown in Fig. 3, keep the $Rec_{1}$ parameters fixed to guide the generator to learn degradation information more accurately, generating more realistic pseudo-blurry images. Subsequently, retrain $G$ , $D$ , and $Rec$ until $Rec$ reaches convergence. Finally, repeat the above process, progressively replacing and enhancing the reconstructor until obtain the best-performing $Rec_{n}$ . Therefore, the complete $Rec$ ’s loss function is as follows:", + "bbox": [ + 75, + 409, + 473, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S E - R e c} = \\mathcal {L} _ {P S N R} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) (4) \\\\ \\begin{array}{l} + \\mathcal {L} _ {P S N R} (\\bar {R e c} (y), y _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (y), y _ {r e c}) + \\mathcal {L} _ {R e c} \\end{array} (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 768, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we obtain the total loss function:", + "bbox": [ + 96, + 853, + 367, + 867 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\min _ {G} \\max _ {D} \\mathcal {L} _ {G A N} + \\Phi_ {R P C ^ {2}} \\mathcal {L} _ {R P C ^ {2}} + \\mathcal {L} _ {S E - R e c} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 880, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6af792334ef6fc286beed1d0fafc7906c45e11028aabe3d45fc7f6173ac5be21.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StrategiesOriginalGeometric AugmentationSelf-Enhancement
PSNR27.6827.7929.06
PSNR Gains—0.111.38
", + "bbox": [ + 506, + 325, + 888, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparing the performance enhancement of the self-enhancement strategy and the geometric augmentation strategy on the GoPro dataset.", + "bbox": [ + 498, + 392, + 893, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{RPC^2}$ represents the hyperparameter for the re-degradation principal component consistency loss.", + "bbox": [ + 496, + 460, + 893, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The basic idea of the SE strategy is to use the results of the previous phase as feedback information to guide and improve the training of subsequent stages. A comparative analysis with existing typical geometric augmentation methods, as presented in Fig. 4, reveals that the SE strategy markedly boosts performance with slightly additional training cost, without imposing extra complexity or testing overhead. As Table 1 indicates, our method improves the PSNR by more than 1dB after incorporating the SE strategy, which demonstrates the effectiveness of SE strategy in image deblurring.", + "bbox": [ + 496, + 491, + 892, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 656, + 633, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 681, + 818, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following the state-of-the-art image deblurring methods [5, 9], we first divide the GoPro dataset [24] (consists of 2103 pairs of blurred and sharp images) into separate sharp and blurred image parts and further constitute unpaired blurred datasets to train the algorithm proposed in this paper. We evaluate our method on the GoPro dataset [24], RealBlur dataset [32], and HIDE dataset [33]. The GoPro dataset consists of 1111 test images, the RealBlur dataset includes 980 images for testing, and the HIDE dataset provides 2025 test images.", + "bbox": [ + 496, + 703, + 893, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We follow the experimental settings described in [5]. We adopt the Adam optimizer [14] $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ , the initial learning rate is set to $10^{-4}$", + "bbox": [ + 498, + 854, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c219b6db781c78e8aeedc64125b6403ddaaf8b1f7b71bde4c09fed522769f19e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Conference/JournalMethodsGoProHIDERealBlur-RRealBlur-J
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Deep supervisedCVPR 2018DeepDeblur [24]29.230.91625.730.87432.510.84127.870.827
CVPR 2018SRN [37]30.260.93428.360.91535.660.94728.560.867
CVPR 2021HiNet [4]32.710.95930.320.93235.750.94928.170.849
ECCV 2022Stripfromer [38]33.080.96231.030.94036.070.95228.820.876
ECCV 2022MSDI-Net [17]33.280.96431.020.94035.880.95228.590.869
ECCV 2022NAFNet [5]33.690.96731.320.94335.500.95328.320.857
ICCV 2023icDPM [31]33.200.96330.960.938N/AN/A28.810.872
CVPR 2023UFPNet [9]34.060.96831.740.94736.250.95329.870.884
Deep unsupervisedICCV 2017CycleGAN [50]22.540.72021.810.69012.380.24219.790.633
ICCV 2017DualGAN [43]22.860.722N/AN/AN/AN/AN/AN/A
CVPR 2019UIDGAN [23]23.560.73822.700.71516.640.32322.870.671
ACM MM 2022FCLGAN [48]24.840.77123.430.73228.370.66325.350.736
ICCV 2021USR-DA [39]25.490.78723.910.75632.320.82126.390.784
ACM MM 2023USDF [12]25.580.85723.930.82932.570.92326.590.881
CVPR 2023UAUD [36]26.120.86924.370.83732.910.88526.840.792
—SEMGUD (Ours)29.060.92727.640.89235.510.94628.010.844
", + "bbox": [ + 96, + 88, + 875, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The comparison results on the benchmark datasets. All the models are trained on the GoPro dataset.", + "bbox": [ + 161, + 358, + 803, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b14cb85280179695def7ff6151ad0b8bd294f22c940e2c097bdb882330872f07.jpg", + "image_caption": [ + "Figure 5. Visual comparisons on the GoPro dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 91, + 385, + 354, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5784cae6c3f153aec221851357223d943c871104dc1cdfed6151e7b8d88ed60d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 385, + 488, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5ceaf78144a3d24b97fbcba9d572d5a98ae955c060d7845dfbb57b8627d311a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 386, + 617, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3dd219cd63d3fbf8bc849bfa14a850e8c4d9dc903cea8941c2780275ac3bc959.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 386, + 746, + 603 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/682078aaebcd8c3f22395c62ed0cbbc0f7bc77ad6a5e6fd0f6976f3b24183a72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 386, + 875, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the training patch size is $128 \\times 128$ . For the $RPC^2$ loss hyperparameter $(\\Phi_{RPC^2})$ in the loss function of Eq. 5 is set to 3. More details are presented in the supplementary materials.", + "bbox": [ + 75, + 670, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 277, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the most recent unsupervised methods from recent years, including [12, 23, 36, 39, 43, 48, 50]. Additionally, we also list the state-of-the-art supervised methods based on paired images, including [4, 5, 9, 17, 24, 31, 37, 38]. Note that the related models for comparison are limited, since only a few unsupervised deblurring models have been proposed in the field. We evaluate the effectiveness of each method using performance metrics (PSNR and SSIM). The results are directly cited from the", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "original papers or generated using the official models provided by the authors.", + "bbox": [ + 496, + 670, + 893, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. Table 2 presents the PSNR and SSIM results for various single-image deblurring test methods on the GoPro, HIDE, and RealBlur test datasets. It is evident that our proposed method outperforms other unsupervised methods. On the GoPro dataset, our method achieves 2.94dB improvement in PSNR over the state-of-the-art unsupervised methods. To validate the effectiveness and generalization of our method, we compare the results on various real-world blurry datasets. Fig. 1 intuitively demonstrates the performance improvement of our method compared to other unsupervised methods. It is well known that supervised methods usually outperform unsupervised methods. But on the RealBlur-R dataset, the performance of", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25696", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7480356da02edf20a5981789cc9a917c7f6dfe752ce858b58c6795501d697f79.jpg", + "image_caption": [ + "Figure 6. Visual comparisons on the Hide dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 356, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d7422404ba8cac0d6d7de6736fe5403e3516895a6b449a0a21e0d0bac5e9302.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 490, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17fa5c39a9600ecb6a6ca3752d85371b802c82bde777ada67ab668801b63195b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 617, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb353c0a29355a9f9a2bed4460b76f519cd2df7cf074476e3b4ddc4947840762.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 88, + 746, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb064c455cfc5f6b89a65773a60c0d7ef5d06c0290bad75bbf8708ac067f4a08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 88, + 875, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/531aa58098985cbd7eb82b7e9218ce4f0042c5123e9f6e5ade4c9589e9ec2a82.jpg", + "image_caption": [ + "Figure 7. Visual comparisons on the RealBlur dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 89, + 358, + 356, + 577 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12c0ec784942512f63992c1d0b8daaec6309b9286c2268afca203354a10d57ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 358, + 490, + 577 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e262d8497aae3aaafced51757d371fb1e281c0b168607abf453ce5c3f7c09ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 358, + 619, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a0882089ccea1c85ca2154d70f4553305fd8b6c067532d5bae8215600ee3e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 358, + 748, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47b5cdc356f2f7a16ed1c9cb9ddb14a836e102420c742885f1922f6e386352e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 358, + 875, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2c9cadc1222e39aa16f06cdea2eaa986ea98c9660895dba068e53ac0e7698ee.jpg", + "image_caption": [ + "Figure 8. The \"mode collapse\" in NAFNet [5] and UFPNet [9]: trained on the GoPro dataset may output anomalous pixel regions during testing on the RealBlur-J dataset. \"(UFPNet)+\" and \"(NAFNet)+\" denote the results obtained through training with our SEMGUD framework." + ], + "image_footnote": [], + "bbox": [ + 83, + 640, + 176, + 734 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b165c71756d845482603c54b4b3b1db3955c6c1e1738b5d4c286956c4e3788b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 641, + 272, + 733 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/132962a3ed0a86da4023204dfc6ce6481c32b7b74f81792b5800225463da6a39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 641, + 370, + 734 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/346e8d511014b60dbbc77c97d21fea48e589692002274368ad3c4e4303a185bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 641, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method is even higher than NAFNet [5]. This suggests that our method's effectiveness is improving and it's even becoming competitive with certain supervised methods. It's", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcabcb265585a97d96fab9d48375f859019a449d2e2774529ca4143075f13998.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KEφ3,9,15ωφGoPro
3,5,73,7,91,0.2,0.041,0.1,0.01PSNR↑SSIM↑
✓✓✗✗✗✓28.540.919
✓✗✗✓✓✗28.720.921
✓✗✗✓✗✓28.830.923
✓✗✓✗✓✗28.890.922
✓✗✓✗✗✓28.970.924
✓✓✗✗✓✗29.000.925
✓✓✗✗✗✓29.060.927
", + "bbox": [ + 509, + 640, + 885, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the hyperparameters $\\phi$ and ${\\omega }_{\\phi }$ of the ${RPC}^{2}$ loss. The first column indicates whether the kernel estimation network is used.", + "bbox": [ + 498, + 768, + 893, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "worth noting that all the mentioned models are trained on the GoPro dataset, which further demonstrates the excellent generalization performance of our method from GoPro to other real-world datasets.", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25697", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/062ffff7ac238f6012ae3a41484d13304544e8be0c099cff4178b8e09fdc1950.jpg", + "image_caption": [ + "Blurry", + "Figure 9. Crop the blurred region from the ReloBlur dataset for visual comparison." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 179, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5e02569953d6005ca469730a6a3288f46fc6d21fc5076669fa377e219c47b28.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 274, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f9758134b43c40db393afc420516d7c57bd7665fa17240131977323a6ee2cba.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 372, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad08660141929a68e0e18480ffe81b77fd8e34c6bda8004204334d407f1712e6.jpg", + "image_caption": [ + "Sharp" + ], + "image_footnote": [], + "bbox": [ + 375, + 88, + 470, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual comparison. As shown in Figs. 5, 6, and 7, we compare the visual deblurring results of different unsupervised methods on the GoPro, HIDE, and RealBlur datasets, respectively. Note that although UAUD [36] effectively restores images with uniform blur, such performance may not extend to images with non-uniform blur. It can be seen that our proposed method achieves good results in removing motion blur from real blurry images. Additionally, we note that existing unsupervised methods often perform poorly in dealing with more severe blurring as shown in Fig. 6. Furthermore, we observe that while some existing advanced supervised methods like NAFNet and UFPNet achieve the best performance when trained on the GoPro paired datasets, they exhibit \"mode collapse\" when directly applied to other datasets, as illustrated in Fig. 8. In contrast, unsupervised models trained with our framework do not exhibit this phenomenon. Additionally, we directly use our GoPro-trained model to test the ReloBlur [18] dataset, which contains Real World Partly-blurred images, and find it to be effective for some of the images, as shown in Fig. 9.", + "bbox": [ + 75, + 239, + 472, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09370441a3732239eac3107e1dad50b1af5d0a5cb110f74f8f126542ddd4b21e.jpg", + "image_caption": [ + "Figure 10. Performance improvement (PSNR gain) of different methods on various datasets driven by the self-enhancement strategy." + ], + "image_footnote": [], + "bbox": [ + 84, + 553, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 230, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Re-Degradation Principal Component Consistency Loss. In fact, we find that if we remove the $RPC^2$ loss, the generator always struggles to generate satisfactory blurry images. Due to the minimal impact of different kernel estimation networks on the final deblurring performance as illustrated in Section 3.1.3, we can use any well-performed", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2a1f232f7cafda12fda9e33dfe79c43e7290c65418878eff63aa1994cb3a6046.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSEGoProHIDE
PSNR↑SSIM↑PSNR↑SSIM↑
FCLGAN [48]X24.700.76823.490.733
✓25.110.77223.790.741
UFPNet [9]X27.370.90325.910.862
✓28.410.91626.670.875
NAFNet [5] (Ours)X27.680.91126.240.859
✓29.060.92727.640.892
", + "bbox": [ + 511, + 88, + 883, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the superiority and generalization of the self-enhancement strategy for different reconstructors.", + "bbox": [ + 498, + 229, + 890, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "kernel estimation networks, e.g., [1] can be used for testing. Therefore, our ablation study on the $RPC^2$ loss primarily focuses on the blur kernel estimation network and the settings of hyperparameters $\\phi$ and $\\omega$ in Eq. 3. As shown in Table 3, we observe that the blur kernel estimation network significantly improves the performance of our method.", + "bbox": [ + 496, + 282, + 893, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Self-Enhancement Strategy. As shown in Table 4, the effect is improved by $1.38\\mathrm{dB}$ on the GoPro dataset with the SE strategy. Furthermore, to demonstrate the superiority and versatility of our proposed self-enhancement strategy, we also conduct experiments incorporating SE strategy on several deblurring methods, including NAFNet [5], FCLGAN [48], UFPNet [9]. Fig. 10 shows in detail the different network performance improvements before and after using the self-enhancement strategy. Note that due to FCLGAN do not provide the complete network weights, the experimental results in the table are obtained from our retraining. This strongly validates the superiority and significant potential of the self-enhancement strategy.", + "bbox": [ + 496, + 375, + 893, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel unsupervised framework for image deblurring. Our approach incorporates a re-degradation principal component consistency loss, ensuring that the principal components of the synthetically blurred images align closely with those from the re-degraded versions of the original sharp images. We also put forward a self-enhancement strategy that substantially improves performance without modifying the model's architecture or adding to the computational cost in inference. Comprehensive tests on standard datasets reveal that our method surpasses the existing state-of-the-art unsupervised methods with strong generalization capabilities across various real-world blurry image datasets.", + "bbox": [ + 496, + 609, + 893, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and the Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09.", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25698", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guillermo Carbajal, Patricia Vitoria, Mauricio Delbracio, Pablo Muse, and José Lezama. Non-uniform motion blur kernel estimation via adaptive decomposition. arXiv e-prints, pages arXiv-2102, 2021. 4, 8", + "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV), pages 221-235, 2016. 2", + "[3] Huaijin Chen, Jinwei Gu, Orazio Gallo, Ming-Yu Liu, Ashok Veeraraghavan, and Jan Kautz. Reblur2deblur: Deblurring videos via self-supervised learning. In IEEE International Conference on Computational Photography (ICCP), pages 1-9, 2018. 1", + "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 1, 6", + "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 2, 3, 5, 6, 7, 8", + "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1", + "[7] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, pages 53-65, 2018. 2", + "[8] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345-12354, 2023. 2", + "[9] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18105-18114, 2023. 2, 4, 5, 6, 7, 8", + "[10] Amit Goldstein and Raanan Fattal. Blur-kernel estimation from spectral irregularities. In European Conference on Computer Vision (ECCV), pages 622-635, 2012. 1", + "[11] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 2", + "[12] Runhua Jiang and Yahong Han. Uncertainty-aware variate decomposition for self-supervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 252-260, 2023. 1, 2, 6", + "[13] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In IEEE International Conference on Image Processing (ICIP), pages 2761-2765, 2019. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 1, 2", + "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8878-8887, 2019. 1, 2", + "[17] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In European Conference on Computer Vision (ECCV), pages 736-753, 2022. 6", + "[18] Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. Real-world deep local motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1314–1322, 2023. 8", + "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 1833-1844, 2021. 1, 2", + "[20] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4096-4105, 2021. 2, 4", + "[21] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12642-12652, 2023. 2", + "[22] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient Information Modulation Network for Image Super-Resolution. 2023. 2", + "[23] Boyu Lu, Jun-Cheng Chen, and Rama Chellappa. Unsupervised domain-specific deblurring via disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10225-10234, 2019. 1, 2, 6, 7", + "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3883-3891, 2017. 1, 2, 5, 6", + "[25] Thekke Madam Nimisha, Kumar Sunil, and AN Rajagopalan. Unsupervised class-specific deblurring. In European Conference on Computer Vision (ECCV), pages 353-369, 2018. 1, 2", + "[26] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25699", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 2901-2908, 2014. 1", + "[27] Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1628-1636, 2016. 1", + "[28] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2", + "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision (ECCV), pages 327-343, 2020. 1, 2", + "[30] Chao Ren, Xiaohai He, and Truong Q Nguyen. Adjusted non-local regression and directional smoothness for image restoration. IEEE Transactions on Multimedia, pages 731-745, 2018. 2", + "[31] Mengwei Ren, Maurizio Delbracio, Hossein Talebi, Guido Gering, and Peyman Milanfar. Multiscale structure guided diffusion for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10721-10733, 2023. 2, 6", + "[32] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In European Conference on Computer Vision (ECCV), pages 184-201, 2020. 5", + "[33] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5572-5581, 2019. 5", + "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3615, 2020. 1, 2", + "[35] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2", + "[36] Xiaole Tang, Xile Zhao, Jun Liu, Jianli Wang, Yuchun Miao, and Tieyong Zeng. Uncertainty-aware unsupervised image deblurring with deep residual prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9883-9892, 2023. 1, 2, 6, 7, 8", + "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 6", + "[38] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision (ECCV), pages 146-162, 2022. 2, 6" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Wei Wang, Haochen Zhang, Zehuan Yuan, and Changhu Wang. Unsupervised real-world super-resolution: A domain adaptation perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4318-4327, 2021. 1, 2, 6, 7", + "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, pages 168-186, 2012. 1", + "[41] Lin X, Yue J, and Ren C. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. In arXiv, 2023. 2", + "[42] Li Xu, Shicheng Zheng, and Jiaya Jia. Unnatural 10 sparse representation for natural image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1107-1114, 2013. 1", + "[43] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2849-2857, 2017. 6", + "[44] Yuan Yuan, Wei Su, and Dandan Ma. Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3555-3564, 2020. 1, 2", + "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 2", + "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1", + "[47] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2", + "[48] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, and Meng Wang. Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6220-6229, 2022. 1, 2, 6, 7, 8", + "[49] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Hajjun Zhang, Meng Wang, and Shuicheng Yan. Crnet: Unsupervised color retention network for blind motion deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6193-6201, 2022. 1, 2", + "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2223-2232, 2017. 1, 2, 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25700", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_model.json b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aba5fdac46df34ac5ff264180df433def4e9c48d --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Unsupervised Blind Image Deblurring Based on Self-Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.181, + 0.824, + 0.235 + ], + "angle": 0, + "content": "Lufei Chen Xiangpeng Tian Shuhua Xiong Yinjie Lei Chao Ren* College of Electronics and Information Engineering, Sichuan University, China {chenlufei, tianxp}@stu.scu.edu.cn, {xiongh, yinjie, chaoren}@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Significant progress in image deblurring has been achieved by deep learning methods, especially the remarkable performance of supervised models on paired synthetic data. However, real-world quality degradation is more complex than synthetic datasets, and acquiring paired data in real-world scenarios poses significant challenges. To address these challenges, we propose a novel unsupervised image deblurring framework based on self-enhancement. The framework progressively generates improved pseudosharp and blurry image pairs without the need for real paired datasets, and the generated image pairs with higher qualities can be used to enhance the performance of the reconstructor. To ensure the generated blurry images are closer to the real blurry images, we propose a novel re-degradation principal component consistency loss, which enforces the principal components of the generated low-quality images to be similar to those of re-degraded images from the original sharp ones. Furthermore, we introduce the self-enhancement strategy that significantly improves deblurring performance without increasing the computational complexity of network during inference. Through extensive experiments on multiple real-world blurry datasets, we demonstrate the superiority of our approach over other state-of-the-art unsupervised methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Image deblurring is a classical problem in the field of computer vision, which aims to recover a clear image from its blurred version. The image deblurring tasks can be divided into blind and non-blind deblurring, where the blind image deblurring with unknown degradation is more challenging in general. Conventional model-based methods for blind image deblurring typically involve two main steps: estimating the blur kernel, and then reconstructing the sharp image from the blurred input [10, 26, 40, 42]. These methods' performance is largely constrained by the accuracy of blur kernel" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.895, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.44, + 0.895, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Performance comparison of our proposed SEMGUD with other unsupervised methods [12, 23, 36, 39, 48, 50] on different datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.896, + 0.623 + ], + "angle": 0, + "content": "estimation. For instance, in [27], the dark channel prior is used to estimate the blur kernel and reconstruct the sharp image. However, the blurry characteristics in real-world scenarios are quite complex, making it challenging to accurately estimate the optimal blur kernel. In addition, these methods often require complex iterative optimisation processes, which may lead to long inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.897, + 0.868 + ], + "angle": 0, + "content": "In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) have been widely used in deblurring tasks, achieving significant success. The supervised methods [4-6, 15, 16, 19, 29, 34, 44-46] focus on training deep neural network models using a large number of paired sharp and blurry images. This enables the network to learn the mapping from blurry images to sharp images without the need for blur kernel estimation, achieving end-to-end reconstruction of the blurry and sharp images. For example, DeepDeblur [24] proposes a multi-scale CNN to implement a coarse-to-fine processing pipeline and directly restores sharp images. However, in the real world cases, for the supervised learning methods, collecting paired datasets from the real world is challenging, and manually synthesized datasets are difficult to simulate the complex real image degradation processes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Compared to supervised deep learning methods, unsupervised deep learning methods [3, 25, 49, 50] for real world" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "25691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "images typically achieve end-to-end image reconstruction without requiring real world paired sharp and blurry images during training. This allows unsupervised methods to more effectively handle complex real-world scenarios, especially when data collection cost is high, or the blur degradation is complex and challenging to model. Unfortunately, due to the high difficulty and challenges of unsupervised deblurring methods in training, there is still very little relevant research [12, 23, 36, 48] in this field. Specifically, because unsupervised methods lack the strong constraints provided by paired datasets, researchers tend to design more extensive deep models and complex inference processes to establish connections between inputs and outputs. This often leads to longer model inference times. Additionally, deploying large models in practical applications presents significant challenges. Therefore, how to synthesize high-quality pseudopaired datasets and how to improve the performance of unsupervised methods without increasing the complexity of the model have become key research issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.473, + 0.593 + ], + "angle": 0, + "content": "In this paper, we employ Generative Adversarial Network (GAN) [7] to learn the real-world image degradation distribution, addressing the issue of the lack of real-world paired data. We also introduce a novel re-degradation principal component consistency loss to more accurately synthesize blurred images. Considering that progressively update pseudo paired data can lead to higher performance [21], we propose a self-enhancement deblurring strategy within our unsupervised framework to further enhance the network's performance. This strategy significantly improves network deblurring performance without altering the existing network structure and without increasing inference computational complexity. The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.597, + 0.47, + 0.702 + ], + "angle": 0, + "content": "- We propose a novel self-enhancement based unsupervised deblurring framework. This framework can progressively improve the generated pseudo-paired data and reconstruct without the need for real paired datasets, addressing the issues of insufficient paired data in real deblurring tasks, as well as the substantial complexity increase in the conventional geometric augmentation inference." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.702, + 0.471, + 0.838 + ], + "angle": 0, + "content": "- We design a novel loss function called re-degradation principal component consistency (\\(RPC^2\\)) loss. By introducing blur kernel prediction in principal component constraint, the \\(RPC^2\\) loss enforces the principal components of the synthesized low-quality images to be similar to those of re-degraded images from the original sharp ones. It can decrease the effect of noise interfere in blurring images, and also make the synthetic image has more similar blurring degradation to the real data." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.472, + 0.899 + ], + "angle": 0, + "content": "- Extensive experiments are conducted on typical real blurring datasets and the results verify the superior performance and generalization of our method over other existing unsupervised deblurring methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.597, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.808, + 0.133 + ], + "angle": 0, + "content": "2.1. Deep Supervised Image Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.896, + 0.55 + ], + "angle": 0, + "content": "Recently, deep learning methods have achieved significant success in the field of image restoration [20-22, 24, 25, 28, 30, 34, 41]. For image deblurring, DL-MRF [35] proposes a CNN-based model to estimate the blur kernel and eliminate motion blur. In [2], a CNN is employed to compute an estimate of a clear image blurred by an unknown motion kernel. Thanks to the availability of some paired datasets, learning-based methods have gradually shifted their focus towards learning the mapping from given blurry images to the original sharp images, without explicitly estimating the blur kernel [11, 15, 16, 19, 29, 31, 44, 45]. MPRNet [45] introduces a novel multi-level progressive architecture to generate context-rich and spatially accurate output. DBGAN [47] designs an effective GAN-based model for simulating the synthesis of real-world blurry images. Stripformer [38] develops a transformer-based architecture by constructing horizontal and vertical labels to reweight image features. NAFNet [5] introduces a simple baseline network for image deblurring and denoising tasks. MRLPFNet [8] proposes a simple and effective multi-scale residual low-pass filter network that can better model both low and high frequency information. UFPNet [9] represents the motion blur kernel space in the latent space using normalized flows and designs a CNN to predict latent codes instead of motion blur kernels. Overall, supervised real image deblurring methods are dependent on real paired datasets, yet the scarcity of such data in real-world scenarios frequently limits their applicability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.828, + 0.577 + ], + "angle": 0, + "content": "2.2. Deep Unsupervised Image Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Compared to supervised methods, unsupervised real image deblurring methods [25, 39, 49, 50] do not involve real paired data during the training process. Consequently, unsupervised methods have weaker constraints between input and output, making it challenging to accurately learn the blur-to-clear mapping. Building on the basic GAN, UID-GAN [23] entangles the content and blur features of blurry images on a domain-specific dataset and then removes the blur from the blurry images. FCLGAN [48] proposes a lightweight and real-time unsupervised single-image blind deblurring baseline. USDF [12] introduces a multi-step self-supervised deblurring framework that iteratively decomposes and reassembles input images, exploiting the uncertainty of blur artifacts to generate a variety of pseudo-blurred and sharp image pairs. UADU [36] proposes an unsupervised semi-blind deblurring model that can effectively recover uniformly blurry image. However, unsupervised real deblurring methods have only received limited research attention in recent years. Existing methods overlook the potential in performance enhancement, and traditional model augmentation tends to increase inference computational complexity. Ad" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.896, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Multi-Generator Unsupervised Deblurring (MGUD) framework. The whole framework employs four generators and discriminators and uses NAFNet [5] as reconstructor. The red arrows represent the backbone of MGUD, and the blue arrows, purple arrows, and green arrows respectively represent the different generator complementary constraint modules GECM-1, GECM-2, and GECM-3. DGIG Module denotes the degradation guidance information generation module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.536 + ], + "angle": 0, + "content": "ditionally, methods based on synthetic pseudo-real paired images don't address how to continuously improve the quality of synthetic images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.214, + 0.568 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.405, + 0.592 + ], + "angle": 0, + "content": "3.1. Unsupervised Deblurring Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.471, + 0.691 + ], + "angle": 0, + "content": "We provide a detailed explanation of our proposed unsupervised framework in this subsection. Our approach aims to address the issue of insufficient paired data in real-world deblurring applications, and significantly improve network performance without altering the existing network structure and without increasing inference computational complexity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.416, + 0.726 + ], + "angle": 0, + "content": "3.1.1 Overall Framework of Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.839 + ], + "angle": 0, + "content": "In response to the lack of paired datasets in the real world, we propose the Multi-Generator Unsupervised Deblurring (MGUD) framework, as illustrated in Fig. 2. The framework employs ResNet-based generator with 6 residual blocks, PatchGAN [50] discriminator, and NAFNet [5] reconstructor. The detailed structures are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In generator, we effectively utilize unpaired sharp and blurry images to synthesize pseudo-paired datasets. Inspired by CycleGAN [50], we design a GAN structure that transitions from \"original sharp \\(\\rightarrow\\) synthetic blur \\(\\rightarrow\\) reconstructed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.687 + ], + "angle": 0, + "content": "sharp \\(\\rightarrow\\) secondary synthetic blur\", as shown by the red and blue arrows in Fig. 2. The process \\((x, \\mathcal{F}(y)) \\to x_{b1}\\) transforms original sharp image \\(x\\) to synthetic blurry image \\(x_{b1}\\) according to degradation of blurry image \\(y\\), where “\\(\\mathcal{F}(\\cdot)\\)” is the degradation guide information generation (DGIG) module, providing ample pseudo-paired data for deblurring. The process \\(x_{b1} \\to x_{rec}\\) achieves deblurring by transforming \\(x_{b1}\\) to reconstructed sharp image \\(x_{rec}\\). We incorporate a generator complementary constraint module (GECM) following the CycleGAN concept, \\((x_{rec}, \\mathcal{F}(x_{b1})) \\to x_{b2}\\), facilitating the transition from reconstructed sharp image \\(x_{rec}\\) to secondary synthetic blur image \\(x_{b2}\\), adding additional constraints and enhancing training stability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Considering that the unsupervised deblurring framework's performance mainly relies on the blur image generator's capability, and also considering the challenges in GAN training, we further introduce two GECMs: \\((y_{rec},\\mathcal{F}(y))\\to y_{b1}\\) and \\((y_{rec},\\mathcal{F}(x_{b1}))\\to y_{b2}\\), which substantially enhance the qualities of the synthetic images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.895, + 0.832 + ], + "angle": 0, + "content": "3.1.2 Generation of Pseudo Paired Datasets and Reconstructor for Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Compared to supervised methods, the training process of unsupervised methods struggle to accurately learn the mapping from blurry to sharp due to the lack of strong constraints from paired data. GAN is often preferred for unsupervised" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25693" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.473, + 0.382 + ], + "angle": 0, + "content": "Figure 3. The implementation principles of the proposed self-enhancement strategy and re-degradation principal component consistency loss function. \\( k \\) denotes the blur kernel estimated from the real blurry images. \\( \\tilde{Rec} \\) denotes using the fixed reconstructor trained in the last round at the input of DGIG to synthesize better pseudo-paired data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.472, + 0.541 + ], + "angle": 0, + "content": "methods due to its superior data synthesis capability. Current mainstream unsupervised methods focus on developing more capable generators to simulate the image blurring process and generate synthetic images that closely resemble real blur. Theoretically, the closer the synthetic blur images are to actual blur, the better the reconstructor is expected to perform. The structure of our generator is depicted in Fig. 2, incorporating a sharp feature extraction module and 6 residual blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.473, + 0.875 + ], + "angle": 0, + "content": "The synthesis of high-quality realistic blurred images is challenging due to varying image content influences. Considering many potential factors that could lead to undesired blurring artifacts, we initially process the blurred images \\(y\\) using a DGIG module to obtain blurred features. The DGIG module employs a U-Net architecture, comprising a down-sampling layer followed by an upsampling layer. The \\(G\\) learns the blurring characteristics of blurry images to guide the synthesis of sharp images towards realistic blurring. To ensure the generation of more realistic blurred images, we also train a discriminator \\(D\\) to distinguish between synthesized and real-world blurred images, where the generator and discriminator learn collaboratively in an adversarial manner. In the MGUD framework, the adversarial loss \\(\\mathcal{L}_{adv1}\\) of the backbone is constrained between the \\(y\\epsilon Y\\) and \\(x_{b1}\\epsilon S\\), where \\(Y\\) and \\(S\\) denote the real blurred image and synthetic blurred image respectively. Considering that the pseudopaired images generated by the generator are crucial to the reconstructor's performance in this unsupervised framework, and accurate degradation is difficult to obtain, additional three adversarial losses are further introduced to enhance constraints as shown in Fig. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G A N} = \\mathcal {L} _ {a d v 1} + \\mathcal {L} _ {a d v 2} + \\mathcal {L} _ {a d v 3} + \\mathcal {L} _ {a d v 4} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "For the synthesized blurry image \\( x_{b1} \\), the deblurred image \\( x_{rec} \\) is obtained through the reconstructor Rec. In our entire framework, we optimize the following loss function to train the reconstructor:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.167, + 0.892, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R e c} = \\mathcal {L} _ {P S N R} (x, x _ {r e c}) + \\mathcal {L} _ {S S I M} (x, x _ {r e c}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.197, + 0.894, + 0.258 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{PSNR}(\\cdot)\\) represents the PSNR loss used to constrain the peak signal-to-noise ratio of the image, and \\(\\mathcal{L}_{SSIM}(\\cdot)\\) represents the SSIM loss used to constrain the structural information of the image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.277, + 0.895, + 0.307 + ], + "angle": 0, + "content": "3.1.3 Re-Degradation Principal Component Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.895, + 0.603 + ], + "angle": 0, + "content": "Although adversarial losses can effectively improve the performance of the generator and discriminator, leading to more accurately synthesized realistic blurred images, the data synthesis process is highly susceptible to interference from the inherent information in blurred images, such as content and color. To mitigate this, we draw inspiration from [13] and design a novel loss function called re-degradation principal component consistency \\((RPC^2)\\) loss. By introducing blur kernel prediction in principal component constraint, this \\(RPC^2\\) loss ensures the principal components of the synthesized low-quality images align with those of re-degraded images from the original sharp ones, which can decrease the impact of noise interfere in the blurring process. As shown in Fig. 3, we use the kernel estimation network to estimate the blur kernel of the blurry image \\(y\\), and then utilize this kernel to further re-blur the sharp image \\(x\\). Subsequently, we maintain the principal component consistency between \\(x\\) and the synthesized blurred image \\(x_{b1}\\) by introducing the L1 norm loss. The process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.614, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R P C ^ {2}} = \\sum_ {\\phi = 3, 5, 7} \\omega_ {\\phi} \\left\\| \\mathcal {G} _ {\\phi} (x \\otimes K (y)) - \\mathcal {G} _ {\\phi} \\left(x _ {b 1}\\right) \\right\\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.718 + ], + "angle": 0, + "content": "where \\(K(\\cdot)\\) denotes the blur kernel estimation process, \\(\\otimes\\) denotes the blur operation, \\(\\mathcal{G}_{\\phi}(\\cdot)\\) denotes a Gaussian filter operator with a kernel size of \\(\\phi\\), and \\(\\omega_{\\phi}\\) represents the weight for level \\(\\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Specifically, the \\(RPC^2\\) loss's role is to guide \\(G\\) in isolating the main degradation component (e.g., motion blur) from \\(y\\), derived from \\(K\\), with Gaussian filter acting to prevent \\(G\\) from incorporating residual texture interference (like content or color from \\(y\\)) onto \\(x\\). Our test results in Section 4.3 show that without using the kernel estimation, the performance will significant decrease, verifying the necessity of the kernel estimation. In addition, we test three well-performed kernel estimation methods [1, 9, 20] and find that they achieve similar final performance. This is due to the introduction of multi-scale Gaussian low-pass filtering, the impact of noise and kernel estimation biases on the image is mitigated to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.332, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.089, + 0.621, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.885, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Comparison of our proposed self-enhancement strategy with the conventional geometric augmentation strategy in terms of training expenses, testing expenses, and performance improvement respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.47, + 0.36 + ], + "angle": 0, + "content": "some extent, ensuring good consistency of the main content in the image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.47, + 0.402 + ], + "angle": 0, + "content": "3.2. Inference Complexity Invariant Self-Enhancement Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.474, + 0.756 + ], + "angle": 0, + "content": "In order to enable the reconstructor to self-correct and enhance its performance without altering its structure or increasing the complexity of the original network, we propose the Self-Enhancement (SE) strategy. Fig. 3 illustrates the implementation scheme of one of the DGIGs integrating the SE strategy, and the other DGIG with the SE strategy is similar. By employing the SE strategy, a better reconstructor \\((Rec)\\) can be obtained, thereby learning more accurate degradation. This results in the generation of more realistic synthetic blurred-sharp image pairs and progressively improves the performance of the updated \\(Rec\\) with higher quality synthetic samples. It is exciting to note that the performance of the \\(Rec\\) using the SE strategy shows significant improvement over the one without the SE strategy. The implementation of the SE strategy involves several steps. First, we need to train an initial reconstructor \\(Rec_{1}\\). Then, as shown in Fig. 3, keep the \\(Rec_{1}\\) parameters fixed to guide the generator to learn degradation information more accurately, generating more realistic pseudo-blurry images. Subsequently, retrain \\(G\\), \\(D\\), and \\(Rec\\) until \\(Rec\\) reaches convergence. Finally, repeat the above process, progressively replacing and enhancing the reconstructor until obtain the best-performing \\(Rec_{n}\\). Therefore, the complete \\(Rec\\)’s loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.769, + 0.47, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S E - R e c} = \\mathcal {L} _ {P S N R} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) (4) \\\\ \\begin{array}{l} + \\mathcal {L} _ {P S N R} (\\bar {R e c} (y), y _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (y), y _ {r e c}) + \\mathcal {L} _ {R e c} \\end{array} (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.854, + 0.368, + 0.868 + ], + "angle": 0, + "content": "Finally, we obtain the total loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.881, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\min _ {G} \\max _ {D} \\mathcal {L} _ {G A N} + \\Phi_ {R P C ^ {2}} \\mathcal {L} _ {R P C ^ {2}} + \\mathcal {L} _ {S E - R e c} \\tag {5}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.326, + 0.89, + 0.376 + ], + "angle": 0, + "content": "
StrategiesOriginalGeometric AugmentationSelf-Enhancement
PSNR27.6827.7929.06
PSNR Gains—0.111.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.393, + 0.894, + 0.434 + ], + "angle": 0, + "content": "Table 1. Comparing the performance enhancement of the self-enhancement strategy and the geometric augmentation strategy on the GoPro dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.894, + 0.492 + ], + "angle": 0, + "content": "where \\(\\Phi_{RPC^2}\\) represents the hyperparameter for the re-degradation principal component consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.644 + ], + "angle": 0, + "content": "The basic idea of the SE strategy is to use the results of the previous phase as feedback information to guide and improve the training of subsequent stages. A comparative analysis with existing typical geometric augmentation methods, as presented in Fig. 4, reveals that the SE strategy markedly boosts performance with slightly additional training cost, without imposing extra complexity or testing overhead. As Table 1 indicates, our method improves the PSNR by more than 1dB after incorporating the SE strategy, which demonstrates the effectiveness of SE strategy in image deblurring." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.634, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.82, + 0.699 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Datasets. Following the state-of-the-art image deblurring methods [5, 9], we first divide the GoPro dataset [24] (consists of 2103 pairs of blurred and sharp images) into separate sharp and blurred image parts and further constitute unpaired blurred datasets to train the algorithm proposed in this paper. We evaluate our method on the GoPro dataset [24], RealBlur dataset [32], and HIDE dataset [33]. The GoPro dataset consists of 1111 test images, the RealBlur dataset includes 980 images for testing, and the HIDE dataset provides 2025 test images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We follow the experimental settings described in [5]. We adopt the Adam optimizer [14] \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\), the initial learning rate is set to \\(10^{-4}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.877, + 0.336 + ], + "angle": 0, + "content": "
Conference/JournalMethodsGoProHIDERealBlur-RRealBlur-J
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Deep supervisedCVPR 2018DeepDeblur [24]29.230.91625.730.87432.510.84127.870.827
CVPR 2018SRN [37]30.260.93428.360.91535.660.94728.560.867
CVPR 2021HiNet [4]32.710.95930.320.93235.750.94928.170.849
ECCV 2022Stripfromer [38]33.080.96231.030.94036.070.95228.820.876
ECCV 2022MSDI-Net [17]33.280.96431.020.94035.880.95228.590.869
ECCV 2022NAFNet [5]33.690.96731.320.94335.500.95328.320.857
ICCV 2023icDPM [31]33.200.96330.960.938N/AN/A28.810.872
CVPR 2023UFPNet [9]34.060.96831.740.94736.250.95329.870.884
Deep unsupervisedICCV 2017CycleGAN [50]22.540.72021.810.69012.380.24219.790.633
ICCV 2017DualGAN [43]22.860.722N/AN/AN/AN/AN/AN/A
CVPR 2019UIDGAN [23]23.560.73822.700.71516.640.32322.870.671
ACM MM 2022FCLGAN [48]24.840.77123.430.73228.370.66325.350.736
ICCV 2021USR-DA [39]25.490.78723.910.75632.320.82126.390.784
ACM MM 2023USDF [12]25.580.85723.930.82932.570.92326.590.881
CVPR 2023UAUD [36]26.120.86924.370.83732.910.88526.840.792
—SEMGUD (Ours)29.060.92727.640.89235.510.94628.010.844
" + }, + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.359, + 0.804, + 0.373 + ], + "angle": 0, + "content": "Table 2. The comparison results on the benchmark datasets. All the models are trained on the GoPro dataset." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.386, + 0.355, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.386, + 0.49, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.387, + 0.619, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.387, + 0.747, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.387, + 0.877, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.616, + 0.893, + 0.645 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on the GoPro dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.732 + ], + "angle": 0, + "content": "and the training patch size is \\(128 \\times 128\\). For the \\(RPC^2\\) loss hyperparameter \\((\\Phi_{RPC^2})\\) in the loss function of Eq. 5 is set to 3. More details are presented in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.279, + 0.758 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We compare our method with the most recent unsupervised methods from recent years, including [12, 23, 36, 39, 43, 48, 50]. Additionally, we also list the state-of-the-art supervised methods based on paired images, including [4, 5, 9, 17, 24, 31, 37, 38]. Note that the related models for comparison are limited, since only a few unsupervised deblurring models have been proposed in the field. We evaluate the effectiveness of each method using performance metrics (PSNR and SSIM). The results are directly cited from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.701 + ], + "angle": 0, + "content": "original papers or generated using the official models provided by the authors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. Table 2 presents the PSNR and SSIM results for various single-image deblurring test methods on the GoPro, HIDE, and RealBlur test datasets. It is evident that our proposed method outperforms other unsupervised methods. On the GoPro dataset, our method achieves 2.94dB improvement in PSNR over the state-of-the-art unsupervised methods. To validate the effectiveness and generalization of our method, we compare the results on various real-world blurry datasets. Fig. 1 intuitively demonstrates the performance improvement of our method compared to other unsupervised methods. It is well known that supervised methods usually outperform unsupervised methods. But on the RealBlur-R dataset, the performance of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.357, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.491, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.089, + 0.619, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.089, + 0.748, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.089, + 0.877, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons on the Hide dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.359, + 0.357, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.359, + 0.491, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.359, + 0.62, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.359, + 0.749, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.359, + 0.877, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.589, + 0.897, + 0.619 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons on the RealBlur dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.641, + 0.177, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.642, + 0.274, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.642, + 0.372, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.642, + 0.47, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.747, + 0.473, + 0.817 + ], + "angle": 0, + "content": "Figure 8. The \"mode collapse\" in NAFNet [5] and UFPNet [9]: trained on the GoPro dataset may output anomalous pixel regions during testing on the RealBlur-J dataset. \"(UFPNet)+\" and \"(NAFNet)+\" denote the results obtained through training with our SEMGUD framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "our method is even higher than NAFNet [5]. This suggests that our method's effectiveness is improving and it's even becoming competitive with certain supervised methods. It's" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.641, + 0.887, + 0.753 + ], + "angle": 0, + "content": "
KEφ3,9,15ωφGoPro
3,5,73,7,91,0.2,0.041,0.1,0.01PSNR↑SSIM↑
✓✓✗✗✗✓28.540.919
✓✗✗✓✓✗28.720.921
✓✗✗✓✗✓28.830.923
✓✗✓✗✓✗28.890.922
✓✗✓✗✗✓28.970.924
✓✓✗✗✓✗29.000.925
✓✓✗✗✗✓29.060.927
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.769, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Table 3. Ablation study on the hyperparameters \\( \\phi \\) and \\( {\\omega }_{\\phi } \\) of the \\( {RPC}^{2} \\) loss. The first column indicates whether the kernel estimation network is used." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "worth noting that all the mentioned models are trained on the GoPro dataset, which further demonstrates the excellent generalization performance of our method from GoPro to other real-world datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.18, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.164, + 0.153, + 0.177 + ], + "angle": 0, + "content": "Blurry" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.276, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.165, + 0.253, + 0.176 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.374, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.165, + 0.34, + 0.175 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.089, + 0.472, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.165, + 0.441, + 0.177 + ], + "angle": 0, + "content": "Sharp" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.187, + 0.47, + 0.215 + ], + "angle": 0, + "content": "Figure 9. Crop the blurred region from the ReloBlur dataset for visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.473, + 0.543 + ], + "angle": 0, + "content": "Visual comparison. As shown in Figs. 5, 6, and 7, we compare the visual deblurring results of different unsupervised methods on the GoPro, HIDE, and RealBlur datasets, respectively. Note that although UAUD [36] effectively restores images with uniform blur, such performance may not extend to images with non-uniform blur. It can be seen that our proposed method achieves good results in removing motion blur from real blurry images. Additionally, we note that existing unsupervised methods often perform poorly in dealing with more severe blurring as shown in Fig. 6. Furthermore, we observe that while some existing advanced supervised methods like NAFNet and UFPNet achieve the best performance when trained on the GoPro paired datasets, they exhibit \"mode collapse\" when directly applied to other datasets, as illustrated in Fig. 8. In contrast, unsupervised models trained with our framework do not exhibit this phenomenon. Additionally, we directly use our GoPro-trained model to test the ReloBlur [18] dataset, which contains Real World Partly-blurred images, and find it to be effective for some of the images, as shown in Fig. 9." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.554, + 0.47, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.722, + 0.471, + 0.765 + ], + "angle": 0, + "content": "Figure 10. Performance improvement (PSNR gain) of different methods on various datasets driven by the self-enhancement strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The Re-Degradation Principal Component Consistency Loss. In fact, we find that if we remove the \\(RPC^2\\) loss, the generator always struggles to generate satisfactory blurry images. Due to the minimal impact of different kernel estimation networks on the final deblurring performance as illustrated in Section 3.1.3, we can use any well-performed" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.089, + 0.885, + 0.213 + ], + "angle": 0, + "content": "
MethodsSEGoProHIDE
PSNR↑SSIM↑PSNR↑SSIM↑
FCLGAN [48]X24.700.76823.490.733
✓25.110.77223.790.741
UFPNet [9]X27.370.90325.910.862
✓28.410.91626.670.875
NAFNet [5] (Ours)X27.680.91126.240.859
✓29.060.92727.640.892
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Table 4. Ablation study on the superiority and generalization of the self-enhancement strategy for different reconstructors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.894, + 0.375 + ], + "angle": 0, + "content": "kernel estimation networks, e.g., [1] can be used for testing. Therefore, our ablation study on the \\(RPC^2\\) loss primarily focuses on the blur kernel estimation network and the settings of hyperparameters \\(\\phi\\) and \\(\\omega\\) in Eq. 3. As shown in Table 3, we observe that the blur kernel estimation network significantly improves the performance of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.895, + 0.572 + ], + "angle": 0, + "content": "The Self-Enhancement Strategy. As shown in Table 4, the effect is improved by \\(1.38\\mathrm{dB}\\) on the GoPro dataset with the SE strategy. Furthermore, to demonstrate the superiority and versatility of our proposed self-enhancement strategy, we also conduct experiments incorporating SE strategy on several deblurring methods, including NAFNet [5], FCLGAN [48], UFPNet [9]. Fig. 10 shows in detail the different network performance improvements before and after using the self-enhancement strategy. Note that due to FCLGAN do not provide the complete network weights, the experimental results in the table are obtained from our retraining. This strongly validates the superiority and significant potential of the self-enhancement strategy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.619, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.807 + ], + "angle": 0, + "content": "In this paper, we introduce a novel unsupervised framework for image deblurring. Our approach incorporates a re-degradation principal component consistency loss, ensuring that the principal components of the synthetically blurred images align closely with those from the re-degraded versions of the original sharp images. We also put forward a self-enhancement strategy that substantially improves performance without modifying the model's architecture or adding to the computational cost in inference. Comprehensive tests on standard datasets reveal that our method surpasses the existing state-of-the-art unsupervised methods with strong generalization capabilities across various real-world blurry image datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and the Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25698" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.17 + ], + "angle": 0, + "content": "[1] Guillermo Carbajal, Patricia Vitoria, Mauricio Delbracio, Pablo Muse, and José Lezama. Non-uniform motion blur kernel estimation via adaptive decomposition. arXiv e-prints, pages arXiv-2102, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV), pages 221-235, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Huaijin Chen, Jinwei Gu, Orazio Gallo, Ming-Yu Liu, Ashok Veeraraghavan, and Jan Kautz. Reblur2deblur: Deblurring videos via self-supervised learning. In IEEE International Conference on Computational Photography (ICCP), pages 1-9, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.472, + 0.354 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.472, + 0.409 + ], + "angle": 0, + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, pages 53-65, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[8] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345-12354, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[9] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18105-18114, 2023. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[10] Amit Goldstein and Raanan Fattal. Blur-kernel estimation from spectral irregularities. In European Conference on Computer Vision (ECCV), pages 622-635, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[11] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[12] Runhua Jiang and Yahong Han. Uncertainty-aware variate decomposition for self-supervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 252-260, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In IEEE International Conference on Image Processing (ICIP), pages 2761-2765, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8878-8887, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[17] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In European Conference on Computer Vision (ECCV), pages 736-753, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[18] Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. Real-world deep local motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1314–1322, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.893, + 0.474 + ], + "angle": 0, + "content": "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 1833-1844, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[20] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4096-4105, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[21] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12642-12652, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[22] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient Information Modulation Network for Image Super-Resolution. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[23] Boyu Lu, Jun-Cheng Chen, and Rama Chellappa. Unsupervised domain-specific deblurring via disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10225-10234, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.893, + 0.8 + ], + "angle": 0, + "content": "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3883-3891, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[25] Thekke Madam Nimisha, Kumar Sunil, and AN Rajagopalan. Unsupervised class-specific deblurring. In European Conference on Computer Vision (ECCV), pages 353-369, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 2901-2908, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[27] Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1628-1636, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.246 + ], + "angle": 0, + "content": "[28] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision (ECCV), pages 327-343, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[30] Chao Ren, Xiaohai He, and Truong Q Nguyen. Adjusted non-local regression and directional smoothness for image restoration. IEEE Transactions on Multimedia, pages 731-745, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[31] Mengwei Ren, Maurizio Delbracio, Hossein Talebi, Guido Gering, and Peyman Milanfar. Multiscale structure guided diffusion for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10721-10733, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[32] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In European Conference on Computer Vision (ECCV), pages 184-201, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[33] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5572-5581, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3615, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[35] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[36] Xiaole Tang, Xile Zhao, Jun Liu, Jianli Wang, Yuchun Miao, and Tieyong Zeng. Uncertainty-aware unsupervised image deblurring with deep residual prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9883-9892, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision (ECCV), pages 146-162, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[39] Wei Wang, Haochen Zhang, Zehuan Yuan, and Changhu Wang. Unsupervised real-world super-resolution: A domain adaptation perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4318-4327, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, pages 168-186, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.26 + ], + "angle": 0, + "content": "[41] Lin X, Yue J, and Ren C. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. In arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.893, + 0.318 + ], + "angle": 0, + "content": "[42] Li Xu, Shicheng Zheng, and Jiaya Jia. Unnatural 10 sparse representation for natural image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1107-1114, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[43] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2849-2857, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[44] Yuan Yuan, Wei Su, and Dandan Ma. Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3555-3564, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.894, + 0.515 + ], + "angle": 0, + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.894, + 0.585 + ], + "angle": 0, + "content": "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.893, + 0.655 + ], + "angle": 0, + "content": "[47] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.893, + 0.727 + ], + "angle": 0, + "content": "[48] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, and Meng Wang. Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6220-6229, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.894, + 0.797 + ], + "angle": 0, + "content": "[49] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Hajjun Zhang, Meng Wang, and Shuicheng Yan. Crnet: Unsupervised color retention network for blind motion deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6193-6201, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.894, + 0.867 + ], + "angle": 0, + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2223-2232, 2017. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25700" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_origin.pdf b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88992f1cdf2ffb35bc034fa6a1bf63ebdd9b4e35 --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/ae508369-dbac-472c-9eda-522b8f70eff5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dff1ac692f69d5b906aa50fbb94da19991bac0e38fa3843066d23a555310306b +size 5469041 diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/full.md b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e7d39b3f5aae071960a3d8bfe55f82d60b23040 --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/full.md @@ -0,0 +1,290 @@ +# Unsupervised Blind Image Deblurring Based on Self-Enhancement + +Lufei Chen Xiangpeng Tian Shuhua Xiong Yinjie Lei Chao Ren* College of Electronics and Information Engineering, Sichuan University, China {chenlufei, tianxp}@stu.scu.edu.cn, {xiongh, yinjie, chaoren}@scu.edu.cn + +# Abstract + +Significant progress in image deblurring has been achieved by deep learning methods, especially the remarkable performance of supervised models on paired synthetic data. However, real-world quality degradation is more complex than synthetic datasets, and acquiring paired data in real-world scenarios poses significant challenges. To address these challenges, we propose a novel unsupervised image deblurring framework based on self-enhancement. The framework progressively generates improved pseudosharp and blurry image pairs without the need for real paired datasets, and the generated image pairs with higher qualities can be used to enhance the performance of the reconstructor. To ensure the generated blurry images are closer to the real blurry images, we propose a novel re-degradation principal component consistency loss, which enforces the principal components of the generated low-quality images to be similar to those of re-degraded images from the original sharp ones. Furthermore, we introduce the self-enhancement strategy that significantly improves deblurring performance without increasing the computational complexity of network during inference. Through extensive experiments on multiple real-world blurry datasets, we demonstrate the superiority of our approach over other state-of-the-art unsupervised methods. + +# 1. Introduction + +Image deblurring is a classical problem in the field of computer vision, which aims to recover a clear image from its blurred version. The image deblurring tasks can be divided into blind and non-blind deblurring, where the blind image deblurring with unknown degradation is more challenging in general. Conventional model-based methods for blind image deblurring typically involve two main steps: estimating the blur kernel, and then reconstructing the sharp image from the blurred input [10, 26, 40, 42]. These methods' performance is largely constrained by the accuracy of blur kernel + +![](images/d984f4da6052005ac60be58ef7d25742297c3f10fd1fb34706ccec80470f4240.jpg) +Figure 1. Performance comparison of our proposed SEMGUD with other unsupervised methods [12, 23, 36, 39, 48, 50] on different datasets. + +estimation. For instance, in [27], the dark channel prior is used to estimate the blur kernel and reconstruct the sharp image. However, the blurry characteristics in real-world scenarios are quite complex, making it challenging to accurately estimate the optimal blur kernel. In addition, these methods often require complex iterative optimisation processes, which may lead to long inference time. + +In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) have been widely used in deblurring tasks, achieving significant success. The supervised methods [4-6, 15, 16, 19, 29, 34, 44-46] focus on training deep neural network models using a large number of paired sharp and blurry images. This enables the network to learn the mapping from blurry images to sharp images without the need for blur kernel estimation, achieving end-to-end reconstruction of the blurry and sharp images. For example, DeepDeblur [24] proposes a multi-scale CNN to implement a coarse-to-fine processing pipeline and directly restores sharp images. However, in the real world cases, for the supervised learning methods, collecting paired datasets from the real world is challenging, and manually synthesized datasets are difficult to simulate the complex real image degradation processes. + +Compared to supervised deep learning methods, unsupervised deep learning methods [3, 25, 49, 50] for real world + +images typically achieve end-to-end image reconstruction without requiring real world paired sharp and blurry images during training. This allows unsupervised methods to more effectively handle complex real-world scenarios, especially when data collection cost is high, or the blur degradation is complex and challenging to model. Unfortunately, due to the high difficulty and challenges of unsupervised deblurring methods in training, there is still very little relevant research [12, 23, 36, 48] in this field. Specifically, because unsupervised methods lack the strong constraints provided by paired datasets, researchers tend to design more extensive deep models and complex inference processes to establish connections between inputs and outputs. This often leads to longer model inference times. Additionally, deploying large models in practical applications presents significant challenges. Therefore, how to synthesize high-quality pseudopaired datasets and how to improve the performance of unsupervised methods without increasing the complexity of the model have become key research issues. + +In this paper, we employ Generative Adversarial Network (GAN) [7] to learn the real-world image degradation distribution, addressing the issue of the lack of real-world paired data. We also introduce a novel re-degradation principal component consistency loss to more accurately synthesize blurred images. Considering that progressively update pseudo paired data can lead to higher performance [21], we propose a self-enhancement deblurring strategy within our unsupervised framework to further enhance the network's performance. This strategy significantly improves network deblurring performance without altering the existing network structure and without increasing inference computational complexity. The main contributions of this paper are as follows: + +- We propose a novel self-enhancement based unsupervised deblurring framework. This framework can progressively improve the generated pseudo-paired data and reconstruct without the need for real paired datasets, addressing the issues of insufficient paired data in real deblurring tasks, as well as the substantial complexity increase in the conventional geometric augmentation inference. +- We design a novel loss function called re-degradation principal component consistency ( $RPC^2$ ) loss. By introducing blur kernel prediction in principal component constraint, the $RPC^2$ loss enforces the principal components of the synthesized low-quality images to be similar to those of re-degraded images from the original sharp ones. It can decrease the effect of noise interfere in blurring images, and also make the synthetic image has more similar blurring degradation to the real data. +- Extensive experiments are conducted on typical real blurring datasets and the results verify the superior performance and generalization of our method over other existing unsupervised deblurring methods. + +# 2. Related Work + +# 2.1. Deep Supervised Image Deblurring + +Recently, deep learning methods have achieved significant success in the field of image restoration [20-22, 24, 25, 28, 30, 34, 41]. For image deblurring, DL-MRF [35] proposes a CNN-based model to estimate the blur kernel and eliminate motion blur. In [2], a CNN is employed to compute an estimate of a clear image blurred by an unknown motion kernel. Thanks to the availability of some paired datasets, learning-based methods have gradually shifted their focus towards learning the mapping from given blurry images to the original sharp images, without explicitly estimating the blur kernel [11, 15, 16, 19, 29, 31, 44, 45]. MPRNet [45] introduces a novel multi-level progressive architecture to generate context-rich and spatially accurate output. DBGAN [47] designs an effective GAN-based model for simulating the synthesis of real-world blurry images. Stripformer [38] develops a transformer-based architecture by constructing horizontal and vertical labels to reweight image features. NAFNet [5] introduces a simple baseline network for image deblurring and denoising tasks. MRLPFNet [8] proposes a simple and effective multi-scale residual low-pass filter network that can better model both low and high frequency information. UFPNet [9] represents the motion blur kernel space in the latent space using normalized flows and designs a CNN to predict latent codes instead of motion blur kernels. Overall, supervised real image deblurring methods are dependent on real paired datasets, yet the scarcity of such data in real-world scenarios frequently limits their applicability. + +# 2.2. Deep Unsupervised Image Deblurring + +Compared to supervised methods, unsupervised real image deblurring methods [25, 39, 49, 50] do not involve real paired data during the training process. Consequently, unsupervised methods have weaker constraints between input and output, making it challenging to accurately learn the blur-to-clear mapping. Building on the basic GAN, UID-GAN [23] entangles the content and blur features of blurry images on a domain-specific dataset and then removes the blur from the blurry images. FCLGAN [48] proposes a lightweight and real-time unsupervised single-image blind deblurring baseline. USDF [12] introduces a multi-step self-supervised deblurring framework that iteratively decomposes and reassembles input images, exploiting the uncertainty of blur artifacts to generate a variety of pseudo-blurred and sharp image pairs. UADU [36] proposes an unsupervised semi-blind deblurring model that can effectively recover uniformly blurry image. However, unsupervised real deblurring methods have only received limited research attention in recent years. Existing methods overlook the potential in performance enhancement, and traditional model augmentation tends to increase inference computational complexity. Ad + +![](images/cc8e8398487a43dcbb94198192aaa68c665d61ef37a6e72afcf760d96d261471.jpg) +Figure 2. Multi-Generator Unsupervised Deblurring (MGUD) framework. The whole framework employs four generators and discriminators and uses NAFNet [5] as reconstructor. The red arrows represent the backbone of MGUD, and the blue arrows, purple arrows, and green arrows respectively represent the different generator complementary constraint modules GECM-1, GECM-2, and GECM-3. DGIG Module denotes the degradation guidance information generation module. + +ditionally, methods based on synthetic pseudo-real paired images don't address how to continuously improve the quality of synthetic images. + +# 3. Methodology + +# 3.1. Unsupervised Deblurring Framework + +We provide a detailed explanation of our proposed unsupervised framework in this subsection. Our approach aims to address the issue of insufficient paired data in real-world deblurring applications, and significantly improve network performance without altering the existing network structure and without increasing inference computational complexity. + +# 3.1.1 Overall Framework of Proposed Method + +In response to the lack of paired datasets in the real world, we propose the Multi-Generator Unsupervised Deblurring (MGUD) framework, as illustrated in Fig. 2. The framework employs ResNet-based generator with 6 residual blocks, PatchGAN [50] discriminator, and NAFNet [5] reconstructor. The detailed structures are presented in the supplementary materials. + +In generator, we effectively utilize unpaired sharp and blurry images to synthesize pseudo-paired datasets. Inspired by CycleGAN [50], we design a GAN structure that transitions from "original sharp $\rightarrow$ synthetic blur $\rightarrow$ reconstructed + +sharp $\rightarrow$ secondary synthetic blur", as shown by the red and blue arrows in Fig. 2. The process $(x, \mathcal{F}(y)) \to x_{b1}$ transforms original sharp image $x$ to synthetic blurry image $x_{b1}$ according to degradation of blurry image $y$ , where “ $\mathcal{F}(\cdot)$ ” is the degradation guide information generation (DGIG) module, providing ample pseudo-paired data for deblurring. The process $x_{b1} \to x_{rec}$ achieves deblurring by transforming $x_{b1}$ to reconstructed sharp image $x_{rec}$ . We incorporate a generator complementary constraint module (GECM) following the CycleGAN concept, $(x_{rec}, \mathcal{F}(x_{b1})) \to x_{b2}$ , facilitating the transition from reconstructed sharp image $x_{rec}$ to secondary synthetic blur image $x_{b2}$ , adding additional constraints and enhancing training stability. + +Considering that the unsupervised deblurring framework's performance mainly relies on the blur image generator's capability, and also considering the challenges in GAN training, we further introduce two GECMs: $(y_{rec},\mathcal{F}(y))\to y_{b1}$ and $(y_{rec},\mathcal{F}(x_{b1}))\to y_{b2}$ , which substantially enhance the qualities of the synthetic images. + +# 3.1.2 Generation of Pseudo Paired Datasets and Reconstructor for Deblurring + +Compared to supervised methods, the training process of unsupervised methods struggle to accurately learn the mapping from blurry to sharp due to the lack of strong constraints from paired data. GAN is often preferred for unsupervised + +![](images/82cdd8d7d5a86658278f0de09ea08e239e153d6bafa8870cf3eb004fceb285df.jpg) +Figure 3. The implementation principles of the proposed self-enhancement strategy and re-degradation principal component consistency loss function. $k$ denotes the blur kernel estimated from the real blurry images. $\tilde{Rec}$ denotes using the fixed reconstructor trained in the last round at the input of DGIG to synthesize better pseudo-paired data. + +methods due to its superior data synthesis capability. Current mainstream unsupervised methods focus on developing more capable generators to simulate the image blurring process and generate synthetic images that closely resemble real blur. Theoretically, the closer the synthetic blur images are to actual blur, the better the reconstructor is expected to perform. The structure of our generator is depicted in Fig. 2, incorporating a sharp feature extraction module and 6 residual blocks. + +The synthesis of high-quality realistic blurred images is challenging due to varying image content influences. Considering many potential factors that could lead to undesired blurring artifacts, we initially process the blurred images $y$ using a DGIG module to obtain blurred features. The DGIG module employs a U-Net architecture, comprising a down-sampling layer followed by an upsampling layer. The $G$ learns the blurring characteristics of blurry images to guide the synthesis of sharp images towards realistic blurring. To ensure the generation of more realistic blurred images, we also train a discriminator $D$ to distinguish between synthesized and real-world blurred images, where the generator and discriminator learn collaboratively in an adversarial manner. In the MGUD framework, the adversarial loss $\mathcal{L}_{adv1}$ of the backbone is constrained between the $y\epsilon Y$ and $x_{b1}\epsilon S$ , where $Y$ and $S$ denote the real blurred image and synthetic blurred image respectively. Considering that the pseudopaired images generated by the generator are crucial to the reconstructor's performance in this unsupervised framework, and accurate degradation is difficult to obtain, additional three adversarial losses are further introduced to enhance constraints as shown in Fig. 2: + +$$ +\mathcal {L} _ {G A N} = \mathcal {L} _ {a d v 1} + \mathcal {L} _ {a d v 2} + \mathcal {L} _ {a d v 3} + \mathcal {L} _ {a d v 4} \tag {1} +$$ + +For the synthesized blurry image $x_{b1}$ , the deblurred image $x_{rec}$ is obtained through the reconstructor Rec. In our entire framework, we optimize the following loss function to train the reconstructor: + +$$ +\mathcal {L} _ {R e c} = \mathcal {L} _ {P S N R} (x, x _ {r e c}) + \mathcal {L} _ {S S I M} (x, x _ {r e c}) \tag {2} +$$ + +where $\mathcal{L}_{PSNR}(\cdot)$ represents the PSNR loss used to constrain the peak signal-to-noise ratio of the image, and $\mathcal{L}_{SSIM}(\cdot)$ represents the SSIM loss used to constrain the structural information of the image. + +# 3.1.3 Re-Degradation Principal Component Consistency Loss + +Although adversarial losses can effectively improve the performance of the generator and discriminator, leading to more accurately synthesized realistic blurred images, the data synthesis process is highly susceptible to interference from the inherent information in blurred images, such as content and color. To mitigate this, we draw inspiration from [13] and design a novel loss function called re-degradation principal component consistency $(RPC^2)$ loss. By introducing blur kernel prediction in principal component constraint, this $RPC^2$ loss ensures the principal components of the synthesized low-quality images align with those of re-degraded images from the original sharp ones, which can decrease the impact of noise interfere in the blurring process. As shown in Fig. 3, we use the kernel estimation network to estimate the blur kernel of the blurry image $y$ , and then utilize this kernel to further re-blur the sharp image $x$ . Subsequently, we maintain the principal component consistency between $x$ and the synthesized blurred image $x_{b1}$ by introducing the L1 norm loss. The process is defined as follows: + +$$ +\mathcal {L} _ {R P C ^ {2}} = \sum_ {\phi = 3, 5, 7} \omega_ {\phi} \left\| \mathcal {G} _ {\phi} (x \otimes K (y)) - \mathcal {G} _ {\phi} \left(x _ {b 1}\right) \right\| _ {1} \tag {3} +$$ + +where $K(\cdot)$ denotes the blur kernel estimation process, $\otimes$ denotes the blur operation, $\mathcal{G}_{\phi}(\cdot)$ denotes a Gaussian filter operator with a kernel size of $\phi$ , and $\omega_{\phi}$ represents the weight for level $\phi$ . + +Specifically, the $RPC^2$ loss's role is to guide $G$ in isolating the main degradation component (e.g., motion blur) from $y$ , derived from $K$ , with Gaussian filter acting to prevent $G$ from incorporating residual texture interference (like content or color from $y$ ) onto $x$ . Our test results in Section 4.3 show that without using the kernel estimation, the performance will significant decrease, verifying the necessity of the kernel estimation. In addition, we test three well-performed kernel estimation methods [1, 9, 20] and find that they achieve similar final performance. This is due to the introduction of multi-scale Gaussian low-pass filtering, the impact of noise and kernel estimation biases on the image is mitigated to + +![](images/e05c545e939cc3dd152d3ac67dae2211a1d7acfad508c69a9c0e07e59eb8a0bd.jpg) +Figure 4. Comparison of our proposed self-enhancement strategy with the conventional geometric augmentation strategy in terms of training expenses, testing expenses, and performance improvement respectively. + +![](images/c05e55f9d459c976b6735499e4499f297c66ec2fbe9be28bef62e70cd0594e5e.jpg) + +![](images/273981147ef80c1f7e58524d328bb18efcc57eb937ad9bf2206e0aa98b913cbc.jpg) + +some extent, ensuring good consistency of the main content in the image. + +# 3.2. Inference Complexity Invariant Self-Enhancement Strategy + +In order to enable the reconstructor to self-correct and enhance its performance without altering its structure or increasing the complexity of the original network, we propose the Self-Enhancement (SE) strategy. Fig. 3 illustrates the implementation scheme of one of the DGIGs integrating the SE strategy, and the other DGIG with the SE strategy is similar. By employing the SE strategy, a better reconstructor $(Rec)$ can be obtained, thereby learning more accurate degradation. This results in the generation of more realistic synthetic blurred-sharp image pairs and progressively improves the performance of the updated $Rec$ with higher quality synthetic samples. It is exciting to note that the performance of the $Rec$ using the SE strategy shows significant improvement over the one without the SE strategy. The implementation of the SE strategy involves several steps. First, we need to train an initial reconstructor $Rec_{1}$ . Then, as shown in Fig. 3, keep the $Rec_{1}$ parameters fixed to guide the generator to learn degradation information more accurately, generating more realistic pseudo-blurry images. Subsequently, retrain $G$ , $D$ , and $Rec$ until $Rec$ reaches convergence. Finally, repeat the above process, progressively replacing and enhancing the reconstructor until obtain the best-performing $Rec_{n}$ . Therefore, the complete $Rec$ ’s loss function is as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {S E - R e c} = \mathcal {L} _ {P S N R} (\bar {R e c} (x _ {b 1}), x _ {r e c}) \\ + \mathcal {L} _ {S S I M} (\bar {R e c} (x _ {b 1}), x _ {r e c}) (4) \\ \begin{array}{l} + \mathcal {L} _ {P S N R} (\bar {R e c} (y), y _ {r e c}) \\ + \mathcal {L} _ {S S I M} (\bar {R e c} (y), y _ {r e c}) + \mathcal {L} _ {R e c} \end{array} (4) \\ \end{array} +$$ + +Finally, we obtain the total loss function: + +$$ +\mathcal {L} = \min _ {G} \max _ {D} \mathcal {L} _ {G A N} + \Phi_ {R P C ^ {2}} \mathcal {L} _ {R P C ^ {2}} + \mathcal {L} _ {S E - R e c} \tag {5} +$$ + +
StrategiesOriginalGeometric AugmentationSelf-Enhancement
PSNR27.6827.7929.06
PSNR Gains—0.111.38
+ +Table 1. Comparing the performance enhancement of the self-enhancement strategy and the geometric augmentation strategy on the GoPro dataset. + +where $\Phi_{RPC^2}$ represents the hyperparameter for the re-degradation principal component consistency loss. + +The basic idea of the SE strategy is to use the results of the previous phase as feedback information to guide and improve the training of subsequent stages. A comparative analysis with existing typical geometric augmentation methods, as presented in Fig. 4, reveals that the SE strategy markedly boosts performance with slightly additional training cost, without imposing extra complexity or testing overhead. As Table 1 indicates, our method improves the PSNR by more than 1dB after incorporating the SE strategy, which demonstrates the effectiveness of SE strategy in image deblurring. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +Datasets. Following the state-of-the-art image deblurring methods [5, 9], we first divide the GoPro dataset [24] (consists of 2103 pairs of blurred and sharp images) into separate sharp and blurred image parts and further constitute unpaired blurred datasets to train the algorithm proposed in this paper. We evaluate our method on the GoPro dataset [24], RealBlur dataset [32], and HIDE dataset [33]. The GoPro dataset consists of 1111 test images, the RealBlur dataset includes 980 images for testing, and the HIDE dataset provides 2025 test images. + +Implementation Details. We follow the experimental settings described in [5]. We adopt the Adam optimizer [14] $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ , the initial learning rate is set to $10^{-4}$ + +
Conference/JournalMethodsGoProHIDERealBlur-RRealBlur-J
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Deep supervisedCVPR 2018DeepDeblur [24]29.230.91625.730.87432.510.84127.870.827
CVPR 2018SRN [37]30.260.93428.360.91535.660.94728.560.867
CVPR 2021HiNet [4]32.710.95930.320.93235.750.94928.170.849
ECCV 2022Stripfromer [38]33.080.96231.030.94036.070.95228.820.876
ECCV 2022MSDI-Net [17]33.280.96431.020.94035.880.95228.590.869
ECCV 2022NAFNet [5]33.690.96731.320.94335.500.95328.320.857
ICCV 2023icDPM [31]33.200.96330.960.938N/AN/A28.810.872
CVPR 2023UFPNet [9]34.060.96831.740.94736.250.95329.870.884
Deep unsupervisedICCV 2017CycleGAN [50]22.540.72021.810.69012.380.24219.790.633
ICCV 2017DualGAN [43]22.860.722N/AN/AN/AN/AN/AN/A
CVPR 2019UIDGAN [23]23.560.73822.700.71516.640.32322.870.671
ACM MM 2022FCLGAN [48]24.840.77123.430.73228.370.66325.350.736
ICCV 2021USR-DA [39]25.490.78723.910.75632.320.82126.390.784
ACM MM 2023USDF [12]25.580.85723.930.82932.570.92326.590.881
CVPR 2023UAUD [36]26.120.86924.370.83732.910.88526.840.792
—SEMGUD (Ours)29.060.92727.640.89235.510.94628.010.844
+ +Table 2. The comparison results on the benchmark datasets. All the models are trained on the GoPro dataset. + +![](images/b14cb85280179695def7ff6151ad0b8bd294f22c940e2c097bdb882330872f07.jpg) +Figure 5. Visual comparisons on the GoPro dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth. + +![](images/5784cae6c3f153aec221851357223d943c871104dc1cdfed6151e7b8d88ed60d.jpg) + +![](images/d5ceaf78144a3d24b97fbcba9d572d5a98ae955c060d7845dfbb57b8627d311a.jpg) + +![](images/3dd219cd63d3fbf8bc849bfa14a850e8c4d9dc903cea8941c2780275ac3bc959.jpg) + +![](images/682078aaebcd8c3f22395c62ed0cbbc0f7bc77ad6a5e6fd0f6976f3b24183a72.jpg) + +and the training patch size is $128 \times 128$ . For the $RPC^2$ loss hyperparameter $(\Phi_{RPC^2})$ in the loss function of Eq. 5 is set to 3. More details are presented in the supplementary materials. + +# 4.2. Experimental Results + +We compare our method with the most recent unsupervised methods from recent years, including [12, 23, 36, 39, 43, 48, 50]. Additionally, we also list the state-of-the-art supervised methods based on paired images, including [4, 5, 9, 17, 24, 31, 37, 38]. Note that the related models for comparison are limited, since only a few unsupervised deblurring models have been proposed in the field. We evaluate the effectiveness of each method using performance metrics (PSNR and SSIM). The results are directly cited from the + +original papers or generated using the official models provided by the authors. + +Quantitative comparison. Table 2 presents the PSNR and SSIM results for various single-image deblurring test methods on the GoPro, HIDE, and RealBlur test datasets. It is evident that our proposed method outperforms other unsupervised methods. On the GoPro dataset, our method achieves 2.94dB improvement in PSNR over the state-of-the-art unsupervised methods. To validate the effectiveness and generalization of our method, we compare the results on various real-world blurry datasets. Fig. 1 intuitively demonstrates the performance improvement of our method compared to other unsupervised methods. It is well known that supervised methods usually outperform unsupervised methods. But on the RealBlur-R dataset, the performance of + +![](images/7480356da02edf20a5981789cc9a917c7f6dfe752ce858b58c6795501d697f79.jpg) +Figure 6. Visual comparisons on the Hide dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth. + +![](images/1d7422404ba8cac0d6d7de6736fe5403e3516895a6b449a0a21e0d0bac5e9302.jpg) + +![](images/17fa5c39a9600ecb6a6ca3752d85371b802c82bde777ada67ab668801b63195b.jpg) + +![](images/bb353c0a29355a9f9a2bed4460b76f519cd2df7cf074476e3b4ddc4947840762.jpg) + +![](images/bb064c455cfc5f6b89a65773a60c0d7ef5d06c0290bad75bbf8708ac067f4a08.jpg) + +![](images/531aa58098985cbd7eb82b7e9218ce4f0042c5123e9f6e5ade4c9589e9ec2a82.jpg) +Figure 7. Visual comparisons on the RealBlur dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth. + +![](images/12c0ec784942512f63992c1d0b8daaec6309b9286c2268afca203354a10d57ca.jpg) + +![](images/e262d8497aae3aaafced51757d371fb1e281c0b168607abf453ce5c3f7c09ffd.jpg) + +![](images/9a0882089ccea1c85ca2154d70f4553305fd8b6c067532d5bae8215600ee3e45.jpg) + +![](images/47b5cdc356f2f7a16ed1c9cb9ddb14a836e102420c742885f1922f6e386352e4.jpg) + +![](images/d2c9cadc1222e39aa16f06cdea2eaa986ea98c9660895dba068e53ac0e7698ee.jpg) +Figure 8. The "mode collapse" in NAFNet [5] and UFPNet [9]: trained on the GoPro dataset may output anomalous pixel regions during testing on the RealBlur-J dataset. "(UFPNet)+" and "(NAFNet)+" denote the results obtained through training with our SEMGUD framework. + +![](images/b165c71756d845482603c54b4b3b1db3955c6c1e1738b5d4c286956c4e3788b7.jpg) + +![](images/132962a3ed0a86da4023204dfc6ce6481c32b7b74f81792b5800225463da6a39.jpg) + +![](images/346e8d511014b60dbbc77c97d21fea48e589692002274368ad3c4e4303a185bd.jpg) + +our method is even higher than NAFNet [5]. This suggests that our method's effectiveness is improving and it's even becoming competitive with certain supervised methods. It's + +
KEφ3,9,15ωφGoPro
3,5,73,7,91,0.2,0.041,0.1,0.01PSNR↑SSIM↑
✓✓✗✗✗✓28.540.919
✓✗✗✓✓✗28.720.921
✓✗✗✓✗✓28.830.923
✓✗✓✗✓✗28.890.922
✓✗✓✗✗✓28.970.924
✓✓✗✗✓✗29.000.925
✓✓✗✗✗✓29.060.927
+ +Table 3. Ablation study on the hyperparameters $\phi$ and ${\omega }_{\phi }$ of the ${RPC}^{2}$ loss. The first column indicates whether the kernel estimation network is used. + +worth noting that all the mentioned models are trained on the GoPro dataset, which further demonstrates the excellent generalization performance of our method from GoPro to other real-world datasets. + +![](images/062ffff7ac238f6012ae3a41484d13304544e8be0c099cff4178b8e09fdc1950.jpg) +Blurry +Figure 9. Crop the blurred region from the ReloBlur dataset for visual comparison. + +![](images/d5e02569953d6005ca469730a6a3288f46fc6d21fc5076669fa377e219c47b28.jpg) +NAFNet + +![](images/5f9758134b43c40db393afc420516d7c57bd7665fa17240131977323a6ee2cba.jpg) +Ours + +![](images/ad08660141929a68e0e18480ffe81b77fd8e34c6bda8004204334d407f1712e6.jpg) +Sharp + +Visual comparison. As shown in Figs. 5, 6, and 7, we compare the visual deblurring results of different unsupervised methods on the GoPro, HIDE, and RealBlur datasets, respectively. Note that although UAUD [36] effectively restores images with uniform blur, such performance may not extend to images with non-uniform blur. It can be seen that our proposed method achieves good results in removing motion blur from real blurry images. Additionally, we note that existing unsupervised methods often perform poorly in dealing with more severe blurring as shown in Fig. 6. Furthermore, we observe that while some existing advanced supervised methods like NAFNet and UFPNet achieve the best performance when trained on the GoPro paired datasets, they exhibit "mode collapse" when directly applied to other datasets, as illustrated in Fig. 8. In contrast, unsupervised models trained with our framework do not exhibit this phenomenon. Additionally, we directly use our GoPro-trained model to test the ReloBlur [18] dataset, which contains Real World Partly-blurred images, and find it to be effective for some of the images, as shown in Fig. 9. + +![](images/09370441a3732239eac3107e1dad50b1af5d0a5cb110f74f8f126542ddd4b21e.jpg) +Figure 10. Performance improvement (PSNR gain) of different methods on various datasets driven by the self-enhancement strategy. + +# 4.3. Ablation Study + +The Re-Degradation Principal Component Consistency Loss. In fact, we find that if we remove the $RPC^2$ loss, the generator always struggles to generate satisfactory blurry images. Due to the minimal impact of different kernel estimation networks on the final deblurring performance as illustrated in Section 3.1.3, we can use any well-performed + +
MethodsSEGoProHIDE
PSNR↑SSIM↑PSNR↑SSIM↑
FCLGAN [48]X24.700.76823.490.733
✓25.110.77223.790.741
UFPNet [9]X27.370.90325.910.862
✓28.410.91626.670.875
NAFNet [5] (Ours)X27.680.91126.240.859
✓29.060.92727.640.892
+ +Table 4. Ablation study on the superiority and generalization of the self-enhancement strategy for different reconstructors. + +kernel estimation networks, e.g., [1] can be used for testing. Therefore, our ablation study on the $RPC^2$ loss primarily focuses on the blur kernel estimation network and the settings of hyperparameters $\phi$ and $\omega$ in Eq. 3. As shown in Table 3, we observe that the blur kernel estimation network significantly improves the performance of our method. + +The Self-Enhancement Strategy. As shown in Table 4, the effect is improved by $1.38\mathrm{dB}$ on the GoPro dataset with the SE strategy. Furthermore, to demonstrate the superiority and versatility of our proposed self-enhancement strategy, we also conduct experiments incorporating SE strategy on several deblurring methods, including NAFNet [5], FCLGAN [48], UFPNet [9]. Fig. 10 shows in detail the different network performance improvements before and after using the self-enhancement strategy. Note that due to FCLGAN do not provide the complete network weights, the experimental results in the table are obtained from our retraining. This strongly validates the superiority and significant potential of the self-enhancement strategy. + +# 5. Conclusion + +In this paper, we introduce a novel unsupervised framework for image deblurring. Our approach incorporates a re-degradation principal component consistency loss, ensuring that the principal components of the synthetically blurred images align closely with those from the re-degraded versions of the original sharp images. We also put forward a self-enhancement strategy that substantially improves performance without modifying the model's architecture or adding to the computational cost in inference. Comprehensive tests on standard datasets reveal that our method surpasses the existing state-of-the-art unsupervised methods with strong generalization capabilities across various real-world blurry image datasets. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and the Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09. + +# References + +[1] Guillermo Carbajal, Patricia Vitoria, Mauricio Delbracio, Pablo Muse, and José Lezama. Non-uniform motion blur kernel estimation via adaptive decomposition. arXiv e-prints, pages arXiv-2102, 2021. 4, 8 +[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV), pages 221-235, 2016. 2 +[3] Huaijin Chen, Jinwei Gu, Orazio Gallo, Ming-Yu Liu, Ashok Veeraraghavan, and Jan Kautz. Reblur2deblur: Deblurring videos via self-supervised learning. In IEEE International Conference on Computational Photography (ICCP), pages 1-9, 2018. 1 +[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 1, 6 +[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 2, 3, 5, 6, 7, 8 +[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1 +[7] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, pages 53-65, 2018. 2 +[8] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345-12354, 2023. 2 +[9] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18105-18114, 2023. 2, 4, 5, 6, 7, 8 +[10] Amit Goldstein and Raanan Fattal. Blur-kernel estimation from spectral irregularities. In European Conference on Computer Vision (ECCV), pages 622-635, 2012. 1 +[11] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 2 +[12] Runhua Jiang and Yahong Han. Uncertainty-aware variate decomposition for self-supervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 252-260, 2023. 1, 2, 6 +[13] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In IEEE International Conference on Image Processing (ICIP), pages 2761-2765, 2019. 4 + +[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 1, 2 +[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8878-8887, 2019. 1, 2 +[17] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In European Conference on Computer Vision (ECCV), pages 736-753, 2022. 6 +[18] Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. Real-world deep local motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1314–1322, 2023. 8 +[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 1833-1844, 2021. 1, 2 +[20] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4096-4105, 2021. 2, 4 +[21] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12642-12652, 2023. 2 +[22] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient Information Modulation Network for Image Super-Resolution. 2023. 2 +[23] Boyu Lu, Jun-Cheng Chen, and Rama Chellappa. Unsupervised domain-specific deblurring via disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10225-10234, 2019. 1, 2, 6, 7 +[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3883-3891, 2017. 1, 2, 5, 6 +[25] Thekke Madam Nimisha, Kumar Sunil, and AN Rajagopalan. Unsupervised class-specific deblurring. In European Conference on Computer Vision (ECCV), pages 353-369, 2018. 1, 2 +[26] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 2901-2908, 2014. 1 +[27] Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1628-1636, 2016. 1 +[28] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2 +[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision (ECCV), pages 327-343, 2020. 1, 2 +[30] Chao Ren, Xiaohai He, and Truong Q Nguyen. Adjusted non-local regression and directional smoothness for image restoration. IEEE Transactions on Multimedia, pages 731-745, 2018. 2 +[31] Mengwei Ren, Maurizio Delbracio, Hossein Talebi, Guido Gering, and Peyman Milanfar. Multiscale structure guided diffusion for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10721-10733, 2023. 2, 6 +[32] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In European Conference on Computer Vision (ECCV), pages 184-201, 2020. 5 +[33] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5572-5581, 2019. 5 +[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3615, 2020. 1, 2 +[35] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2 +[36] Xiaole Tang, Xile Zhao, Jun Liu, Jianli Wang, Yuchun Miao, and Tieyong Zeng. Uncertainty-aware unsupervised image deblurring with deep residual prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9883-9892, 2023. 1, 2, 6, 7, 8 +[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 6 +[38] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision (ECCV), pages 146-162, 2022. 2, 6 + +[39] Wei Wang, Haochen Zhang, Zehuan Yuan, and Changhu Wang. Unsupervised real-world super-resolution: A domain adaptation perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4318-4327, 2021. 1, 2, 6, 7 +[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, pages 168-186, 2012. 1 +[41] Lin X, Yue J, and Ren C. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. In arXiv, 2023. 2 +[42] Li Xu, Shicheng Zheng, and Jiaya Jia. Unnatural 10 sparse representation for natural image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1107-1114, 2013. 1 +[43] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2849-2857, 2017. 6 +[44] Yuan Yuan, Wei Su, and Dandan Ma. Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3555-3564, 2020. 1, 2 +[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 2 +[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1 +[47] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2 +[48] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, and Meng Wang. Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6220-6229, 2022. 1, 2, 6, 7, 8 +[49] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Hajjun Zhang, Meng Wang, and Shuicheng Yan. Crnet: Unsupervised color retention network for blind motion deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6193-6201, 2022. 1, 2 +[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2223-2232, 2017. 1, 2, 3, 6, 7 \ No newline at end of file diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/images.zip b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fbfea3cbef0aa3861b083589c4578568c95ade4f --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b71f106c0dce0d3e419260358db672b69c3130c2210be97bbe1245188788142a +size 778150 diff --git a/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/layout.json b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f4742dd9fbbee23ea416617b6757c36f96b015d4 --- /dev/null +++ b/2024/Unsupervised Blind Image Deblurring Based on Self-Enhancement/layout.json @@ -0,0 +1,8218 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "text", + "content": "Unsupervised Blind Image Deblurring Based on Self-Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 143, + 504, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 143, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 93, + 143, + 504, + 186 + ], + "type": "text", + "content": "Lufei Chen Xiangpeng Tian Shuhua Xiong Yinjie Lei Chao Ren* College of Electronics and Information Engineering, Sichuan University, China {chenlufei, tianxp}@stu.scu.edu.cn, {xiongh, yinjie, chaoren}@scu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "text", + "content": "Significant progress in image deblurring has been achieved by deep learning methods, especially the remarkable performance of supervised models on paired synthetic data. However, real-world quality degradation is more complex than synthetic datasets, and acquiring paired data in real-world scenarios poses significant challenges. To address these challenges, we propose a novel unsupervised image deblurring framework based on self-enhancement. The framework progressively generates improved pseudosharp and blurry image pairs without the need for real paired datasets, and the generated image pairs with higher qualities can be used to enhance the performance of the reconstructor. To ensure the generated blurry images are closer to the real blurry images, we propose a novel re-degradation principal component consistency loss, which enforces the principal components of the generated low-quality images to be similar to those of re-degraded images from the original sharp ones. Furthermore, we introduce the self-enhancement strategy that significantly improves deblurring performance without increasing the computational complexity of network during inference. Through extensive experiments on multiple real-world blurry datasets, we demonstrate the superiority of our approach over other state-of-the-art unsupervised methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 288, + 693 + ], + "type": "text", + "content": "Image deblurring is a classical problem in the field of computer vision, which aims to recover a clear image from its blurred version. The image deblurring tasks can be divided into blind and non-blind deblurring, where the blind image deblurring with unknown degradation is more challenging in general. Conventional model-based methods for blind image deblurring typically involve two main steps: estimating the blur kernel, and then reconstructing the sharp image from the blurred input [10, 26, 40, 42]. These methods' performance is largely constrained by the accuracy of blur kernel" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 547, + 342 + ], + "type": "image", + "image_path": "d984f4da6052005ac60be58ef7d25742297c3f10fd1fb34706ccec80470f4240.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 348, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 348, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 547, + 381 + ], + "type": "text", + "content": "Figure 1. Performance comparison of our proposed SEMGUD with other unsupervised methods [12, 23, 36, 39, 48, 50] on different datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 408, + 548, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 548, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 548, + 493 + ], + "type": "text", + "content": "estimation. For instance, in [27], the dark channel prior is used to estimate the blur kernel and reconstruct the sharp image. However, the blurry characteristics in real-world scenarios are quite complex, making it challenging to accurately estimate the optimal blur kernel. In addition, these methods often require complex iterative optimisation processes, which may lead to long inference time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 548, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 548, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 548, + 687 + ], + "type": "text", + "content": "In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) have been widely used in deblurring tasks, achieving significant success. The supervised methods [4-6, 15, 16, 19, 29, 34, 44-46] focus on training deep neural network models using a large number of paired sharp and blurry images. This enables the network to learn the mapping from blurry images to sharp images without the need for blur kernel estimation, achieving end-to-end reconstruction of the blurry and sharp images. For example, DeepDeblur [24] proposes a multi-scale CNN to implement a coarse-to-fine processing pipeline and directly restores sharp images. However, in the real world cases, for the supervised learning methods, collecting paired datasets from the real world is challenging, and manually synthesized datasets are difficult to simulate the complex real image degradation processes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Compared to supervised deep learning methods, unsupervised deep learning methods [3, 25, 49, 50] for real world" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25691" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "images typically achieve end-to-end image reconstruction without requiring real world paired sharp and blurry images during training. This allows unsupervised methods to more effectively handle complex real-world scenarios, especially when data collection cost is high, or the blur degradation is complex and challenging to model. Unfortunately, due to the high difficulty and challenges of unsupervised deblurring methods in training, there is still very little relevant research [12, 23, 36, 48] in this field. Specifically, because unsupervised methods lack the strong constraints provided by paired datasets, researchers tend to design more extensive deep models and complex inference processes to establish connections between inputs and outputs. This often leads to longer model inference times. Additionally, deploying large models in practical applications presents significant challenges. Therefore, how to synthesize high-quality pseudopaired datasets and how to improve the performance of unsupervised methods without increasing the complexity of the model have become key research issues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 289, + 469 + ], + "type": "text", + "content": "In this paper, we employ Generative Adversarial Network (GAN) [7] to learn the real-world image degradation distribution, addressing the issue of the lack of real-world paired data. We also introduce a novel re-degradation principal component consistency loss to more accurately synthesize blurred images. Considering that progressively update pseudo paired data can lead to higher performance [21], we propose a self-enhancement deblurring strategy within our unsupervised framework to further enhance the network's performance. This strategy significantly improves network deblurring performance without altering the existing network structure and without increasing inference computational complexity. The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 472, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 472, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 555 + ], + "type": "text", + "content": "- We propose a novel self-enhancement based unsupervised deblurring framework. This framework can progressively improve the generated pseudo-paired data and reconstruct without the need for real paired datasets, addressing the issues of insufficient paired data in real deblurring tasks, as well as the substantial complexity increase in the conventional geometric augmentation inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "text", + "content": "- We design a novel loss function called re-degradation principal component consistency (" + }, + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "text", + "content": ") loss. By introducing blur kernel prediction in principal component constraint, the " + }, + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 47, + 555, + 288, + 663 + ], + "type": "text", + "content": " loss enforces the principal components of the synthesized low-quality images to be similar to those of re-degraded images from the original sharp ones. It can decrease the effect of noise interfere in blurring images, and also make the synthetic image has more similar blurring degradation to the real data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 663, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 288, + 712 + ], + "type": "text", + "content": "- Extensive experiments are conducted on typical real blurring datasets and the results verify the superior performance and generalization of our method over other existing unsupervised deblurring methods." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 91, + 494, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 494, + 105 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 494, + 105 + ], + "type": "text", + "content": "2.1. Deep Supervised Image Deblurring" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 110, + 548, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 548, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 548, + 435 + ], + "type": "text", + "content": "Recently, deep learning methods have achieved significant success in the field of image restoration [20-22, 24, 25, 28, 30, 34, 41]. For image deblurring, DL-MRF [35] proposes a CNN-based model to estimate the blur kernel and eliminate motion blur. In [2], a CNN is employed to compute an estimate of a clear image blurred by an unknown motion kernel. Thanks to the availability of some paired datasets, learning-based methods have gradually shifted their focus towards learning the mapping from given blurry images to the original sharp images, without explicitly estimating the blur kernel [11, 15, 16, 19, 29, 31, 44, 45]. MPRNet [45] introduces a novel multi-level progressive architecture to generate context-rich and spatially accurate output. DBGAN [47] designs an effective GAN-based model for simulating the synthesis of real-world blurry images. Stripformer [38] develops a transformer-based architecture by constructing horizontal and vertical labels to reweight image features. NAFNet [5] introduces a simple baseline network for image deblurring and denoising tasks. MRLPFNet [8] proposes a simple and effective multi-scale residual low-pass filter network that can better model both low and high frequency information. UFPNet [9] represents the motion blur kernel space in the latent space using normalized flows and designs a CNN to predict latent codes instead of motion blur kernels. Overall, supervised real image deblurring methods are dependent on real paired datasets, yet the scarcity of such data in real-world scenarios frequently limits their applicability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 443, + 506, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 506, + 456 + ], + "type": "text", + "content": "2.2. Deep Unsupervised Image Deblurring" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 548, + 715 + ], + "type": "text", + "content": "Compared to supervised methods, unsupervised real image deblurring methods [25, 39, 49, 50] do not involve real paired data during the training process. Consequently, unsupervised methods have weaker constraints between input and output, making it challenging to accurately learn the blur-to-clear mapping. Building on the basic GAN, UID-GAN [23] entangles the content and blur features of blurry images on a domain-specific dataset and then removes the blur from the blurry images. FCLGAN [48] proposes a lightweight and real-time unsupervised single-image blind deblurring baseline. USDF [12] introduces a multi-step self-supervised deblurring framework that iteratively decomposes and reassembles input images, exploiting the uncertainty of blur artifacts to generate a variety of pseudo-blurred and sharp image pairs. UADU [36] proposes an unsupervised semi-blind deblurring model that can effectively recover uniformly blurry image. However, unsupervised real deblurring methods have only received limited research attention in recent years. Existing methods overlook the potential in performance enhancement, and traditional model augmentation tends to increase inference computational complexity. Ad" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25692" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 315 + ], + "type": "image", + "image_path": "cc8e8398487a43dcbb94198192aaa68c665d61ef37a6e72afcf760d96d261471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 548, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 548, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 548, + 369 + ], + "type": "text", + "content": "Figure 2. Multi-Generator Unsupervised Deblurring (MGUD) framework. The whole framework employs four generators and discriminators and uses NAFNet [5] as reconstructor. The red arrows represent the backbone of MGUD, and the blue arrows, purple arrows, and green arrows respectively represent the different generator complementary constraint modules GECM-1, GECM-2, and GECM-3. DGIG Module denotes the degradation guidance information generation module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 388, + 288, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 424 + ], + "type": "text", + "content": "ditionally, methods based on synthetic pseudo-real paired images don't address how to continuously improve the quality of synthetic images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 436, + 130, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 130, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 130, + 449 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 247, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 247, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 247, + 468 + ], + "type": "text", + "content": "3.1. Unsupervised Deblurring Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 475, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 547 + ], + "type": "text", + "content": "We provide a detailed explanation of our proposed unsupervised framework in this subsection. Our approach aims to address the issue of insufficient paired data in real-world deblurring applications, and significantly improve network performance without altering the existing network structure and without increasing inference computational complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 254, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 254, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 254, + 574 + ], + "type": "text", + "content": "3.1.1 Overall Framework of Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 664 + ], + "type": "text", + "content": "In response to the lack of paired datasets in the real world, we propose the Multi-Generator Unsupervised Deblurring (MGUD) framework, as illustrated in Fig. 2. The framework employs ResNet-based generator with 6 residual blocks, PatchGAN [50] discriminator, and NAFNet [5] reconstructor. The detailed structures are presented in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In generator, we effectively utilize unpaired sharp and blurry images to synthesize pseudo-paired datasets. Inspired by CycleGAN [50], we design a GAN structure that transitions from \"original sharp " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " synthetic blur " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " reconstructed" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": "sharp " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " secondary synthetic blur\", as shown by the red and blue arrows in Fig. 2. The process " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "(x, \\mathcal{F}(y)) \\to x_{b1}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " transforms original sharp image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " to synthetic blurry image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{b1}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " according to degradation of blurry image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": ", where “" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": "” is the degradation guide information generation (DGIG) module, providing ample pseudo-paired data for deblurring. The process " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{b1} \\to x_{rec}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " achieves deblurring by transforming " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{b1}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " to reconstructed sharp image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{rec}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": ". We incorporate a generator complementary constraint module (GECM) following the CycleGAN concept, " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "(x_{rec}, \\mathcal{F}(x_{b1})) \\to x_{b2}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": ", facilitating the transition from reconstructed sharp image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{rec}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": " to secondary synthetic blur image " + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{b2}" + }, + { + "bbox": [ + 304, + 388, + 547, + 544 + ], + "type": "text", + "content": ", adding additional constraints and enhancing training stability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "text", + "content": "Considering that the unsupervised deblurring framework's performance mainly relies on the blur image generator's capability, and also considering the challenges in GAN training, we further introduce two GECMs: " + }, + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "inline_equation", + "content": "(y_{rec},\\mathcal{F}(y))\\to y_{b1}" + }, + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "inline_equation", + "content": "(y_{rec},\\mathcal{F}(x_{b1}))\\to y_{b2}" + }, + { + "bbox": [ + 304, + 544, + 547, + 617 + ], + "type": "text", + "content": ", which substantially enhance the qualities of the synthetic images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 547, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 547, + 658 + ], + "type": "text", + "content": "3.1.2 Generation of Pseudo Paired Datasets and Reconstructor for Deblurring" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Compared to supervised methods, the training process of unsupervised methods struggle to accurately learn the mapping from blurry to sharp due to the lack of strong constraints from paired data. GAN is often preferred for unsupervised" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25693" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 228 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 228 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 228 + ], + "type": "image", + "image_path": "82cdd8d7d5a86658278f0de09ea08e239e153d6bafa8870cf3eb004fceb285df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "type": "text", + "content": "Figure 3. The implementation principles of the proposed self-enhancement strategy and re-degradation principal component consistency loss function. " + }, + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "type": "text", + "content": " denotes the blur kernel estimated from the real blurry images. " + }, + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "type": "inline_equation", + "content": "\\tilde{Rec}" + }, + { + "bbox": [ + 46, + 235, + 289, + 302 + ], + "type": "text", + "content": " denotes using the fixed reconstructor trained in the last round at the input of DGIG to synthesize better pseudo-paired data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 288, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 428 + ], + "type": "text", + "content": "methods due to its superior data synthesis capability. Current mainstream unsupervised methods focus on developing more capable generators to simulate the image blurring process and generate synthetic images that closely resemble real blur. Theoretically, the closer the synthetic blur images are to actual blur, the better the reconstructor is expected to perform. The structure of our generator is depicted in Fig. 2, incorporating a sharp feature extraction module and 6 residual blocks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": "The synthesis of high-quality realistic blurred images is challenging due to varying image content influences. Considering many potential factors that could lead to undesired blurring artifacts, we initially process the blurred images " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " using a DGIG module to obtain blurred features. The DGIG module employs a U-Net architecture, comprising a down-sampling layer followed by an upsampling layer. The " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " learns the blurring characteristics of blurry images to guide the synthesis of sharp images towards realistic blurring. To ensure the generation of more realistic blurred images, we also train a discriminator " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " to distinguish between synthesized and real-world blurred images, where the generator and discriminator learn collaboratively in an adversarial manner. In the MGUD framework, the adversarial loss " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv1}" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " of the backbone is constrained between the " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "y\\epsilon Y" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "x_{b1}\\epsilon S" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 430, + 289, + 693 + ], + "type": "text", + "content": " denote the real blurred image and synthetic blurred image respectively. Considering that the pseudopaired images generated by the generator are crucial to the reconstructor's performance in this unsupervised framework, and accurate degradation is difficult to obtain, additional three adversarial losses are further introduced to enhance constraints as shown in Fig. 2:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 701, + 288, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 79, + 701, + 288, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G A N} = \\mathcal {L} _ {a d v 1} + \\mathcal {L} _ {a d v 2} + \\mathcal {L} _ {a d v 3} + \\mathcal {L} _ {a d v 4} \\tag {1}", + "image_path": "a893d3858eea3b57dfeb7c5072113e7da45155fa2ed251637d15e4ed246f4b87.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "For the synthesized blurry image " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "x_{b1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": ", the deblurred image " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "x_{rec}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " is obtained through the reconstructor Rec. In our entire framework, we optimize the following loss function to train the reconstructor:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 334, + 132, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 132, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 334, + 132, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R e c} = \\mathcal {L} _ {P S N R} (x, x _ {r e c}) + \\mathcal {L} _ {S S I M} (x, x _ {r e c}) \\tag {2}", + "image_path": "27af971646cbac25d5070c7d2caac5e985e3217f7310287449a8b6fbf60ed183.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PSNR}(\\cdot)" + }, + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "text", + "content": " represents the PSNR loss used to constrain the peak signal-to-noise ratio of the image, and " + }, + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSIM}(\\cdot)" + }, + { + "bbox": [ + 305, + 156, + 547, + 204 + ], + "type": "text", + "content": " represents the SSIM loss used to constrain the structural information of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 219, + 547, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 547, + 243 + ], + "type": "text", + "content": "3.1.3 Re-Degradation Principal Component Consistency Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": "Although adversarial losses can effectively improve the performance of the generator and discriminator, leading to more accurately synthesized realistic blurred images, the data synthesis process is highly susceptible to interference from the inherent information in blurred images, such as content and color. To mitigate this, we draw inspiration from [13] and design a novel loss function called re-degradation principal component consistency " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "(RPC^2)" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": " loss. By introducing blur kernel prediction in principal component constraint, this " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": " loss ensures the principal components of the synthesized low-quality images align with those of re-degraded images from the original sharp ones, which can decrease the impact of noise interfere in the blurring process. As shown in Fig. 3, we use the kernel estimation network to estimate the blur kernel of the blurry image " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": ", and then utilize this kernel to further re-blur the sharp image " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": ". Subsequently, we maintain the principal component consistency between " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": " and the synthesized blurred image " + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "inline_equation", + "content": "x_{b1}" + }, + { + "bbox": [ + 304, + 250, + 547, + 477 + ], + "type": "text", + "content": " by introducing the L1 norm loss. The process is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 486, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 486, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 317, + 486, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R P C ^ {2}} = \\sum_ {\\phi = 3, 5, 7} \\omega_ {\\phi} \\left\\| \\mathcal {G} _ {\\phi} (x \\otimes K (y)) - \\mathcal {G} _ {\\phi} \\left(x _ {b 1}\\right) \\right\\| _ {1} \\tag {3}", + "image_path": "9d64b6542836d7a0fe5e58e43559791ba7874d355b9273572413780b810e085d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "K(\\cdot)" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": " denotes the blur kernel estimation process, " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": " denotes the blur operation, " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": " denotes a Gaussian filter operator with a kernel size of " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\omega_{\\phi}" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": " represents the weight for level " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Specifically, the " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " loss's role is to guide " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in isolating the main degradation component (e.g., motion blur) from " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", with Gaussian filter acting to prevent " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " from incorporating residual texture interference (like content or color from " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ") onto " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ". Our test results in Section 4.3 show that without using the kernel estimation, the performance will significant decrease, verifying the necessity of the kernel estimation. In addition, we test three well-performed kernel estimation methods [1, 9, 20] and find that they achieve similar final performance. This is due to the introduction of multi-scale Gaussian low-pass filtering, the impact of noise and kernel estimation biases on the image is mitigated to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25694" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 203, + 206 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 203, + 206 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 203, + 206 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 203, + 206 + ], + "type": "image", + "image_path": "e05c545e939cc3dd152d3ac67dae2211a1d7acfad508c69a9c0e07e59eb8a0bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": "Figure 4. Comparison of our proposed self-enhancement strategy with the conventional geometric augmentation strategy in terms of training expenses, testing expenses, and performance improvement respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 380, + 206 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 380, + 206 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 380, + 206 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 380, + 206 + ], + "type": "image", + "image_path": "c05e55f9d459c976b6735499e4499f297c66ec2fbe9be28bef62e70cd0594e5e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 541, + 206 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 541, + 206 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 541, + 206 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 541, + 206 + ], + "type": "image", + "image_path": "273981147ef80c1f7e58524d328bb18efcc57eb937ad9bf2206e0aa98b913cbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 261, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 287, + 285 + ], + "type": "text", + "content": "some extent, ensuring good consistency of the main content in the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 318 + ], + "type": "text", + "content": "3.2. Inference Complexity Invariant Self-Enhancement Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": "In order to enable the reconstructor to self-correct and enhance its performance without altering its structure or increasing the complexity of the original network, we propose the Self-Enhancement (SE) strategy. Fig. 3 illustrates the implementation scheme of one of the DGIGs integrating the SE strategy, and the other DGIG with the SE strategy is similar. By employing the SE strategy, a better reconstructor " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "(Rec)" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " can be obtained, thereby learning more accurate degradation. This results in the generation of more realistic synthetic blurred-sharp image pairs and progressively improves the performance of the updated " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " with higher quality synthetic samples. It is exciting to note that the performance of the " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " using the SE strategy shows significant improvement over the one without the SE strategy. The implementation of the SE strategy involves several steps. First, we need to train an initial reconstructor " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec_{1}" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": ". Then, as shown in Fig. 3, keep the " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec_{1}" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " parameters fixed to guide the generator to learn degradation information more accurately, generating more realistic pseudo-blurry images. Subsequently, retrain " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": " reaches convergence. Finally, repeat the above process, progressively replacing and enhancing the reconstructor until obtain the best-performing " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec_{n}" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": ". Therefore, the complete " + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "inline_equation", + "content": "Rec" + }, + { + "bbox": [ + 46, + 324, + 290, + 598 + ], + "type": "text", + "content": "’s loss function is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 609, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 609, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 81, + 609, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S E - R e c} = \\mathcal {L} _ {P S N R} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (x _ {b 1}), x _ {r e c}) (4) \\\\ \\begin{array}{l} + \\mathcal {L} _ {P S N R} (\\bar {R e c} (y), y _ {r e c}) \\\\ + \\mathcal {L} _ {S S I M} (\\bar {R e c} (y), y _ {r e c}) + \\mathcal {L} _ {R e c} \\end{array} (4) \\\\ \\end{array}", + "image_path": "df97f25b60a914bca220b182cad693517b3f46e5a314d31838d86617c4abc676.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 676, + 225, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 676, + 225, + 687 + ], + "spans": [ + { + "bbox": [ + 59, + 676, + 225, + 687 + ], + "type": "text", + "content": "Finally, we obtain the total loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 697, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 697, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 697, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\min _ {G} \\max _ {D} \\mathcal {L} _ {G A N} + \\Phi_ {R P C ^ {2}} \\mathcal {L} _ {R P C ^ {2}} + \\mathcal {L} _ {S E - R e c} \\tag {5}", + "image_path": "e3444c9637189f67127ae857bb3e5c902e9affc94501bd52fc3bb1b07c19aa3f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 258, + 544, + 297 + ], + "blocks": [ + { + "bbox": [ + 310, + 258, + 544, + 297 + ], + "lines": [ + { + "bbox": [ + 310, + 258, + 544, + 297 + ], + "spans": [ + { + "bbox": [ + 310, + 258, + 544, + 297 + ], + "type": "table", + "html": "
StrategiesOriginalGeometric AugmentationSelf-Enhancement
PSNR27.6827.7929.06
PSNR Gains—0.111.38
", + "image_path": "6af792334ef6fc286beed1d0fafc7906c45e11028aabe3d45fc7f6173ac5be21.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 311, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 305, + 311, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 547, + 343 + ], + "type": "text", + "content": "Table 1. Comparing the performance enhancement of the self-enhancement strategy and the geometric augmentation strategy on the GoPro dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 365, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\Phi_{RPC^2}" + }, + { + "bbox": [ + 304, + 365, + 547, + 389 + ], + "type": "text", + "content": " represents the hyperparameter for the re-degradation principal component consistency loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 389, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 510 + ], + "type": "text", + "content": "The basic idea of the SE strategy is to use the results of the previous phase as feedback information to guide and improve the training of subsequent stages. A comparative analysis with existing typical geometric augmentation methods, as presented in Fig. 4, reveals that the SE strategy markedly boosts performance with slightly additional training cost, without imposing extra complexity or testing overhead. As Table 1 indicates, our method improves the PSNR by more than 1dB after incorporating the SE strategy, which demonstrates the effectiveness of SE strategy in image deblurring." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 540, + 501, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 501, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 501, + 553 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 676 + ], + "type": "text", + "content": "Datasets. Following the state-of-the-art image deblurring methods [5, 9], we first divide the GoPro dataset [24] (consists of 2103 pairs of blurred and sharp images) into separate sharp and blurred image parts and further constitute unpaired blurred datasets to train the algorithm proposed in this paper. We evaluate our method on the GoPro dataset [24], RealBlur dataset [32], and HIDE dataset [33]. The GoPro dataset consists of 1111 test images, the RealBlur dataset includes 980 images for testing, and the HIDE dataset provides 2025 test images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": "Implementation Details. We follow the experimental settings described in [5]. We adopt the Adam optimizer [14] " + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": ", the initial learning rate is set to " + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25695" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 536, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 536, + 266 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 536, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 536, + 266 + ], + "type": "table", + "html": "
Conference/JournalMethodsGoProHIDERealBlur-RRealBlur-J
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Deep supervisedCVPR 2018DeepDeblur [24]29.230.91625.730.87432.510.84127.870.827
CVPR 2018SRN [37]30.260.93428.360.91535.660.94728.560.867
CVPR 2021HiNet [4]32.710.95930.320.93235.750.94928.170.849
ECCV 2022Stripfromer [38]33.080.96231.030.94036.070.95228.820.876
ECCV 2022MSDI-Net [17]33.280.96431.020.94035.880.95228.590.869
ECCV 2022NAFNet [5]33.690.96731.320.94335.500.95328.320.857
ICCV 2023icDPM [31]33.200.96330.960.938N/AN/A28.810.872
CVPR 2023UFPNet [9]34.060.96831.740.94736.250.95329.870.884
Deep unsupervisedICCV 2017CycleGAN [50]22.540.72021.810.69012.380.24219.790.633
ICCV 2017DualGAN [43]22.860.722N/AN/AN/AN/AN/AN/A
CVPR 2019UIDGAN [23]23.560.73822.700.71516.640.32322.870.671
ACM MM 2022FCLGAN [48]24.840.77123.430.73228.370.66325.350.736
ICCV 2021USR-DA [39]25.490.78723.910.75632.320.82126.390.784
ACM MM 2023USDF [12]25.580.85723.930.82932.570.92326.590.881
CVPR 2023UAUD [36]26.120.86924.370.83732.910.88526.840.792
—SEMGUD (Ours)29.060.92727.640.89235.510.94628.010.844
", + "image_path": "c219b6db781c78e8aeedc64125b6403ddaaf8b1f7b71bde4c09fed522769f19e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 284, + 492, + 295 + ], + "lines": [ + { + "bbox": [ + 99, + 284, + 492, + 295 + ], + "spans": [ + { + "bbox": [ + 99, + 284, + 492, + 295 + ], + "type": "text", + "content": "Table 2. The comparison results on the benchmark datasets. All the models are trained on the GoPro dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 305, + 217, + 479 + ], + "blocks": [ + { + "bbox": [ + 56, + 305, + 217, + 479 + ], + "lines": [ + { + "bbox": [ + 56, + 305, + 217, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 217, + 479 + ], + "type": "image", + "image_path": "b14cb85280179695def7ff6151ad0b8bd294f22c940e2c097bdb882330872f07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 546, + 510 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on the GoPro dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 305, + 299, + 479 + ], + "blocks": [ + { + "bbox": [ + 225, + 305, + 299, + 479 + ], + "lines": [ + { + "bbox": [ + 225, + 305, + 299, + 479 + ], + "spans": [ + { + "bbox": [ + 225, + 305, + 299, + 479 + ], + "type": "image", + "image_path": "5784cae6c3f153aec221851357223d943c871104dc1cdfed6151e7b8d88ed60d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 306, + 378, + 478 + ], + "blocks": [ + { + "bbox": [ + 304, + 306, + 378, + 478 + ], + "lines": [ + { + "bbox": [ + 304, + 306, + 378, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 378, + 478 + ], + "type": "image", + "image_path": "d5ceaf78144a3d24b97fbcba9d572d5a98ae955c060d7845dfbb57b8627d311a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 306, + 457, + 478 + ], + "blocks": [ + { + "bbox": [ + 383, + 306, + 457, + 478 + ], + "lines": [ + { + "bbox": [ + 383, + 306, + 457, + 478 + ], + "spans": [ + { + "bbox": [ + 383, + 306, + 457, + 478 + ], + "type": "image", + "image_path": "3dd219cd63d3fbf8bc849bfa14a850e8c4d9dc903cea8941c2780275ac3bc959.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 462, + 306, + 536, + 478 + ], + "blocks": [ + { + "bbox": [ + 462, + 306, + 536, + 478 + ], + "lines": [ + { + "bbox": [ + 462, + 306, + 536, + 478 + ], + "spans": [ + { + "bbox": [ + 462, + 306, + 536, + 478 + ], + "type": "image", + "image_path": "682078aaebcd8c3f22395c62ed0cbbc0f7bc77ad6a5e6fd0f6976f3b24183a72.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": "and the training patch size is " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": " loss hyperparameter " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "(\\Phi_{RPC^2})" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": " in the loss function of Eq. 5 is set to 3. More details are presented in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 586, + 170, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 170, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 170, + 600 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "We compare our method with the most recent unsupervised methods from recent years, including [12, 23, 36, 39, 43, 48, 50]. Additionally, we also list the state-of-the-art supervised methods based on paired images, including [4, 5, 9, 17, 24, 31, 37, 38]. Note that the related models for comparison are limited, since only a few unsupervised deblurring models have been proposed in the field. We evaluate the effectiveness of each method using performance metrics (PSNR and SSIM). The results are directly cited from the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 531, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 555 + ], + "type": "text", + "content": "original papers or generated using the official models provided by the authors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "Quantitative comparison. Table 2 presents the PSNR and SSIM results for various single-image deblurring test methods on the GoPro, HIDE, and RealBlur test datasets. It is evident that our proposed method outperforms other unsupervised methods. On the GoPro dataset, our method achieves 2.94dB improvement in PSNR over the state-of-the-art unsupervised methods. To validate the effectiveness and generalization of our method, we compare the results on various real-world blurry datasets. Fig. 1 intuitively demonstrates the performance improvement of our method compared to other unsupervised methods. It is well known that supervised methods usually outperform unsupervised methods. But on the RealBlur-R dataset, the performance of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25696" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 218, + 243 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 218, + 243 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 218, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 218, + 243 + ], + "type": "image", + "image_path": "7480356da02edf20a5981789cc9a917c7f6dfe752ce858b58c6795501d697f79.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 275 + ], + "type": "text", + "content": "Figure 6. Visual comparisons on the Hide dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 300, + 242 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 300, + 242 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 300, + 242 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 300, + 242 + ], + "type": "image", + "image_path": "1d7422404ba8cac0d6d7de6736fe5403e3516895a6b449a0a21e0d0bac5e9302.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 378, + 242 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 378, + 242 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 378, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 378, + 242 + ], + "type": "image", + "image_path": "17fa5c39a9600ecb6a6ca3752d85371b802c82bde777ada67ab668801b63195b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 457, + 242 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 457, + 242 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 457, + 242 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 457, + 242 + ], + "type": "image", + "image_path": "bb353c0a29355a9f9a2bed4460b76f519cd2df7cf074476e3b4ddc4947840762.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 462, + 70, + 536, + 242 + ], + "blocks": [ + { + "bbox": [ + 462, + 70, + 536, + 242 + ], + "lines": [ + { + "bbox": [ + 462, + 70, + 536, + 242 + ], + "spans": [ + { + "bbox": [ + 462, + 70, + 536, + 242 + ], + "type": "image", + "image_path": "bb064c455cfc5f6b89a65773a60c0d7ef5d06c0290bad75bbf8708ac067f4a08.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 284, + 218, + 457 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 218, + 457 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 218, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 218, + 457 + ], + "type": "image", + "image_path": "531aa58098985cbd7eb82b7e9218ce4f0042c5123e9f6e5ade4c9589e9ec2a82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 466, + 548, + 490 + ], + "lines": [ + { + "bbox": [ + 46, + 466, + 548, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 548, + 490 + ], + "type": "text", + "content": "Figure 7. Visual comparisons on the RealBlur dataset. From left to right: blurry image, results from CycleGAN [50], UIDGAN [23], USR-DA [39], FCLGAN [48], UAUD [36], SEMGUD (ours), and ground-truth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 284, + 300, + 457 + ], + "blocks": [ + { + "bbox": [ + 225, + 284, + 300, + 457 + ], + "lines": [ + { + "bbox": [ + 225, + 284, + 300, + 457 + ], + "spans": [ + { + "bbox": [ + 225, + 284, + 300, + 457 + ], + "type": "image", + "image_path": "12c0ec784942512f63992c1d0b8daaec6309b9286c2268afca203354a10d57ca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 304, + 284, + 379, + 456 + ], + "blocks": [ + { + "bbox": [ + 304, + 284, + 379, + 456 + ], + "lines": [ + { + "bbox": [ + 304, + 284, + 379, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 379, + 456 + ], + "type": "image", + "image_path": "e262d8497aae3aaafced51757d371fb1e281c0b168607abf453ce5c3f7c09ffd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 383, + 284, + 458, + 456 + ], + "blocks": [ + { + "bbox": [ + 383, + 284, + 458, + 456 + ], + "lines": [ + { + "bbox": [ + 383, + 284, + 458, + 456 + ], + "spans": [ + { + "bbox": [ + 383, + 284, + 458, + 456 + ], + "type": "image", + "image_path": "9a0882089ccea1c85ca2154d70f4553305fd8b6c067532d5bae8215600ee3e45.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 462, + 284, + 536, + 456 + ], + "blocks": [ + { + "bbox": [ + 462, + 284, + 536, + 456 + ], + "lines": [ + { + "bbox": [ + 462, + 284, + 536, + 456 + ], + "spans": [ + { + "bbox": [ + 462, + 284, + 536, + 456 + ], + "type": "image", + "image_path": "47b5cdc356f2f7a16ed1c9cb9ddb14a836e102420c742885f1922f6e386352e4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 51, + 507, + 108, + 582 + ], + "blocks": [ + { + "bbox": [ + 51, + 507, + 108, + 582 + ], + "lines": [ + { + "bbox": [ + 51, + 507, + 108, + 582 + ], + "spans": [ + { + "bbox": [ + 51, + 507, + 108, + 582 + ], + "type": "image", + "image_path": "d2c9cadc1222e39aa16f06cdea2eaa986ea98c9660895dba068e53ac0e7698ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 591, + 289, + 647 + ], + "lines": [ + { + "bbox": [ + 46, + 591, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 289, + 647 + ], + "type": "text", + "content": "Figure 8. The \"mode collapse\" in NAFNet [5] and UFPNet [9]: trained on the GoPro dataset may output anomalous pixel regions during testing on the RealBlur-J dataset. \"(UFPNet)+\" and \"(NAFNet)+\" denote the results obtained through training with our SEMGUD framework." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 111, + 508, + 167, + 581 + ], + "blocks": [ + { + "bbox": [ + 111, + 508, + 167, + 581 + ], + "lines": [ + { + "bbox": [ + 111, + 508, + 167, + 581 + ], + "spans": [ + { + "bbox": [ + 111, + 508, + 167, + 581 + ], + "type": "image", + "image_path": "b165c71756d845482603c54b4b3b1db3955c6c1e1738b5d4c286956c4e3788b7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 170, + 508, + 227, + 582 + ], + "blocks": [ + { + "bbox": [ + 170, + 508, + 227, + 582 + ], + "lines": [ + { + "bbox": [ + 170, + 508, + 227, + 582 + ], + "spans": [ + { + "bbox": [ + 170, + 508, + 227, + 582 + ], + "type": "image", + "image_path": "132962a3ed0a86da4023204dfc6ce6481c32b7b74f81792b5800225463da6a39.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 230, + 508, + 287, + 582 + ], + "blocks": [ + { + "bbox": [ + 230, + 508, + 287, + 582 + ], + "lines": [ + { + "bbox": [ + 230, + 508, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 230, + 508, + 287, + 582 + ], + "type": "image", + "image_path": "346e8d511014b60dbbc77c97d21fea48e589692002274368ad3c4e4303a185bd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "our method is even higher than NAFNet [5]. This suggests that our method's effectiveness is improving and it's even becoming competitive with certain supervised methods. It's" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 312, + 507, + 542, + 596 + ], + "blocks": [ + { + "bbox": [ + 312, + 507, + 542, + 596 + ], + "lines": [ + { + "bbox": [ + 312, + 507, + 542, + 596 + ], + "spans": [ + { + "bbox": [ + 312, + 507, + 542, + 596 + ], + "type": "table", + "html": "
KEφ3,9,15ωφGoPro
3,5,73,7,91,0.2,0.041,0.1,0.01PSNR↑SSIM↑
✓✓✗✗✗✓28.540.919
✓✗✗✓✓✗28.720.921
✓✗✗✓✗✓28.830.923
✓✗✓✗✓✗28.890.922
✓✗✓✗✗✓28.970.924
✓✓✗✗✓✗29.000.925
✓✓✗✗✗✓29.060.927
", + "image_path": "dcabcb265585a97d96fab9d48375f859019a449d2e2774529ca4143075f13998.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "lines": [ + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "text", + "content": "Table 3. Ablation study on the hyperparameters " + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "inline_equation", + "content": "{\\omega }_{\\phi }" + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "inline_equation", + "content": "{RPC}^{2}" + }, + { + "bbox": [ + 305, + 609, + 547, + 641 + ], + "type": "text", + "content": " loss. The first column indicates whether the kernel estimation network is used." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "worth noting that all the mentioned models are trained on the GoPro dataset, which further demonstrates the excellent generalization performance of our method from GoPro to other real-world datasets." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25697" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 110, + 129 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 110, + 129 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 110, + 129 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 110, + 129 + ], + "type": "image", + "image_path": "062ffff7ac238f6012ae3a41484d13304544e8be0c099cff4178b8e09fdc1950.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 129, + 93, + 140 + ], + "lines": [ + { + "bbox": [ + 70, + 129, + 93, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 129, + 93, + 140 + ], + "type": "text", + "content": "Blurry" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "type": "text", + "content": "Figure 9. Crop the blurred region from the ReloBlur dataset for visual comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 70, + 168, + 129 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 168, + 129 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 168, + 129 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 168, + 129 + ], + "type": "image", + "image_path": "d5e02569953d6005ca469730a6a3288f46fc6d21fc5076669fa377e219c47b28.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 130, + 154, + 139 + ], + "lines": [ + { + "bbox": [ + 125, + 130, + 154, + 139 + ], + "spans": [ + { + "bbox": [ + 125, + 130, + 154, + 139 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 228, + 129 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 228, + 129 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 228, + 129 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 228, + 129 + ], + "type": "image", + "image_path": "5f9758134b43c40db393afc420516d7c57bd7665fa17240131977323a6ee2cba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 130, + 208, + 138 + ], + "lines": [ + { + "bbox": [ + 190, + 130, + 208, + 138 + ], + "spans": [ + { + "bbox": [ + 190, + 130, + 208, + 138 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 70, + 288, + 129 + ], + "blocks": [ + { + "bbox": [ + 230, + 70, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 230, + 70, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 230, + 70, + 288, + 129 + ], + "type": "image", + "image_path": "ad08660141929a68e0e18480ffe81b77fd8e34c6bda8004204334d407f1712e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 130, + 269, + 140 + ], + "lines": [ + { + "bbox": [ + 249, + 130, + 269, + 140 + ], + "spans": [ + { + "bbox": [ + 249, + 130, + 269, + 140 + ], + "type": "text", + "content": "Sharp" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 190, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 289, + 430 + ], + "type": "text", + "content": "Visual comparison. As shown in Figs. 5, 6, and 7, we compare the visual deblurring results of different unsupervised methods on the GoPro, HIDE, and RealBlur datasets, respectively. Note that although UAUD [36] effectively restores images with uniform blur, such performance may not extend to images with non-uniform blur. It can be seen that our proposed method achieves good results in removing motion blur from real blurry images. Additionally, we note that existing unsupervised methods often perform poorly in dealing with more severe blurring as shown in Fig. 6. Furthermore, we observe that while some existing advanced supervised methods like NAFNet and UFPNet achieve the best performance when trained on the GoPro paired datasets, they exhibit \"mode collapse\" when directly applied to other datasets, as illustrated in Fig. 8. In contrast, unsupervised models trained with our framework do not exhibit this phenomenon. Additionally, we directly use our GoPro-trained model to test the ReloBlur [18] dataset, which contains Real World Partly-blurred images, and find it to be effective for some of the images, as shown in Fig. 9." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 438, + 287, + 563 + ], + "blocks": [ + { + "bbox": [ + 52, + 438, + 287, + 563 + ], + "lines": [ + { + "bbox": [ + 52, + 438, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 52, + 438, + 287, + 563 + ], + "type": "image", + "image_path": "09370441a3732239eac3107e1dad50b1af5d0a5cb110f74f8f126542ddd4b21e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 571, + 288, + 605 + ], + "lines": [ + { + "bbox": [ + 46, + 571, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 288, + 605 + ], + "type": "text", + "content": "Figure 10. Performance improvement (PSNR gain) of different methods on various datasets driven by the self-enhancement strategy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "The Re-Degradation Principal Component Consistency Loss. In fact, we find that if we remove the " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " loss, the generator always struggles to generate satisfactory blurry images. Due to the minimal impact of different kernel estimation networks on the final deblurring performance as illustrated in Section 3.1.3, we can use any well-performed" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 541, + 168 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 168 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 168 + ], + "type": "table", + "html": "
MethodsSEGoProHIDE
PSNR↑SSIM↑PSNR↑SSIM↑
FCLGAN [48]X24.700.76823.490.733
✓25.110.77223.790.741
UFPNet [9]X27.370.90325.910.862
✓28.410.91626.670.875
NAFNet [5] (Ours)X27.680.91126.240.859
✓29.060.92727.640.892
", + "image_path": "2a1f232f7cafda12fda9e33dfe79c43e7290c65418878eff63aa1994cb3a6046.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 182, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 204 + ], + "type": "text", + "content": "Table 4. Ablation study on the superiority and generalization of the self-enhancement strategy for different reconstructors." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "text", + "content": "kernel estimation networks, e.g., [1] can be used for testing. Therefore, our ablation study on the " + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "inline_equation", + "content": "RPC^2" + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "text", + "content": " loss primarily focuses on the blur kernel estimation network and the settings of hyperparameters " + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 224, + 547, + 297 + ], + "type": "text", + "content": " in Eq. 3. As shown in Table 3, we observe that the blur kernel estimation network significantly improves the performance of our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 297, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 453 + ], + "type": "text", + "content": "The Self-Enhancement Strategy. As shown in Table 4, the effect is improved by " + }, + { + "bbox": [ + 304, + 297, + 547, + 453 + ], + "type": "inline_equation", + "content": "1.38\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 297, + 547, + 453 + ], + "type": "text", + "content": " on the GoPro dataset with the SE strategy. Furthermore, to demonstrate the superiority and versatility of our proposed self-enhancement strategy, we also conduct experiments incorporating SE strategy on several deblurring methods, including NAFNet [5], FCLGAN [48], UFPNet [9]. Fig. 10 shows in detail the different network performance improvements before and after using the self-enhancement strategy. Note that due to FCLGAN do not provide the complete network weights, the experimental results in the table are obtained from our retraining. This strongly validates the superiority and significant potential of the self-enhancement strategy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 483, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 639 + ], + "type": "text", + "content": "In this paper, we introduce a novel unsupervised framework for image deblurring. Our approach incorporates a re-degradation principal component consistency loss, ensuring that the principal components of the synthetically blurred images align closely with those from the re-degraded versions of the original sharp images. We also put forward a self-enhancement strategy that substantially improves performance without modifying the model's architecture or adding to the computational cost in inference. Comprehensive tests on standard datasets reveal that our method surpasses the existing state-of-the-art unsupervised methods with strong generalization capabilities across various real-world blurry image datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and the Cooperation Science and Technology Project of Sichuan University and Dazhou City under Grant 2022CDDZ-09." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25698" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 134 + ], + "type": "text", + "content": "[1] Guillermo Carbajal, Patricia Vitoria, Mauricio Delbracio, Pablo Muse, and José Lezama. Non-uniform motion blur kernel estimation via adaptive decomposition. arXiv e-prints, pages arXiv-2102, 2021. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "text", + "content": "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV), pages 221-235, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[3] Huaijin Chen, Jinwei Gu, Orazio Gallo, Ming-Yu Liu, Ashok Veeraraghavan, and Jan Kautz. Reblur2deblur: Deblurring videos via self-supervised learning. In IEEE International Conference on Computational Photography (ICCP), pages 1-9, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 182-192, 2021. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "text", + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "text", + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "text", + "content": "[7] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE signal processing magazine, pages 53-65, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[8] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345-12354, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "type": "text", + "content": "[9] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18105-18114, 2023. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "text", + "content": "[10] Amit Goldstein and Raanan Fattal. Blur-kernel estimation from spectral irregularities. In European Conference on Computer Vision (ECCV), pages 622-635, 2012. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[11] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[12] Runhua Jiang and Yahong Han. Uncertainty-aware variate decomposition for self-supervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 252-260, 2023. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In IEEE International Conference on Image Processing (ICIP), pages 2761-2765, 2019. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "text", + "content": "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "text", + "content": "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8878-8887, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "text", + "content": "[17] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In European Conference on Computer Vision (ECCV), pages 736-753, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "text", + "content": "[18] Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. Real-world deep local motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1314–1322, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 321, + 546, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 546, + 375 + ], + "type": "text", + "content": "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 1833-1844, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "type": "text", + "content": "[20] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4096-4105, 2021. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 546, + 487 + ], + "type": "text", + "content": "[21] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12642-12652, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 488, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 546, + 521 + ], + "type": "text", + "content": "[22] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient Information Modulation Network for Image Super-Resolution. 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "text", + "content": "[23] Boyu Lu, Jun-Cheng Chen, and Rama Chellappa. Unsupervised domain-specific deblurring via disentangled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10225-10234, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 546, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 546, + 633 + ], + "type": "text", + "content": "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3883-3891, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "text", + "content": "[25] Thekke Madam Nimisha, Kumar Sunil, and AN Rajagopalan. Unsupervised class-specific deblurring. In European Conference on Computer Vision (ECCV), pages 353-369, 2018. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[26] Jinshan Pan, Zhe Hu, Zhixun Su, and Ming-Hsuan Yang. Deblurring text images via 10-regularized intensity and gradient prior. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25699" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 2901-2908, 2014. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "text", + "content": "[27] Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1628-1636, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "text", + "content": "[28] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 249 + ], + "type": "text", + "content": "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision (ECCV), pages 327-343, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 293 + ], + "type": "text", + "content": "[30] Chao Ren, Xiaohai He, and Truong Q Nguyen. Adjusted non-local regression and directional smoothness for image restoration. IEEE Transactions on Multimedia, pages 731-745, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 288, + 348 + ], + "type": "text", + "content": "[31] Mengwei Ren, Maurizio Delbracio, Hossein Talebi, Guido Gering, and Peyman Milanfar. Multiscale structure guided diffusion for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10721-10733, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "text", + "content": "[32] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In European Conference on Computer Vision (ECCV), pages 184-201, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "type": "text", + "content": "[33] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5572-5581, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "text", + "content": "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606–3615, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 505, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 288, + 557 + ], + "type": "text", + "content": "[35] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[36] Xiaole Tang, Xile Zhao, Jun Liu, Jianli Wang, Yuchun Miao, and Tieyong Zeng. Uncertainty-aware unsupervised image deblurring with deep residual prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9883-9892, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 615, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 288, + 667 + ], + "type": "text", + "content": "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[38] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European Conference on Computer Vision (ECCV), pages 146-162, 2022. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 686 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[39] Wei Wang, Haochen Zhang, Zehuan Yuan, and Changhu Wang. Unsupervised real-world super-resolution: A domain adaptation perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4318-4327, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, pages 168-186, 2012. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 205 + ], + "type": "text", + "content": "[41] Lin X, Yue J, and Ren C. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. In arXiv, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "type": "text", + "content": "[42] Li Xu, Shicheng Zheng, and Jiaya Jia. Unnatural 10 sparse representation for natural image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1107-1114, 2013. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 297 + ], + "type": "text", + "content": "[43] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2849-2857, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "type": "text", + "content": "[44] Yuan Yuan, Wei Su, and Dandan Ma. Efficient dynamic scene deblurring using spatially variant deconvolution network with optical flow guided training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3555-3564, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 547, + 407 + ], + "type": "text", + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 410, + 547, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 547, + 463 + ], + "type": "text", + "content": "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 465, + 546, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 546, + 518 + ], + "type": "text", + "content": "[47] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 520, + 546, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 546, + 575 + ], + "type": "text", + "content": "[48] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, and Meng Wang. Fcl-gan: A lightweight and real-time baseline for unsupervised blind image deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6220-6229, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 577, + 547, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 547, + 631 + ], + "type": "text", + "content": "[49] Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Hajjun Zhang, Meng Wang, and Shuicheng Yan. Crnet: Unsupervised color retention network for blind motion deblurring. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 6193-6201, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 632, + 547, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 547, + 686 + ], + "type": "text", + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2223-2232, 2017. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25700" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_content_list.json b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d803c892a45a83cdcb02428da83e5df22951e994 --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_content_list.json @@ -0,0 +1,1857 @@ +[ + { + "type": "text", + "text": "Unsupervised Deep Unrolling Networks for Phase Unwrapping", + "text_level": 1, + "bbox": [ + 168, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhile Chen $^{1,2}$ Yuhui Quan $^{1,2*}$ Hui Ji $^{3}$", + "bbox": [ + 285, + 179, + 681, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China $^{2}$ Pazhou Lab, Guangzhou 510320, China", + "bbox": [ + 107, + 200, + 861, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Mathematics, National University of Singapore, 119076, Singapore", + "bbox": [ + 207, + 234, + 759, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cs.zhilechen@mail.scut.edu.cn, csyhquan@scut.edu.cn, matjh@nus.edu.sg", + "bbox": [ + 176, + 253, + 785, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phase unwrapping (PU) is a technique to reconstruct original phase images from their noisy wrapped counterparts, finding many applications in scientific imaging. Although supervised learning has shown promise in PU, its utility is limited in ground-truth (GT) scarce scenarios. This paper presents an unsupervised learning approach that eliminates the need for GTs during end-to-end training. Our approach leverages the insight that both the gradients and wrapped gradients of wrapped phases serve as noisy labels for GT phase gradients, along with sparse outliers induced by the wrapping operation. A recorruption-based self-reconstruction loss in the gradient domain is proposed to mitigate the adverse effects of label noise, complemented with a self-distillation loss for improved generalization. Additionally, by unfolding a variational model of PU that utilizes wrapped gradients of wrapped phases for its data-fitting term, we develop a deep unrolling network that encodes physics of phase wrapping and incorporates special treatments on outliers. In the experiments on three types of phase data, our approach outperforms existing GT-free methods and competes well against the supervised ones.", + "bbox": [ + 75, + 335, + 473, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 207, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many imaging systems, the direct acquisition of original phases of a target signal is both challenging and costly. Typically, initial measurement yields a phase image wrapped in $[- \\pi, \\pi)$ . Let $X \\in \\mathbb{R}^{M \\times N}$ and $Y \\in [-\\pi, \\pi)^{M \\times N}$ denote the true phase image and its noisy wrapped counterpart, respectively. The wrapping process can be formulated as:", + "bbox": [ + 75, + 704, + 468, + 796 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} = \\mathcal {W} (\\boldsymbol {X} + \\boldsymbol {N}), \\quad \\mathcal {W}: \\mathbb {R} ^ {M \\times N} \\rightarrow [ - \\pi , \\pi) ^ {M \\times N}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 801, + 468, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ denotes the wrapping operator with its entry-wise operation defined as $\\mathcal{W}(\\theta) = ((\\theta + \\pi) \\bmod 2\\pi) - \\pi$ for $\\theta \\in \\mathbb{R}$ , and $N \\in \\mathbb{R}^{M \\times N}$ denotes the measurement noise, typically assumed to be Gaussian [22]. Consequently, the original phase image, which exhibits strong smoothness, is degraded to a wrapped version with both incorrect values and artificial discontinuities induced by the modulo operation, which adversely affects subsequent processing.", + "bbox": [ + 496, + 305, + 892, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phase unwrapping (PU) is about recovering the original unwrapped phase image $\\mathbf{X}$ (up to an additive constant) from its noisy wrapped counterpart $\\mathbf{Y}$ . Serving as a critical step in many optical imaging techniques for generating clear and coherent phase images, PU finds extensive applications in diverse domains such as quantitative phase imaging, magnetic resonance imaging, synthetic aperture radar interferometry, 3D depth sensing, phase contrast microscopy, fringe projection, and digital holographic interferometry; see e.g. [10, 14, 16, 23, 41, 47, 56, 57].", + "bbox": [ + 496, + 426, + 892, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The model (1) is also usually formulated as follows:", + "bbox": [ + 517, + 577, + 864, + 592 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} = \\boldsymbol {X} - 2 \\pi \\boldsymbol {K} + \\boldsymbol {N} ^ {\\prime}, \\quad \\boldsymbol {K} \\in \\mathbb {Z} ^ {M \\times N}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 599, + 890, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $K \\in \\mathbb{Z}^{M \\times N}$ denotes the map of wrap counts, indicating the number of times a phase value has been wrapped around by $2\\pi$ , and $N'$ denotes the noise dependent on both $X$ and $N$ , whose distribution can be complex even when $N$ follows a simple distribution such as the normal distribution. It can be seen that PU is a challenging ill-posed inverse problem involving a continuous variable $X$ and a discrete variable $K$ . Direct solutions by integrating wrapped phase differences are insufficient for ensuring spatial consistency in all directions of the unwrapped phases and may accumulate significant errors due to substantial noise.", + "bbox": [ + 496, + 625, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Limitations of existing works: Conventional approaches for PU include reliability-guided methods (e.g. [3, 9, 11, 23, 36]), filtering-based methods (e.g. [2, 45]), and model-driven methods (e.g. [1, 12, 15]). These approaches are either sensitive to measurement noise, due to their heuristic rules, or overly simplistic for complex phase structures, due to their handcrafted image priors. In contrast, deep learning", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Yuhui Quan.", + "bbox": [ + 94, + 827, + 292, + 840 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is supported by National Natural Science Foundation of China (Grant No. 62372186), Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755, 2023A1515012841), Fundamental Research Funds for Central Universities (Grant No. x2jsD2230220), and Singapore MOE AcRF Tier 1 (Grant No. A-8000981-00-00).", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25182", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5d1b7d488142b56bf5730f32ea270b2b72b03793c60e50c6631347304ee3c2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 205, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/568acda83cbd389b33f62b98b25702fd3e47a9cbb7758f873558114cc829e178.jpg", + "image_caption": [ + "Randomly generated Set" + ], + "image_footnote": [], + "bbox": [ + 230, + 107, + 313, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4408d9a135bed6a749e9f4540b0b86708297caeb14ddeceebc847a3b081ec5dd.jpg", + "image_caption": [ + "Wrapped Phase Images" + ], + "image_footnote": [], + "bbox": [ + 341, + 99, + 441, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d53fd8067bf8677d978b49aba8118a13484dcbc1d67b863b9c35a4f787d6a00f.jpg", + "image_caption": [ + "Supervised Learning", + "Figure 1. Training data setting of existing PU methods versus ours." + ], + "image_footnote": [ + "- GT collection is hard" + ], + "bbox": [ + 101, + 178, + 200, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9658c25f4c65824b637632f84c479f946678ea5a960627a6cf26b955fe309a78.jpg", + "image_caption": [ + "Wrappered Phase Image (Single Test Sample)", + "Internal Learning" + ], + "image_footnote": [ + "Time consuming" + ], + "bbox": [ + 230, + 198, + 313, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/492b3f000e6d48e83b7fc1417bc5f9d74aa02c3e93ec7733f559883528364f94.jpg", + "image_caption": [ + "Wrapped Phase Images", + "End-to-End", + "Unsupervised Learning" + ], + "image_footnote": [ + "Bypassing both issues" + ], + "bbox": [ + 341, + 189, + 439, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(DL) with neural networks (NNs) has emerged as a powerful tool for PU [41]. A pre-trained end-to-end NN offers data adaptivity, rapid inference, and excellent expressivity.", + "bbox": [ + 75, + 352, + 467, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing DL-based PU methods often rely on supervised learning with ground-truth (GT) phases. However, acquiring GT phases is often impractical due to physical limits in imaging, destructive imaging techniques, and ethical/safety concerns. This challenge limits the applicability of supervised DL in many practical scenarios. Internal unsupervised DL-based methods (e.g. [47]) use an untrained NN to parameterize latent phases, adjusting it to fit observed measurements with certain regularization. Though flexible, this kind of methods requires time-costly per-sample test-time training, resulting in poor computational efficiency compared to supervised DL-based methods.", + "bbox": [ + 75, + 398, + 467, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach: To bypass the difficulty of GT acquisition in supervised DL and avoid the computational inefficiency of internal unsupervised DL, we present an external unsupervised $DL$ approach for PU. It trains an end-to-end NN solely on a dataset of noisy wrapped measurements. See Fig. 1 for the comparison in training data configuration between existing works and ours. Indeed, our study is the first one on external (end-to-end) unsupervised DL for PU.", + "bbox": [ + 75, + 582, + 467, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is inspired by the insight that applying the gradient operator or the wrapping gradient operator to a wrapped phase image will result in an output identical to the gradient map of the GT phase image, with the exception of a sparse outlier set. Concretely, we have that", + "bbox": [ + 75, + 704, + 467, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) $\\nabla Y[m,n] = \\nabla X[m,n] + \\nabla N[m,n]$ holds for the points whose neighbors share the same wrap count.", + "(b) $\\mathcal{W}(\\nabla Y[m,n]) = \\nabla X[m,n] + \\nabla N[m,n]$ is applicable for the points that meet the 2D Itoh's continuity condition [13]." + ], + "bbox": [ + 81, + 779, + 467, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The majority of points in a GT phase image fall under the aforementioned categories. Therefore, on these points, $\\nabla Y$ or $\\mathcal{W}(\\nabla Y)$ can serve as a noisy label of $\\nabla X$ . Once $\\nabla X$", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is predicted, reconstructing $X$ is relatively simple.", + "bbox": [ + 498, + 90, + 833, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach utilizes $\\nabla Y$ for unsupervised training and tackles its label noise via a recorruption-based self-reconstruction loss function in the gradient domain, which is inspired by existing recorruption-based denoisers [19, 21, 46]. In addition, to mitigate the potential ambiguity and overfitting caused by the absence of GTs, a self-distillation loss function is proposed that leverages intermediate predictions from the NN for pseudo-supervision.", + "bbox": [ + 496, + 107, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To encode physical model/priors of phase wrapping into the NN-learned PU process, which is beneficial for mitigating possible overfitting particularly when training on a limited amount of data, we develop a deep unrolling network (DUN) by unfolding a variational model that leverages $\\mathcal{W}(\\nabla Y)$ for data fitting. To our knowledge, this is the first work to utilize deep unrolling for PU. Indeed, most existing studies rely on U-shaped convolutional NNs (CNNs) or recurrent NNs (RNNs) that ignore physical priors of phase wrapping. Additionally, our DUN is tailored to handle the outliers that do not satisfy the Itoh's continuity condition, thereby mitigating their potential adverse effects.", + "bbox": [ + 496, + 229, + 890, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main contributions: Our contributions include", + "text_level": 1, + "bbox": [ + 500, + 414, + 821, + 428 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first external unsupervised DL approach for end-to-end PU, circumventing the challenge of GT acquisition in supervised DL and the computational inefficiency of internal unsupervised DL.", + "- We perform the first exploration of deep unfolding for PU, founded on a variational model that utilizes wrapped gradients of wrapped phases for data fitting under Itoh's continuity condition. The resulting DUN is outlier-aware and enjoys advantages in interpretability and performance.", + "- We propose a recorruption-based self-reconstruction loss function with noise tolerance to leverage Itoh's continuity condition, as well as a self-distillation loss function for improved generalization." + ], + "bbox": [ + 500, + 431, + 890, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In extensive experiments, our proposed approach shows superior performance over existing GT-free PU methods and competes well against the supervised ones.", + "bbox": [ + 498, + 630, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 691, + 640, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-learning-based PU: Methods along this line fall into three types: reliability-guided, filtering-based, and model-driven. Reliability-guided methods [3, 9, 11, 23, 36, 54] use a manually-designed metric to quantify pixel-wise reliability for path selection or region expansion during phase correction. Due to unreliable quality scores in noisy regions, quality-guided methods are noise-sensitive. Filtering-based methods [2, 45] treat PU as a smoothing problem, leveraging non-linear filters for denoising and unwrapping. However, their limited-support filters fail to exploit global consistency constraints within a phase image. Model-driven methods [1, 8, 12, 15] construct and solve an optimization", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model of PU regularized by some handcrafted priors. While achieving strong global consistency brought by the global optimization, these methods often fall short for complex phase patterns, due to the use of handcrafted priors.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised DL for PU: Supervised DL-based PU methods differ mainly in NN architectures. Existing methods can be roughly classified as CNN-based or RNN-based.", + "bbox": [ + 75, + 154, + 468, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging the effectiveness of CNNs in segmentation and classification, many CNN-based methods (e.g. [10, 34, 35, 49, 50, 52, 56]) cast PU as a wrap count prediction problem, using a pixel-wise classification loss with wrap counts as labels. For robustness to noise, pre-denoising or post-processing modules are often included; see e.g. [34, 35, 52]. However, these methods face scalability issues with widerange phase data, where the candidate values for wrap counts are overwhelmed. Instead, some CNN-based methods [6, 25, 40, 57]. Typically, CNN-based methods are inefficient in exploiting global spatial dependencies. In contrast, the RNN-based methods [22, 32] excel at capturing long-range dependencies. Viewing a phase pixel as an element in a sequence, these methods predefine several paths across all phase pixels to form an RNN. However, due to computational costs, only a few paths can be utilized. This constraint on path selection results in suboptimal outcomes.", + "bbox": [ + 75, + 199, + 468, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The NN designs of most existing methods overlook the physical priors of phase wrapping, which are beneficial for generalization, especially when training data is limited. Our proposed DUN incorporates these priors, enhancing interpretability, reducing overfitting, and boosting performance.", + "bbox": [ + 75, + 455, + 468, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised DL for PU: Despite significant advances in unsupervised DL for various inverse problems in imaging, e.g., denoising [17, 21, 26], deblurring [5, 18, 27, 31], compressed sensing [28, 38, 42] and phase retrieval [4, 29], unsupervised DL for PU has been much less studied. One related work is Yang et al. [47], which utilizes the deep image prior [39] encoded by an untrained CNN to address the absence of GT data. Specifically, the phase image is reparameterized by a CNN, with its weights learned to fit the measurements. As shown in [39, 47], the CNN prefers smooth patterns and requires early stopping for smooth phase unwrapping, akin to conventional model-based methods but employing a CNN-based prior. A drawback is the need for per-sample retraining, which is computationally intensive. In contrast, our approach allows end-to-end unsupervised training on a dataset, eliminating test-time learning and significantly speeding up test.", + "bbox": [ + 75, + 535, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DUNs for imaging: DUNs have been studied for many inverse problems in imaging, e.g., denoising [20, 55], super-resolution [51], deblurring [20], compressed sensing [24, 53] and medical imaging [7]. However, the study on deep unrolling for PU is scant. Our study is the first to develop DUNs for PU, in addition to an unsupervised training scheme.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 635, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Basics and Keys", + "text_level": 1, + "bbox": [ + 500, + 114, + 660, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider estimating $\\mathbf{X}$ from its wrapped version $\\mathcal{W}(\\mathbf{X}) = \\mathbf{X} - 2\\pi \\mathbf{K}, \\mathbf{K} \\in \\mathbb{Z}^{M \\times N}$ . Let $\\nabla : \\mathbf{Z} \\to (\\nabla_x \\mathbf{Z}; \\nabla_y \\mathbf{Z})$ denote the spatial gradient operator. By definition,", + "bbox": [ + 500, + 137, + 893, + 184 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\boldsymbol {\\nabla} \\boldsymbol {X} [ m, n ] + 2 \\pi \\boldsymbol {\\nabla} \\boldsymbol {K} [ m, n ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 189, + 890, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For any location $[m,n]$ whose adjacent pixels for gradient calculation have the same wrap count, we have", + "bbox": [ + 498, + 213, + 893, + 244 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 250, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In other words, the wrap count can be nullified by the gradient operator on many points. For the points not meeting the condition, we have the following Proposition.", + "bbox": [ + 498, + 273, + 890, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. [Two-dimensional extension of Itoh's continuity condition [13]] For any point $[m, n]$ satisfying", + "bbox": [ + 498, + 325, + 890, + 358 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\nabla \\boldsymbol {X} [ m, n ] \\right\\| _ {\\infty} = \\max \\left\\{\\left| \\nabla_ {x} \\boldsymbol {X} [ m, n ] \\right|, \\left| \\nabla_ {y} \\boldsymbol {X} [ m, n ] \\right| \\right\\} < \\pi , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 364, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we have that $\\mathcal{W}(\\nabla \\mathcal{W}(\\pmb {X}))[m,n] = \\nabla \\pmb {X}[m,n]$", + "bbox": [ + 500, + 395, + 823, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. See supplemental material for the proof.", + "bbox": [ + 500, + 416, + 816, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1 shows that the wrapped gradients of a wrapped phase image are identical to the gradients of the GT, for the points satisfying the 2D Itoh's continuity condition (5) (called Itoh's condition for brevity in the following text). In general, GT phases exhibit gradual changes in most points, and their rapid changes only occur at a few points sparsely distributed over the image. Then, the outlier points not meeting Itoh's condition are sparse.", + "bbox": [ + 498, + 440, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now consider a noisy wrapped image $\\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N})$ . We can directly extend Eq. (4) to the noisy case:", + "bbox": [ + 498, + 561, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\boldsymbol {Y} [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ] + \\nabla \\boldsymbol {N} [ m, n ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 598, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for any location $[m, n]$ in $(\\mathbf{X} + \\mathbf{N})$ whose adjacent points share the same wrap count. Similarly, we have the following corollary of Proposition 1.", + "bbox": [ + 498, + 621, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Corollary 1.1. For points $[m,n]$ satisfying $\\| \\nabla X[m,n] + \\nabla N[m,n]\\|_{\\infty} < \\pi$ , we have", + "bbox": [ + 500, + 672, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} (\\nabla \\boldsymbol {Y} [ m, n ]) = \\mathcal {W} (\\nabla \\boldsymbol {X} [ m, n ]) + \\nabla \\boldsymbol {N} [ m, n ]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 710, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is clear that both $\\nabla Y$ and $\\mathcal{W}(\\nabla Y)$ can act as a noisy version for $\\nabla X$ , except on some generally-sparse outlier points that do not satisfy the corresponding conditions. If we can mitigate those noise and outliers to predict $\\nabla X$ , the remaining task, reconstructing $X$ from $\\nabla X$ , is straightforward, as $X$ can be reconstructed by integrating its spatial gradients $\\nabla X$ with appropriate boundary conditions. Indeed, the primary concern in PU revolves around relative phases rather than the absolute ones. The observations described by Eq. (7) and Eq. (6) underpin our DUN and unsupervised loss function, respectively.", + "bbox": [ + 498, + 734, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25184", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/abeb0ac61fa480e7f289556dfc002ba42f92838af551153f06b61b11a8c7f8d7.jpg", + "image_caption": [ + "Figure 2. Architecture of our proposed DUN for PU." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 861, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Deep Unrolling Network for PU", + "text_level": 1, + "bbox": [ + 76, + 426, + 357, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our DUN is based on a regularization model inspired by Eq. (7). Though Eq. (7) is inapplicable to the pixels not meeting Itoh's condition, these outliers are typically sparse in a phase image with gradual changes in most areas. Consequently, we introduce a sparse variable $\\mathbf{E} \\in \\mathbb{R}^{M \\times N \\times 2}$ into the regularization model, which counteracts the errors introduced by those sparse outliers. Specifically, we consider the following regularization model of PU:", + "bbox": [ + 75, + 450, + 470, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {X}, \\boldsymbol {E}} \\| \\boldsymbol {\\nabla} \\boldsymbol {X} - \\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) + \\boldsymbol {E} \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\boldsymbol {X}) + \\psi (\\boldsymbol {E}), \\quad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 582, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi, \\psi$ are some regularization functions for the latent phase image $X$ and the error matrix $E$ , respectively. It can be seen that $E$ essentially neutralizes outlier effects, preserving correctness of the data fitting term.", + "bbox": [ + 75, + 616, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our DUN is constructed via the unrolled iterative solver for (8), which reads as follows: for $t = 1,\\dots ,T$", + "bbox": [ + 76, + 676, + 470, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\pmb {X} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {X}} \\| \\pmb {\\nabla} \\pmb {X} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {E} ^ {(t - 1)}) \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\pmb {X}), \\\\ \\pmb {E} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {E}} \\| \\pmb {E} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {\\nabla} \\pmb {X} ^ {(t)}) \\| _ {\\mathrm {F}} ^ {2} + \\psi (\\pmb {E}). \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 717, + 493, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, the DUN predicts $\\mathbf{X}$ from $\\mathcal{W}(\\nabla \\mathbf{Y})$ . It alternatively connects two types of sub-NNs that act as the estimators of $\\mathbf{X}^{(t)}$ and $\\mathbf{E}^{(t)}$ , respectively. See Fig. 2 for an illustration of this architecture.", + "bbox": [ + 76, + 762, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sub-NN for estimating $\\boldsymbol{X}^{(t)}$ : The sub-problem for $\\boldsymbol{X}^{(t)}$ is solved using proximal gradient descent (PGD), which combines gradient descent with proximal mapping. Considering the limited number of modules (i.e., iteration number $T$ ) in our DUN for computational efficiency, we implement", + "bbox": [ + 76, + 823, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multiple gradient descent steps within a stage to bridge the gap. Specifically, Nesterov's acceleration scheme is used to enhance the gradient descent, called AGD. We initialize with $\\mathbf{X}_{(0)}^{(t)} = \\mathbf{X}^{(t - 1)}, \\mathbf{X}^{(0)} = \\mathbf{1}, \\mathbf{E}^{(0)} = \\mathbf{0}$ and $\\alpha_{(0)} = 0$ . For $j = 1,\\dots ,J$ ,", + "bbox": [ + 496, + 428, + 892, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\boldsymbol {X} _ {(j)} ^ {(t)} = \\boldsymbol {V} _ {(j - 1)} ^ {(t)} + \\lambda^ {(t)} \\mathrm {d i v} \\big (\\boldsymbol {\\nabla} \\boldsymbol {V} _ {(j - 1)} ^ {(t)} - (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) - \\boldsymbol {E} ^ {(t - 1)}) \\big), \\\\ \\alpha_ {(j)} = \\frac {1}{2} (1 + \\sqrt {1 + 4 \\alpha_ {(j - 1)} ^ {2}}), \\\\ \\boldsymbol {V} _ {(j)} ^ {(t)} = \\boldsymbol {X} _ {(j)} ^ {(t)} + \\frac {\\alpha_ {(j - 1)} - 1}{\\alpha_ {(j)}} (\\boldsymbol {X} _ {(j)} ^ {(t)} - \\boldsymbol {X} _ {(j - 1)} ^ {(t)}). \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 513, + 924, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here the term $\\mathrm{div}\\bigl (\\pmb {\\nabla}V_{(j - 1)}^{(t)} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y}) - \\pmb{E}^{(t - 1)})\\bigr)$ corresponds to the gradient $\\mathcal{G}_t(\\pmb {X}) = \\pmb {\\nabla}_\\pmb {X}\\| \\pmb {\\nabla}\\pmb {X} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y})-$ $E^{(t - 1)})\\|_{\\mathrm{F}}^2$ . Afterward, the proximal operator, defined as $\\mathrm{Prox}_f(\\bar{Z}):= \\mathrm{argmin}_{\\pmb{Z}}\\| \\pmb {Z} - \\bar{\\pmb{Z}}\\| _2^2 +f(\\pmb {Z})$ , is applied:", + "bbox": [ + 498, + 593, + 892, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {(t)} = \\operatorname {P r o x} _ {\\phi} \\left(\\boldsymbol {X} _ {(J)} ^ {(t)}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 669, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This proximal operator is implemented using a sub-NN with an attentive U-shape structure. As illustrated in Fig. 2, it has a $16 \\times$ down-scaling encoder and a corresponding up-scaling decoder, connected by skip connections. At the coarsest scale, a standard self-attention block is used for exploiting global features. The sub-NN also takes the gradient term $\\mathcal{G}_t(\\boldsymbol{X}_{(J)}^{(t)})$ as an additional input via concatenation. A long skip connection from the input to output is added, enhancing feature flow with a weighted summation.", + "bbox": [ + 496, + 699, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sub-NN for estimating $E^{(t)}$ : The sub-problem for $E^{(t)}$ aims to address the outliers presented in the data fitting term defined on $\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}$ , guided by the prior encoded in $\\psi (E)$ . We employ a sub-NN with $\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}$", + "bbox": [ + 496, + 838, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as the input to implement the solver of this sub-problem. This sub-NN, illustrated in Fig. 2, is composed of sequential convolutional layers, each followed by a parametric rectified linear unit (PReLU). The sparsity prior of $\\pmb{E}$ is induced by attaching a soft thresholding activation function, parametrized by a non-negative parameter $\\mathrm{ReLU}(d^{(t)})$ .", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condition-aware parameter estimation: The noise level in the wrapped phases impacts the optimal selection of the parameters $\\{\\lambda^{(t)}, w^{(t)}, d^{(t)}\\}_{t=1}^T$ . Existing DUNs (e.g. [51]) often include imaging condition values in the input. In our DUN, a condition-aware module (CAM) is used to predict these parameters from noise level, using a chain of fully-connected layers with PReLU or Sigmoid activations.", + "bbox": [ + 75, + 184, + 467, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Unsupervised End-to-End Training Scheme", + "text_level": 1, + "bbox": [ + 76, + 297, + 447, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{F}$ denote our DUN that maps $\\mathcal{W}(\\nabla Y)$ to $X$ . Our proposed loss function $\\mathcal{L}_{\\mathrm{total}}$ for the unsupervised training of $\\mathcal{F}$ comprises a noise-resistant self-reconstruction loss $\\mathcal{L}_{\\mathrm{sr}}$ and a self-distillation loss $\\mathcal{L}_{\\mathrm{sd}}$ , applied to each stage of the DUN:", + "bbox": [ + 75, + 321, + 467, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {s r}} + \\eta \\mathcal {L} _ {\\mathrm {s d}}, \\eta \\in \\mathbb {R} ^ {+}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 392, + 467, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noise-resistant self-reconstruction loss: This loss utilizes $\\nabla Y$ as a noisy label for $\\nabla X$ , as shown in Eq. (6). To mitigate the impact of label noise, the loss is defined as", + "bbox": [ + 76, + 421, + 467, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) - (\\nabla \\boldsymbol {Y} - \\nabla \\boldsymbol {U}) \\| _ {\\mathrm {F}} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 476, + 467, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U$ is random noise drawn from the distribution of $N$ . The introduction of the noise pair $\\{+\\nabla U, -\\nabla U\\}$ is to achieve the resilience to measurement noise in $\\nabla Y$ in the training loss, as justified in Proposition 2.", + "bbox": [ + 76, + 503, + 467, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. Let $\\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N})$ . Suppose Eq. (6) is satisfied at all points. Assume that $\\mathbf{N}, \\mathbf{U} \\sim \\mathcal{P}$ are independent. Then, we have that", + "bbox": [ + 75, + 571, + 467, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {Y}} \\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {X}, \\boldsymbol {N}, \\boldsymbol {U}} \\| \\boldsymbol {\\nabla} \\mathcal {F} (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})) - \\boldsymbol {\\nabla} \\boldsymbol {X} \\| _ {F} ^ {2} + C _ {0}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 627, + 467, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_0$ is a constant.", + "bbox": [ + 76, + 659, + 233, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. See the proof in supplemental material.", + "bbox": [ + 76, + 683, + 383, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2 shows that if one can independently draw instances $\\pmb{U}$ from the same distribution as $N$ , the loss $\\mathcal{L}$ effectively simulates the loss supervised by $\\nabla X$ , nullifying the label noise as shown in Eq. (12). Note that Proposition 2 requires all points satisfying Eq. (6), which is not generally true. On the outliers, $\\nabla Y$ provides wrong supervision with superfluous multiples of $2\\pi$ , as seen in Eq. (3). To counteract the impact of those outlier points, we exploit the property $\\mathcal{W}(\\theta + 2k\\pi) = \\theta$ to modify the self-reconstruction loss as follows:", + "bbox": [ + 75, + 708, + 467, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s r}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\mathcal {W} \\left[ \\boldsymbol {\\nabla} \\mathcal {F} \\left(\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})\\right) - (\\boldsymbol {\\nabla} \\boldsymbol {Y} - \\boldsymbol {\\nabla} \\boldsymbol {U}) \\right] \\| _ {\\mathrm {F}} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 867, + 478, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "That is, an extra $\\mathcal{W}$ is applied to the prediction residual within the Frobenius norm, which negates the superfluous multiples of $2\\pi$ during the loss calculation.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-distillation loss: Although $\\mathcal{L}_{\\mathrm{sr}}$ reduces the impact of measurement noise, the ambiguity in solutions due to wrapping may remain, particularly at those outlier points. Fortunately, CNNs including unrolling CNNs [30, 39], show strong induced bias towards regular structures, which aids in reducing such ambiguity during the training with $\\mathcal{L}_{\\mathrm{sr}}$ .", + "bbox": [ + 496, + 138, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further reduce the ambiguity, we introduce a self-distillation loss defined as", + "bbox": [ + 498, + 229, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s d}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y})) - \\nabla \\bar {\\mathcal {F}} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) \\| _ {\\mathrm {F}} ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 266, + 890, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\mathcal{F}}$ denotes the NN detached from the previous iteration with stop gradient. The loss leverages pseudo supervision from the NN's varied predictions with noise-altered inputs, implicitly aggregating those predictions for self-training. This self-distillation scheme can reduce the NN's prediction variance, enhancing the PU accuracy. Moreover, recall that $\\mathcal{L}_{\\mathrm{sr}}$ inputs $\\mathcal{W}(\\nabla Y + \\nabla U)$ to $\\mathcal{F}$ , which does not fully align with $\\mathcal{W}(\\nabla Y)$ in test. This inconsistency can be reconciled by $\\mathcal{L}_{\\mathrm{sd}}$ , as it uses $\\mathcal{W}(\\nabla Y)$ for the input of $\\mathcal{F}$ .", + "bbox": [ + 496, + 297, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 446, + 630, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance evaluation is conducted on both simulated and real-world phase patterns. Following [22], we use NRMSE (Normalized Root Mean Squared Error) as the performance metric. As relative phase is the main focus in PU, the mean of unwrapped phase values is first aligned to that of the GT before the NRMSE calculation. Computational complexity is measured by three metrics: the number of parameters, the number of FLOPs (floating point of operations), and the average test time per sample of size $256 \\times 256$ .", + "bbox": [ + 496, + 472, + 890, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details: Our proposed approach, called U3Net (U3 for Unsupervised, Unrolling and Unwrapping) for convenience, is implemented in Python and run on a single NVIDIA RTX 3080 GPU. We set $\\eta = 0.5$ in $\\mathcal{L}_{\\mathrm{total}}$ and initialize model weights via kaiming Uniform. The training employs the Adam optimizer with 500 epochs and a batch size of 10. The learning rate is initialized as $1\\times 10^{-3}$ and exponentially decayed by a factor of 0.99 at every epoch.", + "bbox": [ + 496, + 611, + 890, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Methods for comparison: We select eight representative methods of general PU as baselines, including LS [8], QGPU [11], Ryu. et al. [32], PhaseNet2.0 [35], PUDIP [47], SQD-LSTM [22], EESANet [49], and TriNet [37]. Additionally, two representative transformer NNs for general image restoration, UFormer [43] and Restormer [48], are also included for performance comparison. In all the baseline methods, LS and QGPU are conventional non-learning iterative methods, built upon least squares and path selection, respectively; PUDIP is an internal DL-based method; and the others are supervised DL-", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25186", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c7e27bf5099fe744de7f3ed2b8e2968b77b82b544a70a1b1c14cd59d70843531.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMoGRRME#Param. (M)#FLOPs (G)Time (msec.)
SNR(dB)0510203005102030
Non-LearningLS7.912.281.200.370.157.852.501.250.400.12--8.39
QGPU17.122.121.150.360.1117.262.291.200.390.12--14.94
SupervisedRyu. et al.1.341.111.091.051.051.511.051.010.980.931.0721.85303.96
PhaseNet2.08.358.198.078.068.049.298.537.536.956.861.1511.9320.87
SQD-LSTM0.870.710.700.690.681.571.131.121.101.090.904.0713.04
EESANet0.780.770.760.760.741.311.311.251.241.0661.6875.279.85
TriNet5.655.555.465.405.345.405.335.215.105.0513.6165.4811.16
UFormer0.500.470.470.460.450.580.510.490.490.4820.6040.9842.00
Restormer0.450.380.360.360.350.500.430.420.410.413.0217.2344.69
UnsupervisedPUDIP17.5315.167.870.340.1113.107.222.620.380.122.3321.5899695.01
U3Net0.690.250.190.160.101.120.380.270.170.120.748.7710.28
", + "bbox": [ + 109, + 88, + 861, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison on synthetic datasets in terms of NRMSE(%) . Bold for Best and underline for second-best.", + "bbox": [ + 112, + 289, + 854, + 304 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/913f40f396c5ace88c33506d39f066d69e0c7e06faf6b07a14c1edefdca222ff.jpg", + "image_caption": [ + "Figure 3. Visualization of PU results on MoGR (top) and RME (bottom), showing residual images to facilitate visual comparison." + ], + "image_footnote": [], + "bbox": [ + 86, + 316, + 888, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based methods with their NNs retrained on our generated paired phase data. Specifically, as UFormer and Restormer are not originally designed for PU, we train their models using the supervised loss of SQD-LSTM [22] for better performance.", + "bbox": [ + 75, + 547, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation on Simulated Phase Patterns", + "text_level": 1, + "bbox": [ + 76, + 633, + 421, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data preparation: Two types of simulated data are generated for training and evaluation. The first type follows [22], where a GT phase image of size $256 \\times 256$ is generated through the superposition of a mixture of Gaussians (MoG) with $c$ clusters and randomly selected ramps. The phase range is defined within $[-2p\\pi, 2p\\pi]$ . The resulting dataset is called MoGR (MoG with Ramp). The noisy wrapped phase images are synthesized via Eq. (1), with $N$ defined as Gaussian white noise. For each GT, only one noisy wrapped phase image is generated. With $c$ and $p$ uniformly selected from $\\{1, \\dots, 4\\}$ and $\\{1, \\dots, 7\\}$ respectively, we generate 5000 noisy wrapped phase images with SNR uniformly selected from $\\{0, 5, 10, 20, 30, 60\\}$ to construct the training set. Using the same way, 1000 noisy wrapped phase images are generated for test with SNR in $\\{0, 5, 10, 20, 30\\}$ .", + "bbox": [ + 75, + 657, + 468, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The other dataset follows a different construction scheme", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "named random matrix enlargement [40], thus called RME dataset. Concretely, a GT phase image is formed through enlarging the square matrix of a distribution from an initial size $s \\times s$ to size $256 \\times 256$ via interpolation. The size parameter $s$ is selected uniformly from $\\{2, \\dots, 10\\}$ , the distribution type is chosen from the {Uniform, Normal} distributions, and the interpolation method is chosen from the {Bilinear, Bicubic} interpolations. Then the phase image is re-scaled to the range varying in $[-2p\\pi, 2p\\pi]$ with $p$ randomly chosen from $\\{1, \\dots, 7\\}$ . The noisy wrapped phase images are obtained by the same way as the MoGR dataset.", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and analysis: Table 1 lists the quantitative results on both datasets, including those in computational complexity. Even trained without GT data, U3Net achieves the best results in 8/10 settings, using a lightweight model. Concretely, U3Net outperforms the two conventional non-learning-based methods, LS and QGPU. In comparison to PUDIP, an unsupervised internal DL-based method, U3Net also shows significant performance gain. The performance advantage of U3Net over these three GT-free methods be", + "bbox": [ + 496, + 719, + 892, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ PUDIP involves many forward passes for test-time learning. We report its #FLOPs in one forward pass and running time of the whole process.", + "bbox": [ + 500, + 875, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/60156f64c4db38f79bf98e996d9115531d42cff561d39a647c4f5feeee597555.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SNR(dB)LSQGPURyu. et al.PhaseNet2.0SQD-LSTMEESANetTriNetPU-GANPUNetUFormerRestormerPUDIPU3Net
53.313.131.412.281.762.455.0413.739.591.461.069.991.00
101.841.961.271.691.521.994.6911.849.201.280.935.300.82
200.941.131.241.431.481.794.4611.629.010.970.910.470.46
", + "bbox": [ + 83, + 89, + 885, + 154 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance comparison on InSAR phase data in terms of NRMSE(%) . Bold for Best and underline for second-best.", + "bbox": [ + 109, + 162, + 854, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/397a852b44ad372cb3af2a5dc71bdf6a42355fa060ed9de63d0fdbf20df78d5a.jpg", + "image_caption": [ + "Figure 4. Visualization of unwrapped phase images on InSAR data, showing residual images to facilitate visual comparison." + ], + "image_footnote": [], + "bbox": [ + 155, + 191, + 816, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "comes more noticeable for lower SNR (i.e. heavier noise), particularly compared to PUDIP. This is because U3Net has an explicit mechanism to deal with noise during training, thereby achieving higher noise robustness. In addition, U3Net is much faster than PUDIP, due to that U3Net employs end-to-end pre-trained models, avoiding the time-costly per-sample training of PUDIP.", + "bbox": [ + 75, + 426, + 467, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Surprisingly, U3Net also surpasses the supervised DL-based methods in many settings. In the lowest-SNR case, U3Net performs worse than UFormer and Restormer, two supervised transformer models. This is probably due to that severe noise increases the number of outlier points, lowering the effectiveness of the training function. Even that, U3Net still outperforms other supervised methods. The model of U3Net is the smallest among all the DL-based methods, indicating that the excellent performance of U3Net is mainly attributed to its interpretable physics-encoded NN architecture and its effective training scheme, rather than the use of larger models. Meanwhile, U3Net is computationally efficient, with the second smallest number of FLOPs and the third fastest inference speed.", + "bbox": [ + 75, + 532, + 467, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "See Fig. 3 for the visualization of PU results from different methods. The residuals to GTs produced by U3Net are much more insignificant compared to the internal DL-based PUDIP. In comparison to supervised methods, U3Net outperforms all of them on the image from RME dataset and shows comparable results on the image of MoGR dataset.", + "bbox": [ + 75, + 744, + 467, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation on InSAR Phase Patterns", + "text_level": 1, + "bbox": [ + 76, + 845, + 395, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The InSAR phase data are generated by converting the digital elevation model at Texas, U.S. into absolute phases,", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "following [33]. Totally, 4000 (1000) GT phase images are obtained by cropping the $256 \\times 256$ patches from the absolute phases for training (test). Each phase value is scaled into $[-6\\pi, 6\\pi]$ . The noisy wrapped phases are generated by adding Gaussian noise following the experiments in Sec. 4.1, and we test on three noise levels: 5, 10, 20. In addition to the previously-compared general PU methods and general-purpose transformer NNs, we also compare with two specific methods designed for InSAR PU: PUNet [44] and PU-GAN [57]. All the compared models are retrained using the same data as ours.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "See Table 2 for the quantitative comparison on this In-SAR dataset. Our U3Net ranks the first among all compared methods over all noise settings, noticeably surpassing other GT-free methods as well as many supervised methods in many cases. The visual comparison provided in Fig. 4 also shows the advantage of U3Net. All these results further verify the effectiveness of our approach in dealing with phase patterns from real scenes.", + "bbox": [ + 496, + 593, + 890, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 723, + 663, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We form some baselines from U3Net for ablation studies, which are retrained and tested on the MoGR dataset. Their results are summarized in Table 3.", + "bbox": [ + 496, + 746, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Study on loss function: We design the following schemes to analyze the loss function: (a) $\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}$ : remove the outer $\\mathcal{W}$ applied to the residual, from the self-reconstruction loss $\\mathcal{L}_{\\mathrm{sr}}$ , i.e., replacing $\\mathcal{L}_{\\mathrm{sr}}$ by the loss $\\mathcal{L}$ defined in Eq. (11); (b) w/o $\\nabla U$ : remove the noise re-corruption mechanism in $\\mathcal{L}_{\\mathrm{sr}}$ by discarding the $\\pm \\nabla U$ terms; (c) $\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}$ w/o $\\nabla U$ : combine the former cases, yielding a plain gradient loss", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25188", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3386b00feb973f36340c31d4fc46e6f562717d8855e7a1f5d122be5eb8e7d4bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingSNR=0SNR=5SNR=10SNR=20SNR=30
Lsr→L25.4523.4823.2623.2323.20
w/o ∇U6.162.401.360.690.60
Lsr→L w/o ∇U25.8523.5723.4223.3823.34
w/o Lsd1.140.330.230.220.17
Supervised0.290.160.130.110.05
w/o E7.160.750.440.360.31
AGD→GD2.140.350.210.170.12
w/o CAM0.980.310.220.210.15
W(∇Y)→∇Y2.881.741.501.461.39
DUN→UFormer1.551.411.311.301.30
U3Net (original)0.690.250.190.160.10
", + "bbox": [ + 80, + 88, + 472, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results in ablation studies in terms of NRMSE(%)", + "bbox": [ + 94, + 258, + 450, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/595e81c47fec259ebc2b19f954dc73e1904385058746d70dd9a5601d4749c892.jpg", + "image_caption": [ + "Figure 5. Visualization of unwrapped phase images using models trained by different loss functions. $\\mathcal{L}_{\\mathrm{sr}} \\rightarrow \\mathcal{L}$ and w/o $\\nabla U$ produce noticeable errors than Original. Zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 91, + 290, + 455, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla Y\\|_{\\mathrm{F}}^2$ (d) w/o $\\mathcal{L}_{\\mathrm{sd}}$ : remove the selfdistillation loss $\\mathcal{L}_{\\mathrm{sd}}$ , only employing $\\mathcal{L}_{\\mathrm{sr}}$ for training; and (e) Supervised: train with a gradient-domain supervised loss $\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla X\\|_{\\mathrm{F}}^2$", + "bbox": [ + 76, + 530, + 468, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to Table 3, we also provide some visual comparisons in Fig. 5 for demonstration. From both the quantitative and qualitative results, we have following observations. (a) The outer $\\mathcal{W}$ in $\\mathcal{L}_{\\mathrm{sr}}$ is critical to the performance of U3Net. (b) The re-corruption mechanism in $\\mathcal{L}_{\\mathrm{sr}}$ does help eliminate the measurement noise. (c) A simple gradient loss function does not work, with further performance decrease. (d) While the self-reconstruction loss already yields promising results, the self-distillation loss benefits further performance gain. (e) U3Net performs closely to its supervised counterpart when $\\mathrm{SNR} \\geq 5$ . We also see that U3Net has its supervised counterpart outperforming the supervised UFormer reported in Table 1, demonstrating its superiority.", + "bbox": [ + 75, + 594, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on DUN: To analyze the components in the DUN architecture of U3Net, we design the following baseline models: (a) w/o $\\pmb{E}$ : abandon the error term $\\pmb{E}$ during unrolling, forming a DUN alternating the accelerated gradient descent step and the sub-NN for estimating $\\pmb{X}^{(t)}$ ; (b) AGD→GD: replace the Nesterov's AGD steps by one-step standard gradient; (c) w/o CAM: remove the CAM from U3Net; (d)", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c81c67944f6bb3c22d8d09aff4f14d9fe2a38190002fd4075c37d807c26cec9.jpg", + "image_caption": [ + "Figure 6. Visualization of error matrices $\\pmb{E}$ estimated from three wrapped phase images. The outlier maps use white pixels to indicate the points $[x,y]$ satisfying $\\| \\nabla X[x,y] + \\nabla N[x,y]\\|_{\\infty} \\geq \\pi$ and black pixels otherwise." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 880, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathcal{W}(\\nabla \\pmb {Y})\\rightarrow \\nabla \\pmb {Y}$ : use $\\nabla Y$ to replace $\\mathcal{W}(\\nabla Y)$ in the input of U3Net as well in the total loss; (e) DUN $\\rightarrow$ UFormer: replace the DUN by UFormer, keeping the loss unchanged.", + "bbox": [ + 498, + 409, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As demonstrated in Table 3, each component in the DUN of U3Net is useful. (a) One critical part is the introduction of the error matrix $\\pmb{E}$ for unrolling. It brings significant performance gain, particularly in dealing with low-SNR data, as the absorption of outliers breaking the Itoh's condition is more desired for heavier noise. See Fig. 6 for a visualization of estimated error matrices, which successfully identify the outliers. (b)&(c) The AGD and CAM have similar contribution to the performance of U3Net. (d) Using $\\mathcal{W}(\\nabla Y)$ for DUN is also crucial, serving as a better choice compared to using $\\nabla Y$ . (e) Our proposed loss function also works for UFormer, but yielding worse results than U3Net, which indicates the higher effectiveness of our physics-encoded DUN for unsupervised PU, compared to general transformer models.", + "bbox": [ + 496, + 455, + 892, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 619, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposed a fully unsupervised DL approach for PU, eliminating the dependence on GTs in end-to-end training. Utilizing the connection between (wrapped) gradients of wrapped phases and gradients of original phases, we constructed a physics-encoded DUN, trained via a noise-resistant self-reconstruction loss complemented by a self-distillation loss. Empirical evaluations confirmed that our approach surpassed existing GT-free methods and rivaled supervised counterparts. Similar to current unsupervised and synthetic data-dependent supervised techniques, our approach presumes known noise statistics. Future research will focus on dealing with unknown noise profiles.", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25189", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Liheng Bian, Xin Wang, Daoyu Li, Qiuling Ren, and Dezhi Zheng. Robust phase unwrapping via non-local regularization. Optics Letters, 48(6):1399-1402, 2023. 1, 2", + "[2] David Blinder, Heidi Ottevaere, Adrian Munteanu, and Peter Schelkens. Efficient multiscale phase unwrapping methodology with modulo wavelet transform. Optics Express, 24(20): 23094-23108, 2016. 1, 2", + "[3] Donald J Bone. Fourier fringe analysis: the two-dimensional phase unwrapping problem. Applied Optics, 30(25):3627-3632, 1991. 1, 2", + "[4] Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate mmse estimation. IEEE Transactions on Signal Processing, 70:2239-2252, 2022. 3", + "[5] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 3", + "[6] Gili Dardikman and Natan T Shaked. Phase unwrapping using residual neural networks. In Computational Optical Sensing and Imaging, pages CW3B-5. Optica Publishing Group, 2018. 3", + "[7] Qiaoqiao Ding, Gaoyu Chen, Xiaoqun Zhang, Qiu Huang, Hui Ji, and Hao Gao. Low-dose ct with deep learning regularization via proximal forward-backward splitting. Physics in Medicine & Biology, 65(12):125009, 2020. 3", + "[8] Dennis C Ghiglia and Louis A Romero. Robust two-dimensional weighted and unweighted phase unwrapping that uses fast transforms and iterative methods. Journal of the Optical Society of America A, 11(1):107-117, 1994. 2, 5", + "[9] Richard M Goldstein, Howard A Zebker, and Charles L Werner. Satellite radar interferometry: Two-dimensional phase unwrapping. Radio Science, 23(4):713-720, 1988. 1, 2", + "[10] Michal Gontarz, Vibekananda Dutta, Malgorzata Kujawinska, and Wojciech Krauze. Phase unwrapping using deep learning in holographic tomography. Optics Express, 31(12):18964-18992, 2023. 1, 3", + "[11] Miguel Arevallilo Herraez, David R Burton, Michael J Lolor, and Munther A Gdeisat. Fast two-dimensional phaseunwrapping algorithm based on sorting by reliability following a noncontinuous path. Applied Optics, 41(35):7437-7444, 2002. 1, 2, 5", + "[12] Howard YH Huang, L Tian, Z Zhang, Y Liu, Z Chen, and G Barbastathis. Path-independent phase unwrapping using phase gradient and total-variation (tv) denoising. Optics Express, 20(13):14075–14089, 2012. 1, 2", + "[13] Kazuyoshi Itoh. Analysis of the phase unwrapping algorithm. Applied Optics, 21(14):2470-2470, 1982. 2, 3", + "[14] Charles VJ Jakowitz, Daniel E Wahl, Paul H Eichel, Dennis C Ghiglia, and Paul A Thompson. Spotlight-mode synthetic aperture radar: a signal processing approach. Springer Science & Business Media, 2012. 1", + "[15] Ulugbek S Kamilov, Ioannis N Papadopoulos, Morteza H Shoreh, Demetri Psaltis, and Michael Unser. Isotropic" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inverse-problem approach for two-dimensional phase unwrapping. Journal of the Optical Society of America A, 32 (6):1092-1100, 2015. 1, 2", + "[16] Thomas Kreis. Handbook of holographic interferometry: optical and digital methods. John Wiley & Sons, 2006. 1", + "[17] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2137, 2019. 3", + "[18] Ji Li, Weixi Wang, Yuesong Nan, and Hui Ji. Self-supervised blind motion deblurring with deep expectation maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13986-13996, 2023. 3", + "[19] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 2", + "[20] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 3", + "[21] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2043-2052, 2021. 2, 3", + "[22] Malsha V Perera and Ashwin De Silva. A joint convolutional and spatial quad-directional LSTM network for phase unwrapping. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4055-4059. IEEE, 2021. 1, 3, 5, 6", + "[23] Mark D Pritt. Phase unwrapping by means of multigrid techniques for interferometric sar. IEEE Transactions on Geoscience and Remote Sensing, 34(3):728-738, 1996. 1, 2", + "[24] Xinran Qin, Yuhui Quan, and Hui Ji. Enhanced deep unrolling networks for snapshot compressive hyperspectral imaging. Neural Networks, 174:106250, 2024. 3", + "[25] Yi Qin, Shujia Wan, Yuhong Wan, Jiawen Weng, Wei Liu, and Qiong Gong. Direct and accurate phase unwrapping with deep neural network. Applied Optics, 59(24):7258-7267, 2020. 3", + "[26] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1898, 2020. 3", + "[27] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In European Conference on Computer Vision, pages 642-659. Springer, 2022. 3", + "[28] Yuhui Quan, Xinran Qin, Tongyao Pang, and Hui Ji. Dual-domain self-supervised learning and model adaption for deep compressive imaging. In European Conference on Computer Vision, pages 409-426. Springer, 2022. 3", + "[29] Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25190", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 3", + "[30] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5", + "[31] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3341-3350, 2020. 3", + "[32] Kanghyun Ryu, Sung-Min Gho, Yoonho Nam, Kevin Koch, and Dong-Hyun Kim. Development of a deep learning method for phase unwrapping mr images. In Proceedings of the International Society for Magnetic Resonance in Medicine, page 4707, 2019. 3, 5", + "[33] Francescopaolo Sica, Francesco Calvanese, Giuseppe Scarpa, and Paola Rizzoli. A cnn-based coherence-driven approach for insar phase unwrapping. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2020. 7", + "[34] GE Spoorthi, Subrahmanyam Gorthi, and Rama Krishna Sai Subrahmanyam Gorthi. Phasenet: A deep convolutional neural network for two-dimensional phase unwrapping. IEEE Signal Processing Letters, 26(1):54-58, 2018. 3", + "[35] GE Spoorthi, Rama Krishna Sai Subrahmanyam Gorthi, and Subrahmanyam Gorthi. Phasenet 2.0: Phase unwrapping of noisy data based on deep learning approach. IEEE Transactions on Image Processing, 29:4862-4872, 2020. 3, 5", + "[36] Xianyu Su and Wenjing Chen. Reliability-guided phase unwrapping algorithm: a review. Optics and Lasers in Engineering, 42(3):245-261, 2004. 1, 2", + "[37] Krishna Sumanth, Vaishnavi Ravi, and Rama Krishna Gorthi. A multi-task learning for 2d phase unwrapping in fringe projection. IEEE Signal Processing Letters, 29:797-801, 2022. 5", + "[38] Julián Tachella, Dongdong Chen, and Mike Davies. Unsupervised learning from incomplete measurements for inverse problems. Advances in Neural Information Processing Systems, 35:4983-4995, 2022. 3", + "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 3, 5", + "[40] Kaiqiang Wang, Ying Li, Qian Kemao, Jianglei Di, and Jianlin Zhao. One-step robust deep learning phase unwrapping. Optics Express, 27(10):15100-15115, 2019. 3, 6", + "[41] Kaiqiang Wang, Qian Kemao, Jianglei Di, and Jianlin Zhao. Deep learning spatial phase unwrapping: a comparative review. Advanced Photonics Nexus, 1(1):014001, 2022. 1, 2", + "[42] Weixi Wang, Ji Li, and Hui Ji. Self-supervised deep image restoration via adaptive stochastic gradient Langevin dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2022. 3", + "[43] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 5", + "[44] Zhipeng Wu, Teng Wang, Yingjie Wang, Robert Wang, and Daqing Ge. Deep learning for the detection and phase unwrapping of mining-induced deformation in large-scale interferograms. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2021. 7", + "[45] Xianming Xie. Iterated unscented kalman filter for phase unwrapping of interferometric fringes. Optics Express, 24 (17):18872-18897, 2016. 1, 2", + "[46] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 2", + "[47] Fangshu Yang, Thanh-An Pham, Nathalie Brandenberg, Matthias P Lütolf, Jianwei Ma, and Michael Unser. Robust phase unwrapping via deep image prior for quantitative phase imaging. IEEE Transactions on Image Processing, 30: 7025-7037, 2021. 1, 2, 3, 5", + "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5", + "[49] Junkang Zhang and Qingguang Li. Eesanet: edge-enhanced self-attention network for two-dimensional phase unwrapping. Optics Express, 30(7):10470-10490, 2022. 3, 5", + "[50] Junchao Zhang, Xiaobo Tian, Jianbo Shao, Haibo Luo, and Rongguang Liang. Phase unwrapping in optical metrology via denoised and convolutional segmentation networks. Optics Express, 27(10):14903-14912, 2019. 3", + "[51] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3217-3226, 2020. 3, 5", + "[52] Teng Zhang, Shaowei Jiang, Zixin Zhao, Krishna Dixit, Xiaofei Zhou, Jia Hou, Yongbing Zhang, and Chenggang Yan. Rapid and robust two-dimensional phase unwrapping via deep learning. Optics Express, 27(16):23173-23185, 2019. 3", + "[53] Zhonghao Zhang, Yipeng Liu, Jiani Liu, Fei Wen, and Ce Zhu. Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing, 30:1487-1500, 2020. 3", + "[54] Ming Zhao, Lei Huang, Qican Zhang, Xianyu Su, Anand Asundi, and Qian Kemo. Quality-guided phase unwrapping technique: comparison of quality maps and guiding strategies. Applied Optics, 50(33):6214-6224, 2011. 2", + "[55] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 630-641, 2021. 3", + "[56] Hongyu Zhou, Chuanli Cheng, Hao Peng, Dong Liang, Xin Liu, Hairong Zheng, and Chao Zou. The phu-net: A robust phase unwrapping method for mri based on deep learning." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25191", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Magnetic Resonance in Medicine, 86(6):3321-3333, 2021. 1, 3", + "[57] Lifan Zhou, Hanwen Yu, Vito Pascazio, and Mengdao Xing. Pu-gan: a one-step 2-d insar phase unwrapping based on conditional generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 60:1-10, 2022. 1, 3, 7" + ], + "bbox": [ + 78, + 90, + 467, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25192", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_model.json b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b12245434303141f53f61ee237056d477d9d0b4b --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Unsupervised Deep Unrolling Networks for Phase Unwrapping" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.18, + 0.682, + 0.199 + ], + "angle": 0, + "content": "Zhile Chen\\(^{1,2}\\) Yuhui Quan\\(^{1,2*}\\) Hui Ji\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.201, + 0.862, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China \\(^{2}\\)Pazhou Lab, Guangzhou 510320, China" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.235, + 0.761, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Mathematics, National University of Singapore, 119076, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.255, + 0.786, + 0.27 + ], + "angle": 0, + "content": "cs.zhilechen@mail.scut.edu.cn, csyhquan@scut.edu.cn, matjh@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.654 + ], + "angle": 0, + "content": "Phase unwrapping (PU) is a technique to reconstruct original phase images from their noisy wrapped counterparts, finding many applications in scientific imaging. Although supervised learning has shown promise in PU, its utility is limited in ground-truth (GT) scarce scenarios. This paper presents an unsupervised learning approach that eliminates the need for GTs during end-to-end training. Our approach leverages the insight that both the gradients and wrapped gradients of wrapped phases serve as noisy labels for GT phase gradients, along with sparse outliers induced by the wrapping operation. A recorruption-based self-reconstruction loss in the gradient domain is proposed to mitigate the adverse effects of label noise, complemented with a self-distillation loss for improved generalization. Additionally, by unfolding a variational model of PU that utilizes wrapped gradients of wrapped phases for its data-fitting term, we develop a deep unrolling network that encodes physics of phase wrapping and incorporates special treatments on outliers. In the experiments on three types of phase data, our approach outperforms existing GT-free methods and competes well against the supervised ones." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.208, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.797 + ], + "angle": 0, + "content": "In many imaging systems, the direct acquisition of original phases of a target signal is both challenging and costly. Typically, initial measurement yields a phase image wrapped in \\([- \\pi, \\pi)\\). Let \\( X \\in \\mathbb{R}^{M \\times N} \\) and \\( Y \\in [-\\pi, \\pi)^{M \\times N} \\) denote the true phase image and its noisy wrapped counterpart, respectively. The wrapping process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.803, + 0.469, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} = \\mathcal {W} (\\boldsymbol {X} + \\boldsymbol {N}), \\quad \\mathcal {W}: \\mathbb {R} ^ {M \\times N} \\rightarrow [ - \\pi , \\pi) ^ {M \\times N}, \\tag {1}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.828, + 0.294, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author: Yuhui Quan." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 62372186), Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755, 2023A1515012841), Fundamental Research Funds for Central Universities (Grant No. x2jsD2230220), and Singapore MOE AcRF Tier 1 (Grant No. A-8000981-00-00)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.828, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.893, + 0.427 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) denotes the wrapping operator with its entry-wise operation defined as \\(\\mathcal{W}(\\theta) = ((\\theta + \\pi) \\bmod 2\\pi) - \\pi\\) for \\(\\theta \\in \\mathbb{R}\\), and \\(N \\in \\mathbb{R}^{M \\times N}\\) denotes the measurement noise, typically assumed to be Gaussian [22]. Consequently, the original phase image, which exhibits strong smoothness, is degraded to a wrapped version with both incorrect values and artificial discontinuities induced by the modulo operation, which adversely affects subsequent processing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.893, + 0.578 + ], + "angle": 0, + "content": "Phase unwrapping (PU) is about recovering the original unwrapped phase image \\( \\mathbf{X} \\) (up to an additive constant) from its noisy wrapped counterpart \\( \\mathbf{Y} \\). Serving as a critical step in many optical imaging techniques for generating clear and coherent phase images, PU finds extensive applications in diverse domains such as quantitative phase imaging, magnetic resonance imaging, synthetic aperture radar interferometry, 3D depth sensing, phase contrast microscopy, fringe projection, and digital holographic interferometry; see e.g. [10, 14, 16, 23, 41, 47, 56, 57]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.578, + 0.866, + 0.593 + ], + "angle": 0, + "content": "The model (1) is also usually formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.6, + 0.892, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} = \\boldsymbol {X} - 2 \\pi \\boldsymbol {K} + \\boldsymbol {N} ^ {\\prime}, \\quad \\boldsymbol {K} \\in \\mathbb {Z} ^ {M \\times N}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.793 + ], + "angle": 0, + "content": "where \\( K \\in \\mathbb{Z}^{M \\times N} \\) denotes the map of wrap counts, indicating the number of times a phase value has been wrapped around by \\( 2\\pi \\), and \\( N' \\) denotes the noise dependent on both \\( X \\) and \\( N \\), whose distribution can be complex even when \\( N \\) follows a simple distribution such as the normal distribution. It can be seen that PU is a challenging ill-posed inverse problem involving a continuous variable \\( X \\) and a discrete variable \\( K \\). Direct solutions by integrating wrapped phase differences are insufficient for ensuring spatial consistency in all directions of the unwrapped phases and may accumulate significant errors due to substantial noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Limitations of existing works: Conventional approaches for PU include reliability-guided methods (e.g. [3, 9, 11, 23, 36]), filtering-based methods (e.g. [2, 45]), and model-driven methods (e.g. [1, 12, 15]). These approaches are either sensitive to measurement noise, due to their heuristic rules, or overly simplistic for complex phase structures, due to their handcrafted image priors. In contrast, deep learning" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.206, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.089, + 0.296, + 0.107 + ], + "angle": 0, + "content": "Randomly generated Set" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.108, + 0.315, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.09, + 0.444, + 0.1 + ], + "angle": 0, + "content": "Wrapped Phase Images" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.101, + 0.442, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.179, + 0.202, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.273, + 0.201, + 0.283 + ], + "angle": 0, + "content": "Supervised Learning" + }, + { + "type": "image_footnote", + "bbox": [ + 0.1, + 0.29, + 0.208, + 0.3 + ], + "angle": 0, + "content": "- GT collection is hard" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.179, + 0.322, + 0.198 + ], + "angle": 0, + "content": "Wrappered Phase Image (Single Test Sample)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.199, + 0.314, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.273, + 0.315, + 0.283 + ], + "angle": 0, + "content": "Internal Learning" + }, + { + "type": "image_footnote", + "bbox": [ + 0.227, + 0.29, + 0.319, + 0.301 + ], + "angle": 0, + "content": "Time consuming" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.179, + 0.444, + 0.189 + ], + "angle": 0, + "content": "Wrapped Phase Images" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.19, + 0.441, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.267, + 0.423, + 0.276 + ], + "angle": 0, + "content": "End-to-End" + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.278, + 0.451, + 0.288 + ], + "angle": 0, + "content": "Unsupervised Learning" + }, + { + "type": "image_footnote", + "bbox": [ + 0.34, + 0.29, + 0.451, + 0.3 + ], + "angle": 0, + "content": "Bypassing both issues" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.312, + 0.468, + 0.326 + ], + "angle": 0, + "content": "Figure 1. Training data setting of existing PU methods versus ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.468, + 0.399 + ], + "angle": 0, + "content": "(DL) with neural networks (NNs) has emerged as a powerful tool for PU [41]. A pre-trained end-to-end NN offers data adaptivity, rapid inference, and excellent expressivity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.468, + 0.58 + ], + "angle": 0, + "content": "Existing DL-based PU methods often rely on supervised learning with ground-truth (GT) phases. However, acquiring GT phases is often impractical due to physical limits in imaging, destructive imaging techniques, and ethical/safety concerns. This challenge limits the applicability of supervised DL in many practical scenarios. Internal unsupervised DL-based methods (e.g. [47]) use an untrained NN to parameterize latent phases, adjusting it to fit observed measurements with certain regularization. Though flexible, this kind of methods requires time-costly per-sample test-time training, resulting in poor computational efficiency compared to supervised DL-based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Our approach: To bypass the difficulty of GT acquisition in supervised DL and avoid the computational inefficiency of internal unsupervised DL, we present an external unsupervised \\(DL\\) approach for PU. It trains an end-to-end NN solely on a dataset of noisy wrapped measurements. See Fig. 1 for the comparison in training data configuration between existing works and ours. Indeed, our study is the first one on external (end-to-end) unsupervised DL for PU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.779 + ], + "angle": 0, + "content": "Our approach is inspired by the insight that applying the gradient operator or the wrapping gradient operator to a wrapped phase image will result in an output identical to the gradient map of the GT phase image, with the exception of a sparse outlier set. Concretely, we have that" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.78, + 0.468, + 0.81 + ], + "angle": 0, + "content": "(a) \\(\\nabla Y[m,n] = \\nabla X[m,n] + \\nabla N[m,n]\\) holds for the points whose neighbors share the same wrap count." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.81, + 0.468, + 0.856 + ], + "angle": 0, + "content": "(b) \\(\\mathcal{W}(\\nabla Y[m,n]) = \\nabla X[m,n] + \\nabla N[m,n]\\) is applicable for the points that meet the 2D Itoh's continuity condition [13]." + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.78, + 0.468, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The majority of points in a GT phase image fall under the aforementioned categories. Therefore, on these points, \\(\\nabla Y\\) or \\(\\mathcal{W}(\\nabla Y)\\) can serve as a noisy label of \\(\\nabla X\\). Once \\(\\nabla X\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.834, + 0.107 + ], + "angle": 0, + "content": "is predicted, reconstructing \\(X\\) is relatively simple." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Our approach utilizes \\(\\nabla Y\\) for unsupervised training and tackles its label noise via a recorruption-based self-reconstruction loss function in the gradient domain, which is inspired by existing recorruption-based denoisers [19, 21, 46]. In addition, to mitigate the potential ambiguity and overfitting caused by the absence of GTs, a self-distillation loss function is proposed that leverages intermediate predictions from the NN for pseudo-supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.412 + ], + "angle": 0, + "content": "To encode physical model/priors of phase wrapping into the NN-learned PU process, which is beneficial for mitigating possible overfitting particularly when training on a limited amount of data, we develop a deep unrolling network (DUN) by unfolding a variational model that leverages \\(\\mathcal{W}(\\nabla Y)\\) for data fitting. To our knowledge, this is the first work to utilize deep unrolling for PU. Indeed, most existing studies rely on U-shaped convolutional NNs (CNNs) or recurrent NNs (RNNs) that ignore physical priors of phase wrapping. Additionally, our DUN is tailored to handle the outliers that do not satisfy the Itoh's continuity condition, thereby mitigating their potential adverse effects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.415, + 0.822, + 0.429 + ], + "angle": 0, + "content": "Main contributions: Our contributions include" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.432, + 0.892, + 0.492 + ], + "angle": 0, + "content": "- We introduce the first external unsupervised DL approach for end-to-end PU, circumventing the challenge of GT acquisition in supervised DL and the computational inefficiency of internal unsupervised DL." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.492, + 0.892, + 0.567 + ], + "angle": 0, + "content": "- We perform the first exploration of deep unfolding for PU, founded on a variational model that utilizes wrapped gradients of wrapped phases for data fitting under Itoh's continuity condition. The resulting DUN is outlier-aware and enjoys advantages in interpretability and performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.568, + 0.892, + 0.627 + ], + "angle": 0, + "content": "- We propose a recorruption-based self-reconstruction loss function with noise tolerance to leverage Itoh's continuity condition, as well as a self-distillation loss function for improved generalization." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.432, + 0.892, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.631, + 0.892, + 0.676 + ], + "angle": 0, + "content": "In extensive experiments, our proposed approach shows superior performance over existing GT-free PU methods and competes well against the supervised ones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.641, + 0.708 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Non-learning-based PU: Methods along this line fall into three types: reliability-guided, filtering-based, and model-driven. Reliability-guided methods [3, 9, 11, 23, 36, 54] use a manually-designed metric to quantify pixel-wise reliability for path selection or region expansion during phase correction. Due to unreliable quality scores in noisy regions, quality-guided methods are noise-sensitive. Filtering-based methods [2, 45] treat PU as a smoothing problem, leveraging non-linear filters for denoising and unwrapping. However, their limited-support filters fail to exploit global consistency constraints within a phase image. Model-driven methods [1, 8, 12, 15] construct and solve an optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "model of PU regularized by some handcrafted priors. While achieving strong global consistency brought by the global optimization, these methods often fall short for complex phase patterns, due to the use of handcrafted priors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.47, + 0.2 + ], + "angle": 0, + "content": "Supervised DL for PU: Supervised DL-based PU methods differ mainly in NN architectures. Existing methods can be roughly classified as CNN-based or RNN-based." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Leveraging the effectiveness of CNNs in segmentation and classification, many CNN-based methods (e.g. [10, 34, 35, 49, 50, 52, 56]) cast PU as a wrap count prediction problem, using a pixel-wise classification loss with wrap counts as labels. For robustness to noise, pre-denoising or post-processing modules are often included; see e.g. [34, 35, 52]. However, these methods face scalability issues with widerange phase data, where the candidate values for wrap counts are overwhelmed. Instead, some CNN-based methods [6, 25, 40, 57]. Typically, CNN-based methods are inefficient in exploiting global spatial dependencies. In contrast, the RNN-based methods [22, 32] excel at capturing long-range dependencies. Viewing a phase pixel as an element in a sequence, these methods predefine several paths across all phase pixels to form an RNN. However, due to computational costs, only a few paths can be utilized. This constraint on path selection results in suboptimal outcomes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.534 + ], + "angle": 0, + "content": "The NN designs of most existing methods overlook the physical priors of phase wrapping, which are beneficial for generalization, especially when training data is limited. Our proposed DUN incorporates these priors, enhancing interpretability, reducing overfitting, and boosting performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Unsupervised DL for PU: Despite significant advances in unsupervised DL for various inverse problems in imaging, e.g., denoising [17, 21, 26], deblurring [5, 18, 27, 31], compressed sensing [28, 38, 42] and phase retrieval [4, 29], unsupervised DL for PU has been much less studied. One related work is Yang et al. [47], which utilizes the deep image prior [39] encoded by an untrained CNN to address the absence of GT data. Specifically, the phase image is reparameterized by a CNN, with its weights learned to fit the measurements. As shown in [39, 47], the CNN prefers smooth patterns and requires early stopping for smooth phase unwrapping, akin to conventional model-based methods but employing a CNN-based prior. A drawback is the need for per-sample retraining, which is computationally intensive. In contrast, our approach allows end-to-end unsupervised training on a dataset, eliminating test-time learning and significantly speeding up test." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "DUNs for imaging: DUNs have been studied for many inverse problems in imaging, e.g., denoising [20, 55], super-resolution [51], deblurring [20], compressed sensing [24, 53] and medical imaging [7]. However, the study on deep unrolling for PU is scant. Our study is the first to develop DUNs for PU, in addition to an unsupervised training scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.636, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.661, + 0.131 + ], + "angle": 0, + "content": "3.1. Basics and Keys" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.138, + 0.894, + 0.185 + ], + "angle": 0, + "content": "Consider estimating \\( \\mathbf{X} \\) from its wrapped version \\( \\mathcal{W}(\\mathbf{X}) = \\mathbf{X} - 2\\pi \\mathbf{K}, \\mathbf{K} \\in \\mathbb{Z}^{M \\times N} \\). Let \\( \\nabla : \\mathbf{Z} \\to (\\nabla_x \\mathbf{Z}; \\nabla_y \\mathbf{Z}) \\) denote the spatial gradient operator. By definition," + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.19, + 0.892, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\boldsymbol {\\nabla} \\boldsymbol {X} [ m, n ] + 2 \\pi \\boldsymbol {\\nabla} \\boldsymbol {K} [ m, n ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.214, + 0.894, + 0.245 + ], + "angle": 0, + "content": "For any location \\([m,n]\\) whose adjacent pixels for gradient calculation have the same wrap count, we have" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.251, + 0.892, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.321 + ], + "angle": 0, + "content": "In other words, the wrap count can be nullified by the gradient operator on many points. For the points not meeting the condition, we have the following Proposition." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.359 + ], + "angle": 0, + "content": "Proposition 1. [Two-dimensional extension of Itoh's continuity condition [13]] For any point \\([m, n]\\) satisfying" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.365, + 0.894, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\nabla \\boldsymbol {X} [ m, n ] \\right\\| _ {\\infty} = \\max \\left\\{\\left| \\nabla_ {x} \\boldsymbol {X} [ m, n ] \\right|, \\left| \\nabla_ {y} \\boldsymbol {X} [ m, n ] \\right| \\right\\} < \\pi , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.396, + 0.824, + 0.412 + ], + "angle": 0, + "content": "we have that \\(\\mathcal{W}(\\nabla \\mathcal{W}(\\pmb {X}))[m,n] = \\nabla \\pmb {X}[m,n]\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.417, + 0.817, + 0.433 + ], + "angle": 0, + "content": "Proof. See supplemental material for the proof." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.562 + ], + "angle": 0, + "content": "Proposition 1 shows that the wrapped gradients of a wrapped phase image are identical to the gradients of the GT, for the points satisfying the 2D Itoh's continuity condition (5) (called Itoh's condition for brevity in the following text). In general, GT phases exhibit gradual changes in most points, and their rapid changes only occur at a few points sparsely distributed over the image. Then, the outlier points not meeting Itoh's condition are sparse." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Now consider a noisy wrapped image \\(\\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N})\\). We can directly extend Eq. (4) to the noisy case:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.599, + 0.892, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\nabla \\boldsymbol {Y} [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ] + \\nabla \\boldsymbol {N} [ m, n ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.622, + 0.892, + 0.668 + ], + "angle": 0, + "content": "for any location \\([m, n]\\) in \\((\\mathbf{X} + \\mathbf{N})\\) whose adjacent points share the same wrap count. Similarly, we have the following corollary of Proposition 1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.673, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Corollary 1.1. For points \\([m,n]\\) satisfying \\(\\| \\nabla X[m,n] + \\nabla N[m,n]\\|_{\\infty} < \\pi\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.711, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} (\\nabla \\boldsymbol {Y} [ m, n ]) = \\mathcal {W} (\\nabla \\boldsymbol {X} [ m, n ]) + \\nabla \\boldsymbol {N} [ m, n ]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.903 + ], + "angle": 0, + "content": "It is clear that both \\(\\nabla Y\\) and \\(\\mathcal{W}(\\nabla Y)\\) can act as a noisy version for \\(\\nabla X\\), except on some generally-sparse outlier points that do not satisfy the corresponding conditions. If we can mitigate those noise and outliers to predict \\(\\nabla X\\), the remaining task, reconstructing \\(X\\) from \\(\\nabla X\\), is straightforward, as \\(X\\) can be reconstructed by integrating its spatial gradients \\(\\nabla X\\) with appropriate boundary conditions. Indeed, the primary concern in PU revolves around relative phases rather than the absolute ones. The observations described by Eq. (7) and Eq. (6) underpin our DUN and unsupervised loss function, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25184" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.862, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.388, + 0.644, + 0.403 + ], + "angle": 0, + "content": "Figure 2. Architecture of our proposed DUN for PU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.428, + 0.358, + 0.444 + ], + "angle": 0, + "content": "3.2. Deep Unrolling Network for PU" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.572 + ], + "angle": 0, + "content": "Our DUN is based on a regularization model inspired by Eq. (7). Though Eq. (7) is inapplicable to the pixels not meeting Itoh's condition, these outliers are typically sparse in a phase image with gradual changes in most areas. Consequently, we introduce a sparse variable \\( \\mathbf{E} \\in \\mathbb{R}^{M \\times N \\times 2} \\) into the regularization model, which counteracts the errors introduced by those sparse outliers. Specifically, we consider the following regularization model of PU:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.583, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {X}, \\boldsymbol {E}} \\| \\boldsymbol {\\nabla} \\boldsymbol {X} - \\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) + \\boldsymbol {E} \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\boldsymbol {X}) + \\psi (\\boldsymbol {E}), \\quad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.678 + ], + "angle": 0, + "content": "where \\(\\phi, \\psi\\) are some regularization functions for the latent phase image \\(X\\) and the error matrix \\(E\\), respectively. It can be seen that \\(E\\) essentially neutralizes outlier effects, preserving correctness of the data fitting term." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.471, + 0.709 + ], + "angle": 0, + "content": "Our DUN is constructed via the unrolled iterative solver for (8), which reads as follows: for \\(t = 1,\\dots ,T\\)" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.718, + 0.495, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\pmb {X} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {X}} \\| \\pmb {\\nabla} \\pmb {X} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {E} ^ {(t - 1)}) \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\pmb {X}), \\\\ \\pmb {E} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {E}} \\| \\pmb {E} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {\\nabla} \\pmb {X} ^ {(t)}) \\| _ {\\mathrm {F}} ^ {2} + \\psi (\\pmb {E}). \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Accordingly, the DUN predicts \\( \\mathbf{X} \\) from \\( \\mathcal{W}(\\nabla \\mathbf{Y}) \\). It alternatively connects two types of sub-NNs that act as the estimators of \\( \\mathbf{X}^{(t)} \\) and \\( \\mathbf{E}^{(t)} \\), respectively. See Fig. 2 for an illustration of this architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.824, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Sub-NN for estimating \\( \\boldsymbol{X}^{(t)} \\): The sub-problem for \\( \\boldsymbol{X}^{(t)} \\) is solved using proximal gradient descent (PGD), which combines gradient descent with proximal mapping. Considering the limited number of modules (i.e., iteration number \\( T \\)) in our DUN for computational efficiency, we implement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.507 + ], + "angle": 0, + "content": "multiple gradient descent steps within a stage to bridge the gap. Specifically, Nesterov's acceleration scheme is used to enhance the gradient descent, called AGD. We initialize with \\( \\mathbf{X}_{(0)}^{(t)} = \\mathbf{X}^{(t - 1)}, \\mathbf{X}^{(0)} = \\mathbf{1}, \\mathbf{E}^{(0)} = \\mathbf{0} \\) and \\( \\alpha_{(0)} = 0 \\). For \\( j = 1,\\dots ,J \\)," + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.515, + 0.925, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\boldsymbol {X} _ {(j)} ^ {(t)} = \\boldsymbol {V} _ {(j - 1)} ^ {(t)} + \\lambda^ {(t)} \\mathrm {d i v} \\big (\\boldsymbol {\\nabla} \\boldsymbol {V} _ {(j - 1)} ^ {(t)} - (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) - \\boldsymbol {E} ^ {(t - 1)}) \\big), \\\\ \\alpha_ {(j)} = \\frac {1}{2} (1 + \\sqrt {1 + 4 \\alpha_ {(j - 1)} ^ {2}}), \\\\ \\boldsymbol {V} _ {(j)} ^ {(t)} = \\boldsymbol {X} _ {(j)} ^ {(t)} + \\frac {\\alpha_ {(j - 1)} - 1}{\\alpha_ {(j)}} (\\boldsymbol {X} _ {(j)} ^ {(t)} - \\boldsymbol {X} _ {(j - 1)} ^ {(t)}). \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.594, + 0.893, + 0.661 + ], + "angle": 0, + "content": "Here the term \\(\\mathrm{div}\\bigl (\\pmb {\\nabla}V_{(j - 1)}^{(t)} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y}) - \\pmb{E}^{(t - 1)})\\bigr)\\) corresponds to the gradient \\(\\mathcal{G}_t(\\pmb {X}) = \\pmb {\\nabla}_\\pmb {X}\\| \\pmb {\\nabla}\\pmb {X} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y})-\\) \\(E^{(t - 1)})\\|_{\\mathrm{F}}^2\\) . Afterward, the proximal operator, defined as \\(\\mathrm{Prox}_f(\\bar{Z}):= \\mathrm{argmin}_{\\pmb{Z}}\\| \\pmb {Z} - \\bar{\\pmb{Z}}\\| _2^2 +f(\\pmb {Z})\\) , is applied:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.67, + 0.892, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {(t)} = \\operatorname {P r o x} _ {\\phi} \\left(\\boldsymbol {X} _ {(J)} ^ {(t)}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.838 + ], + "angle": 0, + "content": "This proximal operator is implemented using a sub-NN with an attentive U-shape structure. As illustrated in Fig. 2, it has a \\(16 \\times\\) down-scaling encoder and a corresponding up-scaling decoder, connected by skip connections. At the coarsest scale, a standard self-attention block is used for exploiting global features. The sub-NN also takes the gradient term \\(\\mathcal{G}_t(\\boldsymbol{X}_{(J)}^{(t)})\\) as an additional input via concatenation. A long skip connection from the input to output is added, enhancing feature flow with a weighted summation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Sub-NN for estimating \\(E^{(t)}\\): The sub-problem for \\(E^{(t)}\\) aims to address the outliers presented in the data fitting term defined on \\(\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}\\), guided by the prior encoded in \\(\\psi (E)\\). We employ a sub-NN with \\(\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.183 + ], + "angle": 0, + "content": "as the input to implement the solver of this sub-problem. This sub-NN, illustrated in Fig. 2, is composed of sequential convolutional layers, each followed by a parametric rectified linear unit (PReLU). The sparsity prior of \\(\\pmb{E}\\) is induced by attaching a soft thresholding activation function, parametrized by a non-negative parameter \\(\\mathrm{ReLU}(d^{(t)})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.468, + 0.29 + ], + "angle": 0, + "content": "Condition-aware parameter estimation: The noise level in the wrapped phases impacts the optimal selection of the parameters \\(\\{\\lambda^{(t)}, w^{(t)}, d^{(t)}\\}_{t=1}^T\\). Existing DUNs (e.g. [51]) often include imaging condition values in the input. In our DUN, a condition-aware module (CAM) is used to predict these parameters from noise level, using a chain of fully-connected layers with PReLU or Sigmoid activations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.448, + 0.315 + ], + "angle": 0, + "content": "3.3. Unsupervised End-to-End Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.468, + 0.381 + ], + "angle": 0, + "content": "Let \\(\\mathcal{F}\\) denote our DUN that maps \\(\\mathcal{W}(\\nabla Y)\\) to \\(X\\). Our proposed loss function \\(\\mathcal{L}_{\\mathrm{total}}\\) for the unsupervised training of \\(\\mathcal{F}\\) comprises a noise-resistant self-reconstruction loss \\(\\mathcal{L}_{\\mathrm{sr}}\\) and a self-distillation loss \\(\\mathcal{L}_{\\mathrm{sd}}\\), applied to each stage of the DUN:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.393, + 0.468, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {s r}} + \\eta \\mathcal {L} _ {\\mathrm {s d}}, \\eta \\in \\mathbb {R} ^ {+}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.468, + 0.467 + ], + "angle": 0, + "content": "Noise-resistant self-reconstruction loss: This loss utilizes \\(\\nabla Y\\) as a noisy label for \\(\\nabla X\\), as shown in Eq. (6). To mitigate the impact of label noise, the loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.477, + 0.468, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) - (\\nabla \\boldsymbol {Y} - \\nabla \\boldsymbol {U}) \\| _ {\\mathrm {F}} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.468, + 0.565 + ], + "angle": 0, + "content": "where \\(U\\) is random noise drawn from the distribution of \\(N\\). The introduction of the noise pair \\(\\{+\\nabla U, -\\nabla U\\}\\) is to achieve the resilience to measurement noise in \\(\\nabla Y\\) in the training loss, as justified in Proposition 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.468, + 0.618 + ], + "angle": 0, + "content": "Proposition 2. Let \\( \\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N}) \\). Suppose Eq. (6) is satisfied at all points. Assume that \\( \\mathbf{N}, \\mathbf{U} \\sim \\mathcal{P} \\) are independent. Then, we have that" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.628, + 0.468, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {Y}} \\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {X}, \\boldsymbol {N}, \\boldsymbol {U}} \\| \\boldsymbol {\\nabla} \\mathcal {F} (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})) - \\boldsymbol {\\nabla} \\boldsymbol {X} \\| _ {F} ^ {2} + C _ {0}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.234, + 0.674 + ], + "angle": 0, + "content": "where \\(C_0\\) is a constant." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.385, + 0.699 + ], + "angle": 0, + "content": "Proof. See the proof in supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.468, + 0.858 + ], + "angle": 0, + "content": "Proposition 2 shows that if one can independently draw instances \\( \\pmb{U} \\) from the same distribution as \\( N \\), the loss \\( \\mathcal{L} \\) effectively simulates the loss supervised by \\( \\nabla X \\), nullifying the label noise as shown in Eq. (12). Note that Proposition 2 requires all points satisfying Eq. (6), which is not generally true. On the outliers, \\( \\nabla Y \\) provides wrong supervision with superfluous multiples of \\( 2\\pi \\), as seen in Eq. (3). To counteract the impact of those outlier points, we exploit the property \\( \\mathcal{W}(\\theta + 2k\\pi) = \\theta \\) to modify the self-reconstruction loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.868, + 0.479, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s r}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\mathcal {W} \\left[ \\boldsymbol {\\nabla} \\mathcal {F} \\left(\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})\\right) - (\\boldsymbol {\\nabla} \\boldsymbol {Y} - \\boldsymbol {\\nabla} \\boldsymbol {U}) \\right] \\| _ {\\mathrm {F}} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.136 + ], + "angle": 0, + "content": "That is, an extra \\(\\mathcal{W}\\) is applied to the prediction residual within the Frobenius norm, which negates the superfluous multiples of \\(2\\pi\\) during the loss calculation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Self-distillation loss: Although \\(\\mathcal{L}_{\\mathrm{sr}}\\) reduces the impact of measurement noise, the ambiguity in solutions due to wrapping may remain, particularly at those outlier points. Fortunately, CNNs including unrolling CNNs [30, 39], show strong induced bias towards regular structures, which aids in reducing such ambiguity during the training with \\(\\mathcal{L}_{\\mathrm{sr}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.23, + 0.891, + 0.259 + ], + "angle": 0, + "content": "To further reduce the ambiguity, we introduce a self-distillation loss defined as" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.267, + 0.891, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s d}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y})) - \\nabla \\bar {\\mathcal {F}} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) \\| _ {\\mathrm {F}} ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.435 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathcal{F}}\\) denotes the NN detached from the previous iteration with stop gradient. The loss leverages pseudo supervision from the NN's varied predictions with noise-altered inputs, implicitly aggregating those predictions for self-training. This self-distillation scheme can reduce the NN's prediction variance, enhancing the PU accuracy. Moreover, recall that \\(\\mathcal{L}_{\\mathrm{sr}}\\) inputs \\(\\mathcal{W}(\\nabla Y + \\nabla U)\\) to \\(\\mathcal{F}\\), which does not fully align with \\(\\mathcal{W}(\\nabla Y)\\) in test. This inconsistency can be reconciled by \\(\\mathcal{L}_{\\mathrm{sd}}\\), as it uses \\(\\mathcal{W}(\\nabla Y)\\) for the input of \\(\\mathcal{F}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.632, + 0.465 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Performance evaluation is conducted on both simulated and real-world phase patterns. Following [22], we use NRMSE (Normalized Root Mean Squared Error) as the performance metric. As relative phase is the main focus in PU, the mean of unwrapped phase values is first aligned to that of the GT before the NRMSE calculation. Computational complexity is measured by three metrics: the number of parameters, the number of FLOPs (floating point of operations), and the average test time per sample of size \\( 256 \\times 256 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.733 + ], + "angle": 0, + "content": "Implementation details: Our proposed approach, called U3Net (U3 for Unsupervised, Unrolling and Unwrapping) for convenience, is implemented in Python and run on a single NVIDIA RTX 3080 GPU. We set \\(\\eta = 0.5\\) in \\(\\mathcal{L}_{\\mathrm{total}}\\) and initialize model weights via kaiming Uniform. The training employs the Adam optimizer with 500 epochs and a batch size of 10. The learning rate is initialized as \\(1\\times 10^{-3}\\) and exponentially decayed by a factor of 0.99 at every epoch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Methods for comparison: We select eight representative methods of general PU as baselines, including LS [8], QGPU [11], Ryu. et al. [32], PhaseNet2.0 [35], PUDIP [47], SQD-LSTM [22], EESANet [49], and TriNet [37]. Additionally, two representative transformer NNs for general image restoration, UFormer [43] and Restormer [48], are also included for performance comparison. In all the baseline methods, LS and QGPU are conventional non-learning iterative methods, built upon least squares and path selection, respectively; PUDIP is an internal DL-based method; and the others are supervised DL-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25186" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.862, + 0.281 + ], + "angle": 0, + "content": "
DatasetMoGRRME#Param. (M)#FLOPs (G)Time (msec.)
SNR(dB)0510203005102030
Non-LearningLS7.912.281.200.370.157.852.501.250.400.12--8.39
QGPU17.122.121.150.360.1117.262.291.200.390.12--14.94
SupervisedRyu. et al.1.341.111.091.051.051.511.051.010.980.931.0721.85303.96
PhaseNet2.08.358.198.078.068.049.298.537.536.956.861.1511.9320.87
SQD-LSTM0.870.710.700.690.681.571.131.121.101.090.904.0713.04
EESANet0.780.770.760.760.741.311.311.251.241.0661.6875.279.85
TriNet5.655.555.465.405.345.405.335.215.105.0513.6165.4811.16
UFormer0.500.470.470.460.450.580.510.490.490.4820.6040.9842.00
Restormer0.450.380.360.360.350.500.430.420.410.413.0217.2344.69
UnsupervisedPUDIP17.5315.167.870.340.1113.107.222.620.380.122.3321.5899695.01
U3Net0.690.250.190.160.101.120.380.270.170.120.748.7710.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.29, + 0.856, + 0.305 + ], + "angle": 0, + "content": "Table 1. Performance comparison on synthetic datasets in terms of NRMSE(%) . Bold for Best and underline for second-best." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.317, + 0.889, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.508, + 0.871, + 0.523 + ], + "angle": 0, + "content": "Figure 3. Visualization of PU results on MoGR (top) and RME (bottom), showing residual images to facilitate visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.625 + ], + "angle": 0, + "content": "based methods with their NNs retrained on our generated paired phase data. Specifically, as UFormer and Restormer are not originally designed for PU, we train their models using the supervised loss of SQD-LSTM [22] for better performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.422, + 0.65 + ], + "angle": 0, + "content": "4.1. Evaluation on Simulated Phase Patterns" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Data preparation: Two types of simulated data are generated for training and evaluation. The first type follows [22], where a GT phase image of size \\( 256 \\times 256 \\) is generated through the superposition of a mixture of Gaussians (MoG) with \\( c \\) clusters and randomly selected ramps. The phase range is defined within \\( [-2p\\pi, 2p\\pi] \\). The resulting dataset is called MoGR (MoG with Ramp). The noisy wrapped phase images are synthesized via Eq. (1), with \\( N \\) defined as Gaussian white noise. For each GT, only one noisy wrapped phase image is generated. With \\( c \\) and \\( p \\) uniformly selected from \\( \\{1, \\dots, 4\\} \\) and \\( \\{1, \\dots, 7\\} \\) respectively, we generate 5000 noisy wrapped phase images with SNR uniformly selected from \\( \\{0, 5, 10, 20, 30, 60\\} \\) to construct the training set. Using the same way, 1000 noisy wrapped phase images are generated for test with SNR in \\( \\{0, 5, 10, 20, 30\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.9 + ], + "angle": 0, + "content": "The other dataset follows a different construction scheme" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "named random matrix enlargement [40], thus called RME dataset. Concretely, a GT phase image is formed through enlarging the square matrix of a distribution from an initial size \\( s \\times s \\) to size \\( 256 \\times 256 \\) via interpolation. The size parameter \\( s \\) is selected uniformly from \\( \\{2, \\dots, 10\\} \\), the distribution type is chosen from the {Uniform, Normal} distributions, and the interpolation method is chosen from the {Bilinear, Bicubic} interpolations. Then the phase image is re-scaled to the range varying in \\( [-2p\\pi, 2p\\pi] \\) with \\( p \\) randomly chosen from \\( \\{1, \\dots, 7\\} \\). The noisy wrapped phase images are obtained by the same way as the MoGR dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.856 + ], + "angle": 0, + "content": "Results and analysis: Table 1 lists the quantitative results on both datasets, including those in computational complexity. Even trained without GT data, U3Net achieves the best results in 8/10 settings, using a lightweight model. Concretely, U3Net outperforms the two conventional non-learning-based methods, LS and QGPU. In comparison to PUDIP, an unsupervised internal DL-based method, U3Net also shows significant performance gain. The performance advantage of U3Net over these three GT-free methods be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.893, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)PUDIP involves many forward passes for test-time learning. We report its #FLOPs in one forward pass and running time of the whole process." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25187" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.887, + 0.155 + ], + "angle": 0, + "content": "
SNR(dB)LSQGPURyu. et al.PhaseNet2.0SQD-LSTMEESANetTriNetPU-GANPUNetUFormerRestormerPUDIPU3Net
53.313.131.412.281.762.455.0413.739.591.461.069.991.00
101.841.961.271.691.521.994.6911.849.201.280.935.300.82
200.941.131.241.431.481.794.4611.629.010.970.910.470.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.164, + 0.856, + 0.178 + ], + "angle": 0, + "content": "Table 2. Performance comparison on InSAR phase data in terms of NRMSE(%) . Bold for Best and underline for second-best." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.192, + 0.818, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.387, + 0.854, + 0.402 + ], + "angle": 0, + "content": "Figure 4. Visualization of unwrapped phase images on InSAR data, showing residual images to facilitate visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.468, + 0.532 + ], + "angle": 0, + "content": "comes more noticeable for lower SNR (i.e. heavier noise), particularly compared to PUDIP. This is because U3Net has an explicit mechanism to deal with noise during training, thereby achieving higher noise robustness. In addition, U3Net is much faster than PUDIP, due to that U3Net employs end-to-end pre-trained models, avoiding the time-costly per-sample training of PUDIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.468, + 0.744 + ], + "angle": 0, + "content": "Surprisingly, U3Net also surpasses the supervised DL-based methods in many settings. In the lowest-SNR case, U3Net performs worse than UFormer and Restormer, two supervised transformer models. This is probably due to that severe noise increases the number of outlier points, lowering the effectiveness of the training function. Even that, U3Net still outperforms other supervised methods. The model of U3Net is the smallest among all the DL-based methods, indicating that the excellent performance of U3Net is mainly attributed to its interpretable physics-encoded NN architecture and its effective training scheme, rather than the use of larger models. Meanwhile, U3Net is computationally efficient, with the second smallest number of FLOPs and the third fastest inference speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.468, + 0.835 + ], + "angle": 0, + "content": "See Fig. 3 for the visualization of PU results from different methods. The residuals to GTs produced by U3Net are much more insignificant compared to the internal DL-based PUDIP. In comparison to supervised methods, U3Net outperforms all of them on the image from RME dataset and shows comparable results on the image of MoGR dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.396, + 0.862 + ], + "angle": 0, + "content": "4.2. Evaluation on InSAR Phase Patterns" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The InSAR phase data are generated by converting the digital elevation model at Texas, U.S. into absolute phases," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.593 + ], + "angle": 0, + "content": "following [33]. Totally, 4000 (1000) GT phase images are obtained by cropping the \\(256 \\times 256\\) patches from the absolute phases for training (test). Each phase value is scaled into \\([-6\\pi, 6\\pi]\\). The noisy wrapped phases are generated by adding Gaussian noise following the experiments in Sec. 4.1, and we test on three noise levels: 5, 10, 20. In addition to the previously-compared general PU methods and general-purpose transformer NNs, we also compare with two specific methods designed for InSAR PU: PUNet [44] and PU-GAN [57]. All the compared models are retrained using the same data as ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.714 + ], + "angle": 0, + "content": "See Table 2 for the quantitative comparison on this In-SAR dataset. Our U3Net ranks the first among all compared methods over all noise settings, noticeably surpassing other GT-free methods as well as many supervised methods in many cases. The visual comparison provided in Fig. 4 also shows the advantage of U3Net. All these results further verify the effectiveness of our approach in dealing with phase patterns from real scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.664, + 0.739 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.791 + ], + "angle": 0, + "content": "We form some baselines from U3Net for ablation studies, which are retrained and tested on the MoGR dataset. Their results are summarized in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Study on loss function: We design the following schemes to analyze the loss function: (a) \\(\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}\\): remove the outer \\(\\mathcal{W}\\) applied to the residual, from the self-reconstruction loss \\(\\mathcal{L}_{\\mathrm{sr}}\\), i.e., replacing \\(\\mathcal{L}_{\\mathrm{sr}}\\) by the loss \\(\\mathcal{L}\\) defined in Eq. (11); (b) w/o \\(\\nabla U\\): remove the noise re-corruption mechanism in \\(\\mathcal{L}_{\\mathrm{sr}}\\) by discarding the \\(\\pm \\nabla U\\) terms; (c) \\(\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}\\) w/o \\(\\nabla U\\): combine the former cases, yielding a plain gradient loss" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25188" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.25 + ], + "angle": 0, + "content": "
SettingSNR=0SNR=5SNR=10SNR=20SNR=30
Lsr→L25.4523.4823.2623.2323.20
w/o ∇U6.162.401.360.690.60
Lsr→L w/o ∇U25.8523.5723.4223.3823.34
w/o Lsd1.140.330.230.220.17
Supervised0.290.160.130.110.05
w/o E7.160.750.440.360.31
AGD→GD2.140.350.210.170.12
w/o CAM0.980.310.220.210.15
W(∇Y)→∇Y2.881.741.501.461.39
DUN→UFormer1.551.411.311.301.30
U3Net (original)0.690.250.190.160.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.26, + 0.452, + 0.273 + ], + "angle": 0, + "content": "Table 3. Results in ablation studies in terms of NRMSE(%)" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.291, + 0.457, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.503 + ], + "angle": 0, + "content": "Figure 5. Visualization of unwrapped phase images using models trained by different loss functions. \\(\\mathcal{L}_{\\mathrm{sr}} \\rightarrow \\mathcal{L}\\) and w/o \\(\\nabla U\\) produce noticeable errors than Original. Zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.594 + ], + "angle": 0, + "content": "\\(\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla Y\\|_{\\mathrm{F}}^2\\) (d) w/o \\(\\mathcal{L}_{\\mathrm{sd}}\\) : remove the selfdistillation loss \\(\\mathcal{L}_{\\mathrm{sd}}\\) , only employing \\(\\mathcal{L}_{\\mathrm{sr}}\\) for training; and (e) Supervised: train with a gradient-domain supervised loss \\(\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla X\\|_{\\mathrm{F}}^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.793 + ], + "angle": 0, + "content": "In addition to Table 3, we also provide some visual comparisons in Fig. 5 for demonstration. From both the quantitative and qualitative results, we have following observations. (a) The outer \\(\\mathcal{W}\\) in \\(\\mathcal{L}_{\\mathrm{sr}}\\) is critical to the performance of U3Net. (b) The re-corruption mechanism in \\(\\mathcal{L}_{\\mathrm{sr}}\\) does help eliminate the measurement noise. (c) A simple gradient loss function does not work, with further performance decrease. (d) While the self-reconstruction loss already yields promising results, the self-distillation loss benefits further performance gain. (e) U3Net performs closely to its supervised counterpart when \\(\\mathrm{SNR} \\geq 5\\). We also see that U3Net has its supervised counterpart outperforming the supervised UFormer reported in Table 1, demonstrating its superiority." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Study on DUN: To analyze the components in the DUN architecture of U3Net, we design the following baseline models: (a) w/o \\( \\pmb{E} \\): abandon the error term \\( \\pmb{E} \\) during unrolling, forming a DUN alternating the accelerated gradient descent step and the sub-NN for estimating \\( \\pmb{X}^{(t)} \\); (b) AGD→GD: replace the Nesterov's AGD steps by one-step standard gradient; (c) w/o CAM: remove the CAM from U3Net; (d)" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.882, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Visualization of error matrices \\(\\pmb{E}\\) estimated from three wrapped phase images. The outlier maps use white pixels to indicate the points \\([x,y]\\) satisfying \\(\\| \\nabla X[x,y] + \\nabla N[x,y]\\|_{\\infty} \\geq \\pi\\) and black pixels otherwise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.456 + ], + "angle": 0, + "content": "\\(\\mathcal{W}(\\nabla \\pmb {Y})\\rightarrow \\nabla \\pmb {Y}\\) : use \\(\\nabla Y\\) to replace \\(\\mathcal{W}(\\nabla Y)\\) in the input of U3Net as well in the total loss; (e) DUN \\(\\rightarrow\\) UFormer: replace the DUN by UFormer, keeping the loss unchanged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.893, + 0.681 + ], + "angle": 0, + "content": "As demonstrated in Table 3, each component in the DUN of U3Net is useful. (a) One critical part is the introduction of the error matrix \\(\\pmb{E}\\) for unrolling. It brings significant performance gain, particularly in dealing with low-SNR data, as the absorption of outliers breaking the Itoh's condition is more desired for heavier noise. See Fig. 6 for a visualization of estimated error matrices, which successfully identify the outliers. (b)&(c) The AGD and CAM have similar contribution to the performance of U3Net. (d) Using \\(\\mathcal{W}(\\nabla Y)\\) for DUN is also crucial, serving as a better choice compared to using \\(\\nabla Y\\). (e) Our proposed loss function also works for UFormer, but yielding worse results than U3Net, which indicates the higher effectiveness of our physics-encoded DUN for unsupervised PU, compared to general transformer models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.62, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This work proposed a fully unsupervised DL approach for PU, eliminating the dependence on GTs in end-to-end training. Utilizing the connection between (wrapped) gradients of wrapped phases and gradients of original phases, we constructed a physics-encoded DUN, trained via a noise-resistant self-reconstruction loss complemented by a self-distillation loss. Empirical evaluations confirmed that our approach surpassed existing GT-free methods and rivaled supervised counterparts. Similar to current unsupervised and synthetic data-dependent supervised techniques, our approach presumes known noise statistics. Future research will focus on dealing with unknown noise profiles." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25189" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Liheng Bian, Xin Wang, Daoyu Li, Qiuling Ren, and Dezhi Zheng. Robust phase unwrapping via non-local regularization. Optics Letters, 48(6):1399-1402, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] David Blinder, Heidi Ottevaere, Adrian Munteanu, and Peter Schelkens. Efficient multiscale phase unwrapping methodology with modulo wavelet transform. Optics Express, 24(20): 23094-23108, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Donald J Bone. Fourier fringe analysis: the two-dimensional phase unwrapping problem. Applied Optics, 30(25):3627-3632, 1991. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[4] Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate mmse estimation. IEEE Transactions on Signal Processing, 70:2239-2252, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.468, + 0.371 + ], + "angle": 0, + "content": "[5] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.468, + 0.428 + ], + "angle": 0, + "content": "[6] Gili Dardikman and Natan T Shaked. Phase unwrapping using residual neural networks. In Computational Optical Sensing and Imaging, pages CW3B-5. Optica Publishing Group, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.485 + ], + "angle": 0, + "content": "[7] Qiaoqiao Ding, Gaoyu Chen, Xiaoqun Zhang, Qiu Huang, Hui Ji, and Hao Gao. Low-dose ct with deep learning regularization via proximal forward-backward splitting. Physics in Medicine & Biology, 65(12):125009, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.468, + 0.542 + ], + "angle": 0, + "content": "[8] Dennis C Ghiglia and Louis A Romero. Robust two-dimensional weighted and unweighted phase unwrapping that uses fast transforms and iterative methods. Journal of the Optical Society of America A, 11(1):107-117, 1994. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.468, + 0.598 + ], + "angle": 0, + "content": "[9] Richard M Goldstein, Howard A Zebker, and Charles L Werner. Satellite radar interferometry: Two-dimensional phase unwrapping. Radio Science, 23(4):713-720, 1988. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[10] Michal Gontarz, Vibekananda Dutta, Malgorzata Kujawinska, and Wojciech Krauze. Phase unwrapping using deep learning in holographic tomography. Optics Express, 31(12):18964-18992, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.468, + 0.727 + ], + "angle": 0, + "content": "[11] Miguel Arevallilo Herraez, David R Burton, Michael J Lolor, and Munther A Gdeisat. Fast two-dimensional phaseunwrapping algorithm based on sorting by reliability following a noncontinuous path. Applied Optics, 41(35):7437-7444, 2002. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[12] Howard YH Huang, L Tian, Z Zhang, Y Liu, Z Chen, and G Barbastathis. Path-independent phase unwrapping using phase gradient and total-variation (tv) denoising. Optics Express, 20(13):14075–14089, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[13] Kazuyoshi Itoh. Analysis of the phase unwrapping algorithm. Applied Optics, 21(14):2470-2470, 1982. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[14] Charles VJ Jakowitz, Daniel E Wahl, Paul H Eichel, Dennis C Ghiglia, and Paul A Thompson. Spotlight-mode synthetic aperture radar: a signal processing approach. Springer Science & Business Media, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Ulugbek S Kamilov, Ioannis N Papadopoulos, Morteza H Shoreh, Demetri Psaltis, and Michael Unser. Isotropic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.134 + ], + "angle": 0, + "content": "inverse-problem approach for two-dimensional phase unwrapping. Journal of the Optical Society of America A, 32 (6):1092-1100, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Thomas Kreis. Handbook of holographic interferometry: optical and digital methods. John Wiley & Sons, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[17] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2137, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[18] Ji Li, Weixi Wang, Yuesong Nan, and Hui Ji. Self-supervised blind motion deblurring with deep expectation maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13986-13996, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[19] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[20] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[21] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2043-2052, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[22] Malsha V Perera and Ashwin De Silva. A joint convolutional and spatial quad-directional LSTM network for phase unwrapping. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4055-4059. IEEE, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[23] Mark D Pritt. Phase unwrapping by means of multigrid techniques for interferometric sar. IEEE Transactions on Geoscience and Remote Sensing, 34(3):728-738, 1996. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[24] Xinran Qin, Yuhui Quan, and Hui Ji. Enhanced deep unrolling networks for snapshot compressive hyperspectral imaging. Neural Networks, 174:106250, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[25] Yi Qin, Shujia Wan, Yuhong Wan, Jiawen Weng, Wei Liu, and Qiong Gong. Direct and accurate phase unwrapping with deep neural network. Applied Optics, 59(24):7258-7267, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[26] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1898, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[27] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In European Conference on Computer Vision, pages 642-659. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Yuhui Quan, Xinran Qin, Tongyao Pang, and Hui Ji. Dual-domain self-supervised learning and model adaption for deep compressive imaging. In European Conference on Computer Vision, pages 409-426. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25190" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[30] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.258 + ], + "angle": 0, + "content": "[31] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3341-3350, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[32] Kanghyun Ryu, Sung-Min Gho, Yoonho Nam, Kevin Koch, and Dong-Hyun Kim. Development of a deep learning method for phase unwrapping mr images. In Proceedings of the International Society for Magnetic Resonance in Medicine, page 4707, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[33] Francescopaolo Sica, Francesco Calvanese, Giuseppe Scarpa, and Paola Rizzoli. A cnn-based coherence-driven approach for insar phase unwrapping. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[34] GE Spoorthi, Subrahmanyam Gorthi, and Rama Krishna Sai Subrahmanyam Gorthi. Phasenet: A deep convolutional neural network for two-dimensional phase unwrapping. IEEE Signal Processing Letters, 26(1):54-58, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[35] GE Spoorthi, Rama Krishna Sai Subrahmanyam Gorthi, and Subrahmanyam Gorthi. Phasenet 2.0: Phase unwrapping of noisy data based on deep learning approach. IEEE Transactions on Image Processing, 29:4862-4872, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[36] Xianyu Su and Wenjing Chen. Reliability-guided phase unwrapping algorithm: a review. Optics and Lasers in Engineering, 42(3):245-261, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[37] Krishna Sumanth, Vaishnavi Ravi, and Rama Krishna Gorthi. A multi-task learning for 2d phase unwrapping in fringe projection. IEEE Signal Processing Letters, 29:797-801, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[38] Julián Tachella, Dongdong Chen, and Mike Davies. Unsupervised learning from incomplete measurements for inverse problems. Advances in Neural Information Processing Systems, 35:4983-4995, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[40] Kaiqiang Wang, Ying Li, Qian Kemao, Jianglei Di, and Jianlin Zhao. One-step robust deep learning phase unwrapping. Optics Express, 27(10):15100-15115, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[41] Kaiqiang Wang, Qian Kemao, Jianglei Di, and Jianlin Zhao. Deep learning spatial phase unwrapping: a comparative review. Advanced Photonics Nexus, 1(1):014001, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[42] Weixi Wang, Ji Li, and Hui Ji. Self-supervised deep image restoration via adaptive stochastic gradient Langevin dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Zhipeng Wu, Teng Wang, Yingjie Wang, Robert Wang, and Daqing Ge. Deep learning for the detection and phase unwrapping of mining-induced deformation in large-scale interferograms. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[45] Xianming Xie. Iterated unscented kalman filter for phase unwrapping of interferometric fringes. Optics Express, 24 (17):18872-18897, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[46] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[47] Fangshu Yang, Thanh-An Pham, Nathalie Brandenberg, Matthias P Lütolf, Jianwei Ma, and Michael Unser. Robust phase unwrapping via deep image prior for quantitative phase imaging. IEEE Transactions on Image Processing, 30: 7025-7037, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[49] Junkang Zhang and Qingguang Li. Eesanet: edge-enhanced self-attention network for two-dimensional phase unwrapping. Optics Express, 30(7):10470-10490, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[50] Junchao Zhang, Xiaobo Tian, Jianbo Shao, Haibo Luo, and Rongguang Liang. Phase unwrapping in optical metrology via denoised and convolutional segmentation networks. Optics Express, 27(10):14903-14912, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[51] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3217-3226, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[52] Teng Zhang, Shaowei Jiang, Zixin Zhao, Krishna Dixit, Xiaofei Zhou, Jia Hou, Yongbing Zhang, and Chenggang Yan. Rapid and robust two-dimensional phase unwrapping via deep learning. Optics Express, 27(16):23173-23185, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[53] Zhonghao Zhang, Yipeng Liu, Jiani Liu, Fei Wen, and Ce Zhu. Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing, 30:1487-1500, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[54] Ming Zhao, Lei Huang, Qican Zhang, Xianyu Su, Anand Asundi, and Qian Kemo. Quality-guided phase unwrapping technique: comparison of quality maps and guiding strategies. Applied Optics, 50(33):6214-6224, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[55] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 630-641, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Hongyu Zhou, Chuanli Cheng, Hao Peng, Dong Liang, Xin Liu, Hairong Zheng, and Chao Zou. The phu-net: A robust phase unwrapping method for mri based on deep learning." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25191" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Magnetic Resonance in Medicine, 86(6):3321-3333, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[57] Lifan Zhou, Hanwen Yu, Vito Pascazio, and Mengdao Xing. Pu-gan: a one-step 2-d insar phase unwrapping based on conditional generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 60:1-10, 2022. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25192" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_origin.pdf b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..979edb53ade13b6274e3f09f471deee33adb2079 --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/a918881b-c4ff-499e-bbcd-69ac8904aaf8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c32d62dfee41d35be66c0317ce224263fc95bf09a40ef129336a6f849dcd6716 +size 4575427 diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/full.md b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d90a6b1739e33d8250f63eb5b474c63a1f22c0eb --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/full.md @@ -0,0 +1,381 @@ +# Unsupervised Deep Unrolling Networks for Phase Unwrapping + +Zhile Chen $^{1,2}$ Yuhui Quan $^{1,2*}$ Hui Ji $^{3}$ + +$^{1}$ School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China $^{2}$ Pazhou Lab, Guangzhou 510320, China + +$^{3}$ Department of Mathematics, National University of Singapore, 119076, Singapore + +cs.zhilechen@mail.scut.edu.cn, csyhquan@scut.edu.cn, matjh@nus.edu.sg + +# Abstract + +Phase unwrapping (PU) is a technique to reconstruct original phase images from their noisy wrapped counterparts, finding many applications in scientific imaging. Although supervised learning has shown promise in PU, its utility is limited in ground-truth (GT) scarce scenarios. This paper presents an unsupervised learning approach that eliminates the need for GTs during end-to-end training. Our approach leverages the insight that both the gradients and wrapped gradients of wrapped phases serve as noisy labels for GT phase gradients, along with sparse outliers induced by the wrapping operation. A recorruption-based self-reconstruction loss in the gradient domain is proposed to mitigate the adverse effects of label noise, complemented with a self-distillation loss for improved generalization. Additionally, by unfolding a variational model of PU that utilizes wrapped gradients of wrapped phases for its data-fitting term, we develop a deep unrolling network that encodes physics of phase wrapping and incorporates special treatments on outliers. In the experiments on three types of phase data, our approach outperforms existing GT-free methods and competes well against the supervised ones. + +# 1. Introduction + +In many imaging systems, the direct acquisition of original phases of a target signal is both challenging and costly. Typically, initial measurement yields a phase image wrapped in $[- \pi, \pi)$ . Let $X \in \mathbb{R}^{M \times N}$ and $Y \in [-\pi, \pi)^{M \times N}$ denote the true phase image and its noisy wrapped counterpart, respectively. The wrapping process can be formulated as: + +$$ +\boldsymbol {Y} = \mathcal {W} (\boldsymbol {X} + \boldsymbol {N}), \quad \mathcal {W}: \mathbb {R} ^ {M \times N} \rightarrow [ - \pi , \pi) ^ {M \times N}, \tag {1} +$$ + +where $\mathcal{W}$ denotes the wrapping operator with its entry-wise operation defined as $\mathcal{W}(\theta) = ((\theta + \pi) \bmod 2\pi) - \pi$ for $\theta \in \mathbb{R}$ , and $N \in \mathbb{R}^{M \times N}$ denotes the measurement noise, typically assumed to be Gaussian [22]. Consequently, the original phase image, which exhibits strong smoothness, is degraded to a wrapped version with both incorrect values and artificial discontinuities induced by the modulo operation, which adversely affects subsequent processing. + +Phase unwrapping (PU) is about recovering the original unwrapped phase image $\mathbf{X}$ (up to an additive constant) from its noisy wrapped counterpart $\mathbf{Y}$ . Serving as a critical step in many optical imaging techniques for generating clear and coherent phase images, PU finds extensive applications in diverse domains such as quantitative phase imaging, magnetic resonance imaging, synthetic aperture radar interferometry, 3D depth sensing, phase contrast microscopy, fringe projection, and digital holographic interferometry; see e.g. [10, 14, 16, 23, 41, 47, 56, 57]. + +The model (1) is also usually formulated as follows: + +$$ +\boldsymbol {Y} = \boldsymbol {X} - 2 \pi \boldsymbol {K} + \boldsymbol {N} ^ {\prime}, \quad \boldsymbol {K} \in \mathbb {Z} ^ {M \times N}, \tag {2} +$$ + +where $K \in \mathbb{Z}^{M \times N}$ denotes the map of wrap counts, indicating the number of times a phase value has been wrapped around by $2\pi$ , and $N'$ denotes the noise dependent on both $X$ and $N$ , whose distribution can be complex even when $N$ follows a simple distribution such as the normal distribution. It can be seen that PU is a challenging ill-posed inverse problem involving a continuous variable $X$ and a discrete variable $K$ . Direct solutions by integrating wrapped phase differences are insufficient for ensuring spatial consistency in all directions of the unwrapped phases and may accumulate significant errors due to substantial noise. + +Limitations of existing works: Conventional approaches for PU include reliability-guided methods (e.g. [3, 9, 11, 23, 36]), filtering-based methods (e.g. [2, 45]), and model-driven methods (e.g. [1, 12, 15]). These approaches are either sensitive to measurement noise, due to their heuristic rules, or overly simplistic for complex phase structures, due to their handcrafted image priors. In contrast, deep learning + +![](images/e5d1b7d488142b56bf5730f32ea270b2b72b03793c60e50c6631347304ee3c2c.jpg) + +![](images/568acda83cbd389b33f62b98b25702fd3e47a9cbb7758f873558114cc829e178.jpg) +Randomly generated Set + +![](images/4408d9a135bed6a749e9f4540b0b86708297caeb14ddeceebc847a3b081ec5dd.jpg) +Wrapped Phase Images + +Supervised Learning +Figure 1. Training data setting of existing PU methods versus ours. +![](images/d53fd8067bf8677d978b49aba8118a13484dcbc1d67b863b9c35a4f787d6a00f.jpg) +- GT collection is hard + +Wrappered Phase Image (Single Test Sample) +Internal Learning +![](images/9658c25f4c65824b637632f84c479f946678ea5a960627a6cf26b955fe309a78.jpg) +Time consuming + +Wrapped Phase Images +End-to-End +Unsupervised Learning +![](images/492b3f000e6d48e83b7fc1417bc5f9d74aa02c3e93ec7733f559883528364f94.jpg) +Bypassing both issues + +(DL) with neural networks (NNs) has emerged as a powerful tool for PU [41]. A pre-trained end-to-end NN offers data adaptivity, rapid inference, and excellent expressivity. + +Existing DL-based PU methods often rely on supervised learning with ground-truth (GT) phases. However, acquiring GT phases is often impractical due to physical limits in imaging, destructive imaging techniques, and ethical/safety concerns. This challenge limits the applicability of supervised DL in many practical scenarios. Internal unsupervised DL-based methods (e.g. [47]) use an untrained NN to parameterize latent phases, adjusting it to fit observed measurements with certain regularization. Though flexible, this kind of methods requires time-costly per-sample test-time training, resulting in poor computational efficiency compared to supervised DL-based methods. + +Our approach: To bypass the difficulty of GT acquisition in supervised DL and avoid the computational inefficiency of internal unsupervised DL, we present an external unsupervised $DL$ approach for PU. It trains an end-to-end NN solely on a dataset of noisy wrapped measurements. See Fig. 1 for the comparison in training data configuration between existing works and ours. Indeed, our study is the first one on external (end-to-end) unsupervised DL for PU. + +Our approach is inspired by the insight that applying the gradient operator or the wrapping gradient operator to a wrapped phase image will result in an output identical to the gradient map of the GT phase image, with the exception of a sparse outlier set. Concretely, we have that + +(a) $\nabla Y[m,n] = \nabla X[m,n] + \nabla N[m,n]$ holds for the points whose neighbors share the same wrap count. +(b) $\mathcal{W}(\nabla Y[m,n]) = \nabla X[m,n] + \nabla N[m,n]$ is applicable for the points that meet the 2D Itoh's continuity condition [13]. + +The majority of points in a GT phase image fall under the aforementioned categories. Therefore, on these points, $\nabla Y$ or $\mathcal{W}(\nabla Y)$ can serve as a noisy label of $\nabla X$ . Once $\nabla X$ + +is predicted, reconstructing $X$ is relatively simple. + +Our approach utilizes $\nabla Y$ for unsupervised training and tackles its label noise via a recorruption-based self-reconstruction loss function in the gradient domain, which is inspired by existing recorruption-based denoisers [19, 21, 46]. In addition, to mitigate the potential ambiguity and overfitting caused by the absence of GTs, a self-distillation loss function is proposed that leverages intermediate predictions from the NN for pseudo-supervision. + +To encode physical model/priors of phase wrapping into the NN-learned PU process, which is beneficial for mitigating possible overfitting particularly when training on a limited amount of data, we develop a deep unrolling network (DUN) by unfolding a variational model that leverages $\mathcal{W}(\nabla Y)$ for data fitting. To our knowledge, this is the first work to utilize deep unrolling for PU. Indeed, most existing studies rely on U-shaped convolutional NNs (CNNs) or recurrent NNs (RNNs) that ignore physical priors of phase wrapping. Additionally, our DUN is tailored to handle the outliers that do not satisfy the Itoh's continuity condition, thereby mitigating their potential adverse effects. + +# Main contributions: Our contributions include + +- We introduce the first external unsupervised DL approach for end-to-end PU, circumventing the challenge of GT acquisition in supervised DL and the computational inefficiency of internal unsupervised DL. +- We perform the first exploration of deep unfolding for PU, founded on a variational model that utilizes wrapped gradients of wrapped phases for data fitting under Itoh's continuity condition. The resulting DUN is outlier-aware and enjoys advantages in interpretability and performance. +- We propose a recorruption-based self-reconstruction loss function with noise tolerance to leverage Itoh's continuity condition, as well as a self-distillation loss function for improved generalization. + +In extensive experiments, our proposed approach shows superior performance over existing GT-free PU methods and competes well against the supervised ones. + +# 2. Related Work + +Non-learning-based PU: Methods along this line fall into three types: reliability-guided, filtering-based, and model-driven. Reliability-guided methods [3, 9, 11, 23, 36, 54] use a manually-designed metric to quantify pixel-wise reliability for path selection or region expansion during phase correction. Due to unreliable quality scores in noisy regions, quality-guided methods are noise-sensitive. Filtering-based methods [2, 45] treat PU as a smoothing problem, leveraging non-linear filters for denoising and unwrapping. However, their limited-support filters fail to exploit global consistency constraints within a phase image. Model-driven methods [1, 8, 12, 15] construct and solve an optimization + +model of PU regularized by some handcrafted priors. While achieving strong global consistency brought by the global optimization, these methods often fall short for complex phase patterns, due to the use of handcrafted priors. + +Supervised DL for PU: Supervised DL-based PU methods differ mainly in NN architectures. Existing methods can be roughly classified as CNN-based or RNN-based. + +Leveraging the effectiveness of CNNs in segmentation and classification, many CNN-based methods (e.g. [10, 34, 35, 49, 50, 52, 56]) cast PU as a wrap count prediction problem, using a pixel-wise classification loss with wrap counts as labels. For robustness to noise, pre-denoising or post-processing modules are often included; see e.g. [34, 35, 52]. However, these methods face scalability issues with widerange phase data, where the candidate values for wrap counts are overwhelmed. Instead, some CNN-based methods [6, 25, 40, 57]. Typically, CNN-based methods are inefficient in exploiting global spatial dependencies. In contrast, the RNN-based methods [22, 32] excel at capturing long-range dependencies. Viewing a phase pixel as an element in a sequence, these methods predefine several paths across all phase pixels to form an RNN. However, due to computational costs, only a few paths can be utilized. This constraint on path selection results in suboptimal outcomes. + +The NN designs of most existing methods overlook the physical priors of phase wrapping, which are beneficial for generalization, especially when training data is limited. Our proposed DUN incorporates these priors, enhancing interpretability, reducing overfitting, and boosting performance. + +Unsupervised DL for PU: Despite significant advances in unsupervised DL for various inverse problems in imaging, e.g., denoising [17, 21, 26], deblurring [5, 18, 27, 31], compressed sensing [28, 38, 42] and phase retrieval [4, 29], unsupervised DL for PU has been much less studied. One related work is Yang et al. [47], which utilizes the deep image prior [39] encoded by an untrained CNN to address the absence of GT data. Specifically, the phase image is reparameterized by a CNN, with its weights learned to fit the measurements. As shown in [39, 47], the CNN prefers smooth patterns and requires early stopping for smooth phase unwrapping, akin to conventional model-based methods but employing a CNN-based prior. A drawback is the need for per-sample retraining, which is computationally intensive. In contrast, our approach allows end-to-end unsupervised training on a dataset, eliminating test-time learning and significantly speeding up test. + +DUNs for imaging: DUNs have been studied for many inverse problems in imaging, e.g., denoising [20, 55], super-resolution [51], deblurring [20], compressed sensing [24, 53] and medical imaging [7]. However, the study on deep unrolling for PU is scant. Our study is the first to develop DUNs for PU, in addition to an unsupervised training scheme. + +# 3. Methodology + +# 3.1. Basics and Keys + +Consider estimating $\mathbf{X}$ from its wrapped version $\mathcal{W}(\mathbf{X}) = \mathbf{X} - 2\pi \mathbf{K}, \mathbf{K} \in \mathbb{Z}^{M \times N}$ . Let $\nabla : \mathbf{Z} \to (\nabla_x \mathbf{Z}; \nabla_y \mathbf{Z})$ denote the spatial gradient operator. By definition, + +$$ +\nabla \mathcal {W} (\boldsymbol {X}) [ m, n ] = \boldsymbol {\nabla} \boldsymbol {X} [ m, n ] + 2 \pi \boldsymbol {\nabla} \boldsymbol {K} [ m, n ]. \tag {3} +$$ + +For any location $[m,n]$ whose adjacent pixels for gradient calculation have the same wrap count, we have + +$$ +\nabla \mathcal {W} (\boldsymbol {X}) [ m, n ] = \nabla \boldsymbol {X} [ m, n ]. \tag {4} +$$ + +In other words, the wrap count can be nullified by the gradient operator on many points. For the points not meeting the condition, we have the following Proposition. + +Proposition 1. [Two-dimensional extension of Itoh's continuity condition [13]] For any point $[m, n]$ satisfying + +$$ +\left\| \nabla \boldsymbol {X} [ m, n ] \right\| _ {\infty} = \max \left\{\left| \nabla_ {x} \boldsymbol {X} [ m, n ] \right|, \left| \nabla_ {y} \boldsymbol {X} [ m, n ] \right| \right\} < \pi , \tag {5} +$$ + +we have that $\mathcal{W}(\nabla \mathcal{W}(\pmb {X}))[m,n] = \nabla \pmb {X}[m,n]$ + +Proof. See supplemental material for the proof. + +Proposition 1 shows that the wrapped gradients of a wrapped phase image are identical to the gradients of the GT, for the points satisfying the 2D Itoh's continuity condition (5) (called Itoh's condition for brevity in the following text). In general, GT phases exhibit gradual changes in most points, and their rapid changes only occur at a few points sparsely distributed over the image. Then, the outlier points not meeting Itoh's condition are sparse. + +Now consider a noisy wrapped image $\mathbf{Y} = \mathcal{W}(\mathbf{X} + \mathbf{N})$ . We can directly extend Eq. (4) to the noisy case: + +$$ +\nabla \boldsymbol {Y} [ m, n ] = \nabla \boldsymbol {X} [ m, n ] + \nabla \boldsymbol {N} [ m, n ], \tag {6} +$$ + +for any location $[m, n]$ in $(\mathbf{X} + \mathbf{N})$ whose adjacent points share the same wrap count. Similarly, we have the following corollary of Proposition 1. + +Corollary 1.1. For points $[m,n]$ satisfying $\| \nabla X[m,n] + \nabla N[m,n]\|_{\infty} < \pi$ , we have + +$$ +\mathcal {W} (\nabla \boldsymbol {Y} [ m, n ]) = \mathcal {W} (\nabla \boldsymbol {X} [ m, n ]) + \nabla \boldsymbol {N} [ m, n ]. \tag {7} +$$ + +It is clear that both $\nabla Y$ and $\mathcal{W}(\nabla Y)$ can act as a noisy version for $\nabla X$ , except on some generally-sparse outlier points that do not satisfy the corresponding conditions. If we can mitigate those noise and outliers to predict $\nabla X$ , the remaining task, reconstructing $X$ from $\nabla X$ , is straightforward, as $X$ can be reconstructed by integrating its spatial gradients $\nabla X$ with appropriate boundary conditions. Indeed, the primary concern in PU revolves around relative phases rather than the absolute ones. The observations described by Eq. (7) and Eq. (6) underpin our DUN and unsupervised loss function, respectively. + +![](images/abeb0ac61fa480e7f289556dfc002ba42f92838af551153f06b61b11a8c7f8d7.jpg) +Figure 2. Architecture of our proposed DUN for PU. + +# 3.2. Deep Unrolling Network for PU + +Our DUN is based on a regularization model inspired by Eq. (7). Though Eq. (7) is inapplicable to the pixels not meeting Itoh's condition, these outliers are typically sparse in a phase image with gradual changes in most areas. Consequently, we introduce a sparse variable $\mathbf{E} \in \mathbb{R}^{M \times N \times 2}$ into the regularization model, which counteracts the errors introduced by those sparse outliers. Specifically, we consider the following regularization model of PU: + +$$ +\min _ {\boldsymbol {X}, \boldsymbol {E}} \| \boldsymbol {\nabla} \boldsymbol {X} - \mathcal {W} (\boldsymbol {\nabla} \boldsymbol {Y}) + \boldsymbol {E} \| _ {\mathrm {F}} ^ {2} + \phi (\boldsymbol {X}) + \psi (\boldsymbol {E}), \quad (8) +$$ + +where $\phi, \psi$ are some regularization functions for the latent phase image $X$ and the error matrix $E$ , respectively. It can be seen that $E$ essentially neutralizes outlier effects, preserving correctness of the data fitting term. + +Our DUN is constructed via the unrolled iterative solver for (8), which reads as follows: for $t = 1,\dots ,T$ + +$$ +\left\{ \begin{array}{l} \pmb {X} ^ {(t)} := \operatorname * {a r g m i n} _ {\pmb {X}} \| \pmb {\nabla} \pmb {X} - (\mathcal {W} (\pmb {\nabla} \pmb {Y}) - \pmb {E} ^ {(t - 1)}) \| _ {\mathrm {F}} ^ {2} + \phi (\pmb {X}), \\ \pmb {E} ^ {(t)} := \operatorname * {a r g m i n} _ {\pmb {E}} \| \pmb {E} - (\mathcal {W} (\pmb {\nabla} \pmb {Y}) - \pmb {\nabla} \pmb {X} ^ {(t)}) \| _ {\mathrm {F}} ^ {2} + \psi (\pmb {E}). \end{array} \right. +$$ + +Accordingly, the DUN predicts $\mathbf{X}$ from $\mathcal{W}(\nabla \mathbf{Y})$ . It alternatively connects two types of sub-NNs that act as the estimators of $\mathbf{X}^{(t)}$ and $\mathbf{E}^{(t)}$ , respectively. See Fig. 2 for an illustration of this architecture. + +Sub-NN for estimating $\boldsymbol{X}^{(t)}$ : The sub-problem for $\boldsymbol{X}^{(t)}$ is solved using proximal gradient descent (PGD), which combines gradient descent with proximal mapping. Considering the limited number of modules (i.e., iteration number $T$ ) in our DUN for computational efficiency, we implement + +multiple gradient descent steps within a stage to bridge the gap. Specifically, Nesterov's acceleration scheme is used to enhance the gradient descent, called AGD. We initialize with $\mathbf{X}_{(0)}^{(t)} = \mathbf{X}^{(t - 1)}, \mathbf{X}^{(0)} = \mathbf{1}, \mathbf{E}^{(0)} = \mathbf{0}$ and $\alpha_{(0)} = 0$ . For $j = 1,\dots ,J$ , + +$$ +\left\{ \begin{array}{l} \boldsymbol {X} _ {(j)} ^ {(t)} = \boldsymbol {V} _ {(j - 1)} ^ {(t)} + \lambda^ {(t)} \mathrm {d i v} \big (\boldsymbol {\nabla} \boldsymbol {V} _ {(j - 1)} ^ {(t)} - (\mathcal {W} (\boldsymbol {\nabla} \boldsymbol {Y}) - \boldsymbol {E} ^ {(t - 1)}) \big), \\ \alpha_ {(j)} = \frac {1}{2} (1 + \sqrt {1 + 4 \alpha_ {(j - 1)} ^ {2}}), \\ \boldsymbol {V} _ {(j)} ^ {(t)} = \boldsymbol {X} _ {(j)} ^ {(t)} + \frac {\alpha_ {(j - 1)} - 1}{\alpha_ {(j)}} (\boldsymbol {X} _ {(j)} ^ {(t)} - \boldsymbol {X} _ {(j - 1)} ^ {(t)}). \end{array} \right. +$$ + +Here the term $\mathrm{div}\bigl (\pmb {\nabla}V_{(j - 1)}^{(t)} - (\mathcal{W}(\pmb {\nabla}\pmb {Y}) - \pmb{E}^{(t - 1)})\bigr)$ corresponds to the gradient $\mathcal{G}_t(\pmb {X}) = \pmb {\nabla}_\pmb {X}\| \pmb {\nabla}\pmb {X} - (\mathcal{W}(\pmb {\nabla}\pmb {Y})-$ $E^{(t - 1)})\|_{\mathrm{F}}^2$ . Afterward, the proximal operator, defined as $\mathrm{Prox}_f(\bar{Z}):= \mathrm{argmin}_{\pmb{Z}}\| \pmb {Z} - \bar{\pmb{Z}}\| _2^2 +f(\pmb {Z})$ , is applied: + +$$ +\boldsymbol {X} ^ {(t)} = \operatorname {P r o x} _ {\phi} \left(\boldsymbol {X} _ {(J)} ^ {(t)}\right). \tag {9} +$$ + +This proximal operator is implemented using a sub-NN with an attentive U-shape structure. As illustrated in Fig. 2, it has a $16 \times$ down-scaling encoder and a corresponding up-scaling decoder, connected by skip connections. At the coarsest scale, a standard self-attention block is used for exploiting global features. The sub-NN also takes the gradient term $\mathcal{G}_t(\boldsymbol{X}_{(J)}^{(t)})$ as an additional input via concatenation. A long skip connection from the input to output is added, enhancing feature flow with a weighted summation. + +Sub-NN for estimating $E^{(t)}$ : The sub-problem for $E^{(t)}$ aims to address the outliers presented in the data fitting term defined on $\mathcal{W}(\nabla Y) - \nabla X^{(t)}$ , guided by the prior encoded in $\psi (E)$ . We employ a sub-NN with $\mathcal{W}(\nabla Y) - \nabla X^{(t)}$ + +as the input to implement the solver of this sub-problem. This sub-NN, illustrated in Fig. 2, is composed of sequential convolutional layers, each followed by a parametric rectified linear unit (PReLU). The sparsity prior of $\pmb{E}$ is induced by attaching a soft thresholding activation function, parametrized by a non-negative parameter $\mathrm{ReLU}(d^{(t)})$ . + +Condition-aware parameter estimation: The noise level in the wrapped phases impacts the optimal selection of the parameters $\{\lambda^{(t)}, w^{(t)}, d^{(t)}\}_{t=1}^T$ . Existing DUNs (e.g. [51]) often include imaging condition values in the input. In our DUN, a condition-aware module (CAM) is used to predict these parameters from noise level, using a chain of fully-connected layers with PReLU or Sigmoid activations. + +# 3.3. Unsupervised End-to-End Training Scheme + +Let $\mathcal{F}$ denote our DUN that maps $\mathcal{W}(\nabla Y)$ to $X$ . Our proposed loss function $\mathcal{L}_{\mathrm{total}}$ for the unsupervised training of $\mathcal{F}$ comprises a noise-resistant self-reconstruction loss $\mathcal{L}_{\mathrm{sr}}$ and a self-distillation loss $\mathcal{L}_{\mathrm{sd}}$ , applied to each stage of the DUN: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {s r}} + \eta \mathcal {L} _ {\mathrm {s d}}, \eta \in \mathbb {R} ^ {+}. \tag {10} +$$ + +Noise-resistant self-reconstruction loss: This loss utilizes $\nabla Y$ as a noisy label for $\nabla X$ , as shown in Eq. (6). To mitigate the impact of label noise, the loss is defined as + +$$ +\mathcal {L} := \mathbb {E} _ {\boldsymbol {U}} \| \nabla \mathcal {F} (\mathcal {W} (\nabla \boldsymbol {Y} + \nabla \boldsymbol {U})) - (\nabla \boldsymbol {Y} - \nabla \boldsymbol {U}) \| _ {\mathrm {F}} ^ {2}, \tag {11} +$$ + +where $U$ is random noise drawn from the distribution of $N$ . The introduction of the noise pair $\{+\nabla U, -\nabla U\}$ is to achieve the resilience to measurement noise in $\nabla Y$ in the training loss, as justified in Proposition 2. + +Proposition 2. Let $\mathbf{Y} = \mathcal{W}(\mathbf{X} + \mathbf{N})$ . Suppose Eq. (6) is satisfied at all points. Assume that $\mathbf{N}, \mathbf{U} \sim \mathcal{P}$ are independent. Then, we have that + +$$ +\mathbb {E} _ {\boldsymbol {Y}} \mathcal {L} = \mathbb {E} _ {\boldsymbol {X}, \boldsymbol {N}, \boldsymbol {U}} \| \boldsymbol {\nabla} \mathcal {F} (\mathcal {W} (\boldsymbol {\nabla} \boldsymbol {Y} + \boldsymbol {\nabla} \boldsymbol {U})) - \boldsymbol {\nabla} \boldsymbol {X} \| _ {F} ^ {2} + C _ {0}, \tag {12} +$$ + +where $C_0$ is a constant. + +Proof. See the proof in supplemental material. + +Proposition 2 shows that if one can independently draw instances $\pmb{U}$ from the same distribution as $N$ , the loss $\mathcal{L}$ effectively simulates the loss supervised by $\nabla X$ , nullifying the label noise as shown in Eq. (12). Note that Proposition 2 requires all points satisfying Eq. (6), which is not generally true. On the outliers, $\nabla Y$ provides wrong supervision with superfluous multiples of $2\pi$ , as seen in Eq. (3). To counteract the impact of those outlier points, we exploit the property $\mathcal{W}(\theta + 2k\pi) = \theta$ to modify the self-reconstruction loss as follows: + +$$ +\mathcal {L} _ {\mathrm {s r}} := \mathbb {E} _ {\boldsymbol {U}} \| \mathcal {W} \left[ \boldsymbol {\nabla} \mathcal {F} \left(\mathcal {W} (\boldsymbol {\nabla} \boldsymbol {Y} + \boldsymbol {\nabla} \boldsymbol {U})\right) - (\boldsymbol {\nabla} \boldsymbol {Y} - \boldsymbol {\nabla} \boldsymbol {U}) \right] \| _ {\mathrm {F}} ^ {2}. \tag {13} +$$ + +That is, an extra $\mathcal{W}$ is applied to the prediction residual within the Frobenius norm, which negates the superfluous multiples of $2\pi$ during the loss calculation. + +Self-distillation loss: Although $\mathcal{L}_{\mathrm{sr}}$ reduces the impact of measurement noise, the ambiguity in solutions due to wrapping may remain, particularly at those outlier points. Fortunately, CNNs including unrolling CNNs [30, 39], show strong induced bias towards regular structures, which aids in reducing such ambiguity during the training with $\mathcal{L}_{\mathrm{sr}}$ . + +To further reduce the ambiguity, we introduce a self-distillation loss defined as + +$$ +\mathcal {L} _ {\mathrm {s d}} := \mathbb {E} _ {\boldsymbol {U}} \| \nabla \mathcal {F} (\mathcal {W} (\nabla \boldsymbol {Y})) - \nabla \bar {\mathcal {F}} (\mathcal {W} (\nabla \boldsymbol {Y} + \nabla \boldsymbol {U})) \| _ {\mathrm {F}} ^ {2}. \tag {14} +$$ + +where $\bar{\mathcal{F}}$ denotes the NN detached from the previous iteration with stop gradient. The loss leverages pseudo supervision from the NN's varied predictions with noise-altered inputs, implicitly aggregating those predictions for self-training. This self-distillation scheme can reduce the NN's prediction variance, enhancing the PU accuracy. Moreover, recall that $\mathcal{L}_{\mathrm{sr}}$ inputs $\mathcal{W}(\nabla Y + \nabla U)$ to $\mathcal{F}$ , which does not fully align with $\mathcal{W}(\nabla Y)$ in test. This inconsistency can be reconciled by $\mathcal{L}_{\mathrm{sd}}$ , as it uses $\mathcal{W}(\nabla Y)$ for the input of $\mathcal{F}$ . + +# 4. Experiments + +Performance evaluation is conducted on both simulated and real-world phase patterns. Following [22], we use NRMSE (Normalized Root Mean Squared Error) as the performance metric. As relative phase is the main focus in PU, the mean of unwrapped phase values is first aligned to that of the GT before the NRMSE calculation. Computational complexity is measured by three metrics: the number of parameters, the number of FLOPs (floating point of operations), and the average test time per sample of size $256 \times 256$ . + +Implementation details: Our proposed approach, called U3Net (U3 for Unsupervised, Unrolling and Unwrapping) for convenience, is implemented in Python and run on a single NVIDIA RTX 3080 GPU. We set $\eta = 0.5$ in $\mathcal{L}_{\mathrm{total}}$ and initialize model weights via kaiming Uniform. The training employs the Adam optimizer with 500 epochs and a batch size of 10. The learning rate is initialized as $1\times 10^{-3}$ and exponentially decayed by a factor of 0.99 at every epoch. + +Methods for comparison: We select eight representative methods of general PU as baselines, including LS [8], QGPU [11], Ryu. et al. [32], PhaseNet2.0 [35], PUDIP [47], SQD-LSTM [22], EESANet [49], and TriNet [37]. Additionally, two representative transformer NNs for general image restoration, UFormer [43] and Restormer [48], are also included for performance comparison. In all the baseline methods, LS and QGPU are conventional non-learning iterative methods, built upon least squares and path selection, respectively; PUDIP is an internal DL-based method; and the others are supervised DL- + +
DatasetMoGRRME#Param. (M)#FLOPs (G)Time (msec.)
SNR(dB)0510203005102030
Non-LearningLS7.912.281.200.370.157.852.501.250.400.12--8.39
QGPU17.122.121.150.360.1117.262.291.200.390.12--14.94
SupervisedRyu. et al.1.341.111.091.051.051.511.051.010.980.931.0721.85303.96
PhaseNet2.08.358.198.078.068.049.298.537.536.956.861.1511.9320.87
SQD-LSTM0.870.710.700.690.681.571.131.121.101.090.904.0713.04
EESANet0.780.770.760.760.741.311.311.251.241.0661.6875.279.85
TriNet5.655.555.465.405.345.405.335.215.105.0513.6165.4811.16
UFormer0.500.470.470.460.450.580.510.490.490.4820.6040.9842.00
Restormer0.450.380.360.360.350.500.430.420.410.413.0217.2344.69
UnsupervisedPUDIP17.5315.167.870.340.1113.107.222.620.380.122.3321.5899695.01
U3Net0.690.250.190.160.101.120.380.270.170.120.748.7710.28
+ +Table 1. Performance comparison on synthetic datasets in terms of NRMSE(%) . Bold for Best and underline for second-best. + +![](images/913f40f396c5ace88c33506d39f066d69e0c7e06faf6b07a14c1edefdca222ff.jpg) +Figure 3. Visualization of PU results on MoGR (top) and RME (bottom), showing residual images to facilitate visual comparison. + +based methods with their NNs retrained on our generated paired phase data. Specifically, as UFormer and Restormer are not originally designed for PU, we train their models using the supervised loss of SQD-LSTM [22] for better performance. + +# 4.1. Evaluation on Simulated Phase Patterns + +Data preparation: Two types of simulated data are generated for training and evaluation. The first type follows [22], where a GT phase image of size $256 \times 256$ is generated through the superposition of a mixture of Gaussians (MoG) with $c$ clusters and randomly selected ramps. The phase range is defined within $[-2p\pi, 2p\pi]$ . The resulting dataset is called MoGR (MoG with Ramp). The noisy wrapped phase images are synthesized via Eq. (1), with $N$ defined as Gaussian white noise. For each GT, only one noisy wrapped phase image is generated. With $c$ and $p$ uniformly selected from $\{1, \dots, 4\}$ and $\{1, \dots, 7\}$ respectively, we generate 5000 noisy wrapped phase images with SNR uniformly selected from $\{0, 5, 10, 20, 30, 60\}$ to construct the training set. Using the same way, 1000 noisy wrapped phase images are generated for test with SNR in $\{0, 5, 10, 20, 30\}$ . + +The other dataset follows a different construction scheme + +named random matrix enlargement [40], thus called RME dataset. Concretely, a GT phase image is formed through enlarging the square matrix of a distribution from an initial size $s \times s$ to size $256 \times 256$ via interpolation. The size parameter $s$ is selected uniformly from $\{2, \dots, 10\}$ , the distribution type is chosen from the {Uniform, Normal} distributions, and the interpolation method is chosen from the {Bilinear, Bicubic} interpolations. Then the phase image is re-scaled to the range varying in $[-2p\pi, 2p\pi]$ with $p$ randomly chosen from $\{1, \dots, 7\}$ . The noisy wrapped phase images are obtained by the same way as the MoGR dataset. + +Results and analysis: Table 1 lists the quantitative results on both datasets, including those in computational complexity. Even trained without GT data, U3Net achieves the best results in 8/10 settings, using a lightweight model. Concretely, U3Net outperforms the two conventional non-learning-based methods, LS and QGPU. In comparison to PUDIP, an unsupervised internal DL-based method, U3Net also shows significant performance gain. The performance advantage of U3Net over these three GT-free methods be + +
SNR(dB)LSQGPURyu. et al.PhaseNet2.0SQD-LSTMEESANetTriNetPU-GANPUNetUFormerRestormerPUDIPU3Net
53.313.131.412.281.762.455.0413.739.591.461.069.991.00
101.841.961.271.691.521.994.6911.849.201.280.935.300.82
200.941.131.241.431.481.794.4611.629.010.970.910.470.46
+ +Table 2. Performance comparison on InSAR phase data in terms of NRMSE(%) . Bold for Best and underline for second-best. + +![](images/397a852b44ad372cb3af2a5dc71bdf6a42355fa060ed9de63d0fdbf20df78d5a.jpg) +Figure 4. Visualization of unwrapped phase images on InSAR data, showing residual images to facilitate visual comparison. + +comes more noticeable for lower SNR (i.e. heavier noise), particularly compared to PUDIP. This is because U3Net has an explicit mechanism to deal with noise during training, thereby achieving higher noise robustness. In addition, U3Net is much faster than PUDIP, due to that U3Net employs end-to-end pre-trained models, avoiding the time-costly per-sample training of PUDIP. + +Surprisingly, U3Net also surpasses the supervised DL-based methods in many settings. In the lowest-SNR case, U3Net performs worse than UFormer and Restormer, two supervised transformer models. This is probably due to that severe noise increases the number of outlier points, lowering the effectiveness of the training function. Even that, U3Net still outperforms other supervised methods. The model of U3Net is the smallest among all the DL-based methods, indicating that the excellent performance of U3Net is mainly attributed to its interpretable physics-encoded NN architecture and its effective training scheme, rather than the use of larger models. Meanwhile, U3Net is computationally efficient, with the second smallest number of FLOPs and the third fastest inference speed. + +See Fig. 3 for the visualization of PU results from different methods. The residuals to GTs produced by U3Net are much more insignificant compared to the internal DL-based PUDIP. In comparison to supervised methods, U3Net outperforms all of them on the image from RME dataset and shows comparable results on the image of MoGR dataset. + +# 4.2. Evaluation on InSAR Phase Patterns + +The InSAR phase data are generated by converting the digital elevation model at Texas, U.S. into absolute phases, + +following [33]. Totally, 4000 (1000) GT phase images are obtained by cropping the $256 \times 256$ patches from the absolute phases for training (test). Each phase value is scaled into $[-6\pi, 6\pi]$ . The noisy wrapped phases are generated by adding Gaussian noise following the experiments in Sec. 4.1, and we test on three noise levels: 5, 10, 20. In addition to the previously-compared general PU methods and general-purpose transformer NNs, we also compare with two specific methods designed for InSAR PU: PUNet [44] and PU-GAN [57]. All the compared models are retrained using the same data as ours. + +See Table 2 for the quantitative comparison on this In-SAR dataset. Our U3Net ranks the first among all compared methods over all noise settings, noticeably surpassing other GT-free methods as well as many supervised methods in many cases. The visual comparison provided in Fig. 4 also shows the advantage of U3Net. All these results further verify the effectiveness of our approach in dealing with phase patterns from real scenes. + +# 4.3. Ablation Studies + +We form some baselines from U3Net for ablation studies, which are retrained and tested on the MoGR dataset. Their results are summarized in Table 3. + +Study on loss function: We design the following schemes to analyze the loss function: (a) $\mathcal{L}_{\mathrm{sr}} \to \mathcal{L}$ : remove the outer $\mathcal{W}$ applied to the residual, from the self-reconstruction loss $\mathcal{L}_{\mathrm{sr}}$ , i.e., replacing $\mathcal{L}_{\mathrm{sr}}$ by the loss $\mathcal{L}$ defined in Eq. (11); (b) w/o $\nabla U$ : remove the noise re-corruption mechanism in $\mathcal{L}_{\mathrm{sr}}$ by discarding the $\pm \nabla U$ terms; (c) $\mathcal{L}_{\mathrm{sr}} \to \mathcal{L}$ w/o $\nabla U$ : combine the former cases, yielding a plain gradient loss + +
SettingSNR=0SNR=5SNR=10SNR=20SNR=30
Lsr→L25.4523.4823.2623.2323.20
w/o ∇U6.162.401.360.690.60
Lsr→L w/o ∇U25.8523.5723.4223.3823.34
w/o Lsd1.140.330.230.220.17
Supervised0.290.160.130.110.05
w/o E7.160.750.440.360.31
AGD→GD2.140.350.210.170.12
w/o CAM0.980.310.220.210.15
W(∇Y)→∇Y2.881.741.501.461.39
DUN→UFormer1.551.411.311.301.30
U3Net (original)0.690.250.190.160.10
+ +Table 3. Results in ablation studies in terms of NRMSE(%) + +![](images/595e81c47fec259ebc2b19f954dc73e1904385058746d70dd9a5601d4749c892.jpg) +Figure 5. Visualization of unwrapped phase images using models trained by different loss functions. $\mathcal{L}_{\mathrm{sr}} \rightarrow \mathcal{L}$ and w/o $\nabla U$ produce noticeable errors than Original. Zoom in for a better view. + +$\| \nabla \mathcal{F}(\mathcal{W}(\nabla Y)) - \nabla Y\|_{\mathrm{F}}^2$ (d) w/o $\mathcal{L}_{\mathrm{sd}}$ : remove the selfdistillation loss $\mathcal{L}_{\mathrm{sd}}$ , only employing $\mathcal{L}_{\mathrm{sr}}$ for training; and (e) Supervised: train with a gradient-domain supervised loss $\| \nabla \mathcal{F}(\mathcal{W}(\nabla Y)) - \nabla X\|_{\mathrm{F}}^2$ + +In addition to Table 3, we also provide some visual comparisons in Fig. 5 for demonstration. From both the quantitative and qualitative results, we have following observations. (a) The outer $\mathcal{W}$ in $\mathcal{L}_{\mathrm{sr}}$ is critical to the performance of U3Net. (b) The re-corruption mechanism in $\mathcal{L}_{\mathrm{sr}}$ does help eliminate the measurement noise. (c) A simple gradient loss function does not work, with further performance decrease. (d) While the self-reconstruction loss already yields promising results, the self-distillation loss benefits further performance gain. (e) U3Net performs closely to its supervised counterpart when $\mathrm{SNR} \geq 5$ . We also see that U3Net has its supervised counterpart outperforming the supervised UFormer reported in Table 1, demonstrating its superiority. + +Study on DUN: To analyze the components in the DUN architecture of U3Net, we design the following baseline models: (a) w/o $\pmb{E}$ : abandon the error term $\pmb{E}$ during unrolling, forming a DUN alternating the accelerated gradient descent step and the sub-NN for estimating $\pmb{X}^{(t)}$ ; (b) AGD→GD: replace the Nesterov's AGD steps by one-step standard gradient; (c) w/o CAM: remove the CAM from U3Net; (d) + +![](images/6c81c67944f6bb3c22d8d09aff4f14d9fe2a38190002fd4075c37d807c26cec9.jpg) +Figure 6. Visualization of error matrices $\pmb{E}$ estimated from three wrapped phase images. The outlier maps use white pixels to indicate the points $[x,y]$ satisfying $\| \nabla X[x,y] + \nabla N[x,y]\|_{\infty} \geq \pi$ and black pixels otherwise. + +$\mathcal{W}(\nabla \pmb {Y})\rightarrow \nabla \pmb {Y}$ : use $\nabla Y$ to replace $\mathcal{W}(\nabla Y)$ in the input of U3Net as well in the total loss; (e) DUN $\rightarrow$ UFormer: replace the DUN by UFormer, keeping the loss unchanged. + +As demonstrated in Table 3, each component in the DUN of U3Net is useful. (a) One critical part is the introduction of the error matrix $\pmb{E}$ for unrolling. It brings significant performance gain, particularly in dealing with low-SNR data, as the absorption of outliers breaking the Itoh's condition is more desired for heavier noise. See Fig. 6 for a visualization of estimated error matrices, which successfully identify the outliers. (b)&(c) The AGD and CAM have similar contribution to the performance of U3Net. (d) Using $\mathcal{W}(\nabla Y)$ for DUN is also crucial, serving as a better choice compared to using $\nabla Y$ . (e) Our proposed loss function also works for UFormer, but yielding worse results than U3Net, which indicates the higher effectiveness of our physics-encoded DUN for unsupervised PU, compared to general transformer models. + +# 5. Conclusion + +This work proposed a fully unsupervised DL approach for PU, eliminating the dependence on GTs in end-to-end training. Utilizing the connection between (wrapped) gradients of wrapped phases and gradients of original phases, we constructed a physics-encoded DUN, trained via a noise-resistant self-reconstruction loss complemented by a self-distillation loss. Empirical evaluations confirmed that our approach surpassed existing GT-free methods and rivaled supervised counterparts. Similar to current unsupervised and synthetic data-dependent supervised techniques, our approach presumes known noise statistics. Future research will focus on dealing with unknown noise profiles. + +# References + +[1] Liheng Bian, Xin Wang, Daoyu Li, Qiuling Ren, and Dezhi Zheng. Robust phase unwrapping via non-local regularization. Optics Letters, 48(6):1399-1402, 2023. 1, 2 +[2] David Blinder, Heidi Ottevaere, Adrian Munteanu, and Peter Schelkens. Efficient multiscale phase unwrapping methodology with modulo wavelet transform. Optics Express, 24(20): 23094-23108, 2016. 1, 2 +[3] Donald J Bone. Fourier fringe analysis: the two-dimensional phase unwrapping problem. Applied Optics, 30(25):3627-3632, 1991. 1, 2 +[4] Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate mmse estimation. IEEE Transactions on Signal Processing, 70:2239-2252, 2022. 3 +[5] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 3 +[6] Gili Dardikman and Natan T Shaked. Phase unwrapping using residual neural networks. In Computational Optical Sensing and Imaging, pages CW3B-5. Optica Publishing Group, 2018. 3 +[7] Qiaoqiao Ding, Gaoyu Chen, Xiaoqun Zhang, Qiu Huang, Hui Ji, and Hao Gao. Low-dose ct with deep learning regularization via proximal forward-backward splitting. Physics in Medicine & Biology, 65(12):125009, 2020. 3 +[8] Dennis C Ghiglia and Louis A Romero. Robust two-dimensional weighted and unweighted phase unwrapping that uses fast transforms and iterative methods. Journal of the Optical Society of America A, 11(1):107-117, 1994. 2, 5 +[9] Richard M Goldstein, Howard A Zebker, and Charles L Werner. Satellite radar interferometry: Two-dimensional phase unwrapping. Radio Science, 23(4):713-720, 1988. 1, 2 +[10] Michal Gontarz, Vibekananda Dutta, Malgorzata Kujawinska, and Wojciech Krauze. Phase unwrapping using deep learning in holographic tomography. Optics Express, 31(12):18964-18992, 2023. 1, 3 +[11] Miguel Arevallilo Herraez, David R Burton, Michael J Lolor, and Munther A Gdeisat. Fast two-dimensional phaseunwrapping algorithm based on sorting by reliability following a noncontinuous path. Applied Optics, 41(35):7437-7444, 2002. 1, 2, 5 +[12] Howard YH Huang, L Tian, Z Zhang, Y Liu, Z Chen, and G Barbastathis. Path-independent phase unwrapping using phase gradient and total-variation (tv) denoising. Optics Express, 20(13):14075–14089, 2012. 1, 2 +[13] Kazuyoshi Itoh. Analysis of the phase unwrapping algorithm. Applied Optics, 21(14):2470-2470, 1982. 2, 3 +[14] Charles VJ Jakowitz, Daniel E Wahl, Paul H Eichel, Dennis C Ghiglia, and Paul A Thompson. Spotlight-mode synthetic aperture radar: a signal processing approach. Springer Science & Business Media, 2012. 1 +[15] Ulugbek S Kamilov, Ioannis N Papadopoulos, Morteza H Shoreh, Demetri Psaltis, and Michael Unser. Isotropic + +inverse-problem approach for two-dimensional phase unwrapping. Journal of the Optical Society of America A, 32 (6):1092-1100, 2015. 1, 2 +[16] Thomas Kreis. Handbook of holographic interferometry: optical and digital methods. John Wiley & Sons, 2006. 1 +[17] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2137, 2019. 3 +[18] Ji Li, Weixi Wang, Yuesong Nan, and Hui Ji. Self-supervised blind motion deblurring with deep expectation maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13986-13996, 2023. 3 +[19] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 2 +[20] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 3 +[21] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2043-2052, 2021. 2, 3 +[22] Malsha V Perera and Ashwin De Silva. A joint convolutional and spatial quad-directional LSTM network for phase unwrapping. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4055-4059. IEEE, 2021. 1, 3, 5, 6 +[23] Mark D Pritt. Phase unwrapping by means of multigrid techniques for interferometric sar. IEEE Transactions on Geoscience and Remote Sensing, 34(3):728-738, 1996. 1, 2 +[24] Xinran Qin, Yuhui Quan, and Hui Ji. Enhanced deep unrolling networks for snapshot compressive hyperspectral imaging. Neural Networks, 174:106250, 2024. 3 +[25] Yi Qin, Shujia Wan, Yuhong Wan, Jiawen Weng, Wei Liu, and Qiong Gong. Direct and accurate phase unwrapping with deep neural network. Applied Optics, 59(24):7258-7267, 2020. 3 +[26] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1898, 2020. 3 +[27] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In European Conference on Computer Vision, pages 642-659. Springer, 2022. 3 +[28] Yuhui Quan, Xinran Qin, Tongyao Pang, and Hui Ji. Dual-domain self-supervised learning and model adaption for deep compressive imaging. In European Conference on Computer Vision, pages 409-426. Springer, 2022. 3 +[29] Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student + +distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 3 +[30] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5 +[31] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3341-3350, 2020. 3 +[32] Kanghyun Ryu, Sung-Min Gho, Yoonho Nam, Kevin Koch, and Dong-Hyun Kim. Development of a deep learning method for phase unwrapping mr images. In Proceedings of the International Society for Magnetic Resonance in Medicine, page 4707, 2019. 3, 5 +[33] Francescopaolo Sica, Francesco Calvanese, Giuseppe Scarpa, and Paola Rizzoli. A cnn-based coherence-driven approach for insar phase unwrapping. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2020. 7 +[34] GE Spoorthi, Subrahmanyam Gorthi, and Rama Krishna Sai Subrahmanyam Gorthi. Phasenet: A deep convolutional neural network for two-dimensional phase unwrapping. IEEE Signal Processing Letters, 26(1):54-58, 2018. 3 +[35] GE Spoorthi, Rama Krishna Sai Subrahmanyam Gorthi, and Subrahmanyam Gorthi. Phasenet 2.0: Phase unwrapping of noisy data based on deep learning approach. IEEE Transactions on Image Processing, 29:4862-4872, 2020. 3, 5 +[36] Xianyu Su and Wenjing Chen. Reliability-guided phase unwrapping algorithm: a review. Optics and Lasers in Engineering, 42(3):245-261, 2004. 1, 2 +[37] Krishna Sumanth, Vaishnavi Ravi, and Rama Krishna Gorthi. A multi-task learning for 2d phase unwrapping in fringe projection. IEEE Signal Processing Letters, 29:797-801, 2022. 5 +[38] Julián Tachella, Dongdong Chen, and Mike Davies. Unsupervised learning from incomplete measurements for inverse problems. Advances in Neural Information Processing Systems, 35:4983-4995, 2022. 3 +[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 3, 5 +[40] Kaiqiang Wang, Ying Li, Qian Kemao, Jianglei Di, and Jianlin Zhao. One-step robust deep learning phase unwrapping. Optics Express, 27(10):15100-15115, 2019. 3, 6 +[41] Kaiqiang Wang, Qian Kemao, Jianglei Di, and Jianlin Zhao. Deep learning spatial phase unwrapping: a comparative review. Advanced Photonics Nexus, 1(1):014001, 2022. 1, 2 +[42] Weixi Wang, Ji Li, and Hui Ji. Self-supervised deep image restoration via adaptive stochastic gradient Langevin dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2022. 3 +[43] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general + +u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 5 +[44] Zhipeng Wu, Teng Wang, Yingjie Wang, Robert Wang, and Daqing Ge. Deep learning for the detection and phase unwrapping of mining-induced deformation in large-scale interferograms. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2021. 7 +[45] Xianming Xie. Iterated unscented kalman filter for phase unwrapping of interferometric fringes. Optics Express, 24 (17):18872-18897, 2016. 1, 2 +[46] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 2 +[47] Fangshu Yang, Thanh-An Pham, Nathalie Brandenberg, Matthias P Lütolf, Jianwei Ma, and Michael Unser. Robust phase unwrapping via deep image prior for quantitative phase imaging. IEEE Transactions on Image Processing, 30: 7025-7037, 2021. 1, 2, 3, 5 +[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5 +[49] Junkang Zhang and Qingguang Li. Eesanet: edge-enhanced self-attention network for two-dimensional phase unwrapping. Optics Express, 30(7):10470-10490, 2022. 3, 5 +[50] Junchao Zhang, Xiaobo Tian, Jianbo Shao, Haibo Luo, and Rongguang Liang. Phase unwrapping in optical metrology via denoised and convolutional segmentation networks. Optics Express, 27(10):14903-14912, 2019. 3 +[51] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3217-3226, 2020. 3, 5 +[52] Teng Zhang, Shaowei Jiang, Zixin Zhao, Krishna Dixit, Xiaofei Zhou, Jia Hou, Yongbing Zhang, and Chenggang Yan. Rapid and robust two-dimensional phase unwrapping via deep learning. Optics Express, 27(16):23173-23185, 2019. 3 +[53] Zhonghao Zhang, Yipeng Liu, Jiani Liu, Fei Wen, and Ce Zhu. Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing, 30:1487-1500, 2020. 3 +[54] Ming Zhao, Lei Huang, Qican Zhang, Xianyu Su, Anand Asundi, and Qian Kemo. Quality-guided phase unwrapping technique: comparison of quality maps and guiding strategies. Applied Optics, 50(33):6214-6224, 2011. 2 +[55] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 630-641, 2021. 3 +[56] Hongyu Zhou, Chuanli Cheng, Hao Peng, Dong Liang, Xin Liu, Hairong Zheng, and Chao Zou. The phu-net: A robust phase unwrapping method for mri based on deep learning. + +Magnetic Resonance in Medicine, 86(6):3321-3333, 2021. 1, 3 +[57] Lifan Zhou, Hanwen Yu, Vito Pascazio, and Mengdao Xing. Pu-gan: a one-step 2-d insar phase unwrapping based on conditional generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 60:1-10, 2022. 1, 3, 7 \ No newline at end of file diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/images.zip b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d533be61bffb31287ac8d7afd6a9ab24b03a9df --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f03c6b6dd35d816134a2f76b4198e4abf0b321c9bc610791d214ab1d089f5463 +size 743437 diff --git a/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/layout.json b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c004c021e57d2222092804884b8a5c949cd5091 --- /dev/null +++ b/2024/Unsupervised Deep Unrolling Networks for Phase Unwrapping/layout.json @@ -0,0 +1,11349 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 103, + 103, + 492, + 121 + ], + "type": "text", + "content": "Unsupervised Deep Unrolling Networks for Phase Unwrapping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "spans": [ + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "text", + "content": "Zhile Chen" + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "text", + "content": " Yuhui Quan" + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "text", + "content": " Hui Ji" + }, + { + "bbox": [ + 175, + 142, + 417, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "type": "text", + "content": "School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China " + }, + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 159, + 527, + 185 + ], + "type": "text", + "content": "Pazhou Lab, Guangzhou 510320, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 186, + 465, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 186, + 465, + 200 + ], + "spans": [ + { + "bbox": [ + 127, + 186, + 465, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 186, + 465, + 200 + ], + "type": "text", + "content": "Department of Mathematics, National University of Singapore, 119076, Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 201, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 201, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 108, + 201, + 481, + 213 + ], + "type": "text", + "content": "cs.zhilechen@mail.scut.edu.cn, csyhquan@scut.edu.cn, matjh@nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 517 + ], + "type": "text", + "content": "Phase unwrapping (PU) is a technique to reconstruct original phase images from their noisy wrapped counterparts, finding many applications in scientific imaging. Although supervised learning has shown promise in PU, its utility is limited in ground-truth (GT) scarce scenarios. This paper presents an unsupervised learning approach that eliminates the need for GTs during end-to-end training. Our approach leverages the insight that both the gradients and wrapped gradients of wrapped phases serve as noisy labels for GT phase gradients, along with sparse outliers induced by the wrapping operation. A recorruption-based self-reconstruction loss in the gradient domain is proposed to mitigate the adverse effects of label noise, complemented with a self-distillation loss for improved generalization. Additionally, by unfolding a variational model of PU that utilizes wrapped gradients of wrapped phases for its data-fitting term, we develop a deep unrolling network that encodes physics of phase wrapping and incorporates special treatments on outliers. In the experiments on three types of phase data, our approach outperforms existing GT-free methods and competes well against the supervised ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 539, + 127, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 127, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 127, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "text", + "content": "In many imaging systems, the direct acquisition of original phases of a target signal is both challenging and costly. Typically, initial measurement yields a phase image wrapped in " + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "inline_equation", + "content": "[- \\pi, \\pi)" + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{M \\times N}" + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "inline_equation", + "content": "Y \\in [-\\pi, \\pi)^{M \\times N}" + }, + { + "bbox": [ + 46, + 558, + 287, + 631 + ], + "type": "text", + "content": " denote the true phase image and its noisy wrapped counterpart, respectively. The wrapping process can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 635, + 287, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 287, + 651 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} = \\mathcal {W} (\\boldsymbol {X} + \\boldsymbol {N}), \\quad \\mathcal {W}: \\mathbb {R} ^ {M \\times N} \\rightarrow [ - \\pi , \\pi) ^ {M \\times N}, \\tag {1}", + "image_path": "2d7be6da8374cd42eea6deca69dd23f563cd45201bbaf32f090aab6d0127dc7b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "content": " denotes the wrapping operator with its entry-wise operation defined as " + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\theta) = ((\\theta + \\pi) \\bmod 2\\pi) - \\pi" + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "inline_equation", + "content": "N \\in \\mathbb{R}^{M \\times N}" + }, + { + "bbox": [ + 304, + 242, + 546, + 338 + ], + "type": "text", + "content": " denotes the measurement noise, typically assumed to be Gaussian [22]. Consequently, the original phase image, which exhibits strong smoothness, is degraded to a wrapped version with both incorrect values and artificial discontinuities induced by the modulo operation, which adversely affects subsequent processing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "text", + "content": "Phase unwrapping (PU) is about recovering the original unwrapped phase image " + }, + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "text", + "content": " (up to an additive constant) from its noisy wrapped counterpart " + }, + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 338, + 546, + 457 + ], + "type": "text", + "content": ". Serving as a critical step in many optical imaging techniques for generating clear and coherent phase images, PU finds extensive applications in diverse domains such as quantitative phase imaging, magnetic resonance imaging, synthetic aperture radar interferometry, 3D depth sensing, phase contrast microscopy, fringe projection, and digital holographic interferometry; see e.g. [10, 14, 16, 23, 41, 47, 56, 57]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 457, + 529, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 457, + 529, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 457, + 529, + 469 + ], + "type": "text", + "content": "The model (1) is also usually formulated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 346, + 475, + 545, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 475, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 346, + 475, + 545, + 489 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} = \\boldsymbol {X} - 2 \\pi \\boldsymbol {K} + \\boldsymbol {N} ^ {\\prime}, \\quad \\boldsymbol {K} \\in \\mathbb {Z} ^ {M \\times N}, \\tag {2}", + "image_path": "38f63bc39168fc58049fe39f42daebbe83e8a5d7d75e5c5f546f4786f51e914f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{Z}^{M \\times N}" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": " denotes the map of wrap counts, indicating the number of times a phase value has been wrapped around by " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": " denotes the noise dependent on both " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": ", whose distribution can be complex even when " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": " follows a simple distribution such as the normal distribution. It can be seen that PU is a challenging ill-posed inverse problem involving a continuous variable " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": " and a discrete variable " + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 495, + 547, + 628 + ], + "type": "text", + "content": ". Direct solutions by integrating wrapped phase differences are insufficient for ensuring spatial consistency in all directions of the unwrapped phases and may accumulate significant errors due to substantial noise." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Limitations of existing works: Conventional approaches for PU include reliability-guided methods (e.g. [3, 9, 11, 23, 36]), filtering-based methods (e.g. [2, 45]), and model-driven methods (e.g. [1, 12, 15]). These approaches are either sensitive to measurement noise, due to their heuristic rules, or overly simplistic for complex phase structures, due to their handcrafted image priors. In contrast, deep learning" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 655, + 179, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 655, + 179, + 666 + ], + "spans": [ + { + "bbox": [ + 58, + 655, + 179, + 666 + ], + "type": "text", + "content": "* Corresponding author: Yuhui Quan." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 62372186), Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755, 2023A1515012841), Fundamental Research Funds for Central Universities (Grant No. x2jsD2230220), and Singapore MOE AcRF Tier 1 (Grant No. A-8000981-00-00)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25182" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 126, + 137 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 126, + 137 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 126, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 126, + 137 + ], + "type": "image", + "image_path": "e5d1b7d488142b56bf5730f32ea270b2b72b03793c60e50c6631347304ee3c2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 85, + 192, + 135 + ], + "blocks": [ + { + "bbox": [ + 152, + 70, + 181, + 84 + ], + "lines": [ + { + "bbox": [ + 152, + 70, + 181, + 84 + ], + "spans": [ + { + "bbox": [ + 152, + 70, + 181, + 84 + ], + "type": "text", + "content": "Randomly generated Set" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 85, + 192, + 135 + ], + "lines": [ + { + "bbox": [ + 141, + 85, + 192, + 135 + ], + "spans": [ + { + "bbox": [ + 141, + 85, + 192, + 135 + ], + "type": "image", + "image_path": "568acda83cbd389b33f62b98b25702fd3e47a9cbb7758f873558114cc829e178.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 79, + 270, + 137 + ], + "blocks": [ + { + "bbox": [ + 208, + 71, + 271, + 79 + ], + "lines": [ + { + "bbox": [ + 208, + 71, + 271, + 79 + ], + "spans": [ + { + "bbox": [ + 208, + 71, + 271, + 79 + ], + "type": "text", + "content": "Wrapped Phase Images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 79, + 270, + 137 + ], + "lines": [ + { + "bbox": [ + 209, + 79, + 270, + 137 + ], + "spans": [ + { + "bbox": [ + 209, + 79, + 270, + 137 + ], + "type": "image", + "image_path": "4408d9a135bed6a749e9f4540b0b86708297caeb14ddeceebc847a3b081ec5dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 141, + 123, + 210 + ], + "blocks": [ + { + "bbox": [ + 62, + 141, + 123, + 210 + ], + "lines": [ + { + "bbox": [ + 62, + 141, + 123, + 210 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 123, + 210 + ], + "type": "image", + "image_path": "d53fd8067bf8677d978b49aba8118a13484dcbc1d67b863b9c35a4f787d6a00f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 216, + 123, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 216, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 123, + 224 + ], + "type": "text", + "content": "Supervised Learning" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 229, + 127, + 237 + ], + "lines": [ + { + "bbox": [ + 61, + 229, + 127, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 127, + 237 + ], + "type": "text", + "content": "- GT collection is hard" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 47, + 247, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 286, + 258 + ], + "type": "text", + "content": "Figure 1. Training data setting of existing PU methods versus ours." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 157, + 192, + 206 + ], + "blocks": [ + { + "bbox": [ + 136, + 141, + 197, + 156 + ], + "lines": [ + { + "bbox": [ + 136, + 141, + 197, + 156 + ], + "spans": [ + { + "bbox": [ + 136, + 141, + 197, + 156 + ], + "type": "text", + "content": "Wrappered Phase Image (Single Test Sample)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 157, + 192, + 206 + ], + "lines": [ + { + "bbox": [ + 141, + 157, + 192, + 206 + ], + "spans": [ + { + "bbox": [ + 141, + 157, + 192, + 206 + ], + "type": "image", + "image_path": "9658c25f4c65824b637632f84c479f946678ea5a960627a6cf26b955fe309a78.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 216, + 192, + 224 + ], + "lines": [ + { + "bbox": [ + 140, + 216, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 216, + 192, + 224 + ], + "type": "text", + "content": "Internal Learning" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 229, + 195, + 238 + ], + "lines": [ + { + "bbox": [ + 138, + 229, + 195, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 195, + 238 + ], + "type": "text", + "content": "Time consuming" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 209, + 150, + 269, + 208 + ], + "blocks": [ + { + "bbox": [ + 209, + 141, + 271, + 149 + ], + "lines": [ + { + "bbox": [ + 209, + 141, + 271, + 149 + ], + "spans": [ + { + "bbox": [ + 209, + 141, + 271, + 149 + ], + "type": "text", + "content": "Wrapped Phase Images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 150, + 269, + 208 + ], + "lines": [ + { + "bbox": [ + 209, + 150, + 269, + 208 + ], + "spans": [ + { + "bbox": [ + 209, + 150, + 269, + 208 + ], + "type": "image", + "image_path": "492b3f000e6d48e83b7fc1417bc5f9d74aa02c3e93ec7733f559883528364f94.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 211, + 258, + 218 + ], + "lines": [ + { + "bbox": [ + 224, + 211, + 258, + 218 + ], + "spans": [ + { + "bbox": [ + 224, + 211, + 258, + 218 + ], + "type": "text", + "content": "End-to-End" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 220, + 276, + 228 + ], + "lines": [ + { + "bbox": [ + 208, + 220, + 276, + 228 + ], + "spans": [ + { + "bbox": [ + 208, + 220, + 276, + 228 + ], + "type": "text", + "content": "Unsupervised Learning" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 229, + 276, + 237 + ], + "lines": [ + { + "bbox": [ + 208, + 229, + 276, + 237 + ], + "spans": [ + { + "bbox": [ + 208, + 229, + 276, + 237 + ], + "type": "text", + "content": "Bypassing both issues" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 279, + 286, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 286, + 316 + ], + "type": "text", + "content": "(DL) with neural networks (NNs) has emerged as a powerful tool for PU [41]. A pre-trained end-to-end NN offers data adaptivity, rapid inference, and excellent expressivity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 316, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 286, + 459 + ], + "type": "text", + "content": "Existing DL-based PU methods often rely on supervised learning with ground-truth (GT) phases. However, acquiring GT phases is often impractical due to physical limits in imaging, destructive imaging techniques, and ethical/safety concerns. This challenge limits the applicability of supervised DL in many practical scenarios. Internal unsupervised DL-based methods (e.g. [47]) use an untrained NN to parameterize latent phases, adjusting it to fit observed measurements with certain regularization. Though flexible, this kind of methods requires time-costly per-sample test-time training, resulting in poor computational efficiency compared to supervised DL-based methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 461, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 557 + ], + "type": "text", + "content": "Our approach: To bypass the difficulty of GT acquisition in supervised DL and avoid the computational inefficiency of internal unsupervised DL, we present an external unsupervised " + }, + { + "bbox": [ + 46, + 461, + 286, + 557 + ], + "type": "inline_equation", + "content": "DL" + }, + { + "bbox": [ + 46, + 461, + 286, + 557 + ], + "type": "text", + "content": " approach for PU. It trains an end-to-end NN solely on a dataset of noisy wrapped measurements. See Fig. 1 for the comparison in training data configuration between existing works and ours. Indeed, our study is the first one on external (end-to-end) unsupervised DL for PU." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 558, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 616 + ], + "type": "text", + "content": "Our approach is inspired by the insight that applying the gradient operator or the wrapping gradient operator to a wrapped phase image will result in an output identical to the gradient map of the GT phase image, with the exception of a sparse outlier set. Concretely, we have that" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 50, + 617, + 286, + 677 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 50, + 617, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 617, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 50, + 617, + 286, + 641 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 50, + 617, + 286, + 641 + ], + "type": "inline_equation", + "content": "\\nabla Y[m,n] = \\nabla X[m,n] + \\nabla N[m,n]" + }, + { + "bbox": [ + 50, + 617, + 286, + 641 + ], + "type": "text", + "content": " holds for the points whose neighbors share the same wrap count." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 50, + 641, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 641, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 50, + 641, + 286, + 677 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 50, + 641, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y[m,n]) = \\nabla X[m,n] + \\nabla N[m,n]" + }, + { + "bbox": [ + 50, + 641, + 286, + 677 + ], + "type": "text", + "content": " is applicable for the points that meet the 2D Itoh's continuity condition [13]." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The majority of points in a GT phase image fall under the aforementioned categories. Therefore, on these points, " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " can serve as a noisy label of " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": ". Once " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\nabla X" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 72, + 510, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 510, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 510, + 84 + ], + "type": "text", + "content": "is predicted, reconstructing " + }, + { + "bbox": [ + 305, + 72, + 510, + 84 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 72, + 510, + 84 + ], + "type": "text", + "content": " is relatively simple." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 85, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 180 + ], + "type": "text", + "content": "Our approach utilizes " + }, + { + "bbox": [ + 304, + 85, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 304, + 85, + 545, + 180 + ], + "type": "text", + "content": " for unsupervised training and tackles its label noise via a recorruption-based self-reconstruction loss function in the gradient domain, which is inspired by existing recorruption-based denoisers [19, 21, 46]. In addition, to mitigate the potential ambiguity and overfitting caused by the absence of GTs, a self-distillation loss function is proposed that leverages intermediate predictions from the NN for pseudo-supervision." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 182, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 326 + ], + "type": "text", + "content": "To encode physical model/priors of phase wrapping into the NN-learned PU process, which is beneficial for mitigating possible overfitting particularly when training on a limited amount of data, we develop a deep unrolling network (DUN) by unfolding a variational model that leverages " + }, + { + "bbox": [ + 304, + 182, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 304, + 182, + 545, + 326 + ], + "type": "text", + "content": " for data fitting. To our knowledge, this is the first work to utilize deep unrolling for PU. Indeed, most existing studies rely on U-shaped convolutional NNs (CNNs) or recurrent NNs (RNNs) that ignore physical priors of phase wrapping. Additionally, our DUN is tailored to handle the outliers that do not satisfy the Itoh's continuity condition, thereby mitigating their potential adverse effects." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 328, + 503, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 328, + 503, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 503, + 339 + ], + "type": "text", + "content": "Main contributions: Our contributions include" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 342, + 545, + 496 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 306, + 342, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 389 + ], + "type": "text", + "content": "- We introduce the first external unsupervised DL approach for end-to-end PU, circumventing the challenge of GT acquisition in supervised DL and the computational inefficiency of internal unsupervised DL." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 389, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 545, + 449 + ], + "type": "text", + "content": "- We perform the first exploration of deep unfolding for PU, founded on a variational model that utilizes wrapped gradients of wrapped phases for data fitting under Itoh's continuity condition. The resulting DUN is outlier-aware and enjoys advantages in interpretability and performance." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "type": "text", + "content": "- We propose a recorruption-based self-reconstruction loss function with noise tolerance to leverage Itoh's continuity condition, as well as a self-distillation loss function for improved generalization." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 499, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 499, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 545, + 535 + ], + "type": "text", + "content": "In extensive experiments, our proposed approach shows superior performance over existing GT-free PU methods and competes well against the supervised ones." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Non-learning-based PU: Methods along this line fall into three types: reliability-guided, filtering-based, and model-driven. Reliability-guided methods [3, 9, 11, 23, 36, 54] use a manually-designed metric to quantify pixel-wise reliability for path selection or region expansion during phase correction. Due to unreliable quality scores in noisy regions, quality-guided methods are noise-sensitive. Filtering-based methods [2, 45] treat PU as a smoothing problem, leveraging non-linear filters for denoising and unwrapping. However, their limited-support filters fail to exploit global consistency constraints within a phase image. Model-driven methods [1, 8, 12, 15] construct and solve an optimization" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25183" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "model of PU regularized by some handcrafted priors. While achieving strong global consistency brought by the global optimization, these methods often fall short for complex phase patterns, due to the use of handcrafted priors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 287, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 287, + 158 + ], + "type": "text", + "content": "Supervised DL for PU: Supervised DL-based PU methods differ mainly in NN architectures. Existing methods can be roughly classified as CNN-based or RNN-based." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 361 + ], + "type": "text", + "content": "Leveraging the effectiveness of CNNs in segmentation and classification, many CNN-based methods (e.g. [10, 34, 35, 49, 50, 52, 56]) cast PU as a wrap count prediction problem, using a pixel-wise classification loss with wrap counts as labels. For robustness to noise, pre-denoising or post-processing modules are often included; see e.g. [34, 35, 52]. However, these methods face scalability issues with widerange phase data, where the candidate values for wrap counts are overwhelmed. Instead, some CNN-based methods [6, 25, 40, 57]. Typically, CNN-based methods are inefficient in exploiting global spatial dependencies. In contrast, the RNN-based methods [22, 32] excel at capturing long-range dependencies. Viewing a phase pixel as an element in a sequence, these methods predefine several paths across all phase pixels to form an RNN. However, due to computational costs, only a few paths can be utilized. This constraint on path selection results in suboptimal outcomes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "content": "The NN designs of most existing methods overlook the physical priors of phase wrapping, which are beneficial for generalization, especially when training data is limited. Our proposed DUN incorporates these priors, enhancing interpretability, reducing overfitting, and boosting performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 628 + ], + "type": "text", + "content": "Unsupervised DL for PU: Despite significant advances in unsupervised DL for various inverse problems in imaging, e.g., denoising [17, 21, 26], deblurring [5, 18, 27, 31], compressed sensing [28, 38, 42] and phase retrieval [4, 29], unsupervised DL for PU has been much less studied. One related work is Yang et al. [47], which utilizes the deep image prior [39] encoded by an untrained CNN to address the absence of GT data. Specifically, the phase image is reparameterized by a CNN, with its weights learned to fit the measurements. As shown in [39, 47], the CNN prefers smooth patterns and requires early stopping for smooth phase unwrapping, akin to conventional model-based methods but employing a CNN-based prior. A drawback is the need for per-sample retraining, which is computationally intensive. In contrast, our approach allows end-to-end unsupervised training on a dataset, eliminating test-time learning and significantly speeding up test." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "DUNs for imaging: DUNs have been studied for many inverse problems in imaging, e.g., denoising [20, 55], super-resolution [51], deblurring [20], compressed sensing [24, 53] and medical imaging [7]. However, the study on deep unrolling for PU is scant. Our study is the first to develop DUNs for PU, in addition to an unsupervised training scheme." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 91, + 404, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 404, + 103 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 404, + 103 + ], + "type": "text", + "content": "3.1. Basics and Keys" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "text", + "content": "Consider estimating " + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "text", + "content": " from its wrapped version " + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\mathbf{X}) = \\mathbf{X} - 2\\pi \\mathbf{K}, \\mathbf{K} \\in \\mathbb{Z}^{M \\times N}" + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "inline_equation", + "content": "\\nabla : \\mathbf{Z} \\to (\\nabla_x \\mathbf{Z}; \\nabla_y \\mathbf{Z})" + }, + { + "bbox": [ + 306, + 109, + 547, + 146 + ], + "type": "text", + "content": " denote the spatial gradient operator. By definition," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 150, + 545, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 150, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 323, + 150, + 545, + 164 + ], + "type": "interline_equation", + "content": "\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\boldsymbol {\\nabla} \\boldsymbol {X} [ m, n ] + 2 \\pi \\boldsymbol {\\nabla} \\boldsymbol {K} [ m, n ]. \\tag {3}", + "image_path": "54e44a0ae97c3619277364f5519c943d675a0aa156d8f694f4904fa369d96663.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 169, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 169, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 547, + 194 + ], + "type": "text", + "content": "For any location " + }, + { + "bbox": [ + 305, + 169, + 547, + 194 + ], + "type": "inline_equation", + "content": "[m,n]" + }, + { + "bbox": [ + 305, + 169, + 547, + 194 + ], + "type": "text", + "content": " whose adjacent pixels for gradient calculation have the same wrap count, we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 362, + 198, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 198, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 362, + 198, + 545, + 213 + ], + "type": "interline_equation", + "content": "\\nabla \\mathcal {W} (\\boldsymbol {X}) [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ]. \\tag {4}", + "image_path": "93f8950322df2ad00d95b18e4621d622ad403447209b0eb57ee8c66a3c3aba3c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 217, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 254 + ], + "type": "text", + "content": "In other words, the wrap count can be nullified by the gradient operator on many points. For the points not meeting the condition, we have the following Proposition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 258, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 284 + ], + "type": "text", + "content": "Proposition 1. [Two-dimensional extension of Itoh's continuity condition [13]] For any point " + }, + { + "bbox": [ + 305, + 258, + 545, + 284 + ], + "type": "inline_equation", + "content": "[m, n]" + }, + { + "bbox": [ + 305, + 258, + 545, + 284 + ], + "type": "text", + "content": " satisfying" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 289, + 547, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 547, + 312 + ], + "type": "interline_equation", + "content": "\\left\\| \\nabla \\boldsymbol {X} [ m, n ] \\right\\| _ {\\infty} = \\max \\left\\{\\left| \\nabla_ {x} \\boldsymbol {X} [ m, n ] \\right|, \\left| \\nabla_ {y} \\boldsymbol {X} [ m, n ] \\right| \\right\\} < \\pi , \\tag {5}", + "image_path": "f106a084d7209eda0723746d1ce0e705564d07b03bae21dd1949e1188691929b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 313, + 504, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 504, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 504, + 326 + ], + "type": "text", + "content": "we have that " + }, + { + "bbox": [ + 306, + 313, + 504, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla \\mathcal{W}(\\pmb {X}))[m,n] = \\nabla \\pmb {X}[m,n]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 330, + 500, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 500, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 500, + 342 + ], + "type": "text", + "content": "Proof. See supplemental material for the proof." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 349, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 445 + ], + "type": "text", + "content": "Proposition 1 shows that the wrapped gradients of a wrapped phase image are identical to the gradients of the GT, for the points satisfying the 2D Itoh's continuity condition (5) (called Itoh's condition for brevity in the following text). In general, GT phases exhibit gradual changes in most points, and their rapid changes only occur at a few points sparsely distributed over the image. Then, the outlier points not meeting Itoh's condition are sparse." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "text", + "content": "Now consider a noisy wrapped image " + }, + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N})" + }, + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "text", + "content": ". We can directly extend Eq. (4) to the noisy case:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 343, + 474, + 545, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 474, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 343, + 474, + 545, + 488 + ], + "type": "interline_equation", + "content": "\\nabla \\boldsymbol {Y} [ m, n ] = \\nabla \\boldsymbol {X} [ m, n ] + \\nabla \\boldsymbol {N} [ m, n ], \\tag {6}", + "image_path": "ea05d8a06a01f638dbbdd39555de34b5726003464f749cf79199f6680d27edfa.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "content": "for any location " + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "[m, n]" + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} + \\mathbf{N})" + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "content": " whose adjacent points share the same wrap count. Similarly, we have the following corollary of Proposition 1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "text", + "content": "Corollary 1.1. For points " + }, + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "inline_equation", + "content": "[m,n]" + }, + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\| \\nabla X[m,n] + \\nabla N[m,n]\\|_{\\infty} < \\pi" + }, + { + "bbox": [ + 306, + 533, + 545, + 559 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {W} (\\nabla \\boldsymbol {Y} [ m, n ]) = \\mathcal {W} (\\nabla \\boldsymbol {X} [ m, n ]) + \\nabla \\boldsymbol {N} [ m, n ]. \\tag {7}", + "image_path": "07c31560bcdc7d8f70230e5826a98acce72028d1349333c42c7860e1a9129d0c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": "It is clear that both " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": " can act as a noisy version for " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": ", except on some generally-sparse outlier points that do not satisfy the corresponding conditions. If we can mitigate those noise and outliers to predict " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": ", the remaining task, reconstructing " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": ", is straightforward, as " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": " can be reconstructed by integrating its spatial gradients " + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 305, + 582, + 545, + 715 + ], + "type": "text", + "content": " with appropriate boundary conditions. Indeed, the primary concern in PU revolves around relative phases rather than the absolute ones. The observations described by Eq. (7) and Eq. (6) underpin our DUN and unsupervised loss function, respectively." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25184" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 527, + 300 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 527, + 300 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 527, + 300 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 527, + 300 + ], + "type": "image", + "image_path": "abeb0ac61fa480e7f289556dfc002ba42f92838af551153f06b61b11a8c7f8d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 307, + 394, + 319 + ], + "lines": [ + { + "bbox": [ + 200, + 307, + 394, + 319 + ], + "spans": [ + { + "bbox": [ + 200, + 307, + 394, + 319 + ], + "type": "text", + "content": "Figure 2. Architecture of our proposed DUN for PU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 338, + 219, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 219, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 219, + 351 + ], + "type": "text", + "content": "3.2. Deep Unrolling Network for PU" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "text", + "content": "Our DUN is based on a regularization model inspired by Eq. (7). Though Eq. (7) is inapplicable to the pixels not meeting Itoh's condition, these outliers are typically sparse in a phase image with gradual changes in most areas. Consequently, we introduce a sparse variable " + }, + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{M \\times N \\times 2}" + }, + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "text", + "content": " into the regularization model, which counteracts the errors introduced by those sparse outliers. Specifically, we consider the following regularization model of PU:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 461, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 461, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 60, + 461, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {X}, \\boldsymbol {E}} \\| \\boldsymbol {\\nabla} \\boldsymbol {X} - \\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) + \\boldsymbol {E} \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\boldsymbol {X}) + \\psi (\\boldsymbol {E}), \\quad (8)", + "image_path": "24d373e0434fa6ebe6e2c3ae510c6089a4f8b6d7778c5fefedf75811d21f5632.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\phi, \\psi" + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "content": " are some regularization functions for the latent phase image " + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "content": " and the error matrix " + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "content": ", respectively. It can be seen that " + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 488, + 287, + 536 + ], + "type": "text", + "content": " essentially neutralizes outlier effects, preserving correctness of the data fitting term." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 288, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 288, + 561 + ], + "type": "text", + "content": "Our DUN is constructed via the unrolled iterative solver for (8), which reads as follows: for " + }, + { + "bbox": [ + 47, + 536, + 288, + 561 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 568, + 302, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 302, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 302, + 596 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\pmb {X} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {X}} \\| \\pmb {\\nabla} \\pmb {X} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {E} ^ {(t - 1)}) \\| _ {\\mathrm {F}} ^ {2} + \\phi (\\pmb {X}), \\\\ \\pmb {E} ^ {(t)} := \\operatorname * {a r g m i n} _ {\\pmb {E}} \\| \\pmb {E} - (\\mathcal {W} (\\pmb {\\nabla} \\pmb {Y}) - \\pmb {\\nabla} \\pmb {X} ^ {(t)}) \\| _ {\\mathrm {F}} ^ {2} + \\psi (\\pmb {E}). \\end{array} \\right.", + "image_path": "7f866cbdef15dce349124508a21c82554858347d3896b47cfd461b3b5e19e39b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "content": "Accordingly, the DUN predicts " + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla \\mathbf{Y})" + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "content": ". It alternatively connects two types of sub-NNs that act as the estimators of " + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(t)}" + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{(t)}" + }, + { + "bbox": [ + 47, + 604, + 287, + 651 + ], + "type": "text", + "content": ", respectively. See Fig. 2 for an illustration of this architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "text", + "content": "Sub-NN for estimating " + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{X}^{(t)}" + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "text", + "content": ": The sub-problem for " + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{X}^{(t)}" + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "text", + "content": " is solved using proximal gradient descent (PGD), which combines gradient descent with proximal mapping. Considering the limited number of modules (i.e., iteration number " + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 652, + 288, + 714 + ], + "type": "text", + "content": ") in our DUN for computational efficiency, we implement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "text", + "content": "multiple gradient descent steps within a stage to bridge the gap. Specifically, Nesterov's acceleration scheme is used to enhance the gradient descent, called AGD. We initialize with " + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{(0)}^{(t)} = \\mathbf{X}^{(t - 1)}, \\mathbf{X}^{(0)} = \\mathbf{1}, \\mathbf{E}^{(0)} = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "inline_equation", + "content": "\\alpha_{(0)} = 0" + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,J" + }, + { + "bbox": [ + 304, + 339, + 546, + 401 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 407, + 566, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 566, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 566, + 462 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\boldsymbol {X} _ {(j)} ^ {(t)} = \\boldsymbol {V} _ {(j - 1)} ^ {(t)} + \\lambda^ {(t)} \\mathrm {d i v} \\big (\\boldsymbol {\\nabla} \\boldsymbol {V} _ {(j - 1)} ^ {(t)} - (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y}) - \\boldsymbol {E} ^ {(t - 1)}) \\big), \\\\ \\alpha_ {(j)} = \\frac {1}{2} (1 + \\sqrt {1 + 4 \\alpha_ {(j - 1)} ^ {2}}), \\\\ \\boldsymbol {V} _ {(j)} ^ {(t)} = \\boldsymbol {X} _ {(j)} ^ {(t)} + \\frac {\\alpha_ {(j - 1)} - 1}{\\alpha_ {(j)}} (\\boldsymbol {X} _ {(j)} ^ {(t)} - \\boldsymbol {X} _ {(j - 1)} ^ {(t)}). \\end{array} \\right.", + "image_path": "8d7edd7b56fdefd8fc0ebd0e8c0fd08ff05b3f49f24e85a8d3852086bc64c3eb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "text", + "content": "Here the term " + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{div}\\bigl (\\pmb {\\nabla}V_{(j - 1)}^{(t)} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y}) - \\pmb{E}^{(t - 1)})\\bigr)" + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "text", + "content": " corresponds to the gradient " + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t(\\pmb {X}) = \\pmb {\\nabla}_\\pmb {X}\\| \\pmb {\\nabla}\\pmb {X} - (\\mathcal{W}(\\pmb {\\nabla}\\pmb {Y})-" + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "inline_equation", + "content": "E^{(t - 1)})\\|_{\\mathrm{F}}^2" + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "text", + "content": " . Afterward, the proximal operator, defined as " + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{Prox}_f(\\bar{Z}):= \\mathrm{argmin}_{\\pmb{Z}}\\| \\pmb {Z} - \\bar{\\pmb{Z}}\\| _2^2 +f(\\pmb {Z})" + }, + { + "bbox": [ + 305, + 470, + 546, + 523 + ], + "type": "text", + "content": " , is applied:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 530, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 530, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 380, + 530, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {(t)} = \\operatorname {P r o x} _ {\\phi} \\left(\\boldsymbol {X} _ {(J)} ^ {(t)}\\right). \\tag {9}", + "image_path": "60d487e8ceb91cf637097b0fc6efd510744829925e6dcae0d25313420a01bc2d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "text", + "content": "This proximal operator is implemented using a sub-NN with an attentive U-shape structure. As illustrated in Fig. 2, it has a " + }, + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "text", + "content": " down-scaling encoder and a corresponding up-scaling decoder, connected by skip connections. At the coarsest scale, a standard self-attention block is used for exploiting global features. The sub-NN also takes the gradient term " + }, + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t(\\boldsymbol{X}_{(J)}^{(t)})" + }, + { + "bbox": [ + 304, + 554, + 545, + 663 + ], + "type": "text", + "content": " as an additional input via concatenation. A long skip connection from the input to output is added, enhancing feature flow with a weighted summation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "content": "Sub-NN for estimating " + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "inline_equation", + "content": "E^{(t)}" + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "content": ": The sub-problem for " + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "inline_equation", + "content": "E^{(t)}" + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "content": " aims to address the outliers presented in the data fitting term defined on " + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}" + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "content": ", guided by the prior encoded in " + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\psi (E)" + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "text", + "content": ". We employ a sub-NN with " + }, + { + "bbox": [ + 304, + 664, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y) - \\nabla X^{(t)}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25185" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "as the input to implement the solver of this sub-problem. This sub-NN, illustrated in Fig. 2, is composed of sequential convolutional layers, each followed by a parametric rectified linear unit (PReLU). The sparsity prior of " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " is induced by attaching a soft thresholding activation function, parametrized by a non-negative parameter " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}(d^{(t)})" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 286, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 286, + 229 + ], + "type": "text", + "content": "Condition-aware parameter estimation: The noise level in the wrapped phases impacts the optimal selection of the parameters " + }, + { + "bbox": [ + 46, + 146, + 286, + 229 + ], + "type": "inline_equation", + "content": "\\{\\lambda^{(t)}, w^{(t)}, d^{(t)}\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 146, + 286, + 229 + ], + "type": "text", + "content": ". Existing DUNs (e.g. [51]) often include imaging condition values in the input. In our DUN, a condition-aware module (CAM) is used to predict these parameters from noise level, using a chain of fully-connected layers with PReLU or Sigmoid activations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 274, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 274, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 274, + 249 + ], + "type": "text", + "content": "3.3. Unsupervised End-to-End Training Scheme" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": " denote our DUN that maps " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": ". Our proposed loss function " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": " for the unsupervised training of " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": " comprises a noise-resistant self-reconstruction loss " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": " and a self-distillation loss " + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sd}}" + }, + { + "bbox": [ + 46, + 255, + 286, + 301 + ], + "type": "text", + "content": ", applied to each stage of the DUN:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 311, + 286, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 311, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 286, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {s r}} + \\eta \\mathcal {L} _ {\\mathrm {s d}}, \\eta \\in \\mathbb {R} ^ {+}. \\tag {10}", + "image_path": "42ed07e66c8b5f18ac8f1c3cbdb7074f2b8879e07a1376f188c48576dbe1b39a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "content": "Noise-resistant self-reconstruction loss: This loss utilizes " + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "content": " as a noisy label for " + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "content": ", as shown in Eq. (6). To mitigate the impact of label noise, the loss is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 377, + 286, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 377, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 52, + 377, + 286, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) - (\\nabla \\boldsymbol {Y} - \\nabla \\boldsymbol {U}) \\| _ {\\mathrm {F}} ^ {2}, \\tag {11}", + "image_path": "54606427a591d4f892b76c4bd094a7baedcd4ac83240fccdb5cccd1af00c4a93.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "content": " is random noise drawn from the distribution of " + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "content": ". The introduction of the noise pair " + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "inline_equation", + "content": "\\{+\\nabla U, -\\nabla U\\}" + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "content": " is to achieve the resilience to measurement noise in " + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 47, + 399, + 286, + 447 + ], + "type": "text", + "content": " in the training loss, as justified in Proposition 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\mathcal{W}(\\mathbf{X} + \\mathbf{N})" + }, + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "text", + "content": ". Suppose Eq. (6) is satisfied at all points. Assume that " + }, + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{N}, \\mathbf{U} \\sim \\mathcal{P}" + }, + { + "bbox": [ + 46, + 453, + 286, + 489 + ], + "type": "text", + "content": " are independent. Then, we have that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 497, + 286, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 286, + 521 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {Y}} \\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {X}, \\boldsymbol {N}, \\boldsymbol {U}} \\| \\boldsymbol {\\nabla} \\mathcal {F} (\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})) - \\boldsymbol {\\nabla} \\boldsymbol {X} \\| _ {F} ^ {2} + C _ {0}, \\tag {12}", + "image_path": "b3f1655f245101a81a74832433e4e25e958e9d3fb4287cdc315ff22b18e1fb05.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 522, + 143, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 143, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 143, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 522, + 143, + 533 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 47, + 522, + 143, + 533 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 541, + 235, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 235, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 235, + 553 + ], + "type": "text", + "content": "Proof. See the proof in supplemental material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": "Proposition 2 shows that if one can independently draw instances " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": " from the same distribution as " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": ", the loss " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": " effectively simulates the loss supervised by " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "\\nabla X" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": ", nullifying the label noise as shown in Eq. (12). Note that Proposition 2 requires all points satisfying Eq. (6), which is not generally true. On the outliers, " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": " provides wrong supervision with superfluous multiples of " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": ", as seen in Eq. (3). To counteract the impact of those outlier points, we exploit the property " + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\theta + 2k\\pi) = \\theta" + }, + { + "bbox": [ + 46, + 561, + 286, + 679 + ], + "type": "text", + "content": " to modify the self-reconstruction loss as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 687, + 293, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 293, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 293, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s r}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\mathcal {W} \\left[ \\boldsymbol {\\nabla} \\mathcal {F} \\left(\\mathcal {W} (\\boldsymbol {\\nabla} \\boldsymbol {Y} + \\boldsymbol {\\nabla} \\boldsymbol {U})\\right) - (\\boldsymbol {\\nabla} \\boldsymbol {Y} - \\boldsymbol {\\nabla} \\boldsymbol {U}) \\right] \\| _ {\\mathrm {F}} ^ {2}. \\tag {13}", + "image_path": "0456c4dbd105689caaf314eb691e862ce97bdbe52aa8aca7f79cc3d9655194c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "That is, an extra " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " is applied to the prediction residual within the Frobenius norm, which negates the superfluous multiples of " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " during the loss calculation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "content": "Self-distillation loss: Although " + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "content": " reduces the impact of measurement noise, the ambiguity in solutions due to wrapping may remain, particularly at those outlier points. Fortunately, CNNs including unrolling CNNs [30, 39], show strong induced bias towards regular structures, which aids in reducing such ambiguity during the training with " + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "type": "text", + "content": "To further reduce the ambiguity, we introduce a self-distillation loss defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 211, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s d}} := \\mathbb {E} _ {\\boldsymbol {U}} \\| \\nabla \\mathcal {F} (\\mathcal {W} (\\nabla \\boldsymbol {Y})) - \\nabla \\bar {\\mathcal {F}} (\\mathcal {W} (\\nabla \\boldsymbol {Y} + \\nabla \\boldsymbol {U})) \\| _ {\\mathrm {F}} ^ {2}. \\tag {14}", + "image_path": "3b75a3a2add1450d2333b3779c2117b6a648b9db575ca5e2a69729a3554f26f6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " denotes the NN detached from the previous iteration with stop gradient. The loss leverages pseudo supervision from the NN's varied predictions with noise-altered inputs, implicitly aggregating those predictions for self-training. This self-distillation scheme can reduce the NN's prediction variance, enhancing the PU accuracy. Moreover, recall that " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " inputs " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y + \\nabla U)" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": ", which does not fully align with " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " in test. This inconsistency can be reconciled by " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sd}}" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": ", as it uses " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " for the input of " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 354, + 386, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 386, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 386, + 368 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 374, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 482 + ], + "type": "text", + "content": "Performance evaluation is conducted on both simulated and real-world phase patterns. Following [22], we use NRMSE (Normalized Root Mean Squared Error) as the performance metric. As relative phase is the main focus in PU, the mean of unwrapped phase values is first aligned to that of the GT before the NRMSE calculation. Computational complexity is measured by three metrics: the number of parameters, the number of FLOPs (floating point of operations), and the average test time per sample of size " + }, + { + "bbox": [ + 304, + 374, + 545, + 482 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 374, + 545, + 482 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": "Implementation details: Our proposed approach, called U3Net (U3 for Unsupervised, Unrolling and Unwrapping) for convenience, is implemented in Python and run on a single NVIDIA RTX 3080 GPU. We set " + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\eta = 0.5" + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": " and initialize model weights via kaiming Uniform. The training employs the Adam optimizer with 500 epochs and a batch size of 10. The learning rate is initialized as " + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": " and exponentially decayed by a factor of 0.99 at every epoch." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "Methods for comparison: We select eight representative methods of general PU as baselines, including LS [8], QGPU [11], Ryu. et al. [32], PhaseNet2.0 [35], PUDIP [47], SQD-LSTM [22], EESANet [49], and TriNet [37]. Additionally, two representative transformer NNs for general image restoration, UFormer [43] and Restormer [48], are also included for performance comparison. In all the baseline methods, LS and QGPU are conventional non-learning iterative methods, built upon least squares and path selection, respectively; PUDIP is an internal DL-based method; and the others are supervised DL-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25186" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 527, + 222 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 527, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 527, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 527, + 222 + ], + "type": "table", + "html": "
DatasetMoGRRME#Param. (M)#FLOPs (G)Time (msec.)
SNR(dB)0510203005102030
Non-LearningLS7.912.281.200.370.157.852.501.250.400.12--8.39
QGPU17.122.121.150.360.1117.262.291.200.390.12--14.94
SupervisedRyu. et al.1.341.111.091.051.051.511.051.010.980.931.0721.85303.96
PhaseNet2.08.358.198.078.068.049.298.537.536.956.861.1511.9320.87
SQD-LSTM0.870.710.700.690.681.571.131.121.101.090.904.0713.04
EESANet0.780.770.760.760.741.311.311.251.241.0661.6875.279.85
TriNet5.655.555.465.405.345.405.335.215.105.0513.6165.4811.16
UFormer0.500.470.470.460.450.580.510.490.490.4820.6040.9842.00
Restormer0.450.380.360.360.350.500.430.420.410.413.0217.2344.69
UnsupervisedPUDIP17.5315.167.870.340.1113.107.222.620.380.122.3321.5899695.01
U3Net0.690.250.190.160.101.120.380.270.170.120.748.7710.28
", + "image_path": "c7e27bf5099fe744de7f3ed2b8e2968b77b82b544a70a1b1c14cd59d70843531.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 229, + 523, + 241 + ], + "lines": [ + { + "bbox": [ + 69, + 229, + 523, + 241 + ], + "spans": [ + { + "bbox": [ + 69, + 229, + 523, + 241 + ], + "type": "text", + "content": "Table 1. Performance comparison on synthetic datasets in terms of NRMSE(%) . Bold for Best and underline for second-best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 53, + 251, + 544, + 395 + ], + "blocks": [ + { + "bbox": [ + 53, + 251, + 544, + 395 + ], + "lines": [ + { + "bbox": [ + 53, + 251, + 544, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 544, + 395 + ], + "type": "image", + "image_path": "913f40f396c5ace88c33506d39f066d69e0c7e06faf6b07a14c1edefdca222ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 402, + 533, + 414 + ], + "lines": [ + { + "bbox": [ + 59, + 402, + 533, + 414 + ], + "spans": [ + { + "bbox": [ + 59, + 402, + 533, + 414 + ], + "type": "text", + "content": "Figure 3. Visualization of PU results on MoGR (top) and RME (bottom), showing residual images to facilitate visual comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "content": "based methods with their NNs retrained on our generated paired phase data. Specifically, as UFormer and Restormer are not originally designed for PU, we train their models using the supervised loss of SQD-LSTM [22] for better performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 258, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 258, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 258, + 514 + ], + "type": "text", + "content": "4.1. Evaluation on Simulated Phase Patterns" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": "Data preparation: Two types of simulated data are generated for training and evaluation. The first type follows [22], where a GT phase image of size " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " is generated through the superposition of a mixture of Gaussians (MoG) with " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " clusters and randomly selected ramps. The phase range is defined within " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "[-2p\\pi, 2p\\pi]" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": ". The resulting dataset is called MoGR (MoG with Ramp). The noisy wrapped phase images are synthesized via Eq. (1), with " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " defined as Gaussian white noise. For each GT, only one noisy wrapped phase image is generated. With " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " uniformly selected from " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, 4\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, 7\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " respectively, we generate 5000 noisy wrapped phase images with SNR uniformly selected from " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "\\{0, 5, 10, 20, 30, 60\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": " to construct the training set. Using the same way, 1000 noisy wrapped phase images are generated for test with SNR in " + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "inline_equation", + "content": "\\{0, 5, 10, 20, 30\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "The other dataset follows a different construction scheme" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "named random matrix enlargement [40], thus called RME dataset. Concretely, a GT phase image is formed through enlarging the square matrix of a distribution from an initial size " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " to size " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " via interpolation. The size parameter " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " is selected uniformly from " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\{2, \\dots, 10\\}" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": ", the distribution type is chosen from the {Uniform, Normal} distributions, and the interpolation method is chosen from the {Bilinear, Bicubic} interpolations. Then the phase image is re-scaled to the range varying in " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "[-2p\\pi, 2p\\pi]" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " randomly chosen from " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, 7\\}" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": ". The noisy wrapped phase images are obtained by the same way as the MoGR dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 677 + ], + "type": "text", + "content": "Results and analysis: Table 1 lists the quantitative results on both datasets, including those in computational complexity. Even trained without GT data, U3Net achieves the best results in 8/10 settings, using a lightweight model. Concretely, U3Net outperforms the two conventional non-learning-based methods, LS and QGPU. In comparison to PUDIP, an unsupervised internal DL-based method, U3Net also shows significant performance gain. The performance advantage of U3Net over these three GT-free methods be" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "text", + "content": "PUDIP involves many forward passes for test-time learning. We report its #FLOPs in one forward pass and running time of the whole process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 542, + 122 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 542, + 122 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 542, + 122 + ], + "type": "table", + "html": "
SNR(dB)LSQGPURyu. et al.PhaseNet2.0SQD-LSTMEESANetTriNetPU-GANPUNetUFormerRestormerPUDIPU3Net
53.313.131.412.281.762.455.0413.739.591.461.069.991.00
101.841.961.271.691.521.994.6911.849.201.280.935.300.82
200.941.131.241.431.481.794.4611.629.010.970.910.470.46
", + "image_path": "60156f64c4db38f79bf98e996d9115531d42cff561d39a647c4f5feeee597555.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 129, + 523, + 140 + ], + "lines": [ + { + "bbox": [ + 67, + 129, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 129, + 523, + 140 + ], + "type": "text", + "content": "Table 2. Performance comparison on InSAR phase data in terms of NRMSE(%) . Bold for Best and underline for second-best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 95, + 152, + 500, + 296 + ], + "blocks": [ + { + "bbox": [ + 95, + 152, + 500, + 296 + ], + "lines": [ + { + "bbox": [ + 95, + 152, + 500, + 296 + ], + "spans": [ + { + "bbox": [ + 95, + 152, + 500, + 296 + ], + "type": "image", + "image_path": "397a852b44ad372cb3af2a5dc71bdf6a42355fa060ed9de63d0fdbf20df78d5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 306, + 522, + 318 + ], + "lines": [ + { + "bbox": [ + 69, + 306, + 522, + 318 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 522, + 318 + ], + "type": "text", + "content": "Figure 4. Visualization of unwrapped phase images on InSAR data, showing residual images to facilitate visual comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 286, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 286, + 421 + ], + "type": "text", + "content": "comes more noticeable for lower SNR (i.e. heavier noise), particularly compared to PUDIP. This is because U3Net has an explicit mechanism to deal with noise during training, thereby achieving higher noise robustness. In addition, U3Net is much faster than PUDIP, due to that U3Net employs end-to-end pre-trained models, avoiding the time-costly per-sample training of PUDIP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "type": "text", + "content": "Surprisingly, U3Net also surpasses the supervised DL-based methods in many settings. In the lowest-SNR case, U3Net performs worse than UFormer and Restormer, two supervised transformer models. This is probably due to that severe noise increases the number of outlier points, lowering the effectiveness of the training function. Even that, U3Net still outperforms other supervised methods. The model of U3Net is the smallest among all the DL-based methods, indicating that the excellent performance of U3Net is mainly attributed to its interpretable physics-encoded NN architecture and its effective training scheme, rather than the use of larger models. Meanwhile, U3Net is computationally efficient, with the second smallest number of FLOPs and the third fastest inference speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 590, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 286, + 661 + ], + "type": "text", + "content": "See Fig. 3 for the visualization of PU results from different methods. The residuals to GTs produced by U3Net are much more insignificant compared to the internal DL-based PUDIP. In comparison to supervised methods, U3Net outperforms all of them on the image from RME dataset and shows comparable results on the image of MoGR dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 242, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 242, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 242, + 682 + ], + "type": "text", + "content": "4.2. Evaluation on InSAR Phase Patterns" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "The InSAR phase data are generated by converting the digital elevation model at Texas, U.S. into absolute phases," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "following [33]. Totally, 4000 (1000) GT phase images are obtained by cropping the " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " patches from the absolute phases for training (test). Each phase value is scaled into " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "[-6\\pi, 6\\pi]" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": ". The noisy wrapped phases are generated by adding Gaussian noise following the experiments in Sec. 4.1, and we test on three noise levels: 5, 10, 20. In addition to the previously-compared general PU methods and general-purpose transformer NNs, we also compare with two specific methods designed for InSAR PU: PUNet [44] and PU-GAN [57]. All the compared models are retrained using the same data as ours." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 470, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 565 + ], + "type": "text", + "content": "See Table 2 for the quantitative comparison on this In-SAR dataset. Our U3Net ranks the first among all compared methods over all noise settings, noticeably surpassing other GT-free methods as well as many supervised methods in many cases. The visual comparison provided in Fig. 4 also shows the advantage of U3Net. All these results further verify the effectiveness of our approach in dealing with phase patterns from real scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 573, + 406, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 406, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 406, + 585 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 626 + ], + "type": "text", + "content": "We form some baselines from U3Net for ablation studies, which are retrained and tested on the MoGR dataset. Their results are summarized in Table 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Study on loss function: We design the following schemes to analyze the loss function: (a) " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ": remove the outer " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " applied to the residual, from the self-reconstruction loss " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", i.e., replacing " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " by the loss " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " defined in Eq. (11); (b) w/o " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nabla U" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ": remove the noise re-corruption mechanism in " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " by discarding the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pm \\nabla U" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " terms; (c) " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}} \\to \\mathcal{L}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " w/o " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nabla U" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ": combine the former cases, yielding a plain gradient loss" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25188" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 198 + ], + "type": "table", + "html": "
SettingSNR=0SNR=5SNR=10SNR=20SNR=30
Lsr→L25.4523.4823.2623.2323.20
w/o ∇U6.162.401.360.690.60
Lsr→L w/o ∇U25.8523.5723.4223.3823.34
w/o Lsd1.140.330.230.220.17
Supervised0.290.160.130.110.05
w/o E7.160.750.440.360.31
AGD→GD2.140.350.210.170.12
w/o CAM0.980.310.220.210.15
W(∇Y)→∇Y2.881.741.501.461.39
DUN→UFormer1.551.411.311.301.30
U3Net (original)0.690.250.190.160.10
", + "image_path": "3386b00feb973f36340c31d4fc46e6f562717d8855e7a1f5d122be5eb8e7d4bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 205, + 276, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 205, + 276, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 276, + 216 + ], + "type": "text", + "content": "Table 3. Results in ablation studies in terms of NRMSE(%)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 230, + 279, + 355 + ], + "blocks": [ + { + "bbox": [ + 56, + 230, + 279, + 355 + ], + "lines": [ + { + "bbox": [ + 56, + 230, + 279, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 279, + 355 + ], + "type": "image", + "image_path": "595e81c47fec259ebc2b19f954dc73e1904385058746d70dd9a5601d4749c892.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "text", + "content": "Figure 5. Visualization of unwrapped phase images using models trained by different loss functions. " + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}} \\rightarrow \\mathcal{L}" + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "text", + "content": " and w/o " + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\nabla U" + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "text", + "content": " produce noticeable errors than Original. Zoom in for a better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla Y\\|_{\\mathrm{F}}^2" + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "text", + "content": " (d) w/o " + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sd}}" + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "text", + "content": " : remove the selfdistillation loss " + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sd}}" + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "text", + "content": " , only employing " + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "text", + "content": " for training; and (e) Supervised: train with a gradient-domain supervised loss " + }, + { + "bbox": [ + 47, + 420, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\| \\nabla \\mathcal{F}(\\mathcal{W}(\\nabla Y)) - \\nabla X\\|_{\\mathrm{F}}^2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "content": "In addition to Table 3, we also provide some visual comparisons in Fig. 5 for demonstration. From both the quantitative and qualitative results, we have following observations. (a) The outer " + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "content": " is critical to the performance of U3Net. (b) The re-corruption mechanism in " + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sr}}" + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "content": " does help eliminate the measurement noise. (c) A simple gradient loss function does not work, with further performance decrease. (d) While the self-reconstruction loss already yields promising results, the self-distillation loss benefits further performance gain. (e) U3Net performs closely to its supervised counterpart when " + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SNR} \\geq 5" + }, + { + "bbox": [ + 46, + 471, + 287, + 628 + ], + "type": "text", + "content": ". We also see that U3Net has its supervised counterpart outperforming the supervised UFormer reported in Table 1, demonstrating its superiority." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Study on DUN: To analyze the components in the DUN architecture of U3Net, we design the following baseline models: (a) w/o " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ": abandon the error term " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " during unrolling, forming a DUN alternating the accelerated gradient descent step and the sub-NN for estimating " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{X}^{(t)}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "; (b) AGD→GD: replace the Nesterov's AGD steps by one-step standard gradient; (c) w/o CAM: remove the CAM from U3Net; (d)" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 539, + 252 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 539, + 252 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 539, + 252 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 539, + 252 + ], + "type": "image", + "image_path": "6c81c67944f6bb3c22d8d09aff4f14d9fe2a38190002fd4075c37d807c26cec9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "text", + "content": "Figure 6. Visualization of error matrices " + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "text", + "content": " estimated from three wrapped phase images. The outlier maps use white pixels to indicate the points " + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "inline_equation", + "content": "[x,y]" + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\| \\nabla X[x,y] + \\nabla N[x,y]\\|_{\\infty} \\geq \\pi" + }, + { + "bbox": [ + 305, + 260, + 545, + 304 + ], + "type": "text", + "content": " and black pixels otherwise." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla \\pmb {Y})\\rightarrow \\nabla \\pmb {Y}" + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "content": " : use " + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "content": " to replace " + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "content": " in the input of U3Net as well in the total loss; (e) DUN " + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "content": " UFormer: replace the DUN by UFormer, keeping the loss unchanged." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "text", + "content": "As demonstrated in Table 3, each component in the DUN of U3Net is useful. (a) One critical part is the introduction of the error matrix " + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "text", + "content": " for unrolling. It brings significant performance gain, particularly in dealing with low-SNR data, as the absorption of outliers breaking the Itoh's condition is more desired for heavier noise. See Fig. 6 for a visualization of estimated error matrices, which successfully identify the outliers. (b)&(c) The AGD and CAM have similar contribution to the performance of U3Net. (d) Using " + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\nabla Y)" + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "text", + "content": " for DUN is also crucial, serving as a better choice compared to using " + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "inline_equation", + "content": "\\nabla Y" + }, + { + "bbox": [ + 304, + 361, + 546, + 539 + ], + "type": "text", + "content": ". (e) Our proposed loss function also works for UFormer, but yielding worse results than U3Net, which indicates the higher effectiveness of our physics-encoded DUN for unsupervised PU, compared to general transformer models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 550, + 379, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 379, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "This work proposed a fully unsupervised DL approach for PU, eliminating the dependence on GTs in end-to-end training. Utilizing the connection between (wrapped) gradients of wrapped phases and gradients of original phases, we constructed a physics-encoded DUN, trained via a noise-resistant self-reconstruction loss complemented by a self-distillation loss. Empirical evaluations confirmed that our approach surpassed existing GT-free methods and rivaled supervised counterparts. Similar to current unsupervised and synthetic data-dependent supervised techniques, our approach presumes known noise statistics. Future research will focus on dealing with unknown noise profiles." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25189" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Liheng Bian, Xin Wang, Daoyu Li, Qiuling Ren, and Dezhi Zheng. Robust phase unwrapping via non-local regularization. Optics Letters, 48(6):1399-1402, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "type": "text", + "content": "[2] David Blinder, Heidi Ottevaere, Adrian Munteanu, and Peter Schelkens. Efficient multiscale phase unwrapping methodology with modulo wavelet transform. Optics Express, 24(20): 23094-23108, 2016. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "text", + "content": "[3] Donald J Bone. Fourier fringe analysis: the two-dimensional phase unwrapping problem. Applied Optics, 30(25):3627-3632, 1991. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "text", + "content": "[4] Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate mmse estimation. IEEE Transactions on Signal Processing, 70:2239-2252, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "type": "text", + "content": "[5] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 286, + 338 + ], + "type": "text", + "content": "[6] Gili Dardikman and Natan T Shaked. Phase unwrapping using residual neural networks. In Computational Optical Sensing and Imaging, pages CW3B-5. Optica Publishing Group, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "type": "text", + "content": "[7] Qiaoqiao Ding, Gaoyu Chen, Xiaoqun Zhang, Qiu Huang, Hui Ji, and Hao Gao. Low-dose ct with deep learning regularization via proximal forward-backward splitting. Physics in Medicine & Biology, 65(12):125009, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 286, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 286, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 286, + 429 + ], + "type": "text", + "content": "[8] Dennis C Ghiglia and Louis A Romero. Robust two-dimensional weighted and unweighted phase unwrapping that uses fast transforms and iterative methods. Journal of the Optical Society of America A, 11(1):107-117, 1994. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 430, + 286, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 286, + 473 + ], + "type": "text", + "content": "[9] Richard M Goldstein, Howard A Zebker, and Charles L Werner. Satellite radar interferometry: Two-dimensional phase unwrapping. Radio Science, 23(4):713-720, 1988. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 518 + ], + "type": "text", + "content": "[10] Michal Gontarz, Vibekananda Dutta, Malgorzata Kujawinska, and Wojciech Krauze. Phase unwrapping using deep learning in holographic tomography. Optics Express, 31(12):18964-18992, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 286, + 575 + ], + "type": "text", + "content": "[11] Miguel Arevallilo Herraez, David R Burton, Michael J Lolor, and Munther A Gdeisat. Fast two-dimensional phaseunwrapping algorithm based on sorting by reliability following a noncontinuous path. Applied Optics, 41(35):7437-7444, 2002. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "text", + "content": "[12] Howard YH Huang, L Tian, Z Zhang, Y Liu, Z Chen, and G Barbastathis. Path-independent phase unwrapping using phase gradient and total-variation (tv) denoising. Optics Express, 20(13):14075–14089, 2012. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "type": "text", + "content": "[13] Kazuyoshi Itoh. Analysis of the phase unwrapping algorithm. Applied Optics, 21(14):2470-2470, 1982. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[14] Charles VJ Jakowitz, Daniel E Wahl, Paul H Eichel, Dennis C Ghiglia, and Paul A Thompson. Spotlight-mode synthetic aperture radar: a signal processing approach. Springer Science & Business Media, 2012. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[15] Ulugbek S Kamilov, Ioannis N Papadopoulos, Morteza H Shoreh, Demetri Psaltis, and Michael Unser. Isotropic" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "text", + "content": "inverse-problem approach for two-dimensional phase unwrapping. Journal of the Optical Society of America A, 32 (6):1092-1100, 2015. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 128 + ], + "type": "text", + "content": "[16] Thomas Kreis. Handbook of holographic interferometry: optical and digital methods. John Wiley & Sons, 2006. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[17] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2137, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 174, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 226 + ], + "type": "text", + "content": "[18] Ji Li, Weixi Wang, Yuesong Nan, and Hui Ji. Self-supervised blind motion deblurring with deep expectation maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13986-13996, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[19] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "type": "text", + "content": "[20] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[21] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2043-2052, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 382, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 436 + ], + "type": "text", + "content": "[22] Malsha V Perera and Ashwin De Silva. A joint convolutional and spatial quad-directional LSTM network for phase unwrapping. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4055-4059. IEEE, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[23] Mark D Pritt. Phase unwrapping by means of multigrid techniques for interferometric sar. IEEE Transactions on Geoscience and Remote Sensing, 34(3):728-738, 1996. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[24] Xinran Qin, Yuhui Quan, and Hui Ji. Enhanced deep unrolling networks for snapshot compressive hyperspectral imaging. Neural Networks, 174:106250, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[25] Yi Qin, Shujia Wan, Yuhong Wan, Jiawen Weng, Wei Liu, and Qiong Gong. Direct and accurate phase unwrapping with deep neural network. Applied Optics, 59(24):7258-7267, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[26] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1898, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "text", + "content": "[27] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In European Conference on Computer Vision, pages 642-659. Springer, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[28] Yuhui Quan, Xinran Qin, Tongyao Pang, and Hui Ji. Dual-domain self-supervised learning and model adaption for deep compressive imaging. In European Conference on Computer Vision, pages 409-426. Springer, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[29] Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25190" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[30] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "type": "text", + "content": "[31] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3341-3350, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "text", + "content": "[32] Kanghyun Ryu, Sung-Min Gho, Yoonho Nam, Kevin Koch, and Dong-Hyun Kim. Development of a deep learning method for phase unwrapping mr images. In Proceedings of the International Society for Magnetic Resonance in Medicine, page 4707, 2019. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[33] Francescopaolo Sica, Francesco Calvanese, Giuseppe Scarpa, and Paola Rizzoli. A cnn-based coherence-driven approach for insar phase unwrapping. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "type": "text", + "content": "[34] GE Spoorthi, Subrahmanyam Gorthi, and Rama Krishna Sai Subrahmanyam Gorthi. Phasenet: A deep convolutional neural network for two-dimensional phase unwrapping. IEEE Signal Processing Letters, 26(1):54-58, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[35] GE Spoorthi, Rama Krishna Sai Subrahmanyam Gorthi, and Subrahmanyam Gorthi. Phasenet 2.0: Phase unwrapping of noisy data based on deep learning approach. IEEE Transactions on Image Processing, 29:4862-4872, 2020. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "text", + "content": "[36] Xianyu Su and Wenjing Chen. Reliability-guided phase unwrapping algorithm: a review. Optics and Lasers in Engineering, 42(3):245-261, 2004. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[37] Krishna Sumanth, Vaishnavi Ravi, and Rama Krishna Gorthi. A multi-task learning for 2d phase unwrapping in fringe projection. IEEE Signal Processing Letters, 29:797-801, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[38] Julián Tachella, Dongdong Chen, and Mike Davies. Unsupervised learning from incomplete measurements for inverse problems. Advances in Neural Information Processing Systems, 35:4983-4995, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[39] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9446-9454, 2018. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[40] Kaiqiang Wang, Ying Li, Qian Kemao, Jianglei Di, and Jianlin Zhao. One-step robust deep learning phase unwrapping. Optics Express, 27(10):15100-15115, 2019. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[41] Kaiqiang Wang, Qian Kemao, Jianglei Di, and Jianlin Zhao. Deep learning spatial phase unwrapping: a comparative review. Advanced Photonics Nexus, 1(1):014001, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[42] Weixi Wang, Ji Li, and Hui Ji. Self-supervised deep image restoration via adaptive stochastic gradient Langevin dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[43] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[44] Zhipeng Wu, Teng Wang, Yingjie Wang, Robert Wang, and Daqing Ge. Deep learning for the detection and phase unwrapping of mining-induced deformation in large-scale interferograms. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[45] Xianming Xie. Iterated unscented kalman filter for phase unwrapping of interferometric fringes. Optics Express, 24 (17):18872-18897, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[46] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "text", + "content": "[47] Fangshu Yang, Thanh-An Pham, Nathalie Brandenberg, Matthias P Lütolf, Jianwei Ma, and Michael Unser. Robust phase unwrapping via deep image prior for quantitative phase imaging. IEEE Transactions on Image Processing, 30: 7025-7037, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "text", + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[49] Junkang Zhang and Qingguang Li. Eesanet: edge-enhanced self-attention network for two-dimensional phase unwrapping. Optics Express, 30(7):10470-10490, 2022. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[50] Junchao Zhang, Xiaobo Tian, Jianbo Shao, Haibo Luo, and Rongguang Liang. Phase unwrapping in optical metrology via denoised and convolutional segmentation networks. Optics Express, 27(10):14903-14912, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[51] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3217-3226, 2020. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "text", + "content": "[52] Teng Zhang, Shaowei Jiang, Zixin Zhao, Krishna Dixit, Xiaofei Zhou, Jia Hou, Yongbing Zhang, and Chenggang Yan. Rapid and robust two-dimensional phase unwrapping via deep learning. Optics Express, 27(16):23173-23185, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "text", + "content": "[53] Zhonghao Zhang, Yipeng Liu, Jiani Liu, Fei Wen, and Ce Zhu. Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing, 30:1487-1500, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[54] Ming Zhao, Lei Huang, Qican Zhang, Xianyu Su, Anand Asundi, and Qian Kemo. Quality-guided phase unwrapping technique: comparison of quality maps and guiding strategies. Applied Optics, 50(33):6214-6224, 2011. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[55] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 630-641, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[56] Hongyu Zhou, Chuanli Cheng, Hao Peng, Dong Liang, Xin Liu, Hairong Zheng, and Chao Zou. The phu-net: A robust phase unwrapping method for mri based on deep learning." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25191" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "Magnetic Resonance in Medicine, 86(6):3321-3333, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "text", + "content": "[57] Lifan Zhou, Hanwen Yu, Vito Pascazio, and Mengdao Xing. Pu-gan: a one-step 2-d insar phase unwrapping based on conditional generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 60:1-10, 2022. 1, 3, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25192" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_content_list.json b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c644dcad968cfd4e5682c74600e2e864843c0e0 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "Unsupervised Feature Learning with Emergent Data-Driven Prototypicality", + "text_level": 1, + "bbox": [ + 102, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunhui Guo1 Youren Zhang2 Yubei Chen3 Stella X. Yu2,4 \n1The University of Texas at Dallas 2University of Michigan 3UC Davis 4UC Berkeley", + "bbox": [ + 99, + 178, + 869, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a set of images, our goal is to map each image to a point in a feature space such that, not only point proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset.", + "bbox": [ + 75, + 282, + 470, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflects image similarity, yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space.", + "bbox": [ + 75, + 345, + 468, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKKing. HACK first generates uniformly packed particles in the Poincaré ball of hyperbolic space and then assigns each image uniquely to a particle. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing - a warping process that aligns all the images and makes them appear more common and similar to each other, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones.", + "bbox": [ + 75, + 496, + 470, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 738, + 207, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Not all instances are created equal. For example, the MNIST dataset of handwritten digits contain almost 6,000 samples of 2's; some are close to textbook versions that we are taught to follow, whereas others have idiosyncratic cursive styles, varying in proportions and stroke weights (Fig. 1). Given such a set of natural data, we are interested in dataset summarization and organization such that we can automatically discover which instances are more representative and which ones are anomalies. In other words, we aim to computation-", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5593ab71496867a84926998f37b4ca5d63a2740aa848718bd400c1b52cae2c4.jpg", + "image_caption": [ + "Figure 1. Given a dataset (left), we aim to learn an image feature that encodes not only visual similarity between instances but also data-driven prototypicality (right). Additionally, the angular arrangement of the features can naturally serve as a measure of diversity. Our feature encoder (in 2D hyperbolic space) is learned without any labels. We can then learn a decoder to map each point in the feature space back to an image. The right plot visualizes images located at the origin and those moving away in different directions, automatically revealing that 2's with loops are most common and the whole dataset can be grasped as the cursive style systematically varies." + ], + "image_footnote": [], + "bbox": [ + 509, + 250, + 883, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ally rationalize graded membership [8, 36] of a category in a purely data-driven manner, putting each instance on a feature map that reflects not only their prototypicality in a particular dataset but also their visual similarity with each other.", + "bbox": [ + 496, + 597, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Such unsupervised feature learning is useful for not only data organization, but those discovered representative instances or prototypes can be used for interpretable machine learning [4], curriculum learning [3], and learning better decision boundaries [5]. Prototypes also allow us to classify with as few as or even one example [31].", + "bbox": [ + 496, + 657, + 890, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "If the image feature is given, it is relatively easy to find prototypes: We just need to identify density peaks of the feature distribution of the image set. Otherwise, discovering prototypical instances without supervision is difficult: There is no universal definition or simple metric to assess the prototypicality of the examples.", + "bbox": [ + 496, + 750, + 893, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One way to address this problem is to examine the gradient magnitude [5]. However, this approach is shown to have a high variance which results from different training setups [5]. Some methods address this problem using adversarial", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23199", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/276794e3216f93c4a263cd67fa8d95a3c6d5fcd181c22e76a70a9e90799c3caf.jpg", + "image_caption": [ + "Figure 2. Existing self-supervised feature learning methods focus on visual similarity only in Euclidean space, whereas our unsupervised feature learning method embeds images in hyperbolic space which automatically encodes prototypicality with respect to the origin: Images are organized hierarchically, with typical images at the center and atypical ones near the boundary of the Poincaré ball." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 433, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "robustness [5, 39]: Prototypical examples should be more adversarially robust. However, selecting these examples is dependent on the adversarial method and the metric used in the adversarial attack. Several methods exist but they are either based on heuristics or lack a proper justification [5].", + "bbox": [ + 75, + 409, + 468, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Naturally, given a feature space, prototypical examples can be identified as density peaks. However, prototypicality varies with the feature. We propose an unsupervised feature learning method in hyperbolic space, which, unlike the shift-invariant Euclidean space, naturally embeds a continuous version of hierarchical trees rooted at the origin of the space.", + "bbox": [ + 75, + 487, + 468, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hyperbolic space is non-Euclidean space with constant non-negative curvature [1]. Different from Euclidean space, hyperbolic space can represent hierarchical relations with low distortion. Poincaré ball model is one of the most commonly used models for hyperbolic space [35]. One notable property of Poincaré ball model is that the distance to the origin grows exponentially as we move towards the boundary. Thus, points located in the center of the ball are close to all the other points while the points located close to the boundary are infinitely far away from other points. With unsupervised learning in hyperbolic space, we can learn features that capture both visual similarity and hierarchical proximity among instances.", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is that a typical image is similar (closest) to more nearby instances than atypical ones, and such an image would be at the center of the neighbourhood in Euclidean space but the parent/root node in a tree in hyperbolic space, closer to the origin. We develop a new learning procedure that first places all the target locations evenly in the Poincaré ball to reflect the desire of mapping uniformity [43], and then we optimize which images should be mapped to which", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "target locations in a batch-wise Hungarian matching manner, where those similar to most instances naturally moving closer to the origin.", + "bbox": [ + 496, + 90, + 893, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work makes the following contributions. 1) We propose the first unsupervised feature learning method to learn features which capture visual similarity with distance between features and prototypicality with the distance to the origin. 2) We develop a new learning paradigm that sits between supervised learning (with known targets) and unsupervised metric learning (with unknown targets and constrained metric distances). We want to map images to known targets uniformly packed and maximally distant in hyperbolic space, but we learn to optimize the specific image to target assignment. 3) We validate our joint feature learning and data prototypicality discovery on congealing [31], where the consensus is that images after congealing are perceived to be more typical images. 4) We demonstrate two practical usages of data-driven prototypicality: Prototypical and atypical examples are shown to reduce sample complexity for learning and increase the robustness of the model respectively.", + "bbox": [ + 496, + 136, + 893, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 407, + 640, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prototypicality. The study of prototypical examples in machine learning has a long history. In Zhang [47], the authors select typical instances based on the fact that typical instances should be representative of the cluster. In Kim et al. [22], prototypical examples are defined as the examples that have maximum mean discrepancy within the data. Li et al. [27] propose to discover prototypical examples by architectural modifications: project the dataset onto a low-dimensional manifold and use a prototype layer to minimize the distance between inputs and the prototypes on the manifold. The robustness to adversarial attacks is also used as a criterion for prototypicality [39]. In Carlini et al. [5], the authors propose multiple metrics for prototypicality discovery. For example, the features of prototypical examples should be consistent across different training setups. However, these metrics usually depend heavily on the training setups and hyperparameters. The idea of prototypicality is also extensively studied in meta-learning for one-shot or few-shot classification [38]. No existing works address the prototypicality discovery problem in a data-driven fashion. Our proposed HACK naturally exploits hyperbolic space to organize the images based on prototypicality.", + "bbox": [ + 496, + 431, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Learning in Hyperbolic Space. Learning features in hyperbolic space have shown to be useful for many machine learning problems [11, 34]. One useful property is that hierarchical relations can be embedded in hyperbolic space with low distortion [34]. Wrapped normal distribution, which is a generalized version of the normal distribution for modeling the distribution of points in hyperbolic space [33], is used as the latent space for constructing hyperbolic variational autoencoders (VAEs)", + "bbox": [ + 496, + 765, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23200", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f15b21f57154c3757da9e0d08a4609b178121bcecf28e67c56360693977f2471.jpg", + "image_caption": [ + "Figure 3. Congeaed images are more typical than the original images. First row: sampled original images. Second row: the corresponding congealed images." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 467, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[23]. Poincaré VAEs is constructed in Mathieu et al. [30] with a similar idea to Nagano et al. [33] by replacing the standard normal distribution with hyperbolic normal distribution. Unsupervised 3D segmentation [20] and instance segmentation [44] are conducted in hyperbolic space via hierarchical hyperbolic triplet loss. CO-SNE [15] is recently proposed to visualize high-dimensional hyperbolic features in a two-dimensional hyperbolic space. Although hyperbolic distance facilitates the learning of hierarchical structure, how to leverage hyperbolic space for unsupervised prototypicality discovery is not explored in the current literature.", + "bbox": [ + 75, + 253, + 472, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sample Hierarchy", + "text_level": 1, + "bbox": [ + 76, + 450, + 253, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sample Hierarchy VS. Class Hierarchy. While most of the existing works in hierarchical image classification focus on using label hierarchy [9, 14], there also exists a natural hierarchy among different samples. In Khrulkov et al. [21], the authors conducted an experiment to measure the $\\delta$ -hyperbolicity of the various image datasets and showed that common image datasets such as CIFAR10 and CUB exhibit natural hierarchical structure among the samples. Amongst a collection of images representing digit 1, suppose $\\mathbf{x}$ is used for representing an image with a digit '1' that is upright, $\\mathbf{x}'$ is used for representing an image with a digit 1 that leaning left and $\\mathbf{x}''$ is used for representing an image with a digit '1' that leaning right. Given a metric $d(\\cdot, \\cdot)$ , if we assume that $d(\\mathbf{x}''', \\mathbf{x}') \\approx d(\\mathbf{x}''', \\mathbf{x}) + d(\\mathbf{x}', \\mathbf{x})$ , in this context, we can naturally view the sample $\\mathbf{x}$ as the root, and consider the other samples as its children in an underlying tree.", + "bbox": [ + 75, + 476, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with class hierarchy which can be extracted based on the pre-defined label relations, sample hierarchy is much harder to construct due to the lack of ground truth. Once a sample hierarchy is established, there are currently no existing methods available for verifying the accuracy of the hierarchy. Additionally, just like with class hierarchies, there may be ambiguities when constructing a sample hierarchy since multiple samples could potentially serve as the root.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building Sample Hierarchy from Density Peaks. Given existing features $\\{f(v_{i})\\}$ obtained by applying a feature extractor for each instance $v_{i}$ , prototypical examples can be found by examining the density peaks via techniques from", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "density estimation. For example, the K-nearest neighbor density (K-NN) estimation [10] is defined as $p_{knn}(v_i, k) = \\frac{k}{n} \\frac{1}{A_d \\cdot D^d(v_i, v_{k(i)})}$ , where $d$ is the feature dimension, $A_d = \\pi^{d/2} / \\Gamma(d/2 + 1)$ , $\\Gamma(x)$ is the Gamma function and $k(i)$ is the $k$ th nearest neighbor of example $v_i$ . The nearest neighbors can be found by computing the distance between the features. Therefore, the process of constructing sample hierarchy through density estimation can be conceptualized as a two-step procedure involving: 1) feature learning and 2) detecting density peaks.", + "bbox": [ + 496, + 90, + 893, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the density estimation approach outlined above, the level of prototypicality depends on the learned features. Varying training setups can induce diverse feature spaces, resulting in differing conclusions on prototypicality. Nevertheless, prototypicality is an inherent attribute of the dataset and should remain consistent across various features. The aim of this paper is to extract features that intrinsically showcase the hierarchical organization of the samples. Specifically, by examining the feature alone within the feature space, we should be able to identify the example's prototypicality.", + "bbox": [ + 496, + 247, + 893, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Construct a Sample Hierarchy from Congealing. To determine whether the feature truly captures prototypicality, it is necessary to identify which sample is the prototype. We ground our concept of prototypicality based on congealing [31]. In particular, we define prototypical examples in the pixel space by examining the distance of the images to the average image in the corresponding class. Our idea is based on a traditional computer vision technique called image alignment [40] that aims to find correspondences across images. During congealing [31], a set of images are transformed to be jointly aligned by minimizing the joint pixel-wise entropies. The congealed images are more prototypical: they are better aligned with the average image. Thus, we have a simple way to transform an atypical example into a typical example (see Figure 3). This is useful since given an unlabeled image dataset the typicality of the examples is unknown, congealing examples can be naturally served as examples with known typicality and be used as a validation for the effectiveness of our method.", + "bbox": [ + 496, + 398, + 893, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Unsupervised Hyperbolic Feature Learning", + "text_level": 1, + "bbox": [ + 498, + 699, + 887, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Poincaré Ball Model for Hyperbolic Space", + "text_level": 1, + "bbox": [ + 498, + 724, + 859, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Euclidean space has a curvature of zero and a hyperbolic space is a Riemannian manifold with constant negative curvature.. There are several isometrically equivalent models for visualizing hyperbolic space with Euclidean representation. The Poincaré ball model is the commonly used one in hyperbolic representation learning [35]. The $n$ -dimensional Poincaré ball model is defined as $(\\mathbb{B}^n, \\mathfrak{g}_{\\mathbf{x}})$ , where $\\mathbb{B}^n = \\{\\mathbf{x} \\in \\mathbb{R}^n : \\| \\mathbf{x} \\| < 1\\}$ and $\\mathfrak{g}_{\\mathbf{x}} = (\\gamma_{\\mathbf{x}})^2 I_n$ is the Riemannian metric tensor. $\\gamma_{\\mathbf{x}} = \\frac{2}{1 - \\| \\mathbf{x} \\|^2}$ is the conformal factor and $I_n$ is the Euclidean metric tensor.", + "bbox": [ + 496, + 748, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea7971f0a81d7aebf36b97a9e34f3d5b5af9b96b25a3652bb9f957912665893c.jpg", + "image_caption": [ + "a) Supervised classification with fixed known targets" + ], + "image_footnote": [], + "bbox": [ + 147, + 90, + 334, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a944af3952c3dc778a0d0d2b62ba17bfab10fd3a134687c0c48faaeccf4799c2.jpg", + "image_caption": [ + "b) Our unsupervised feature learning with fixed but unknown targets", + "Figure 4. The proposed HACK has a predefined geometrical arrangement and allows the images to be freely assigned to any particle. a) Standard supervised learning has predefined targets. The image is only allowed to be assigned to the corresponding target. b) HACK packs particles uniformly in hyperbolic space to create initial seeds for the organization. The images are assigned to the particles based on their prototypicality and semantic similarities. c) Standard unsupervised learning has no predefined targets and images are clustered based on their semantic similarities." + ], + "image_footnote": [], + "bbox": [ + 405, + 87, + 566, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a0056c4cda138179dc44ec507a89dfa6785467f6ab0b84624bb34a96f4607a1.jpg", + "image_caption": [ + "c) Metric feature learning with unknown targets" + ], + "image_footnote": [], + "bbox": [ + 635, + 87, + 810, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic Distance. Given two points $\\pmb{u} \\in \\mathbb{B}^n$ and $\\pmb{v} \\in \\mathbb{B}^n$ , the hyperbolic distance is defined as,", + "bbox": [ + 76, + 354, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathbb {B} ^ {n}} (\\boldsymbol {u}, \\boldsymbol {v}) = \\operatorname {a r c o s h} \\left(1 + 2 \\frac {\\| \\boldsymbol {u} - \\boldsymbol {v} \\| ^ {2}}{(1 - \\| \\boldsymbol {u} \\| ^ {2}) (1 - \\| \\boldsymbol {v} \\| ^ {2})}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 395, + 468, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{arcosh}$ is the inverse hyperbolic cosine function and $\\|\\cdot\\|$ is the usual Euclidean norm.", + "bbox": [ + 76, + 441, + 468, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic distance has the unique property that it grows exponentially as we move towards the boundary of the Poincaré ball. In particular, the points on the circle represent points in infinity. Hyperbolic space is naturally suitable for embedding hierarchical structure [35, 37] and can be regarded as a continuous representation of trees [6]. The hyperbolic distance between samples implicitly reflects their hierarchical relation. Thus, by embedding images in hyperbolic space we can naturally organize images based on their semantic similarity and prototypicality.", + "bbox": [ + 75, + 472, + 468, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Build Instance Hierarchy in Hyperbolic Space", + "text_level": 1, + "bbox": [ + 76, + 633, + 465, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic space is naturally suitable for embedding tree structure. However, in order to leverage hyperbolic space to build a sample hierarchy in an unsupervised manner, a suitable objective function is still missing. There are two challenges in designing the objective function. First, the underlying tree structure of the samples is unknown. Second, how to perform feature learning such that hierarchy can naturally emerge is unclear. In this paper, we propose Hyperbolic space with sphere pACKing, also called HACK, to address the two challenges.", + "bbox": [ + 75, + 657, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the first challenge, instead of creating a predefined tree structure that might not faithfully represent the genuine hierarchical organization, we leverage sphere packing in hyperbolic space for building a skeleton for placing the samples. We specify where the particles should be located ahead of training with uniform packing, which by design are", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maximally evenly spread out in hyperbolic space. The uniformly distributed particles guide feature learning to achieve maximum instance discrimination [45] while enabling us to extract a tree structure from the samples.", + "bbox": [ + 496, + 354, + 893, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the second challenge, HACK figures out which instance should be mapped to which target through bipartite graph matching as a global optimization procedure. During training, HACK minimizes the total hyperbolic distances between the mapped image point (in the feature space) and the target, those that are more typical naturally emerge closer to the origin of Poincaré ball. HACK differs from the existing learning methods in several aspects (Figure 4). Different from supervised learning, HACK allows the image to be assigned to any target (particle). This enables the exploration of the natural organization of the data. Different from unsupervised learning method, HACK specifies a predefined geometrical organization which encourages the corresponding structure to be emerged from the dataset.", + "bbox": [ + 496, + 416, + 893, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Sphere Packing in Hyperbolic Space", + "text_level": 1, + "bbox": [ + 500, + 643, + 815, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $n$ particles, our goal is to pack the particles into a two-dimensional hyperbolic space as densely as possible. We derive a simple repulsion loss function to encourage the particles to be equally distant from each other. The loss is derived via the following steps. First, we need to determine the radius of the Poincaré ball used for packing. We use a curvature of 1.0 so the radius of the Poincaré ball is 1.0. The whole Poincaré ball cannot be used for packing since the volume is infinite. We use $r < 1$ to denote the actual radius used for packing. Thus, our goal is to pack $n$ particles in a compact subspace of Poincaré ball. Then, the Euclidean radius $r$ is further converted into hyperbolic radius $r_{\\mathbb{B}}$ . Let $s = \\frac{1}{\\sqrt{c}}$ , where $c$ is the curvature. The relation between $r$ and $r_{\\mathbb{B}}$ is $r_{\\mathbb{B}} = s \\log \\frac{s + r}{s - r}$ . Next, the total hyperbolic area $A_{\\mathbb{B}}$ of a Poincaré ball of radius $r_{\\mathbb{B}}$ can", + "bbox": [ + 496, + 670, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23202", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/55191b4684c72402edc1035939655b1c1b6991e08d2593be404a320cc4060c43.jpg", + "image_caption": [ + "Figure 5. HACK conducts unsupervised learning in hyperbolic space with sphere packing. The images are mapped to particles by minimizing the total hyperbolic distance. HACK learns features that can capture both visual similarities and prototypicality." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 478, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be computed as $A_{\\mathbb{B}} = 4\\pi s^2 \\sinh^2 \\left( \\frac{r_{\\mathbb{B}}}{2s} \\right)$ , where $\\sinh$ is the hyperbolic sine function. Finally, the area per point $A_{n}$ can be easily computed as $\\frac{A_{\\mathbb{B}}}{n}$ , where $n$ is the total number of particles. Given $A_{n}$ , the radius per point can be computed as $r_{n} = 2s \\sinh^{-1} \\left( \\sqrt{\\frac{A_{n}}{4\\pi s^{2}}} \\right)$ . We use the following loss to generate uniform packing in hyperbolic space. Given two particles $i$ and $j$ , the repulsion loss $V$ is defined as,", + "bbox": [ + 75, + 287, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV (i, j) = \\left\\{\\frac {1}{\\left[ 2 r _ {n} - \\max \\left(0 , 2 r _ {n} - d _ {\\mathbb {B}} (i , j)\\right) \\right] ^ {k}} - \\frac {1}{\\left(2 r _ {n}\\right) ^ {k}} \\right\\} \\cdot C (k) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 410, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C(k) = \\frac{(2r_n)^{k + 1}}{k}$ and $k$ is a hyperparameter. Intuitively, if the particle $i$ and the particle $j$ are within $2r_n$ , the repulsion loss is positive. Minimizing the repulsion loss would push the particles $i$ and $j$ away. If the repulsion is zero, this indicates all the particles are equally distant. Also the repulsion loss grows significantly when two particles become close. We also adopt the following boundary loss to prevent the particles from escaping the ball,", + "bbox": [ + 75, + 449, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB (i; r) = \\max \\left(0, \\operatorname {n o r m} _ {i} - r + \\text {m a r g i n}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 585, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{norm}_i$ is the $\\ell_2$ norm of the representation of the particle $i$ . Figure 4 b) shows an example of the generated particles that are uniformly packed in hyperbolic space.", + "bbox": [ + 75, + 612, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Hyperbolic Instance Assignment", + "text_level": 1, + "bbox": [ + 76, + 665, + 362, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HACK learns the features by optimizing the assignments of the images to particles (Figure 5). The assignment should be one-to-one, i.e., each image is assigned to one particle and each particle is allowed to be associated with one image. We cast the instance assignment problem as a bipartite matching problem [12] and solve it with Hungarian algorithm [32].", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, we randomly assign the particles to the images, thus there is a random one-to-one correspondence between the images to the particles (not optimized). Given a batch of samples $\\{(\\mathbf{x}_1,\\mathbf{s}_1),(\\mathbf{x}_2,\\mathbf{s}_2),\\dots,(\\mathbf{x}_B,\\mathbf{s}_B)\\}$ , where $\\mathbf{x}_i$ is an image and $\\mathbf{s}_i$ is the corresponding particle, and an encoder $f_{\\theta}$ , we generate the hyperbolic feature for each image $\\mathbf{x}_i$ as $f_{\\theta}(\\mathbf{x}_i) \\in \\mathbb{B}^2$ , where $\\mathbb{B}^2$ is a two-dimensional Poincaré ball. For a given hyperbolic feature $f_{\\theta}(\\mathbf{x})$ , with fixed particle", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 HACK: Unsupervised Learning in Hyperbolic Space. ised Learning in Hyperbolic Space.", + "bbox": [ + 519, + 104, + 893, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: # of images: $n \\geq 0$ . Radius for packing: $r < 1$ . An encoder with parameters $\\theta$ : $f_{\\theta}$", + "bbox": [ + 521, + 137, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Generate uniformly distributed particles in hyperbolic space by minimizing the repulsion loss in Equation 2", + "2: Given $\\{(\\mathbf{x}_1, s_1), (\\mathbf{x}_2, s_2), \\dots, (\\mathbf{x}_b, s_b)\\}$ , optimize $f_{\\theta}$ by minimizing the total hyperbolic distance via Hungarian algorithm." + ], + "bbox": [ + 531, + 167, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "locations, the distance between the hyperbolic feature and the particles signifies the hierarchical level of the associated sample. Thus, to determine the hierarchical levels for all samples within the hierarchy, we must establish a one-to-one mapping between all the samples and the particles. This can be cast as the following bipartite matching problem in hyperbolic space,", + "bbox": [ + 496, + 289, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\theta , \\pi) = \\sum_ {i = 1} ^ {B} d _ {\\mathbb {B} ^ {n}} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {s} _ {\\pi \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right)\\right)}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 404, + 890, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi : f_{\\theta}(\\mathbf{x}) \\to \\mathbb{N}$ is a projection function which projects hyperbolic features to a particle index. Minimizing $\\ell(\\theta, \\pi)$ with respect to $\\pi$ is a combinatorial optimization problem, which can hardly be optimized with $\\theta$ using gradient-based algorithms. Thus, we adopt a joint optimization strategy which optimizes $\\theta$ and $\\pi$ alternatively. In each batch, we first leverage the Hungarian algorithm [32] to find the optimal matching $\\pi^{*}$ based on the current hyperbolic features. Then we minimize Eq. 4 with respect to $\\theta$ based on the current assignment $\\pi^{*}$ . This process is repeated for a certain number of epochs until convergence is achieved. On the other hand, the feature encoder can serve as an image prior for assigning similar images to nearby particles [41].", + "bbox": [ + 496, + 454, + 892, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Hungarian algorithm [32] has a complexity of $\\mathcal{O}(x^3)$ , where $x$ is the number of items. As we perform the particle assignment in the batch level, the time and memory complexity is tolerable. Also, the one-to-one correspondence between the images and particles is always maintained during training. After training, based on the assigned particle, the level of the sample in the hierarchy can be easily retrieved. The details of HACK are shown in Algorithm 1.", + "bbox": [ + 496, + 650, + 893, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 784, + 632, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design several experiments to show the effectiveness of HACK for the semantic and hierarchical organization. First, we first construct a dataset with known hierarchical structure using the congealing algorithm [31]. Then, we apply HACK to datasets with unknown hierarchical structure to organize the samples based on the semantic and prototypical structure.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we show that the prototypical structure can be used to reduce sample complexity and increase model robustness.", + "bbox": [ + 76, + 90, + 470, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We first construct a dataset called Congeaed MNIST. To verify the efficacy of HACK for unsupervised prototypicality discovery, we need a benchmark with known prototypical examples. However, currently there is no standard benchmark for this purpose. To construct the benchmark, we use the congealing algorithm from Miller et al. [31] to align the images in each class of MNIST [25]. The congealing algorithm is initially used for one-shot classification. During congealing, the images are brought into correspondence with each other jointly. The congealed images are more prototypical: they are better aligned with the average image. The synthetic data is generated by replacing 500 original images with the corresponding congealed images. In the Appendix, we show the results of changing the number of replaced original images. We expect HACK to discover the congealed images and place them in the center of the Poincaré ball. We also aim to discover the prototypical examples from each class of the standard MNIST dataset [25] and CIFAR10 [24]. CIFAR10 consists of 60000 from 10 object categories ranging from airplane to truck. CIFAR10 is more challenging than MNIST since it has larger intra-class variations. Moreover, to better visualize how HACK arranges the samples according to their prototypicality, we run HACK on 10k US Adult Faces [2] (hereafter referred to as USA10kF), which contains 10,168 natural face photographs.", + "bbox": [ + 76, + 125, + 472, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider several existing metrics proposed in Carlini et al. [5] for prototypicality discovery, the details can be found in the Appendix.", + "bbox": [ + 76, + 501, + 470, + 547 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Holdout Retraining [5]: We consider the Holdout Retraining proposed in Carlini et al. [5]. The idea is that the distance of features of prototypical examples obtained from models trained on different datasets should be close.", + "- Model Confidence [5]: Intuitively, the model should be confident in prototypical examples. Thus, it is natural to use the confidence of the model prediction as the criterion for prototypicality.", + "- UHML [46]: UHML is an unsupervised hyperbolic learning method that aims to discover the natural hierarchies of data by taking advantage of hyperbolic metric learning and hierarchical clustering." + ], + "bbox": [ + 76, + 550, + 470, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement HACK in PyTorch and the code will be made public. To generate uniform particles, we first randomly initialize the particles and then run the training for 1000 epochs with a 0.01 learning rate to minimize the repulsion loss and boundary loss. The curvature of the Poincaré ball is 1.0 and the $r$ is 0.76 which is used to alleviate the numerical issues [16]. The hyperparameter $k$ is 1.55 which is shown to generate uniform particles well. For the assignment, we use a LeNet [26] for MNIST, a ResNet20 [17] for CIFAR10, and a ResNet18 for USA10kF as the encoder. We apply HACK to each class separately.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54cedbae1e2a22363ac63e66e749e4faffb04e42e2a4d99b18a936698fe67fd5.jpg", + "image_caption": [ + "Figure 6. Hyperbolic space can capture the prototypicality inherently. The error bar of each point is given by the variance of density within the corresponding portion, and the width of the shaded band indicates the number of features within the portion." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 844, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5b57a190c1e8e84c5405ed046eb5cf8fa8a9351c9f5949bceb2b4f513a62d6a.jpg", + "image_caption": [ + "a)", + "Figure 7. Congeaed images are located in the center of the Poincaré ball. a) Red dots denote congealed images and cyan dots denote original images. b) Typical images are in the center and atypical images are close to the boundary. Images are also clustered together based on visual similarity. Congeaed images are shown in red boxes." + ], + "image_footnote": [], + "bbox": [ + 513, + 299, + 681, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/baefb88089f2b9b7a214982cb21013a009ab59dca4fbbfa0c46761a2468ebb6e.jpg", + "image_caption": [ + "b)" + ], + "image_footnote": [], + "bbox": [ + 681, + 299, + 852, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We attach a fully connected layer to project the feature into a two-dimensional Euclidean space. The image features are then projected onto hyperbolic space via an exponential map. We run the training for 200 epochs using a cosine learning rate scheduler [29] with an initial learning rate of 0.1. We optimize the assignment every other epoch. All the experiments are run on an NVIDIA Titan RTX GPU.", + "bbox": [ + 496, + 575, + 893, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Prototypicality in the Hyperbolic Feature Norm", + "text_level": 1, + "bbox": [ + 498, + 695, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We explicitly show that the hyperbolic space can capture prototypicality by analyzing the relation between hyperbolic norms and the K-NN density estimation. Taking the learned hyperbolic features, we first divide the range of norms of hyperbolic features into numerous portions with equal length (50 portions for this plot). The mean K-NN density is calculated by averaging the density estimation of features within each portion. Figure 6 shows that the mean density drops as the norm increases, which shows that the prototypicality emerges automatically within the norms, the inherent characteristic of hyperbolic space. This validates that prototypicality is reflected in the hyperbolic feature norm.", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5290b646bd25251ae3bba89047cb5c40f749d28c5f59b32db759504005a724d9.jpg", + "image_caption": [ + "Figure 8. Our unsupervised learning methods conform to our visual perception. The images are organized from left to right, top to bottom to cover the 360 degrees at the same radius." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 477, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ced45d0b8401eb597f4e9ac3ca5a7636710f36f4b0a359fe299da78cfd610503.jpg", + "image_caption": [ + "Figure 9. HACK discovers typical and atypical images from the data. First row: atypical images with large hyperbolic norm. Second row: typical images with small hyperbolic norm." + ], + "image_footnote": [], + "bbox": [ + 96, + 318, + 460, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Visual Prototypicality", + "text_level": 1, + "bbox": [ + 76, + 465, + 279, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Congeaed MNIST. We further apply HACK for visual feature learning on congealed MNIST. Figure 7 shows that HACK can discover the congealed images from all images. In Figure 7 a), the red particles denote the congealed images and cyan particles denote the original images. We can observe that the congealed images are assigned to the particles located in the center of the Poincaré ball. This verifies that HACK can indeed discover prototypical examples from the original dataset. In the Appendix, we show that the features of atypical examples gradually move to the boundary of the Poincaré ball during training. In Figure 7 b), we show the actual images that are embedded in the two-dimensional hyperbolic space. We can observe that the images in the center of Poincaré ball are more prototypical and images close to the boundary are more atypical. Also, the images are naturally organized by their semantic similarity. In summary, HACK can discover prototypicality and also organize the images based on their semantic and hierarchical structure. To the best of our knowledge, this is the first unsupervised learning method that can be used to discover prototypical examples in a data-driven fashion.", + "bbox": [ + 75, + 491, + 470, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "USA10kF. Figure 10 a) shows the assignment of 2000 images sampled from USA10kF. Compared to MNIST, the variation in faces is much more complex. A facial image is also subject to various factors (such as race, facial expres", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fb203d23b82dadb6c754b9fc394d4ebcc1286b17e22c7abf5b827b58812a213.jpg", + "image_caption": [ + "Figure 10. HACK captures prototypicality of faces. a) The assignment of 2000 images sampled from USA10kF. b) The angles indicate the uniform division of the space. For each orientation, 8 equidistant sample points were selected and the nearest faces are shown at each point. The face located in the center is most symmetrical with clear smiling" + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 705, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/182ae3765979e2e5ba3fd745f897d65d6c093e11dfc4624c5b218a10383bff9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 94, + 897, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sion, environmental condition, etc.). Therefore, the results from USA10kF are less intuitive than those from MNIST. However, Figure 10 b) illustrates the evolutionary process of images originating from the center of hyperbolic space and progressing along different directions. In the space, we selected 12 directions at equal angular intervals and chose five equally spaced sampling points in each direction. The images closest to these sampling points are displayed at their respective locations. Although the detailed organization is unclear, the evolutionary process reveals a tendency of HACK to cluster different features together, such as darker skin tones appearing in the 0-90 degree range and lighter skin tones in the 180-270 degree range.", + "bbox": [ + 496, + 383, + 892, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MNIST and CIFAR10. Figure 8 shows the embedding of class 0 from MNIST and class \"airplane\" from CIFAR10 arranged to cover 360 degrees of at the same radius. The visual similarity of the images has a smooth transition as we move around angularly. Figure 9 shows the typical images and atypical images discovered by HACK. This further illustrates that HACK captures the semantic similarity of the images which enables prototypicality discovery. Please see the Appendix for more results.", + "bbox": [ + 496, + 579, + 893, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Prototypicality for Instance Selection", + "text_level": 1, + "bbox": [ + 500, + 726, + 818, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 12 shows the comparison of the baselines with HACK. With HACK, typical images are characterized by the smallest hyperbolic norms, whereas atypical images are associated with the largest hyperbolic norms. We can observe that both HACK and Model Confidence (MC) can discover typical and atypical images. Compared with MC, HACK defines prototypicality as the distance of the sample to other samples which is more aligned with human intuition. Moreover, in addition to prototypicality, HACK can also be used to organize examples by semantic similarities. Holdout Retraining", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c05788586425e73d576dd78e4898a73937d90946ee5911e53b131820a60dbdb.jpg", + "image_caption": [ + "a) Reducing sample complexity" + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 289, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d85ab442833a112a00ad3d812c1c3c141d6095c964e89a8d21563522f3569089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 90, + 483, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/386b08f9ce6773421c28f95dfd7aa2bc350035b9715c24859845bbac42578507.jpg", + "image_caption": [ + "b) Increasing model robustness" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 679, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88350d94a95e075163cd1a55a0e1f34571fd89ebd9016c3ee610ac6f2c79c569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 90, + 872, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b83c9dba7df38a388ffe1157b9adef0de0c2497ea12d41b52cb29ce2debf743e.jpg", + "image_caption": [ + "Figure 11. HACK can be used to construct sample prototypical hierarchy which is useful for several downstream tasks. a) Training with atypical examples achieves higher accuracy than training with typical examples. b) The adversarial accuracy greatly improves after removing the X% of most atypical examples.", + "Figure 12. HACK can capture both prototypicality and diversity in the dataset. Each row shows the typical and atypical images discovered by HACK/HR [5]/MC [5]." + ], + "image_footnote": [], + "bbox": [ + 86, + 297, + 468, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(HR) is not effective for prototypicality discovery due to the randomness of model training.", + "bbox": [ + 75, + 474, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Application of Prototypicality", + "text_level": 1, + "bbox": [ + 76, + 513, + 341, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reducing Sample Complexity. The proposed HACK can discover prototypical images as well as atypical images. We show that with atypical images we can reduce the sample complexity for training the model. Prototypical images are representative of the dataset but lack variations. Atypical examples contain more variations and it is intuitive that models trained on atypical examples should generalize better to the test samples. To verify this hypothesis, we select a subset of samples based on the norm of the features which indicates prototypicality of the examples. In particular, typical samples correspond to the samples with smaller norms and atypical samples correspond to the samples with larger norms. The angular layout of the hyperbolic features naturally captures sample diversity, thus for selecting atypical examples, we consider introducing more diversity by sampling images with large norms along the angular direction.", + "bbox": [ + 75, + 537, + 470, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 11 a) shows that training with atypical images can achieve much higher accuracy than training with typical images. In particular, training with the most atypical $10\\%$ of the images achieves $22.67\\%$ higher accuracy than with the most typical $10\\%$ of the images on CIFAR10. Similar results can be observed on MNIST. Thus, HACK provides an easy solution to reduce sample complexity. We also compared UHML [46], which is an unsupervised metric learning in", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hyperbolic space, with HACK on the MNIST dataset. By incorporating $10\\%$ atypical samples based on feature norm, HACK outperformed UHML by $10.2\\%$ . Also by excluding the $1\\%$ atypical examples, HACK achieved an additional $5.7\\%$ improvement over UHML.", + "bbox": [ + 496, + 299, + 893, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Increasing Model Robustness. Training models with atypical examples can lead to a vulnerable model to adversarial attacks [5, 28]. Intuitively, atypical examples lead to a less smooth decision boundary thus a small perturbation to examples is likely to change the prediction. With HACK, we can easily identify atypical samples to improve the robustness of the model. We use MNIST and CIFAR 10 as the benchmark and use FGSM [13] to attack the model with an $\\epsilon$ of 0.07 on MNIST and $8/(255^{*}std)$ on CIFAR10, where $std$ is the standard deviation used for normalization. More details of the attack settings can be found in the appendix. We identify the atypical examples based on the norm of the features with HACK and remove the most atypical $X\\%$ of the examples. Figure 11 b) shows that discarding atypically examples greatly improves the robustness of the model: the adversarial accuracy is improved by $8.7\\%$ when excluding the most atypical $1\\%$ of examples on MNIST and $7.3\\%$ on CIFAR10. It is worth noting that the clean accuracy remains the same after removing a small number of atypical examples.", + "bbox": [ + 496, + 380, + 893, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Summary", + "text_level": 1, + "bbox": [ + 500, + 688, + 607, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose HACK, an unsupervised learning method that organizes images in hyperbolic space using sphere packing. By optimizing image assignments to uniformly distributed particles, HACK leverages the inherent properties of hyperbolic space, leading to the natural emergence of prototypical and semantic structures through feature learning. We validate HACK on synthetic data and standard datasets, demonstrating its ability to discover prototypical examples for reducing sample complexity and increasing model robustness.", + "bbox": [ + 496, + 714, + 893, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project was supported, in part, by NSF 2131111, NSF 2215542, and NSF 2313151 to S. Yu, and a grant from UT Dallas to Y. Guo.", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23206", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James W Anderson. Hyperbolic geometry. Springer Science & Business Media, 2006. 2", + "[2] Wilma A Bainbridge, Phillip Isola, and Aude Oliva. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General, 142(4):1323, 2013. 6", + "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 1", + "[4] Jacob Bien and Robert Tibshirani. Prototype selection for interpretable classification. The Annals of Applied Statistics, 5(4):2403-2424, 2011. 1", + "[5] Nicholas Carlini, Ulfar Erlingsson, and Nicolas Papernot. Prototypical examples in deep learning: Metrics, characteristics, and utility. 2018. 1, 2, 6, 8", + "[6] Ines Chami, Albert Gu, Vaggos Chatziafratis, and Christopher Ré. From trees to continuous embeddings and back: Hyperbolic hierarchical clustering. Advances in Neural Information Processing Systems, 33:15065-15076, 2020. 4", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[8] Lieven Decock and Igor Douven. What is graded membership? Nous, 48(4):653-82, 2014. 1", + "[9] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 836-837, 2020. 3", + "[10] Evelyn Fix and Joseph Lawson Hodges. Discriminatory analysis. nonparametric discrimination: Consistency properties. International Statistical Review/Revue Internationale de Statistique, 57(3):238-247, 1989. 3", + "[11] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. arXiv preprint arXiv:1805.09112, 2018. 2", + "[12] Alan Gibbons. Algorithmic graph theory. Cambridge university press, 1985. 5", + "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 8, 1", + "[14] Yanming Guo, Yu Liu, Erwin M Bakker, Yuanhao Guo, and Michael S Lew. Cnn-rnn: a large-scale hierarchical image classification framework. Multimedia tools and applications, 77(8):10251-10271, 2018. 3", + "[15] Yunhui Guo, Haoran Guo, and Stella Yu. Co-sne: Dimensionality reduction and visualization for hyperbolic data. arXiv preprint arXiv:2111.15037, 2021. 3", + "[16] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 6", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2019. 1", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729–9738, 2020. 2", + "[20] Joy Hsu, Jeffrey Gu, Gong-Her Wu, Wah Chiu, and Serena Yeung. Learning hyperbolic representations for unsupervised 3d segmentation. arXiv preprint arXiv:2012.01644, 2020. 3", + "[21] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 3", + "[22] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in neural information processing systems, 29, 2016. 2", + "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[25] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 6, 1", + "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 1", + "[27] Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2", + "[28] Yongshuai Liu, Jiyu Chen, and Hao Chen. Less is more: Culling the training set to improve robustness of deep neural networks. In International Conference on Decision and Game Theory for Security, pages 102-114. Springer, 2018. 8", + "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6, 1", + "[30] Emile Mathieu, Charline Le Lan, Chris J Maddison, Ryota Tomioka, and Yee Whye Teh. Continuous hierarchical representations with poincar\\`e variational auto-encoders. arXiv preprint arXiv:1901.06033, 2019. 3", + "[31] Erik G Miller, Nicholas E Matsakis, and Paul A Viola. Learning from one example through shared densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 464-471. IEEE, 2000. 1, 2, 3, 5, 6", + "[32] James Munkres. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 5(1):32-38, 1957. 5", + "[33] Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A wrapped normal distribution on" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hyperbolic space for gradient-based learning. In International Conference on Machine Learning, pages 4693-4702. PMLR, 2019. 2, 3", + "[34] Maximilian Nickel and Douwe Kiela. Poincar\\`e embeddings for learning hierarchical representations. arXiv preprint arXiv:1705.08039, 2017. 2", + "[35] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 2, 3, 4", + "[36] Eleanor Rosch. Cognitive representations of semantic categories. Journal of experimental psychology: General, 104(3): 192, 1975. 1", + "[37] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International Symposium on Graph Drawing, pages 355-366. Springer, 2011. 4", + "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2", + "[39] Pierre Stock and Moustapha Cisse. Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In Proceedings of the European Conference on Computer Vision (ECCV), pages 498-512, 2018. 2", + "[40] Richard Szeliski et al. Image alignment and stitching: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2(1):1-104, 2007. 3", + "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 5", + "[42] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 1", + "[43] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 2", + "[44] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2603-2612, 2021. 3", + "[45] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 4, 2", + "[46] Jiei Xi Yan, Lei Luo, Cheng Deng, and Heng Huang. Unsupervised hyperbolic metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12465-12474, 2021. 6, 8", + "[47] Jianping Zhang. Selecting typical instances in instance-based learning. In Machine learning proceedings 1992, pages 470-479. Elsevier, 1992. 2" + ], + "bbox": [ + 78, + 90, + 470, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23208", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_model.json b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac5977b65d21004a520b745d271572b7eff4a1e6 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.867, + 0.153 + ], + "angle": 0, + "content": "Unsupervised Feature Learning with Emergent Data-Driven Prototypicality" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.179, + 0.87, + 0.218 + ], + "angle": 0, + "content": "Yunhui Guo1 Youren Zhang2 Yubei Chen3 Stella X. Yu2,4 \n1The University of Texas at Dallas 2University of Michigan 3UC Davis 4UC Berkeley" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.471, + 0.345 + ], + "angle": 0, + "content": "Given a set of images, our goal is to map each image to a point in a feature space such that, not only point proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflects image similarity, yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.471, + 0.709 + ], + "angle": 0, + "content": "We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKKing. HACK first generates uniformly packed particles in the Poincaré ball of hyperbolic space and then assigns each image uniquely to a particle. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing - a warping process that aligns all the images and makes them appear more common and similar to each other, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.208, + 0.754 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Not all instances are created equal. For example, the MNIST dataset of handwritten digits contain almost 6,000 samples of 2's; some are close to textbook versions that we are taught to follow, whereas others have idiosyncratic cursive styles, varying in proportions and stroke weights (Fig. 1). Given such a set of natural data, we are interested in dataset summarization and organization such that we can automatically discover which instances are more representative and which ones are anomalies. In other words, we aim to computation-" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.25, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.418, + 0.895, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Given a dataset (left), we aim to learn an image feature that encodes not only visual similarity between instances but also data-driven prototypicality (right). Additionally, the angular arrangement of the features can naturally serve as a measure of diversity. Our feature encoder (in 2D hyperbolic space) is learned without any labels. We can then learn a decoder to map each point in the feature space back to an image. The right plot visualizes images located at the origin and those moving away in different directions, automatically revealing that 2's with loops are most common and the whole dataset can be grasped as the cursive style systematically varies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.891, + 0.658 + ], + "angle": 0, + "content": "ally rationalize graded membership [8, 36] of a category in a purely data-driven manner, putting each instance on a feature map that reflects not only their prototypicality in a particular dataset but also their visual similarity with each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Such unsupervised feature learning is useful for not only data organization, but those discovered representative instances or prototypes can be used for interpretable machine learning [4], curriculum learning [3], and learning better decision boundaries [5]. Prototypes also allow us to classify with as few as or even one example [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.84 + ], + "angle": 0, + "content": "If the image feature is given, it is relatively easy to find prototypes: We just need to identify density peaks of the feature distribution of the image set. Otherwise, discovering prototypical instances without supervision is difficult: There is no universal definition or simple metric to assess the prototypicality of the examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "One way to address this problem is to examine the gradient magnitude [5]. However, this approach is shown to have a high variance which results from different training setups [5]. Some methods address this problem using adversarial" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.434, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.471, + 0.379 + ], + "angle": 0, + "content": "Figure 2. Existing self-supervised feature learning methods focus on visual similarity only in Euclidean space, whereas our unsupervised feature learning method embeds images in hyperbolic space which automatically encodes prototypicality with respect to the origin: Images are organized hierarchically, with typical images at the center and atypical ones near the boundary of the Poincaré ball." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.485 + ], + "angle": 0, + "content": "robustness [5, 39]: Prototypical examples should be more adversarially robust. However, selecting these examples is dependent on the adversarial method and the metric used in the adversarial attack. Several methods exist but they are either based on heuristics or lack a proper justification [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Naturally, given a feature space, prototypical examples can be identified as density peaks. However, prototypicality varies with the feature. We propose an unsupervised feature learning method in hyperbolic space, which, unlike the shift-invariant Euclidean space, naturally embeds a continuous version of hierarchical trees rooted at the origin of the space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Hyperbolic space is non-Euclidean space with constant non-negative curvature [1]. Different from Euclidean space, hyperbolic space can represent hierarchical relations with low distortion. Poincaré ball model is one of the most commonly used models for hyperbolic space [35]. One notable property of Poincaré ball model is that the distance to the origin grows exponentially as we move towards the boundary. Thus, points located in the center of the ball are close to all the other points while the points located close to the boundary are infinitely far away from other points. With unsupervised learning in hyperbolic space, we can learn features that capture both visual similarity and hierarchical proximity among instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our key insight is that a typical image is similar (closest) to more nearby instances than atypical ones, and such an image would be at the center of the neighbourhood in Euclidean space but the parent/root node in a tree in hyperbolic space, closer to the origin. We develop a new learning procedure that first places all the target locations evenly in the Poincaré ball to reflect the desire of mapping uniformity [43], and then we optimize which images should be mapped to which" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.136 + ], + "angle": 0, + "content": "target locations in a batch-wise Hungarian matching manner, where those similar to most instances naturally moving closer to the origin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Our work makes the following contributions. 1) We propose the first unsupervised feature learning method to learn features which capture visual similarity with distance between features and prototypicality with the distance to the origin. 2) We develop a new learning paradigm that sits between supervised learning (with known targets) and unsupervised metric learning (with unknown targets and constrained metric distances). We want to map images to known targets uniformly packed and maximally distant in hyperbolic space, but we learn to optimize the specific image to target assignment. 3) We validate our joint feature learning and data prototypicality discovery on congealing [31], where the consensus is that images after congealing are perceived to be more typical images. 4) We demonstrate two practical usages of data-driven prototypicality: Prototypical and atypical examples are shown to reduce sample complexity for learning and increase the robustness of the model respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.642, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Prototypicality. The study of prototypical examples in machine learning has a long history. In Zhang [47], the authors select typical instances based on the fact that typical instances should be representative of the cluster. In Kim et al. [22], prototypical examples are defined as the examples that have maximum mean discrepancy within the data. Li et al. [27] propose to discover prototypical examples by architectural modifications: project the dataset onto a low-dimensional manifold and use a prototype layer to minimize the distance between inputs and the prototypes on the manifold. The robustness to adversarial attacks is also used as a criterion for prototypicality [39]. In Carlini et al. [5], the authors propose multiple metrics for prototypicality discovery. For example, the features of prototypical examples should be consistent across different training setups. However, these metrics usually depend heavily on the training setups and hyperparameters. The idea of prototypicality is also extensively studied in meta-learning for one-shot or few-shot classification [38]. No existing works address the prototypicality discovery problem in a data-driven fashion. Our proposed HACK naturally exploits hyperbolic space to organize the images based on prototypicality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Unsupervised Learning in Hyperbolic Space. Learning features in hyperbolic space have shown to be useful for many machine learning problems [11, 34]. One useful property is that hierarchical relations can be embedded in hyperbolic space with low distortion [34]. Wrapped normal distribution, which is a generalized version of the normal distribution for modeling the distribution of points in hyperbolic space [33], is used as the latent space for constructing hyperbolic variational autoencoders (VAEs)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.09, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.228 + ], + "angle": 0, + "content": "Figure 3. Congeaed images are more typical than the original images. First row: sampled original images. Second row: the corresponding congealed images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.473, + 0.422 + ], + "angle": 0, + "content": "[23]. Poincaré VAEs is constructed in Mathieu et al. [30] with a similar idea to Nagano et al. [33] by replacing the standard normal distribution with hyperbolic normal distribution. Unsupervised 3D segmentation [20] and instance segmentation [44] are conducted in hyperbolic space via hierarchical hyperbolic triplet loss. CO-SNE [15] is recently proposed to visualize high-dimensional hyperbolic features in a two-dimensional hyperbolic space. Although hyperbolic distance facilitates the learning of hierarchical structure, how to leverage hyperbolic space for unsupervised prototypicality discovery is not explored in the current literature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.254, + 0.469 + ], + "angle": 0, + "content": "3. Sample Hierarchy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Sample Hierarchy VS. Class Hierarchy. While most of the existing works in hierarchical image classification focus on using label hierarchy [9, 14], there also exists a natural hierarchy among different samples. In Khrulkov et al. [21], the authors conducted an experiment to measure the \\(\\delta\\)-hyperbolicity of the various image datasets and showed that common image datasets such as CIFAR10 and CUB exhibit natural hierarchical structure among the samples. Amongst a collection of images representing digit 1, suppose \\(\\mathbf{x}\\) is used for representing an image with a digit '1' that is upright, \\(\\mathbf{x}'\\) is used for representing an image with a digit 1 that leaning left and \\(\\mathbf{x}''\\) is used for representing an image with a digit '1' that leaning right. Given a metric \\(d(\\cdot, \\cdot)\\), if we assume that \\(d(\\mathbf{x}''', \\mathbf{x}') \\approx d(\\mathbf{x}''', \\mathbf{x}) + d(\\mathbf{x}', \\mathbf{x})\\), in this context, we can naturally view the sample \\(\\mathbf{x}\\) as the root, and consider the other samples as its children in an underlying tree." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Compared with class hierarchy which can be extracted based on the pre-defined label relations, sample hierarchy is much harder to construct due to the lack of ground truth. Once a sample hierarchy is established, there are currently no existing methods available for verifying the accuracy of the hierarchy. Additionally, just like with class hierarchies, there may be ambiguities when constructing a sample hierarchy since multiple samples could potentially serve as the root." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Building Sample Hierarchy from Density Peaks. Given existing features \\(\\{f(v_{i})\\}\\) obtained by applying a feature extractor for each instance \\(v_{i}\\), prototypical examples can be found by examining the density peaks via techniques from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.248 + ], + "angle": 0, + "content": "density estimation. For example, the K-nearest neighbor density (K-NN) estimation [10] is defined as \\( p_{knn}(v_i, k) = \\frac{k}{n} \\frac{1}{A_d \\cdot D^d(v_i, v_{k(i)})} \\), where \\( d \\) is the feature dimension, \\( A_d = \\pi^{d/2} / \\Gamma(d/2 + 1) \\), \\( \\Gamma(x) \\) is the Gamma function and \\( k(i) \\) is the \\( k \\)th nearest neighbor of example \\( v_i \\). The nearest neighbors can be found by computing the distance between the features. Therefore, the process of constructing sample hierarchy through density estimation can be conceptualized as a two-step procedure involving: 1) feature learning and 2) detecting density peaks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.895, + 0.399 + ], + "angle": 0, + "content": "In the density estimation approach outlined above, the level of prototypicality depends on the learned features. Varying training setups can induce diverse feature spaces, resulting in differing conclusions on prototypicality. Nevertheless, prototypicality is an inherent attribute of the dataset and should remain consistent across various features. The aim of this paper is to extract features that intrinsically showcase the hierarchical organization of the samples. Specifically, by examining the feature alone within the feature space, we should be able to identify the example's prototypicality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Construct a Sample Hierarchy from Congealing. To determine whether the feature truly captures prototypicality, it is necessary to identify which sample is the prototype. We ground our concept of prototypicality based on congealing [31]. In particular, we define prototypical examples in the pixel space by examining the distance of the images to the average image in the corresponding class. Our idea is based on a traditional computer vision technique called image alignment [40] that aims to find correspondences across images. During congealing [31], a set of images are transformed to be jointly aligned by minimizing the joint pixel-wise entropies. The congealed images are more prototypical: they are better aligned with the average image. Thus, we have a simple way to transform an atypical example into a typical example (see Figure 3). This is useful since given an unlabeled image dataset the typicality of the examples is unknown, congealing examples can be naturally served as examples with known typicality and be used as a validation for the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.7, + 0.888, + 0.718 + ], + "angle": 0, + "content": "4. Unsupervised Hyperbolic Feature Learning" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.725, + 0.86, + 0.742 + ], + "angle": 0, + "content": "4.1. Poincaré Ball Model for Hyperbolic Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Euclidean space has a curvature of zero and a hyperbolic space is a Riemannian manifold with constant negative curvature.. There are several isometrically equivalent models for visualizing hyperbolic space with Euclidean representation. The Poincaré ball model is the commonly used one in hyperbolic representation learning [35]. The \\(n\\)-dimensional Poincaré ball model is defined as \\((\\mathbb{B}^n, \\mathfrak{g}_{\\mathbf{x}})\\), where \\(\\mathbb{B}^n = \\{\\mathbf{x} \\in \\mathbb{R}^n : \\| \\mathbf{x} \\| < 1\\}\\) and \\(\\mathfrak{g}_{\\mathbf{x}} = (\\gamma_{\\mathbf{x}})^2 I_n\\) is the Riemannian metric tensor. \\(\\gamma_{\\mathbf{x}} = \\frac{2}{1 - \\| \\mathbf{x} \\|^2}\\) is the conformal factor and \\(I_n\\) is the Euclidean metric tensor." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23201" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.091, + 0.335, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.215, + 0.331, + 0.246 + ], + "angle": 0, + "content": "a) Supervised classification with fixed known targets" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.088, + 0.568, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.215, + 0.608, + 0.246 + ], + "angle": 0, + "content": "b) Our unsupervised feature learning with fixed but unknown targets" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.088, + 0.811, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.215, + 0.811, + 0.246 + ], + "angle": 0, + "content": "c) Metric feature learning with unknown targets" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 4. The proposed HACK has a predefined geometrical arrangement and allows the images to be freely assigned to any particle. a) Standard supervised learning has predefined targets. The image is only allowed to be assigned to the corresponding target. b) HACK packs particles uniformly in hyperbolic space to create initial seeds for the organization. The images are assigned to the particles based on their prototypicality and semantic similarities. c) Standard unsupervised learning has no predefined targets and images are clustered based on their semantic similarities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.469, + 0.385 + ], + "angle": 0, + "content": "Hyperbolic Distance. Given two points \\(\\pmb{u} \\in \\mathbb{B}^n\\) and \\(\\pmb{v} \\in \\mathbb{B}^n\\), the hyperbolic distance is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.396, + 0.47, + 0.432 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathbb {B} ^ {n}} (\\boldsymbol {u}, \\boldsymbol {v}) = \\operatorname {a r c o s h} \\left(1 + 2 \\frac {\\| \\boldsymbol {u} - \\boldsymbol {v} \\| ^ {2}}{(1 - \\| \\boldsymbol {u} \\| ^ {2}) (1 - \\| \\boldsymbol {v} \\| ^ {2})}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.471 + ], + "angle": 0, + "content": "where \\(\\operatorname{arcosh}\\) is the inverse hyperbolic cosine function and \\(\\|\\cdot\\|\\) is the usual Euclidean norm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.624 + ], + "angle": 0, + "content": "Hyperbolic distance has the unique property that it grows exponentially as we move towards the boundary of the Poincaré ball. In particular, the points on the circle represent points in infinity. Hyperbolic space is naturally suitable for embedding hierarchical structure [35, 37] and can be regarded as a continuous representation of trees [6]. The hyperbolic distance between samples implicitly reflects their hierarchical relation. Thus, by embedding images in hyperbolic space we can naturally organize images based on their semantic similarity and prototypicality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.467, + 0.651 + ], + "angle": 0, + "content": "4.2. Build Instance Hierarchy in Hyperbolic Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Hyperbolic space is naturally suitable for embedding tree structure. However, in order to leverage hyperbolic space to build a sample hierarchy in an unsupervised manner, a suitable objective function is still missing. There are two challenges in designing the objective function. First, the underlying tree structure of the samples is unknown. Second, how to perform feature learning such that hierarchy can naturally emerge is unclear. In this paper, we propose Hyperbolic space with sphere pACKing, also called HACK, to address the two challenges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To address the first challenge, instead of creating a predefined tree structure that might not faithfully represent the genuine hierarchical organization, we leverage sphere packing in hyperbolic space for building a skeleton for placing the samples. We specify where the particles should be located ahead of training with uniform packing, which by design are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.894, + 0.414 + ], + "angle": 0, + "content": "maximally evenly spread out in hyperbolic space. The uniformly distributed particles guide feature learning to achieve maximum instance discrimination [45] while enabling us to extract a tree structure from the samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.629 + ], + "angle": 0, + "content": "To address the second challenge, HACK figures out which instance should be mapped to which target through bipartite graph matching as a global optimization procedure. During training, HACK minimizes the total hyperbolic distances between the mapped image point (in the feature space) and the target, those that are more typical naturally emerge closer to the origin of Poincaré ball. HACK differs from the existing learning methods in several aspects (Figure 4). Different from supervised learning, HACK allows the image to be assigned to any target (particle). This enables the exploration of the natural organization of the data. Different from unsupervised learning method, HACK specifies a predefined geometrical organization which encourages the corresponding structure to be emerged from the dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.645, + 0.816, + 0.663 + ], + "angle": 0, + "content": "4.3. Sphere Packing in Hyperbolic Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Given \\( n \\) particles, our goal is to pack the particles into a two-dimensional hyperbolic space as densely as possible. We derive a simple repulsion loss function to encourage the particles to be equally distant from each other. The loss is derived via the following steps. First, we need to determine the radius of the Poincaré ball used for packing. We use a curvature of 1.0 so the radius of the Poincaré ball is 1.0. The whole Poincaré ball cannot be used for packing since the volume is infinite. We use \\( r < 1 \\) to denote the actual radius used for packing. Thus, our goal is to pack \\( n \\) particles in a compact subspace of Poincaré ball. Then, the Euclidean radius \\( r \\) is further converted into hyperbolic radius \\( r_{\\mathbb{B}} \\). Let \\( s = \\frac{1}{\\sqrt{c}} \\), where \\( c \\) is the curvature. The relation between \\( r \\) and \\( r_{\\mathbb{B}} \\) is \\( r_{\\mathbb{B}} = s \\log \\frac{s + r}{s - r} \\). Next, the total hyperbolic area \\( A_{\\mathbb{B}} \\) of a Poincaré ball of radius \\( r_{\\mathbb{B}} \\) can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.48, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.201, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Figure 5. HACK conducts unsupervised learning in hyperbolic space with sphere packing. The images are mapped to particles by minimizing the total hyperbolic distance. HACK learns features that can capture both visual similarities and prototypicality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.404 + ], + "angle": 0, + "content": "be computed as \\( A_{\\mathbb{B}} = 4\\pi s^2 \\sinh^2 \\left( \\frac{r_{\\mathbb{B}}}{2s} \\right) \\), where \\( \\sinh \\) is the hyperbolic sine function. Finally, the area per point \\( A_{n} \\) can be easily computed as \\( \\frac{A_{\\mathbb{B}}}{n} \\), where \\( n \\) is the total number of particles. Given \\( A_{n} \\), the radius per point can be computed as \\( r_{n} = 2s \\sinh^{-1} \\left( \\sqrt{\\frac{A_{n}}{4\\pi s^{2}}} \\right) \\). We use the following loss to generate uniform packing in hyperbolic space. Given two particles \\( i \\) and \\( j \\), the repulsion loss \\( V \\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.411, + 0.469, + 0.45 + ], + "angle": 0, + "content": "\\[\nV (i, j) = \\left\\{\\frac {1}{\\left[ 2 r _ {n} - \\max \\left(0 , 2 r _ {n} - d _ {\\mathbb {B}} (i , j)\\right) \\right] ^ {k}} - \\frac {1}{\\left(2 r _ {n}\\right) ^ {k}} \\right\\} \\cdot C (k) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.575 + ], + "angle": 0, + "content": "where \\( C(k) = \\frac{(2r_n)^{k + 1}}{k} \\) and \\( k \\) is a hyperparameter. Intuitively, if the particle \\( i \\) and the particle \\( j \\) are within \\( 2r_n \\), the repulsion loss is positive. Minimizing the repulsion loss would push the particles \\( i \\) and \\( j \\) away. If the repulsion is zero, this indicates all the particles are equally distant. Also the repulsion loss grows significantly when two particles become close. We also adopt the following boundary loss to prevent the particles from escaping the ball," + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.586, + 0.469, + 0.601 + ], + "angle": 0, + "content": "\\[\nB (i; r) = \\max \\left(0, \\operatorname {n o r m} _ {i} - r + \\text {m a r g i n}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.658 + ], + "angle": 0, + "content": "where \\(\\mathrm{norm}_i\\) is the \\(\\ell_2\\) norm of the representation of the particle \\(i\\). Figure 4 b) shows an example of the generated particles that are uniformly packed in hyperbolic space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.364, + 0.683 + ], + "angle": 0, + "content": "4.4. Hyperbolic Instance Assignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.78 + ], + "angle": 0, + "content": "HACK learns the features by optimizing the assignments of the images to particles (Figure 5). The assignment should be one-to-one, i.e., each image is assigned to one particle and each particle is allowed to be associated with one image. We cast the instance assignment problem as a bipartite matching problem [12] and solve it with Hungarian algorithm [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Initially, we randomly assign the particles to the images, thus there is a random one-to-one correspondence between the images to the particles (not optimized). Given a batch of samples \\(\\{(\\mathbf{x}_1,\\mathbf{s}_1),(\\mathbf{x}_2,\\mathbf{s}_2),\\dots,(\\mathbf{x}_B,\\mathbf{s}_B)\\}\\), where \\(\\mathbf{x}_i\\) is an image and \\(\\mathbf{s}_i\\) is the corresponding particle, and an encoder \\(f_{\\theta}\\), we generate the hyperbolic feature for each image \\(\\mathbf{x}_i\\) as \\(f_{\\theta}(\\mathbf{x}_i) \\in \\mathbb{B}^2\\), where \\(\\mathbb{B}^2\\) is a two-dimensional Poincaré ball. For a given hyperbolic feature \\(f_{\\theta}(\\mathbf{x})\\), with fixed particle" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.106, + 0.894, + 0.137 + ], + "angle": 0, + "content": "Algorithm 1 HACK: Unsupervised Learning in Hyperbolic Space. ised Learning in Hyperbolic Space." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.138, + 0.894, + 0.168 + ], + "angle": 0, + "content": "Require: # of images: \\(n \\geq 0\\). Radius for packing: \\(r < 1\\). An encoder with parameters \\(\\theta\\): \\(f_{\\theta}\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.169, + 0.894, + 0.211 + ], + "angle": 0, + "content": "1: Generate uniformly distributed particles in hyperbolic space by minimizing the repulsion loss in Equation 2" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.214, + 0.894, + 0.259 + ], + "angle": 0, + "content": "2: Given \\(\\{(\\mathbf{x}_1, s_1), (\\mathbf{x}_2, s_2), \\dots, (\\mathbf{x}_b, s_b)\\}\\), optimize \\(f_{\\theta}\\) by minimizing the total hyperbolic distance via Hungarian algorithm." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.169, + 0.894, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.396 + ], + "angle": 0, + "content": "locations, the distance between the hyperbolic feature and the particles signifies the hierarchical level of the associated sample. Thus, to determine the hierarchical levels for all samples within the hierarchy, we must establish a one-to-one mapping between all the samples and the particles. This can be cast as the following bipartite matching problem in hyperbolic space," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\ell (\\theta , \\pi) = \\sum_ {i = 1} ^ {B} d _ {\\mathbb {B} ^ {n}} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {s} _ {\\pi \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right)\\right)}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.652 + ], + "angle": 0, + "content": "where \\(\\pi : f_{\\theta}(\\mathbf{x}) \\to \\mathbb{N}\\) is a projection function which projects hyperbolic features to a particle index. Minimizing \\(\\ell(\\theta, \\pi)\\) with respect to \\(\\pi\\) is a combinatorial optimization problem, which can hardly be optimized with \\(\\theta\\) using gradient-based algorithms. Thus, we adopt a joint optimization strategy which optimizes \\(\\theta\\) and \\(\\pi\\) alternatively. In each batch, we first leverage the Hungarian algorithm [32] to find the optimal matching \\(\\pi^{*}\\) based on the current hyperbolic features. Then we minimize Eq. 4 with respect to \\(\\theta\\) based on the current assignment \\(\\pi^{*}\\). This process is repeated for a certain number of epochs until convergence is achieved. On the other hand, the feature encoder can serve as an image prior for assigning similar images to nearby particles [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.894, + 0.773 + ], + "angle": 0, + "content": "The Hungarian algorithm [32] has a complexity of \\(\\mathcal{O}(x^3)\\), where \\(x\\) is the number of items. As we perform the particle assignment in the batch level, the time and memory complexity is tolerable. Also, the one-to-one correspondence between the images and particles is always maintained during training. After training, based on the assigned particle, the level of the sample in the hierarchy can be easily retrieved. The details of HACK are shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.633, + 0.802 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We design several experiments to show the effectiveness of HACK for the semantic and hierarchical organization. First, we first construct a dataset with known hierarchical structure using the congealing algorithm [31]. Then, we apply HACK to datasets with unknown hierarchical structure to organize the samples based on the semantic and prototypical structure." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23203" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.122 + ], + "angle": 0, + "content": "Finally, we show that the prototypical structure can be used to reduce sample complexity and increase model robustness." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.125, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Datasets. We first construct a dataset called Congeaed MNIST. To verify the efficacy of HACK for unsupervised prototypicality discovery, we need a benchmark with known prototypical examples. However, currently there is no standard benchmark for this purpose. To construct the benchmark, we use the congealing algorithm from Miller et al. [31] to align the images in each class of MNIST [25]. The congealing algorithm is initially used for one-shot classification. During congealing, the images are brought into correspondence with each other jointly. The congealed images are more prototypical: they are better aligned with the average image. The synthetic data is generated by replacing 500 original images with the corresponding congealed images. In the Appendix, we show the results of changing the number of replaced original images. We expect HACK to discover the congealed images and place them in the center of the Poincaré ball. We also aim to discover the prototypical examples from each class of the standard MNIST dataset [25] and CIFAR10 [24]. CIFAR10 consists of 60000 from 10 object categories ranging from airplane to truck. CIFAR10 is more challenging than MNIST since it has larger intra-class variations. Moreover, to better visualize how HACK arranges the samples according to their prototypicality, we run HACK on 10k US Adult Faces [2] (hereafter referred to as USA10kF), which contains 10,168 natural face photographs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.471, + 0.548 + ], + "angle": 0, + "content": "Baselines. We consider several existing metrics proposed in Carlini et al. [5] for prototypicality discovery, the details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.551, + 0.472, + 0.61 + ], + "angle": 0, + "content": "- Holdout Retraining [5]: We consider the Holdout Retraining proposed in Carlini et al. [5]. The idea is that the distance of features of prototypical examples obtained from models trained on different datasets should be close." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.611, + 0.472, + 0.671 + ], + "angle": 0, + "content": "- Model Confidence [5]: Intuitively, the model should be confident in prototypical examples. Thus, it is natural to use the confidence of the model prediction as the criterion for prototypicality." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.671, + 0.472, + 0.732 + ], + "angle": 0, + "content": "- UHML [46]: UHML is an unsupervised hyperbolic learning method that aims to discover the natural hierarchies of data by taking advantage of hyperbolic metric learning and hierarchical clustering." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.551, + 0.472, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We implement HACK in PyTorch and the code will be made public. To generate uniform particles, we first randomly initialize the particles and then run the training for 1000 epochs with a 0.01 learning rate to minimize the repulsion loss and boundary loss. The curvature of the Poincaré ball is 1.0 and the \\(r\\) is 0.76 which is used to alleviate the numerical issues [16]. The hyperparameter \\(k\\) is 1.55 which is shown to generate uniform particles well. For the assignment, we use a LeNet [26] for MNIST, a ResNet20 [17] for CIFAR10, and a ResNet18 for USA10kF as the encoder. We apply HACK to each class separately." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.846, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Figure 6. Hyperbolic space can capture the prototypicality inherently. The error bar of each point is given by the variance of density within the corresponding portion, and the width of the shaded band indicates the number of features within the portion." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.3, + 0.682, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.437, + 0.608, + 0.45 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.3, + 0.853, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.436, + 0.775, + 0.45 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.463, + 0.882, + 0.545 + ], + "angle": 0, + "content": "Figure 7. Congeaed images are located in the center of the Poincaré ball. a) Red dots denote congealed images and cyan dots denote original images. b) Typical images are in the center and atypical images are close to the boundary. Images are also clustered together based on visual similarity. Congeaed images are shown in red boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.682 + ], + "angle": 0, + "content": "We attach a fully connected layer to project the feature into a two-dimensional Euclidean space. The image features are then projected onto hyperbolic space via an exponential map. We run the training for 200 epochs using a cosine learning rate scheduler [29] with an initial learning rate of 0.1. We optimize the assignment every other epoch. All the experiments are run on an NVIDIA Titan RTX GPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.696, + 0.892, + 0.712 + ], + "angle": 0, + "content": "5.1. Prototypicality in the Hyperbolic Feature Norm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "We explicitly show that the hyperbolic space can capture prototypicality by analyzing the relation between hyperbolic norms and the K-NN density estimation. Taking the learned hyperbolic features, we first divide the range of norms of hyperbolic features into numerous portions with equal length (50 portions for this plot). The mean K-NN density is calculated by averaging the density estimation of features within each portion. Figure 6 shows that the mean density drops as the norm increases, which shows that the prototypicality emerges automatically within the norms, the inherent characteristic of hyperbolic space. This validates that prototypicality is reflected in the hyperbolic feature norm." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23204" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.478, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.471, + 0.302 + ], + "angle": 0, + "content": "Figure 8. Our unsupervised learning methods conform to our visual perception. The images are organized from left to right, top to bottom to cover the 360 degrees at the same radius." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.319, + 0.462, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.438 + ], + "angle": 0, + "content": "Figure 9. HACK discovers typical and atypical images from the data. First row: atypical images with large hyperbolic norm. Second row: typical images with small hyperbolic norm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.281, + 0.483 + ], + "angle": 0, + "content": "5.2. Visual Prototypicality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Congeaed MNIST. We further apply HACK for visual feature learning on congealed MNIST. Figure 7 shows that HACK can discover the congealed images from all images. In Figure 7 a), the red particles denote the congealed images and cyan particles denote the original images. We can observe that the congealed images are assigned to the particles located in the center of the Poincaré ball. This verifies that HACK can indeed discover prototypical examples from the original dataset. In the Appendix, we show that the features of atypical examples gradually move to the boundary of the Poincaré ball during training. In Figure 7 b), we show the actual images that are embedded in the two-dimensional hyperbolic space. We can observe that the images in the center of Poincaré ball are more prototypical and images close to the boundary are more atypical. Also, the images are naturally organized by their semantic similarity. In summary, HACK can discover prototypicality and also organize the images based on their semantic and hierarchical structure. To the best of our knowledge, this is the first unsupervised learning method that can be used to discover prototypical examples in a data-driven fashion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "USA10kF. Figure 10 a) shows the assignment of 2000 images sampled from USA10kF. Compared to MNIST, the variation in faces is much more complex. A facial image is also subject to various factors (such as race, facial expres" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.707, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.095, + 0.898, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.273, + 0.894, + 0.357 + ], + "angle": 0, + "content": "Figure 10. HACK captures prototypicality of faces. a) The assignment of 2000 images sampled from USA10kF. b) The angles indicate the uniform division of the space. For each orientation, 8 equidistant sample points were selected and the nearest faces are shown at each point. The face located in the center is most symmetrical with clear smiling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.58 + ], + "angle": 0, + "content": "sion, environmental condition, etc.). Therefore, the results from USA10kF are less intuitive than those from MNIST. However, Figure 10 b) illustrates the evolutionary process of images originating from the center of hyperbolic space and progressing along different directions. In the space, we selected 12 directions at equal angular intervals and chose five equally spaced sampling points in each direction. The images closest to these sampling points are displayed at their respective locations. Although the detailed organization is unclear, the evolutionary process reveals a tendency of HACK to cluster different features together, such as darker skin tones appearing in the 0-90 degree range and lighter skin tones in the 180-270 degree range." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.894, + 0.716 + ], + "angle": 0, + "content": "MNIST and CIFAR10. Figure 8 shows the embedding of class 0 from MNIST and class \"airplane\" from CIFAR10 arranged to cover 360 degrees of at the same radius. The visual similarity of the images has a smooth transition as we move around angularly. Figure 9 shows the typical images and atypical images discovered by HACK. This further illustrates that HACK captures the semantic similarity of the images which enables prototypicality discovery. Please see the Appendix for more results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.82, + 0.743 + ], + "angle": 0, + "content": "5.3. Prototypicality for Instance Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Figure 12 shows the comparison of the baselines with HACK. With HACK, typical images are characterized by the smallest hyperbolic norms, whereas atypical images are associated with the largest hyperbolic norms. We can observe that both HACK and Model Confidence (MC) can discover typical and atypical images. Compared with MC, HACK defines prototypicality as the distance of the sample to other samples which is more aligned with human intuition. Moreover, in addition to prototypicality, HACK can also be used to organize examples by semantic similarities. Holdout Retraining" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23205" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.29, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.205, + 0.396, + 0.22 + ], + "angle": 0, + "content": "a) Reducing sample complexity" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.091, + 0.484, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.092, + 0.68, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.091, + 0.874, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.205, + 0.782, + 0.22 + ], + "angle": 0, + "content": "b) Increasing model robustness" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Figure 11. HACK can be used to construct sample prototypical hierarchy which is useful for several downstream tasks. a) Training with atypical examples achieves higher accuracy than training with typical examples. b) The adversarial accuracy greatly improves after removing the X% of most atypical examples." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.298, + 0.47, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.448 + ], + "angle": 0, + "content": "Figure 12. HACK can capture both prototypicality and diversity in the dataset. Each row shows the typical and atypical images discovered by HACK/HR [5]/MC [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.506 + ], + "angle": 0, + "content": "(HR) is not effective for prototypicality discovery due to the randomness of model training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.342, + 0.533 + ], + "angle": 0, + "content": "5.4. Application of Prototypicality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Reducing Sample Complexity. The proposed HACK can discover prototypical images as well as atypical images. We show that with atypical images we can reduce the sample complexity for training the model. Prototypical images are representative of the dataset but lack variations. Atypical examples contain more variations and it is intuitive that models trained on atypical examples should generalize better to the test samples. To verify this hypothesis, we select a subset of samples based on the norm of the features which indicates prototypicality of the examples. In particular, typical samples correspond to the samples with smaller norms and atypical samples correspond to the samples with larger norms. The angular layout of the hyperbolic features naturally captures sample diversity, thus for selecting atypical examples, we consider introducing more diversity by sampling images with large norms along the angular direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Figure 11 a) shows that training with atypical images can achieve much higher accuracy than training with typical images. In particular, training with the most atypical \\(10\\%\\) of the images achieves \\(22.67\\%\\) higher accuracy than with the most typical \\(10\\%\\) of the images on CIFAR10. Similar results can be observed on MNIST. Thus, HACK provides an easy solution to reduce sample complexity. We also compared UHML [46], which is an unsupervised metric learning in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.376 + ], + "angle": 0, + "content": "hyperbolic space, with HACK on the MNIST dataset. By incorporating \\(10\\%\\) atypical samples based on feature norm, HACK outperformed UHML by \\(10.2\\%\\). Also by excluding the \\(1\\%\\) atypical examples, HACK achieved an additional \\(5.7\\%\\) improvement over UHML." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Increasing Model Robustness. Training models with atypical examples can lead to a vulnerable model to adversarial attacks [5, 28]. Intuitively, atypical examples lead to a less smooth decision boundary thus a small perturbation to examples is likely to change the prediction. With HACK, we can easily identify atypical samples to improve the robustness of the model. We use MNIST and CIFAR 10 as the benchmark and use FGSM [13] to attack the model with an \\(\\epsilon\\) of 0.07 on MNIST and \\(8/(255^{*}std)\\) on CIFAR10, where \\(std\\) is the standard deviation used for normalization. More details of the attack settings can be found in the appendix. We identify the atypical examples based on the norm of the features with HACK and remove the most atypical \\(X\\%\\) of the examples. Figure 11 b) shows that discarding atypically examples greatly improves the robustness of the model: the adversarial accuracy is improved by \\(8.7\\%\\) when excluding the most atypical \\(1\\%\\) of examples on MNIST and \\(7.3\\%\\) on CIFAR10. It is worth noting that the clean accuracy remains the same after removing a small number of atypical examples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.689, + 0.608, + 0.706 + ], + "angle": 0, + "content": "6. Summary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.895, + 0.852 + ], + "angle": 0, + "content": "We propose HACK, an unsupervised learning method that organizes images in hyperbolic space using sphere packing. By optimizing image assignments to uniformly distributed particles, HACK leverages the inherent properties of hyperbolic space, leading to the natural emergence of prototypical and semantic structures through feature learning. We validate HACK on synthetic data and standard datasets, demonstrating its ability to discover prototypical examples for reducing sample complexity and increasing model robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project was supported, in part, by NSF 2131111, NSF 2215542, and NSF 2313151 to S. Yu, and a grant from UT Dallas to Y. Guo." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23206" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.142 + ], + "angle": 0, + "content": "[1] James W Anderson. Hyperbolic geometry. Springer Science & Business Media, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.471, + 0.186 + ], + "angle": 0, + "content": "[2] Wilma A Bainbridge, Phillip Isola, and Aude Oliva. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General, 142(4):1323, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[4] Jacob Bien and Robert Tibshirani. Prototype selection for interpretable classification. The Annals of Applied Statistics, 5(4):2403-2424, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[5] Nicholas Carlini, Ulfar Erlingsson, and Nicolas Papernot. Prototypical examples in deep learning: Metrics, characteristics, and utility. 2018. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[6] Ines Chami, Albert Gu, Vaggos Chatziafratis, and Christopher Ré. From trees to continuous embeddings and back: Hyperbolic hierarchical clustering. Advances in Neural Information Processing Systems, 33:15065-15076, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[8] Lieven Decock and Igor Douven. What is graded membership? Nous, 48(4):653-82, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[9] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 836-837, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[10] Evelyn Fix and Joseph Lawson Hodges. Discriminatory analysis. nonparametric discrimination: Consistency properties. International Statistical Review/Revue Internationale de Statistique, 57(3):238-247, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[11] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. arXiv preprint arXiv:1805.09112, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[12] Alan Gibbons. Algorithmic graph theory. Cambridge university press, 1985. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[14] Yanming Guo, Yu Liu, Erwin M Bakker, Yuanhao Guo, and Michael S Lew. Cnn-rnn: a large-scale hierarchical image classification framework. Multimedia tools and applications, 77(8):10251-10271, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[15] Yunhui Guo, Haoran Guo, and Stella Yu. Co-sne: Dimensionality reduction and visualization for hyperbolic data. arXiv preprint arXiv:2111.15037, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[16] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729–9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[20] Joy Hsu, Jeffrey Gu, Gong-Her Wu, Wah Chiu, and Serena Yeung. Learning hyperbolic representations for unsupervised 3d segmentation. arXiv preprint arXiv:2012.01644, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[21] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[22] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[25] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[27] Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[28] Yongshuai Liu, Jiyu Chen, and Hao Chen. Less is more: Culling the training set to improve robustness of deep neural networks. In International Conference on Decision and Game Theory for Security, pages 102-114. Springer, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[30] Emile Mathieu, Charline Le Lan, Chris J Maddison, Ryota Tomioka, and Yee Whye Teh. Continuous hierarchical representations with poincar\\`e variational auto-encoders. arXiv preprint arXiv:1901.06033, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[31] Erik G Miller, Nicholas E Matsakis, and Paul A Viola. Learning from one example through shared densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 464-471. IEEE, 2000. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[32] James Munkres. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 5(1):32-38, 1957. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A wrapped normal distribution on" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "hyperbolic space for gradient-based learning. In International Conference on Machine Learning, pages 4693-4702. PMLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[34] Maximilian Nickel and Douwe Kiela. Poincar\\`e embeddings for learning hierarchical representations. arXiv preprint arXiv:1705.08039, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[35] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[36] Eleanor Rosch. Cognitive representations of semantic categories. Journal of experimental psychology: General, 104(3): 192, 1975. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[37] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International Symposium on Graph Drawing, pages 355-366. Springer, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.471, + 0.405 + ], + "angle": 0, + "content": "[39] Pierre Stock and Moustapha Cisse. Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In Proceedings of the European Conference on Computer Vision (ECCV), pages 498-512, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[40] Richard Szeliski et al. Image alignment and stitching: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2(1):1-104, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[42] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[43] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[44] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2603-2612, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[45] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[46] Jiei Xi Yan, Lei Luo, Cheng Deng, and Heng Huang. Unsupervised hyperbolic metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12465-12474, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[47] Jianping Zhang. Selecting typical instances in instance-based learning. In Machine learning proceedings 1992, pages 470-479. Elsevier, 1992. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23208" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_origin.pdf b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4928aea212826e0e1c85b744c77dc7bcb8ec7623 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/b2742b78-cc7b-4e1c-b277-447debc213bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6310ba22c879c047c9640b5182cfd8b8eae44867643ed6a10ee4a2ed2e82a819 +size 9689660 diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/full.md b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df0d1bb37badd73666902ad763a15997c51a9ef3 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/full.md @@ -0,0 +1,290 @@ +# Unsupervised Feature Learning with Emergent Data-Driven Prototypicality + +Yunhui Guo1 Youren Zhang2 Yubei Chen3 Stella X. Yu2,4 +1The University of Texas at Dallas 2University of Michigan 3UC Davis 4UC Berkeley + +# Abstract + +Given a set of images, our goal is to map each image to a point in a feature space such that, not only point proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset. + +Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflects image similarity, yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space. + +We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKKing. HACK first generates uniformly packed particles in the Poincaré ball of hyperbolic space and then assigns each image uniquely to a particle. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing - a warping process that aligns all the images and makes them appear more common and similar to each other, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones. + +# 1. Introduction + +Not all instances are created equal. For example, the MNIST dataset of handwritten digits contain almost 6,000 samples of 2's; some are close to textbook versions that we are taught to follow, whereas others have idiosyncratic cursive styles, varying in proportions and stroke weights (Fig. 1). Given such a set of natural data, we are interested in dataset summarization and organization such that we can automatically discover which instances are more representative and which ones are anomalies. In other words, we aim to computation- + +![](images/a5593ab71496867a84926998f37b4ca5d63a2740aa848718bd400c1b52cae2c4.jpg) +Figure 1. Given a dataset (left), we aim to learn an image feature that encodes not only visual similarity between instances but also data-driven prototypicality (right). Additionally, the angular arrangement of the features can naturally serve as a measure of diversity. Our feature encoder (in 2D hyperbolic space) is learned without any labels. We can then learn a decoder to map each point in the feature space back to an image. The right plot visualizes images located at the origin and those moving away in different directions, automatically revealing that 2's with loops are most common and the whole dataset can be grasped as the cursive style systematically varies. + +ally rationalize graded membership [8, 36] of a category in a purely data-driven manner, putting each instance on a feature map that reflects not only their prototypicality in a particular dataset but also their visual similarity with each other. + +Such unsupervised feature learning is useful for not only data organization, but those discovered representative instances or prototypes can be used for interpretable machine learning [4], curriculum learning [3], and learning better decision boundaries [5]. Prototypes also allow us to classify with as few as or even one example [31]. + +If the image feature is given, it is relatively easy to find prototypes: We just need to identify density peaks of the feature distribution of the image set. Otherwise, discovering prototypical instances without supervision is difficult: There is no universal definition or simple metric to assess the prototypicality of the examples. + +One way to address this problem is to examine the gradient magnitude [5]. However, this approach is shown to have a high variance which results from different training setups [5]. Some methods address this problem using adversarial + +![](images/276794e3216f93c4a263cd67fa8d95a3c6d5fcd181c22e76a70a9e90799c3caf.jpg) +Figure 2. Existing self-supervised feature learning methods focus on visual similarity only in Euclidean space, whereas our unsupervised feature learning method embeds images in hyperbolic space which automatically encodes prototypicality with respect to the origin: Images are organized hierarchically, with typical images at the center and atypical ones near the boundary of the Poincaré ball. + +robustness [5, 39]: Prototypical examples should be more adversarially robust. However, selecting these examples is dependent on the adversarial method and the metric used in the adversarial attack. Several methods exist but they are either based on heuristics or lack a proper justification [5]. + +Naturally, given a feature space, prototypical examples can be identified as density peaks. However, prototypicality varies with the feature. We propose an unsupervised feature learning method in hyperbolic space, which, unlike the shift-invariant Euclidean space, naturally embeds a continuous version of hierarchical trees rooted at the origin of the space. + +Hyperbolic space is non-Euclidean space with constant non-negative curvature [1]. Different from Euclidean space, hyperbolic space can represent hierarchical relations with low distortion. Poincaré ball model is one of the most commonly used models for hyperbolic space [35]. One notable property of Poincaré ball model is that the distance to the origin grows exponentially as we move towards the boundary. Thus, points located in the center of the ball are close to all the other points while the points located close to the boundary are infinitely far away from other points. With unsupervised learning in hyperbolic space, we can learn features that capture both visual similarity and hierarchical proximity among instances. + +Our key insight is that a typical image is similar (closest) to more nearby instances than atypical ones, and such an image would be at the center of the neighbourhood in Euclidean space but the parent/root node in a tree in hyperbolic space, closer to the origin. We develop a new learning procedure that first places all the target locations evenly in the Poincaré ball to reflect the desire of mapping uniformity [43], and then we optimize which images should be mapped to which + +target locations in a batch-wise Hungarian matching manner, where those similar to most instances naturally moving closer to the origin. + +Our work makes the following contributions. 1) We propose the first unsupervised feature learning method to learn features which capture visual similarity with distance between features and prototypicality with the distance to the origin. 2) We develop a new learning paradigm that sits between supervised learning (with known targets) and unsupervised metric learning (with unknown targets and constrained metric distances). We want to map images to known targets uniformly packed and maximally distant in hyperbolic space, but we learn to optimize the specific image to target assignment. 3) We validate our joint feature learning and data prototypicality discovery on congealing [31], where the consensus is that images after congealing are perceived to be more typical images. 4) We demonstrate two practical usages of data-driven prototypicality: Prototypical and atypical examples are shown to reduce sample complexity for learning and increase the robustness of the model respectively. + +# 2. Related Work + +Prototypicality. The study of prototypical examples in machine learning has a long history. In Zhang [47], the authors select typical instances based on the fact that typical instances should be representative of the cluster. In Kim et al. [22], prototypical examples are defined as the examples that have maximum mean discrepancy within the data. Li et al. [27] propose to discover prototypical examples by architectural modifications: project the dataset onto a low-dimensional manifold and use a prototype layer to minimize the distance between inputs and the prototypes on the manifold. The robustness to adversarial attacks is also used as a criterion for prototypicality [39]. In Carlini et al. [5], the authors propose multiple metrics for prototypicality discovery. For example, the features of prototypical examples should be consistent across different training setups. However, these metrics usually depend heavily on the training setups and hyperparameters. The idea of prototypicality is also extensively studied in meta-learning for one-shot or few-shot classification [38]. No existing works address the prototypicality discovery problem in a data-driven fashion. Our proposed HACK naturally exploits hyperbolic space to organize the images based on prototypicality. + +Unsupervised Learning in Hyperbolic Space. Learning features in hyperbolic space have shown to be useful for many machine learning problems [11, 34]. One useful property is that hierarchical relations can be embedded in hyperbolic space with low distortion [34]. Wrapped normal distribution, which is a generalized version of the normal distribution for modeling the distribution of points in hyperbolic space [33], is used as the latent space for constructing hyperbolic variational autoencoders (VAEs) + +![](images/f15b21f57154c3757da9e0d08a4609b178121bcecf28e67c56360693977f2471.jpg) +Figure 3. Congeaed images are more typical than the original images. First row: sampled original images. Second row: the corresponding congealed images. + +[23]. Poincaré VAEs is constructed in Mathieu et al. [30] with a similar idea to Nagano et al. [33] by replacing the standard normal distribution with hyperbolic normal distribution. Unsupervised 3D segmentation [20] and instance segmentation [44] are conducted in hyperbolic space via hierarchical hyperbolic triplet loss. CO-SNE [15] is recently proposed to visualize high-dimensional hyperbolic features in a two-dimensional hyperbolic space. Although hyperbolic distance facilitates the learning of hierarchical structure, how to leverage hyperbolic space for unsupervised prototypicality discovery is not explored in the current literature. + +# 3. Sample Hierarchy + +Sample Hierarchy VS. Class Hierarchy. While most of the existing works in hierarchical image classification focus on using label hierarchy [9, 14], there also exists a natural hierarchy among different samples. In Khrulkov et al. [21], the authors conducted an experiment to measure the $\delta$ -hyperbolicity of the various image datasets and showed that common image datasets such as CIFAR10 and CUB exhibit natural hierarchical structure among the samples. Amongst a collection of images representing digit 1, suppose $\mathbf{x}$ is used for representing an image with a digit '1' that is upright, $\mathbf{x}'$ is used for representing an image with a digit 1 that leaning left and $\mathbf{x}''$ is used for representing an image with a digit '1' that leaning right. Given a metric $d(\cdot, \cdot)$ , if we assume that $d(\mathbf{x}''', \mathbf{x}') \approx d(\mathbf{x}''', \mathbf{x}) + d(\mathbf{x}', \mathbf{x})$ , in this context, we can naturally view the sample $\mathbf{x}$ as the root, and consider the other samples as its children in an underlying tree. + +Compared with class hierarchy which can be extracted based on the pre-defined label relations, sample hierarchy is much harder to construct due to the lack of ground truth. Once a sample hierarchy is established, there are currently no existing methods available for verifying the accuracy of the hierarchy. Additionally, just like with class hierarchies, there may be ambiguities when constructing a sample hierarchy since multiple samples could potentially serve as the root. + +Building Sample Hierarchy from Density Peaks. Given existing features $\{f(v_{i})\}$ obtained by applying a feature extractor for each instance $v_{i}$ , prototypical examples can be found by examining the density peaks via techniques from + +density estimation. For example, the K-nearest neighbor density (K-NN) estimation [10] is defined as $p_{knn}(v_i, k) = \frac{k}{n} \frac{1}{A_d \cdot D^d(v_i, v_{k(i)})}$ , where $d$ is the feature dimension, $A_d = \pi^{d/2} / \Gamma(d/2 + 1)$ , $\Gamma(x)$ is the Gamma function and $k(i)$ is the $k$ th nearest neighbor of example $v_i$ . The nearest neighbors can be found by computing the distance between the features. Therefore, the process of constructing sample hierarchy through density estimation can be conceptualized as a two-step procedure involving: 1) feature learning and 2) detecting density peaks. + +In the density estimation approach outlined above, the level of prototypicality depends on the learned features. Varying training setups can induce diverse feature spaces, resulting in differing conclusions on prototypicality. Nevertheless, prototypicality is an inherent attribute of the dataset and should remain consistent across various features. The aim of this paper is to extract features that intrinsically showcase the hierarchical organization of the samples. Specifically, by examining the feature alone within the feature space, we should be able to identify the example's prototypicality. + +Construct a Sample Hierarchy from Congealing. To determine whether the feature truly captures prototypicality, it is necessary to identify which sample is the prototype. We ground our concept of prototypicality based on congealing [31]. In particular, we define prototypical examples in the pixel space by examining the distance of the images to the average image in the corresponding class. Our idea is based on a traditional computer vision technique called image alignment [40] that aims to find correspondences across images. During congealing [31], a set of images are transformed to be jointly aligned by minimizing the joint pixel-wise entropies. The congealed images are more prototypical: they are better aligned with the average image. Thus, we have a simple way to transform an atypical example into a typical example (see Figure 3). This is useful since given an unlabeled image dataset the typicality of the examples is unknown, congealing examples can be naturally served as examples with known typicality and be used as a validation for the effectiveness of our method. + +# 4. Unsupervised Hyperbolic Feature Learning + +# 4.1. Poincaré Ball Model for Hyperbolic Space + +Euclidean space has a curvature of zero and a hyperbolic space is a Riemannian manifold with constant negative curvature.. There are several isometrically equivalent models for visualizing hyperbolic space with Euclidean representation. The Poincaré ball model is the commonly used one in hyperbolic representation learning [35]. The $n$ -dimensional Poincaré ball model is defined as $(\mathbb{B}^n, \mathfrak{g}_{\mathbf{x}})$ , where $\mathbb{B}^n = \{\mathbf{x} \in \mathbb{R}^n : \| \mathbf{x} \| < 1\}$ and $\mathfrak{g}_{\mathbf{x}} = (\gamma_{\mathbf{x}})^2 I_n$ is the Riemannian metric tensor. $\gamma_{\mathbf{x}} = \frac{2}{1 - \| \mathbf{x} \|^2}$ is the conformal factor and $I_n$ is the Euclidean metric tensor. + +![](images/ea7971f0a81d7aebf36b97a9e34f3d5b5af9b96b25a3652bb9f957912665893c.jpg) +a) Supervised classification with fixed known targets + +![](images/a944af3952c3dc778a0d0d2b62ba17bfab10fd3a134687c0c48faaeccf4799c2.jpg) +b) Our unsupervised feature learning with fixed but unknown targets +Figure 4. The proposed HACK has a predefined geometrical arrangement and allows the images to be freely assigned to any particle. a) Standard supervised learning has predefined targets. The image is only allowed to be assigned to the corresponding target. b) HACK packs particles uniformly in hyperbolic space to create initial seeds for the organization. The images are assigned to the particles based on their prototypicality and semantic similarities. c) Standard unsupervised learning has no predefined targets and images are clustered based on their semantic similarities. + +![](images/3a0056c4cda138179dc44ec507a89dfa6785467f6ab0b84624bb34a96f4607a1.jpg) +c) Metric feature learning with unknown targets + +Hyperbolic Distance. Given two points $\pmb{u} \in \mathbb{B}^n$ and $\pmb{v} \in \mathbb{B}^n$ , the hyperbolic distance is defined as, + +$$ +d _ {\mathbb {B} ^ {n}} (\boldsymbol {u}, \boldsymbol {v}) = \operatorname {a r c o s h} \left(1 + 2 \frac {\| \boldsymbol {u} - \boldsymbol {v} \| ^ {2}}{(1 - \| \boldsymbol {u} \| ^ {2}) (1 - \| \boldsymbol {v} \| ^ {2})}\right) \tag {1} +$$ + +where $\operatorname{arcosh}$ is the inverse hyperbolic cosine function and $\|\cdot\|$ is the usual Euclidean norm. + +Hyperbolic distance has the unique property that it grows exponentially as we move towards the boundary of the Poincaré ball. In particular, the points on the circle represent points in infinity. Hyperbolic space is naturally suitable for embedding hierarchical structure [35, 37] and can be regarded as a continuous representation of trees [6]. The hyperbolic distance between samples implicitly reflects their hierarchical relation. Thus, by embedding images in hyperbolic space we can naturally organize images based on their semantic similarity and prototypicality. + +# 4.2. Build Instance Hierarchy in Hyperbolic Space + +Hyperbolic space is naturally suitable for embedding tree structure. However, in order to leverage hyperbolic space to build a sample hierarchy in an unsupervised manner, a suitable objective function is still missing. There are two challenges in designing the objective function. First, the underlying tree structure of the samples is unknown. Second, how to perform feature learning such that hierarchy can naturally emerge is unclear. In this paper, we propose Hyperbolic space with sphere pACKing, also called HACK, to address the two challenges. + +To address the first challenge, instead of creating a predefined tree structure that might not faithfully represent the genuine hierarchical organization, we leverage sphere packing in hyperbolic space for building a skeleton for placing the samples. We specify where the particles should be located ahead of training with uniform packing, which by design are + +maximally evenly spread out in hyperbolic space. The uniformly distributed particles guide feature learning to achieve maximum instance discrimination [45] while enabling us to extract a tree structure from the samples. + +To address the second challenge, HACK figures out which instance should be mapped to which target through bipartite graph matching as a global optimization procedure. During training, HACK minimizes the total hyperbolic distances between the mapped image point (in the feature space) and the target, those that are more typical naturally emerge closer to the origin of Poincaré ball. HACK differs from the existing learning methods in several aspects (Figure 4). Different from supervised learning, HACK allows the image to be assigned to any target (particle). This enables the exploration of the natural organization of the data. Different from unsupervised learning method, HACK specifies a predefined geometrical organization which encourages the corresponding structure to be emerged from the dataset. + +# 4.3. Sphere Packing in Hyperbolic Space + +Given $n$ particles, our goal is to pack the particles into a two-dimensional hyperbolic space as densely as possible. We derive a simple repulsion loss function to encourage the particles to be equally distant from each other. The loss is derived via the following steps. First, we need to determine the radius of the Poincaré ball used for packing. We use a curvature of 1.0 so the radius of the Poincaré ball is 1.0. The whole Poincaré ball cannot be used for packing since the volume is infinite. We use $r < 1$ to denote the actual radius used for packing. Thus, our goal is to pack $n$ particles in a compact subspace of Poincaré ball. Then, the Euclidean radius $r$ is further converted into hyperbolic radius $r_{\mathbb{B}}$ . Let $s = \frac{1}{\sqrt{c}}$ , where $c$ is the curvature. The relation between $r$ and $r_{\mathbb{B}}$ is $r_{\mathbb{B}} = s \log \frac{s + r}{s - r}$ . Next, the total hyperbolic area $A_{\mathbb{B}}$ of a Poincaré ball of radius $r_{\mathbb{B}}$ can + +![](images/55191b4684c72402edc1035939655b1c1b6991e08d2593be404a320cc4060c43.jpg) +Figure 5. HACK conducts unsupervised learning in hyperbolic space with sphere packing. The images are mapped to particles by minimizing the total hyperbolic distance. HACK learns features that can capture both visual similarities and prototypicality. + +be computed as $A_{\mathbb{B}} = 4\pi s^2 \sinh^2 \left( \frac{r_{\mathbb{B}}}{2s} \right)$ , where $\sinh$ is the hyperbolic sine function. Finally, the area per point $A_{n}$ can be easily computed as $\frac{A_{\mathbb{B}}}{n}$ , where $n$ is the total number of particles. Given $A_{n}$ , the radius per point can be computed as $r_{n} = 2s \sinh^{-1} \left( \sqrt{\frac{A_{n}}{4\pi s^{2}}} \right)$ . We use the following loss to generate uniform packing in hyperbolic space. Given two particles $i$ and $j$ , the repulsion loss $V$ is defined as, + +$$ +V (i, j) = \left\{\frac {1}{\left[ 2 r _ {n} - \max \left(0 , 2 r _ {n} - d _ {\mathbb {B}} (i , j)\right) \right] ^ {k}} - \frac {1}{\left(2 r _ {n}\right) ^ {k}} \right\} \cdot C (k) \tag {2} +$$ + +where $C(k) = \frac{(2r_n)^{k + 1}}{k}$ and $k$ is a hyperparameter. Intuitively, if the particle $i$ and the particle $j$ are within $2r_n$ , the repulsion loss is positive. Minimizing the repulsion loss would push the particles $i$ and $j$ away. If the repulsion is zero, this indicates all the particles are equally distant. Also the repulsion loss grows significantly when two particles become close. We also adopt the following boundary loss to prevent the particles from escaping the ball, + +$$ +B (i; r) = \max \left(0, \operatorname {n o r m} _ {i} - r + \text {m a r g i n}\right) \tag {3} +$$ + +where $\mathrm{norm}_i$ is the $\ell_2$ norm of the representation of the particle $i$ . Figure 4 b) shows an example of the generated particles that are uniformly packed in hyperbolic space. + +# 4.4. Hyperbolic Instance Assignment + +HACK learns the features by optimizing the assignments of the images to particles (Figure 5). The assignment should be one-to-one, i.e., each image is assigned to one particle and each particle is allowed to be associated with one image. We cast the instance assignment problem as a bipartite matching problem [12] and solve it with Hungarian algorithm [32]. + +Initially, we randomly assign the particles to the images, thus there is a random one-to-one correspondence between the images to the particles (not optimized). Given a batch of samples $\{(\mathbf{x}_1,\mathbf{s}_1),(\mathbf{x}_2,\mathbf{s}_2),\dots,(\mathbf{x}_B,\mathbf{s}_B)\}$ , where $\mathbf{x}_i$ is an image and $\mathbf{s}_i$ is the corresponding particle, and an encoder $f_{\theta}$ , we generate the hyperbolic feature for each image $\mathbf{x}_i$ as $f_{\theta}(\mathbf{x}_i) \in \mathbb{B}^2$ , where $\mathbb{B}^2$ is a two-dimensional Poincaré ball. For a given hyperbolic feature $f_{\theta}(\mathbf{x})$ , with fixed particle + +Algorithm 1 HACK: Unsupervised Learning in Hyperbolic Space. ised Learning in Hyperbolic Space. + +Require: # of images: $n \geq 0$ . Radius for packing: $r < 1$ . An encoder with parameters $\theta$ : $f_{\theta}$ + +1: Generate uniformly distributed particles in hyperbolic space by minimizing the repulsion loss in Equation 2 +2: Given $\{(\mathbf{x}_1, s_1), (\mathbf{x}_2, s_2), \dots, (\mathbf{x}_b, s_b)\}$ , optimize $f_{\theta}$ by minimizing the total hyperbolic distance via Hungarian algorithm. + +locations, the distance between the hyperbolic feature and the particles signifies the hierarchical level of the associated sample. Thus, to determine the hierarchical levels for all samples within the hierarchy, we must establish a one-to-one mapping between all the samples and the particles. This can be cast as the following bipartite matching problem in hyperbolic space, + +$$ +\ell (\theta , \pi) = \sum_ {i = 1} ^ {B} d _ {\mathbb {B} ^ {n}} \left(f _ {\theta} \left(\mathbf {x} _ {i}\right), \mathbf {s} _ {\pi \left(f _ {\theta} \left(\mathbf {x} _ {i}\right)\right)}\right) \tag {4} +$$ + +where $\pi : f_{\theta}(\mathbf{x}) \to \mathbb{N}$ is a projection function which projects hyperbolic features to a particle index. Minimizing $\ell(\theta, \pi)$ with respect to $\pi$ is a combinatorial optimization problem, which can hardly be optimized with $\theta$ using gradient-based algorithms. Thus, we adopt a joint optimization strategy which optimizes $\theta$ and $\pi$ alternatively. In each batch, we first leverage the Hungarian algorithm [32] to find the optimal matching $\pi^{*}$ based on the current hyperbolic features. Then we minimize Eq. 4 with respect to $\theta$ based on the current assignment $\pi^{*}$ . This process is repeated for a certain number of epochs until convergence is achieved. On the other hand, the feature encoder can serve as an image prior for assigning similar images to nearby particles [41]. + +The Hungarian algorithm [32] has a complexity of $\mathcal{O}(x^3)$ , where $x$ is the number of items. As we perform the particle assignment in the batch level, the time and memory complexity is tolerable. Also, the one-to-one correspondence between the images and particles is always maintained during training. After training, based on the assigned particle, the level of the sample in the hierarchy can be easily retrieved. The details of HACK are shown in Algorithm 1. + +# 5. Experiments + +We design several experiments to show the effectiveness of HACK for the semantic and hierarchical organization. First, we first construct a dataset with known hierarchical structure using the congealing algorithm [31]. Then, we apply HACK to datasets with unknown hierarchical structure to organize the samples based on the semantic and prototypical structure. + +Finally, we show that the prototypical structure can be used to reduce sample complexity and increase model robustness. + +Datasets. We first construct a dataset called Congeaed MNIST. To verify the efficacy of HACK for unsupervised prototypicality discovery, we need a benchmark with known prototypical examples. However, currently there is no standard benchmark for this purpose. To construct the benchmark, we use the congealing algorithm from Miller et al. [31] to align the images in each class of MNIST [25]. The congealing algorithm is initially used for one-shot classification. During congealing, the images are brought into correspondence with each other jointly. The congealed images are more prototypical: they are better aligned with the average image. The synthetic data is generated by replacing 500 original images with the corresponding congealed images. In the Appendix, we show the results of changing the number of replaced original images. We expect HACK to discover the congealed images and place them in the center of the Poincaré ball. We also aim to discover the prototypical examples from each class of the standard MNIST dataset [25] and CIFAR10 [24]. CIFAR10 consists of 60000 from 10 object categories ranging from airplane to truck. CIFAR10 is more challenging than MNIST since it has larger intra-class variations. Moreover, to better visualize how HACK arranges the samples according to their prototypicality, we run HACK on 10k US Adult Faces [2] (hereafter referred to as USA10kF), which contains 10,168 natural face photographs. + +Baselines. We consider several existing metrics proposed in Carlini et al. [5] for prototypicality discovery, the details can be found in the Appendix. + +- Holdout Retraining [5]: We consider the Holdout Retraining proposed in Carlini et al. [5]. The idea is that the distance of features of prototypical examples obtained from models trained on different datasets should be close. +- Model Confidence [5]: Intuitively, the model should be confident in prototypical examples. Thus, it is natural to use the confidence of the model prediction as the criterion for prototypicality. +- UHML [46]: UHML is an unsupervised hyperbolic learning method that aims to discover the natural hierarchies of data by taking advantage of hyperbolic metric learning and hierarchical clustering. + +Implementation Details. We implement HACK in PyTorch and the code will be made public. To generate uniform particles, we first randomly initialize the particles and then run the training for 1000 epochs with a 0.01 learning rate to minimize the repulsion loss and boundary loss. The curvature of the Poincaré ball is 1.0 and the $r$ is 0.76 which is used to alleviate the numerical issues [16]. The hyperparameter $k$ is 1.55 which is shown to generate uniform particles well. For the assignment, we use a LeNet [26] for MNIST, a ResNet20 [17] for CIFAR10, and a ResNet18 for USA10kF as the encoder. We apply HACK to each class separately. + +![](images/54cedbae1e2a22363ac63e66e749e4faffb04e42e2a4d99b18a936698fe67fd5.jpg) +Figure 6. Hyperbolic space can capture the prototypicality inherently. The error bar of each point is given by the variance of density within the corresponding portion, and the width of the shaded band indicates the number of features within the portion. + +![](images/f5b57a190c1e8e84c5405ed046eb5cf8fa8a9351c9f5949bceb2b4f513a62d6a.jpg) +a) +Figure 7. Congeaed images are located in the center of the Poincaré ball. a) Red dots denote congealed images and cyan dots denote original images. b) Typical images are in the center and atypical images are close to the boundary. Images are also clustered together based on visual similarity. Congeaed images are shown in red boxes. + +![](images/baefb88089f2b9b7a214982cb21013a009ab59dca4fbbfa0c46761a2468ebb6e.jpg) +b) + +We attach a fully connected layer to project the feature into a two-dimensional Euclidean space. The image features are then projected onto hyperbolic space via an exponential map. We run the training for 200 epochs using a cosine learning rate scheduler [29] with an initial learning rate of 0.1. We optimize the assignment every other epoch. All the experiments are run on an NVIDIA Titan RTX GPU. + +# 5.1. Prototypicality in the Hyperbolic Feature Norm + +We explicitly show that the hyperbolic space can capture prototypicality by analyzing the relation between hyperbolic norms and the K-NN density estimation. Taking the learned hyperbolic features, we first divide the range of norms of hyperbolic features into numerous portions with equal length (50 portions for this plot). The mean K-NN density is calculated by averaging the density estimation of features within each portion. Figure 6 shows that the mean density drops as the norm increases, which shows that the prototypicality emerges automatically within the norms, the inherent characteristic of hyperbolic space. This validates that prototypicality is reflected in the hyperbolic feature norm. + +![](images/5290b646bd25251ae3bba89047cb5c40f749d28c5f59b32db759504005a724d9.jpg) +Figure 8. Our unsupervised learning methods conform to our visual perception. The images are organized from left to right, top to bottom to cover the 360 degrees at the same radius. + +![](images/ced45d0b8401eb597f4e9ac3ca5a7636710f36f4b0a359fe299da78cfd610503.jpg) +Figure 9. HACK discovers typical and atypical images from the data. First row: atypical images with large hyperbolic norm. Second row: typical images with small hyperbolic norm. + +# 5.2. Visual Prototypicality + +Congeaed MNIST. We further apply HACK for visual feature learning on congealed MNIST. Figure 7 shows that HACK can discover the congealed images from all images. In Figure 7 a), the red particles denote the congealed images and cyan particles denote the original images. We can observe that the congealed images are assigned to the particles located in the center of the Poincaré ball. This verifies that HACK can indeed discover prototypical examples from the original dataset. In the Appendix, we show that the features of atypical examples gradually move to the boundary of the Poincaré ball during training. In Figure 7 b), we show the actual images that are embedded in the two-dimensional hyperbolic space. We can observe that the images in the center of Poincaré ball are more prototypical and images close to the boundary are more atypical. Also, the images are naturally organized by their semantic similarity. In summary, HACK can discover prototypicality and also organize the images based on their semantic and hierarchical structure. To the best of our knowledge, this is the first unsupervised learning method that can be used to discover prototypical examples in a data-driven fashion. + +USA10kF. Figure 10 a) shows the assignment of 2000 images sampled from USA10kF. Compared to MNIST, the variation in faces is much more complex. A facial image is also subject to various factors (such as race, facial expres + +![](images/7fb203d23b82dadb6c754b9fc394d4ebcc1286b17e22c7abf5b827b58812a213.jpg) +Figure 10. HACK captures prototypicality of faces. a) The assignment of 2000 images sampled from USA10kF. b) The angles indicate the uniform division of the space. For each orientation, 8 equidistant sample points were selected and the nearest faces are shown at each point. The face located in the center is most symmetrical with clear smiling + +![](images/182ae3765979e2e5ba3fd745f897d65d6c093e11dfc4624c5b218a10383bff9e.jpg) + +sion, environmental condition, etc.). Therefore, the results from USA10kF are less intuitive than those from MNIST. However, Figure 10 b) illustrates the evolutionary process of images originating from the center of hyperbolic space and progressing along different directions. In the space, we selected 12 directions at equal angular intervals and chose five equally spaced sampling points in each direction. The images closest to these sampling points are displayed at their respective locations. Although the detailed organization is unclear, the evolutionary process reveals a tendency of HACK to cluster different features together, such as darker skin tones appearing in the 0-90 degree range and lighter skin tones in the 180-270 degree range. + +MNIST and CIFAR10. Figure 8 shows the embedding of class 0 from MNIST and class "airplane" from CIFAR10 arranged to cover 360 degrees of at the same radius. The visual similarity of the images has a smooth transition as we move around angularly. Figure 9 shows the typical images and atypical images discovered by HACK. This further illustrates that HACK captures the semantic similarity of the images which enables prototypicality discovery. Please see the Appendix for more results. + +# 5.3. Prototypicality for Instance Selection + +Figure 12 shows the comparison of the baselines with HACK. With HACK, typical images are characterized by the smallest hyperbolic norms, whereas atypical images are associated with the largest hyperbolic norms. We can observe that both HACK and Model Confidence (MC) can discover typical and atypical images. Compared with MC, HACK defines prototypicality as the distance of the sample to other samples which is more aligned with human intuition. Moreover, in addition to prototypicality, HACK can also be used to organize examples by semantic similarities. Holdout Retraining + +![](images/4c05788586425e73d576dd78e4898a73937d90946ee5911e53b131820a60dbdb.jpg) +a) Reducing sample complexity + +![](images/d85ab442833a112a00ad3d812c1c3c141d6095c964e89a8d21563522f3569089.jpg) + +![](images/386b08f9ce6773421c28f95dfd7aa2bc350035b9715c24859845bbac42578507.jpg) +b) Increasing model robustness + +![](images/88350d94a95e075163cd1a55a0e1f34571fd89ebd9016c3ee610ac6f2c79c569.jpg) + +![](images/b83c9dba7df38a388ffe1157b9adef0de0c2497ea12d41b52cb29ce2debf743e.jpg) +Figure 11. HACK can be used to construct sample prototypical hierarchy which is useful for several downstream tasks. a) Training with atypical examples achieves higher accuracy than training with typical examples. b) The adversarial accuracy greatly improves after removing the X% of most atypical examples. +Figure 12. HACK can capture both prototypicality and diversity in the dataset. Each row shows the typical and atypical images discovered by HACK/HR [5]/MC [5]. + +(HR) is not effective for prototypicality discovery due to the randomness of model training. + +# 5.4. Application of Prototypicality + +Reducing Sample Complexity. The proposed HACK can discover prototypical images as well as atypical images. We show that with atypical images we can reduce the sample complexity for training the model. Prototypical images are representative of the dataset but lack variations. Atypical examples contain more variations and it is intuitive that models trained on atypical examples should generalize better to the test samples. To verify this hypothesis, we select a subset of samples based on the norm of the features which indicates prototypicality of the examples. In particular, typical samples correspond to the samples with smaller norms and atypical samples correspond to the samples with larger norms. The angular layout of the hyperbolic features naturally captures sample diversity, thus for selecting atypical examples, we consider introducing more diversity by sampling images with large norms along the angular direction. + +Figure 11 a) shows that training with atypical images can achieve much higher accuracy than training with typical images. In particular, training with the most atypical $10\%$ of the images achieves $22.67\%$ higher accuracy than with the most typical $10\%$ of the images on CIFAR10. Similar results can be observed on MNIST. Thus, HACK provides an easy solution to reduce sample complexity. We also compared UHML [46], which is an unsupervised metric learning in + +hyperbolic space, with HACK on the MNIST dataset. By incorporating $10\%$ atypical samples based on feature norm, HACK outperformed UHML by $10.2\%$ . Also by excluding the $1\%$ atypical examples, HACK achieved an additional $5.7\%$ improvement over UHML. + +Increasing Model Robustness. Training models with atypical examples can lead to a vulnerable model to adversarial attacks [5, 28]. Intuitively, atypical examples lead to a less smooth decision boundary thus a small perturbation to examples is likely to change the prediction. With HACK, we can easily identify atypical samples to improve the robustness of the model. We use MNIST and CIFAR 10 as the benchmark and use FGSM [13] to attack the model with an $\epsilon$ of 0.07 on MNIST and $8/(255^{*}std)$ on CIFAR10, where $std$ is the standard deviation used for normalization. More details of the attack settings can be found in the appendix. We identify the atypical examples based on the norm of the features with HACK and remove the most atypical $X\%$ of the examples. Figure 11 b) shows that discarding atypically examples greatly improves the robustness of the model: the adversarial accuracy is improved by $8.7\%$ when excluding the most atypical $1\%$ of examples on MNIST and $7.3\%$ on CIFAR10. It is worth noting that the clean accuracy remains the same after removing a small number of atypical examples. + +# 6. Summary + +We propose HACK, an unsupervised learning method that organizes images in hyperbolic space using sphere packing. By optimizing image assignments to uniformly distributed particles, HACK leverages the inherent properties of hyperbolic space, leading to the natural emergence of prototypical and semantic structures through feature learning. We validate HACK on synthetic data and standard datasets, demonstrating its ability to discover prototypical examples for reducing sample complexity and increasing model robustness. + +Acknowledgements. This project was supported, in part, by NSF 2131111, NSF 2215542, and NSF 2313151 to S. Yu, and a grant from UT Dallas to Y. Guo. + +# References + +[1] James W Anderson. Hyperbolic geometry. Springer Science & Business Media, 2006. 2 +[2] Wilma A Bainbridge, Phillip Isola, and Aude Oliva. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General, 142(4):1323, 2013. 6 +[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 1 +[4] Jacob Bien and Robert Tibshirani. Prototype selection for interpretable classification. The Annals of Applied Statistics, 5(4):2403-2424, 2011. 1 +[5] Nicholas Carlini, Ulfar Erlingsson, and Nicolas Papernot. Prototypical examples in deep learning: Metrics, characteristics, and utility. 2018. 1, 2, 6, 8 +[6] Ines Chami, Albert Gu, Vaggos Chatziafratis, and Christopher Ré. From trees to continuous embeddings and back: Hyperbolic hierarchical clustering. Advances in Neural Information Processing Systems, 33:15065-15076, 2020. 4 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[8] Lieven Decock and Igor Douven. What is graded membership? Nous, 48(4):653-82, 2014. 1 +[9] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 836-837, 2020. 3 +[10] Evelyn Fix and Joseph Lawson Hodges. Discriminatory analysis. nonparametric discrimination: Consistency properties. International Statistical Review/Revue Internationale de Statistique, 57(3):238-247, 1989. 3 +[11] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. arXiv preprint arXiv:1805.09112, 2018. 2 +[12] Alan Gibbons. Algorithmic graph theory. Cambridge university press, 1985. 5 +[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 8, 1 +[14] Yanming Guo, Yu Liu, Erwin M Bakker, Yuanhao Guo, and Michael S Lew. Cnn-rnn: a large-scale hierarchical image classification framework. Multimedia tools and applications, 77(8):10251-10271, 2018. 3 +[15] Yunhui Guo, Haoran Guo, and Stella Yu. Co-sne: Dimensionality reduction and visualization for hyperbolic data. arXiv preprint arXiv:2111.15037, 2021. 3 +[16] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 6 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2019. 1 +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729–9738, 2020. 2 +[20] Joy Hsu, Jeffrey Gu, Gong-Her Wu, Wah Chiu, and Serena Yeung. Learning hyperbolic representations for unsupervised 3d segmentation. arXiv preprint arXiv:2012.01644, 2020. 3 +[21] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 3 +[22] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in neural information processing systems, 29, 2016. 2 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[25] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 6, 1 +[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 1 +[27] Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2 +[28] Yongshuai Liu, Jiyu Chen, and Hao Chen. Less is more: Culling the training set to improve robustness of deep neural networks. In International Conference on Decision and Game Theory for Security, pages 102-114. Springer, 2018. 8 +[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6, 1 +[30] Emile Mathieu, Charline Le Lan, Chris J Maddison, Ryota Tomioka, and Yee Whye Teh. Continuous hierarchical representations with poincar\`e variational auto-encoders. arXiv preprint arXiv:1901.06033, 2019. 3 +[31] Erik G Miller, Nicholas E Matsakis, and Paul A Viola. Learning from one example through shared densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 464-471. IEEE, 2000. 1, 2, 3, 5, 6 +[32] James Munkres. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 5(1):32-38, 1957. 5 +[33] Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A wrapped normal distribution on + +hyperbolic space for gradient-based learning. In International Conference on Machine Learning, pages 4693-4702. PMLR, 2019. 2, 3 +[34] Maximilian Nickel and Douwe Kiela. Poincar\`e embeddings for learning hierarchical representations. arXiv preprint arXiv:1705.08039, 2017. 2 +[35] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 2, 3, 4 +[36] Eleanor Rosch. Cognitive representations of semantic categories. Journal of experimental psychology: General, 104(3): 192, 1975. 1 +[37] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International Symposium on Graph Drawing, pages 355-366. Springer, 2011. 4 +[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2 +[39] Pierre Stock and Moustapha Cisse. Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In Proceedings of the European Conference on Computer Vision (ECCV), pages 498-512, 2018. 2 +[40] Richard Szeliski et al. Image alignment and stitching: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2(1):1-104, 2007. 3 +[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 5 +[42] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 1 +[43] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 2 +[44] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2603-2612, 2021. 3 +[45] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 4, 2 +[46] Jiei Xi Yan, Lei Luo, Cheng Deng, and Heng Huang. Unsupervised hyperbolic metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12465-12474, 2021. 6, 8 +[47] Jianping Zhang. Selecting typical instances in instance-based learning. In Machine learning proceedings 1992, pages 470-479. Elsevier, 1992. 2 \ No newline at end of file diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/images.zip b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44e07326923b86baee83bffe99240b17bce83641 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d80c40581e2ba70744843026975762fd34b8396a6bb3bb3a30d1e233a84cd527 +size 471877 diff --git a/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/layout.json b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb37fc8467b20329c842fb55b7bafb6d7439c848 --- /dev/null +++ b/2024/Unsupervised Feature Learning with Emergent Data-Driven Prototypicality/layout.json @@ -0,0 +1,8559 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 530, + 121 + ], + "type": "text", + "content": "Unsupervised Feature Learning with Emergent Data-Driven Prototypicality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 532, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 532, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 532, + 172 + ], + "type": "text", + "content": "Yunhui Guo1 Youren Zhang2 Yubei Chen3 Stella X. Yu2,4 \n1The University of Texas at Dallas 2University of Michigan 3UC Davis 4UC Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 288, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 273 + ], + "type": "text", + "content": "Given a set of images, our goal is to map each image to a point in a feature space such that, not only point proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 274, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 392 + ], + "type": "text", + "content": "Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflects image similarity, yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 393, + 288, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 561 + ], + "type": "text", + "content": "We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKKing. HACK first generates uniformly packed particles in the Poincaré ball of hyperbolic space and then assigns each image uniquely to a particle. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing - a warping process that aligns all the images and makes them appear more common and similar to each other, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Not all instances are created equal. For example, the MNIST dataset of handwritten digits contain almost 6,000 samples of 2's; some are close to textbook versions that we are taught to follow, whereas others have idiosyncratic cursive styles, varying in proportions and stroke weights (Fig. 1). Given such a set of natural data, we are interested in dataset summarization and organization such that we can automatically discover which instances are more representative and which ones are anomalies. In other words, we aim to computation-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 198, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 312, + 198, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 312, + 198, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 541, + 322 + ], + "type": "image", + "image_path": "a5593ab71496867a84926998f37b4ca5d63a2740aa848718bd400c1b52cae2c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 331, + 547, + 451 + ], + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 451 + ], + "type": "text", + "content": "Figure 1. Given a dataset (left), we aim to learn an image feature that encodes not only visual similarity between instances but also data-driven prototypicality (right). Additionally, the angular arrangement of the features can naturally serve as a measure of diversity. Our feature encoder (in 2D hyperbolic space) is learned without any labels. We can then learn a decoder to map each point in the feature space back to an image. The right plot visualizes images located at the origin and those moving away in different directions, automatically revealing that 2's with loops are most common and the whole dataset can be grasped as the cursive style systematically varies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "type": "text", + "content": "ally rationalize graded membership [8, 36] of a category in a purely data-driven manner, putting each instance on a feature map that reflects not only their prototypicality in a particular dataset but also their visual similarity with each other." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "content": "Such unsupervised feature learning is useful for not only data organization, but those discovered representative instances or prototypes can be used for interpretable machine learning [4], curriculum learning [3], and learning better decision boundaries [5]. Prototypes also allow us to classify with as few as or even one example [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": "If the image feature is given, it is relatively easy to find prototypes: We just need to identify density peaks of the feature distribution of the image set. Otherwise, discovering prototypical instances without supervision is difficult: There is no universal definition or simple metric to assess the prototypicality of the examples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "One way to address this problem is to examine the gradient magnitude [5]. However, this approach is shown to have a high variance which results from different training setups [5]. Some methods address this problem using adversarial" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23199" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 265, + 224 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 265, + 224 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 265, + 224 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 265, + 224 + ], + "type": "image", + "image_path": "276794e3216f93c4a263cd67fa8d95a3c6d5fcd181c22e76a70a9e90799c3caf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 288, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 288, + 300 + ], + "type": "text", + "content": "Figure 2. Existing self-supervised feature learning methods focus on visual similarity only in Euclidean space, whereas our unsupervised feature learning method embeds images in hyperbolic space which automatically encodes prototypicality with respect to the origin: Images are organized hierarchically, with typical images at the center and atypical ones near the boundary of the Poincaré ball." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 384 + ], + "type": "text", + "content": "robustness [5, 39]: Prototypical examples should be more adversarially robust. However, selecting these examples is dependent on the adversarial method and the metric used in the adversarial attack. Several methods exist but they are either based on heuristics or lack a proper justification [5]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 386, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 458 + ], + "type": "text", + "content": "Naturally, given a feature space, prototypical examples can be identified as density peaks. However, prototypicality varies with the feature. We propose an unsupervised feature learning method in hyperbolic space, which, unlike the shift-invariant Euclidean space, naturally embeds a continuous version of hierarchical trees rooted at the origin of the space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "Hyperbolic space is non-Euclidean space with constant non-negative curvature [1]. Different from Euclidean space, hyperbolic space can represent hierarchical relations with low distortion. Poincaré ball model is one of the most commonly used models for hyperbolic space [35]. One notable property of Poincaré ball model is that the distance to the origin grows exponentially as we move towards the boundary. Thus, points located in the center of the ball are close to all the other points while the points located close to the boundary are infinitely far away from other points. With unsupervised learning in hyperbolic space, we can learn features that capture both visual similarity and hierarchical proximity among instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Our key insight is that a typical image is similar (closest) to more nearby instances than atypical ones, and such an image would be at the center of the neighbourhood in Euclidean space but the parent/root node in a tree in hyperbolic space, closer to the origin. We develop a new learning procedure that first places all the target locations evenly in the Poincaré ball to reflect the desire of mapping uniformity [43], and then we optimize which images should be mapped to which" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 107 + ], + "type": "text", + "content": "target locations in a batch-wise Hungarian matching manner, where those similar to most instances naturally moving closer to the origin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 108, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 312 + ], + "type": "text", + "content": "Our work makes the following contributions. 1) We propose the first unsupervised feature learning method to learn features which capture visual similarity with distance between features and prototypicality with the distance to the origin. 2) We develop a new learning paradigm that sits between supervised learning (with known targets) and unsupervised metric learning (with unknown targets and constrained metric distances). We want to map images to known targets uniformly packed and maximally distant in hyperbolic space, but we learn to optimize the specific image to target assignment. 3) We validate our joint feature learning and data prototypicality discovery on congealing [31], where the consensus is that images after congealing are perceived to be more typical images. 4) We demonstrate two practical usages of data-driven prototypicality: Prototypical and atypical examples are shown to reduce sample complexity for learning and increase the robustness of the model respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 342, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 605 + ], + "type": "text", + "content": "Prototypicality. The study of prototypical examples in machine learning has a long history. In Zhang [47], the authors select typical instances based on the fact that typical instances should be representative of the cluster. In Kim et al. [22], prototypical examples are defined as the examples that have maximum mean discrepancy within the data. Li et al. [27] propose to discover prototypical examples by architectural modifications: project the dataset onto a low-dimensional manifold and use a prototype layer to minimize the distance between inputs and the prototypes on the manifold. The robustness to adversarial attacks is also used as a criterion for prototypicality [39]. In Carlini et al. [5], the authors propose multiple metrics for prototypicality discovery. For example, the features of prototypical examples should be consistent across different training setups. However, these metrics usually depend heavily on the training setups and hyperparameters. The idea of prototypicality is also extensively studied in meta-learning for one-shot or few-shot classification [38]. No existing works address the prototypicality discovery problem in a data-driven fashion. Our proposed HACK naturally exploits hyperbolic space to organize the images based on prototypicality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "content": "Unsupervised Learning in Hyperbolic Space. Learning features in hyperbolic space have shown to be useful for many machine learning problems [11, 34]. One useful property is that hierarchical relations can be embedded in hyperbolic space with low distortion [34]. Wrapped normal distribution, which is a generalized version of the normal distribution for modeling the distribution of points in hyperbolic space [33], is used as the latent space for constructing hyperbolic variational autoencoders (VAEs)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23200" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 286, + 137 + ], + "type": "image", + "image_path": "f15b21f57154c3757da9e0d08a4609b178121bcecf28e67c56360693977f2471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 146, + 287, + 180 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 180 + ], + "type": "text", + "content": "Figure 3. Congeaed images are more typical than the original images. First row: sampled original images. Second row: the corresponding congealed images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 289, + 334 + ], + "type": "text", + "content": "[23]. Poincaré VAEs is constructed in Mathieu et al. [30] with a similar idea to Nagano et al. [33] by replacing the standard normal distribution with hyperbolic normal distribution. Unsupervised 3D segmentation [20] and instance segmentation [44] are conducted in hyperbolic space via hierarchical hyperbolic triplet loss. CO-SNE [15] is recently proposed to visualize high-dimensional hyperbolic features in a two-dimensional hyperbolic space. Although hyperbolic distance facilitates the learning of hierarchical structure, how to leverage hyperbolic space for unsupervised prototypicality discovery is not explored in the current literature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 357, + 155, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 155, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 155, + 371 + ], + "type": "text", + "content": "3. Sample Hierarchy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": "Sample Hierarchy VS. Class Hierarchy. While most of the existing works in hierarchical image classification focus on using label hierarchy [9, 14], there also exists a natural hierarchy among different samples. In Khrulkov et al. [21], the authors conducted an experiment to measure the " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": "-hyperbolicity of the various image datasets and showed that common image datasets such as CIFAR10 and CUB exhibit natural hierarchical structure among the samples. Amongst a collection of images representing digit 1, suppose " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": " is used for representing an image with a digit '1' that is upright, " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": " is used for representing an image with a digit 1 that leaning left and " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}''" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": " is used for representing an image with a digit '1' that leaning right. Given a metric " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": ", if we assume that " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}''', \\mathbf{x}') \\approx d(\\mathbf{x}''', \\mathbf{x}) + d(\\mathbf{x}', \\mathbf{x})" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": ", in this context, we can naturally view the sample " + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 377, + 287, + 568 + ], + "type": "text", + "content": " as the root, and consider the other samples as its children in an underlying tree." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "Compared with class hierarchy which can be extracted based on the pre-defined label relations, sample hierarchy is much harder to construct due to the lack of ground truth. Once a sample hierarchy is established, there are currently no existing methods available for verifying the accuracy of the hierarchy. Additionally, just like with class hierarchies, there may be ambiguities when constructing a sample hierarchy since multiple samples could potentially serve as the root." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Building Sample Hierarchy from Density Peaks. Given existing features " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{f(v_{i})\\}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " obtained by applying a feature extractor for each instance " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": ", prototypical examples can be found by examining the density peaks via techniques from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": "density estimation. For example, the K-nearest neighbor density (K-NN) estimation [10] is defined as " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "p_{knn}(v_i, k) = \\frac{k}{n} \\frac{1}{A_d \\cdot D^d(v_i, v_{k(i)})}" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": " is the feature dimension, " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "A_d = \\pi^{d/2} / \\Gamma(d/2 + 1)" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\Gamma(x)" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": " is the Gamma function and " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "k(i)" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": "th nearest neighbor of example " + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 196 + ], + "type": "text", + "content": ". The nearest neighbors can be found by computing the distance between the features. Therefore, the process of constructing sample hierarchy through density estimation can be conceptualized as a two-step procedure involving: 1) feature learning and 2) detecting density peaks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 196, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 316 + ], + "type": "text", + "content": "In the density estimation approach outlined above, the level of prototypicality depends on the learned features. Varying training setups can induce diverse feature spaces, resulting in differing conclusions on prototypicality. Nevertheless, prototypicality is an inherent attribute of the dataset and should remain consistent across various features. The aim of this paper is to extract features that intrinsically showcase the hierarchical organization of the samples. Specifically, by examining the feature alone within the feature space, we should be able to identify the example's prototypicality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 316, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 542 + ], + "type": "text", + "content": "Construct a Sample Hierarchy from Congealing. To determine whether the feature truly captures prototypicality, it is necessary to identify which sample is the prototype. We ground our concept of prototypicality based on congealing [31]. In particular, we define prototypical examples in the pixel space by examining the distance of the images to the average image in the corresponding class. Our idea is based on a traditional computer vision technique called image alignment [40] that aims to find correspondences across images. During congealing [31], a set of images are transformed to be jointly aligned by minimizing the joint pixel-wise entropies. The congealed images are more prototypical: they are better aligned with the average image. Thus, we have a simple way to transform an atypical example into a typical example (see Figure 3). This is useful since given an unlabeled image dataset the typicality of the examples is unknown, congealing examples can be naturally served as examples with known typicality and be used as a validation for the effectiveness of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 554, + 543, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 543, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 543, + 568 + ], + "type": "text", + "content": "4. Unsupervised Hyperbolic Feature Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 574, + 526, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 526, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 526, + 587 + ], + "type": "text", + "content": "4.1. Poincaré Ball Model for Hyperbolic Space" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": "Euclidean space has a curvature of zero and a hyperbolic space is a Riemannian manifold with constant negative curvature.. There are several isometrically equivalent models for visualizing hyperbolic space with Euclidean representation. The Poincaré ball model is the commonly used one in hyperbolic representation learning [35]. The " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": "-dimensional Poincaré ball model is defined as " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbb{B}^n, \\mathfrak{g}_{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{B}^n = \\{\\mathbf{x} \\in \\mathbb{R}^n : \\| \\mathbf{x} \\| < 1\\}" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathfrak{g}_{\\mathbf{x}} = (\\gamma_{\\mathbf{x}})^2 I_n" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": " is the Riemannian metric tensor. " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathbf{x}} = \\frac{2}{1 - \\| \\mathbf{x} \\|^2}" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": " is the conformal factor and " + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 304, + 593, + 547, + 713 + ], + "type": "text", + "content": " is the Euclidean metric tensor." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23201" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 72, + 205, + 167 + ], + "blocks": [ + { + "bbox": [ + 90, + 72, + 205, + 167 + ], + "lines": [ + { + "bbox": [ + 90, + 72, + 205, + 167 + ], + "spans": [ + { + "bbox": [ + 90, + 72, + 205, + 167 + ], + "type": "image", + "image_path": "ea7971f0a81d7aebf36b97a9e34f3d5b5af9b96b25a3652bb9f957912665893c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 170, + 202, + 194 + ], + "lines": [ + { + "bbox": [ + 90, + 170, + 202, + 194 + ], + "spans": [ + { + "bbox": [ + 90, + 170, + 202, + 194 + ], + "type": "text", + "content": "a) Supervised classification with fixed known targets" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 248, + 69, + 347, + 167 + ], + "blocks": [ + { + "bbox": [ + 248, + 69, + 347, + 167 + ], + "lines": [ + { + "bbox": [ + 248, + 69, + 347, + 167 + ], + "spans": [ + { + "bbox": [ + 248, + 69, + 347, + 167 + ], + "type": "image", + "image_path": "a944af3952c3dc778a0d0d2b62ba17bfab10fd3a134687c0c48faaeccf4799c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 170, + 372, + 194 + ], + "lines": [ + { + "bbox": [ + 222, + 170, + 372, + 194 + ], + "spans": [ + { + "bbox": [ + 222, + 170, + 372, + 194 + ], + "type": "text", + "content": "b) Our unsupervised feature learning with fixed but unknown targets" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 204, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 258 + ], + "type": "text", + "content": "Figure 4. The proposed HACK has a predefined geometrical arrangement and allows the images to be freely assigned to any particle. a) Standard supervised learning has predefined targets. The image is only allowed to be assigned to the corresponding target. b) HACK packs particles uniformly in hyperbolic space to create initial seeds for the organization. The images are assigned to the particles based on their prototypicality and semantic similarities. c) Standard unsupervised learning has no predefined targets and images are clustered based on their semantic similarities." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 69, + 496, + 167 + ], + "blocks": [ + { + "bbox": [ + 389, + 69, + 496, + 167 + ], + "lines": [ + { + "bbox": [ + 389, + 69, + 496, + 167 + ], + "spans": [ + { + "bbox": [ + 389, + 69, + 496, + 167 + ], + "type": "image", + "image_path": "3a0056c4cda138179dc44ec507a89dfa6785467f6ab0b84624bb34a96f4607a1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 170, + 496, + 194 + ], + "lines": [ + { + "bbox": [ + 391, + 170, + 496, + 194 + ], + "spans": [ + { + "bbox": [ + 391, + 170, + 496, + 194 + ], + "type": "text", + "content": "c) Metric feature learning with unknown targets" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "content": "Hyperbolic Distance. Given two points " + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{u} \\in \\mathbb{B}^n" + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{v} \\in \\mathbb{B}^n" + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "content": ", the hyperbolic distance is defined as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 313, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 313, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 52, + 313, + 287, + 342 + ], + "type": "interline_equation", + "content": "d _ {\\mathbb {B} ^ {n}} (\\boldsymbol {u}, \\boldsymbol {v}) = \\operatorname {a r c o s h} \\left(1 + 2 \\frac {\\| \\boldsymbol {u} - \\boldsymbol {v} \\| ^ {2}}{(1 - \\| \\boldsymbol {u} \\| ^ {2}) (1 - \\| \\boldsymbol {v} \\| ^ {2})}\\right) \\tag {1}", + "image_path": "4b6fec15145bdb96b6a42a8b0296514d2e1780e6fff232e1f9c290787bc7c3c0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\operatorname{arcosh}" + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "content": " is the inverse hyperbolic cosine function and " + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "content": " is the usual Euclidean norm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "content": "Hyperbolic distance has the unique property that it grows exponentially as we move towards the boundary of the Poincaré ball. In particular, the points on the circle represent points in infinity. Hyperbolic space is naturally suitable for embedding hierarchical structure [35, 37] and can be regarded as a continuous representation of trees [6]. The hyperbolic distance between samples implicitly reflects their hierarchical relation. Thus, by embedding images in hyperbolic space we can naturally organize images based on their semantic similarity and prototypicality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 502, + 285, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 285, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 285, + 515 + ], + "type": "text", + "content": "4.2. Build Instance Hierarchy in Hyperbolic Space" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": "Hyperbolic space is naturally suitable for embedding tree structure. However, in order to leverage hyperbolic space to build a sample hierarchy in an unsupervised manner, a suitable objective function is still missing. There are two challenges in designing the objective function. First, the underlying tree structure of the samples is unknown. Second, how to perform feature learning such that hierarchy can naturally emerge is unclear. In this paper, we propose Hyperbolic space with sphere pACKing, also called HACK, to address the two challenges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "To address the first challenge, instead of creating a predefined tree structure that might not faithfully represent the genuine hierarchical organization, we leverage sphere packing in hyperbolic space for building a skeleton for placing the samples. We specify where the particles should be located ahead of training with uniform packing, which by design are" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 327 + ], + "type": "text", + "content": "maximally evenly spread out in hyperbolic space. The uniformly distributed particles guide feature learning to achieve maximum instance discrimination [45] while enabling us to extract a tree structure from the samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 330, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 498 + ], + "type": "text", + "content": "To address the second challenge, HACK figures out which instance should be mapped to which target through bipartite graph matching as a global optimization procedure. During training, HACK minimizes the total hyperbolic distances between the mapped image point (in the feature space) and the target, those that are more typical naturally emerge closer to the origin of Poincaré ball. HACK differs from the existing learning methods in several aspects (Figure 4). Different from supervised learning, HACK allows the image to be assigned to any target (particle). This enables the exploration of the natural organization of the data. Different from unsupervised learning method, HACK specifies a predefined geometrical organization which encourages the corresponding structure to be emerged from the dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 510, + 499, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 499, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 499, + 525 + ], + "type": "text", + "content": "4.3. Sphere Packing in Hyperbolic Space" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " particles, our goal is to pack the particles into a two-dimensional hyperbolic space as densely as possible. We derive a simple repulsion loss function to encourage the particles to be equally distant from each other. The loss is derived via the following steps. First, we need to determine the radius of the Poincaré ball used for packing. We use a curvature of 1.0 so the radius of the Poincaré ball is 1.0. The whole Poincaré ball cannot be used for packing since the volume is infinite. We use " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r < 1" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " to denote the actual radius used for packing. Thus, our goal is to pack " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " particles in a compact subspace of Poincaré ball. Then, the Euclidean radius " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " is further converted into hyperbolic radius " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{B}}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "s = \\frac{1}{\\sqrt{c}}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " is the curvature. The relation between " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{B}}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{B}} = s \\log \\frac{s + r}{s - r}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": ". Next, the total hyperbolic area " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathbb{B}}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " of a Poincaré ball of radius " + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{B}}" + }, + { + "bbox": [ + 304, + 531, + 547, + 713 + ], + "type": "text", + "content": " can" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23202" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 293, + 148 + ], + "type": "image", + "image_path": "55191b4684c72402edc1035939655b1c1b6991e08d2593be404a320cc4060c43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 289, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 204 + ], + "type": "text", + "content": "Figure 5. HACK conducts unsupervised learning in hyperbolic space with sphere packing. The images are mapped to particles by minimizing the total hyperbolic distance. HACK learns features that can capture both visual similarities and prototypicality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": "be computed as " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "A_{\\mathbb{B}} = 4\\pi s^2 \\sinh^2 \\left( \\frac{r_{\\mathbb{B}}}{2s} \\right)" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\sinh" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": " is the hyperbolic sine function. Finally, the area per point " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": " can be easily computed as " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\frac{A_{\\mathbb{B}}}{n}" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": " is the total number of particles. Given " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": ", the radius per point can be computed as " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "r_{n} = 2s \\sinh^{-1} \\left( \\sqrt{\\frac{A_{n}}{4\\pi s^{2}}} \\right)" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": ". We use the following loss to generate uniform packing in hyperbolic space. Given two particles " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": ", the repulsion loss " + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 228, + 287, + 319 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 325, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 287, + 356 + ], + "type": "interline_equation", + "content": "V (i, j) = \\left\\{\\frac {1}{\\left[ 2 r _ {n} - \\max \\left(0 , 2 r _ {n} - d _ {\\mathbb {B}} (i , j)\\right) \\right] ^ {k}} - \\frac {1}{\\left(2 r _ {n}\\right) ^ {k}} \\right\\} \\cdot C (k) \\tag {2}", + "image_path": "e877f78ba372bbcfbd6dafb388c4ddd74b7ddd2099e7da9fa1cb18940036834d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "C(k) = \\frac{(2r_n)^{k + 1}}{k}" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " is a hyperparameter. Intuitively, if the particle " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " and the particle " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " are within " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "2r_n" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": ", the repulsion loss is positive. Minimizing the repulsion loss would push the particles " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 356, + 287, + 455 + ], + "type": "text", + "content": " away. If the repulsion is zero, this indicates all the particles are equally distant. Also the repulsion loss grows significantly when two particles become close. We also adopt the following boundary loss to prevent the particles from escaping the ball," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 464, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 464, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 86, + 464, + 287, + 475 + ], + "type": "interline_equation", + "content": "B (i; r) = \\max \\left(0, \\operatorname {n o r m} _ {i} - r + \\text {m a r g i n}\\right) \\tag {3}", + "image_path": "a1ff5cc1b4947661aba31517bc05c8c627d942c295db36dfbced22058b6080fc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{norm}_i" + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "content": " norm of the representation of the particle " + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "content": ". Figure 4 b) shows an example of the generated particles that are uniformly packed in hyperbolic space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 527, + 222, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 222, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 222, + 540 + ], + "type": "text", + "content": "4.4. Hyperbolic Instance Assignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "HACK learns the features by optimizing the assignments of the images to particles (Figure 5). The assignment should be one-to-one, i.e., each image is assigned to one particle and each particle is allowed to be associated with one image. We cast the instance assignment problem as a bipartite matching problem [12] and solve it with Hungarian algorithm [32]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Initially, we randomly assign the particles to the images, thus there is a random one-to-one correspondence between the images to the particles (not optimized). Given a batch of samples " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_1,\\mathbf{s}_1),(\\mathbf{x}_2,\\mathbf{s}_2),\\dots,(\\mathbf{x}_B,\\mathbf{s}_B)\\}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is an image and " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is the corresponding particle, and an encoder " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", we generate the hyperbolic feature for each image " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_i) \\in \\mathbb{B}^2" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{B}^2" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is a two-dimensional Poincaré ball. For a given hyperbolic feature " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", with fixed particle" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 83, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 83, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 547, + 108 + ], + "type": "text", + "content": "Algorithm 1 HACK: Unsupervised Learning in Hyperbolic Space. ised Learning in Hyperbolic Space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "text", + "content": "Require: # of images: " + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "inline_equation", + "content": "n \\geq 0" + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "text", + "content": ". Radius for packing: " + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "inline_equation", + "content": "r < 1" + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "text", + "content": ". An encoder with parameters " + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 319, + 109, + 547, + 133 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 133, + 547, + 205 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 325, + 133, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 133, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 325, + 133, + 547, + 167 + ], + "type": "text", + "content": "1: Generate uniformly distributed particles in hyperbolic space by minimizing the repulsion loss in Equation 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "text", + "content": "2: Given " + }, + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_1, s_1), (\\mathbf{x}_2, s_2), \\dots, (\\mathbf{x}_b, s_b)\\}" + }, + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "text", + "content": ", optimize " + }, + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 325, + 169, + 547, + 205 + ], + "type": "text", + "content": " by minimizing the total hyperbolic distance via Hungarian algorithm." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 229, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 313 + ], + "type": "text", + "content": "locations, the distance between the hyperbolic feature and the particles signifies the hierarchical level of the associated sample. Thus, to determine the hierarchical levels for all samples within the hierarchy, we must establish a one-to-one mapping between all the samples and the particles. This can be cast as the following bipartite matching problem in hyperbolic space," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 320, + 545, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 349, + 320, + 545, + 353 + ], + "type": "interline_equation", + "content": "\\ell (\\theta , \\pi) = \\sum_ {i = 1} ^ {B} d _ {\\mathbb {B} ^ {n}} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {s} _ {\\pi \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right)\\right)}\\right) \\tag {4}", + "image_path": "3e1752556f863155a31def2fa12f8f24c477f86755131fc95112f38532c437f6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\pi : f_{\\theta}(\\mathbf{x}) \\to \\mathbb{N}" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " is a projection function which projects hyperbolic features to a particle index. Minimizing " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\ell(\\theta, \\pi)" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " is a combinatorial optimization problem, which can hardly be optimized with " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " using gradient-based algorithms. Thus, we adopt a joint optimization strategy which optimizes " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " alternatively. In each batch, we first leverage the Hungarian algorithm [32] to find the optimal matching " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\pi^{*}" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " based on the current hyperbolic features. Then we minimize Eq. 4 with respect to " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": " based on the current assignment " + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\pi^{*}" + }, + { + "bbox": [ + 304, + 360, + 546, + 516 + ], + "type": "text", + "content": ". This process is repeated for a certain number of epochs until convergence is achieved. On the other hand, the feature encoder can serve as an image prior for assigning similar images to nearby particles [41]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "text", + "content": "The Hungarian algorithm [32] has a complexity of " + }, + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(x^3)" + }, + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 515, + 547, + 612 + ], + "type": "text", + "content": " is the number of items. As we perform the particle assignment in the batch level, the time and memory complexity is tolerable. Also, the one-to-one correspondence between the images and particles is always maintained during training. After training, based on the assigned particle, the level of the sample in the hierarchy can be easily retrieved. The details of HACK are shown in Algorithm 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "We design several experiments to show the effectiveness of HACK for the semantic and hierarchical organization. First, we first construct a dataset with known hierarchical structure using the congealing algorithm [31]. Then, we apply HACK to datasets with unknown hierarchical structure to organize the samples based on the semantic and prototypical structure." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23203" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "content": "Finally, we show that the prototypical structure can be used to reduce sample complexity and increase model robustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 289, + 397 + ], + "type": "text", + "content": "Datasets. We first construct a dataset called Congeaed MNIST. To verify the efficacy of HACK for unsupervised prototypicality discovery, we need a benchmark with known prototypical examples. However, currently there is no standard benchmark for this purpose. To construct the benchmark, we use the congealing algorithm from Miller et al. [31] to align the images in each class of MNIST [25]. The congealing algorithm is initially used for one-shot classification. During congealing, the images are brought into correspondence with each other jointly. The congealed images are more prototypical: they are better aligned with the average image. The synthetic data is generated by replacing 500 original images with the corresponding congealed images. In the Appendix, we show the results of changing the number of replaced original images. We expect HACK to discover the congealed images and place them in the center of the Poincaré ball. We also aim to discover the prototypical examples from each class of the standard MNIST dataset [25] and CIFAR10 [24]. CIFAR10 consists of 60000 from 10 object categories ranging from airplane to truck. CIFAR10 is more challenging than MNIST since it has larger intra-class variations. Moreover, to better visualize how HACK arranges the samples according to their prototypicality, we run HACK on 10k US Adult Faces [2] (hereafter referred to as USA10kF), which contains 10,168 natural face photographs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 397, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 288, + 434 + ], + "type": "text", + "content": "Baselines. We consider several existing metrics proposed in Carlini et al. [5] for prototypicality discovery, the details can be found in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 436, + 288, + 579 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 436, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 288, + 483 + ], + "type": "text", + "content": "- Holdout Retraining [5]: We consider the Holdout Retraining proposed in Carlini et al. [5]. The idea is that the distance of features of prototypical examples obtained from models trained on different datasets should be close." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 483, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 288, + 531 + ], + "type": "text", + "content": "- Model Confidence [5]: Intuitively, the model should be confident in prototypical examples. Thus, it is natural to use the confidence of the model prediction as the criterion for prototypicality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 531, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 288, + 579 + ], + "type": "text", + "content": "- UHML [46]: UHML is an unsupervised hyperbolic learning method that aims to discover the natural hierarchies of data by taking advantage of hyperbolic metric learning and hierarchical clustering." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Implementation Details. We implement HACK in PyTorch and the code will be made public. To generate uniform particles, we first randomly initialize the particles and then run the training for 1000 epochs with a 0.01 learning rate to minimize the repulsion loss and boundary loss. The curvature of the Poincaré ball is 1.0 and the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is 0.76 which is used to alleviate the numerical issues [16]. The hyperparameter " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is 1.55 which is shown to generate uniform particles well. For the assignment, we use a LeNet [26] for MNIST, a ResNet20 [17] for CIFAR10, and a ResNet18 for USA10kF as the encoder. We apply HACK to each class separately." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 517, + 173 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 517, + 173 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 517, + 173 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 517, + 173 + ], + "type": "image", + "image_path": "54cedbae1e2a22363ac63e66e749e4faffb04e42e2a4d99b18a936698fe67fd5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "type": "text", + "content": "Figure 6. Hyperbolic space can capture the prototypicality inherently. The error bar of each point is given by the variance of density within the corresponding portion, and the width of the shaded band indicates the number of features within the portion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 237, + 417, + 342 + ], + "blocks": [ + { + "bbox": [ + 314, + 237, + 417, + 342 + ], + "lines": [ + { + "bbox": [ + 314, + 237, + 417, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 417, + 342 + ], + "type": "image", + "image_path": "f5b57a190c1e8e84c5405ed046eb5cf8fa8a9351c9f5949bceb2b4f513a62d6a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 346, + 372, + 356 + ], + "lines": [ + { + "bbox": [ + 361, + 346, + 372, + 356 + ], + "spans": [ + { + "bbox": [ + 361, + 346, + 372, + 356 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 366, + 539, + 431 + ], + "lines": [ + { + "bbox": [ + 312, + 366, + 539, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 539, + 431 + ], + "type": "text", + "content": "Figure 7. Congeaed images are located in the center of the Poincaré ball. a) Red dots denote congealed images and cyan dots denote original images. b) Typical images are in the center and atypical images are close to the boundary. Images are also clustered together based on visual similarity. Congeaed images are shown in red boxes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 237, + 522, + 342 + ], + "blocks": [ + { + "bbox": [ + 417, + 237, + 522, + 342 + ], + "lines": [ + { + "bbox": [ + 417, + 237, + 522, + 342 + ], + "spans": [ + { + "bbox": [ + 417, + 237, + 522, + 342 + ], + "type": "image", + "image_path": "baefb88089f2b9b7a214982cb21013a009ab59dca4fbbfa0c46761a2468ebb6e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 345, + 474, + 356 + ], + "lines": [ + { + "bbox": [ + 463, + 345, + 474, + 356 + ], + "spans": [ + { + "bbox": [ + 463, + 345, + 474, + 356 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 456, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 540 + ], + "type": "text", + "content": "We attach a fully connected layer to project the feature into a two-dimensional Euclidean space. The image features are then projected onto hyperbolic space via an exponential map. We run the training for 200 epochs using a cosine learning rate scheduler [29] with an initial learning rate of 0.1. We optimize the assignment every other epoch. All the experiments are run on an NVIDIA Titan RTX GPU." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "type": "text", + "content": "5.1. Prototypicality in the Hyperbolic Feature Norm" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "We explicitly show that the hyperbolic space can capture prototypicality by analyzing the relation between hyperbolic norms and the K-NN density estimation. Taking the learned hyperbolic features, we first divide the range of norms of hyperbolic features into numerous portions with equal length (50 portions for this plot). The mean K-NN density is calculated by averaging the density estimation of features within each portion. Figure 6 shows that the mean density drops as the norm increases, which shows that the prototypicality emerges automatically within the norms, the inherent characteristic of hyperbolic space. This validates that prototypicality is reflected in the hyperbolic feature norm." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23204" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 292, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 292, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 292, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 292, + 195 + ], + "type": "image", + "image_path": "5290b646bd25251ae3bba89047cb5c40f749d28c5f59b32db759504005a724d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 239 + ], + "type": "text", + "content": "Figure 8. Our unsupervised learning methods conform to our visual perception. The images are organized from left to right, top to bottom to cover the 360 degrees at the same radius." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 252, + 282, + 300 + ], + "blocks": [ + { + "bbox": [ + 59, + 252, + 282, + 300 + ], + "lines": [ + { + "bbox": [ + 59, + 252, + 282, + 300 + ], + "spans": [ + { + "bbox": [ + 59, + 252, + 282, + 300 + ], + "type": "image", + "image_path": "ced45d0b8401eb597f4e9ac3ca5a7636710f36f4b0a359fe299da78cfd610503.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 288, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 346 + ], + "type": "text", + "content": "Figure 9. HACK discovers typical and atypical images from the data. First row: atypical images with large hyperbolic norm. Second row: typical images with small hyperbolic norm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 171, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 171, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 171, + 382 + ], + "type": "text", + "content": "5.2. Visual Prototypicality" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 389, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 651 + ], + "type": "text", + "content": "Congeaed MNIST. We further apply HACK for visual feature learning on congealed MNIST. Figure 7 shows that HACK can discover the congealed images from all images. In Figure 7 a), the red particles denote the congealed images and cyan particles denote the original images. We can observe that the congealed images are assigned to the particles located in the center of the Poincaré ball. This verifies that HACK can indeed discover prototypical examples from the original dataset. In the Appendix, we show that the features of atypical examples gradually move to the boundary of the Poincaré ball during training. In Figure 7 b), we show the actual images that are embedded in the two-dimensional hyperbolic space. We can observe that the images in the center of Poincaré ball are more prototypical and images close to the boundary are more atypical. Also, the images are naturally organized by their semantic similarity. In summary, HACK can discover prototypicality and also organize the images based on their semantic and hierarchical structure. To the best of our knowledge, this is the first unsupervised learning method that can be used to discover prototypical examples in a data-driven fashion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "USA10kF. Figure 10 a) shows the assignment of 2000 images sampled from USA10kF. Compared to MNIST, the variation in faces is much more complex. A facial image is also subject to various factors (such as race, facial expres" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 432, + 205 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 432, + 205 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 432, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 432, + 205 + ], + "type": "image", + "image_path": "7fb203d23b82dadb6c754b9fc394d4ebcc1286b17e22c7abf5b827b58812a213.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 216, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 282 + ], + "type": "text", + "content": "Figure 10. HACK captures prototypicality of faces. a) The assignment of 2000 images sampled from USA10kF. b) The angles indicate the uniform division of the space. For each orientation, 8 equidistant sample points were selected and the nearest faces are shown at each point. The face located in the center is most symmetrical with clear smiling" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 75, + 549, + 205 + ], + "blocks": [ + { + "bbox": [ + 432, + 75, + 549, + 205 + ], + "lines": [ + { + "bbox": [ + 432, + 75, + 549, + 205 + ], + "spans": [ + { + "bbox": [ + 432, + 75, + 549, + 205 + ], + "type": "image", + "image_path": "182ae3765979e2e5ba3fd745f897d65d6c093e11dfc4624c5b218a10383bff9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 546, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 459 + ], + "type": "text", + "content": "sion, environmental condition, etc.). Therefore, the results from USA10kF are less intuitive than those from MNIST. However, Figure 10 b) illustrates the evolutionary process of images originating from the center of hyperbolic space and progressing along different directions. In the space, we selected 12 directions at equal angular intervals and chose five equally spaced sampling points in each direction. The images closest to these sampling points are displayed at their respective locations. Although the detailed organization is unclear, the evolutionary process reveals a tendency of HACK to cluster different features together, such as darker skin tones appearing in the 0-90 degree range and lighter skin tones in the 180-270 degree range." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 567 + ], + "type": "text", + "content": "MNIST and CIFAR10. Figure 8 shows the embedding of class 0 from MNIST and class \"airplane\" from CIFAR10 arranged to cover 360 degrees of at the same radius. The visual similarity of the images has a smooth transition as we move around angularly. Figure 9 shows the typical images and atypical images discovered by HACK. This further illustrates that HACK captures the semantic similarity of the images which enables prototypicality discovery. Please see the Appendix for more results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 575, + 501, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 501, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 501, + 588 + ], + "type": "text", + "content": "5.3. Prototypicality for Instance Selection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Figure 12 shows the comparison of the baselines with HACK. With HACK, typical images are characterized by the smallest hyperbolic norms, whereas atypical images are associated with the largest hyperbolic norms. We can observe that both HACK and Model Confidence (MC) can discover typical and atypical images. Compared with MC, HACK defines prototypicality as the distance of the sample to other samples which is more aligned with human intuition. Moreover, in addition to prototypicality, HACK can also be used to organize examples by semantic similarities. Holdout Retraining" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23205" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 177, + 158 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 177, + 158 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 177, + 158 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 177, + 158 + ], + "type": "image", + "image_path": "4c05788586425e73d576dd78e4898a73937d90946ee5911e53b131820a60dbdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 162, + 242, + 174 + ], + "lines": [ + { + "bbox": [ + 113, + 162, + 242, + 174 + ], + "spans": [ + { + "bbox": [ + 113, + 162, + 242, + 174 + ], + "type": "text", + "content": "a) Reducing sample complexity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 296, + 158 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 296, + 158 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 296, + 158 + ], + "type": "image", + "image_path": "d85ab442833a112a00ad3d812c1c3c141d6095c964e89a8d21563522f3569089.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 416, + 158 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 416, + 158 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 416, + 158 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 416, + 158 + ], + "type": "image", + "image_path": "386b08f9ce6773421c28f95dfd7aa2bc350035b9715c24859845bbac42578507.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 162, + 478, + 174 + ], + "lines": [ + { + "bbox": [ + 352, + 162, + 478, + 174 + ], + "spans": [ + { + "bbox": [ + 352, + 162, + 478, + 174 + ], + "type": "text", + "content": "b) Increasing model robustness" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 417, + 72, + 534, + 158 + ], + "blocks": [ + { + "bbox": [ + 417, + 72, + 534, + 158 + ], + "lines": [ + { + "bbox": [ + 417, + 72, + 534, + 158 + ], + "spans": [ + { + "bbox": [ + 417, + 72, + 534, + 158 + ], + "type": "image", + "image_path": "88350d94a95e075163cd1a55a0e1f34571fd89ebd9016c3ee610ac6f2c79c569.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 236, + 287, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 217 + ], + "type": "text", + "content": "Figure 11. HACK can be used to construct sample prototypical hierarchy which is useful for several downstream tasks. a) Training with atypical examples achieves higher accuracy than training with typical examples. b) The adversarial accuracy greatly improves after removing the X% of most atypical examples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "type": "image", + "image_path": "b83c9dba7df38a388ffe1157b9adef0de0c2497ea12d41b52cb29ce2debf743e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 288, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 354 + ], + "type": "text", + "content": "Figure 12. HACK can capture both prototypicality and diversity in the dataset. Each row shows the typical and atypical images discovered by HACK/HR [5]/MC [5]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 400 + ], + "type": "text", + "content": "(HR) is not effective for prototypicality discovery due to the randomness of model training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 407, + 209, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 209, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 209, + 422 + ], + "type": "text", + "content": "5.4. Application of Prototypicality" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 426, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 617 + ], + "type": "text", + "content": "Reducing Sample Complexity. The proposed HACK can discover prototypical images as well as atypical images. We show that with atypical images we can reduce the sample complexity for training the model. Prototypical images are representative of the dataset but lack variations. Atypical examples contain more variations and it is intuitive that models trained on atypical examples should generalize better to the test samples. To verify this hypothesis, we select a subset of samples based on the norm of the features which indicates prototypicality of the examples. In particular, typical samples correspond to the samples with smaller norms and atypical samples correspond to the samples with larger norms. The angular layout of the hyperbolic features naturally captures sample diversity, thus for selecting atypical examples, we consider introducing more diversity by sampling images with large norms along the angular direction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Figure 11 a) shows that training with atypical images can achieve much higher accuracy than training with typical images. In particular, training with the most atypical " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " of the images achieves " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "22.67\\%" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " higher accuracy than with the most typical " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " of the images on CIFAR10. Similar results can be observed on MNIST. Thus, HACK provides an easy solution to reduce sample complexity. We also compared UHML [46], which is an unsupervised metric learning in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "content": "hyperbolic space, with HACK on the MNIST dataset. By incorporating " + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "content": " atypical samples based on feature norm, HACK outperformed UHML by " + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "content": ". Also by excluding the " + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "content": " atypical examples, HACK achieved an additional " + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 297 + ], + "type": "text", + "content": " improvement over UHML." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": "Increasing Model Robustness. Training models with atypical examples can lead to a vulnerable model to adversarial attacks [5, 28]. Intuitively, atypical examples lead to a less smooth decision boundary thus a small perturbation to examples is likely to change the prediction. With HACK, we can easily identify atypical samples to improve the robustness of the model. We use MNIST and CIFAR 10 as the benchmark and use FGSM [13] to attack the model with an " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " of 0.07 on MNIST and " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "8/(255^{*}std)" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " on CIFAR10, where " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "std" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " is the standard deviation used for normalization. More details of the attack settings can be found in the appendix. We identify the atypical examples based on the norm of the features with HACK and remove the most atypical " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "X\\%" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " of the examples. Figure 11 b) shows that discarding atypically examples greatly improves the robustness of the model: the adversarial accuracy is improved by " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " when excluding the most atypical " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " of examples on MNIST and " + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 304, + 301, + 547, + 529 + ], + "type": "text", + "content": " on CIFAR10. It is worth noting that the clean accuracy remains the same after removing a small number of atypical examples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 545, + 372, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 372, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 372, + 559 + ], + "type": "text", + "content": "6. Summary" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "type": "text", + "content": "We propose HACK, an unsupervised learning method that organizes images in hyperbolic space using sphere packing. By optimizing image assignments to uniformly distributed particles, HACK leverages the inherent properties of hyperbolic space, leading to the natural emergence of prototypical and semantic structures through feature learning. We validate HACK on synthetic data and standard datasets, demonstrating its ability to discover prototypical examples for reducing sample complexity and increasing model robustness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project was supported, in part, by NSF 2131111, NSF 2215542, and NSF 2313151 to S. Yu, and a grant from UT Dallas to Y. Guo." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23206" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 112 + ], + "type": "text", + "content": "[1] James W Anderson. Hyperbolic geometry. Springer Science & Business Media, 2006. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 147 + ], + "type": "text", + "content": "[2] Wilma A Bainbridge, Phillip Isola, and Aude Oliva. The intrinsic memorability of face photographs. Journal of Experimental Psychology: General, 142(4):1323, 2013. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "text", + "content": "[4] Jacob Bien and Robert Tibshirani. Prototype selection for interpretable classification. The Annals of Applied Statistics, 5(4):2403-2424, 2011. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 259 + ], + "type": "text", + "content": "[5] Nicholas Carlini, Ulfar Erlingsson, and Nicolas Papernot. Prototypical examples in deep learning: Metrics, characteristics, and utility. 2018. 1, 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "text", + "content": "[6] Ines Chami, Albert Gu, Vaggos Chatziafratis, and Christopher Ré. From trees to continuous embeddings and back: Hyperbolic hierarchical clustering. Advances in Neural Information Processing Systems, 33:15065-15076, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 373 + ], + "type": "text", + "content": "[8] Lieven Decock and Igor Douven. What is graded membership? Nous, 48(4):653-82, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 430 + ], + "type": "text", + "content": "[9] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 836-837, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "text", + "content": "[10] Evelyn Fix and Joseph Lawson Hodges. Discriminatory analysis. nonparametric discrimination: Consistency properties. International Statistical Review/Revue Internationale de Statistique, 57(3):238-247, 1989. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 508 + ], + "type": "text", + "content": "[11] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. arXiv preprint arXiv:1805.09112, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 510, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 532 + ], + "type": "text", + "content": "[12] Alan Gibbons. Algorithmic graph theory. Cambridge university press, 1985. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 566 + ], + "type": "text", + "content": "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 8, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "text", + "content": "[14] Yanming Guo, Yu Liu, Erwin M Bakker, Yuanhao Guo, and Michael S Lew. Cnn-rnn: a large-scale hierarchical image classification framework. Multimedia tools and applications, 77(8):10251-10271, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "text", + "content": "[15] Yunhui Guo, Haoran Guo, and Stella Yu. Co-sne: Dimensionality reduction and visualization for hyperbolic data. arXiv preprint arXiv:2111.15037, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[16] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729–9738, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "text", + "content": "[20] Joy Hsu, Jeffrey Gu, Gong-Her Wu, Wah Chiu, and Serena Yeung. Learning hyperbolic representations for unsupervised 3d segmentation. arXiv preprint arXiv:2012.01644, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 220, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 273 + ], + "type": "text", + "content": "[21] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "text", + "content": "[22] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 320, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 342 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 343, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 365 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "text", + "content": "[25] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 6, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "text", + "content": "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "text", + "content": "[27] Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[28] Yongshuai Liu, Jiyu Chen, and Hao Chen. Less is more: Culling the training set to improve robustness of deep neural networks. In International Conference on Decision and Game Theory for Security, pages 102-114. Springer, 2018. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6, 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 556, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 601 + ], + "type": "text", + "content": "[30] Emile Mathieu, Charline Le Lan, Chris J Maddison, Ryota Tomioka, and Yee Whye Teh. Continuous hierarchical representations with poincar\\`e variational auto-encoders. arXiv preprint arXiv:1901.06033, 2019. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 601, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 655 + ], + "type": "text", + "content": "[31] Erik G Miller, Nicholas E Matsakis, and Paul A Viola. Learning from one example through shared densities on transforms. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 464-471. IEEE, 2000. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[32] James Munkres. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 5(1):32-38, 1957. 5" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A wrapped normal distribution on" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23207" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 667 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "hyperbolic space for gradient-based learning. In International Conference on Machine Learning, pages 4693-4702. PMLR, 2019. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "text", + "content": "[34] Maximilian Nickel and Douwe Kiela. Poincar\\`e embeddings for learning hierarchical representations. arXiv preprint arXiv:1705.08039, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "type": "text", + "content": "[35] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 206 + ], + "type": "text", + "content": "[36] Eleanor Rosch. Cognitive representations of semantic categories. Journal of experimental psychology: General, 104(3): 192, 1975. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "text", + "content": "[37] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International Symposium on Graph Drawing, pages 355-366. Springer, 2011. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 275 + ], + "type": "text", + "content": "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 288, + 320 + ], + "type": "text", + "content": "[39] Pierre Stock and Moustapha Cisse. Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In Proceedings of the European Conference on Computer Vision (ECCV), pages 498-512, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 322, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 288, + 354 + ], + "type": "text", + "content": "[40] Richard Szeliski et al. Image alignment and stitching: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2(1):1-104, 2007. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "type": "text", + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "type": "text", + "content": "[42] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "text", + "content": "[43] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "type": "text", + "content": "[44] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2603-2612, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "type": "text", + "content": "[45] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 4, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "text", + "content": "[46] Jiei Xi Yan, Lei Luo, Cheng Deng, and Heng Huang. Unsupervised hyperbolic metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12465-12474, 2021. 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "text", + "content": "[47] Jianping Zhang. Selecting typical instances in instance-based learning. In Machine learning proceedings 1992, pages 470-479. Elsevier, 1992. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23208" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_content_list.json b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e2eed468b29bbed7944e7ca929126a7c3752ab4a --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "Unsupervised Gaze Representation Learning from Multi-view Face Images", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiwei Bao Feng Lu* \nState Key Laboratory of VR Technology and Systems, School of CSE, Beihang University \n{baoyiwei, lufeng}@buaa.edu.cn", + "bbox": [ + 127, + 180, + 841, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Annotating gaze is an expensive and time-consuming endeavor, requiring costly eye-trackers or complex geometric calibration procedures. Although some eye-based unsupervised gaze representation learning methods have been proposed, the quality of gaze representation extracted by these methods degrades severely when the head pose is large. In this paper, we present the Multi-View Dual-Encoder (MV-DE), a framework designed to learn gaze representations from unlabeled multi-view face images. Through the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy, the MV-DE successfully disentangles gaze from general facial information and derives gaze representations closely tied to the subject's eyeball rotation without gaze label. Experimental results illustrate that the gaze representations learned by the MV-DE can be used in downstream tasks, including gaze estimation and redirection. Gaze estimation results indicate that the proposed MV-DE displays notably higher robustness to uncontrolled head movements when compared to state-of-the-art (SOTA) unsupervised learning methods.", + "bbox": [ + 75, + 300, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 631, + 209, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision is one of the most important sense for humans. Human gaze reveals the direction of visual attention, which is an important cue for understanding how humans perceive the surrounding world. Thus, gaze estimation techniques have become an vital tool in numerous applications, such as Virtual Realty and Augmented Reality[2, 30, 34], automotive safety [13, 26, 27] and healthcare [3, 20, 23]. In recent years, appearance-based gaze estimation methods have drawn a lot of attention, since these methods only require simple web cameras, eliminating the need for expensive eye trackers with dedicated devices such as infrared cameras. Among these methods, Convolutional Neural Networks (CNN) based approaches exhibit exceptional performance in unconstrained environments.", + "bbox": [ + 75, + 656, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec4df738f571d5cbf031c0bc8d3afe79f18ae2ad4b8c9653e838d5bae6114280.jpg", + "image_caption": [ + "Figure 1. We propose the Multi-View Dual-Encoder, a face-based unsupervised gaze representation learning framework that is robust to large head poses." + ], + "image_footnote": [], + "bbox": [ + 503, + 268, + 890, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CNN-based gaze estimation methods are usually trained in an end-to-end manner with substantial amount of labeled data. The performance of CNN-based gaze estimation methods highly relies on the quantity and diversity of the training dataset. Unfortunately, annotating gaze directions is difficult. Unlike common Computer Vision tasks like Object Detection, gaze cannot be reliably annotated without specialized hardware. Gaze direction is either measured by costly eye trackers under controlled environment [29], or obtained by the line connecting the 3D face center to the 3D position of the gaze target [39]. It takes great effort and time to collect diverse labeled training data, since the participants need to stare at large amount of gaze targets as instructed.", + "bbox": [ + 496, + 506, + 892, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the challenge of annotation, a number of unsupervised representation learning approaches have been proposed [4, 14, 19]. These approaches extract common visual representations from the input image without annotation. However, such methods do not perform well in gaze estimation, as gaze estimation differs significantly from common visual tasks. Common visual tasks like object classification and detection require representations of overall appearance of the subject. On the other hand, gaze is an direction vector that manifested as rotations of the subtle eye structure in the image. Efforts have been made to design unsupervised gaze representation learning methods", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. This work was supported by the National Natural Science Foundation of China (NSFC) under Grant 62372019.", + "bbox": [ + 75, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1419", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[33, 35], which incorporate cropped eye images as input to eliminate undesired visual contents. However, existing unsupervised gaze representation learning methods could only handle samples within limited head pose range, since the eye appearance changes dramatically when the head rotation approaches $90^{\\circ}$ .", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent supervised gaze estimation methods usually employ face images as their input [1, 6, 8, 21, 37] to handle samples with large head poses. Unfortunately, the rich information in the face images becomes a double-edged sword in unsupervised settings. These pieces of information become interference due to the absence of gaze label as constraint. For example, head pose and gaze are strongly coupled. They behave exactly the same in common data augmentation methods such as image flipping and rotation, makes it difficult to separate them even for latest contrastive learning approaches. Thus, unsupervised gaze representation learning method with face images is a challenging task that remain to be solved.", + "bbox": [ + 75, + 183, + 467, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A potential solution to tackle this challenge is to harness the extra constraint introduced by multi-view settings. Multi-view gaze estimation is a recent hot research topic. A number of supervised multi-view gaze estimation methods [7, 24] have been proposed. Multi-view image pairs provide a set of images with the same eyeball rotation, i.e. gaze direction within the Head Coordinate System (HCS), as they are captured simultaneously. Our aim is to design a face-based multi-view unsupervised gaze representation learning method that works well on samples with large head poses.", + "bbox": [ + 75, + 380, + 467, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework that employs face images as input. The proposed MV-DE framework extracts gaze representation by separating eyeball rotation from general facial information. The MV-DE framework consists of two encoders: the Gaze Encoder and the Face Encoder. First, we train the Face Encoder to extract face representations including head pose and appearance, while excluding gaze information. Then, we freeze the Face Encoder and introduce the gaze feature extracted by the Gaze Encoder to compensate for the missing gaze information. Based on the consistent eyeball rotation across different views within the same frame, we integrate the face representations with gaze representations from different views to reconstruct the original image. At the inference time, the MV-DE extracts gaze representations from single-view face images.", + "bbox": [ + 75, + 532, + 467, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results illustrate that the MV-DE framework successfully derives gaze representations across a wide range of head poses. The learned gaze representation can be used for varies downstream tasks such as gaze estimation and gaze redirection. The contribution of this paper are summarized in three folds:", + "bbox": [ + 75, + 789, + 467, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose the Multi-View Dual-Encoder, a face-based", + "bbox": [ + 76, + 885, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unsupervised gaze representation learning framework under uncalibrated multi-view settings. The MV-DE is robust to free head movements.", + "bbox": [ + 509, + 90, + 888, + 133 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We uncouple gaze from other interference such as head pose by the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy.", + "- Extensive experiments demonstrate the effectiveness of the MV-DE framework in unconstrained environments. Qualitative analyses prove that the extracted gaze representation is disentangled from head pose and appearance." + ], + "bbox": [ + 500, + 136, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 257, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Supervised Gaze Estimation", + "text_level": 1, + "bbox": [ + 500, + 282, + 751, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Appearance-based gaze estimation approaches aim to estimate gaze from eye or face appearance directly. Early methods estimate gaze directions from eye images [9, 28, 36]. Zhang et al. first propose to utilize full face images and outperform eye-image-based methods [37]. Since then, most CNN-based gaze estimation methods employ face images as input. Chen et al. propose to utilize dilated convolution for gaze estimation [5]. Cheng et al. propose to employ the Transformer architecture for gaze estimation [6]. Some recent methods utilize both face and eye images as input for better estimation performance [1, 8, 17, 22]. Above methods are trained in an end-to-end manner with gaze annotations. A number of gaze estimation datasets are published. These datasets are collected under different scenario with varies devices, including web cameras [11, 37], $360^{\\circ}$ cameras [21], high-resolution cameras [39], eye trackers [29] and mobile devices [22]. With sufficient labeled data, gaze feature disentanglement is achieved by using GAN [10] and Nerf [32].", + "bbox": [ + 496, + 306, + 890, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-view Gaze Estimation", + "text_level": 1, + "bbox": [ + 500, + 604, + 751, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before the development of Deep Learning methods, most conventional model-based methods utilize multiple cameras for gaze estimation [16]. Model-based methods reconstruct 3D eyeball models and obtain the gaze direction based on the optical geometry of human eye structures [15]. Model-based methods achieve remarkable estimation accuracy. However, these methods have high requirements for the shooting angle of cameras. Thus, these methods are usually employed in the Head Mounted Devices, such as Meta Quest Pro and Microsoft Hololens, where the cameras are approximately stationary relative to the eye of users.", + "bbox": [ + 496, + 628, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view settings have also drawn a lot of attention in CNN-based gaze estimation methods. A few multiview gaze estimation datasets have been proposed. Zhang et al. propose the ETH-XGaze dataset, which employs 18 high-resolution cameras to capture face images with large head pose range [39]. Park et al. propose the EVE dataset [29]. They use 3 web cameras and a industrial camera to", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1420", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capture user face images. Qin et al. propose to utilize 3D face alignment approaches to generate multiple rotated images from a given sample [31]. These rotated images can be regarded as samples from different virtual cameras. Based on these datasets, CNN-based multi-view gaze estimation approaches have been proposed. These methods integrate features from different views by concatenating [24] or self attention [7] to improve estimation accuracy", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Unsupervised Gaze Representation Learning", + "text_level": 1, + "bbox": [ + 76, + 220, + 457, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised representation learning has always been a hot topic in Computer Vision community. Recently, Contrastive Learning approaches have achieved satisfying performance in common Computer Vision tasks. Contrastive Learning methods generate multiple views of a given sample by different data augmentation methods and constrain the model to extract similar representation from these views [4, 14]. He et al. propose the Masked Autoencoders, a self-supervised learner trained by simply reconstructing masked images. However, these methods are designed to extract common visual representations, which do not perform well in the gaze estimation task.", + "bbox": [ + 75, + 243, + 468, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn gaze representations in an unsupervised manner, Yu et al. propose to utilize the gaze redirection task to extract two-dimensional representations that relate to gaze pitch and yaw angles [35]. Sun et al. propose the CrossEncoder, which learn gaze representations by a latent-codeswapping mechanism on eye-consistent image pairs and gaze-similar pairs [33]. Gideon et al. further propose to adapt the Cross-Encoder architecture to the multi-view setting [12]. Above methods all employ eye images as their input, as there are less gaze-irrelevant visual contents. However, appearance of eyes changes dramatically as head pose increases, makes it difficult for these methods to handle samples with large head pose distribution. Our aim is to introduce an unsupervised gaze representation learning approach that employs face images as input, which is robust to unconstrained head movements.", + "bbox": [ + 75, + 425, + 470, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 679, + 168, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Multi-View Dual-Encoder, a face-based multi-view unsupervised gaze representation learning method. Our aim is to extract gaze representations from the input face images without using the gaze label. The main challenge of this task is to separate gaze from the other facial information, especially head pose. Since gaze and head pose are both physical directions, common data augmentations like rotation and flipping are not able to separate them.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve our goal, the proposed MV-DE framework introduces two encoders: a Face Encoder and a Gaze Encoder. We first train the Face Encoder to extract general facial representation including head pose and appearance, while excluding gaze information. Then, we freeze the Face", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1276427211a33707343175ed36949157e06ea13cd622fc3e9857089f8339e46c.jpg", + "image_caption": [ + "Figure 2. Training process of the Face Encoder. The proposed Face Encoder extracts face representations that contain general facial information while excluding gaze information. It is trained by two tasks: the head pose estimation task and the image reconstruction task." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 883, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder and train the Gaze Encoder to extract gaze representation which compensates for the missing gaze information. We constrain the Gaze Encoder to derive similar representation from different views within a given frame since the eyeball rotation is the same. In this way, we isolate gaze representation from other facial features without using the gaze label. In the following sections, we introduce the training pipeline of two encoders in detail.", + "bbox": [ + 496, + 338, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Training of the Face Encoder", + "text_level": 1, + "bbox": [ + 498, + 468, + 759, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The target of the Face Encoder is to extract representation of general facial information except gaze. We design two training strategies to achieve this goal: eye masking and multi-task learning including head pose estimation and image reconstruction, as shown in Fig. 2.", + "bbox": [ + 496, + 492, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a training face image $\\boldsymbol{x}_{i,j}$ where $i$ is the frame index and $j$ is the camera index, we mask the two eye areas by the average pixel value. We further add three random masks to prevent the Face Encoder from directly estimating head pose based on the position of eye masks. Given that input face images are normalized to $224 \\times 224$ pixels, we set the size of masks to $55 \\times 33$ according to the average size of eyes. Then, we input the masked face images to the Face Encoder for the proposed multi-task learning.", + "bbox": [ + 496, + 568, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, the Face Encoder extracts the face representation $e_{i,j}^{f}$ from the masked face image. Then, we add a regression MLP and a decoder for the head pose estimation task and the image reconstruction task, respectively. In the head pose estimation task, the target of the Face Encoder is to estimate the 3D Euler Angles of the subject's head pose. We use $\\mathcal{L}_1$ Loss function for the head pose estimation task: $\\mathcal{L}_{hp} = \\mathcal{L}_1(\\hat{\\boldsymbol{y}}_{i,j},\\boldsymbol{y}_{i,j})$ , where $\\hat{\\boldsymbol{y}}_{i,j}$ is the estimated head pose and $\\boldsymbol{y}_{i,j}$ is the ground truth head pose. The purpose of the head pose estimation task is to ensure that the face representation $e_{i,j}^{f}$ encodes head pose information.", + "bbox": [ + 496, + 704, + 890, + 871 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the image reconstruction task, the Face Encoder and the decoder are trained in an adversarial way. The target", + "bbox": [ + 498, + 871, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1421", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a5d3fd66b99129d7b31e54f96247b820f52d3d893c6857a34a67454ca0c58b0.jpg", + "image_caption": [ + "Figure 3. Training process of the Gaze Encoder. The Gaze Encoder extracts gaze representations that represent the eyeball rotation, i.e. gaze direction within HCS. The Gaze Encoder is trained by the proposed multi-view gaze representation swapping strategy." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 883, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the Face Encoder is to reconstruct the image with eyemasks. The target of the decoder is to reconstruct the original image from the face representation $e_{i,j}^{f}$ , including the eye area. Both the Face Encoder and the decoder are constrained by the Image Reconstruction Loss $\\mathcal{L}_{re}$ :", + "bbox": [ + 75, + 340, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) = \\alpha_ {1} \\mathcal {L} _ {2} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) + \\alpha_ {2} \\mathcal {L} _ {p c p t} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 429, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{x}}$ is the reconstructed image, $\\pmb{x}$ is the target image, $\\mathcal{L}_2$ is the Mean Squared Error (MSE) Loss function and $\\mathcal{L}_{pcpt}$ is the Perception Loss function. $\\alpha_{1},\\alpha_{2}$ are the coefficients of loss functions. The purpose of the image reconstruction tasks is to encode general facial information in the face representation while excluding gaze information. Images constructed from the $e_{i,j}^{f}$ are basically the original face images without iris and pupil.", + "bbox": [ + 75, + 455, + 468, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, we train the Face Encoder by minimizing $\\mathcal{L}_{FE} = \\beta_1\\mathcal{L}_{hp} + \\mathcal{L}_{re}$ , where $\\beta_{1}$ is coefficient to balance two tasks. Once the training of the Face Encoder is completed, we freeze the parameters of the Face Encoder and proceed to train the Gaze Encoder via the multi-view constraint.", + "bbox": [ + 75, + 578, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training of the Gaze Encoder", + "text_level": 1, + "bbox": [ + 76, + 662, + 341, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The target of the Gaze Encoder is to extract gaze representation $e_{i,j}^{g}$ that represents the direction of gaze in the HCS without any gaze label. Since the face representation $e_{i,j}^{f}$ encodes general face information except gaze, we combine the gaze representation $e_{i,j}^{g}$ with $e_{i,j}^{f}$ to fully reconstruct the original image, so that $e_{i,j}^{g}$ compensates for the missing gaze information. The training strategy of the Gaze Encoder is shown in Fig. 3.", + "bbox": [ + 75, + 685, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a training sample $\\pmb{x}_{i,j}$ , we extract the gaze representation from the original image and the face representation from the eye-masked image. Then, we input both representations to the decoder to reconstruct the original image. To reconstruct the original image, the Gaze Encoder has to capture gaze information which is missing in the face repre", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentation. Considering that the region occupied by two eyes represents only a small portion of the face image, we additionally compute the Image Reconstruction Loss for both eye regions:", + "bbox": [ + 498, + 340, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {i, j} = G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j} ^ {g}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 414, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G E} = \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}) + \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}),\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 438, + 831, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(\\cdot)$ is the decoder, $\\hat{\\pmb{x}}_{i,j}^{e}$ and $\\pmb{x}_{i,j}^{e}$ are the two eye regions of the reconstructed image and the original image, respectively.", + "bbox": [ + 498, + 467, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the gaze representation can also encode other facial information such as head pose during the training process. In an extreme case, the decoder might theoretically reconstruct the entire image solely based on the gaze representation, since the Gaze Encoder utilizes the full face image as input. We utilize the consistency of eyeball rotation in the multi-view settings to exclude gaze-irrelevant information. First, we randomly sample an image $\\pmb{x}_{i,j'}$ from another view within the same frame where $j' \\neq j$ . Since the eyeball rotations from both views are consistent, we swap the gaze representations of two views and reconstruct the eye regions of the original views. For a training image pair $\\{\\pmb{x}_{i,j}, \\pmb{x}_{i,j'}\\}$ , the final loss function of the Gaze Encoder $\\mathcal{L}_{GE}$ is:", + "bbox": [ + 498, + 513, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G E} = \\beta_ {2} \\mathcal {L} _ {r e} ^ {\\text {f a c e}} + \\beta_ {3} \\mathcal {L} _ {r e} ^ {\\text {e y e s}} + \\beta_ {4} \\mathcal {L} _ {r e} ^ {\\text {s w a p}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 737, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{2},\\beta_{3},\\beta_{4}$ are the coefficients, $\\mathcal{L}_{re}^{face}$ is the Image Reconstruction Loss of full face images from both view:", + "bbox": [ + 500, + 767, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} ^ {\\text {f a c e}} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}}, \\boldsymbol {x} _ {i, j ^ {\\prime}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 810, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{re}^{eyes}$ is the Image Reconstruction Loss of eye regions from both view:", + "bbox": [ + 500, + 840, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} ^ {e y s} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}} ^ {e}, \\boldsymbol {x} _ {i, j ^ {\\prime}} ^ {e}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1422", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{re}^{swap}$ is the Image Reconstruction Loss of the generated eye regions after gaze representation swapping:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} ^ {s w a p} = \\mathcal {L} _ {r e} \\left(G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j ^ {\\prime}} ^ {g}\\right) ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 127, + 468, + 155 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} (G (\\pmb {e} _ {i, j ^ {\\prime}} ^ {f}, \\pmb {e} _ {i, j} ^ {g}) ^ {e}, \\pmb {x} _ {i, j ^ {\\prime}} ^ {e}).\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 151, + 388, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At test time, the Gaze Encoder extracts gaze representations from the given single-view face images, similar to common supervised single-view gaze estimation methods. We rotate the estimated HCS gaze to CCS by the head pose label for evaluation.", + "bbox": [ + 76, + 176, + 468, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 258, + 294, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details: The proposed method is implemented by PyTorch using two RTX 3090 GPU. We employ the Adam optimizer with a leaning rate of $10^{-3}$ for all the encoders and decoders. The Face Encoder is trained for 13 epochs with a batch size of 100. Learning rate is decayed by 0.2 every 4 epochs. The Gaze Encoder is trained for 5 epochs with a batch size of 50. Learning rate is decayed by 0.2 every 2 epochs. We use the data rectification method from [37] and histogram equalization to normalize the input face images. $(\\alpha_{1},\\alpha_{2})$ are set to (10,1) and (10,0.5) during the training of the Face Encoder and Gaze Encoder, respectively. $\\beta_{1}$ is set to 0.1 and $(\\beta_{2},\\beta_{3},\\beta_{4})$ is set to (0.3,0.35,0.35). We use the first 6 convolutional layers of a ImageNet pretrained VGG-16 to calculate the Perceptual Loss.", + "bbox": [ + 75, + 281, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture: Both the Face Encoder and the Gaze Encoder use the ResNet-18 [18] as backbone. For the Face Encoder, we take the $(512*7*7)$ feature map after the last residual block as the Face Representation. The output channels of the last Linear Layer is set to 3 for head pose estimation. For the Gaze Encoder, the Linear layers is replaced by a $1\\times 1$ Convolutionallayer to compress the $(512*7*7)$ feature map to $(1*7*7)$ gaze representation. In the decoder, the channel of the gaze representation is expanded to 10 by a $1\\times 1$ Convolutionallayer and then concatenated with the face representation. The decoder also has four residual blocks with target channel sizes of [256, 128, 63, 32]. The feature maps are up-scaled to twice their size before each residual block. We upscale the feature map again and use two Convolutional layers to generate the final $(3\\times 224\\times 224)$ image. During the training of the Face Encoder, we set the gaze representation to zeros. Note that we train the decoder from scratch during the training of the Gaze Encoder, as the decoder has learned to ignore the 0 replaced gaze representation during the training of the Face Encoder.", + "bbox": [ + 75, + 493, + 470, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data Preparation", + "text_level": 1, + "bbox": [ + 76, + 847, + 246, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on four different gaze estimation datasets: ETH-XGaze [39], MPII-NV [31], EVE [29] and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ad9c9523511db8eea09f9f50b7b468a82d042fa3c2be6f35c983209b12e2bec.jpg", + "image_caption": [ + "Figure 4. Head pose (top row) and gaze direction distributions (bottom row) of 4 different datasets. The EVE and the EyeDiap datasets only provide very limited head pose and gaze range with frontal faces." + ], + "image_footnote": [], + "bbox": [ + 511, + 89, + 883, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8229af6ef1af697d193bafbe8f5648567464d8f4e11356e5358083355d725502.jpg", + "image_caption": [ + "Figure 5. Image reconstruction results of combining face representations with different gaze representations in the ETH-XGaze dataset." + ], + "image_footnote": [], + "bbox": [ + 509, + 359, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EyeDiap [11]. We visualize the head pose and gaze distribution of each dataset in Fig. 4. The distribution of these datasets varies significantly. below we introduce each dataset in detail.", + "bbox": [ + 496, + 643, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ETH-XGaze $(\\mathcal{D}_E)$ : A multi-view dataset collected by 18 high-resolution cameras. It provides very large head pose and gaze distributions. ETH-XGaze dataset contains 80 subjects, results in over 750,000 images in total. We divide the last 5 subjects as the labled test set and the first 75 subjects as the multi-view unsupervised training set. We follow [25] to optimize the gaze and head pose label to ensure multi-view consistency.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MPII-NV $(\\mathcal{D}_M)$ : A synthesized multi-view dataset. We follow [31] to reconstruct 3D faces from MPIIFaceGaze [37] dataset and rotate every 3D face to generate 18 different views, referring to the setting of ETH-XGaze. MPII-NV dataset contains 15 subjects and 359,984 images in total.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1423", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5735b2ddf355088be6d462f6580ee02f1d68692edd0ab3380d4af5c9c0998d65.jpg", + "table_caption": [ + "Table 1. Gaze estimation error of few-shot experiments in degrees. Bold numbers are the best results and underline numbers are the second best results." + ], + "table_footnote": [], + "table_body": "
DatasetETH-XGazeMPII-NVEVEAverage
Calibration Num50100200Avg.50100200Avg.50100200Avg.
BYOL[14]25.6726.0425.2925.6710.1912.6511.2611.3710.369.749.749.9515.66
SimCLR[4]26.8124.9823.8625.2212.5512.1312.1412.2712.5311.3611.4811.7916.43
Cross-Encoder[33]16.2314.7214.5415.1611.4711.3611.7511.538.077.817.487.7911.49
DE (ours w/o MV)23.3219.9317.7620.3411.6410.7310.7411.0410.799.739.449.9913.79
MV-DE (ours)8.668.087.778.177.286.526.396.739.188.68.688.827.91
Gain▼ 7.57▼ 6.64▼ 6.77▼ 6.99▼ 2.91▼ 4.21▼ 4.35▼ 4.31▲ 1.11▲ 0.79▲ 1.20▲ 1.03▼ 3.58
", + "bbox": [ + 78, + 121, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f255b2f83446306ebc5c21857e12ddde86ec889cfa975e5b136ff728a1c57ed.jpg", + "image_caption": [ + "Figure 6. Gaze estimation error relative to the gaze yaw angle in the ETH-XGaze dataset. The proposed MV-DE performs significantly better when the gaze yaw angle is large." + ], + "image_footnote": [], + "bbox": [ + 86, + 270, + 460, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the last subject as the labeled test set.", + "bbox": [ + 76, + 520, + 375, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EVE $(\\mathcal{D}_V)$ : A multi-view dataset with 4 camera views. As shown in the Fig. 4, the EVE dataset only provides frontal faces with limited head pose and gaze distribution. We use the training and testing set split as defined in the original dataset and sample 3 images per second from the original videos, result in 527,896 images for unsupervised training and 64,464 images for testing.", + "bbox": [ + 75, + 536, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EyeDiap $(\\mathcal{D}_D)$ : A single-view dataset with limited head pose and gaze distribution with 16,674 images in total. We only use the EyeDiap in the cross-domain experiments. Thus, we use the whole dataset for training or testing.", + "bbox": [ + 75, + 642, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow [38] to normalize the face images. We employ Histogram Equalization and normalize the pixel values to $[-1, 1]$ .", + "bbox": [ + 75, + 702, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation of the Learned Representations", + "text_level": 1, + "bbox": [ + 76, + 756, + 442, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness of the learned face and gaze representations through image reconstruction. The propose MV-DE is supposed to decouple representations of the eyeball rotation from general facial information. In Fig. 5, we combine the face representation with different gaze representations and reconstruct the face image for validation. As shown in row 2, the decoder effectively reconstructs the input face image from the original face and", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gaze representation. When replacing the original gaze representation with gaze representation from another view, the reconstruction results are similar with the original one. It proves that the MV-DE successfully learned the multi-view consistency of the eyeball rotation. In the last row, the gaze representation is replaced by zero vectors. The head pose and appearance of the reconstructed images are almost identical with the original image, while the eyeball rotations are rather random. These results confirm that the MV-DE successfully separate eyeball rotation from general facial information.", + "bbox": [ + 496, + 273, + 893, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Application of the MV-DE: Gaze Estimation", + "text_level": 1, + "bbox": [ + 498, + 452, + 875, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The learned gaze representation can be used for gaze estimation by adding a MLP head calibrated under the few-shot learning setting. The estimation error is also an evaluation of the learned gaze representation. In the MV-DE, the MLP takes the gaze representation and the head pose estimation as input. The estimated head pose is first encoded to a 30 dimensional embedding and then concatenated with the gaze representation through 3 linear layers with target dimensions of (64, 64, 2). The MLP head is trained to estimate the HCS gaze. Then, we rotate the HCS gaze by head rotation to obtain CCS gaze. Sigmoid function is employed in the activation layer.", + "bbox": [ + 496, + 476, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare the MV-DE with 3 SOTA unsupervised learning methods. BYOL [14] and SimCLR [4] are two SOTA contrastive learning methods. We modify them to adapt the multi-view gaze representation learning task. We employ face images from different camera views in the same frame instead of data augmentation methods to generate the positive pairs. Cross-Encoder [33] is a SOTA eye-based unsupervised gaze representation learning method. Gideon et al. also propose an eye-based method [12], but their method requires sample pairs with different eyeball rotation while keeping the head stable within a short video clip. Since such sample pairs are not available in $\\mathcal{D}_E$ and $\\mathcal{D}_M$ , we exclude their moethod in comparison. We also remove the multi-view constraint in the MV-DE as an ablation study (named DE). Without the multi-view constraint, the Gaze Encoder is trained by the image reconstruc", + "bbox": [ + 496, + 657, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1424", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e32304ba061a793fc9da458d60918a11dc1bba4f2a5d00d548343d62e068f6be.jpg", + "table_caption": [ + "Table 2. Estimation error of different designs of the MLP head with 200 calibration samples." + ], + "table_footnote": [], + "table_body": "
DimensionMLP inputs & outputsD_ED_MD_V
d=32eg→CCS Gaze24.3442.829.83
eg+HP→CCS Gaze20.5942.299.57
eg→HCS Gaze8.8910.498.66
eg+HP→HCS Gaze8.637.268.60
d=64eg→CCS Gaze22.2741.709.98
eg+HP→CCS Gaze18.1134.569.55
eg→HCS Gaze8.437.448.68
eg+HP→HCS Gaze7.846.398.67
d=256eg→CCS Gaze20.9740.9910.32
eg+HP→CCS Gaze18.8834.0510.39
eg→HCS Gaze8.076.508.72
eg+HP→HCS Gaze7.595.718.87
", + "bbox": [ + 81, + 128, + 468, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion task described in Eq. (2).", + "bbox": [ + 76, + 339, + 274, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the average accuracy of the MV-DE outperforms other SOTA methods significantly. BYOL and SimCLR perform worse than the gaze-specialized CrossEncoder and the MV-DE, proves that common visual representations are not suitable for the gaze estimation task.", + "bbox": [ + 75, + 354, + 468, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Most importantly, the MV-DE outperforms the CrossEncoder significantly in the ETH-XGaze and MPII-NV datasets, where the head pose distributes across a wide range. On the other hand, the MV-DE achieves comparable performance with the Cross-Encoder in the EVE dataset, because the EVE dataset only contains frontal face images with limited head pose range.", + "bbox": [ + 75, + 431, + 468, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further investigate the performance gap between the Cross-Encoder and the MV-DE, we visualize the gaze estimation error relative to the gaze yaw angle within the ETH-XGaze dataset in Fig. 6. Obviously, the estimation error of the Cross-Encoder increases severely when the gaze yaw angle approach $100^{\\circ}$ . Samples with such extreme gaze yaw angle are generally associated with large head poses. Above results prove the point that the MV-DE are more robust to large head poses.", + "bbox": [ + 75, + 537, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results from the last 2 rows demonstrate the importance of the proposed multi-view gaze representation swapping strategy. The multi-view constraint brought a improvement as large as $5.88^{\\circ}$ . But the Dual-Encoder (DE) still achieves better overall performance than SOTA constrastive learning methods, primary attributes to the proposed dual-encoder architecture.", + "bbox": [ + 75, + 672, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Ablation Study: Design of the MLP Head", + "text_level": 1, + "bbox": [ + 76, + 800, + 419, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The MLP head is responsible for estimating gaze from the learned gaze representations under the few-shot learning setting. In Tab. 2, we verify the performances of different MLP designs with 200 calibration samples. For example, $[d = 64, e_g + HP \\rightarrow HCSGaze]$ denotes that the MLP", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c15153e5600cd6639e8175a779be45bb5c052c3e96264685bf62082c7fc3940a.jpg", + "table_caption": [ + "Table 3. Cross-dataset gaze estimation errors of using the MV-DE as an unsupervised pretrain method. We first pretrain the model on the ETH-XGaze or the MPII-NV dataset unsupervisely. Then, we use the labeled training dataset to calibrate the MLP head." + ], + "table_footnote": [], + "table_body": "
MethodUnsupervisedTrainTest Dataset
PretrainDatasetETH-XgazeMPII-NVEyeDiap
Baseline-EVE31.2732.5212.55
MV-DEETH-XGazeEVE19.67 ▼ 37.10%10.45▼ 67.87%11.36 ▼ 9.47%
MV-DEMPII-NVEVE25.42 ▼ 18.71%10.33 ▼ 68.23%11.58 ▼ 7.73%
ETH-XgazeMPII-NVEVE
Baseline-EyeDiap44.0639.4623.58
MV-DEETH-XGazeEyeDiap25.15 ▼ 42.92%12.11 ▼ 69.31%10.4 ▼ 55.89%
MV-DEMPII-NVEyeDiap25.58 ▼ 41.94%11.51 ▼ 70.83%10.58 ▼ 55.13%
", + "bbox": [ + 503, + 170, + 893, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "includes 64-dimensional hidden layers, takes gaze representations and head poses as the input vector, and predicts gaze directions within Head Coordinate System (HCS). Note that in the MV-DE, we rotate the estimated HCS gaze with head pose label physically.", + "bbox": [ + 496, + 309, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Tab. 2 lead to 3 conclusions: (1) The MLP performs significantly better on predicting HCS gaze, since the learned gaze representations represent eyeball rotation instead of the CCS Gaze. Theoretically, combination of $e_{g} + HP$ should be enough to predict the CCS gaze, but the estimation error is still huge, probably due to the limited number of calibration samples and MLP parameters. (2) Head pose seems helpful for predicting the HCS gaze. This observation is consistent with previous research [37] that the additional information of the full face images helps in the gaze estimation task. (3) 256-dimensional hidden layers achieve the best performance in the 200-shot settings. But larger number of parameters require more calibration samples, the performance of 256-dimensional hidden layers degrades in 50 and 100 shot calibration settings. Considering the balance between the number of parameters and required calibration samples, we choose $d = 64$ for the MV-DE.", + "bbox": [ + 496, + 385, + 892, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Additional Experiments: MV-DE as an Unsupervised Pretrain Method", + "text_level": 1, + "bbox": [ + 498, + 662, + 890, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another potential application of the MV-DE is to use as an unsupervised pretrain method, since it is more cost-effective to collect large number of unlabeled face images than labeled ones. In Tab. 3, we employ the MV-DE to pretrain the Gaze Encoder without gaze label in the ETH-XGaze or the MPII-NV datasets. Then, we freeze the Gaze Encoder and only train the MLP head in the EVE or the EyeDiap dataset with gaze label.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To establish a baseline, we directly train the Gaze Encoder and the regression MLP in an end-to-end manner with gaze label in the labeled training dataset, just like common supervised gaze estimation methods. The cross-dataset errors of the baseline method are extremely large on the ETH-", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1425", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52e4c07bb8017484a4de72094b875728ce102f1c8a14428f8d349d34544887ca.jpg", + "image_caption": [ + "Figure 7. Distributions of cross-domain gaze estimation error with (right) and without (left) the proposed unsupervised pretrain on the ETH-XGaze. Blue rectangle indicates the gaze distribution range of the labeled training dataset (EyeDiap)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b53fd10394894cae9d6784fb2cce1ab00c90d5e5bfd0eec807b183b3723681e7.jpg", + "image_caption": [ + "Figure 8. Results of using the MV-DE for gaze redirection. Spheres on the bottom-right corner indicates the gaze direction ground truth within the HCS, i.e. the eyeball rotation." + ], + "image_footnote": [], + "bbox": [ + 94, + 335, + 457, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "XGaze and MPII-NV dataset, since the gaze distributions of the training sets are significantly smaller than the test sets. When used as a unsupervised pretrain method, the MV-DE effectively improves the cross-dataset performance of the baseline model. The distributions of estimation errors in Fig. 7 demonstrate that with the pretrain of the MV-DE, the model generates reasonable estimations in a significantly larger range. Above results indicate that the MV-DE can be used as a pretrain method to improve cross-dataset performances when large amount of unlabeled face images are available.", + "bbox": [ + 75, + 650, + 472, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Application of the MV-DE: Gaze Redirection", + "text_level": 1, + "bbox": [ + 76, + 830, + 459, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The learned representations of the MV-DE can also be used for the gaze redirection task. Results in Fig. 8 show that the MV-DE effectively redirects the gaze direction accord", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/181f90429d1f451985d8b251c349af38e4053dc4f1e5ec651c1171466b09f0b2.jpg", + "image_caption": [ + "Figure 9. Gaze redirection results of combining different gaze representations linearly. Results demonstrate that the gaze representations exhibit good linearity invariance. Spheres on the bottom-right corner indicate the ground truth or the target eyeball rotation." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing to the HCS gaze of the reference images, while keeping the head pose and identity unchanged. Fig. 8 demonstrate that the learned gaze representation is subject-independent. The head pose and identity of the reconstructed images are controlled by the face representations, where the eyeabll rotation is controlled by the gaze representations as expected.", + "bbox": [ + 498, + 330, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1 Linearity Invariance of the Gaze Representation", + "text_level": 1, + "bbox": [ + 498, + 431, + 888, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gaze, as a physical direction vector, conforms to the principles of additivity. Ideally, well-learned gaze representations should possess similar characteristics. In Fig. 9, we combine gaze representations from different samples linearly with different ratios to reconstruct face images. The sphere in the bottom-left corner of reconstructed the images indicates the target eyeball rotation angle, which corresponds to the linear combination of eyeball rotations from the HCS gaze source images. Results demonstrate that the learned gaze representation satisfies Linearity Invariance, i.e. the eyeball rotation of reconstructed images undergo the same linear combination as the input gaze representations. Experiments in Fig. 9 prove that the MV-DE successfully captures and retains part of the physical properties of gaze during the unsupervised learning process.", + "bbox": [ + 496, + 455, + 892, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 695, + 619, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework based on multi-view face images. We propose the Dual-Encoder architecture and the multiview gaze representation swapping strategy to learn gaze representations that are separated from general facial information without gaze label. Experiments show that the learned gaze representations can be used for downstream tasks like gaze estimation and gaze redirection. Gaze estimation results show that the quality of learned gaze representations from the MV-DE are significantly better than other SOTA methods with unconstrained head movements.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1426", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yiwei Bao, Yihua Cheng, Yunfei Liu, and Feng Lu. Adaptive feature fusion network for gaze tracking in mobile tablets. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9936-9943. IEEE, 2021. 2", + "[2] Yiwei Bao, Jiaxi Wang, Zhimin Wang, and Feng Lu. Exploring 3d interaction with gaze guidance in augmented reality. In 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 22-32. IEEE, 2023. 1", + "[3] Ariella Fornachari Ribeiro Belan, Marcos Vasconcelos Pais, Marina von Zuben de Arruda Camargo, Livea Carla Fidalgo Garcez Sant'Ana, Marcia Radanovic, and Orestes Vicente Forlenza. Diagnostic performance of an eye-tracking assisted visual inference language test in the assessment of cognitive decline due to alzheimer's disease. Journal of Alzheimer's Disease, (Preprint):1-15, 2023. 1", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6", + "[5] Zhaokang Chen and Bertram E Shi. Appearance-based gaze estimation using dilated-convolutions. In Asian Conference on Computer Vision, pages 309–324. Springer, 2018. 2", + "[6] Yihua Cheng and Feng Lu. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3341-3347. IEEE, 2022. 2", + "[7] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20632-20641, 2023. 2, 3", + "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10623-10630, 2020. 2", + "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2", + "[10] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 436-443, 2022. 2", + "[11] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 255–258, 2014. 2, 5", + "[12] John Gideon, Shan Su, and Simon Stent. Unsupervised multi-view gaze representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5001-5009, 2022. 3, 6", + "[13] Mohamed Waheed Gomaa, Rasha O Mahmoud, and Amany M Sarhan. A cnn-lstm-based deep learning approach for driver drowsiness prediction. Journal of Engineering Research, 6(3):59-70, 2022. 1", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 3, 6", + "[15] Elias Daniel Guestrin and Moshe Eizenman. General theory of remote gaze estimation using the pupil center and corneal reflections. IEEE Transactions on biomedical engineering, 53(6):1124-1133, 2006. 2", + "[16] Dan Witzner Hansen and Qiang Ji. In the eye of the beholder: A survey of models for eyes and gaze. IEEE transactions on pattern analysis and machine intelligence, 32(3):478-500, 2009. 2", + "[17] Junfeng He, Khoi Pham, Nachiappan Valliappan, Pingmei Xu, Chase Roberts, Dmitry Lagun, and Vidhya Navalpakkam. On-device few-shot personalization for real-time gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1", + "[20] Daniel A Hofmaenner and Philipp K Buehler. The impact of eye-tracking on patient safety in critical care. Journal of Clinical Monitoring and Computing, 36(6):1577-1579, 2022. 1", + "[21] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2", + "[22] Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhandarkar, Wojciech Matusik, and Antonio Torralba. Eye tracking for everyone. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2176-2184, 2016. 2", + "[23] Astar Lev, Yoram Braw, Tomer Elbaum, Michael Wagner, and Yuri Rassovsky. Eye tracking during a continuous performance test: Utility for assessing ADHD patients. Journal of Attention Disorders, 26(2):245-255, 2022. 1", + "[24] Dongze Lian, Lina Hu, Weixin Luo, Yanyu Xu, Lixin Duan, Jingyi Yu, and Shenghua Gao. Multiview multitask gaze estimation with deep convolutional neural networks. IEEE transactions on neural networks and learning systems, 30 (10):3010-3023, 2018. 2, 3", + "[25] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. Unpublished manuscript. 5", + "[26] SI Lyapunov, II Shoshina, and IS Lyapunov. Tremor eye movements as an objective marker of driver's fatigue. Human Physiology, 48(1):71-77, 2022. 1", + "[27] Nada Mohammed Murad, Lilia Rejeb, and Lamjed Ben Said. Computing driver tiredness and fatigue in automobile via eye" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1427", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracking and body movements. Periodicals of Engineering and Natural Sciences, 10(1):573-586, 2022. 1", + "[28] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In Proceedings of the European conference on computer vision (ECCV), pages 721-738, 2018. 2", + "[29] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5", + "[30] Thammathip Piumsomboon, Gun Lee, Robert W Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In 2017 IEEE symposium on 3D user interfaces (3DUI), pages 36-39. IEEE, 2017. 1", + "[31] Jiawei Qin, Takuru Shimoyama, and Yusuke Sugano. Learning-by-novel-view-synthesis for full-face appearance-based 3d gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2022. 3, 5", + "[32] Alessandro Ruzzi, Xiangwei Shi, Xi Wang, Gengyan Li, Shalini De Mello, Hyung Jin Chang, Xuong Zhang, and Otmar Hilliges. Gazenerf: 3d-aware gaze redirection with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9676-9685, 2023. 2", + "[33] Yunjia Sun, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Cross-encoder for unsupervised gaze representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3702-3711, 2021. 2, 3, 6", + "[34] Xin Yi, Yiqin Lu, Ziyin Cai, Zihan Wu, Yuntao Wang, and Yuanchun Shi. Gazedock: Gaze-only menu selection in virtual reality using auto-triggering peripheral menu. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 832-842. IEEE, 2022. 1", + "[35] Yu Yu and Jean-Marc Odobez. Unsupervised representation learning for gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7314-7324, 2020. 2, 3", + "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4511-4520, 2015. 2", + "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-60, 2017. 2, 5, 7", + "[38] Xuong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM symposium on eye tracking research & applications, pages 1-9, 2018. 6", + "[39] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In European Conference on Computer Vision, pages 365-381. Springer, 2020. 1, 2, 5" + ], + "bbox": [ + 78, + 92, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1428", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_model.json b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9049cc6cff6e1e4e55bd9c1e8030ad50542ffd9e --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "Unsupervised Gaze Representation Learning from Multi-view Face Images" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.181, + 0.842, + 0.234 + ], + "angle": 0, + "content": "Yiwei Bao Feng Lu* \nState Key Laboratory of VR Technology and Systems, School of CSE, Beihang University \n{baoyiwei, lufeng}@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Annotating gaze is an expensive and time-consuming endeavor, requiring costly eye-trackers or complex geometric calibration procedures. Although some eye-based unsupervised gaze representation learning methods have been proposed, the quality of gaze representation extracted by these methods degrades severely when the head pose is large. In this paper, we present the Multi-View Dual-Encoder (MV-DE), a framework designed to learn gaze representations from unlabeled multi-view face images. Through the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy, the MV-DE successfully disentangles gaze from general facial information and derives gaze representations closely tied to the subject's eyeball rotation without gaze label. Experimental results illustrate that the gaze representations learned by the MV-DE can be used in downstream tasks, including gaze estimation and redirection. Gaze estimation results indicate that the proposed MV-DE displays notably higher robustness to uncontrolled head movements when compared to state-of-the-art (SOTA) unsupervised learning methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.632, + 0.21, + 0.647 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Vision is one of the most important sense for humans. Human gaze reveals the direction of visual attention, which is an important cue for understanding how humans perceive the surrounding world. Thus, gaze estimation techniques have become an vital tool in numerous applications, such as Virtual Realty and Augmented Reality[2, 30, 34], automotive safety [13, 26, 27] and healthcare [3, 20, 23]. In recent years, appearance-based gaze estimation methods have drawn a lot of attention, since these methods only require simple web cameras, eliminating the need for expensive eye trackers with dedicated devices such as infrared cameras. Among these methods, Convolutional Neural Networks (CNN) based approaches exhibit exceptional performance in unconstrained environments." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.27, + 0.891, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.443, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 1. We propose the Multi-View Dual-Encoder, a face-based unsupervised gaze representation learning framework that is robust to large head poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.718 + ], + "angle": 0, + "content": "CNN-based gaze estimation methods are usually trained in an end-to-end manner with substantial amount of labeled data. The performance of CNN-based gaze estimation methods highly relies on the quantity and diversity of the training dataset. Unfortunately, annotating gaze directions is difficult. Unlike common Computer Vision tasks like Object Detection, gaze cannot be reliably annotated without specialized hardware. Gaze direction is either measured by costly eye trackers under controlled environment [29], or obtained by the line connecting the 3D face center to the 3D position of the gaze target [39]. It takes great effort and time to collect diverse labeled training data, since the participants need to stare at large amount of gaze targets as instructed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address the challenge of annotation, a number of unsupervised representation learning approaches have been proposed [4, 14, 19]. These approaches extract common visual representations from the input image without annotation. However, such methods do not perform well in gaze estimation, as gaze estimation differs significantly from common visual tasks. Common visual tasks like object classification and detection require representations of overall appearance of the subject. On the other hand, gaze is an direction vector that manifested as rotations of the subtle eye structure in the image. Efforts have been made to design unsupervised gaze representation learning methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. This work was supported by the National Natural Science Foundation of China (NSFC) under Grant 62372019." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "[33, 35], which incorporate cropped eye images as input to eliminate undesired visual contents. However, existing unsupervised gaze representation learning methods could only handle samples within limited head pose range, since the eye appearance changes dramatically when the head rotation approaches \\(90^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.379 + ], + "angle": 0, + "content": "Recent supervised gaze estimation methods usually employ face images as their input [1, 6, 8, 21, 37] to handle samples with large head poses. Unfortunately, the rich information in the face images becomes a double-edged sword in unsupervised settings. These pieces of information become interference due to the absence of gaze label as constraint. For example, head pose and gaze are strongly coupled. They behave exactly the same in common data augmentation methods such as image flipping and rotation, makes it difficult to separate them even for latest contrastive learning approaches. Thus, unsupervised gaze representation learning method with face images is a challenging task that remain to be solved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.468, + 0.532 + ], + "angle": 0, + "content": "A potential solution to tackle this challenge is to harness the extra constraint introduced by multi-view settings. Multi-view gaze estimation is a recent hot research topic. A number of supervised multi-view gaze estimation methods [7, 24] have been proposed. Multi-view image pairs provide a set of images with the same eyeball rotation, i.e. gaze direction within the Head Coordinate System (HCS), as they are captured simultaneously. Our aim is to design a face-based multi-view unsupervised gaze representation learning method that works well on samples with large head poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.468, + 0.789 + ], + "angle": 0, + "content": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework that employs face images as input. The proposed MV-DE framework extracts gaze representation by separating eyeball rotation from general facial information. The MV-DE framework consists of two encoders: the Gaze Encoder and the Face Encoder. First, we train the Face Encoder to extract face representations including head pose and appearance, while excluding gaze information. Then, we freeze the Face Encoder and introduce the gaze feature extracted by the Gaze Encoder to compensate for the missing gaze information. Based on the consistent eyeball rotation across different views within the same frame, we integrate the face representations with gaze representations from different views to reconstruct the original image. At the inference time, the MV-DE extracts gaze representations from single-view face images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.468, + 0.88 + ], + "angle": 0, + "content": "Experimental results illustrate that the MV-DE framework successfully derives gaze representations across a wide range of head poses. The learned gaze representation can be used for varies downstream tasks such as gaze estimation and gaze redirection. The contribution of this paper are summarized in three folds:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We propose the Multi-View Dual-Encoder, a face-based" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.89, + 0.135 + ], + "angle": 0, + "content": "unsupervised gaze representation learning framework under uncalibrated multi-view settings. The MV-DE is robust to free head movements." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.89, + 0.182 + ], + "angle": 0, + "content": "- We uncouple gaze from other interference such as head pose by the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.183, + 0.891, + 0.243 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness of the MV-DE framework in unconstrained environments. Qualitative analyses prove that the extracted gaze representation is disentangled from head pose and appearance." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.137, + 0.891, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.258, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.753, + 0.299 + ], + "angle": 0, + "content": "2.1. Supervised Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Appearance-based gaze estimation approaches aim to estimate gaze from eye or face appearance directly. Early methods estimate gaze directions from eye images [9, 28, 36]. Zhang et al. first propose to utilize full face images and outperform eye-image-based methods [37]. Since then, most CNN-based gaze estimation methods employ face images as input. Chen et al. propose to utilize dilated convolution for gaze estimation [5]. Cheng et al. propose to employ the Transformer architecture for gaze estimation [6]. Some recent methods utilize both face and eye images as input for better estimation performance [1, 8, 17, 22]. Above methods are trained in an end-to-end manner with gaze annotations. A number of gaze estimation datasets are published. These datasets are collected under different scenario with varies devices, including web cameras [11, 37], \\(360^{\\circ}\\) cameras [21], high-resolution cameras [39], eye trackers [29] and mobile devices [22]. With sufficient labeled data, gaze feature disentanglement is achieved by using GAN [10] and Nerf [32]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.753, + 0.62 + ], + "angle": 0, + "content": "2.2. Multi-view Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Before the development of Deep Learning methods, most conventional model-based methods utilize multiple cameras for gaze estimation [16]. Model-based methods reconstruct 3D eyeball models and obtain the gaze direction based on the optical geometry of human eye structures [15]. Model-based methods achieve remarkable estimation accuracy. However, these methods have high requirements for the shooting angle of cameras. Thus, these methods are usually employed in the Head Mounted Devices, such as Meta Quest Pro and Microsoft Hololens, where the cameras are approximately stationary relative to the eye of users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Multi-view settings have also drawn a lot of attention in CNN-based gaze estimation methods. A few multiview gaze estimation datasets have been proposed. Zhang et al. propose the ETH-XGaze dataset, which employs 18 high-resolution cameras to capture face images with large head pose range [39]. Park et al. propose the EVE dataset [29]. They use 3 web cameras and a industrial camera to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "capture user face images. Qin et al. propose to utilize 3D face alignment approaches to generate multiple rotated images from a given sample [31]. These rotated images can be regarded as samples from different virtual cameras. Based on these datasets, CNN-based multi-view gaze estimation approaches have been proposed. These methods integrate features from different views by concatenating [24] or self attention [7] to improve estimation accuracy" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.221, + 0.459, + 0.238 + ], + "angle": 0, + "content": "2.3. Unsupervised Gaze Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.425 + ], + "angle": 0, + "content": "Unsupervised representation learning has always been a hot topic in Computer Vision community. Recently, Contrastive Learning approaches have achieved satisfying performance in common Computer Vision tasks. Contrastive Learning methods generate multiple views of a given sample by different data augmentation methods and constrain the model to extract similar representation from these views [4, 14]. He et al. propose the Masked Autoencoders, a self-supervised learner trained by simply reconstructing masked images. However, these methods are designed to extract common visual representations, which do not perform well in the gaze estimation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.668 + ], + "angle": 0, + "content": "To learn gaze representations in an unsupervised manner, Yu et al. propose to utilize the gaze redirection task to extract two-dimensional representations that relate to gaze pitch and yaw angles [35]. Sun et al. propose the CrossEncoder, which learn gaze representations by a latent-codeswapping mechanism on eye-consistent image pairs and gaze-similar pairs [33]. Gideon et al. further propose to adapt the Cross-Encoder architecture to the multi-view setting [12]. Above methods all employ eye images as their input, as there are less gaze-irrelevant visual contents. However, appearance of eyes changes dramatically as head pose increases, makes it difficult for these methods to handle samples with large head pose distribution. Our aim is to introduce an unsupervised gaze representation learning approach that employs face images as input, which is robust to unconstrained head movements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.169, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.825 + ], + "angle": 0, + "content": "We propose the Multi-View Dual-Encoder, a face-based multi-view unsupervised gaze representation learning method. Our aim is to extract gaze representations from the input face images without using the gaze label. The main challenge of this task is to separate gaze from the other facial information, especially head pose. Since gaze and head pose are both physical directions, common data augmentations like rotation and flipping are not able to separate them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To achieve our goal, the proposed MV-DE framework introduces two encoders: a Face Encoder and a Gaze Encoder. We first train the Face Encoder to extract general facial representation including head pose and appearance, while excluding gaze information. Then, we freeze the Face" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.885, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.242, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Training process of the Face Encoder. The proposed Face Encoder extracts face representations that contain general facial information while excluding gaze information. It is trained by two tasks: the head pose estimation task and the image reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Encoder and train the Gaze Encoder to extract gaze representation which compensates for the missing gaze information. We constrain the Gaze Encoder to derive similar representation from different views within a given frame since the eyeball rotation is the same. In this way, we isolate gaze representation from other facial features without using the gaze label. In the following sections, we introduce the training pipeline of two encoders in detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.469, + 0.761, + 0.485 + ], + "angle": 0, + "content": "3.1. Training of the Face Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.568 + ], + "angle": 0, + "content": "The target of the Face Encoder is to extract representation of general facial information except gaze. We design two training strategies to achieve this goal: eye masking and multi-task learning including head pose estimation and image reconstruction, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Given a training face image \\( \\boldsymbol{x}_{i,j} \\) where \\( i \\) is the frame index and \\( j \\) is the camera index, we mask the two eye areas by the average pixel value. We further add three random masks to prevent the Face Encoder from directly estimating head pose based on the position of eye masks. Given that input face images are normalized to \\( 224 \\times 224 \\) pixels, we set the size of masks to \\( 55 \\times 33 \\) according to the average size of eyes. Then, we input the masked face images to the Face Encoder for the proposed multi-task learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.872 + ], + "angle": 0, + "content": "First, the Face Encoder extracts the face representation \\( e_{i,j}^{f} \\) from the masked face image. Then, we add a regression MLP and a decoder for the head pose estimation task and the image reconstruction task, respectively. In the head pose estimation task, the target of the Face Encoder is to estimate the 3D Euler Angles of the subject's head pose. We use \\( \\mathcal{L}_1 \\) Loss function for the head pose estimation task: \\( \\mathcal{L}_{hp} = \\mathcal{L}_1(\\hat{\\boldsymbol{y}}_{i,j},\\boldsymbol{y}_{i,j}) \\), where \\( \\hat{\\boldsymbol{y}}_{i,j} \\) is the estimated head pose and \\( \\boldsymbol{y}_{i,j} \\) is the ground truth head pose. The purpose of the head pose estimation task is to ensure that the face representation \\( e_{i,j}^{f} \\) encodes head pose information." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the image reconstruction task, the Face Encoder and the decoder are trained in an adversarial way. The target" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1421" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.885, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Training process of the Gaze Encoder. The Gaze Encoder extracts gaze representations that represent the eyeball rotation, i.e. gaze direction within HCS. The Gaze Encoder is trained by the proposed multi-view gaze representation swapping strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.419 + ], + "angle": 0, + "content": "of the Face Encoder is to reconstruct the image with eyemasks. The target of the decoder is to reconstruct the original image from the face representation \\( e_{i,j}^{f} \\), including the eye area. Both the Face Encoder and the decoder are constrained by the Image Reconstruction Loss \\( \\mathcal{L}_{re} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.43, + 0.469, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) = \\alpha_ {1} \\mathcal {L} _ {2} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) + \\alpha_ {2} \\mathcal {L} _ {p c p t} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.469, + 0.578 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{x}}\\) is the reconstructed image, \\(\\pmb{x}\\) is the target image, \\(\\mathcal{L}_2\\) is the Mean Squared Error (MSE) Loss function and \\(\\mathcal{L}_{pcpt}\\) is the Perception Loss function. \\(\\alpha_{1},\\alpha_{2}\\) are the coefficients of loss functions. The purpose of the image reconstruction tasks is to encode general facial information in the face representation while excluding gaze information. Images constructed from the \\(e_{i,j}^{f}\\) are basically the original face images without iris and pupil." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.653 + ], + "angle": 0, + "content": "Overall, we train the Face Encoder by minimizing \\(\\mathcal{L}_{FE} = \\beta_1\\mathcal{L}_{hp} + \\mathcal{L}_{re}\\), where \\(\\beta_{1}\\) is coefficient to balance two tasks. Once the training of the Face Encoder is completed, we freeze the parameters of the Face Encoder and proceed to train the Gaze Encoder via the multi-view constraint." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.342, + 0.68 + ], + "angle": 0, + "content": "3.2. Training of the Gaze Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.81 + ], + "angle": 0, + "content": "The target of the Gaze Encoder is to extract gaze representation \\( e_{i,j}^{g} \\) that represents the direction of gaze in the HCS without any gaze label. Since the face representation \\( e_{i,j}^{f} \\) encodes general face information except gaze, we combine the gaze representation \\( e_{i,j}^{g} \\) with \\( e_{i,j}^{f} \\) to fully reconstruct the original image, so that \\( e_{i,j}^{g} \\) compensates for the missing gaze information. The training strategy of the Gaze Encoder is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Given a training sample \\( \\pmb{x}_{i,j} \\), we extract the gaze representation from the original image and the face representation from the eye-masked image. Then, we input both representations to the decoder to reconstruct the original image. To reconstruct the original image, the Gaze Encoder has to capture gaze information which is missing in the face repre" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.404 + ], + "angle": 0, + "content": "sentation. Considering that the region occupied by two eyes represents only a small portion of the face image, we additionally compute the Image Reconstruction Loss for both eye regions:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.415, + 0.891, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {i, j} = G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j} ^ {g}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.439, + 0.832, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G E} = \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}) + \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.891, + 0.514 + ], + "angle": 0, + "content": "where \\(G(\\cdot)\\) is the decoder, \\(\\hat{\\pmb{x}}_{i,j}^{e}\\) and \\(\\pmb{x}_{i,j}^{e}\\) are the two eye regions of the reconstructed image and the original image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.726 + ], + "angle": 0, + "content": "However, the gaze representation can also encode other facial information such as head pose during the training process. In an extreme case, the decoder might theoretically reconstruct the entire image solely based on the gaze representation, since the Gaze Encoder utilizes the full face image as input. We utilize the consistency of eyeball rotation in the multi-view settings to exclude gaze-irrelevant information. First, we randomly sample an image \\( \\pmb{x}_{i,j'} \\) from another view within the same frame where \\( j' \\neq j \\). Since the eyeball rotations from both views are consistent, we swap the gaze representations of two views and reconstruct the eye regions of the original views. For a training image pair \\( \\{\\pmb{x}_{i,j}, \\pmb{x}_{i,j'}\\} \\), the final loss function of the Gaze Encoder \\( \\mathcal{L}_{GE} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.738, + 0.891, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G E} = \\beta_ {2} \\mathcal {L} _ {r e} ^ {\\text {f a c e}} + \\beta_ {3} \\mathcal {L} _ {r e} ^ {\\text {e y e s}} + \\beta_ {4} \\mathcal {L} _ {r e} ^ {\\text {s w a p}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.768, + 0.891, + 0.8 + ], + "angle": 0, + "content": "where \\(\\beta_{2},\\beta_{3},\\beta_{4}\\) are the coefficients, \\(\\mathcal{L}_{re}^{face}\\) is the Image Reconstruction Loss of full face images from both view:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.811, + 0.891, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} ^ {\\text {f a c e}} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}}, \\boldsymbol {x} _ {i, j ^ {\\prime}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.842, + 0.891, + 0.87 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{re}^{eyes}\\) is the Image Reconstruction Loss of eye regions from both view:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.885, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} ^ {e y s} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}} ^ {e}, \\boldsymbol {x} _ {i, j ^ {\\prime}} ^ {e}\\right). \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{re}^{swap}\\) is the Image Reconstruction Loss of the generated eye regions after gaze representation swapping:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.128, + 0.47, + 0.156 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} ^ {s w a p} = \\mathcal {L} _ {r e} \\left(G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j ^ {\\prime}} ^ {g}\\right) ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.152, + 0.39, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} (G (\\pmb {e} _ {i, j ^ {\\prime}} ^ {f}, \\pmb {e} _ {i, j} ^ {g}) ^ {e}, \\pmb {x} _ {i, j ^ {\\prime}} ^ {e}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.252 + ], + "angle": 0, + "content": "At test time, the Gaze Encoder extracts gaze representations from the given single-view face images, similar to common supervised single-view gaze estimation methods. We rotate the estimated HCS gaze to CCS by the head pose label for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.259, + 0.295, + 0.276 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Training Details: The proposed method is implemented by PyTorch using two RTX 3090 GPU. We employ the Adam optimizer with a leaning rate of \\(10^{-3}\\) for all the encoders and decoders. The Face Encoder is trained for 13 epochs with a batch size of 100. Learning rate is decayed by 0.2 every 4 epochs. The Gaze Encoder is trained for 5 epochs with a batch size of 50. Learning rate is decayed by 0.2 every 2 epochs. We use the data rectification method from [37] and histogram equalization to normalize the input face images. \\((\\alpha_{1},\\alpha_{2})\\) are set to (10,1) and (10,0.5) during the training of the Face Encoder and Gaze Encoder, respectively. \\(\\beta_{1}\\) is set to 0.1 and \\((\\beta_{2},\\beta_{3},\\beta_{4})\\) is set to (0.3,0.35,0.35). We use the first 6 convolutional layers of a ImageNet pretrained VGG-16 to calculate the Perceptual Loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.811 + ], + "angle": 0, + "content": "Network Architecture: Both the Face Encoder and the Gaze Encoder use the ResNet-18 [18] as backbone. For the Face Encoder, we take the \\((512*7*7)\\) feature map after the last residual block as the Face Representation. The output channels of the last Linear Layer is set to 3 for head pose estimation. For the Gaze Encoder, the Linear layers is replaced by a \\(1\\times 1\\) Convolutionallayer to compress the \\((512*7*7)\\) feature map to \\((1*7*7)\\) gaze representation. In the decoder, the channel of the gaze representation is expanded to 10 by a \\(1\\times 1\\) Convolutionallayer and then concatenated with the face representation. The decoder also has four residual blocks with target channel sizes of [256, 128, 63, 32]. The feature maps are up-scaled to twice their size before each residual block. We upscale the feature map again and use two Convolutional layers to generate the final \\((3\\times 224\\times 224)\\) image. During the training of the Face Encoder, we set the gaze representation to zeros. Note that we train the decoder from scratch during the training of the Gaze Encoder, as the decoder has learned to ignore the 0 replaced gaze representation during the training of the Face Encoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.248, + 0.864 + ], + "angle": 0, + "content": "4.1. Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on four different gaze estimation datasets: ETH-XGaze [39], MPII-NV [31], EVE [29] and" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.09, + 0.885, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.285, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Head pose (top row) and gaze direction distributions (bottom row) of 4 different datasets. The EVE and the EyeDiap datasets only provide very limited head pose and gaze range with frontal faces." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.361, + 0.891, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Figure 5. Image reconstruction results of combining face representations with different gaze representations in the ETH-XGaze dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.703 + ], + "angle": 0, + "content": "EyeDiap [11]. We visualize the head pose and gaze distribution of each dataset in Fig. 4. The distribution of these datasets varies significantly. below we introduce each dataset in detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "ETH-XGaze \\((\\mathcal{D}_E)\\): A multi-view dataset collected by 18 high-resolution cameras. It provides very large head pose and gaze distributions. ETH-XGaze dataset contains 80 subjects, results in over 750,000 images in total. We divide the last 5 subjects as the labled test set and the first 75 subjects as the multi-view unsupervised training set. We follow [25] to optimize the gaze and head pose label to ensure multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "MPII-NV \\((\\mathcal{D}_M)\\): A synthesized multi-view dataset. We follow [31] to reconstruct 3D faces from MPIIFaceGaze [37] dataset and rotate every 3D face to generate 18 different views, referring to the setting of ETH-XGaze. MPII-NV dataset contains 15 subjects and 359,984 images in total." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1423" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 1. Gaze estimation error of few-shot experiments in degrees. Bold numbers are the best results and underline numbers are the second best results." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.122, + 0.894, + 0.251 + ], + "angle": 0, + "content": "
DatasetETH-XGazeMPII-NVEVEAverage
Calibration Num50100200Avg.50100200Avg.50100200Avg.
BYOL[14]25.6726.0425.2925.6710.1912.6511.2611.3710.369.749.749.9515.66
SimCLR[4]26.8124.9823.8625.2212.5512.1312.1412.2712.5311.3611.4811.7916.43
Cross-Encoder[33]16.2314.7214.5415.1611.4711.3611.7511.538.077.817.487.7911.49
DE (ours w/o MV)23.3219.9317.7620.3411.6410.7310.7411.0410.799.739.449.9913.79
MV-DE (ours)8.668.087.778.177.286.526.396.739.188.68.688.827.91
Gain▼ 7.57▼ 6.64▼ 6.77▼ 6.99▼ 2.91▼ 4.21▼ 4.35▼ 4.31▲ 1.11▲ 0.79▲ 1.20▲ 1.03▼ 3.58
" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.271, + 0.462, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.453, + 0.471, + 0.497 + ], + "angle": 0, + "content": "Figure 6. Gaze estimation error relative to the gaze yaw angle in the ETH-XGaze dataset. The proposed MV-DE performs significantly better when the gaze yaw angle is large." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.377, + 0.536 + ], + "angle": 0, + "content": "We use the last subject as the labeled test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.643 + ], + "angle": 0, + "content": "EVE \\((\\mathcal{D}_V)\\): A multi-view dataset with 4 camera views. As shown in the Fig. 4, the EVE dataset only provides frontal faces with limited head pose and gaze distribution. We use the training and testing set split as defined in the original dataset and sample 3 images per second from the original videos, result in 527,896 images for unsupervised training and 64,464 images for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.703 + ], + "angle": 0, + "content": "EyeDiap \\((\\mathcal{D}_D)\\): A single-view dataset with limited head pose and gaze distribution with 16,674 images in total. We only use the EyeDiap in the cross-domain experiments. Thus, we use the whole dataset for training or testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We follow [38] to normalize the face images. We employ Histogram Equalization and normalize the pixel values to \\([-1, 1]\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.443, + 0.773 + ], + "angle": 0, + "content": "4.2. Evaluation of the Learned Representations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness of the learned face and gaze representations through image reconstruction. The propose MV-DE is supposed to decouple representations of the eyeball rotation from general facial information. In Fig. 5, we combine the face representation with different gaze representations and reconstruct the face image for validation. As shown in row 2, the decoder effectively reconstructs the input face image from the original face and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.895, + 0.44 + ], + "angle": 0, + "content": "gaze representation. When replacing the original gaze representation with gaze representation from another view, the reconstruction results are similar with the original one. It proves that the MV-DE successfully learned the multi-view consistency of the eyeball rotation. In the last row, the gaze representation is replaced by zero vectors. The head pose and appearance of the reconstructed images are almost identical with the original image, while the eyeball rotations are rather random. These results confirm that the MV-DE successfully separate eyeball rotation from general facial information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.453, + 0.877, + 0.469 + ], + "angle": 0, + "content": "4.3. Application of the MV-DE: Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.658 + ], + "angle": 0, + "content": "The learned gaze representation can be used for gaze estimation by adding a MLP head calibrated under the few-shot learning setting. The estimation error is also an evaluation of the learned gaze representation. In the MV-DE, the MLP takes the gaze representation and the head pose estimation as input. The estimated head pose is first encoded to a 30 dimensional embedding and then concatenated with the gaze representation through 3 linear layers with target dimensions of (64, 64, 2). The MLP head is trained to estimate the HCS gaze. Then, we rotate the HCS gaze by head rotation to obtain CCS gaze. Sigmoid function is employed in the activation layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In Tab. 1, we compare the MV-DE with 3 SOTA unsupervised learning methods. BYOL [14] and SimCLR [4] are two SOTA contrastive learning methods. We modify them to adapt the multi-view gaze representation learning task. We employ face images from different camera views in the same frame instead of data augmentation methods to generate the positive pairs. Cross-Encoder [33] is a SOTA eye-based unsupervised gaze representation learning method. Gideon et al. also propose an eye-based method [12], but their method requires sample pairs with different eyeball rotation while keeping the head stable within a short video clip. Since such sample pairs are not available in \\(\\mathcal{D}_E\\) and \\(\\mathcal{D}_M\\), we exclude their moethod in comparison. We also remove the multi-view constraint in the MV-DE as an ablation study (named DE). Without the multi-view constraint, the Gaze Encoder is trained by the image reconstruc" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1424" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.119 + ], + "angle": 0, + "content": "Table 2. Estimation error of different designs of the MLP head with 200 calibration samples." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.13, + 0.47, + 0.313 + ], + "angle": 0, + "content": "
DimensionMLP inputs & outputsD_ED_MD_V
d=32eg→CCS Gaze24.3442.829.83
eg+HP→CCS Gaze20.5942.299.57
eg→HCS Gaze8.8910.498.66
eg+HP→HCS Gaze8.637.268.60
d=64eg→CCS Gaze22.2741.709.98
eg+HP→CCS Gaze18.1134.569.55
eg→HCS Gaze8.437.448.68
eg+HP→HCS Gaze7.846.398.67
d=256eg→CCS Gaze20.9740.9910.32
eg+HP→CCS Gaze18.8834.0510.39
eg→HCS Gaze8.076.508.72
eg+HP→HCS Gaze7.595.718.87
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.275, + 0.355 + ], + "angle": 0, + "content": "tion task described in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.431 + ], + "angle": 0, + "content": "Overall, the average accuracy of the MV-DE outperforms other SOTA methods significantly. BYOL and SimCLR perform worse than the gaze-specialized CrossEncoder and the MV-DE, proves that common visual representations are not suitable for the gaze estimation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Most importantly, the MV-DE outperforms the CrossEncoder significantly in the ETH-XGaze and MPII-NV datasets, where the head pose distributes across a wide range. On the other hand, the MV-DE achieves comparable performance with the Cross-Encoder in the EVE dataset, because the EVE dataset only contains frontal face images with limited head pose range." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.674 + ], + "angle": 0, + "content": "To further investigate the performance gap between the Cross-Encoder and the MV-DE, we visualize the gaze estimation error relative to the gaze yaw angle within the ETH-XGaze dataset in Fig. 6. Obviously, the estimation error of the Cross-Encoder increases severely when the gaze yaw angle approach \\(100^{\\circ}\\). Samples with such extreme gaze yaw angle are generally associated with large head poses. Above results prove the point that the MV-DE are more robust to large head poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Results from the last 2 rows demonstrate the importance of the proposed multi-view gaze representation swapping strategy. The multi-view constraint brought a improvement as large as \\(5.88^{\\circ}\\). But the Dual-Encoder (DE) still achieves better overall performance than SOTA constrastive learning methods, primary attributes to the proposed dual-encoder architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.421, + 0.816 + ], + "angle": 0, + "content": "4.3.1 Ablation Study: Design of the MLP Head" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The MLP head is responsible for estimating gaze from the learned gaze representations under the few-shot learning setting. In Tab. 2, we verify the performances of different MLP designs with 200 calibration samples. For example, \\([d = 64, e_g + HP \\rightarrow HCSGaze]\\) denotes that the MLP" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.894, + 0.158 + ], + "angle": 0, + "content": "Table 3. Cross-dataset gaze estimation errors of using the MV-DE as an unsupervised pretrain method. We first pretrain the model on the ETH-XGaze or the MPII-NV dataset unsupervisely. Then, we use the labeled training dataset to calibrate the MLP head." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.171, + 0.894, + 0.283 + ], + "angle": 0, + "content": "
MethodUnsupervisedTrainTest Dataset
PretrainDatasetETH-XgazeMPII-NVEyeDiap
Baseline-EVE31.2732.5212.55
MV-DEETH-XGazeEVE19.67 ▼ 37.10%10.45▼ 67.87%11.36 ▼ 9.47%
MV-DEMPII-NVEVE25.42 ▼ 18.71%10.33 ▼ 68.23%11.58 ▼ 7.73%
ETH-XgazeMPII-NVEVE
Baseline-EyeDiap44.0639.4623.58
MV-DEETH-XGazeEyeDiap25.15 ▼ 42.92%12.11 ▼ 69.31%10.4 ▼ 55.89%
MV-DEMPII-NVEyeDiap25.58 ▼ 41.94%11.51 ▼ 70.83%10.58 ▼ 55.13%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.385 + ], + "angle": 0, + "content": "includes 64-dimensional hidden layers, takes gaze representations and head poses as the input vector, and predicts gaze directions within Head Coordinate System (HCS). Note that in the MV-DE, we rotate the estimated HCS gaze with head pose label physically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.893, + 0.643 + ], + "angle": 0, + "content": "Results of Tab. 2 lead to 3 conclusions: (1) The MLP performs significantly better on predicting HCS gaze, since the learned gaze representations represent eyeball rotation instead of the CCS Gaze. Theoretically, combination of \\( e_{g} + HP \\) should be enough to predict the CCS gaze, but the estimation error is still huge, probably due to the limited number of calibration samples and MLP parameters. (2) Head pose seems helpful for predicting the HCS gaze. This observation is consistent with previous research [37] that the additional information of the full face images helps in the gaze estimation task. (3) 256-dimensional hidden layers achieve the best performance in the 200-shot settings. But larger number of parameters require more calibration samples, the performance of 256-dimensional hidden layers degrades in 50 and 100 shot calibration settings. Considering the balance between the number of parameters and required calibration samples, we choose \\( d = 64 \\) for the MV-DE." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.664, + 0.892, + 0.695 + ], + "angle": 0, + "content": "4.3.2 Additional Experiments: MV-DE as an Unsupervised Pretrain Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Another potential application of the MV-DE is to use as an unsupervised pretrain method, since it is more cost-effective to collect large number of unlabeled face images than labeled ones. In Tab. 3, we employ the MV-DE to pretrain the Gaze Encoder without gaze label in the ETH-XGaze or the MPII-NV datasets. Then, we freeze the Gaze Encoder and only train the MLP head in the EVE or the EyeDiap dataset with gaze label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To establish a baseline, we directly train the Gaze Encoder and the regression MLP in an end-to-end manner with gaze label in the labeled training dataset, just like common supervised gaze estimation methods. The cross-dataset errors of the baseline method are extremely large on the ETH-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.471, + 0.325 + ], + "angle": 0, + "content": "Figure 7. Distributions of cross-domain gaze estimation error with (right) and without (left) the proposed unsupervised pretrain on the ETH-XGaze. Blue rectangle indicates the gaze distribution range of the labeled training dataset (EyeDiap)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.337, + 0.458, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.63 + ], + "angle": 0, + "content": "Figure 8. Results of using the MV-DE for gaze redirection. Spheres on the bottom-right corner indicates the gaze direction ground truth within the HCS, i.e. the eyeball rotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.473, + 0.817 + ], + "angle": 0, + "content": "XGaze and MPII-NV dataset, since the gaze distributions of the training sets are significantly smaller than the test sets. When used as a unsupervised pretrain method, the MV-DE effectively improves the cross-dataset performance of the baseline model. The distributions of estimation errors in Fig. 7 demonstrate that with the pretrain of the MV-DE, the model generates reasonable estimations in a significantly larger range. Above results indicate that the MV-DE can be used as a pretrain method to improve cross-dataset performances when large amount of unlabeled face images are available." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.46, + 0.848 + ], + "angle": 0, + "content": "4.4. Application of the MV-DE: Gaze Redirection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The learned representations of the MV-DE can also be used for the gaze redirection task. Results in Fig. 8 show that the MV-DE effectively redirects the gaze direction accord" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.264, + 0.893, + 0.321 + ], + "angle": 0, + "content": "Figure 9. Gaze redirection results of combining different gaze representations linearly. Results demonstrate that the gaze representations exhibit good linearity invariance. Spheres on the bottom-right corner indicate the ground truth or the target eyeball rotation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.422 + ], + "angle": 0, + "content": "ing to the HCS gaze of the reference images, while keeping the head pose and identity unchanged. Fig. 8 demonstrate that the learned gaze representation is subject-independent. The head pose and identity of the reconstructed images are controlled by the face representations, where the eyeabll rotation is controlled by the gaze representations as expected." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.433, + 0.89, + 0.449 + ], + "angle": 0, + "content": "4.4.1 Linearity Invariance of the Gaze Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.683 + ], + "angle": 0, + "content": "Gaze, as a physical direction vector, conforms to the principles of additivity. Ideally, well-learned gaze representations should possess similar characteristics. In Fig. 9, we combine gaze representations from different samples linearly with different ratios to reconstruct face images. The sphere in the bottom-left corner of reconstructed the images indicates the target eyeball rotation angle, which corresponds to the linear combination of eyeball rotations from the HCS gaze source images. Results demonstrate that the learned gaze representation satisfies Linearity Invariance, i.e. the eyeball rotation of reconstructed images undergo the same linear combination as the input gaze representations. Experiments in Fig. 9 prove that the MV-DE successfully captures and retains part of the physical properties of gaze during the unsupervised learning process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.62, + 0.712 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework based on multi-view face images. We propose the Dual-Encoder architecture and the multiview gaze representation swapping strategy to learn gaze representations that are separated from general facial information without gaze label. Experiments show that the learned gaze representations can be used for downstream tasks like gaze estimation and gaze redirection. Gaze estimation results show that the quality of learned gaze representations from the MV-DE are significantly better than other SOTA methods with unconstrained head movements." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1426" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Yiwei Bao, Yihua Cheng, Yunfei Liu, and Feng Lu. Adaptive feature fusion network for gaze tracking in mobile tablets. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9936-9943. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Yiwei Bao, Jiaxi Wang, Zhimin Wang, and Feng Lu. Exploring 3d interaction with gaze guidance in augmented reality. In 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 22-32. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.473, + 0.327 + ], + "angle": 0, + "content": "[3] Ariella Fornachari Ribeiro Belan, Marcos Vasconcelos Pais, Marina von Zuben de Arruda Camargo, Livea Carla Fidalgo Garcez Sant'Ana, Marcia Radanovic, and Orestes Vicente Forlenza. Diagnostic performance of an eye-tracking assisted visual inference language test in the assessment of cognitive decline due to alzheimer's disease. Journal of Alzheimer's Disease, (Preprint):1-15, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[5] Zhaokang Chen and Bertram E Shi. Appearance-based gaze estimation using dilated-convolutions. In Asian Conference on Computer Vision, pages 309–324. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[6] Yihua Cheng and Feng Lu. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3341-3347. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[7] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20632-20641, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10623-10630, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[10] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 436-443, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[11] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 255–258, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] John Gideon, Shan Su, and Simon Stent. Unsupervised multi-view gaze representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5001-5009, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Mohamed Waheed Gomaa, Rasha O Mahmoud, and Amany M Sarhan. A cnn-lstm-based deep learning approach for driver drowsiness prediction. Journal of Engineering Research, 6(3):59-70, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Elias Daniel Guestrin and Moshe Eizenman. General theory of remote gaze estimation using the pupil center and corneal reflections. IEEE Transactions on biomedical engineering, 53(6):1124-1133, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Dan Witzner Hansen and Qiang Ji. In the eye of the beholder: A survey of models for eyes and gaze. IEEE transactions on pattern analysis and machine intelligence, 32(3):478-500, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Junfeng He, Khoi Pham, Nachiappan Valliappan, Pingmei Xu, Chase Roberts, Dmitry Lagun, and Vidhya Navalpakkam. On-device few-shot personalization for real-time gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] Daniel A Hofmaenner and Philipp K Buehler. The impact of eye-tracking on patient safety in critical care. Journal of Clinical Monitoring and Computing, 36(6):1577-1579, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhandarkar, Wojciech Matusik, and Antonio Torralba. Eye tracking for everyone. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2176-2184, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Astar Lev, Yoram Braw, Tomer Elbaum, Michael Wagner, and Yuri Rassovsky. Eye tracking during a continuous performance test: Utility for assessing ADHD patients. Journal of Attention Disorders, 26(2):245-255, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Dongze Lian, Lina Hu, Weixin Luo, Yanyu Xu, Lixin Duan, Jingyi Yu, and Shenghua Gao. Multiview multitask gaze estimation with deep convolutional neural networks. IEEE transactions on neural networks and learning systems, 30 (10):3010-3023, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. Unpublished manuscript. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] SI Lyapunov, II Shoshina, and IS Lyapunov. Tremor eye movements as an objective marker of driver's fatigue. Human Physiology, 48(1):71-77, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Nada Mohammed Murad, Lilia Rejeb, and Lamjed Ben Said. Computing driver tiredness and fatigue in automobile via eye" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "tracking and body movements. Periodicals of Engineering and Natural Sciences, 10(1):573-586, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.174 + ], + "angle": 0, + "content": "[28] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In Proceedings of the European conference on computer vision (ECCV), pages 721-738, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[29] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[30] Thammathip Piumsomboon, Gun Lee, Robert W Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In 2017 IEEE symposium on 3D user interfaces (3DUI), pages 36-39. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[31] Jiawei Qin, Takuru Shimoyama, and Yusuke Sugano. Learning-by-novel-view-synthesis for full-face appearance-based 3d gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[32] Alessandro Ruzzi, Xiangwei Shi, Xi Wang, Gengyan Li, Shalini De Mello, Hyung Jin Chang, Xuong Zhang, and Otmar Hilliges. Gazenerf: 3d-aware gaze redirection with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9676-9685, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[33] Yunjia Sun, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Cross-encoder for unsupervised gaze representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3702-3711, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[34] Xin Yi, Yiqin Lu, Ziyin Cai, Zihan Wu, Yuntao Wang, and Yuanchun Shi. Gazedock: Gaze-only menu selection in virtual reality using auto-triggering peripheral menu. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 832-842. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.468, + 0.65 + ], + "angle": 0, + "content": "[35] Yu Yu and Jean-Marc Odobez. Unsupervised representation learning for gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7314-7324, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.652, + 0.468, + 0.705 + ], + "angle": 0, + "content": "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4511-4520, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.468, + 0.775 + ], + "angle": 0, + "content": "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-60, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.468, + 0.831 + ], + "angle": 0, + "content": "[38] Xuong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM symposium on eye tracking research & applications, pages 1-9, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[39] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In European Conference on Computer Vision, pages 365-381. Springer, 2020. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1428" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_origin.pdf b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26dc614a05ed5fcad6eb49b4374960d6cb9f8fba --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/c85bb360-36d7-45dd-b02d-b8b6c1267338_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22d7043d06f7958947c38b9215339b798d670d288368d9a158e375a2fbd3522f +size 3361471 diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/full.md b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..772d2452be06285764f9e0b013f07934b9f97afc --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/full.md @@ -0,0 +1,293 @@ +# Unsupervised Gaze Representation Learning from Multi-view Face Images + +Yiwei Bao Feng Lu* +State Key Laboratory of VR Technology and Systems, School of CSE, Beihang University +{baoyiwei, lufeng}@buaa.edu.cn + +# Abstract + +Annotating gaze is an expensive and time-consuming endeavor, requiring costly eye-trackers or complex geometric calibration procedures. Although some eye-based unsupervised gaze representation learning methods have been proposed, the quality of gaze representation extracted by these methods degrades severely when the head pose is large. In this paper, we present the Multi-View Dual-Encoder (MV-DE), a framework designed to learn gaze representations from unlabeled multi-view face images. Through the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy, the MV-DE successfully disentangles gaze from general facial information and derives gaze representations closely tied to the subject's eyeball rotation without gaze label. Experimental results illustrate that the gaze representations learned by the MV-DE can be used in downstream tasks, including gaze estimation and redirection. Gaze estimation results indicate that the proposed MV-DE displays notably higher robustness to uncontrolled head movements when compared to state-of-the-art (SOTA) unsupervised learning methods. + +# 1. Introduction + +Vision is one of the most important sense for humans. Human gaze reveals the direction of visual attention, which is an important cue for understanding how humans perceive the surrounding world. Thus, gaze estimation techniques have become an vital tool in numerous applications, such as Virtual Realty and Augmented Reality[2, 30, 34], automotive safety [13, 26, 27] and healthcare [3, 20, 23]. In recent years, appearance-based gaze estimation methods have drawn a lot of attention, since these methods only require simple web cameras, eliminating the need for expensive eye trackers with dedicated devices such as infrared cameras. Among these methods, Convolutional Neural Networks (CNN) based approaches exhibit exceptional performance in unconstrained environments. + +![](images/ec4df738f571d5cbf031c0bc8d3afe79f18ae2ad4b8c9653e838d5bae6114280.jpg) +Figure 1. We propose the Multi-View Dual-Encoder, a face-based unsupervised gaze representation learning framework that is robust to large head poses. + +CNN-based gaze estimation methods are usually trained in an end-to-end manner with substantial amount of labeled data. The performance of CNN-based gaze estimation methods highly relies on the quantity and diversity of the training dataset. Unfortunately, annotating gaze directions is difficult. Unlike common Computer Vision tasks like Object Detection, gaze cannot be reliably annotated without specialized hardware. Gaze direction is either measured by costly eye trackers under controlled environment [29], or obtained by the line connecting the 3D face center to the 3D position of the gaze target [39]. It takes great effort and time to collect diverse labeled training data, since the participants need to stare at large amount of gaze targets as instructed. + +To address the challenge of annotation, a number of unsupervised representation learning approaches have been proposed [4, 14, 19]. These approaches extract common visual representations from the input image without annotation. However, such methods do not perform well in gaze estimation, as gaze estimation differs significantly from common visual tasks. Common visual tasks like object classification and detection require representations of overall appearance of the subject. On the other hand, gaze is an direction vector that manifested as rotations of the subtle eye structure in the image. Efforts have been made to design unsupervised gaze representation learning methods + +[33, 35], which incorporate cropped eye images as input to eliminate undesired visual contents. However, existing unsupervised gaze representation learning methods could only handle samples within limited head pose range, since the eye appearance changes dramatically when the head rotation approaches $90^{\circ}$ . + +Recent supervised gaze estimation methods usually employ face images as their input [1, 6, 8, 21, 37] to handle samples with large head poses. Unfortunately, the rich information in the face images becomes a double-edged sword in unsupervised settings. These pieces of information become interference due to the absence of gaze label as constraint. For example, head pose and gaze are strongly coupled. They behave exactly the same in common data augmentation methods such as image flipping and rotation, makes it difficult to separate them even for latest contrastive learning approaches. Thus, unsupervised gaze representation learning method with face images is a challenging task that remain to be solved. + +A potential solution to tackle this challenge is to harness the extra constraint introduced by multi-view settings. Multi-view gaze estimation is a recent hot research topic. A number of supervised multi-view gaze estimation methods [7, 24] have been proposed. Multi-view image pairs provide a set of images with the same eyeball rotation, i.e. gaze direction within the Head Coordinate System (HCS), as they are captured simultaneously. Our aim is to design a face-based multi-view unsupervised gaze representation learning method that works well on samples with large head poses. + +In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework that employs face images as input. The proposed MV-DE framework extracts gaze representation by separating eyeball rotation from general facial information. The MV-DE framework consists of two encoders: the Gaze Encoder and the Face Encoder. First, we train the Face Encoder to extract face representations including head pose and appearance, while excluding gaze information. Then, we freeze the Face Encoder and introduce the gaze feature extracted by the Gaze Encoder to compensate for the missing gaze information. Based on the consistent eyeball rotation across different views within the same frame, we integrate the face representations with gaze representations from different views to reconstruct the original image. At the inference time, the MV-DE extracts gaze representations from single-view face images. + +Experimental results illustrate that the MV-DE framework successfully derives gaze representations across a wide range of head poses. The learned gaze representation can be used for varies downstream tasks such as gaze estimation and gaze redirection. The contribution of this paper are summarized in three folds: + +- We propose the Multi-View Dual-Encoder, a face-based + +unsupervised gaze representation learning framework under uncalibrated multi-view settings. The MV-DE is robust to free head movements. + +- We uncouple gaze from other interference such as head pose by the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy. +- Extensive experiments demonstrate the effectiveness of the MV-DE framework in unconstrained environments. Qualitative analyses prove that the extracted gaze representation is disentangled from head pose and appearance. + +# 2. Related Work + +# 2.1. Supervised Gaze Estimation + +Appearance-based gaze estimation approaches aim to estimate gaze from eye or face appearance directly. Early methods estimate gaze directions from eye images [9, 28, 36]. Zhang et al. first propose to utilize full face images and outperform eye-image-based methods [37]. Since then, most CNN-based gaze estimation methods employ face images as input. Chen et al. propose to utilize dilated convolution for gaze estimation [5]. Cheng et al. propose to employ the Transformer architecture for gaze estimation [6]. Some recent methods utilize both face and eye images as input for better estimation performance [1, 8, 17, 22]. Above methods are trained in an end-to-end manner with gaze annotations. A number of gaze estimation datasets are published. These datasets are collected under different scenario with varies devices, including web cameras [11, 37], $360^{\circ}$ cameras [21], high-resolution cameras [39], eye trackers [29] and mobile devices [22]. With sufficient labeled data, gaze feature disentanglement is achieved by using GAN [10] and Nerf [32]. + +# 2.2. Multi-view Gaze Estimation + +Before the development of Deep Learning methods, most conventional model-based methods utilize multiple cameras for gaze estimation [16]. Model-based methods reconstruct 3D eyeball models and obtain the gaze direction based on the optical geometry of human eye structures [15]. Model-based methods achieve remarkable estimation accuracy. However, these methods have high requirements for the shooting angle of cameras. Thus, these methods are usually employed in the Head Mounted Devices, such as Meta Quest Pro and Microsoft Hololens, where the cameras are approximately stationary relative to the eye of users. + +Multi-view settings have also drawn a lot of attention in CNN-based gaze estimation methods. A few multiview gaze estimation datasets have been proposed. Zhang et al. propose the ETH-XGaze dataset, which employs 18 high-resolution cameras to capture face images with large head pose range [39]. Park et al. propose the EVE dataset [29]. They use 3 web cameras and a industrial camera to + +capture user face images. Qin et al. propose to utilize 3D face alignment approaches to generate multiple rotated images from a given sample [31]. These rotated images can be regarded as samples from different virtual cameras. Based on these datasets, CNN-based multi-view gaze estimation approaches have been proposed. These methods integrate features from different views by concatenating [24] or self attention [7] to improve estimation accuracy + +# 2.3. Unsupervised Gaze Representation Learning + +Unsupervised representation learning has always been a hot topic in Computer Vision community. Recently, Contrastive Learning approaches have achieved satisfying performance in common Computer Vision tasks. Contrastive Learning methods generate multiple views of a given sample by different data augmentation methods and constrain the model to extract similar representation from these views [4, 14]. He et al. propose the Masked Autoencoders, a self-supervised learner trained by simply reconstructing masked images. However, these methods are designed to extract common visual representations, which do not perform well in the gaze estimation task. + +To learn gaze representations in an unsupervised manner, Yu et al. propose to utilize the gaze redirection task to extract two-dimensional representations that relate to gaze pitch and yaw angles [35]. Sun et al. propose the CrossEncoder, which learn gaze representations by a latent-codeswapping mechanism on eye-consistent image pairs and gaze-similar pairs [33]. Gideon et al. further propose to adapt the Cross-Encoder architecture to the multi-view setting [12]. Above methods all employ eye images as their input, as there are less gaze-irrelevant visual contents. However, appearance of eyes changes dramatically as head pose increases, makes it difficult for these methods to handle samples with large head pose distribution. Our aim is to introduce an unsupervised gaze representation learning approach that employs face images as input, which is robust to unconstrained head movements. + +# 3. Method + +We propose the Multi-View Dual-Encoder, a face-based multi-view unsupervised gaze representation learning method. Our aim is to extract gaze representations from the input face images without using the gaze label. The main challenge of this task is to separate gaze from the other facial information, especially head pose. Since gaze and head pose are both physical directions, common data augmentations like rotation and flipping are not able to separate them. + +To achieve our goal, the proposed MV-DE framework introduces two encoders: a Face Encoder and a Gaze Encoder. We first train the Face Encoder to extract general facial representation including head pose and appearance, while excluding gaze information. Then, we freeze the Face + +![](images/1276427211a33707343175ed36949157e06ea13cd622fc3e9857089f8339e46c.jpg) +Figure 2. Training process of the Face Encoder. The proposed Face Encoder extracts face representations that contain general facial information while excluding gaze information. It is trained by two tasks: the head pose estimation task and the image reconstruction task. + +Encoder and train the Gaze Encoder to extract gaze representation which compensates for the missing gaze information. We constrain the Gaze Encoder to derive similar representation from different views within a given frame since the eyeball rotation is the same. In this way, we isolate gaze representation from other facial features without using the gaze label. In the following sections, we introduce the training pipeline of two encoders in detail. + +# 3.1. Training of the Face Encoder + +The target of the Face Encoder is to extract representation of general facial information except gaze. We design two training strategies to achieve this goal: eye masking and multi-task learning including head pose estimation and image reconstruction, as shown in Fig. 2. + +Given a training face image $\boldsymbol{x}_{i,j}$ where $i$ is the frame index and $j$ is the camera index, we mask the two eye areas by the average pixel value. We further add three random masks to prevent the Face Encoder from directly estimating head pose based on the position of eye masks. Given that input face images are normalized to $224 \times 224$ pixels, we set the size of masks to $55 \times 33$ according to the average size of eyes. Then, we input the masked face images to the Face Encoder for the proposed multi-task learning. + +First, the Face Encoder extracts the face representation $e_{i,j}^{f}$ from the masked face image. Then, we add a regression MLP and a decoder for the head pose estimation task and the image reconstruction task, respectively. In the head pose estimation task, the target of the Face Encoder is to estimate the 3D Euler Angles of the subject's head pose. We use $\mathcal{L}_1$ Loss function for the head pose estimation task: $\mathcal{L}_{hp} = \mathcal{L}_1(\hat{\boldsymbol{y}}_{i,j},\boldsymbol{y}_{i,j})$ , where $\hat{\boldsymbol{y}}_{i,j}$ is the estimated head pose and $\boldsymbol{y}_{i,j}$ is the ground truth head pose. The purpose of the head pose estimation task is to ensure that the face representation $e_{i,j}^{f}$ encodes head pose information. + +In the image reconstruction task, the Face Encoder and the decoder are trained in an adversarial way. The target + +![](images/8a5d3fd66b99129d7b31e54f96247b820f52d3d893c6857a34a67454ca0c58b0.jpg) +Figure 3. Training process of the Gaze Encoder. The Gaze Encoder extracts gaze representations that represent the eyeball rotation, i.e. gaze direction within HCS. The Gaze Encoder is trained by the proposed multi-view gaze representation swapping strategy. + +of the Face Encoder is to reconstruct the image with eyemasks. The target of the decoder is to reconstruct the original image from the face representation $e_{i,j}^{f}$ , including the eye area. Both the Face Encoder and the decoder are constrained by the Image Reconstruction Loss $\mathcal{L}_{re}$ : + +$$ +\mathcal {L} _ {r e} (\hat {\boldsymbol {x}}, \boldsymbol {x}) = \alpha_ {1} \mathcal {L} _ {2} (\hat {\boldsymbol {x}}, \boldsymbol {x}) + \alpha_ {2} \mathcal {L} _ {p c p t} (\hat {\boldsymbol {x}}, \boldsymbol {x}), \tag {1} +$$ + +where $\hat{\pmb{x}}$ is the reconstructed image, $\pmb{x}$ is the target image, $\mathcal{L}_2$ is the Mean Squared Error (MSE) Loss function and $\mathcal{L}_{pcpt}$ is the Perception Loss function. $\alpha_{1},\alpha_{2}$ are the coefficients of loss functions. The purpose of the image reconstruction tasks is to encode general facial information in the face representation while excluding gaze information. Images constructed from the $e_{i,j}^{f}$ are basically the original face images without iris and pupil. + +Overall, we train the Face Encoder by minimizing $\mathcal{L}_{FE} = \beta_1\mathcal{L}_{hp} + \mathcal{L}_{re}$ , where $\beta_{1}$ is coefficient to balance two tasks. Once the training of the Face Encoder is completed, we freeze the parameters of the Face Encoder and proceed to train the Gaze Encoder via the multi-view constraint. + +# 3.2. Training of the Gaze Encoder + +The target of the Gaze Encoder is to extract gaze representation $e_{i,j}^{g}$ that represents the direction of gaze in the HCS without any gaze label. Since the face representation $e_{i,j}^{f}$ encodes general face information except gaze, we combine the gaze representation $e_{i,j}^{g}$ with $e_{i,j}^{f}$ to fully reconstruct the original image, so that $e_{i,j}^{g}$ compensates for the missing gaze information. The training strategy of the Gaze Encoder is shown in Fig. 3. + +Given a training sample $\pmb{x}_{i,j}$ , we extract the gaze representation from the original image and the face representation from the eye-masked image. Then, we input both representations to the decoder to reconstruct the original image. To reconstruct the original image, the Gaze Encoder has to capture gaze information which is missing in the face repre + +sentation. Considering that the region occupied by two eyes represents only a small portion of the face image, we additionally compute the Image Reconstruction Loss for both eye regions: + +$$ +\hat {\boldsymbol {x}} _ {i, j} = G \left(\boldsymbol {e} _ {i, j} ^ {f}, \boldsymbol {e} _ {i, j} ^ {g}\right), \tag {2} +$$ + +$$ +\mathcal {L} _ {G E} = \mathcal {L} _ {r e} (\hat {\boldsymbol {x}} _ {i, j}, \boldsymbol {x} _ {i, j}) + \mathcal {L} _ {r e} (\hat {\boldsymbol {x}} _ {i, j} ^ {e}, \boldsymbol {x} _ {i, j} ^ {e}), +$$ + +where $G(\cdot)$ is the decoder, $\hat{\pmb{x}}_{i,j}^{e}$ and $\pmb{x}_{i,j}^{e}$ are the two eye regions of the reconstructed image and the original image, respectively. + +However, the gaze representation can also encode other facial information such as head pose during the training process. In an extreme case, the decoder might theoretically reconstruct the entire image solely based on the gaze representation, since the Gaze Encoder utilizes the full face image as input. We utilize the consistency of eyeball rotation in the multi-view settings to exclude gaze-irrelevant information. First, we randomly sample an image $\pmb{x}_{i,j'}$ from another view within the same frame where $j' \neq j$ . Since the eyeball rotations from both views are consistent, we swap the gaze representations of two views and reconstruct the eye regions of the original views. For a training image pair $\{\pmb{x}_{i,j}, \pmb{x}_{i,j'}\}$ , the final loss function of the Gaze Encoder $\mathcal{L}_{GE}$ is: + +$$ +\mathcal {L} _ {G E} = \beta_ {2} \mathcal {L} _ {r e} ^ {\text {f a c e}} + \beta_ {3} \mathcal {L} _ {r e} ^ {\text {e y e s}} + \beta_ {4} \mathcal {L} _ {r e} ^ {\text {s w a p}}, \tag {3} +$$ + +where $\beta_{2},\beta_{3},\beta_{4}$ are the coefficients, $\mathcal{L}_{re}^{face}$ is the Image Reconstruction Loss of full face images from both view: + +$$ +\mathcal {L} _ {r e} ^ {\text {f a c e}} = \mathcal {L} _ {r e} \left(\hat {\boldsymbol {x}} _ {i, j}, \boldsymbol {x} _ {i, j}\right) + \mathcal {L} _ {r e} \left(\hat {\boldsymbol {x}} _ {i, j ^ {\prime}}, \boldsymbol {x} _ {i, j ^ {\prime}}\right). \tag {4} +$$ + +$\mathcal{L}_{re}^{eyes}$ is the Image Reconstruction Loss of eye regions from both view: + +$$ +\mathcal {L} _ {r e} ^ {e y s} = \mathcal {L} _ {r e} \left(\hat {\boldsymbol {x}} _ {i, j} ^ {e}, \boldsymbol {x} _ {i, j} ^ {e}\right) + \mathcal {L} _ {r e} \left(\hat {\boldsymbol {x}} _ {i, j ^ {\prime}} ^ {e}, \boldsymbol {x} _ {i, j ^ {\prime}} ^ {e}\right). \tag {5} +$$ + +$\mathcal{L}_{re}^{swap}$ is the Image Reconstruction Loss of the generated eye regions after gaze representation swapping: + +$$ +\mathcal {L} _ {r e} ^ {s w a p} = \mathcal {L} _ {r e} \left(G \left(\boldsymbol {e} _ {i, j} ^ {f}, \boldsymbol {e} _ {i, j ^ {\prime}} ^ {g}\right) ^ {e}, \boldsymbol {x} _ {i, j} ^ {e}\right) + \tag {6} +$$ + +$$ +\mathcal {L} _ {r e} (G (\pmb {e} _ {i, j ^ {\prime}} ^ {f}, \pmb {e} _ {i, j} ^ {g}) ^ {e}, \pmb {x} _ {i, j ^ {\prime}} ^ {e}). +$$ + +At test time, the Gaze Encoder extracts gaze representations from the given single-view face images, similar to common supervised single-view gaze estimation methods. We rotate the estimated HCS gaze to CCS by the head pose label for evaluation. + +# 3.3. Implementation Details + +Training Details: The proposed method is implemented by PyTorch using two RTX 3090 GPU. We employ the Adam optimizer with a leaning rate of $10^{-3}$ for all the encoders and decoders. The Face Encoder is trained for 13 epochs with a batch size of 100. Learning rate is decayed by 0.2 every 4 epochs. The Gaze Encoder is trained for 5 epochs with a batch size of 50. Learning rate is decayed by 0.2 every 2 epochs. We use the data rectification method from [37] and histogram equalization to normalize the input face images. $(\alpha_{1},\alpha_{2})$ are set to (10,1) and (10,0.5) during the training of the Face Encoder and Gaze Encoder, respectively. $\beta_{1}$ is set to 0.1 and $(\beta_{2},\beta_{3},\beta_{4})$ is set to (0.3,0.35,0.35). We use the first 6 convolutional layers of a ImageNet pretrained VGG-16 to calculate the Perceptual Loss. + +Network Architecture: Both the Face Encoder and the Gaze Encoder use the ResNet-18 [18] as backbone. For the Face Encoder, we take the $(512*7*7)$ feature map after the last residual block as the Face Representation. The output channels of the last Linear Layer is set to 3 for head pose estimation. For the Gaze Encoder, the Linear layers is replaced by a $1\times 1$ Convolutionallayer to compress the $(512*7*7)$ feature map to $(1*7*7)$ gaze representation. In the decoder, the channel of the gaze representation is expanded to 10 by a $1\times 1$ Convolutionallayer and then concatenated with the face representation. The decoder also has four residual blocks with target channel sizes of [256, 128, 63, 32]. The feature maps are up-scaled to twice their size before each residual block. We upscale the feature map again and use two Convolutional layers to generate the final $(3\times 224\times 224)$ image. During the training of the Face Encoder, we set the gaze representation to zeros. Note that we train the decoder from scratch during the training of the Gaze Encoder, as the decoder has learned to ignore the 0 replaced gaze representation during the training of the Face Encoder. + +# 4. Experiments + +# 4.1. Data Preparation + +We conduct experiments on four different gaze estimation datasets: ETH-XGaze [39], MPII-NV [31], EVE [29] and + +![](images/2ad9c9523511db8eea09f9f50b7b468a82d042fa3c2be6f35c983209b12e2bec.jpg) +Figure 4. Head pose (top row) and gaze direction distributions (bottom row) of 4 different datasets. The EVE and the EyeDiap datasets only provide very limited head pose and gaze range with frontal faces. + +![](images/8229af6ef1af697d193bafbe8f5648567464d8f4e11356e5358083355d725502.jpg) +Figure 5. Image reconstruction results of combining face representations with different gaze representations in the ETH-XGaze dataset. + +EyeDiap [11]. We visualize the head pose and gaze distribution of each dataset in Fig. 4. The distribution of these datasets varies significantly. below we introduce each dataset in detail. + +ETH-XGaze $(\mathcal{D}_E)$ : A multi-view dataset collected by 18 high-resolution cameras. It provides very large head pose and gaze distributions. ETH-XGaze dataset contains 80 subjects, results in over 750,000 images in total. We divide the last 5 subjects as the labled test set and the first 75 subjects as the multi-view unsupervised training set. We follow [25] to optimize the gaze and head pose label to ensure multi-view consistency. + +MPII-NV $(\mathcal{D}_M)$ : A synthesized multi-view dataset. We follow [31] to reconstruct 3D faces from MPIIFaceGaze [37] dataset and rotate every 3D face to generate 18 different views, referring to the setting of ETH-XGaze. MPII-NV dataset contains 15 subjects and 359,984 images in total. + +Table 1. Gaze estimation error of few-shot experiments in degrees. Bold numbers are the best results and underline numbers are the second best results. + +
DatasetETH-XGazeMPII-NVEVEAverage
Calibration Num50100200Avg.50100200Avg.50100200Avg.
BYOL[14]25.6726.0425.2925.6710.1912.6511.2611.3710.369.749.749.9515.66
SimCLR[4]26.8124.9823.8625.2212.5512.1312.1412.2712.5311.3611.4811.7916.43
Cross-Encoder[33]16.2314.7214.5415.1611.4711.3611.7511.538.077.817.487.7911.49
DE (ours w/o MV)23.3219.9317.7620.3411.6410.7310.7411.0410.799.739.449.9913.79
MV-DE (ours)8.668.087.778.177.286.526.396.739.188.68.688.827.91
Gain▼ 7.57▼ 6.64▼ 6.77▼ 6.99▼ 2.91▼ 4.21▼ 4.35▼ 4.31▲ 1.11▲ 0.79▲ 1.20▲ 1.03▼ 3.58
+ +![](images/8f255b2f83446306ebc5c21857e12ddde86ec889cfa975e5b136ff728a1c57ed.jpg) +Figure 6. Gaze estimation error relative to the gaze yaw angle in the ETH-XGaze dataset. The proposed MV-DE performs significantly better when the gaze yaw angle is large. + +We use the last subject as the labeled test set. + +EVE $(\mathcal{D}_V)$ : A multi-view dataset with 4 camera views. As shown in the Fig. 4, the EVE dataset only provides frontal faces with limited head pose and gaze distribution. We use the training and testing set split as defined in the original dataset and sample 3 images per second from the original videos, result in 527,896 images for unsupervised training and 64,464 images for testing. + +EyeDiap $(\mathcal{D}_D)$ : A single-view dataset with limited head pose and gaze distribution with 16,674 images in total. We only use the EyeDiap in the cross-domain experiments. Thus, we use the whole dataset for training or testing. + +We follow [38] to normalize the face images. We employ Histogram Equalization and normalize the pixel values to $[-1, 1]$ . + +# 4.2. Evaluation of the Learned Representations + +In this section, we validate the effectiveness of the learned face and gaze representations through image reconstruction. The propose MV-DE is supposed to decouple representations of the eyeball rotation from general facial information. In Fig. 5, we combine the face representation with different gaze representations and reconstruct the face image for validation. As shown in row 2, the decoder effectively reconstructs the input face image from the original face and + +gaze representation. When replacing the original gaze representation with gaze representation from another view, the reconstruction results are similar with the original one. It proves that the MV-DE successfully learned the multi-view consistency of the eyeball rotation. In the last row, the gaze representation is replaced by zero vectors. The head pose and appearance of the reconstructed images are almost identical with the original image, while the eyeball rotations are rather random. These results confirm that the MV-DE successfully separate eyeball rotation from general facial information. + +# 4.3. Application of the MV-DE: Gaze Estimation + +The learned gaze representation can be used for gaze estimation by adding a MLP head calibrated under the few-shot learning setting. The estimation error is also an evaluation of the learned gaze representation. In the MV-DE, the MLP takes the gaze representation and the head pose estimation as input. The estimated head pose is first encoded to a 30 dimensional embedding and then concatenated with the gaze representation through 3 linear layers with target dimensions of (64, 64, 2). The MLP head is trained to estimate the HCS gaze. Then, we rotate the HCS gaze by head rotation to obtain CCS gaze. Sigmoid function is employed in the activation layer. + +In Tab. 1, we compare the MV-DE with 3 SOTA unsupervised learning methods. BYOL [14] and SimCLR [4] are two SOTA contrastive learning methods. We modify them to adapt the multi-view gaze representation learning task. We employ face images from different camera views in the same frame instead of data augmentation methods to generate the positive pairs. Cross-Encoder [33] is a SOTA eye-based unsupervised gaze representation learning method. Gideon et al. also propose an eye-based method [12], but their method requires sample pairs with different eyeball rotation while keeping the head stable within a short video clip. Since such sample pairs are not available in $\mathcal{D}_E$ and $\mathcal{D}_M$ , we exclude their moethod in comparison. We also remove the multi-view constraint in the MV-DE as an ablation study (named DE). Without the multi-view constraint, the Gaze Encoder is trained by the image reconstruc + +Table 2. Estimation error of different designs of the MLP head with 200 calibration samples. + +
DimensionMLP inputs & outputsD_ED_MD_V
d=32eg→CCS Gaze24.3442.829.83
eg+HP→CCS Gaze20.5942.299.57
eg→HCS Gaze8.8910.498.66
eg+HP→HCS Gaze8.637.268.60
d=64eg→CCS Gaze22.2741.709.98
eg+HP→CCS Gaze18.1134.569.55
eg→HCS Gaze8.437.448.68
eg+HP→HCS Gaze7.846.398.67
d=256eg→CCS Gaze20.9740.9910.32
eg+HP→CCS Gaze18.8834.0510.39
eg→HCS Gaze8.076.508.72
eg+HP→HCS Gaze7.595.718.87
+ +tion task described in Eq. (2). + +Overall, the average accuracy of the MV-DE outperforms other SOTA methods significantly. BYOL and SimCLR perform worse than the gaze-specialized CrossEncoder and the MV-DE, proves that common visual representations are not suitable for the gaze estimation task. + +Most importantly, the MV-DE outperforms the CrossEncoder significantly in the ETH-XGaze and MPII-NV datasets, where the head pose distributes across a wide range. On the other hand, the MV-DE achieves comparable performance with the Cross-Encoder in the EVE dataset, because the EVE dataset only contains frontal face images with limited head pose range. + +To further investigate the performance gap between the Cross-Encoder and the MV-DE, we visualize the gaze estimation error relative to the gaze yaw angle within the ETH-XGaze dataset in Fig. 6. Obviously, the estimation error of the Cross-Encoder increases severely when the gaze yaw angle approach $100^{\circ}$ . Samples with such extreme gaze yaw angle are generally associated with large head poses. Above results prove the point that the MV-DE are more robust to large head poses. + +Results from the last 2 rows demonstrate the importance of the proposed multi-view gaze representation swapping strategy. The multi-view constraint brought a improvement as large as $5.88^{\circ}$ . But the Dual-Encoder (DE) still achieves better overall performance than SOTA constrastive learning methods, primary attributes to the proposed dual-encoder architecture. + +# 4.3.1 Ablation Study: Design of the MLP Head + +The MLP head is responsible for estimating gaze from the learned gaze representations under the few-shot learning setting. In Tab. 2, we verify the performances of different MLP designs with 200 calibration samples. For example, $[d = 64, e_g + HP \rightarrow HCSGaze]$ denotes that the MLP + +Table 3. Cross-dataset gaze estimation errors of using the MV-DE as an unsupervised pretrain method. We first pretrain the model on the ETH-XGaze or the MPII-NV dataset unsupervisely. Then, we use the labeled training dataset to calibrate the MLP head. + +
MethodUnsupervisedTrainTest Dataset
PretrainDatasetETH-XgazeMPII-NVEyeDiap
Baseline-EVE31.2732.5212.55
MV-DEETH-XGazeEVE19.67 ▼ 37.10%10.45▼ 67.87%11.36 ▼ 9.47%
MV-DEMPII-NVEVE25.42 ▼ 18.71%10.33 ▼ 68.23%11.58 ▼ 7.73%
ETH-XgazeMPII-NVEVE
Baseline-EyeDiap44.0639.4623.58
MV-DEETH-XGazeEyeDiap25.15 ▼ 42.92%12.11 ▼ 69.31%10.4 ▼ 55.89%
MV-DEMPII-NVEyeDiap25.58 ▼ 41.94%11.51 ▼ 70.83%10.58 ▼ 55.13%
+ +includes 64-dimensional hidden layers, takes gaze representations and head poses as the input vector, and predicts gaze directions within Head Coordinate System (HCS). Note that in the MV-DE, we rotate the estimated HCS gaze with head pose label physically. + +Results of Tab. 2 lead to 3 conclusions: (1) The MLP performs significantly better on predicting HCS gaze, since the learned gaze representations represent eyeball rotation instead of the CCS Gaze. Theoretically, combination of $e_{g} + HP$ should be enough to predict the CCS gaze, but the estimation error is still huge, probably due to the limited number of calibration samples and MLP parameters. (2) Head pose seems helpful for predicting the HCS gaze. This observation is consistent with previous research [37] that the additional information of the full face images helps in the gaze estimation task. (3) 256-dimensional hidden layers achieve the best performance in the 200-shot settings. But larger number of parameters require more calibration samples, the performance of 256-dimensional hidden layers degrades in 50 and 100 shot calibration settings. Considering the balance between the number of parameters and required calibration samples, we choose $d = 64$ for the MV-DE. + +# 4.3.2 Additional Experiments: MV-DE as an Unsupervised Pretrain Method + +Another potential application of the MV-DE is to use as an unsupervised pretrain method, since it is more cost-effective to collect large number of unlabeled face images than labeled ones. In Tab. 3, we employ the MV-DE to pretrain the Gaze Encoder without gaze label in the ETH-XGaze or the MPII-NV datasets. Then, we freeze the Gaze Encoder and only train the MLP head in the EVE or the EyeDiap dataset with gaze label. + +To establish a baseline, we directly train the Gaze Encoder and the regression MLP in an end-to-end manner with gaze label in the labeled training dataset, just like common supervised gaze estimation methods. The cross-dataset errors of the baseline method are extremely large on the ETH- + +![](images/52e4c07bb8017484a4de72094b875728ce102f1c8a14428f8d349d34544887ca.jpg) +Figure 7. Distributions of cross-domain gaze estimation error with (right) and without (left) the proposed unsupervised pretrain on the ETH-XGaze. Blue rectangle indicates the gaze distribution range of the labeled training dataset (EyeDiap). + +![](images/b53fd10394894cae9d6784fb2cce1ab00c90d5e5bfd0eec807b183b3723681e7.jpg) +Figure 8. Results of using the MV-DE for gaze redirection. Spheres on the bottom-right corner indicates the gaze direction ground truth within the HCS, i.e. the eyeball rotation. + +XGaze and MPII-NV dataset, since the gaze distributions of the training sets are significantly smaller than the test sets. When used as a unsupervised pretrain method, the MV-DE effectively improves the cross-dataset performance of the baseline model. The distributions of estimation errors in Fig. 7 demonstrate that with the pretrain of the MV-DE, the model generates reasonable estimations in a significantly larger range. Above results indicate that the MV-DE can be used as a pretrain method to improve cross-dataset performances when large amount of unlabeled face images are available. + +# 4.4. Application of the MV-DE: Gaze Redirection + +The learned representations of the MV-DE can also be used for the gaze redirection task. Results in Fig. 8 show that the MV-DE effectively redirects the gaze direction accord + +![](images/181f90429d1f451985d8b251c349af38e4053dc4f1e5ec651c1171466b09f0b2.jpg) +Figure 9. Gaze redirection results of combining different gaze representations linearly. Results demonstrate that the gaze representations exhibit good linearity invariance. Spheres on the bottom-right corner indicate the ground truth or the target eyeball rotation. + +ing to the HCS gaze of the reference images, while keeping the head pose and identity unchanged. Fig. 8 demonstrate that the learned gaze representation is subject-independent. The head pose and identity of the reconstructed images are controlled by the face representations, where the eyeabll rotation is controlled by the gaze representations as expected. + +# 4.4.1 Linearity Invariance of the Gaze Representation + +Gaze, as a physical direction vector, conforms to the principles of additivity. Ideally, well-learned gaze representations should possess similar characteristics. In Fig. 9, we combine gaze representations from different samples linearly with different ratios to reconstruct face images. The sphere in the bottom-left corner of reconstructed the images indicates the target eyeball rotation angle, which corresponds to the linear combination of eyeball rotations from the HCS gaze source images. Results demonstrate that the learned gaze representation satisfies Linearity Invariance, i.e. the eyeball rotation of reconstructed images undergo the same linear combination as the input gaze representations. Experiments in Fig. 9 prove that the MV-DE successfully captures and retains part of the physical properties of gaze during the unsupervised learning process. + +# 5. Conclusion + +In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework based on multi-view face images. We propose the Dual-Encoder architecture and the multiview gaze representation swapping strategy to learn gaze representations that are separated from general facial information without gaze label. Experiments show that the learned gaze representations can be used for downstream tasks like gaze estimation and gaze redirection. Gaze estimation results show that the quality of learned gaze representations from the MV-DE are significantly better than other SOTA methods with unconstrained head movements. + +# References + +[1] Yiwei Bao, Yihua Cheng, Yunfei Liu, and Feng Lu. Adaptive feature fusion network for gaze tracking in mobile tablets. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9936-9943. IEEE, 2021. 2 +[2] Yiwei Bao, Jiaxi Wang, Zhimin Wang, and Feng Lu. Exploring 3d interaction with gaze guidance in augmented reality. In 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 22-32. IEEE, 2023. 1 +[3] Ariella Fornachari Ribeiro Belan, Marcos Vasconcelos Pais, Marina von Zuben de Arruda Camargo, Livea Carla Fidalgo Garcez Sant'Ana, Marcia Radanovic, and Orestes Vicente Forlenza. Diagnostic performance of an eye-tracking assisted visual inference language test in the assessment of cognitive decline due to alzheimer's disease. Journal of Alzheimer's Disease, (Preprint):1-15, 2023. 1 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6 +[5] Zhaokang Chen and Bertram E Shi. Appearance-based gaze estimation using dilated-convolutions. In Asian Conference on Computer Vision, pages 309–324. Springer, 2018. 2 +[6] Yihua Cheng and Feng Lu. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3341-3347. IEEE, 2022. 2 +[7] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20632-20641, 2023. 2, 3 +[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10623-10630, 2020. 2 +[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2 +[10] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 436-443, 2022. 2 +[11] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 255–258, 2014. 2, 5 +[12] John Gideon, Shan Su, and Simon Stent. Unsupervised multi-view gaze representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5001-5009, 2022. 3, 6 +[13] Mohamed Waheed Gomaa, Rasha O Mahmoud, and Amany M Sarhan. A cnn-lstm-based deep learning approach for driver drowsiness prediction. Journal of Engineering Research, 6(3):59-70, 2022. 1 +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, + +Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 3, 6 +[15] Elias Daniel Guestrin and Moshe Eizenman. General theory of remote gaze estimation using the pupil center and corneal reflections. IEEE Transactions on biomedical engineering, 53(6):1124-1133, 2006. 2 +[16] Dan Witzner Hansen and Qiang Ji. In the eye of the beholder: A survey of models for eyes and gaze. IEEE transactions on pattern analysis and machine intelligence, 32(3):478-500, 2009. 2 +[17] Junfeng He, Khoi Pham, Nachiappan Valliappan, Pingmei Xu, Chase Roberts, Dmitry Lagun, and Vidhya Navalpakkam. On-device few-shot personalization for real-time gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[20] Daniel A Hofmaenner and Philipp K Buehler. The impact of eye-tracking on patient safety in critical care. Journal of Clinical Monitoring and Computing, 36(6):1577-1579, 2022. 1 +[21] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2 +[22] Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhandarkar, Wojciech Matusik, and Antonio Torralba. Eye tracking for everyone. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2176-2184, 2016. 2 +[23] Astar Lev, Yoram Braw, Tomer Elbaum, Michael Wagner, and Yuri Rassovsky. Eye tracking during a continuous performance test: Utility for assessing ADHD patients. Journal of Attention Disorders, 26(2):245-255, 2022. 1 +[24] Dongze Lian, Lina Hu, Weixin Luo, Yanyu Xu, Lixin Duan, Jingyi Yu, and Shenghua Gao. Multiview multitask gaze estimation with deep convolutional neural networks. IEEE transactions on neural networks and learning systems, 30 (10):3010-3023, 2018. 2, 3 +[25] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. Unpublished manuscript. 5 +[26] SI Lyapunov, II Shoshina, and IS Lyapunov. Tremor eye movements as an objective marker of driver's fatigue. Human Physiology, 48(1):71-77, 2022. 1 +[27] Nada Mohammed Murad, Lilia Rejeb, and Lamjed Ben Said. Computing driver tiredness and fatigue in automobile via eye + +tracking and body movements. Periodicals of Engineering and Natural Sciences, 10(1):573-586, 2022. 1 +[28] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In Proceedings of the European conference on computer vision (ECCV), pages 721-738, 2018. 2 +[29] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5 +[30] Thammathip Piumsomboon, Gun Lee, Robert W Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In 2017 IEEE symposium on 3D user interfaces (3DUI), pages 36-39. IEEE, 2017. 1 +[31] Jiawei Qin, Takuru Shimoyama, and Yusuke Sugano. Learning-by-novel-view-synthesis for full-face appearance-based 3d gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2022. 3, 5 +[32] Alessandro Ruzzi, Xiangwei Shi, Xi Wang, Gengyan Li, Shalini De Mello, Hyung Jin Chang, Xuong Zhang, and Otmar Hilliges. Gazenerf: 3d-aware gaze redirection with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9676-9685, 2023. 2 +[33] Yunjia Sun, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Cross-encoder for unsupervised gaze representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3702-3711, 2021. 2, 3, 6 +[34] Xin Yi, Yiqin Lu, Ziyin Cai, Zihan Wu, Yuntao Wang, and Yuanchun Shi. Gazedock: Gaze-only menu selection in virtual reality using auto-triggering peripheral menu. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 832-842. IEEE, 2022. 1 +[35] Yu Yu and Jean-Marc Odobez. Unsupervised representation learning for gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7314-7324, 2020. 2, 3 +[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4511-4520, 2015. 2 +[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-60, 2017. 2, 5, 7 +[38] Xuong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM symposium on eye tracking research & applications, pages 1-9, 2018. 6 +[39] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In European Conference on Computer Vision, pages 365-381. Springer, 2020. 1, 2, 5 \ No newline at end of file diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/images.zip b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f034ef701218be32d6251d1036db857b7ec2a75 --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22abb1a76687c3780946e77139d0cebfa1a3bdfaa531c0e3f667484d5c246f7e +size 620337 diff --git a/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/layout.json b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..31a3a34bca41f2363675208d955e6b39f478277a --- /dev/null +++ b/2024/Unsupervised Gaze Representation Learning from Multi-view Face Images/layout.json @@ -0,0 +1,7322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "text", + "content": "Unsupervised Gaze Representation Learning from Multi-view Face Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 143, + 515, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 143, + 515, + 185 + ], + "spans": [ + { + "bbox": [ + 78, + 143, + 515, + 185 + ], + "type": "text", + "content": "Yiwei Bao Feng Lu* \nState Key Laboratory of VR Technology and Systems, School of CSE, Beihang University \n{baoyiwei, lufeng}@buaa.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": "Annotating gaze is an expensive and time-consuming endeavor, requiring costly eye-trackers or complex geometric calibration procedures. Although some eye-based unsupervised gaze representation learning methods have been proposed, the quality of gaze representation extracted by these methods degrades severely when the head pose is large. In this paper, we present the Multi-View Dual-Encoder (MV-DE), a framework designed to learn gaze representations from unlabeled multi-view face images. Through the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy, the MV-DE successfully disentangles gaze from general facial information and derives gaze representations closely tied to the subject's eyeball rotation without gaze label. Experimental results illustrate that the gaze representations learned by the MV-DE can be used in downstream tasks, including gaze estimation and redirection. Gaze estimation results indicate that the proposed MV-DE displays notably higher robustness to uncontrolled head movements when compared to state-of-the-art (SOTA) unsupervised learning methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 128, + 512 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 520, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 687 + ], + "type": "text", + "content": "Vision is one of the most important sense for humans. Human gaze reveals the direction of visual attention, which is an important cue for understanding how humans perceive the surrounding world. Thus, gaze estimation techniques have become an vital tool in numerous applications, such as Virtual Realty and Augmented Reality[2, 30, 34], automotive safety [13, 26, 27] and healthcare [3, 20, 23]. In recent years, appearance-based gaze estimation methods have drawn a lot of attention, since these methods only require simple web cameras, eliminating the need for expensive eye trackers with dedicated devices such as infrared cameras. Among these methods, Convolutional Neural Networks (CNN) based approaches exhibit exceptional performance in unconstrained environments." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 213, + 545, + 331 + ], + "blocks": [ + { + "bbox": [ + 308, + 213, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 308, + 213, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 545, + 331 + ], + "type": "image", + "image_path": "ec4df738f571d5cbf031c0bc8d3afe79f18ae2ad4b8c9653e838d5bae6114280.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "type": "text", + "content": "Figure 1. We propose the Multi-View Dual-Encoder, a face-based unsupervised gaze representation learning framework that is robust to large head poses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 401, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 568 + ], + "type": "text", + "content": "CNN-based gaze estimation methods are usually trained in an end-to-end manner with substantial amount of labeled data. The performance of CNN-based gaze estimation methods highly relies on the quantity and diversity of the training dataset. Unfortunately, annotating gaze directions is difficult. Unlike common Computer Vision tasks like Object Detection, gaze cannot be reliably annotated without specialized hardware. Gaze direction is either measured by costly eye trackers under controlled environment [29], or obtained by the line connecting the 3D face center to the 3D position of the gaze target [39]. It takes great effort and time to collect diverse labeled training data, since the participants need to stare at large amount of gaze targets as instructed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "To address the challenge of annotation, a number of unsupervised representation learning approaches have been proposed [4, 14, 19]. These approaches extract common visual representations from the input image without annotation. However, such methods do not perform well in gaze estimation, as gaze estimation differs significantly from common visual tasks. Common visual tasks like object classification and detection require representations of overall appearance of the subject. On the other hand, gaze is an direction vector that manifested as rotations of the subtle eye structure in the image. Efforts have been made to design unsupervised gaze representation learning methods" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "text", + "content": "*Corresponding Author. This work was supported by the National Natural Science Foundation of China (NSFC) under Grant 62372019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1419" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "[33, 35], which incorporate cropped eye images as input to eliminate undesired visual contents. However, existing unsupervised gaze representation learning methods could only handle samples within limited head pose range, since the eye appearance changes dramatically when the head rotation approaches " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 300 + ], + "type": "text", + "content": "Recent supervised gaze estimation methods usually employ face images as their input [1, 6, 8, 21, 37] to handle samples with large head poses. Unfortunately, the rich information in the face images becomes a double-edged sword in unsupervised settings. These pieces of information become interference due to the absence of gaze label as constraint. For example, head pose and gaze are strongly coupled. They behave exactly the same in common data augmentation methods such as image flipping and rotation, makes it difficult to separate them even for latest contrastive learning approaches. Thus, unsupervised gaze representation learning method with face images is a challenging task that remain to be solved." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 286, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 286, + 421 + ], + "type": "text", + "content": "A potential solution to tackle this challenge is to harness the extra constraint introduced by multi-view settings. Multi-view gaze estimation is a recent hot research topic. A number of supervised multi-view gaze estimation methods [7, 24] have been proposed. Multi-view image pairs provide a set of images with the same eyeball rotation, i.e. gaze direction within the Head Coordinate System (HCS), as they are captured simultaneously. Our aim is to design a face-based multi-view unsupervised gaze representation learning method that works well on samples with large head poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 286, + 624 + ], + "type": "text", + "content": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework that employs face images as input. The proposed MV-DE framework extracts gaze representation by separating eyeball rotation from general facial information. The MV-DE framework consists of two encoders: the Gaze Encoder and the Face Encoder. First, we train the Face Encoder to extract face representations including head pose and appearance, while excluding gaze information. Then, we freeze the Face Encoder and introduce the gaze feature extracted by the Gaze Encoder to compensate for the missing gaze information. Based on the consistent eyeball rotation across different views within the same frame, we integrate the face representations with gaze representations from different views to reconstruct the original image. At the inference time, the MV-DE extracts gaze representations from single-view face images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 625, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 286, + 696 + ], + "type": "text", + "content": "Experimental results illustrate that the MV-DE framework successfully derives gaze representations across a wide range of head poses. The learned gaze representation can be used for varies downstream tasks such as gaze estimation and gaze redirection. The contribution of this paper are summarized in three folds:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "content": "- We propose the Multi-View Dual-Encoder, a face-based" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 72, + 544, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 544, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 544, + 106 + ], + "type": "text", + "content": "unsupervised gaze representation learning framework under uncalibrated multi-view settings. The MV-DE is robust to free head movements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 108, + 545, + 192 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 108, + 544, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 544, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 544, + 144 + ], + "type": "text", + "content": "- We uncouple gaze from other interference such as head pose by the proposed Dual-Encoder architecture and the multi-view gaze representation swapping strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 144, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 192 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness of the MV-DE framework in unconstrained environments. Qualitative analyses prove that the extracted gaze representation is disentangled from head pose and appearance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 204, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 224, + 460, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 460, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 460, + 236 + ], + "type": "text", + "content": "2.1. Supervised Gaze Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 469 + ], + "type": "text", + "content": "Appearance-based gaze estimation approaches aim to estimate gaze from eye or face appearance directly. Early methods estimate gaze directions from eye images [9, 28, 36]. Zhang et al. first propose to utilize full face images and outperform eye-image-based methods [37]. Since then, most CNN-based gaze estimation methods employ face images as input. Chen et al. propose to utilize dilated convolution for gaze estimation [5]. Cheng et al. propose to employ the Transformer architecture for gaze estimation [6]. Some recent methods utilize both face and eye images as input for better estimation performance [1, 8, 17, 22]. Above methods are trained in an end-to-end manner with gaze annotations. A number of gaze estimation datasets are published. These datasets are collected under different scenario with varies devices, including web cameras [11, 37], " + }, + { + "bbox": [ + 304, + 243, + 545, + 469 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 243, + 545, + 469 + ], + "type": "text", + "content": " cameras [21], high-resolution cameras [39], eye trackers [29] and mobile devices [22]. With sufficient labeled data, gaze feature disentanglement is achieved by using GAN [10] and Nerf [32]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 479, + 460, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 460, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 460, + 491 + ], + "type": "text", + "content": "2.2. Multi-view Gaze Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 628 + ], + "type": "text", + "content": "Before the development of Deep Learning methods, most conventional model-based methods utilize multiple cameras for gaze estimation [16]. Model-based methods reconstruct 3D eyeball models and obtain the gaze direction based on the optical geometry of human eye structures [15]. Model-based methods achieve remarkable estimation accuracy. However, these methods have high requirements for the shooting angle of cameras. Thus, these methods are usually employed in the Head Mounted Devices, such as Meta Quest Pro and Microsoft Hololens, where the cameras are approximately stationary relative to the eye of users." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Multi-view settings have also drawn a lot of attention in CNN-based gaze estimation methods. A few multiview gaze estimation datasets have been proposed. Zhang et al. propose the ETH-XGaze dataset, which employs 18 high-resolution cameras to capture face images with large head pose range [39]. Park et al. propose the EVE dataset [29]. They use 3 web cameras and a industrial camera to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1420" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "capture user face images. Qin et al. propose to utilize 3D face alignment approaches to generate multiple rotated images from a given sample [31]. These rotated images can be regarded as samples from different virtual cameras. Based on these datasets, CNN-based multi-view gaze estimation approaches have been proposed. These methods integrate features from different views by concatenating [24] or self attention [7] to improve estimation accuracy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 280, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 280, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 280, + 188 + ], + "type": "text", + "content": "2.3. Unsupervised Gaze Representation Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 336 + ], + "type": "text", + "content": "Unsupervised representation learning has always been a hot topic in Computer Vision community. Recently, Contrastive Learning approaches have achieved satisfying performance in common Computer Vision tasks. Contrastive Learning methods generate multiple views of a given sample by different data augmentation methods and constrain the model to extract similar representation from these views [4, 14]. He et al. propose the Masked Autoencoders, a self-supervised learner trained by simply reconstructing masked images. However, these methods are designed to extract common visual representations, which do not perform well in the gaze estimation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 529 + ], + "type": "text", + "content": "To learn gaze representations in an unsupervised manner, Yu et al. propose to utilize the gaze redirection task to extract two-dimensional representations that relate to gaze pitch and yaw angles [35]. Sun et al. propose the CrossEncoder, which learn gaze representations by a latent-codeswapping mechanism on eye-consistent image pairs and gaze-similar pairs [33]. Gideon et al. further propose to adapt the Cross-Encoder architecture to the multi-view setting [12]. Above methods all employ eye images as their input, as there are less gaze-irrelevant visual contents. However, appearance of eyes changes dramatically as head pose increases, makes it difficult for these methods to handle samples with large head pose distribution. Our aim is to introduce an unsupervised gaze representation learning approach that employs face images as input, which is robust to unconstrained head movements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "We propose the Multi-View Dual-Encoder, a face-based multi-view unsupervised gaze representation learning method. Our aim is to extract gaze representations from the input face images without using the gaze label. The main challenge of this task is to separate gaze from the other facial information, especially head pose. Since gaze and head pose are both physical directions, common data augmentations like rotation and flipping are not able to separate them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "To achieve our goal, the proposed MV-DE framework introduces two encoders: a Face Encoder and a Gaze Encoder. We first train the Face Encoder to extract general facial representation including head pose and appearance, while excluding gaze information. Then, we freeze the Face" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 541, + 169 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 541, + 169 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 541, + 169 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 541, + 169 + ], + "type": "image", + "image_path": "1276427211a33707343175ed36949157e06ea13cd622fc3e9857089f8339e46c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 191, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 304, + 191, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 547, + 246 + ], + "type": "text", + "content": "Figure 2. Training process of the Face Encoder. The proposed Face Encoder extracts face representations that contain general facial information while excluding gaze information. It is trained by two tasks: the head pose estimation task and the image reconstruction task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "type": "text", + "content": "Encoder and train the Gaze Encoder to extract gaze representation which compensates for the missing gaze information. We constrain the Gaze Encoder to derive similar representation from different views within a given frame since the eyeball rotation is the same. In this way, we isolate gaze representation from other facial features without using the gaze label. In the following sections, we introduce the training pipeline of two encoders in detail." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 371, + 465, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 465, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 465, + 384 + ], + "type": "text", + "content": "3.1. Training of the Face Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "type": "text", + "content": "The target of the Face Encoder is to extract representation of general facial information except gaze. We design two training strategies to achieve this goal: eye masking and multi-task learning including head pose estimation and image reconstruction, as shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": "Given a training face image " + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{i,j}" + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": " is the frame index and " + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": " is the camera index, we mask the two eye areas by the average pixel value. We further add three random masks to prevent the Face Encoder from directly estimating head pose based on the position of eye masks. Given that input face images are normalized to " + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": " pixels, we set the size of masks to " + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "inline_equation", + "content": "55 \\times 33" + }, + { + "bbox": [ + 304, + 450, + 545, + 557 + ], + "type": "text", + "content": " according to the average size of eyes. Then, we input the masked face images to the Face Encoder for the proposed multi-task learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": "First, the Face Encoder extracts the face representation " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": " from the masked face image. Then, we add a regression MLP and a decoder for the head pose estimation task and the image reconstruction task, respectively. In the head pose estimation task, the target of the Face Encoder is to estimate the 3D Euler Angles of the subject's head pose. We use " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": " Loss function for the head pose estimation task: " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hp} = \\mathcal{L}_1(\\hat{\\boldsymbol{y}}_{i,j},\\boldsymbol{y}_{i,j})" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{y}}_{i,j}" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": " is the estimated head pose and " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}_{i,j}" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": " is the ground truth head pose. The purpose of the head pose estimation task is to ensure that the face representation " + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 304, + 558, + 545, + 690 + ], + "type": "text", + "content": " encodes head pose information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 690, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 690, + 545, + 713 + ], + "type": "text", + "content": "In the image reconstruction task, the Face Encoder and the decoder are trained in an adversarial way. The target" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1421" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 541, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 541, + 224 + ], + "type": "image", + "image_path": "8a5d3fd66b99129d7b31e54f96247b820f52d3d893c6857a34a67454ca0c58b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 546, + 257 + ], + "type": "text", + "content": "Figure 3. Training process of the Gaze Encoder. The Gaze Encoder extracts gaze representations that represent the eyeball rotation, i.e. gaze direction within HCS. The Gaze Encoder is trained by the proposed multi-view gaze representation swapping strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "text", + "content": "of the Face Encoder is to reconstruct the image with eyemasks. The target of the decoder is to reconstruct the original image from the face representation " + }, + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "text", + "content": ", including the eye area. Both the Face Encoder and the decoder are constrained by the Image Reconstruction Loss " + }, + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{re}" + }, + { + "bbox": [ + 46, + 270, + 287, + 331 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 340, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 340, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 78, + 340, + 287, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) = \\alpha_ {1} \\mathcal {L} _ {2} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}) + \\alpha_ {2} \\mathcal {L} _ {p c p t} (\\hat {\\boldsymbol {x}}, \\boldsymbol {x}), \\tag {1}", + "image_path": "25ea6f7d9d32b2f8d38c542161cf3cc96d215b0efa6373a53404f1f1a2b244a1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the reconstructed image, " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the target image, " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the Mean Squared Error (MSE) Loss function and " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pcpt}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the Perception Loss function. " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " are the coefficients of loss functions. The purpose of the image reconstruction tasks is to encode general facial information in the face representation while excluding gaze information. Images constructed from the " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " are basically the original face images without iris and pupil." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "text", + "content": "Overall, we train the Face Encoder by minimizing " + }, + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FE} = \\beta_1\\mathcal{L}_{hp} + \\mathcal{L}_{re}" + }, + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 458, + 287, + 517 + ], + "type": "text", + "content": " is coefficient to balance two tasks. Once the training of the Face Encoder is completed, we freeze the parameters of the Face Encoder and proceed to train the Gaze Encoder via the multi-view constraint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 525, + 209, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 209, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 209, + 538 + ], + "type": "text", + "content": "3.2. Training of the Gaze Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": "The target of the Gaze Encoder is to extract gaze representation " + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "inline_equation", + "content": "e_{i,j}^{g}" + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": " that represents the direction of gaze in the HCS without any gaze label. Since the face representation " + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": " encodes general face information except gaze, we combine the gaze representation " + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "inline_equation", + "content": "e_{i,j}^{g}" + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "inline_equation", + "content": "e_{i,j}^{f}" + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": " to fully reconstruct the original image, so that " + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "inline_equation", + "content": "e_{i,j}^{g}" + }, + { + "bbox": [ + 46, + 543, + 287, + 641 + ], + "type": "text", + "content": " compensates for the missing gaze information. The training strategy of the Gaze Encoder is shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Given a training sample " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i,j}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", we extract the gaze representation from the original image and the face representation from the eye-masked image. Then, we input both representations to the decoder to reconstruct the original image. To reconstruct the original image, the Gaze Encoder has to capture gaze information which is missing in the face repre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "type": "text", + "content": "sentation. Considering that the region occupied by two eyes represents only a small portion of the face image, we additionally compute the Image Reconstruction Loss for both eye regions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 328, + 545, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 328, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 344, + 328, + 545, + 349 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {i, j} = G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j} ^ {g}\\right), \\tag {2}", + "image_path": "b0aa255c3cf0438521c3067087a5d2e95fafa935f23a15ed0ea909e114a533b9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 347, + 509, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 347, + 509, + 361 + ], + "spans": [ + { + "bbox": [ + 343, + 347, + 509, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G E} = \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}) + \\mathcal {L} _ {r e} (\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}),", + "image_path": "9f1a883b4bf49f971db35bb8dfccafff83f95e2720d5d98f5cffde7f1ed81208.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "text", + "content": " is the decoder, " + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{i,j}^{e}" + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i,j}^{e}" + }, + { + "bbox": [ + 305, + 370, + 545, + 407 + ], + "type": "text", + "content": " are the two eye regions of the reconstructed image and the original image, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "content": "However, the gaze representation can also encode other facial information such as head pose during the training process. In an extreme case, the decoder might theoretically reconstruct the entire image solely based on the gaze representation, since the Gaze Encoder utilizes the full face image as input. We utilize the consistency of eyeball rotation in the multi-view settings to exclude gaze-irrelevant information. First, we randomly sample an image " + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i,j'}" + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "content": " from another view within the same frame where " + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "inline_equation", + "content": "j' \\neq j" + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "content": ". Since the eyeball rotations from both views are consistent, we swap the gaze representations of two views and reconstruct the eye regions of the original views. For a training image pair " + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{i,j}, \\pmb{x}_{i,j'}\\}" + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "content": ", the final loss function of the Gaze Encoder " + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GE}" + }, + { + "bbox": [ + 305, + 407, + 545, + 574 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 584, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 584, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 342, + 584, + 545, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G E} = \\beta_ {2} \\mathcal {L} _ {r e} ^ {\\text {f a c e}} + \\beta_ {3} \\mathcal {L} _ {r e} ^ {\\text {e y e s}} + \\beta_ {4} \\mathcal {L} _ {r e} ^ {\\text {s w a p}}, \\tag {3}", + "image_path": "c9c59181711e35206427409394523a392d47d41588b95e56947c4c2c50f05517.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\beta_{2},\\beta_{3},\\beta_{4}" + }, + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "text", + "content": " are the coefficients, " + }, + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{re}^{face}" + }, + { + "bbox": [ + 306, + 608, + 545, + 633 + ], + "type": "text", + "content": " is the Image Reconstruction Loss of full face images from both view:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 642, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 642, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 336, + 642, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} ^ {\\text {f a c e}} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j}, \\boldsymbol {x} _ {i, j}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}}, \\boldsymbol {x} _ {i, j ^ {\\prime}}\\right). \\tag {4}", + "image_path": "8651712632bf6144f6e4f0a39ee5b8d394f200a8b333ef8cf50f6d5db2e23203.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 666, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{re}^{eyes}" + }, + { + "bbox": [ + 306, + 666, + 545, + 689 + ], + "type": "text", + "content": " is the Image Reconstruction Loss of eye regions from both view:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} ^ {e y s} = \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j} ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\mathcal {L} _ {r e} \\left(\\hat {\\boldsymbol {x}} _ {i, j ^ {\\prime}} ^ {e}, \\boldsymbol {x} _ {i, j ^ {\\prime}} ^ {e}\\right). \\tag {5}", + "image_path": "dfe4774e0ebedd8bf82567b0c609b9ede459d37a5822ae1e3cce50b386218b58.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1422" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{re}^{swap}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the Image Reconstruction Loss of the generated eye regions after gaze representation swapping:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 101, + 287, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 101, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 93, + 101, + 287, + 123 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} ^ {s w a p} = \\mathcal {L} _ {r e} \\left(G \\left(\\boldsymbol {e} _ {i, j} ^ {f}, \\boldsymbol {e} _ {i, j ^ {\\prime}} ^ {g}\\right) ^ {e}, \\boldsymbol {x} _ {i, j} ^ {e}\\right) + \\tag {6}", + "image_path": "a4ba0b88a5bdb1773ac30a711a58f4d1e0a59f2b6f5f3333b8c2b6f5b05126a7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 120, + 238, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 120, + 238, + 136 + ], + "spans": [ + { + "bbox": [ + 131, + 120, + 238, + 136 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} (G (\\pmb {e} _ {i, j ^ {\\prime}} ^ {f}, \\pmb {e} _ {i, j} ^ {g}) ^ {e}, \\pmb {x} _ {i, j ^ {\\prime}} ^ {e}).", + "image_path": "3d77483e4cbe82e0dde911650a23e532c427386f16fb2e15789468a942d0b310.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 140, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 199 + ], + "type": "text", + "content": "At test time, the Gaze Encoder extracts gaze representations from the given single-view face images, similar to common supervised single-view gaze estimation methods. We rotate the estimated HCS gaze to CCS by the head pose label for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 205, + 180, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 180, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 180, + 218 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": "Training Details: The proposed method is implemented by PyTorch using two RTX 3090 GPU. We employ the Adam optimizer with a leaning rate of " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " for all the encoders and decoders. The Face Encoder is trained for 13 epochs with a batch size of 100. Learning rate is decayed by 0.2 every 4 epochs. The Gaze Encoder is trained for 5 epochs with a batch size of 50. Learning rate is decayed by 0.2 every 2 epochs. We use the data rectification method from [37] and histogram equalization to normalize the input face images. " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "(\\alpha_{1},\\alpha_{2})" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " are set to (10,1) and (10,0.5) during the training of the Face Encoder and Gaze Encoder, respectively. " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " is set to 0.1 and " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "(\\beta_{2},\\beta_{3},\\beta_{4})" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " is set to (0.3,0.35,0.35). We use the first 6 convolutional layers of a ImageNet pretrained VGG-16 to calculate the Perceptual Loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": "Network Architecture: Both the Face Encoder and the Gaze Encoder use the ResNet-18 [18] as backbone. For the Face Encoder, we take the " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "(512*7*7)" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " feature map after the last residual block as the Face Representation. The output channels of the last Linear Layer is set to 3 for head pose estimation. For the Gaze Encoder, the Linear layers is replaced by a " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " Convolutionallayer to compress the " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "(512*7*7)" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " feature map to " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "(1*7*7)" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " gaze representation. In the decoder, the channel of the gaze representation is expanded to 10 by a " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " Convolutionallayer and then concatenated with the face representation. The decoder also has four residual blocks with target channel sizes of [256, 128, 63, 32]. The feature maps are up-scaled to twice their size before each residual block. We upscale the feature map again and use two Convolutional layers to generate the final " + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "inline_equation", + "content": "(3\\times 224\\times 224)" + }, + { + "bbox": [ + 46, + 391, + 288, + 642 + ], + "type": "text", + "content": " image. During the training of the Face Encoder, we set the gaze representation to zeros. Note that we train the decoder from scratch during the training of the Gaze Encoder, as the decoder has learned to ignore the 0 replaced gaze representation during the training of the Face Encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 151, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 151, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 151, + 684 + ], + "type": "text", + "content": "4.1. Data Preparation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We conduct experiments on four different gaze estimation datasets: ETH-XGaze [39], MPII-NV [31], EVE [29] and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 71, + 541, + 207 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 541, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 541, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 541, + 207 + ], + "type": "image", + "image_path": "2ad9c9523511db8eea09f9f50b7b468a82d042fa3c2be6f35c983209b12e2bec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 225, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 225, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 545, + 269 + ], + "type": "text", + "content": "Figure 4. Head pose (top row) and gaze direction distributions (bottom row) of 4 different datasets. The EVE and the EyeDiap datasets only provide very limited head pose and gaze range with frontal faces." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 285, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 312, + 285, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 312, + 285, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 285, + 545, + 434 + ], + "type": "image", + "image_path": "8229af6ef1af697d193bafbe8f5648567464d8f4e11356e5358083355d725502.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 455, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 545, + 487 + ], + "type": "text", + "content": "Figure 5. Image reconstruction results of combining face representations with different gaze representations in the ETH-XGaze dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 556 + ], + "type": "text", + "content": "EyeDiap [11]. We visualize the head pose and gaze distribution of each dataset in Fig. 4. The distribution of these datasets varies significantly. below we introduce each dataset in detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "ETH-XGaze " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_E)" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": ": A multi-view dataset collected by 18 high-resolution cameras. It provides very large head pose and gaze distributions. ETH-XGaze dataset contains 80 subjects, results in over 750,000 images in total. We divide the last 5 subjects as the labled test set and the first 75 subjects as the multi-view unsupervised training set. We follow [25] to optimize the gaze and head pose label to ensure multi-view consistency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "MPII-NV " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_M)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ": A synthesized multi-view dataset. We follow [31] to reconstruct 3D faces from MPIIFaceGaze [37] dataset and rotate every 3D face to generate 18 different views, referring to the setting of ETH-XGaze. MPII-NV dataset contains 15 subjects and 359,984 images in total." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1423" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 96, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Gaze estimation error of few-shot experiments in degrees. Bold numbers are the best results and underline numbers are the second best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 96, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 48, + 96, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 547, + 198 + ], + "type": "table", + "html": "
DatasetETH-XGazeMPII-NVEVEAverage
Calibration Num50100200Avg.50100200Avg.50100200Avg.
BYOL[14]25.6726.0425.2925.6710.1912.6511.2611.3710.369.749.749.9515.66
SimCLR[4]26.8124.9823.8625.2212.5512.1312.1412.2712.5311.3611.4811.7916.43
Cross-Encoder[33]16.2314.7214.5415.1611.4711.3611.7511.538.077.817.487.7911.49
DE (ours w/o MV)23.3219.9317.7620.3411.6410.7310.7411.0410.799.739.449.9913.79
MV-DE (ours)8.668.087.778.177.286.526.396.739.188.68.688.827.91
Gain▼ 7.57▼ 6.64▼ 6.77▼ 6.99▼ 2.91▼ 4.21▼ 4.35▼ 4.31▲ 1.11▲ 0.79▲ 1.20▲ 1.03▼ 3.58
", + "image_path": "5735b2ddf355088be6d462f6580ee02f1d68692edd0ab3380d4af5c9c0998d65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 214, + 282, + 340 + ], + "blocks": [ + { + "bbox": [ + 53, + 214, + 282, + 340 + ], + "lines": [ + { + "bbox": [ + 53, + 214, + 282, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 282, + 340 + ], + "type": "image", + "image_path": "8f255b2f83446306ebc5c21857e12ddde86ec889cfa975e5b136ff728a1c57ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 358, + 288, + 393 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 288, + 393 + ], + "type": "text", + "content": "Figure 6. Gaze estimation error relative to the gaze yaw angle in the ETH-XGaze dataset. The proposed MV-DE performs significantly better when the gaze yaw angle is large." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 412, + 230, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 230, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 230, + 424 + ], + "type": "text", + "content": "We use the last subject as the labeled test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 509 + ], + "type": "text", + "content": "EVE " + }, + { + "bbox": [ + 46, + 425, + 287, + 509 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_V)" + }, + { + "bbox": [ + 46, + 425, + 287, + 509 + ], + "type": "text", + "content": ": A multi-view dataset with 4 camera views. As shown in the Fig. 4, the EVE dataset only provides frontal faces with limited head pose and gaze distribution. We use the training and testing set split as defined in the original dataset and sample 3 images per second from the original videos, result in 527,896 images for unsupervised training and 64,464 images for testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "text", + "content": "EyeDiap " + }, + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_D)" + }, + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "text", + "content": ": A single-view dataset with limited head pose and gaze distribution with 16,674 images in total. We only use the EyeDiap in the cross-domain experiments. Thus, we use the whole dataset for training or testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 593 + ], + "type": "text", + "content": "We follow [38] to normalize the face images. We employ Histogram Equalization and normalize the pixel values to " + }, + { + "bbox": [ + 46, + 556, + 287, + 593 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 556, + 287, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 271, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 271, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 271, + 612 + ], + "type": "text", + "content": "4.2. Evaluation of the Learned Representations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "In this section, we validate the effectiveness of the learned face and gaze representations through image reconstruction. The propose MV-DE is supposed to decouple representations of the eyeball rotation from general facial information. In Fig. 5, we combine the face representation with different gaze representations and reconstruct the face image for validation. As shown in row 2, the decoder effectively reconstructs the input face image from the original face and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 348 + ], + "type": "text", + "content": "gaze representation. When replacing the original gaze representation with gaze representation from another view, the reconstruction results are similar with the original one. It proves that the MV-DE successfully learned the multi-view consistency of the eyeball rotation. In the last row, the gaze representation is replaced by zero vectors. The head pose and appearance of the reconstructed images are almost identical with the original image, while the eyeball rotations are rather random. These results confirm that the MV-DE successfully separate eyeball rotation from general facial information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 358, + 536, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 536, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 536, + 371 + ], + "type": "text", + "content": "4.3. Application of the MV-DE: Gaze Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": "The learned gaze representation can be used for gaze estimation by adding a MLP head calibrated under the few-shot learning setting. The estimation error is also an evaluation of the learned gaze representation. In the MV-DE, the MLP takes the gaze representation and the head pose estimation as input. The estimated head pose is first encoded to a 30 dimensional embedding and then concatenated with the gaze representation through 3 linear layers with target dimensions of (64, 64, 2). The MLP head is trained to estimate the HCS gaze. Then, we rotate the HCS gaze by head rotation to obtain CCS gaze. Sigmoid function is employed in the activation layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": "In Tab. 1, we compare the MV-DE with 3 SOTA unsupervised learning methods. BYOL [14] and SimCLR [4] are two SOTA contrastive learning methods. We modify them to adapt the multi-view gaze representation learning task. We employ face images from different camera views in the same frame instead of data augmentation methods to generate the positive pairs. Cross-Encoder [33] is a SOTA eye-based unsupervised gaze representation learning method. Gideon et al. also propose an eye-based method [12], but their method requires sample pairs with different eyeball rotation while keeping the head stable within a short video clip. Since such sample pairs are not available in " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_M" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": ", we exclude their moethod in comparison. We also remove the multi-view constraint in the MV-DE as an ablation study (named DE). Without the multi-view constraint, the Gaze Encoder is trained by the image reconstruc" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1424" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 102, + 287, + 247 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 94 + ], + "type": "text", + "content": "Table 2. Estimation error of different designs of the MLP head with 200 calibration samples." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 102, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 102, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 287, + 247 + ], + "type": "table", + "html": "
DimensionMLP inputs & outputsD_ED_MD_V
d=32eg→CCS Gaze24.3442.829.83
eg+HP→CCS Gaze20.5942.299.57
eg→HCS Gaze8.8910.498.66
eg+HP→HCS Gaze8.637.268.60
d=64eg→CCS Gaze22.2741.709.98
eg+HP→CCS Gaze18.1134.569.55
eg→HCS Gaze8.437.448.68
eg+HP→HCS Gaze7.846.398.67
d=256eg→CCS Gaze20.9740.9910.32
eg+HP→CCS Gaze18.8834.0510.39
eg→HCS Gaze8.076.508.72
eg+HP→HCS Gaze7.595.718.87
", + "image_path": "e32304ba061a793fc9da458d60918a11dc1bba4f2a5d00d548343d62e068f6be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 269, + 168, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 168, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 168, + 281 + ], + "type": "text", + "content": "tion task described in Eq. (2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 341 + ], + "type": "text", + "content": "Overall, the average accuracy of the MV-DE outperforms other SOTA methods significantly. BYOL and SimCLR perform worse than the gaze-specialized CrossEncoder and the MV-DE, proves that common visual representations are not suitable for the gaze estimation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": "Most importantly, the MV-DE outperforms the CrossEncoder significantly in the ETH-XGaze and MPII-NV datasets, where the head pose distributes across a wide range. On the other hand, the MV-DE achieves comparable performance with the Cross-Encoder in the EVE dataset, because the EVE dataset only contains frontal face images with limited head pose range." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "To further investigate the performance gap between the Cross-Encoder and the MV-DE, we visualize the gaze estimation error relative to the gaze yaw angle within the ETH-XGaze dataset in Fig. 6. Obviously, the estimation error of the Cross-Encoder increases severely when the gaze yaw angle approach " + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "inline_equation", + "content": "100^{\\circ}" + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": ". Samples with such extreme gaze yaw angle are generally associated with large head poses. Above results prove the point that the MV-DE are more robust to large head poses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "content": "Results from the last 2 rows demonstrate the importance of the proposed multi-view gaze representation swapping strategy. The multi-view constraint brought a improvement as large as " + }, + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "5.88^{\\circ}" + }, + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "content": ". But the Dual-Encoder (DE) still achieves better overall performance than SOTA constrastive learning methods, primary attributes to the proposed dual-encoder architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 257, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 257, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 257, + 646 + ], + "type": "text", + "content": "4.3.1 Ablation Study: Design of the MLP Head" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "The MLP head is responsible for estimating gaze from the learned gaze representations under the few-shot learning setting. In Tab. 2, we verify the performances of different MLP designs with 200 calibration samples. For example, " + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "[d = 64, e_g + HP \\rightarrow HCSGaze]" + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": " denotes that the MLP" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 135, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "type": "text", + "content": "Table 3. Cross-dataset gaze estimation errors of using the MV-DE as an unsupervised pretrain method. We first pretrain the model on the ETH-XGaze or the MPII-NV dataset unsupervisely. Then, we use the labeled training dataset to calibrate the MLP head." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 135, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 308, + 135, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 135, + 547, + 224 + ], + "type": "table", + "html": "
MethodUnsupervisedTrainTest Dataset
PretrainDatasetETH-XgazeMPII-NVEyeDiap
Baseline-EVE31.2732.5212.55
MV-DEETH-XGazeEVE19.67 ▼ 37.10%10.45▼ 67.87%11.36 ▼ 9.47%
MV-DEMPII-NVEVE25.42 ▼ 18.71%10.33 ▼ 68.23%11.58 ▼ 7.73%
ETH-XgazeMPII-NVEVE
Baseline-EyeDiap44.0639.4623.58
MV-DEETH-XGazeEyeDiap25.15 ▼ 42.92%12.11 ▼ 69.31%10.4 ▼ 55.89%
MV-DEMPII-NVEyeDiap25.58 ▼ 41.94%11.51 ▼ 70.83%10.58 ▼ 55.13%
", + "image_path": "c15153e5600cd6639e8175a779be45bb5c052c3e96264685bf62082c7fc3940a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": "includes 64-dimensional hidden layers, takes gaze representations and head poses as the input vector, and predicts gaze directions within Head Coordinate System (HCS). Note that in the MV-DE, we rotate the estimated HCS gaze with head pose label physically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "text", + "content": "Results of Tab. 2 lead to 3 conclusions: (1) The MLP performs significantly better on predicting HCS gaze, since the learned gaze representations represent eyeball rotation instead of the CCS Gaze. Theoretically, combination of " + }, + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "inline_equation", + "content": "e_{g} + HP" + }, + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "text", + "content": " should be enough to predict the CCS gaze, but the estimation error is still huge, probably due to the limited number of calibration samples and MLP parameters. (2) Head pose seems helpful for predicting the HCS gaze. This observation is consistent with previous research [37] that the additional information of the full face images helps in the gaze estimation task. (3) 256-dimensional hidden layers achieve the best performance in the 200-shot settings. But larger number of parameters require more calibration samples, the performance of 256-dimensional hidden layers degrades in 50 and 100 shot calibration settings. Considering the balance between the number of parameters and required calibration samples, we choose " + }, + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "inline_equation", + "content": "d = 64" + }, + { + "bbox": [ + 304, + 305, + 546, + 509 + ], + "type": "text", + "content": " for the MV-DE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 550 + ], + "type": "text", + "content": "4.3.2 Additional Experiments: MV-DE as an Unsupervised Pretrain Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "Another potential application of the MV-DE is to use as an unsupervised pretrain method, since it is more cost-effective to collect large number of unlabeled face images than labeled ones. In Tab. 3, we employ the MV-DE to pretrain the Gaze Encoder without gaze label in the ETH-XGaze or the MPII-NV datasets. Then, we freeze the Gaze Encoder and only train the MLP head in the EVE or the EyeDiap dataset with gaze label." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "To establish a baseline, we directly train the Gaze Encoder and the regression MLP in an end-to-end manner with gaze label in the labeled training dataset, just like common supervised gaze estimation methods. The cross-dataset errors of the baseline method are extremely large on the ETH-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1425" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 190 + ], + "type": "image", + "image_path": "52e4c07bb8017484a4de72094b875728ce102f1c8a14428f8d349d34544887ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 288, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 288, + 257 + ], + "type": "text", + "content": "Figure 7. Distributions of cross-domain gaze estimation error with (right) and without (left) the proposed unsupervised pretrain on the ETH-XGaze. Blue rectangle indicates the gaze distribution range of the labeled training dataset (EyeDiap)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 266, + 280, + 445 + ], + "blocks": [ + { + "bbox": [ + 58, + 266, + 280, + 445 + ], + "lines": [ + { + "bbox": [ + 58, + 266, + 280, + 445 + ], + "spans": [ + { + "bbox": [ + 58, + 266, + 280, + 445 + ], + "type": "image", + "image_path": "b53fd10394894cae9d6784fb2cce1ab00c90d5e5bfd0eec807b183b3723681e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 464, + 287, + 498 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 498 + ], + "type": "text", + "content": "Figure 8. Results of using the MV-DE for gaze redirection. Spheres on the bottom-right corner indicates the gaze direction ground truth within the HCS, i.e. the eyeball rotation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 515, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 289, + 647 + ], + "type": "text", + "content": "XGaze and MPII-NV dataset, since the gaze distributions of the training sets are significantly smaller than the test sets. When used as a unsupervised pretrain method, the MV-DE effectively improves the cross-dataset performance of the baseline model. The distributions of estimation errors in Fig. 7 demonstrate that with the pretrain of the MV-DE, the model generates reasonable estimations in a significantly larger range. Above results indicate that the MV-DE can be used as a pretrain method to improve cross-dataset performances when large amount of unlabeled face images are available." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 281, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 281, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 281, + 671 + ], + "type": "text", + "content": "4.4. Application of the MV-DE: Gaze Redirection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The learned representations of the MV-DE can also be used for the gaze redirection task. Results in Fig. 8 show that the MV-DE effectively redirects the gaze direction accord" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 200 + ], + "type": "image", + "image_path": "181f90429d1f451985d8b251c349af38e4053dc4f1e5ec651c1171466b09f0b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 209, + 546, + 254 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 546, + 254 + ], + "type": "text", + "content": "Figure 9. Gaze redirection results of combining different gaze representations linearly. Results demonstrate that the gaze representations exhibit good linearity invariance. Spheres on the bottom-right corner indicate the ground truth or the target eyeball rotation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 334 + ], + "type": "text", + "content": "ing to the HCS gaze of the reference images, while keeping the head pose and identity unchanged. Fig. 8 demonstrate that the learned gaze representation is subject-independent. The head pose and identity of the reconstructed images are controlled by the face representations, where the eyeabll rotation is controlled by the gaze representations as expected." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 342, + 544, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 544, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 544, + 355 + ], + "type": "text", + "content": "4.4.1 Linearity Invariance of the Gaze Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 540 + ], + "type": "text", + "content": "Gaze, as a physical direction vector, conforms to the principles of additivity. Ideally, well-learned gaze representations should possess similar characteristics. In Fig. 9, we combine gaze representations from different samples linearly with different ratios to reconstruct face images. The sphere in the bottom-left corner of reconstructed the images indicates the target eyeball rotation angle, which corresponds to the linear combination of eyeball rotations from the HCS gaze source images. Results demonstrate that the learned gaze representation satisfies Linearity Invariance, i.e. the eyeball rotation of reconstructed images undergo the same linear combination as the input gaze representations. Experiments in Fig. 9 prove that the MV-DE successfully captures and retains part of the physical properties of gaze during the unsupervised learning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 551, + 379, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 379, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 379, + 563 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "In this paper, we present the Multi-View DualEncoder (MV-DE), an unsupervised gaze representation learning framework based on multi-view face images. We propose the Dual-Encoder architecture and the multiview gaze representation swapping strategy to learn gaze representations that are separated from general facial information without gaze label. Experiments show that the learned gaze representations can be used for downstream tasks like gaze estimation and gaze redirection. Gaze estimation results show that the quality of learned gaze representations from the MV-DE are significantly better than other SOTA methods with unconstrained head movements." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1426" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Yiwei Bao, Yihua Cheng, Yunfei Liu, and Feng Lu. Adaptive feature fusion network for gaze tracking in mobile tablets. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9936-9943. IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Yiwei Bao, Jiaxi Wang, Zhimin Wang, and Feng Lu. Exploring 3d interaction with gaze guidance in augmented reality. In 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 22-32. IEEE, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 289, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 289, + 258 + ], + "type": "text", + "content": "[3] Ariella Fornachari Ribeiro Belan, Marcos Vasconcelos Pais, Marina von Zuben de Arruda Camargo, Livea Carla Fidalgo Garcez Sant'Ana, Marcia Radanovic, and Orestes Vicente Forlenza. Diagnostic performance of an eye-tracking assisted visual inference language test in the assessment of cognitive decline due to alzheimer's disease. Journal of Alzheimer's Disease, (Preprint):1-15, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 338 + ], + "type": "text", + "content": "[5] Zhaokang Chen and Bertram E Shi. Appearance-based gaze estimation using dilated-convolutions. In Asian Conference on Computer Vision, pages 309–324. Springer, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 373 + ], + "type": "text", + "content": "[6] Yihua Cheng and Feng Lu. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3341-3347. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "text", + "content": "[7] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20632-20641, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 418, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 464 + ], + "type": "text", + "content": "[8] Yihua Cheng, Shiyao Huang, Fei Wang, Chen Qian, and Feng Lu. A coarse-to-fine adaptive network for appearance-based gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10623-10630, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 497 + ], + "type": "text", + "content": "[9] Yihua Cheng, Xuong Zhang, Feng Lu, and Yoichi Sato. Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing, 29:5259-5272, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 542 + ], + "type": "text", + "content": "[10] Yihua Cheng, Yiwei Bao, and Feng Lu. Puregaze: Purifying gaze feature for generalizable gaze estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 436-443, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[11] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 255–258, 2014. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[12] John Gideon, Shan Su, and Simon Stent. Unsupervised multi-view gaze representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5001-5009, 2022. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Mohamed Waheed Gomaa, Rasha O Mahmoud, and Amany M Sarhan. A cnn-lstm-based deep learning approach for driver drowsiness prediction. Journal of Engineering Research, 6(3):59-70, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Elias Daniel Guestrin and Moshe Eizenman. General theory of remote gaze estimation using the pupil center and corneal reflections. IEEE Transactions on biomedical engineering, 53(6):1124-1133, 2006. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "text", + "content": "[16] Dan Witzner Hansen and Qiang Ji. In the eye of the beholder: A survey of models for eyes and gaze. IEEE transactions on pattern analysis and machine intelligence, 32(3):478-500, 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 205, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 270 + ], + "type": "text", + "content": "[17] Junfeng He, Khoi Pham, Nachiappan Valliappan, Pingmei Xu, Chase Roberts, Dmitry Lagun, and Vidhya Navalpakkam. On-device few-shot personalization for real-time gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 414 + ], + "type": "text", + "content": "[20] Daniel A Hofmaenner and Philipp K Buehler. The impact of eye-tracking on patient safety in critical care. Journal of Clinical Monitoring and Computing, 36(6):1577-1579, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "text", + "content": "[21] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[22] Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhandarkar, Wojciech Matusik, and Antonio Torralba. Eye tracking for everyone. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2176-2184, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[23] Astar Lev, Yoram Braw, Tomer Elbaum, Michael Wagner, and Yuri Rassovsky. Eye tracking during a continuous performance test: Utility for assessing ADHD patients. Journal of Attention Disorders, 26(2):245-255, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[24] Dongze Lian, Lina Hu, Weixin Luo, Yanyu Xu, Lixin Duan, Jingyi Yu, and Shenghua Gao. Multiview multitask gaze estimation with deep convolutional neural networks. IEEE transactions on neural networks and learning systems, 30 (10):3010-3023, 2018. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 658 + ], + "type": "text", + "content": "[25] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. Unpublished manuscript. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[26] SI Lyapunov, II Shoshina, and IS Lyapunov. Tremor eye movements as an objective marker of driver's fatigue. Human Physiology, 48(1):71-77, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Nada Mohammed Murad, Lilia Rejeb, and Lamjed Ben Said. Computing driver tiredness and fatigue in automobile via eye" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1427" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "tracking and body movements. Periodicals of Engineering and Natural Sciences, 10(1):573-586, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 137 + ], + "type": "text", + "content": "[28] Seonwook Park, Adrian Spurr, and Otmar Hilliges. Deep pictorial gaze estimation. In Proceedings of the European conference on computer vision (ECCV), pages 721-738, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "type": "text", + "content": "[29] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 286, + 237 + ], + "type": "text", + "content": "[30] Thammathip Piumsomboon, Gun Lee, Robert W Lindeman, and Mark Billinghurst. Exploring natural eye-gaze-based interaction for immersive virtual reality. In 2017 IEEE symposium on 3D user interfaces (3DUI), pages 36-39. IEEE, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 286, + 293 + ], + "type": "text", + "content": "[31] Jiawei Qin, Takuru Shimoyama, and Yusuke Sugano. Learning-by-novel-view-synthesis for full-face appearance-based 3d gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2022. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 286, + 358 + ], + "type": "text", + "content": "[32] Alessandro Ruzzi, Xiangwei Shi, Xi Wang, Gengyan Li, Shalini De Mello, Hyung Jin Chang, Xuong Zhang, and Otmar Hilliges. Gazenerf: 3d-aware gaze redirection with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9676-9685, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 286, + 413 + ], + "type": "text", + "content": "[33] Yunjia Sun, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Cross-encoder for unsupervised gaze representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3702-3711, 2021. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[34] Xin Yi, Yiqin Lu, Ziyin Cai, Zihan Wu, Yuntao Wang, and Yuanchun Shi. Gazedock: Gaze-only menu selection in virtual reality using auto-triggering peripheral menu. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 832-842. IEEE, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 286, + 514 + ], + "type": "text", + "content": "[35] Yu Yu and Jean-Marc Odobez. Unsupervised representation learning for gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7314-7324, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 516, + 286, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 516, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 516, + 286, + 558 + ], + "type": "text", + "content": "[36] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4511-4520, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 613 + ], + "type": "text", + "content": "[37] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-60, 2017. 2, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 615, + 286, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 286, + 658 + ], + "type": "text", + "content": "[38] Xuong Zhang, Yusuke Sugano, and Andreas Bulling. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM symposium on eye tracking research & applications, pages 1-9, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 659, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 286, + 713 + ], + "type": "text", + "content": "[39] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In European Conference on Computer Vision, pages 365-381. Springer, 2020. 1, 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1428" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_content_list.json b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4dcc1017f0b696f1302028e882ecb1edc527d242 --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "Unsupervised Keypoints from Pretrained Diffusion Models", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Hedlin1 Gopal Sharma1 Shweta Mahajan1,2 Xingzhe He1 Hossam Isack3 Abhishek Kar3 Helge Rhodin1 Andrea Tagliaasacchi4,5,6 Kwang Moo Yi1 1 University of British Columbia 2 Vector Institute for AI 3 Google Research 4 Google DeepMind 5 Simon Fraser University 6 University of Toronto", + "bbox": [ + 148, + 178, + 818, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://stablekeypoints.github.io/", + "bbox": [ + 328, + 253, + 633, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cb4f3fa1d3077ec8e473cca9fcb28f10889daff91cb054896eb7458d1e116ba.jpg", + "image_caption": [ + "Figure 1. Teaser - we propose an unsupervised method to learn keypoints based on optimizing text embeddings of latent diffusion models [44]. Our method is motivated by the fact that random text tokens already respond roughly consistently to semantically similar regions. By promoting localization we obtain unsupervised keypoints that outperform the state-of-the-art." + ], + "image_footnote": [], + "bbox": [ + 86, + 276, + 879, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 558, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised learning of keypoints and landmarks has seen significant progress with the help of modern neural network architectures, but performance is yet to match the supervised counterpart, making their practicability questionable. We leverage the emergent knowledge within text-to-image diffusion models, towards more robust unsupervised keypoints. Our core idea is to find text embeddings that would cause the generative model to consistently attend to compact regions in images (i.e. keypoints). To do so, we simply optimize the text embedding such that the cross-attention maps within the denoising network are localized as Gaussians with small standard deviations. We validate our performance on multiple datasets: the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human3.6m datasets. We achieve significantly improved accuracy, sometimes even outperforming supervised ones, particularly for data that is non-aligned and less curated. Our code is publicly available at the project page.", + "bbox": [ + 75, + 611, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 558, + 630, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keypoints or landmarks have played a critical role in computer vision for various tasks including image matching [31], 3D reconstruction [18], and motion tracking [32, 60]. Similarly to many other areas of computer vision, research has quickly adopted supervised learning to tackle this problem [3, 27]. However, labeling is tedious and sometimes even ambiguous—for example, it is difficult to consistently decide which keypoints on a human face are the \"most important\". Researchers have therefore been investigating unsupervised approaches [12, 13, 19, 30, 55, 67]. These are typically implemented as autoencoders paired with handcrafted intermediate layers, or losses that enforce spatial locality and equivariance of keypoint locations under deformation. However, as we will show later, these methods struggle with non-preprocessed data, and their performance is heavily reliant on knowing the ground truth location of objects, clearly limiting their practical applicability.", + "bbox": [ + 496, + 590, + 890, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance the learning of unsupervised keypoints, we draw inspiration from the demonstrated success of scaling up datasets [52]. For example, in natural language pro", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/030a330169ab53748481597408ce0732f89700674297f49a5a01c43ffb13c6cf.jpg", + "image_caption": [ + "Figure 2. Example attention maps - we show example attention maps for a selected learned keypoint for the CUB-200-2011 dataset, on the CUB-aligned subset. As shown, our keypoint attention map responds consistently across varying images." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 885, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cessing performance has recently improved to a great extent thanks to large models and data [7, 39, 57]. Similarly, in computer vision, the performance of text-to-image models [43, 44, 46] has drastically improved thanks to the availability of extra large datasets [48]. However, unsupervised keypoint learning typically assumes class-specific datasets, e.g., animals that have a shared skeleton that connects keypoints, and these datasets are small in scale.", + "bbox": [ + 75, + 282, + 472, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rather than collecting larger domain-specific datasets, we instead propose to leverage the knowledge stored within large generative models, such as Stable Diffusion [44]. This has been shown to be very effective across a number of tasks [1, 2, 4, 8, 15, 24, 26, 33, 35, 41, 54, 56, 59, 61, 63, 64, 66], but, to the best of our knowledge, it has not yet found application for the task of keypoint learning. Our main idea is to localize \"important\" keypoints by finding text embeddings that consistently correspond to a distinct location in images of a certain object class. This idea is rooted in the observation that, even with random text embeddings, the attention maps for various images roughly correspond to regions that are semantically similar; see Fig. 1. Therefore, text embeddings carry semantic meaning, which could be used to relate collections of images to each other; see Fig. 2.", + "bbox": [ + 75, + 407, + 472, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find embeddings that are specific to certain locations by enforcing localized attention maps. In more detail, we propose to find (i.e., optimize) a set of tokens in a text embedding that locally responds in the Stable Diffusion cross-attention layers. We enforce locality by maximizing the similarity of the attention responses of each token to a single-mode Gaussian distribution. Thanks to the way the cross-attention layers are constructed within Stable Diffusion, this simple objective also prevents the different tokens from attending to the same locations in an image, a common degenerate solution that typically requires explicit workarounds [23].", + "bbox": [ + 75, + 638, + 472, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on established benchmarks: CelebA [28], CUB-200-2011 [58], Tai-Chi-HD [50], DeepFashion [29], and Human3.6m [17]. Our approach yields results on par with state-of-the-art methods for well-curated and aligned datasets, while notably enhancing performance", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for in-the-wild setups, particularly with unaligned data, sometimes even surpassing fully supervised baselines.", + "bbox": [ + 498, + 282, + 890, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 330, + 640, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Below we review the literature of finding keypoints in unsupervised and supervised fashion, along with work that exploits large pre-trained models like stable-diffusion for lower-level computer vision tasks.", + "bbox": [ + 498, + 356, + 893, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning keypoints with supervision. Pose estimation and landmark estimation are fundamental problems in computer vision. They naturally arise in various tasks, including human [69] and animal pose estimation [22], hand [5] and face landmark estimation [62], and object pose tracking [34]. Many fully supervised methods find different ways to induce some prior within the model to better capture the task at hand, such as using part affinity fields [3], temporal consistency for video data [45], spacial relationships [65], and geometry constraints [21] among others. While fully supervised methods have excelled in categories with abundant labeled data, such as human pose estimation, their major drawback is the insatiable need for large and high-quality datasets [22, 37, 42]. The scalability of gathering such extensive and meticulously annotated data for every conceivable object category remains a significant drawback [22, 37, 42].", + "bbox": [ + 496, + 425, + 893, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning keypoints via self-supervision. The amount of unlabeled data far exceeds that of labeled data, so unsupervised keypoint estimation methods attempt to take advantage of this. Self-supervised keypoint detection often relies on tracking how keypoints move with image changes and uses various constraints for known transformations [16, 19, 30, 49, 55, 67], but these methods can struggle with background modeling [49, 67] and pose variations [16]. One can also rely on image reconstruction to learn keypoints. Some methods use GANs to generate images from keypoints [12, 14], but this often results in training instability. Alternatively, auto-encoders can be also used [13, 67], but these require training from scratch on each dataset. Our method neither suffers from GAN", + "bbox": [ + 496, + 689, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22821", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training instability, nor requires dataset fine-tuning. Finally, there exist self-supervised methods that exploit skeletal representations [13, 20, 40]. However, many of these approaches generally require known keypoint connectivity and video data [20, 40], and often face limitations in background handling and generalizability to objects within the same class [13, 20, 40]. Our method has no object-specific priors, and generalizes well due to the large dataset used by the pre-trained diffusion models.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models for image understanding. Recently, large image diffusion models have reached impressive image generation quality [43, 44, 46]. These models learn priors for real images within the latent space of the diffusion model, and provide a useful initialization for many downstream tasks such as image correspondence [15, 33, 54, 66], object detection [4], semantic segmentation [2, 24, 56, 59, 61, 63, 64], and image classification [1, 8]. Interestingly, without requiring any retraining, these models demonstrate an innate ability to understand 3D spatial configurations [26, 35, 41]. Recent work in each of these areas has shown the emergent power of these large models, most of them using the model without any modifications or extra supervision required. More relevant to our work, Mokady et al. [36] found that the pre-trained Stable Diffusion [44] model's cross-attention maps connect text tokens to semantically relevant areas in images.", + "bbox": [ + 75, + 233, + 472, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Correspondences via diffusion models. Among works that re-purpose diffusion models, of high relevance, is the effectiveness of diffusion models in correspondence estimation tasks [15, 33, 54, 66]. Hedlin et al. [15] optimizes the attention map for a specific point in a source image and finds the corresponding activation in a target image. However, this method requires a query to be provided in the source image. While our method shares the same inspiration of utilizing attention maps, critically, rather than optimizing the embedding given a single image, we optimize an embedding given a dataset of images from a given object class. In other words, our method discovers on its own, where to focus, rather than relying on user input. Our task is therefore changed from image matching between two images, to semantic matching across all images within the dataset.", + "bbox": [ + 75, + 497, + 472, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 738, + 169, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To identify a set of representative keypoints across a dataset of images, we formulate our approach in an unsupervised framework leveraging conditional diffusion models; see Fig. 3. In particular, we utilize the cross-attention maps between the text embeddings and the image features, derived from the latent diffusion model [44], and force them to consistently concentrate their activation on highly localized regions within the images. While Hedlin et al. [15] employed a similar mechanism (i.e., given a set of keypoint locations", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in one image, identify correspondences in another image), in this work we seek to identify semantic correspondences across all images within a class-specific dataset (e.g., human faces), without any given knowledge on what and where to focus. We show that this is possible simply by enforcing locality and equivariance to transformations.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us start by quickly reviewing the fundamentals of diffusion models and formalizing the attention maps that we will utilize within these models (Sec. 3.1). We then detail the objectives used to learn the text embeddings that represent keypoints (Sec. 3.2) and discuss important implementation details (Sec. 3.3).", + "bbox": [ + 496, + 181, + 893, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attention maps in diffusion networks", + "text_level": 1, + "bbox": [ + 498, + 284, + 821, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are a class of generative models that approximate the data distribution by denoising a base (typically Gaussian) distribution [38]. A latent diffusion model operates on a latent representation $\\mathbf{z}$ rather than the image itself, with an encoder that maps an image $\\mathbf{X}$ into a latent $\\mathbf{z}$ , and a decoder that maps $\\mathbf{z}$ into $\\mathbf{X}$ . These models define a forward diffusion process, where the latent representation $\\mathbf{z}$ is gradually transformed into Gaussian noise over a series of $T$ time steps. The inverse process, over a sequence denoising steps $t = 1,\\dots,T$ predicts the latent noise $\\epsilon_{\\theta}(\\mathbf{z}_t,t)$ which was gradually added in each iteration in order to recover the original (latent) signal.", + "bbox": [ + 496, + 308, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we are interested in conditional diffusion models, and the explicit attentional relationship between the condition (i.e. text) and the generated outcome (i.e. image) that these models learn. Typically, diffusion models are made conditional on some text $\\mathbf{y}$ , by providing an embedding $\\mathbf{e} = \\tau_{\\theta}(\\mathbf{y})$ from a text encoder $\\tau_{\\theta}$ to the denoiser. They are then trained to optimize", + "bbox": [ + 496, + 489, + 890, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathbf {z}, t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}, t, \\mathbf {e}) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 609, + 890, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the denoiser $\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathbf{e})$ is typically implemented by a transformer architecture [38] involving a combination of self-attention and cross-attention layers. Of our interest here is the cross-attention layers that relate $\\mathbf{e}$ to $\\mathbf{z}_t$ , which we now formalize.", + "bbox": [ + 496, + 646, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, in the transformer part of the model, denote $\\Phi_l^c (\\cdot)$ and $\\Psi_l^c (\\cdot)$ as the $c$ -th head and the $l$ -th linear layers of the U-Net. We calculate the query as $\\mathbf{Q}_l^c = \\Phi_l^c (\\mathbf{z}_{t = 1})\\in \\mathbb{R}^{(H\\times W)\\times D_l}$ , and the key from the language embedding $\\mathbf{K}_l^c = \\Psi_l^c (\\mathbf{e})\\in \\mathbb{R}^{N\\times D_l}$ , where $N$ is the number of tokens, $C$ the number of heads in the transformer attention layer, $H$ and $W$ are image height and width at that specific layer in U-Net, and $D_{l}$ the dimensionality of the layer. Given query and key, the cross-attention map $\\mathbf{M}_l\\in \\mathbb{R}^{(H\\times W)\\times N}$ is then computed via softmax along the $N$ di", + "bbox": [ + 496, + 723, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We choose $t = 1$ where $T = 50$ steps via hyper-parameter tuning.", + "bbox": [ + 517, + 886, + 861, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22822", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/005224cd5e3204ccb776ddb962aebd704a81c8241103c6fe3c7660c71aebf81d.jpg", + "image_caption": [ + "Figure 3. Overview - we pass a randomly initialized text embedding into Stable Diffusion [44] and extract the attention maps. We then optimize the text embedding to have localized attention maps, by supervising them to become a single-mode Gaussian distribution, drawn at the location of their maxima. We also enforce attention maps to be transformation equivariant to small affine transformations on images. We repeat this process over a set of training images, which after optimization provides a set of $K$ keypoints." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 869, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mension, and average pooling along the $C$ dimension:", + "bbox": [ + 76, + 376, + 434, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) = \\mathbb {E} _ {c} \\left[ \\operatorname {s o f t m a x} _ {n} \\left(\\mathbf {Q} _ {l} ^ {c} \\cdot \\mathbf {K} _ {l} / \\sqrt {D _ {l}}\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 402, + 468, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As various layers of the U-Net exhibit distinct levels of semantic understanding, following Hedlin et al. [15], we collect this information by average pooling across a selection of layers:", + "bbox": [ + 76, + 439, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {M}} = \\mathbb {E} _ {l = 7.. 1 0} \\left[ \\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) \\right] \\in \\mathbb {R} ^ {(H \\times W) \\times N}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 511, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In what follows, to lighten the notation, we drop the attention mask arguments $(\\mathbf{e},\\mathbf{X})$ and write the attention map for the n-th token as $\\tilde{\\mathbf{M}}_n$ .", + "bbox": [ + 76, + 547, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Optimizing to find the keypoint embeddings", + "text_level": 1, + "bbox": [ + 76, + 604, + 449, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain a text embedding that can be used to locate keypoints, for each of them, we simply optimize for two objectives that respectively encourages localization and equivariance to geometric transformations. We thus write", + "bbox": [ + 76, + 627, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l i z e}} + \\lambda_ {\\text {e q u i v}} \\mathcal {L} _ {\\text {e q u i v}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 700, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we apply $\\lambda_{\\mathrm{equiv}} = 10$ to balance the two losses to be in a similar operating range. Equivariance is enforced in the typical form of learning to be invariant to transformations. We first quickly detail $\\mathcal{L}_{\\mathrm{equiv}}$ , and then discuss how we enforce localization, which is the core of our method.", + "bbox": [ + 76, + 729, + 468, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equivariance - $\\mathcal{L}_{\\mathrm{equiv}}$ . To ensure our model's attention mechanism remains consistent across different geometric transformations $\\mathcal{T}$ of the input, we use the typical equivariance loss [25]:", + "bbox": [ + 76, + 811, + 468, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e q u i v}} = \\mathbb {E} _ {n} \\| \\mathcal {T} ^ {- 1} \\left(\\mathbf {M} _ {n} (\\mathbf {e}, \\mathcal {T} (\\mathbf {X})) - \\mathbf {M} _ {n} (\\mathbf {e}, \\mathbf {X}) \\right\\| ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $\\mathcal{T}$ we simply utilize minor affine transformations. We use random rotations between $\\pm 15$ degrees, translations between $\\pm 0.25 \\times W$ , and scaling between $100 - 120\\%$ of the original image size.", + "bbox": [ + 498, + 376, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encouraging localization - $\\mathcal{L}_{\\mathrm{localize}}$ . We encourage localization by forcing $\\tilde{\\mathbf{M}}_n$ to be a single-mode Gaussian distribution located at its maximum. In more details, denoting the Gaussian image that shares the same maximum as $\\tilde{\\mathbf{M}}_n$ as $\\mathbf{G}_n$ , we write", + "bbox": [ + 498, + 443, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o c a l i z e}} = \\mathbb {E} _ {n} \\| \\tilde {\\mathbf {M}} _ {n} - \\mathbf {G} _ {n} \\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 526, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To create the Gaussian images $\\mathbf{G}_n$ , we first identify the spatial location exhibiting the maximal response within the heatmap corresponding to each token $n$ by taking the argmax:", + "bbox": [ + 498, + 553, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {n} = \\underset {w, h} {\\arg \\max } \\tilde {\\mathbf {M}} _ {n} [ h, w ]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 613, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then generate a Gaussian image; see Fig. 3:", + "bbox": [ + 500, + 642, + 812, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {n} = \\exp \\left(- \\frac {\\| \\mathbf {X Y} _ {\\mathrm {c o o r d}} - \\boldsymbol {\\mu} _ {n} \\| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 665, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}\\mathbf{Y}_{\\mathrm{coord}}$ is a tensor of image coordinates.", + "bbox": [ + 500, + 705, + 826, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Promoting mutual exclusivity. It is important to note that while $\\mathcal{L}_{\\mathrm{localize}}$ in (6) at first glance seem to only encourage localization, it also enforces $\\tilde{\\mathbf{M}}_n$ to be mutually exclusive for different $n$ because of the softmax operation in (2). Should multiple embeddings become similar, their attention responses in (2) $\\mathbf{Q}_l^c\\cdot \\mathbf{K}_l$ will become similar, resulting in the softmax of the attention map being a flat response (i.e. deviating from a Gaussian shape). In other words, (6) naturally enforces exclusivity with the help of (2).", + "bbox": [ + 498, + 728, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stabilizing optimization by working with a subset. We noticed in our experiments that attention maps $\\mathbf{M}$ for some", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22823", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens can be 'spread' for some images, e.g., due to occlusions, destabilizing optimization. We thus opt for a simple solution of looking into the top-K tokens that are local. Specifically, we apply our losses over $n \\in \\mathcal{N}(\\kappa)$ , which returns the $\\kappa \\in \\mathbb{N}$ entries with the most spatially localized heatmap responses2, as measured by KL divergence:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} (\\kappa) = \\operatorname {A r g T o p} _ {\\kappa} \\left\\{- \\mathrm {K L} \\left(\\mathbf {G} _ {n}, \\tilde {\\mathbf {M}} _ {n}\\right) \\right\\} _ {n = 1} ^ {N}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 191, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final keypoints. While $\\mathcal{L}_{\\mathrm{localize}}$ naturally enforces exclusivity, it does not guarantee a complete coverage of the object. Thus, after we finish optimizing, we refine the set of keypoints through furthest point sampling using the training images. Specifically, for each image we write:", + "bbox": [ + 75, + 228, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} = \\operatorname {F P S} _ {K} (\\{\\boldsymbol {\\mu} _ {i} \\mid i \\in \\mathcal {N} (\\kappa) \\}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 315, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ is the desired number of keypoints $K < \\kappa$ . Then, as the set $\\mathcal{K}$ differs in each image, we simply choose $K$ to tokens that appeared most frequently in $\\mathcal{K}$ within the training image set.", + "bbox": [ + 75, + 343, + 468, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 414, + 290, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-time ensembling. At inference time, as in common literature [3, 10, 15], rather than employing the attention map of the original image, we average the attention maps across multiple augmentations (we use the same transformations for test time augmentation as in (5)) of the image:", + "bbox": [ + 75, + 436, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {M}} \\leftarrow \\sum_ {i} \\mathcal {T} _ {i} ^ {- 1} \\left(\\mathbf {M} \\left(\\mathbf {e}, \\mathcal {T} _ {i} (\\mathbf {X})\\right)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 523, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upsampling attention maps. The attention maps in (2) are typically of low resolution. Specifically, as we use Stable Diffusion 1.5 [44], depending on the layer we extract the attention maps from, they are either $16 \\times 16$ or $32 \\times 32$ . We thus opt to upsample the query $\\mathbf{Q}$ via bicubic interpolation to achieve a standard resolution of $128 \\times 128$ . We have experimented with other upsampling techniques such as the commonly used bilinear sampling or a learned upsampler that is trained alongside, but a simple bicubic upsample was shown to be effective.", + "bbox": [ + 75, + 568, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 732, + 161, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 757, + 263, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on five standard datasets for unsupervised keypoint evaluation:", + "bbox": [ + 75, + 781, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CelebA dataset [28]: A dataset of 202,599 facial images of celebrities. We evaluate both the aligned and nonaligned cases following the standard protocol of omitting images with faces occupying less than $30\\%$ of the image.", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The standard metric for this dataset is to measure the average $\\ell_2$ error normalized by the inter-ocular distance.", + "bbox": [ + 511, + 90, + 888, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CUB-200-2011 dataset [58]: This dataset consists of 11,788 bird images. We use both the aligned (CUB-aligned) and non-aligned (CUB-all) variants. For the nonaligned variants, we further look at CUB-001, CUB-002, and CUB-003, which are specific bird subcategories. Notably, these subsets contain only 30 images each—we only use these 30 for training. We follow the standard protocol [6, 30] and normalize the images to be of $256 \\times 256$ . The standard metric for this dataset is the mean $\\ell_2$ error, normalized by the dimension of the images after normalization.", + "- Tai-Chi-HD dataset [50]: This dataset contains 3049 training videos and 285 test videos of people performing Tai-Chi, which shows more diverse poses compared to the other datasets, and is the most challenging among the human pose-centric datasets that we use. We follow Siarohin et al. [51] and use 500 images for testing and 300 images for training. The standard metric for this dataset is to measure the accumulated $\\ell_2$ error, with the images standardized to $256 \\times 256$ .", + "- DeepFashion dataset [29]: This dataset contains 53k images of fashion models, mostly standing with a white background. We follow Lorenz et al. [30] and only keep full body images. This leaves 10,604 images for training and 1,179 images for testing. Also following the baselines, we use keypoints generated by AlphaPose [11] as ground truth. The standard metric for this dataset is the percentage of correct keypoints (PCK) with a 6-pixel threshold.", + "- Human 3.6M dataset [17]: This dataset is of humans performing various actions, comprised of 3.6 million images. We follow the standard protocol [67] and focus on six activities: direction, discussion, posing, waiting, greeting, and walking. We utilize subjects 1, 5, 6, 7, 8, and 9 for training, while subject 11 is reserved for testing. This division yields a training dataset comprising 796,648 images and a testing dataset containing 87,975 images. The background for this dataset is also simple, and often masked out with ground-truth masks for evaluation. This dataset is also typically heavily pre-processed and aligned when used for unsupervised keypoint evaluation. We experiment with the standard pre-processing [30, 67] and also a relaxed version of our own. To relax the alignment, we crop a square bounding box such that the margin from the bounding box to the person is 100 pixels, which on average corresponds to the person's height being $2/3$ of the crop. We further add a uniform random translation up to 100 pixels (same as the margin) to remove the central bias. Example crops are visualized in Fig. 4f. The standard metric for this dataset is the $\\ell_2$ error after normalizing the image resolution to $128 \\times 128$ ." + ], + "bbox": [ + 500, + 121, + 890, + 890 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2We empirically found that using $\\kappa = 25$ works best in general.", + "bbox": [ + 94, + 886, + 426, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22824", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dab41fa904933c4edf9f54abdc57b85ec32d8a75a353d8d7a3b54d4f0f139ba7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAligned (K=10)↓Wild (K=4) ↓Wild (K=8) ↓
Thewlis et al. [55]7.95-31.30
Zhang et al. [67]3.46-40.82
LatentKeypointGAN [12]5.8525.8121.90
Lorenz et al. [30]3.2415.4911.41
IMM [19]3.1919.428.74
LatentKeypointGAN-tuned [12]3.3112.105.63
Autolink [13]3.927.725.66
Autolink † [13]3.546.115.24
Our method3.605.244.35
", + "bbox": [ + 78, + 88, + 467, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that each dataset comes with its own metric. To make results more comparable across the human pose datasets, we report both their original metrics as well as the $\\ell_2$ error when normalizing the image resolution to $128 \\times 128$ .", + "bbox": [ + 75, + 340, + 468, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regressing human-annotated landmarks. To evaluate the quality of unsupervised keypoints, one must relate them with human-annotated landmarks. As in prior research [55], we use linear regression (without bias) to relate between unsupervised keypoints and human-annotated landmarks.", + "bbox": [ + 75, + 407, + 468, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of keypoints and hyperparameters. For each method, we use the standard number of unsupervised keypoints defined for each evaluation protocol—we denote them in our Tables. We use the same hyperparameter for all our experiments as introduced in Sec. 3.2, except for the number of optimization iterations. We optimize the embeddings for 10k iterations, except for the human pose datasets, for which we optimize 500 iterations. To find the number of optimization rounds we use a $10\\%$ validation subset from the training data. While we observed our results on the validation subset to improve consistently for most datasets, we found 10k to give a reasonable optimization time of two hours on an RTX 3090. For the human pose dataset, we found optimization to have converged already at 500 iterations on our validation split.", + "bbox": [ + 75, + 489, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental results", + "text_level": 1, + "bbox": [ + 76, + 726, + 272, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results – Tabs. 1 to 3. We report our results for each dataset in Tabs. 1 to 3. As shown, except for the case when data is heavily processed and aligned (CelebA aligned in Tab. 1, CUB-aligned in Tab. 2, and Human 3.6M in Tab. 3), our method significantly outperforms the state of the art. The most visible gains are for the Tai-Chi-HD dataset, the most challenging among human pose datasets, and on CUB unaligned datasets. For the CUB dataset and the Tai-Chi-HD datasets, we outperform even those that have been supervised with silhouettes or saliency maps.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/45800b2fb463fa78a80eae26c3583145553f34efd240af4e2a47f6959e4189a9.jpg", + "table_caption": [ + "Table 1. Quantitave results for the CelebA dataset - we report results with the standard metrics. Our method performs best for non-aligned cases and is comparable to the state of the art for the aligned case. $\\dagger$ symbol represents the thickness-tuned variant." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionCUB-aligned (K=10) ↓CUB-001 (K=4) ↓CUB-002 (K=4) ↓CUB-003 (K=4) ↓CUB-all (K=4) ↓
SCOPS [16]GT silhouette-18.317.717.012.6
Choudhury et al. [6]GT silhouette-11.315.010.69.2
DFF [9]testing dataset-22.421.622.0-
SCOPS [16]saliency maps-18.518.821.1-
Lorenz et al. [30]unsupervised3.91----
ULD [55, 67]unsupervised-30.129.428.2-
Zhang et al. [67]unsupervised5.3626.927.627.122.4
LatentKeypointGAN [12]unsupervised5.2122.629.121.214.7
GANSeg [14]unsupervised3.2322.122.321.512.1
Autolink [13]unsupervised4.1520.620.319.711.6
Autolink † [13]unsupervised3.5120.219.218.511.3
Our methodunsupervised5.0610.511.110.35.4
", + "bbox": [ + 501, + 88, + 890, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2729725696cca38ae30dc9049c01befc3f9987e9c35a38649aadc935a0220d34.jpg", + "table_caption": [ + "Table 2. Quantitave results for the CUB-200-2011 dataset - we report results with the standard metrics. Except for the CUB-aligned case, our method performs nearly twice better than the compared methods, even outperforming Choudhury et al. [6], which is supervised with ground-truth silhouettes. $\\dagger$ represents the thickness-tuned variant." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionHuman 3.6M (K=16) \\( {\\ell }_{2} \\) standard / unaligned \\( \\downarrow \\)DeepFashion (K=16) PCK† / Rel. \\( {\\ell }_{2} \\downarrow \\)Tai-Chi-HD (K=10) Cum \\( {\\ell }_{2} \\downarrow \\) / Rel. \\( {\\ell }_{2} \\downarrow \\)
Newell et al. [20]paired gt2.16 / ---
DFF [9]testing dataset--494.48 / 14.78
SCOPS [16]saliency maps--411.38 / 12.29
Jakab et al. [20]video*2.73 / ---
Siarohin et al. [51]videos--389.78 / 11.65
Zhang et al. [68]videos--343.67 / 10.27
Zhang et al. [67]videos4.14 / ---
Schmidtke et al. [47]video*3.31 / ---
Sun et al. [53]videos2.53 / ---
Thewis et al. [55]unsupervised7.51 / ---
Zhang et al. [67]unsupervised4.91 / ---
LatentKeypointGAN [12]unsupervised-49%437.69 / 13.08
Lorenz et al. [30]unsupervised2.79 / -57%-
GANSeg [14]unsupervised-59%417.17 / 12.47
autolink [13]unsupervised2.81 / 7.5965%337.50 / 10.08
autolink † [13]unsupervised2.76 / -66%316.10 / 9.45
Our methodunsupervised4.45 / 5.7770%/6.46234.89 / 7.02
", + "bbox": [ + 501, + 344, + 890, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative results for human pose datasets - We report results for the Human 3.6M dataset, Deep Fashion dataset, and the Challenging Tai-Chi-HD datasets. We report both standard metrics for each dataset and the relative $\\ell_2$ error after normalizing images to $128\\times 128$ . Our method, except for the Human 3.6M dataset that is heavily pre-processed, outperforms all baselines. This includes, for the challenging Tai-Chi-HD dataset, supervised ones. * denotes additional supervision (Jakab et al. [20] uses unpaired ground truth, and Schmidtke et al. [47] use the T-pose). The † symbol represents the thickness-tuned variant for Autolink.", + "bbox": [ + 496, + 532, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that our primary focus is on unaligned cases, as we argue that they represent more how keypoints would be used in real-world applications—most real-world datasets are unaligned except for specific classes of objects. Moreover, methods focusing on aligned settings use strong locational priors, and as shown by their results in the unaligned setup—CelebA in the wild, non-aligned cases of Human 3.6M and CUB-200-2011, and Tai-Chi-HD—may perform significantly worse once this alignment prior is broken. Given that the performance of our method, even in the aligned case, is not too far off from methods that utilize alignment, we suspect a more in-depth tuning of our method may make our method outperform these methods,", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22825", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cfac4e55d86b17a0d932f67f6c95dd867bc2368608c3104b10fde8e9a2774792.jpg", + "image_caption": [ + "(a) CelebA dataset keypoints" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 478, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13ad7f932b8010928b28ab4573b299cd0af97687060db3d59eb7e93e85f34264.jpg", + "image_caption": [ + "(c) Tai-Chi-HD dataset keypoints" + ], + "image_footnote": [], + "bbox": [ + 80, + 167, + 478, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a19a4af91ac82dd089039f3f1165af4a1ef579dec65c28fd286b0beea47f2954.jpg", + "image_caption": [ + "(e) Human 3.6M dataset keypoints" + ], + "image_footnote": [], + "bbox": [ + 80, + 246, + 478, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/775c9a914965358e6990af01426d09bacf9e0b22448aee63c631276642544812.jpg", + "image_caption": [ + "(b) CUB-200-2011 dataset keypoints" + ], + "image_footnote": [], + "bbox": [ + 493, + 89, + 890, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d47aa7dd9374c394e3881732a00422cf7b4acb23986668ceeb0f575f8b3e7c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 167, + 890, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b773bfeef0bda086fd21d31a6a9bb9250a2fbde1f585d9927cb0e65b64bc49fa.jpg", + "image_caption": [ + "(d) DeepFashion dataset keypoints", + "(f) Unaligned Human 3.6M dataset keypoints", + "Figure 4. Qualitative examples of unsupervised keypoints – we show our learned keypoints for the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human 3.6M datasets (both for cropped and masked as well as our relaxed version). Note how our keypoints are consistent despite the variability. Our method significantly outperforms other baselines, especially for the challenging Tai-Chi-HD dataset and the CUB subsets." + ], + "image_footnote": [], + "bbox": [ + 493, + 246, + 890, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e37beba569ae00c50929123cf446078c67414af5850fca7e6172e920e5c0e209.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantNormalized ℓ2
Full (Our method)5.4
Without test time ensembling5.6
Without furthest point sampling6.4
Without upsampling the query Q8.0
Without equivariance22.2
", + "bbox": [ + 81, + 416, + 465, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation results - we report the effect of each of our design choices can be seen on the CUB-all dataset. All components contribute to the final performance.", + "bbox": [ + 75, + 536, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but we leave this as future work.", + "bbox": [ + 75, + 604, + 292, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, also note that for CUB-001, CUB-002, and CUB-003, these datasets are small. These datasets are nonaligned, have a large variability between individual images, and only contain 30 images each in the training set. Our method, just from 30 images, successfully identifies keypoints. These results highlight the potential of leveraging emergent (prior) knowledge within Stable Diffusion [44].", + "bbox": [ + 75, + 619, + 468, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results - Fig. 4. We provide example visualizations of our unsupervised keypoints in Fig. 4. As shown, our method discovers keypoints that are consistently localized across the dataset, despite the wide appearance variety.", + "bbox": [ + 75, + 731, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 801, + 227, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform an ablation study for various design choices of our method on the CUB-all dataset. We report the performance of our method with different components disabled in Tab. 4. As shown, all components contribute to the final performance. Test-time ensembling enhances performance,", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but the computation cost linearly scales. We choose, ten augmentations, which provide a good compromise between computation time and accuracy. To remove furthest point sampling we set $\\kappa = K$ , which then makes furthest point sampling select all samples. While this causes points to be more grouped, it still provides reasonable performance. To remove upsampling we instead upsample $\\mathbf{M}$ to the size of the target image, effectively having the attention map build at lower resolutions, sometimes as low as $16\\times 16$ . This results in significant degradation in performance. $\\mathcal{L}_{\\mathrm{equiv}}$ is essential, as without it, the tokens can 'cheat' and simply opt to learn fixed positions on the image.", + "bbox": [ + 496, + 419, + 892, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of training images.. Inspired by our results for the small subsets of CUB-200-2011 dataset, we investigate the impact that the number of images that we use to find keypoints has on our results. We thus optimized our keypoints only with 100 images for CelebA non-aligned setup. Surprisingly, we achieve $5.33K = 8$ , which is comparable to the state of the art. This demonstrates once more how our method is able to leverage information that is already learned in Stable Diffusion [44] to find keypoints.", + "bbox": [ + 496, + 607, + 890, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization", + "text_level": 1, + "bbox": [ + 500, + 755, + 650, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further test the generalization capacity of our learned keypoints. As they are effectively text embeddings, we can simply apply them to any image, including those completely outside of the training domain. We quantitatively evaluate our method and the previous best-performing method Autolink [13]. We find that even in these generalization experiments, our keypoints reach performance comparable to data-specific keypoints. Applying Tai-Chi-HD tokens to un", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22826", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ba9fff91d38fc99eb008a3134a950e3f5e39abd88f3b1e5f4c6c8ab923e95a5.jpg", + "image_caption": [ + "(a) Applying Tai-Chi-HD tokens to Human 3.6M" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 183, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecf61c3715f5b6152c3764e1de86434edbb8072e2b8e28d4871e009674dde050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 88, + 284, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de0dd9cadffe29e8baa849b8e81201bd8fb53b92e47e97e084378bd8d4750473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 88, + 383, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1305e77d530ed04f30a880baa9c103b549b23b3f961d898a7a697c721edc09d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 89, + 482, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af0296a3506232df22930460365fc59f08435495653c10a7149e6e62f1a98538.jpg", + "image_caption": [ + "(c) Applying CelebA tokens to Tai-Chi-HD" + ], + "image_footnote": [], + "bbox": [ + 83, + 184, + 181, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21936bdc746586d3a09f0775637bdbe1a4b2bab990c1f073f2ed566a9d722cdd.jpg", + "image_caption": [ + "Figure 5. Generalization - we apply our learned text tokens (keypoints) to images from other datasets, including those that are of completely different domains. Our tokens generalize well for data of similar type, and surprisingly well even for some extreme cases." + ], + "image_footnote": [], + "bbox": [ + 181, + 184, + 282, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f0808c04c992684bb235584fdecf88ab03992734f44284785c5a5f8a7a47242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 184, + 380, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f23345f7add49711f23a50d802a8750218690a6c68b7dfdd2309ebd6493810f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 184, + 480, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27b8503d3f55eeb79d519fb91bda66a64a3469ce48f1b85cc6de5982d32225ac.jpg", + "image_caption": [ + "(b) Applying CUB-200-2011 tokens to Tai-Chi-HD" + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 586, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e01427bd5156d67fab12cc8e2809a1854f6ce5245d01f91d0623d25e6f2ef48e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 90, + 684, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a3648aa12019cd974626f865eed21a9314692809bff295b58154556010def7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 90, + 785, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/096ccc605c6dbec7da30f24be5f2d91c54dc932c82c8119ad9e93f3f057c3f88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 90, + 883, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27892a36cc3be66b89859e44332cc3dacc8e0a798c08666ccb1f965ae7861afc.jpg", + "image_caption": [ + "(d) Applying CelebA tokens to CUB-200-2011" + ], + "image_footnote": [], + "bbox": [ + 488, + 184, + 586, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2dde74df91c1a29216fe061ba4e03245b85aa5268c6f25fbe2c67da2750d6b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 184, + 686, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffae4282d36896cc5d7ea4a1a0e22db8e4c7a9efa558d43e45707a4d9c4abe5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 184, + 785, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48d59cb3d0a9089accf108ccd388c5adaa4c2178b91be90031aae28cbcb2df17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 184, + 885, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aligned Human 3.6M achieves state-of-the-art performance despite using fewer keypoints $(\\mathrm{K} = 10$ vs $\\mathrm{K} = 16)$ . We also perform on par with the previous state of the art when we apply CUB-200-2011 tokens to Tai-Chi-HD—a case where the dataset gap is not only about the appearance but also beyond classes. Despite the drastic gap, our keypoints perform extremely well, leveraging the generalization power of large pre-trained diffusion models.", + "bbox": [ + 75, + 344, + 467, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show qualitative examples in Fig. 5. As shown, even when applied to different datasets, they look reasonable. For example, in Fig. 5a, when applying Tai-Chi-HD tokens to Human 3.6M, the tokens respond to the same locations on the human body as in Tai-Chi-HD. Surprising was when we applied CUB-200-2011 tokens to Tai-Chi-HD in Fig. 5b—they still responded to the body of the human being, reasonably consistently, although these tokens were trained to respond to birds. Of note are tokens two and six, which correspond to the front and back of the bird heads in Fig. 4—they also reply to the front and back of human heads. Applying CelebA tokens to Tai-Chi-HD in Fig. 5c also shows interesting outcomes, as tokens generally respond to human faces, despite the scale being drastically different between the two datasets. Finally, applying CelebA tokens to the CUB-200-2011 dataset in Fig. 5d shows mixed results—when it is 'successful' it focuses also on the faces of the bird, when it fails, it fails completely. These results hint that the keypoints (tokens) we have learned carry semantic meanings, as expected. We note that none of the baselines that we compare against are able to generalize beyond the dataset they were trained for.", + "bbox": [ + 75, + 465, + 467, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 814, + 202, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a novel method to find unsupervised keypoints using pre-trained text-to-image diffusion models. Given a set of images of a certain object, we propose to optimize the text embeddings (tokens) such that the cross", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb6bcbe188c6104540a9d1f10ae3fb1778af1f8ef0a56716270f71237850dc51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Tai-Chi-HDCUB-200-2011CelebACelebA
→unaligned→Tai-Chi-HD→Tai-Chi-HD→CUB-200-2011
Human3.6m(K=10)↓(K=8)↓(K=8)↓
(K=10)↓Cum ℓ2 ↓ / Rel. ℓ2 ↓Cum ℓ2 ↓ / Rel. ℓ2 ↓
Ours4.88317.94 / 9.50- / 8.618.60
Autolink [13]16.92535.61 / 16.00- / 28.222.56
", + "bbox": [ + 503, + 340, + 888, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Generalization - we quantitatively evaluate the performance of our keypoints on other datasets. Our Tai-Chi-HD keypoints applied to the unaligned Human3.6m setting reach state-of-the-art performance. Interestingly, our CUB keypoints applied to Tai-Chi-HD are on par with the previous state of the art, despite the differences between these datasets.", + "bbox": [ + 498, + 407, + 890, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "attention maps within diffusion models become localized as Gaussians with a small standard deviation. By doing so, we find text tokens that can be used to extract keypoints by extracting the maxima of the attention maps. We have shown that our method, on multiple datasets, under the challenging un-aligned setup, significantly outperforms the state of the art. We have further demonstrated that these tokens are also generalizable.", + "bbox": [ + 498, + 517, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 652, + 678, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The authors would like to thank Cristina Vasconcelos for her constructive feedback during the preparation of this manuscript. Additionally, we extend our gratitude to David Fleet for his approval and support of this work.", + "bbox": [ + 498, + 676, + 890, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant, NSERC Collaborative Research and Development Grant, Google, Digital Research Alliance of Canada, and Advanced Research Computing at the University of British Columbia.", + "bbox": [ + 498, + 736, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from", + "bbox": [ + 506, + 872, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22827", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion models improves imagenet classification. Transactions on Machine Learning Research, 2023. 2, 3", + "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. International Conference on Learning Representations, 2021. 2, 3", + "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 5", + "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection. Proceedings of the IEEE International Conference on Computer Vision, 2022. 2, 3", + "[5] Weiya Chen, Chenchen Yu, Chenyu Tu, Zehua Lyu, Jing Tang, Shiqi Ou, Yan Fu, and Zhidong Xue. A survey on hand pose estimation with wearable sensors and computer- vision-based methods. Sensors, 2020. 2", + "[6] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 2021. 5, 6", + "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv Preprint, 2022. 2", + "[8] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero-shot classifiers. International Conference on Learning Representations, 2023. 2, 3", + "[9] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision, 2018. 6", + "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 5", + "[11] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time, 2022. 5", + "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. International Conference on Learning Representations, 2021. 1, 2, 6", + "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 7, 8", + "[14] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 6", + "[15] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5", + "[16] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Conference on Computer Vision and Pattern Recognition, 2019. 2, 6", + "[17] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013. 2, 5", + "[18] Marwa Jabberi, Ali Wali, Bidyut Baran Chaudhuri, and Adel M Alimi. 68 landmarks are efficient for 3d face alignment: what about more? 3d face alignment method applied to face recognition. Multimedia Tools and Applications, 2023. 1", + "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in Neural Information Processing Systems, 2018. 1, 2, 6", + "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3, 6", + "[21] You-Yi Jau, Rui Zhu, Hao Su, and Manmohan Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In International Conference on Intelligent Robots and Systems, 2020. 2", + "[22] Le Jiang, Caleb Lee, Divyang Teotia, and Sarah Ostadabbas. Animal pose estimation: A closer look at the state-of-the-art, existing gaps and opportunities. Computer Vision and Image Understanding, 2022. 2", + "[23] Yuhe Jin, Weiwei Sun, Jan Hosang, Eduard Trulls, and Kwang Moo Yi. Tusk: Task-agnostic unsupervised keypoints. Advances in Neural Information Processing Systems, 2022. 2", + "[24] Aliasghar Khani, Saeid Asgari Taghanaki, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. arXiv Preprint, 2023. 2, 3", + "[25] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 4", + "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3", + "[27] Wu Liu, Qian Bao, Yu Sun, and Tao Mei. Recent advances of monocular 2d and 3d human pose estimation: A deep learning perspective. ACM Computing Surveys, 2022. 1", + "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 2, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22828", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 2, 5", + "[30] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 5, 6", + "[31] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 2004. 1", + "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv Preprint, 2023. 1", + "[33] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3", + "[34] Giorgia Marullo, Leonardo Tanzi, Pietro Piazzolla, and Enrico Vezzetti. 6d object position estimation from 2d images: a literature review. Multimedia Tools and Applications, 2023. 2", + "[35] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3", + "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[37] Aditya Nandy, Chenru Duan, and Heather J Kulik. Audacity of huge: overcoming challenges of data scarcity and data quality for machine learning in computational materials discovery. Current Opinion in Chemical Engineering, 2022. 2", + "[38] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 3", + "[39] OpenAI. Gpt-4 technical report, 2023. 2", + "[40] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision, 2018. 3", + "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2022. 2, 3", + "[42] Linhao Qu, Siyu Liu, Xiaoyu Liu, Manning Wang, and Zhijian Song. Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv Preprint, 2022. 2, 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7", + "[45] Helena Russello, Rik van der Tol, and Gert Kootstra. T-leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and Electronics in Agriculture, 2022. 2", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 3", + "[47] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 6", + "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2", + "[49] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[50] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 5", + "[51] Aliaksandr Siarohin, Subhankar Roy, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Motion-supervised co-part segmentation. In International Conference on Pattern Recognition, 2021. 5, 6", + "[52] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1", + "[53] Jennifer J Sun, Serim Ryou, Roni H Goldshmid, Brandon Weissboud, John O Dabiri, David J Anderson, Ann Kennedy, Yisong Yue, and Pietro Perona. Self-supervised keypoint discovery in behavioral videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3", + "[55] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22829", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embeddings. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1, 2, 6", + "[56] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv Preprint, 2023. 2, 3", + "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv Preprint, 2023. 2", + "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 2, 5", + "[59] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv Preprint, 2023. 2, 3", + "[60] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. International Conference on Computer Vision, 2023. 1", + "[61] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. arXiv Preprint, 2023. 2, 3", + "[62] Yue Wu and Qiang Ji. Facial landmark detection: A literature survey. International Journal of Computer Vision, 2019. 2", + "[63] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. arXiv Preprint, 2023. 2, 3", + "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3", + "[65] Lumin Xu, Sheng Jin, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3", + "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6", + "[68] Yanping Zhang, Qiaokang Liang, Kunlin Zou, Zhengwei Li, Wei Sun, and Yaonan Wang. Self-supervised part segmentation via motion imitation. Image and Vision Computing, 2022. 6" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[69] Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 2023. 2", + "bbox": [ + 501, + 90, + 890, + 148 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22830", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_model.json b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d87255284d3bd9ac511859a33fbdff30718838d --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.783, + 0.154 + ], + "angle": 0, + "content": "Unsupervised Keypoints from Pretrained Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.179, + 0.82, + 0.253 + ], + "angle": 0, + "content": "Eric Hedlin1 Gopal Sharma1 Shweta Mahajan1,2 Xingzhe He1 Hossam Isack3 Abhishek Kar3 Helge Rhodin1 Andrea Tagliaasacchi4,5,6 Kwang Moo Yi1 1 University of British Columbia 2 Vector Institute for AI 3 Google Research 4 Google DeepMind 5 Simon Fraser University 6 University of Toronto" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.255, + 0.634, + 0.269 + ], + "angle": 0, + "content": "https://stablekeypoints.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.277, + 0.88, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.503, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Teaser - we propose an unsupervised method to learn keypoints based on optimizing text embeddings of latent diffusion models [44]. Our method is motivated by the fact that random text tokens already respond roughly consistently to semantically similar regions. By promoting localization we obtain unsupervised keypoints that outperform the state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.559, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Unsupervised learning of keypoints and landmarks has seen significant progress with the help of modern neural network architectures, but performance is yet to match the supervised counterpart, making their practicability questionable. We leverage the emergent knowledge within text-to-image diffusion models, towards more robust unsupervised keypoints. Our core idea is to find text embeddings that would cause the generative model to consistently attend to compact regions in images (i.e. keypoints). To do so, we simply optimize the text embedding such that the cross-attention maps within the denoising network are localized as Gaussians with small standard deviations. We validate our performance on multiple datasets: the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human3.6m datasets. We achieve significantly improved accuracy, sometimes even outperforming supervised ones, particularly for data that is non-aligned and less curated. Our code is publicly available at the project page." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.559, + 0.632, + 0.575 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Keypoints or landmarks have played a critical role in computer vision for various tasks including image matching [31], 3D reconstruction [18], and motion tracking [32, 60]. Similarly to many other areas of computer vision, research has quickly adopted supervised learning to tackle this problem [3, 27]. However, labeling is tedious and sometimes even ambiguous—for example, it is difficult to consistently decide which keypoints on a human face are the \"most important\". Researchers have therefore been investigating unsupervised approaches [12, 13, 19, 30, 55, 67]. These are typically implemented as autoencoders paired with handcrafted intermediate layers, or losses that enforce spatial locality and equivariance of keypoint locations under deformation. However, as we will show later, these methods struggle with non-preprocessed data, and their performance is heavily reliant on knowing the ground truth location of objects, clearly limiting their practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To enhance the learning of unsupervised keypoints, we draw inspiration from the demonstrated success of scaling up datasets [52]. For example, in natural language pro" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.887, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 2. Example attention maps - we show example attention maps for a selected learned keypoint for the CUB-200-2011 dataset, on the CUB-aligned subset. As shown, our keypoint attention map responds consistently across varying images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.473, + 0.405 + ], + "angle": 0, + "content": "cessing performance has recently improved to a great extent thanks to large models and data [7, 39, 57]. Similarly, in computer vision, the performance of text-to-image models [43, 44, 46] has drastically improved thanks to the availability of extra large datasets [48]. However, unsupervised keypoint learning typically assumes class-specific datasets, e.g., animals that have a shared skeleton that connects keypoints, and these datasets are small in scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.636 + ], + "angle": 0, + "content": "Rather than collecting larger domain-specific datasets, we instead propose to leverage the knowledge stored within large generative models, such as Stable Diffusion [44]. This has been shown to be very effective across a number of tasks [1, 2, 4, 8, 15, 24, 26, 33, 35, 41, 54, 56, 59, 61, 63, 64, 66], but, to the best of our knowledge, it has not yet found application for the task of keypoint learning. Our main idea is to localize \"important\" keypoints by finding text embeddings that consistently correspond to a distinct location in images of a certain object class. This idea is rooted in the observation that, even with random text embeddings, the attention maps for various images roughly correspond to regions that are semantically similar; see Fig. 1. Therefore, text embeddings carry semantic meaning, which could be used to relate collections of images to each other; see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.821 + ], + "angle": 0, + "content": "We find embeddings that are specific to certain locations by enforcing localized attention maps. In more detail, we propose to find (i.e., optimize) a set of tokens in a text embedding that locally responds in the Stable Diffusion cross-attention layers. We enforce locality by maximizing the similarity of the attention responses of each token to a single-mode Gaussian distribution. Thanks to the way the cross-attention layers are constructed within Stable Diffusion, this simple objective also prevents the different tokens from attending to the same locations in an image, a common degenerate solution that typically requires explicit workarounds [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method on established benchmarks: CelebA [28], CUB-200-2011 [58], Tai-Chi-HD [50], DeepFashion [29], and Human3.6m [17]. Our approach yields results on par with state-of-the-art methods for well-curated and aligned datasets, while notably enhancing performance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.314 + ], + "angle": 0, + "content": "for in-the-wild setups, particularly with unaligned data, sometimes even surpassing fully supervised baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.642, + 0.347 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Below we review the literature of finding keypoints in unsupervised and supervised fashion, along with work that exploits large pre-trained models like stable-diffusion for lower-level computer vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Learning keypoints with supervision. Pose estimation and landmark estimation are fundamental problems in computer vision. They naturally arise in various tasks, including human [69] and animal pose estimation [22], hand [5] and face landmark estimation [62], and object pose tracking [34]. Many fully supervised methods find different ways to induce some prior within the model to better capture the task at hand, such as using part affinity fields [3], temporal consistency for video data [45], spacial relationships [65], and geometry constraints [21] among others. While fully supervised methods have excelled in categories with abundant labeled data, such as human pose estimation, their major drawback is the insatiable need for large and high-quality datasets [22, 37, 42]. The scalability of gathering such extensive and meticulously annotated data for every conceivable object category remains a significant drawback [22, 37, 42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Learning keypoints via self-supervision. The amount of unlabeled data far exceeds that of labeled data, so unsupervised keypoint estimation methods attempt to take advantage of this. Self-supervised keypoint detection often relies on tracking how keypoints move with image changes and uses various constraints for known transformations [16, 19, 30, 49, 55, 67], but these methods can struggle with background modeling [49, 67] and pose variations [16]. One can also rely on image reconstruction to learn keypoints. Some methods use GANs to generate images from keypoints [12, 14], but this often results in training instability. Alternatively, auto-encoders can be also used [13, 67], but these require training from scratch on each dataset. Our method neither suffers from GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "training instability, nor requires dataset fine-tuning. Finally, there exist self-supervised methods that exploit skeletal representations [13, 20, 40]. However, many of these approaches generally require known keypoint connectivity and video data [20, 40], and often face limitations in background handling and generalizability to objects within the same class [13, 20, 40]. Our method has no object-specific priors, and generalizes well due to the large dataset used by the pre-trained diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.473, + 0.492 + ], + "angle": 0, + "content": "Diffusion models for image understanding. Recently, large image diffusion models have reached impressive image generation quality [43, 44, 46]. These models learn priors for real images within the latent space of the diffusion model, and provide a useful initialization for many downstream tasks such as image correspondence [15, 33, 54, 66], object detection [4], semantic segmentation [2, 24, 56, 59, 61, 63, 64], and image classification [1, 8]. Interestingly, without requiring any retraining, these models demonstrate an innate ability to understand 3D spatial configurations [26, 35, 41]. Recent work in each of these areas has shown the emergent power of these large models, most of them using the model without any modifications or extra supervision required. More relevant to our work, Mokady et al. [36] found that the pre-trained Stable Diffusion [44] model's cross-attention maps connect text tokens to semantically relevant areas in images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.473, + 0.725 + ], + "angle": 0, + "content": "Correspondences via diffusion models. Among works that re-purpose diffusion models, of high relevance, is the effectiveness of diffusion models in correspondence estimation tasks [15, 33, 54, 66]. Hedlin et al. [15] optimizes the attention map for a specific point in a source image and finds the corresponding activation in a target image. However, this method requires a query to be provided in the source image. While our method shares the same inspiration of utilizing attention maps, critically, rather than optimizing the embedding given a single image, we optimize an embedding given a dataset of images from a given object class. In other words, our method discovers on its own, where to focus, rather than relying on user input. Our task is therefore changed from image matching between two images, to semantic matching across all images within the dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.17, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To identify a set of representative keypoints across a dataset of images, we formulate our approach in an unsupervised framework leveraging conditional diffusion models; see Fig. 3. In particular, we utilize the cross-attention maps between the text embeddings and the image features, derived from the latent diffusion model [44], and force them to consistently concentrate their activation on highly localized regions within the images. While Hedlin et al. [15] employed a similar mechanism (i.e., given a set of keypoint locations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "in one image, identify correspondences in another image), in this work we seek to identify semantic correspondences across all images within a class-specific dataset (e.g., human faces), without any given knowledge on what and where to focus. We show that this is possible simply by enforcing locality and equivariance to transformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Let us start by quickly reviewing the fundamentals of diffusion models and formalizing the attention maps that we will utilize within these models (Sec. 3.1). We then detail the objectives used to learn the text embeddings that represent keypoints (Sec. 3.2) and discuss important implementation details (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.285, + 0.822, + 0.302 + ], + "angle": 0, + "content": "3.1. Attention maps in diffusion networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Diffusion models are a class of generative models that approximate the data distribution by denoising a base (typically Gaussian) distribution [38]. A latent diffusion model operates on a latent representation \\(\\mathbf{z}\\) rather than the image itself, with an encoder that maps an image \\(\\mathbf{X}\\) into a latent \\(\\mathbf{z}\\), and a decoder that maps \\(\\mathbf{z}\\) into \\(\\mathbf{X}\\). These models define a forward diffusion process, where the latent representation \\(\\mathbf{z}\\) is gradually transformed into Gaussian noise over a series of \\(T\\) time steps. The inverse process, over a sequence denoising steps \\(t = 1,\\dots,T\\) predicts the latent noise \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t)\\) which was gradually added in each iteration in order to recover the original (latent) signal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.598 + ], + "angle": 0, + "content": "In our work, we are interested in conditional diffusion models, and the explicit attentional relationship between the condition (i.e. text) and the generated outcome (i.e. image) that these models learn. Typically, diffusion models are made conditional on some text \\(\\mathbf{y}\\), by providing an embedding \\(\\mathbf{e} = \\tau_{\\theta}(\\mathbf{y})\\) from a text encoder \\(\\tau_{\\theta}\\) to the denoiser. They are then trained to optimize" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.61, + 0.892, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathbf {z}, t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}, t, \\mathbf {e}) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.722 + ], + "angle": 0, + "content": "where the denoiser \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathbf{e})\\) is typically implemented by a transformer architecture [38] involving a combination of self-attention and cross-attention layers. Of our interest here is the cross-attention layers that relate \\(\\mathbf{e}\\) to \\(\\mathbf{z}_t\\), which we now formalize." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Specifically, in the transformer part of the model, denote \\(\\Phi_l^c (\\cdot)\\) and \\(\\Psi_l^c (\\cdot)\\) as the \\(c\\)-th head and the \\(l\\)-th linear layers of the U-Net. We calculate the query as \\(\\mathbf{Q}_l^c = \\Phi_l^c (\\mathbf{z}_{t = 1})\\in \\mathbb{R}^{(H\\times W)\\times D_l}\\), and the key from the language embedding \\(\\mathbf{K}_l^c = \\Psi_l^c (\\mathbf{e})\\in \\mathbb{R}^{N\\times D_l}\\), where \\(N\\) is the number of tokens, \\(C\\) the number of heads in the transformer attention layer, \\(H\\) and \\(W\\) are image height and width at that specific layer in U-Net, and \\(D_{l}\\) the dimensionality of the layer. Given query and key, the cross-attention map \\(\\mathbf{M}_l\\in \\mathbb{R}^{(H\\times W)\\times N}\\) is then computed via softmax along the \\(N\\) di" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.862, + 0.902 + ], + "angle": 0, + "content": "1We choose \\( t = 1 \\) where \\( T = 50 \\) steps via hyper-parameter tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22822" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.87, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Overview - we pass a randomly initialized text embedding into Stable Diffusion [44] and extract the attention maps. We then optimize the text embedding to have localized attention maps, by supervising them to become a single-mode Gaussian distribution, drawn at the location of their maxima. We also enforce attention maps to be transformation equivariant to small affine transformations on images. We repeat this process over a set of training images, which after optimization provides a set of \\(K\\) keypoints." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.377, + 0.435, + 0.392 + ], + "angle": 0, + "content": "mension, and average pooling along the \\(C\\) dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.403, + 0.47, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) = \\mathbb {E} _ {c} \\left[ \\operatorname {s o f t m a x} _ {n} \\left(\\mathbf {Q} _ {l} ^ {c} \\cdot \\mathbf {K} _ {l} / \\sqrt {D _ {l}}\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.47, + 0.501 + ], + "angle": 0, + "content": "As various layers of the U-Net exhibit distinct levels of semantic understanding, following Hedlin et al. [15], we collect this information by average pooling across a selection of layers:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.512, + 0.469, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {M}} = \\mathbb {E} _ {l = 7.. 1 0} \\left[ \\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) \\right] \\in \\mathbb {R} ^ {(H \\times W) \\times N}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.47, + 0.595 + ], + "angle": 0, + "content": "In what follows, to lighten the notation, we drop the attention mask arguments \\((\\mathbf{e},\\mathbf{X})\\) and write the attention map for the n-th token as \\(\\tilde{\\mathbf{M}}_n\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.45, + 0.621 + ], + "angle": 0, + "content": "3.2. Optimizing to find the keypoint embeddings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.688 + ], + "angle": 0, + "content": "To obtain a text embedding that can be used to locate keypoints, for each of them, we simply optimize for two objectives that respectively encourages localization and equivariance to geometric transformations. We thus write" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.702, + 0.469, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l i z e}} + \\lambda_ {\\text {e q u i v}} \\mathcal {L} _ {\\text {e q u i v}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.469, + 0.805 + ], + "angle": 0, + "content": "where we apply \\(\\lambda_{\\mathrm{equiv}} = 10\\) to balance the two losses to be in a similar operating range. Equivariance is enforced in the typical form of learning to be invariant to transformations. We first quickly detail \\(\\mathcal{L}_{\\mathrm{equiv}}\\), and then discuss how we enforce localization, which is the core of our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.812, + 0.469, + 0.872 + ], + "angle": 0, + "content": "Equivariance - \\(\\mathcal{L}_{\\mathrm{equiv}}\\). To ensure our model's attention mechanism remains consistent across different geometric transformations \\(\\mathcal{T}\\) of the input, we use the typical equivariance loss [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e q u i v}} = \\mathbb {E} _ {n} \\| \\mathcal {T} ^ {- 1} \\left(\\mathbf {M} _ {n} (\\mathbf {e}, \\mathcal {T} (\\mathbf {X})) - \\mathbf {M} _ {n} (\\mathbf {e}, \\mathbf {X}) \\right\\| ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.438 + ], + "angle": 0, + "content": "For \\(\\mathcal{T}\\) we simply utilize minor affine transformations. We use random rotations between \\(\\pm 15\\) degrees, translations between \\(\\pm 0.25 \\times W\\), and scaling between \\(100 - 120\\%\\) of the original image size." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.444, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Encouraging localization - \\(\\mathcal{L}_{\\mathrm{localize}}\\). We encourage localization by forcing \\(\\tilde{\\mathbf{M}}_n\\) to be a single-mode Gaussian distribution located at its maximum. In more details, denoting the Gaussian image that shares the same maximum as \\(\\tilde{\\mathbf{M}}_n\\) as \\(\\mathbf{G}_n\\), we write" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.527, + 0.891, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o c a l i z e}} = \\mathbb {E} _ {n} \\| \\tilde {\\mathbf {M}} _ {n} - \\mathbf {G} _ {n} \\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.613 + ], + "angle": 0, + "content": "To create the Gaussian images \\(\\mathbf{G}_n\\), we first identify the spatial location exhibiting the maximal response within the heatmap corresponding to each token \\(n\\) by taking the argmax:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.614, + 0.891, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {n} = \\underset {w, h} {\\arg \\max } \\tilde {\\mathbf {M}} _ {n} [ h, w ]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.813, + 0.659 + ], + "angle": 0, + "content": "We then generate a Gaussian image; see Fig. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.666, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {n} = \\exp \\left(- \\frac {\\| \\mathbf {X Y} _ {\\mathrm {c o o r d}} - \\boldsymbol {\\mu} _ {n} \\| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.707, + 0.828, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}\\mathbf{Y}_{\\mathrm{coord}}\\) is a tensor of image coordinates." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.729, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Promoting mutual exclusivity. It is important to note that while \\(\\mathcal{L}_{\\mathrm{localize}}\\) in (6) at first glance seem to only encourage localization, it also enforces \\(\\tilde{\\mathbf{M}}_n\\) to be mutually exclusive for different \\(n\\) because of the softmax operation in (2). Should multiple embeddings become similar, their attention responses in (2) \\(\\mathbf{Q}_l^c\\cdot \\mathbf{K}_l\\) will become similar, resulting in the softmax of the attention map being a flat response (i.e. deviating from a Gaussian shape). In other words, (6) naturally enforces exclusivity with the help of (2)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Stabilizing optimization by working with a subset. We noticed in our experiments that attention maps \\(\\mathbf{M}\\) for some" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22823" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "tokens can be 'spread' for some images, e.g., due to occlusions, destabilizing optimization. We thus opt for a simple solution of looking into the top-K tokens that are local. Specifically, we apply our losses over \\( n \\in \\mathcal{N}(\\kappa) \\), which returns the \\( \\kappa \\in \\mathbb{N} \\) entries with the most spatially localized heatmap responses2, as measured by KL divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.193, + 0.47, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} (\\kappa) = \\operatorname {A r g T o p} _ {\\kappa} \\left\\{- \\mathrm {K L} \\left(\\mathbf {G} _ {n}, \\tilde {\\mathbf {M}} _ {n}\\right) \\right\\} _ {n = 1} ^ {N}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Final keypoints. While \\(\\mathcal{L}_{\\mathrm{localize}}\\) naturally enforces exclusivity, it does not guarantee a complete coverage of the object. Thus, after we finish optimizing, we refine the set of keypoints through furthest point sampling using the training images. Specifically, for each image we write:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.316, + 0.47, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} = \\operatorname {F P S} _ {K} (\\{\\boldsymbol {\\mu} _ {i} \\mid i \\in \\mathcal {N} (\\kappa) \\}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.405 + ], + "angle": 0, + "content": "where \\(K\\) is the desired number of keypoints \\(K < \\kappa\\). Then, as the set \\(\\mathcal{K}\\) differs in each image, we simply choose \\(K\\) to tokens that appeared most frequently in \\(\\mathcal{K}\\) within the training image set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.292, + 0.43 + ], + "angle": 0, + "content": "3.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Test-time ensembling. At inference time, as in common literature [3, 10, 15], rather than employing the attention map of the original image, we average the attention maps across multiple augmentations (we use the same transformations for test time augmentation as in (5)) of the image:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.525, + 0.469, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {M}} \\leftarrow \\sum_ {i} \\mathcal {T} _ {i} ^ {- 1} \\left(\\mathbf {M} \\left(\\mathbf {e}, \\mathcal {T} _ {i} (\\mathbf {X})\\right)\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Upsampling attention maps. The attention maps in (2) are typically of low resolution. Specifically, as we use Stable Diffusion 1.5 [44], depending on the layer we extract the attention maps from, they are either \\(16 \\times 16\\) or \\(32 \\times 32\\). We thus opt to upsample the query \\(\\mathbf{Q}\\) via bicubic interpolation to achieve a standard resolution of \\(128 \\times 128\\). We have experimented with other upsampling techniques such as the commonly used bilinear sampling or a learned upsampler that is trained alongside, but a simple bicubic upsample was shown to be effective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.163, + 0.749 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.758, + 0.264, + 0.774 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.469, + 0.812 + ], + "angle": 0, + "content": "We evaluate our method on five standard datasets for unsupervised keypoint evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.469, + 0.879 + ], + "angle": 0, + "content": "CelebA dataset [28]: A dataset of 202,599 facial images of celebrities. We evaluate both the aligned and nonaligned cases following the standard protocol of omitting images with faces occupying less than \\(30\\%\\) of the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "The standard metric for this dataset is to measure the average \\(\\ell_2\\) error normalized by the inter-ocular distance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.286 + ], + "angle": 0, + "content": "- CUB-200-2011 dataset [58]: This dataset consists of 11,788 bird images. We use both the aligned (CUB-aligned) and non-aligned (CUB-all) variants. For the nonaligned variants, we further look at CUB-001, CUB-002, and CUB-003, which are specific bird subcategories. Notably, these subsets contain only 30 images each—we only use these 30 for training. We follow the standard protocol [6, 30] and normalize the images to be of \\(256 \\times 256\\). The standard metric for this dataset is the mean \\(\\ell_2\\) error, normalized by the dimension of the images after normalization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.287, + 0.892, + 0.422 + ], + "angle": 0, + "content": "- Tai-Chi-HD dataset [50]: This dataset contains 3049 training videos and 285 test videos of people performing Tai-Chi, which shows more diverse poses compared to the other datasets, and is the most challenging among the human pose-centric datasets that we use. We follow Siarohin et al. [51] and use 500 images for testing and 300 images for training. The standard metric for this dataset is to measure the accumulated \\(\\ell_2\\) error, with the images standardized to \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.424, + 0.892, + 0.558 + ], + "angle": 0, + "content": "- DeepFashion dataset [29]: This dataset contains 53k images of fashion models, mostly standing with a white background. We follow Lorenz et al. [30] and only keep full body images. This leaves 10,604 images for training and 1,179 images for testing. Also following the baselines, we use keypoints generated by AlphaPose [11] as ground truth. The standard metric for this dataset is the percentage of correct keypoints (PCK) with a 6-pixel threshold." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.56, + 0.892, + 0.891 + ], + "angle": 0, + "content": "- Human 3.6M dataset [17]: This dataset is of humans performing various actions, comprised of 3.6 million images. We follow the standard protocol [67] and focus on six activities: direction, discussion, posing, waiting, greeting, and walking. We utilize subjects 1, 5, 6, 7, 8, and 9 for training, while subject 11 is reserved for testing. This division yields a training dataset comprising 796,648 images and a testing dataset containing 87,975 images. The background for this dataset is also simple, and often masked out with ground-truth masks for evaluation. This dataset is also typically heavily pre-processed and aligned when used for unsupervised keypoint evaluation. We experiment with the standard pre-processing [30, 67] and also a relaxed version of our own. To relax the alignment, we crop a square bounding box such that the margin from the bounding box to the person is 100 pixels, which on average corresponds to the person's height being \\(2/3\\) of the crop. We further add a uniform random translation up to 100 pixels (same as the margin) to remove the central bias. Example crops are visualized in Fig. 4f. The standard metric for this dataset is the \\(\\ell_2\\) error after normalizing the image resolution to \\(128 \\times 128\\)." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.427, + 0.901 + ], + "angle": 0, + "content": "2We empirically found that using \\(\\kappa = 25\\) works best in general." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22824" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.247 + ], + "angle": 0, + "content": "
MethodAligned (K=10)↓Wild (K=4) ↓Wild (K=8) ↓
Thewlis et al. [55]7.95-31.30
Zhang et al. [67]3.46-40.82
LatentKeypointGAN [12]5.8525.8121.90
Lorenz et al. [30]3.2415.4911.41
IMM [19]3.1919.428.74
LatentKeypointGAN-tuned [12]3.3112.105.63
Autolink [13]3.927.725.66
Autolink † [13]3.546.115.24
Our method3.605.244.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.258, + 0.47, + 0.314 + ], + "angle": 0, + "content": "Table 1. Quantitave results for the CelebA dataset - we report results with the standard metrics. Our method performs best for non-aligned cases and is comparable to the state of the art for the aligned case. \\(\\dagger\\) symbol represents the thickness-tuned variant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.402 + ], + "angle": 0, + "content": "Note that each dataset comes with its own metric. To make results more comparable across the human pose datasets, we report both their original metrics as well as the \\(\\ell_2\\) error when normalizing the image resolution to \\(128 \\times 128\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.469, + 0.483 + ], + "angle": 0, + "content": "Regressing human-annotated landmarks. To evaluate the quality of unsupervised keypoints, one must relate them with human-annotated landmarks. As in prior research [55], we use linear regression (without bias) to relate between unsupervised keypoints and human-annotated landmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.716 + ], + "angle": 0, + "content": "Number of keypoints and hyperparameters. For each method, we use the standard number of unsupervised keypoints defined for each evaluation protocol—we denote them in our Tables. We use the same hyperparameter for all our experiments as introduced in Sec. 3.2, except for the number of optimization iterations. We optimize the embeddings for 10k iterations, except for the human pose datasets, for which we optimize 500 iterations. To find the number of optimization rounds we use a \\(10\\%\\) validation subset from the training data. While we observed our results on the validation subset to improve consistently for most datasets, we found 10k to give a reasonable optimization time of two hours on an RTX 3090. For the human pose dataset, we found optimization to have converged already at 500 iterations on our validation split." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.274, + 0.743 + ], + "angle": 0, + "content": "4.2. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Quantitative results – Tabs. 1 to 3. We report our results for each dataset in Tabs. 1 to 3. As shown, except for the case when data is heavily processed and aligned (CelebA aligned in Tab. 1, CUB-aligned in Tab. 2, and Human 3.6M in Tab. 3), our method significantly outperforms the state of the art. The most visible gains are for the Tai-Chi-HD dataset, the most challenging among human pose datasets, and on CUB unaligned datasets. For the CUB dataset and the Tai-Chi-HD datasets, we outperform even those that have been supervised with silhouettes or saliency maps." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.892, + 0.231 + ], + "angle": 0, + "content": "
MethodSupervisionCUB-aligned (K=10) ↓CUB-001 (K=4) ↓CUB-002 (K=4) ↓CUB-003 (K=4) ↓CUB-all (K=4) ↓
SCOPS [16]GT silhouette-18.317.717.012.6
Choudhury et al. [6]GT silhouette-11.315.010.69.2
DFF [9]testing dataset-22.421.622.0-
SCOPS [16]saliency maps-18.518.821.1-
Lorenz et al. [30]unsupervised3.91----
ULD [55, 67]unsupervised-30.129.428.2-
Zhang et al. [67]unsupervised5.3626.927.627.122.4
LatentKeypointGAN [12]unsupervised5.2122.629.121.214.7
GANSeg [14]unsupervised3.2322.122.321.512.1
Autolink [13]unsupervised4.1520.620.319.711.6
Autolink † [13]unsupervised3.5120.219.218.511.3
Our methodunsupervised5.0610.511.110.35.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Table 2. Quantitave results for the CUB-200-2011 dataset - we report results with the standard metrics. Except for the CUB-aligned case, our method performs nearly twice better than the compared methods, even outperforming Choudhury et al. [6], which is supervised with ground-truth silhouettes. \\(\\dagger\\) represents the thickness-tuned variant." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.345, + 0.891, + 0.523 + ], + "angle": 0, + "content": "
MethodSupervisionHuman 3.6M (K=16) \\( {\\ell }_{2} \\) standard / unaligned \\( \\downarrow \\)DeepFashion (K=16) PCK† / Rel. \\( {\\ell }_{2} \\downarrow \\)Tai-Chi-HD (K=10) Cum \\( {\\ell }_{2} \\downarrow \\) / Rel. \\( {\\ell }_{2} \\downarrow \\)
Newell et al. [20]paired gt2.16 / ---
DFF [9]testing dataset--494.48 / 14.78
SCOPS [16]saliency maps--411.38 / 12.29
Jakab et al. [20]video*2.73 / ---
Siarohin et al. [51]videos--389.78 / 11.65
Zhang et al. [68]videos--343.67 / 10.27
Zhang et al. [67]videos4.14 / ---
Schmidtke et al. [47]video*3.31 / ---
Sun et al. [53]videos2.53 / ---
Thewis et al. [55]unsupervised7.51 / ---
Zhang et al. [67]unsupervised4.91 / ---
LatentKeypointGAN [12]unsupervised-49%437.69 / 13.08
Lorenz et al. [30]unsupervised2.79 / -57%-
GANSeg [14]unsupervised-59%417.17 / 12.47
autolink [13]unsupervised2.81 / 7.5965%337.50 / 10.08
autolink † [13]unsupervised2.76 / -66%316.10 / 9.45
Our methodunsupervised4.45 / 5.7770%/6.46234.89 / 7.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Table 3. Quantitative results for human pose datasets - We report results for the Human 3.6M dataset, Deep Fashion dataset, and the Challenging Tai-Chi-HD datasets. We report both standard metrics for each dataset and the relative \\(\\ell_2\\) error after normalizing images to \\(128\\times 128\\). Our method, except for the Human 3.6M dataset that is heavily pre-processed, outperforms all baselines. This includes, for the challenging Tai-Chi-HD dataset, supervised ones. * denotes additional supervision (Jakab et al. [20] uses unpaired ground truth, and Schmidtke et al. [47] use the T-pose). The † symbol represents the thickness-tuned variant for Autolink." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We note that our primary focus is on unaligned cases, as we argue that they represent more how keypoints would be used in real-world applications—most real-world datasets are unaligned except for specific classes of objects. Moreover, methods focusing on aligned settings use strong locational priors, and as shown by their results in the unaligned setup—CelebA in the wild, non-aligned cases of Human 3.6M and CUB-200-2011, and Tai-Chi-HD—may perform significantly worse once this alignment prior is broken. Given that the performance of our method, even in the aligned case, is not too far off from methods that utilize alignment, we suspect a more in-depth tuning of our method may make our method outperform these methods," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22825" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.48, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.155, + 0.357, + 0.167 + ], + "angle": 0, + "content": "(a) CelebA dataset keypoints" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.168, + 0.48, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.234, + 0.368, + 0.246 + ], + "angle": 0, + "content": "(c) Tai-Chi-HD dataset keypoints" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.247, + 0.48, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.312, + 0.371, + 0.325 + ], + "angle": 0, + "content": "(e) Human 3.6M dataset keypoints" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.09, + 0.892, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.155, + 0.79, + 0.167 + ], + "angle": 0, + "content": "(b) CUB-200-2011 dataset keypoints" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.168, + 0.892, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.234, + 0.784, + 0.247 + ], + "angle": 0, + "content": "(d) DeepFashion dataset keypoints" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.247, + 0.892, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.312, + 0.812, + 0.325 + ], + "angle": 0, + "content": "(f) Unaligned Human 3.6M dataset keypoints" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.894, + 0.392 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples of unsupervised keypoints – we show our learned keypoints for the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human 3.6M datasets (both for cropped and masked as well as our relaxed version). Note how our keypoints are consistent despite the variability. Our method significantly outperforms other baselines, especially for the challenging Tai-Chi-HD dataset and the CUB subsets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.417, + 0.466, + 0.526 + ], + "angle": 0, + "content": "
VariantNormalized ℓ2
Full (Our method)5.4
Without test time ensembling5.6
Without furthest point sampling6.4
Without upsampling the query Q8.0
Without equivariance22.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Table 4. Ablation results - we report the effect of each of our design choices can be seen on the CUB-all dataset. All components contribute to the final performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.294, + 0.619 + ], + "angle": 0, + "content": "but we leave this as future work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Finally, also note that for CUB-001, CUB-002, and CUB-003, these datasets are small. These datasets are nonaligned, have a large variability between individual images, and only contain 30 images each in the training set. Our method, just from 30 images, successfully identifies keypoints. These results highlight the potential of leveraging emergent (prior) knowledge within Stable Diffusion [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Qualitative results - Fig. 4. We provide example visualizations of our unsupervised keypoints in Fig. 4. As shown, our method discovers keypoints that are consistently localized across the dataset, despite the wide appearance variety." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.228, + 0.818 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We perform an ablation study for various design choices of our method on the CUB-all dataset. We report the performance of our method with different components disabled in Tab. 4. As shown, all components contribute to the final performance. Test-time ensembling enhances performance," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.893, + 0.601 + ], + "angle": 0, + "content": "but the computation cost linearly scales. We choose, ten augmentations, which provide a good compromise between computation time and accuracy. To remove furthest point sampling we set \\(\\kappa = K\\), which then makes furthest point sampling select all samples. While this causes points to be more grouped, it still provides reasonable performance. To remove upsampling we instead upsample \\(\\mathbf{M}\\) to the size of the target image, effectively having the attention map build at lower resolutions, sometimes as low as \\(16\\times 16\\). This results in significant degradation in performance. \\(\\mathcal{L}_{\\mathrm{equiv}}\\) is essential, as without it, the tokens can 'cheat' and simply opt to learn fixed positions on the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Number of training images.. Inspired by our results for the small subsets of CUB-200-2011 dataset, we investigate the impact that the number of images that we use to find keypoints has on our results. We thus optimized our keypoints only with 100 images for CelebA non-aligned setup. Surprisingly, we achieve \\(5.33K = 8\\), which is comparable to the state of the art. This demonstrates once more how our method is able to leverage information that is already learned in Stable Diffusion [44] to find keypoints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.651, + 0.77 + ], + "angle": 0, + "content": "4.4. Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We further test the generalization capacity of our learned keypoints. As they are effectively text embeddings, we can simply apply them to any image, including those completely outside of the training domain. We quantitatively evaluate our method and the previous best-performing method Autolink [13]. We find that even in these generalization experiments, our keypoints reach performance comparable to data-specific keypoints. Applying Tai-Chi-HD tokens to un" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22826" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.184, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.089, + 0.285, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.089, + 0.384, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.09, + 0.483, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.17, + 0.413, + 0.183 + ], + "angle": 0, + "content": "(a) Applying Tai-Chi-HD tokens to Human 3.6M" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.185, + 0.182, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.185, + 0.283, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.185, + 0.382, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.185, + 0.482, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.265, + 0.398, + 0.278 + ], + "angle": 0, + "content": "(c) Applying CelebA tokens to Tai-Chi-HD" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.091, + 0.587, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.091, + 0.686, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.091, + 0.786, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.091, + 0.885, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.17, + 0.823, + 0.183 + ], + "angle": 0, + "content": "(b) Applying CUB-200-2011 tokens to Tai-Chi-HD" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.185, + 0.588, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.185, + 0.687, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.185, + 0.787, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.185, + 0.887, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.265, + 0.81, + 0.278 + ], + "angle": 0, + "content": "(d) Applying CelebA tokens to CUB-200-2011" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.894, + 0.319 + ], + "angle": 0, + "content": "Figure 5. Generalization - we apply our learned text tokens (keypoints) to images from other datasets, including those that are of completely different domains. Our tokens generalize well for data of similar type, and surprisingly well even for some extreme cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.468, + 0.464 + ], + "angle": 0, + "content": "aligned Human 3.6M achieves state-of-the-art performance despite using fewer keypoints \\((\\mathrm{K} = 10\\) vs \\(\\mathrm{K} = 16)\\). We also perform on par with the previous state of the art when we apply CUB-200-2011 tokens to Tai-Chi-HD—a case where the dataset gap is not only about the appearance but also beyond classes. Despite the drastic gap, our keypoints perform extremely well, leveraging the generalization power of large pre-trained diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.468, + 0.798 + ], + "angle": 0, + "content": "We show qualitative examples in Fig. 5. As shown, even when applied to different datasets, they look reasonable. For example, in Fig. 5a, when applying Tai-Chi-HD tokens to Human 3.6M, the tokens respond to the same locations on the human body as in Tai-Chi-HD. Surprising was when we applied CUB-200-2011 tokens to Tai-Chi-HD in Fig. 5b—they still responded to the body of the human being, reasonably consistently, although these tokens were trained to respond to birds. Of note are tokens two and six, which correspond to the front and back of the bird heads in Fig. 4—they also reply to the front and back of human heads. Applying CelebA tokens to Tai-Chi-HD in Fig. 5c also shows interesting outcomes, as tokens generally respond to human faces, despite the scale being drastically different between the two datasets. Finally, applying CelebA tokens to the CUB-200-2011 dataset in Fig. 5d shows mixed results—when it is 'successful' it focuses also on the faces of the bird, when it fails, it fails completely. These results hint that the keypoints (tokens) we have learned carry semantic meanings, as expected. We note that none of the baselines that we compare against are able to generalize beyond the dataset they were trained for." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.204, + 0.83 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We have proposed a novel method to find unsupervised keypoints using pre-trained text-to-image diffusion models. Given a set of images of a certain object, we propose to optimize the text embeddings (tokens) such that the cross" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.341, + 0.89, + 0.399 + ], + "angle": 0, + "content": "
Tai-Chi-HDCUB-200-2011CelebACelebA
→unaligned→Tai-Chi-HD→Tai-Chi-HD→CUB-200-2011
Human3.6m(K=10)↓(K=8)↓(K=8)↓
(K=10)↓Cum ℓ2 ↓ / Rel. ℓ2 ↓Cum ℓ2 ↓ / Rel. ℓ2 ↓
Ours4.88317.94 / 9.50- / 8.618.60
Autolink [13]16.92535.61 / 16.00- / 28.222.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 5. Generalization - we quantitatively evaluate the performance of our keypoints on other datasets. Our Tai-Chi-HD keypoints applied to the unaligned Human3.6m setting reach state-of-the-art performance. Interestingly, our CUB keypoints applied to Tai-Chi-HD are on par with the previous state of the art, despite the differences between these datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.64 + ], + "angle": 0, + "content": "attention maps within diffusion models become localized as Gaussians with a small standard deviation. By doing so, we find text tokens that can be used to extract keypoints by extracting the maxima of the attention maps. We have shown that our method, on multiple datasets, under the challenging un-aligned setup, significantly outperforms the state of the art. We have further demonstrated that these tokens are also generalizable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.653, + 0.679, + 0.67 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.678, + 0.892, + 0.737 + ], + "angle": 0, + "content": "The authors would like to thank Cristina Vasconcelos for her constructive feedback during the preparation of this manuscript. Additionally, we extend our gratitude to David Fleet for his approval and support of this work." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.737, + 0.892, + 0.821 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant, NSERC Collaborative Research and Development Grant, Google, Digital Research Alliance of Canada, and Advanced Research Computing at the University of British Columbia." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "diffusion models improves imagenet classification. Transactions on Machine Learning Research, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.236, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection. Proceedings of the IEEE International Conference on Computer Vision, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.292, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[5] Weiya Chen, Chenchen Yu, Chenyu Tu, Zehua Lyu, Jing Tang, Shiqi Ou, Yan Fu, and Zhidong Xue. A survey on hand pose estimation with wearable sensors and computer- vision-based methods. Sensors, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[6] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.406, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv Preprint, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.477, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[8] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero-shot classifiers. International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[9] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[11] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. International Conference on Learning Representations, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Conference on Computer Vision and Pattern Recognition, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[17] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Marwa Jabberi, Ali Wali, Bidyut Baran Chaudhuri, and Adel M Alimi. 68 landmarks are efficient for 3d face alignment: what about more? 3d face alignment method applied to face recognition. Multimedia Tools and Applications, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in Neural Information Processing Systems, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[21] You-Yi Jau, Rui Zhu, Hao Su, and Manmohan Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In International Conference on Intelligent Robots and Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Le Jiang, Caleb Lee, Divyang Teotia, and Sarah Ostadabbas. Animal pose estimation: A closer look at the state-of-the-art, existing gaps and opportunities. Computer Vision and Image Understanding, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[23] Yuhe Jin, Weiwei Sun, Jan Hosang, Eduard Trulls, and Kwang Moo Yi. Tusk: Task-agnostic unsupervised keypoints. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[24] Aliasghar Khani, Saeid Asgari Taghanaki, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. arXiv Preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[25] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Wu Liu, Qian Bao, Yu Sun, and Tao Mei. Recent advances of monocular 2d and 3d human pose estimation: A deep learning perspective. ACM Computing Surveys, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22828" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[30] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[31] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv Preprint, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[33] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.472, + 0.438 + ], + "angle": 0, + "content": "[34] Giorgia Marullo, Leonardo Tanzi, Pietro Piazzolla, and Enrico Vezzetti. 6d object position estimation from 2d images: a literature review. Multimedia Tools and Applications, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.472, + 0.509 + ], + "angle": 0, + "content": "[35] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[37] Aditya Nandy, Chenru Duan, and Heather J Kulik. Audacity of huge: overcoming challenges of data scarcity and data quality for machine learning in computational materials discovery. Current Opinion in Chemical Engineering, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[38] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.354, + 0.692 + ], + "angle": 0, + "content": "[39] OpenAI. Gpt-4 technical report, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[40] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[42] Linhao Qu, Siyu Liu, Xiaoyu Liu, Manning Wang, and Zhijian Song. Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv Preprint, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Helena Russello, Rik van der Tol, and Gert Kootstra. T-leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and Electronics in Agriculture, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[47] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[49] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[50] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[51] Aliaksandr Siarohin, Subhankar Roy, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Motion-supervised co-part segmentation. In International Conference on Pattern Recognition, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[52] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.894, + 0.814 + ], + "angle": 0, + "content": "[53] Jennifer J Sun, Serim Ryou, Roni H Goldshmid, Brandon Weissboud, John O Dabiri, David J Anderson, Ann Kennedy, Yisong Yue, and Pietro Perona. Self-supervised keypoint discovery in behavioral videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[55] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22829" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "embeddings. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[56] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv Preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv Preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[59] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv Preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.468, + 0.418 + ], + "angle": 0, + "content": "[60] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[61] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. arXiv Preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.468, + 0.504 + ], + "angle": 0, + "content": "[62] Yue Wu and Qiang Ji. Facial landmark detection: A literature survey. International Journal of Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[63] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. arXiv Preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[65] Lumin Xu, Sheng Jin, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[68] Yanping Zhang, Qiaokang Liang, Kunlin Zou, Zhengwei Li, Wei Sun, and Yaonan Wang. Self-supervised part segmentation via motion imitation. Image and Vision Computing, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[69] Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22830" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_origin.pdf b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2188be739742488317289d8a0468462727bf5ca9 --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/4645df0d-9e97-449d-a096-f3522e36e6da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7087f29fa94cfc39d278d943dd37cca5e348f597740a78abd9e121a682058ef +size 9842438 diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/full.md b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f5b378b83d6d718031d723237da215590d158b53 --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/full.md @@ -0,0 +1,368 @@ +# Unsupervised Keypoints from Pretrained Diffusion Models + +Eric Hedlin1 Gopal Sharma1 Shweta Mahajan1,2 Xingzhe He1 Hossam Isack3 Abhishek Kar3 Helge Rhodin1 Andrea Tagliaasacchi4,5,6 Kwang Moo Yi1 1 University of British Columbia 2 Vector Institute for AI 3 Google Research 4 Google DeepMind 5 Simon Fraser University 6 University of Toronto + +https://stablekeypoints.github.io/ + +![](images/3cb4f3fa1d3077ec8e473cca9fcb28f10889daff91cb054896eb7458d1e116ba.jpg) +Figure 1. Teaser - we propose an unsupervised method to learn keypoints based on optimizing text embeddings of latent diffusion models [44]. Our method is motivated by the fact that random text tokens already respond roughly consistently to semantically similar regions. By promoting localization we obtain unsupervised keypoints that outperform the state-of-the-art. + +# Abstract + +Unsupervised learning of keypoints and landmarks has seen significant progress with the help of modern neural network architectures, but performance is yet to match the supervised counterpart, making their practicability questionable. We leverage the emergent knowledge within text-to-image diffusion models, towards more robust unsupervised keypoints. Our core idea is to find text embeddings that would cause the generative model to consistently attend to compact regions in images (i.e. keypoints). To do so, we simply optimize the text embedding such that the cross-attention maps within the denoising network are localized as Gaussians with small standard deviations. We validate our performance on multiple datasets: the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human3.6m datasets. We achieve significantly improved accuracy, sometimes even outperforming supervised ones, particularly for data that is non-aligned and less curated. Our code is publicly available at the project page. + +# 1. Introduction + +Keypoints or landmarks have played a critical role in computer vision for various tasks including image matching [31], 3D reconstruction [18], and motion tracking [32, 60]. Similarly to many other areas of computer vision, research has quickly adopted supervised learning to tackle this problem [3, 27]. However, labeling is tedious and sometimes even ambiguous—for example, it is difficult to consistently decide which keypoints on a human face are the "most important". Researchers have therefore been investigating unsupervised approaches [12, 13, 19, 30, 55, 67]. These are typically implemented as autoencoders paired with handcrafted intermediate layers, or losses that enforce spatial locality and equivariance of keypoint locations under deformation. However, as we will show later, these methods struggle with non-preprocessed data, and their performance is heavily reliant on knowing the ground truth location of objects, clearly limiting their practical applicability. + +To enhance the learning of unsupervised keypoints, we draw inspiration from the demonstrated success of scaling up datasets [52]. For example, in natural language pro + +![](images/030a330169ab53748481597408ce0732f89700674297f49a5a01c43ffb13c6cf.jpg) +Figure 2. Example attention maps - we show example attention maps for a selected learned keypoint for the CUB-200-2011 dataset, on the CUB-aligned subset. As shown, our keypoint attention map responds consistently across varying images. + +cessing performance has recently improved to a great extent thanks to large models and data [7, 39, 57]. Similarly, in computer vision, the performance of text-to-image models [43, 44, 46] has drastically improved thanks to the availability of extra large datasets [48]. However, unsupervised keypoint learning typically assumes class-specific datasets, e.g., animals that have a shared skeleton that connects keypoints, and these datasets are small in scale. + +Rather than collecting larger domain-specific datasets, we instead propose to leverage the knowledge stored within large generative models, such as Stable Diffusion [44]. This has been shown to be very effective across a number of tasks [1, 2, 4, 8, 15, 24, 26, 33, 35, 41, 54, 56, 59, 61, 63, 64, 66], but, to the best of our knowledge, it has not yet found application for the task of keypoint learning. Our main idea is to localize "important" keypoints by finding text embeddings that consistently correspond to a distinct location in images of a certain object class. This idea is rooted in the observation that, even with random text embeddings, the attention maps for various images roughly correspond to regions that are semantically similar; see Fig. 1. Therefore, text embeddings carry semantic meaning, which could be used to relate collections of images to each other; see Fig. 2. + +We find embeddings that are specific to certain locations by enforcing localized attention maps. In more detail, we propose to find (i.e., optimize) a set of tokens in a text embedding that locally responds in the Stable Diffusion cross-attention layers. We enforce locality by maximizing the similarity of the attention responses of each token to a single-mode Gaussian distribution. Thanks to the way the cross-attention layers are constructed within Stable Diffusion, this simple objective also prevents the different tokens from attending to the same locations in an image, a common degenerate solution that typically requires explicit workarounds [23]. + +We evaluate our method on established benchmarks: CelebA [28], CUB-200-2011 [58], Tai-Chi-HD [50], DeepFashion [29], and Human3.6m [17]. Our approach yields results on par with state-of-the-art methods for well-curated and aligned datasets, while notably enhancing performance + +for in-the-wild setups, particularly with unaligned data, sometimes even surpassing fully supervised baselines. + +# 2. Related Work + +Below we review the literature of finding keypoints in unsupervised and supervised fashion, along with work that exploits large pre-trained models like stable-diffusion for lower-level computer vision tasks. + +Learning keypoints with supervision. Pose estimation and landmark estimation are fundamental problems in computer vision. They naturally arise in various tasks, including human [69] and animal pose estimation [22], hand [5] and face landmark estimation [62], and object pose tracking [34]. Many fully supervised methods find different ways to induce some prior within the model to better capture the task at hand, such as using part affinity fields [3], temporal consistency for video data [45], spacial relationships [65], and geometry constraints [21] among others. While fully supervised methods have excelled in categories with abundant labeled data, such as human pose estimation, their major drawback is the insatiable need for large and high-quality datasets [22, 37, 42]. The scalability of gathering such extensive and meticulously annotated data for every conceivable object category remains a significant drawback [22, 37, 42]. + +Learning keypoints via self-supervision. The amount of unlabeled data far exceeds that of labeled data, so unsupervised keypoint estimation methods attempt to take advantage of this. Self-supervised keypoint detection often relies on tracking how keypoints move with image changes and uses various constraints for known transformations [16, 19, 30, 49, 55, 67], but these methods can struggle with background modeling [49, 67] and pose variations [16]. One can also rely on image reconstruction to learn keypoints. Some methods use GANs to generate images from keypoints [12, 14], but this often results in training instability. Alternatively, auto-encoders can be also used [13, 67], but these require training from scratch on each dataset. Our method neither suffers from GAN + +training instability, nor requires dataset fine-tuning. Finally, there exist self-supervised methods that exploit skeletal representations [13, 20, 40]. However, many of these approaches generally require known keypoint connectivity and video data [20, 40], and often face limitations in background handling and generalizability to objects within the same class [13, 20, 40]. Our method has no object-specific priors, and generalizes well due to the large dataset used by the pre-trained diffusion models. + +Diffusion models for image understanding. Recently, large image diffusion models have reached impressive image generation quality [43, 44, 46]. These models learn priors for real images within the latent space of the diffusion model, and provide a useful initialization for many downstream tasks such as image correspondence [15, 33, 54, 66], object detection [4], semantic segmentation [2, 24, 56, 59, 61, 63, 64], and image classification [1, 8]. Interestingly, without requiring any retraining, these models demonstrate an innate ability to understand 3D spatial configurations [26, 35, 41]. Recent work in each of these areas has shown the emergent power of these large models, most of them using the model without any modifications or extra supervision required. More relevant to our work, Mokady et al. [36] found that the pre-trained Stable Diffusion [44] model's cross-attention maps connect text tokens to semantically relevant areas in images. + +Correspondences via diffusion models. Among works that re-purpose diffusion models, of high relevance, is the effectiveness of diffusion models in correspondence estimation tasks [15, 33, 54, 66]. Hedlin et al. [15] optimizes the attention map for a specific point in a source image and finds the corresponding activation in a target image. However, this method requires a query to be provided in the source image. While our method shares the same inspiration of utilizing attention maps, critically, rather than optimizing the embedding given a single image, we optimize an embedding given a dataset of images from a given object class. In other words, our method discovers on its own, where to focus, rather than relying on user input. Our task is therefore changed from image matching between two images, to semantic matching across all images within the dataset. + +# 3. Method + +To identify a set of representative keypoints across a dataset of images, we formulate our approach in an unsupervised framework leveraging conditional diffusion models; see Fig. 3. In particular, we utilize the cross-attention maps between the text embeddings and the image features, derived from the latent diffusion model [44], and force them to consistently concentrate their activation on highly localized regions within the images. While Hedlin et al. [15] employed a similar mechanism (i.e., given a set of keypoint locations + +in one image, identify correspondences in another image), in this work we seek to identify semantic correspondences across all images within a class-specific dataset (e.g., human faces), without any given knowledge on what and where to focus. We show that this is possible simply by enforcing locality and equivariance to transformations. + +Let us start by quickly reviewing the fundamentals of diffusion models and formalizing the attention maps that we will utilize within these models (Sec. 3.1). We then detail the objectives used to learn the text embeddings that represent keypoints (Sec. 3.2) and discuss important implementation details (Sec. 3.3). + +# 3.1. Attention maps in diffusion networks + +Diffusion models are a class of generative models that approximate the data distribution by denoising a base (typically Gaussian) distribution [38]. A latent diffusion model operates on a latent representation $\mathbf{z}$ rather than the image itself, with an encoder that maps an image $\mathbf{X}$ into a latent $\mathbf{z}$ , and a decoder that maps $\mathbf{z}$ into $\mathbf{X}$ . These models define a forward diffusion process, where the latent representation $\mathbf{z}$ is gradually transformed into Gaussian noise over a series of $T$ time steps. The inverse process, over a sequence denoising steps $t = 1,\dots,T$ predicts the latent noise $\epsilon_{\theta}(\mathbf{z}_t,t)$ which was gradually added in each iteration in order to recover the original (latent) signal. + +In our work, we are interested in conditional diffusion models, and the explicit attentional relationship between the condition (i.e. text) and the generated outcome (i.e. image) that these models learn. Typically, diffusion models are made conditional on some text $\mathbf{y}$ , by providing an embedding $\mathbf{e} = \tau_{\theta}(\mathbf{y})$ from a text encoder $\tau_{\theta}$ to the denoiser. They are then trained to optimize + +$$ +\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {\mathbf {z}, t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\boldsymbol {\theta}} (\mathbf {z} _ {t}, t, \mathbf {e}) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where the denoiser $\epsilon_{\theta}(\mathbf{z}_t,t,\mathbf{e})$ is typically implemented by a transformer architecture [38] involving a combination of self-attention and cross-attention layers. Of our interest here is the cross-attention layers that relate $\mathbf{e}$ to $\mathbf{z}_t$ , which we now formalize. + +Specifically, in the transformer part of the model, denote $\Phi_l^c (\cdot)$ and $\Psi_l^c (\cdot)$ as the $c$ -th head and the $l$ -th linear layers of the U-Net. We calculate the query as $\mathbf{Q}_l^c = \Phi_l^c (\mathbf{z}_{t = 1})\in \mathbb{R}^{(H\times W)\times D_l}$ , and the key from the language embedding $\mathbf{K}_l^c = \Psi_l^c (\mathbf{e})\in \mathbb{R}^{N\times D_l}$ , where $N$ is the number of tokens, $C$ the number of heads in the transformer attention layer, $H$ and $W$ are image height and width at that specific layer in U-Net, and $D_{l}$ the dimensionality of the layer. Given query and key, the cross-attention map $\mathbf{M}_l\in \mathbb{R}^{(H\times W)\times N}$ is then computed via softmax along the $N$ di + +![](images/005224cd5e3204ccb776ddb962aebd704a81c8241103c6fe3c7660c71aebf81d.jpg) +Figure 3. Overview - we pass a randomly initialized text embedding into Stable Diffusion [44] and extract the attention maps. We then optimize the text embedding to have localized attention maps, by supervising them to become a single-mode Gaussian distribution, drawn at the location of their maxima. We also enforce attention maps to be transformation equivariant to small affine transformations on images. We repeat this process over a set of training images, which after optimization provides a set of $K$ keypoints. + +mension, and average pooling along the $C$ dimension: + +$$ +\mathbf {M} _ {l} (\mathbf {e}, \mathbf {X}) = \mathbb {E} _ {c} \left[ \operatorname {s o f t m a x} _ {n} \left(\mathbf {Q} _ {l} ^ {c} \cdot \mathbf {K} _ {l} / \sqrt {D _ {l}}\right) \right]. \tag {2} +$$ + +As various layers of the U-Net exhibit distinct levels of semantic understanding, following Hedlin et al. [15], we collect this information by average pooling across a selection of layers: + +$$ +\tilde {\mathbf {M}} = \mathbb {E} _ {l = 7.. 1 0} \left[ \mathbf {M} _ {l} (\mathbf {e}, \mathbf {X}) \right] \in \mathbb {R} ^ {(H \times W) \times N}. \tag {3} +$$ + +In what follows, to lighten the notation, we drop the attention mask arguments $(\mathbf{e},\mathbf{X})$ and write the attention map for the n-th token as $\tilde{\mathbf{M}}_n$ . + +# 3.2. Optimizing to find the keypoint embeddings + +To obtain a text embedding that can be used to locate keypoints, for each of them, we simply optimize for two objectives that respectively encourages localization and equivariance to geometric transformations. We thus write + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {l o c a l i z e}} + \lambda_ {\text {e q u i v}} \mathcal {L} _ {\text {e q u i v}}, \tag {4} +$$ + +where we apply $\lambda_{\mathrm{equiv}} = 10$ to balance the two losses to be in a similar operating range. Equivariance is enforced in the typical form of learning to be invariant to transformations. We first quickly detail $\mathcal{L}_{\mathrm{equiv}}$ , and then discuss how we enforce localization, which is the core of our method. + +Equivariance - $\mathcal{L}_{\mathrm{equiv}}$ . To ensure our model's attention mechanism remains consistent across different geometric transformations $\mathcal{T}$ of the input, we use the typical equivariance loss [25]: + +$$ +\mathcal {L} _ {\text {e q u i v}} = \mathbb {E} _ {n} \| \mathcal {T} ^ {- 1} \left(\mathbf {M} _ {n} (\mathbf {e}, \mathcal {T} (\mathbf {X})) - \mathbf {M} _ {n} (\mathbf {e}, \mathbf {X}) \right\| ^ {2} \tag {5} +$$ + +For $\mathcal{T}$ we simply utilize minor affine transformations. We use random rotations between $\pm 15$ degrees, translations between $\pm 0.25 \times W$ , and scaling between $100 - 120\%$ of the original image size. + +Encouraging localization - $\mathcal{L}_{\mathrm{localize}}$ . We encourage localization by forcing $\tilde{\mathbf{M}}_n$ to be a single-mode Gaussian distribution located at its maximum. In more details, denoting the Gaussian image that shares the same maximum as $\tilde{\mathbf{M}}_n$ as $\mathbf{G}_n$ , we write + +$$ +\mathcal {L} _ {\text {l o c a l i z e}} = \mathbb {E} _ {n} \| \tilde {\mathbf {M}} _ {n} - \mathbf {G} _ {n} \| ^ {2}. \tag {6} +$$ + +To create the Gaussian images $\mathbf{G}_n$ , we first identify the spatial location exhibiting the maximal response within the heatmap corresponding to each token $n$ by taking the argmax: + +$$ +\boldsymbol {\mu} _ {n} = \underset {w, h} {\arg \max } \tilde {\mathbf {M}} _ {n} [ h, w ]. \tag {7} +$$ + +We then generate a Gaussian image; see Fig. 3: + +$$ +\mathbf {G} _ {n} = \exp \left(- \frac {\| \mathbf {X Y} _ {\mathrm {c o o r d}} - \boldsymbol {\mu} _ {n} \| _ {2} ^ {2}}{2 \sigma^ {2}}\right), \tag {8} +$$ + +where $\mathbf{X}\mathbf{Y}_{\mathrm{coord}}$ is a tensor of image coordinates. + +Promoting mutual exclusivity. It is important to note that while $\mathcal{L}_{\mathrm{localize}}$ in (6) at first glance seem to only encourage localization, it also enforces $\tilde{\mathbf{M}}_n$ to be mutually exclusive for different $n$ because of the softmax operation in (2). Should multiple embeddings become similar, their attention responses in (2) $\mathbf{Q}_l^c\cdot \mathbf{K}_l$ will become similar, resulting in the softmax of the attention map being a flat response (i.e. deviating from a Gaussian shape). In other words, (6) naturally enforces exclusivity with the help of (2). + +Stabilizing optimization by working with a subset. We noticed in our experiments that attention maps $\mathbf{M}$ for some + +tokens can be 'spread' for some images, e.g., due to occlusions, destabilizing optimization. We thus opt for a simple solution of looking into the top-K tokens that are local. Specifically, we apply our losses over $n \in \mathcal{N}(\kappa)$ , which returns the $\kappa \in \mathbb{N}$ entries with the most spatially localized heatmap responses2, as measured by KL divergence: + +$$ +\mathcal {N} (\kappa) = \operatorname {A r g T o p} _ {\kappa} \left\{- \mathrm {K L} \left(\mathbf {G} _ {n}, \tilde {\mathbf {M}} _ {n}\right) \right\} _ {n = 1} ^ {N}. \tag {9} +$$ + +Final keypoints. While $\mathcal{L}_{\mathrm{localize}}$ naturally enforces exclusivity, it does not guarantee a complete coverage of the object. Thus, after we finish optimizing, we refine the set of keypoints through furthest point sampling using the training images. Specifically, for each image we write: + +$$ +\mathcal {K} = \operatorname {F P S} _ {K} (\{\boldsymbol {\mu} _ {i} \mid i \in \mathcal {N} (\kappa) \}), \tag {10} +$$ + +where $K$ is the desired number of keypoints $K < \kappa$ . Then, as the set $\mathcal{K}$ differs in each image, we simply choose $K$ to tokens that appeared most frequently in $\mathcal{K}$ within the training image set. + +# 3.3. Implementation details + +Test-time ensembling. At inference time, as in common literature [3, 10, 15], rather than employing the attention map of the original image, we average the attention maps across multiple augmentations (we use the same transformations for test time augmentation as in (5)) of the image: + +$$ +\tilde {\mathbf {M}} \leftarrow \sum_ {i} \mathcal {T} _ {i} ^ {- 1} \left(\mathbf {M} \left(\mathbf {e}, \mathcal {T} _ {i} (\mathbf {X})\right)\right). \tag {11} +$$ + +Upsampling attention maps. The attention maps in (2) are typically of low resolution. Specifically, as we use Stable Diffusion 1.5 [44], depending on the layer we extract the attention maps from, they are either $16 \times 16$ or $32 \times 32$ . We thus opt to upsample the query $\mathbf{Q}$ via bicubic interpolation to achieve a standard resolution of $128 \times 128$ . We have experimented with other upsampling techniques such as the commonly used bilinear sampling or a learned upsampler that is trained alongside, but a simple bicubic upsample was shown to be effective. + +# 4. Results + +# 4.1. Experimental setup + +We evaluate our method on five standard datasets for unsupervised keypoint evaluation: + +CelebA dataset [28]: A dataset of 202,599 facial images of celebrities. We evaluate both the aligned and nonaligned cases following the standard protocol of omitting images with faces occupying less than $30\%$ of the image. + +The standard metric for this dataset is to measure the average $\ell_2$ error normalized by the inter-ocular distance. + +- CUB-200-2011 dataset [58]: This dataset consists of 11,788 bird images. We use both the aligned (CUB-aligned) and non-aligned (CUB-all) variants. For the nonaligned variants, we further look at CUB-001, CUB-002, and CUB-003, which are specific bird subcategories. Notably, these subsets contain only 30 images each—we only use these 30 for training. We follow the standard protocol [6, 30] and normalize the images to be of $256 \times 256$ . The standard metric for this dataset is the mean $\ell_2$ error, normalized by the dimension of the images after normalization. +- Tai-Chi-HD dataset [50]: This dataset contains 3049 training videos and 285 test videos of people performing Tai-Chi, which shows more diverse poses compared to the other datasets, and is the most challenging among the human pose-centric datasets that we use. We follow Siarohin et al. [51] and use 500 images for testing and 300 images for training. The standard metric for this dataset is to measure the accumulated $\ell_2$ error, with the images standardized to $256 \times 256$ . +- DeepFashion dataset [29]: This dataset contains 53k images of fashion models, mostly standing with a white background. We follow Lorenz et al. [30] and only keep full body images. This leaves 10,604 images for training and 1,179 images for testing. Also following the baselines, we use keypoints generated by AlphaPose [11] as ground truth. The standard metric for this dataset is the percentage of correct keypoints (PCK) with a 6-pixel threshold. +- Human 3.6M dataset [17]: This dataset is of humans performing various actions, comprised of 3.6 million images. We follow the standard protocol [67] and focus on six activities: direction, discussion, posing, waiting, greeting, and walking. We utilize subjects 1, 5, 6, 7, 8, and 9 for training, while subject 11 is reserved for testing. This division yields a training dataset comprising 796,648 images and a testing dataset containing 87,975 images. The background for this dataset is also simple, and often masked out with ground-truth masks for evaluation. This dataset is also typically heavily pre-processed and aligned when used for unsupervised keypoint evaluation. We experiment with the standard pre-processing [30, 67] and also a relaxed version of our own. To relax the alignment, we crop a square bounding box such that the margin from the bounding box to the person is 100 pixels, which on average corresponds to the person's height being $2/3$ of the crop. We further add a uniform random translation up to 100 pixels (same as the margin) to remove the central bias. Example crops are visualized in Fig. 4f. The standard metric for this dataset is the $\ell_2$ error after normalizing the image resolution to $128 \times 128$ . + +
MethodAligned (K=10)↓Wild (K=4) ↓Wild (K=8) ↓
Thewlis et al. [55]7.95-31.30
Zhang et al. [67]3.46-40.82
LatentKeypointGAN [12]5.8525.8121.90
Lorenz et al. [30]3.2415.4911.41
IMM [19]3.1919.428.74
LatentKeypointGAN-tuned [12]3.3112.105.63
Autolink [13]3.927.725.66
Autolink † [13]3.546.115.24
Our method3.605.244.35
+ +Note that each dataset comes with its own metric. To make results more comparable across the human pose datasets, we report both their original metrics as well as the $\ell_2$ error when normalizing the image resolution to $128 \times 128$ . + +Regressing human-annotated landmarks. To evaluate the quality of unsupervised keypoints, one must relate them with human-annotated landmarks. As in prior research [55], we use linear regression (without bias) to relate between unsupervised keypoints and human-annotated landmarks. + +Number of keypoints and hyperparameters. For each method, we use the standard number of unsupervised keypoints defined for each evaluation protocol—we denote them in our Tables. We use the same hyperparameter for all our experiments as introduced in Sec. 3.2, except for the number of optimization iterations. We optimize the embeddings for 10k iterations, except for the human pose datasets, for which we optimize 500 iterations. To find the number of optimization rounds we use a $10\%$ validation subset from the training data. While we observed our results on the validation subset to improve consistently for most datasets, we found 10k to give a reasonable optimization time of two hours on an RTX 3090. For the human pose dataset, we found optimization to have converged already at 500 iterations on our validation split. + +# 4.2. Experimental results + +Quantitative results – Tabs. 1 to 3. We report our results for each dataset in Tabs. 1 to 3. As shown, except for the case when data is heavily processed and aligned (CelebA aligned in Tab. 1, CUB-aligned in Tab. 2, and Human 3.6M in Tab. 3), our method significantly outperforms the state of the art. The most visible gains are for the Tai-Chi-HD dataset, the most challenging among human pose datasets, and on CUB unaligned datasets. For the CUB dataset and the Tai-Chi-HD datasets, we outperform even those that have been supervised with silhouettes or saliency maps. + +Table 1. Quantitave results for the CelebA dataset - we report results with the standard metrics. Our method performs best for non-aligned cases and is comparable to the state of the art for the aligned case. $\dagger$ symbol represents the thickness-tuned variant. + +
MethodSupervisionCUB-aligned (K=10) ↓CUB-001 (K=4) ↓CUB-002 (K=4) ↓CUB-003 (K=4) ↓CUB-all (K=4) ↓
SCOPS [16]GT silhouette-18.317.717.012.6
Choudhury et al. [6]GT silhouette-11.315.010.69.2
DFF [9]testing dataset-22.421.622.0-
SCOPS [16]saliency maps-18.518.821.1-
Lorenz et al. [30]unsupervised3.91----
ULD [55, 67]unsupervised-30.129.428.2-
Zhang et al. [67]unsupervised5.3626.927.627.122.4
LatentKeypointGAN [12]unsupervised5.2122.629.121.214.7
GANSeg [14]unsupervised3.2322.122.321.512.1
Autolink [13]unsupervised4.1520.620.319.711.6
Autolink † [13]unsupervised3.5120.219.218.511.3
Our methodunsupervised5.0610.511.110.35.4
+ +Table 2. Quantitave results for the CUB-200-2011 dataset - we report results with the standard metrics. Except for the CUB-aligned case, our method performs nearly twice better than the compared methods, even outperforming Choudhury et al. [6], which is supervised with ground-truth silhouettes. $\dagger$ represents the thickness-tuned variant. + +
MethodSupervisionHuman 3.6M (K=16) \( {\ell }_{2} \) standard / unaligned \( \downarrow \)DeepFashion (K=16) PCK† / Rel. \( {\ell }_{2} \downarrow \)Tai-Chi-HD (K=10) Cum \( {\ell }_{2} \downarrow \) / Rel. \( {\ell }_{2} \downarrow \)
Newell et al. [20]paired gt2.16 / ---
DFF [9]testing dataset--494.48 / 14.78
SCOPS [16]saliency maps--411.38 / 12.29
Jakab et al. [20]video*2.73 / ---
Siarohin et al. [51]videos--389.78 / 11.65
Zhang et al. [68]videos--343.67 / 10.27
Zhang et al. [67]videos4.14 / ---
Schmidtke et al. [47]video*3.31 / ---
Sun et al. [53]videos2.53 / ---
Thewis et al. [55]unsupervised7.51 / ---
Zhang et al. [67]unsupervised4.91 / ---
LatentKeypointGAN [12]unsupervised-49%437.69 / 13.08
Lorenz et al. [30]unsupervised2.79 / -57%-
GANSeg [14]unsupervised-59%417.17 / 12.47
autolink [13]unsupervised2.81 / 7.5965%337.50 / 10.08
autolink † [13]unsupervised2.76 / -66%316.10 / 9.45
Our methodunsupervised4.45 / 5.7770%/6.46234.89 / 7.02
+ +Table 3. Quantitative results for human pose datasets - We report results for the Human 3.6M dataset, Deep Fashion dataset, and the Challenging Tai-Chi-HD datasets. We report both standard metrics for each dataset and the relative $\ell_2$ error after normalizing images to $128\times 128$ . Our method, except for the Human 3.6M dataset that is heavily pre-processed, outperforms all baselines. This includes, for the challenging Tai-Chi-HD dataset, supervised ones. * denotes additional supervision (Jakab et al. [20] uses unpaired ground truth, and Schmidtke et al. [47] use the T-pose). The † symbol represents the thickness-tuned variant for Autolink. + +We note that our primary focus is on unaligned cases, as we argue that they represent more how keypoints would be used in real-world applications—most real-world datasets are unaligned except for specific classes of objects. Moreover, methods focusing on aligned settings use strong locational priors, and as shown by their results in the unaligned setup—CelebA in the wild, non-aligned cases of Human 3.6M and CUB-200-2011, and Tai-Chi-HD—may perform significantly worse once this alignment prior is broken. Given that the performance of our method, even in the aligned case, is not too far off from methods that utilize alignment, we suspect a more in-depth tuning of our method may make our method outperform these methods, + +![](images/cfac4e55d86b17a0d932f67f6c95dd867bc2368608c3104b10fde8e9a2774792.jpg) +(a) CelebA dataset keypoints + +![](images/13ad7f932b8010928b28ab4573b299cd0af97687060db3d59eb7e93e85f34264.jpg) +(c) Tai-Chi-HD dataset keypoints + +![](images/a19a4af91ac82dd089039f3f1165af4a1ef579dec65c28fd286b0beea47f2954.jpg) +(e) Human 3.6M dataset keypoints + +![](images/775c9a914965358e6990af01426d09bacf9e0b22448aee63c631276642544812.jpg) +(b) CUB-200-2011 dataset keypoints + +![](images/d47aa7dd9374c394e3881732a00422cf7b4acb23986668ceeb0f575f8b3e7c67.jpg) + +![](images/b773bfeef0bda086fd21d31a6a9bb9250a2fbde1f585d9927cb0e65b64bc49fa.jpg) +(d) DeepFashion dataset keypoints +(f) Unaligned Human 3.6M dataset keypoints +Figure 4. Qualitative examples of unsupervised keypoints – we show our learned keypoints for the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human 3.6M datasets (both for cropped and masked as well as our relaxed version). Note how our keypoints are consistent despite the variability. Our method significantly outperforms other baselines, especially for the challenging Tai-Chi-HD dataset and the CUB subsets. + +
VariantNormalized ℓ2
Full (Our method)5.4
Without test time ensembling5.6
Without furthest point sampling6.4
Without upsampling the query Q8.0
Without equivariance22.2
+ +Table 4. Ablation results - we report the effect of each of our design choices can be seen on the CUB-all dataset. All components contribute to the final performance. + +but we leave this as future work. + +Finally, also note that for CUB-001, CUB-002, and CUB-003, these datasets are small. These datasets are nonaligned, have a large variability between individual images, and only contain 30 images each in the training set. Our method, just from 30 images, successfully identifies keypoints. These results highlight the potential of leveraging emergent (prior) knowledge within Stable Diffusion [44]. + +Qualitative results - Fig. 4. We provide example visualizations of our unsupervised keypoints in Fig. 4. As shown, our method discovers keypoints that are consistently localized across the dataset, despite the wide appearance variety. + +# 4.3. Ablation study + +We perform an ablation study for various design choices of our method on the CUB-all dataset. We report the performance of our method with different components disabled in Tab. 4. As shown, all components contribute to the final performance. Test-time ensembling enhances performance, + +but the computation cost linearly scales. We choose, ten augmentations, which provide a good compromise between computation time and accuracy. To remove furthest point sampling we set $\kappa = K$ , which then makes furthest point sampling select all samples. While this causes points to be more grouped, it still provides reasonable performance. To remove upsampling we instead upsample $\mathbf{M}$ to the size of the target image, effectively having the attention map build at lower resolutions, sometimes as low as $16\times 16$ . This results in significant degradation in performance. $\mathcal{L}_{\mathrm{equiv}}$ is essential, as without it, the tokens can 'cheat' and simply opt to learn fixed positions on the image. + +Number of training images.. Inspired by our results for the small subsets of CUB-200-2011 dataset, we investigate the impact that the number of images that we use to find keypoints has on our results. We thus optimized our keypoints only with 100 images for CelebA non-aligned setup. Surprisingly, we achieve $5.33K = 8$ , which is comparable to the state of the art. This demonstrates once more how our method is able to leverage information that is already learned in Stable Diffusion [44] to find keypoints. + +# 4.4. Generalization + +We further test the generalization capacity of our learned keypoints. As they are effectively text embeddings, we can simply apply them to any image, including those completely outside of the training domain. We quantitatively evaluate our method and the previous best-performing method Autolink [13]. We find that even in these generalization experiments, our keypoints reach performance comparable to data-specific keypoints. Applying Tai-Chi-HD tokens to un + +![](images/0ba9fff91d38fc99eb008a3134a950e3f5e39abd88f3b1e5f4c6c8ab923e95a5.jpg) +(a) Applying Tai-Chi-HD tokens to Human 3.6M + +![](images/ecf61c3715f5b6152c3764e1de86434edbb8072e2b8e28d4871e009674dde050.jpg) + +![](images/de0dd9cadffe29e8baa849b8e81201bd8fb53b92e47e97e084378bd8d4750473.jpg) + +![](images/1305e77d530ed04f30a880baa9c103b549b23b3f961d898a7a697c721edc09d9.jpg) + +![](images/af0296a3506232df22930460365fc59f08435495653c10a7149e6e62f1a98538.jpg) +(c) Applying CelebA tokens to Tai-Chi-HD + +![](images/21936bdc746586d3a09f0775637bdbe1a4b2bab990c1f073f2ed566a9d722cdd.jpg) +Figure 5. Generalization - we apply our learned text tokens (keypoints) to images from other datasets, including those that are of completely different domains. Our tokens generalize well for data of similar type, and surprisingly well even for some extreme cases. + +![](images/5f0808c04c992684bb235584fdecf88ab03992734f44284785c5a5f8a7a47242.jpg) + +![](images/f23345f7add49711f23a50d802a8750218690a6c68b7dfdd2309ebd6493810f2.jpg) + +![](images/27b8503d3f55eeb79d519fb91bda66a64a3469ce48f1b85cc6de5982d32225ac.jpg) +(b) Applying CUB-200-2011 tokens to Tai-Chi-HD + +![](images/e01427bd5156d67fab12cc8e2809a1854f6ce5245d01f91d0623d25e6f2ef48e.jpg) + +![](images/1a3648aa12019cd974626f865eed21a9314692809bff295b58154556010def7a.jpg) + +![](images/096ccc605c6dbec7da30f24be5f2d91c54dc932c82c8119ad9e93f3f057c3f88.jpg) + +![](images/27892a36cc3be66b89859e44332cc3dacc8e0a798c08666ccb1f965ae7861afc.jpg) +(d) Applying CelebA tokens to CUB-200-2011 + +![](images/a2dde74df91c1a29216fe061ba4e03245b85aa5268c6f25fbe2c67da2750d6b2.jpg) + +![](images/ffae4282d36896cc5d7ea4a1a0e22db8e4c7a9efa558d43e45707a4d9c4abe5b.jpg) + +![](images/48d59cb3d0a9089accf108ccd388c5adaa4c2178b91be90031aae28cbcb2df17.jpg) + +aligned Human 3.6M achieves state-of-the-art performance despite using fewer keypoints $(\mathrm{K} = 10$ vs $\mathrm{K} = 16)$ . We also perform on par with the previous state of the art when we apply CUB-200-2011 tokens to Tai-Chi-HD—a case where the dataset gap is not only about the appearance but also beyond classes. Despite the drastic gap, our keypoints perform extremely well, leveraging the generalization power of large pre-trained diffusion models. + +We show qualitative examples in Fig. 5. As shown, even when applied to different datasets, they look reasonable. For example, in Fig. 5a, when applying Tai-Chi-HD tokens to Human 3.6M, the tokens respond to the same locations on the human body as in Tai-Chi-HD. Surprising was when we applied CUB-200-2011 tokens to Tai-Chi-HD in Fig. 5b—they still responded to the body of the human being, reasonably consistently, although these tokens were trained to respond to birds. Of note are tokens two and six, which correspond to the front and back of the bird heads in Fig. 4—they also reply to the front and back of human heads. Applying CelebA tokens to Tai-Chi-HD in Fig. 5c also shows interesting outcomes, as tokens generally respond to human faces, despite the scale being drastically different between the two datasets. Finally, applying CelebA tokens to the CUB-200-2011 dataset in Fig. 5d shows mixed results—when it is 'successful' it focuses also on the faces of the bird, when it fails, it fails completely. These results hint that the keypoints (tokens) we have learned carry semantic meanings, as expected. We note that none of the baselines that we compare against are able to generalize beyond the dataset they were trained for. + +# 5. Conclusions + +We have proposed a novel method to find unsupervised keypoints using pre-trained text-to-image diffusion models. Given a set of images of a certain object, we propose to optimize the text embeddings (tokens) such that the cross + +
Tai-Chi-HDCUB-200-2011CelebACelebA
→unaligned→Tai-Chi-HD→Tai-Chi-HD→CUB-200-2011
Human3.6m(K=10)↓(K=8)↓(K=8)↓
(K=10)↓Cum ℓ2 ↓ / Rel. ℓ2 ↓Cum ℓ2 ↓ / Rel. ℓ2 ↓
Ours4.88317.94 / 9.50- / 8.618.60
Autolink [13]16.92535.61 / 16.00- / 28.222.56
+ +Table 5. Generalization - we quantitatively evaluate the performance of our keypoints on other datasets. Our Tai-Chi-HD keypoints applied to the unaligned Human3.6m setting reach state-of-the-art performance. Interestingly, our CUB keypoints applied to Tai-Chi-HD are on par with the previous state of the art, despite the differences between these datasets. + +attention maps within diffusion models become localized as Gaussians with a small standard deviation. By doing so, we find text tokens that can be used to extract keypoints by extracting the maxima of the attention maps. We have shown that our method, on multiple datasets, under the challenging un-aligned setup, significantly outperforms the state of the art. We have further demonstrated that these tokens are also generalizable. + +# 6. Acknowledgments + +The authors would like to thank Cristina Vasconcelos for her constructive feedback during the preparation of this manuscript. Additionally, we extend our gratitude to David Fleet for his approval and support of this work. + +This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant, NSERC Collaborative Research and Development Grant, Google, Digital Research Alliance of Canada, and Advanced Research Computing at the University of British Columbia. + +# References + +[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from + +diffusion models improves imagenet classification. Transactions on Machine Learning Research, 2023. 2, 3 +[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. International Conference on Learning Representations, 2021. 2, 3 +[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 5 +[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection. Proceedings of the IEEE International Conference on Computer Vision, 2022. 2, 3 +[5] Weiya Chen, Chenchen Yu, Chenyu Tu, Zehua Lyu, Jing Tang, Shiqi Ou, Yan Fu, and Zhidong Xue. A survey on hand pose estimation with wearable sensors and computer- vision-based methods. Sensors, 2020. 2 +[6] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 2021. 5, 6 +[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv Preprint, 2022. 2 +[8] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero-shot classifiers. International Conference on Learning Representations, 2023. 2, 3 +[9] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision, 2018. 6 +[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 5 +[11] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time, 2022. 5 +[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. International Conference on Learning Representations, 2021. 1, 2, 6 +[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 7, 8 +[14] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 6 +[15] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. + +Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5 +[16] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Conference on Computer Vision and Pattern Recognition, 2019. 2, 6 +[17] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013. 2, 5 +[18] Marwa Jabberi, Ali Wali, Bidyut Baran Chaudhuri, and Adel M Alimi. 68 landmarks are efficient for 3d face alignment: what about more? 3d face alignment method applied to face recognition. Multimedia Tools and Applications, 2023. 1 +[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in Neural Information Processing Systems, 2018. 1, 2, 6 +[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3, 6 +[21] You-Yi Jau, Rui Zhu, Hao Su, and Manmohan Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In International Conference on Intelligent Robots and Systems, 2020. 2 +[22] Le Jiang, Caleb Lee, Divyang Teotia, and Sarah Ostadabbas. Animal pose estimation: A closer look at the state-of-the-art, existing gaps and opportunities. Computer Vision and Image Understanding, 2022. 2 +[23] Yuhe Jin, Weiwei Sun, Jan Hosang, Eduard Trulls, and Kwang Moo Yi. Tusk: Task-agnostic unsupervised keypoints. Advances in Neural Information Processing Systems, 2022. 2 +[24] Aliasghar Khani, Saeid Asgari Taghanaki, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. arXiv Preprint, 2023. 2, 3 +[25] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 4 +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3 +[27] Wu Liu, Qian Bao, Yu Sun, and Tao Mei. Recent advances of monocular 2d and 3d human pose estimation: A deep learning perspective. ACM Computing Surveys, 2022. 1 +[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 2, 5 + +[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 2, 5 +[30] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 5, 6 +[31] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 2004. 1 +[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv Preprint, 2023. 1 +[33] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3 +[34] Giorgia Marullo, Leonardo Tanzi, Pietro Piazzolla, and Enrico Vezzetti. 6d object position estimation from 2d images: a literature review. Multimedia Tools and Applications, 2023. 2 +[35] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3 +[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3 +[37] Aditya Nandy, Chenru Duan, and Heather J Kulik. Audacity of huge: overcoming challenges of data scarcity and data quality for machine learning in computational materials discovery. Current Opinion in Chemical Engineering, 2022. 2 +[38] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 3 +[39] OpenAI. Gpt-4 technical report, 2023. 2 +[40] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision, 2018. 3 +[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2022. 2, 3 +[42] Linhao Qu, Siyu Liu, Xiaoyu Liu, Manning Wang, and Zhijian Song. Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology, 2022. 2 + +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv Preprint, 2022. 2, 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7 +[45] Helena Russello, Rik van der Tol, and Gert Kootstra. T-leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and Electronics in Agriculture, 2022. 2 +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 3 +[47] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 6 +[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2 +[49] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Conference on Computer Vision and Pattern Recognition, 2019. 2 +[50] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 5 +[51] Aliaksandr Siarohin, Subhankar Roy, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Motion-supervised co-part segmentation. In International Conference on Pattern Recognition, 2021. 5, 6 +[52] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1 +[53] Jennifer J Sun, Serim Ryou, Roni H Goldshmid, Brandon Weissboud, John O Dabiri, David J Anderson, Ann Kennedy, Yisong Yue, and Pietro Perona. Self-supervised keypoint discovery in behavioral videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6 +[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3 +[55] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial + +embeddings. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1, 2, 6 +[56] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv Preprint, 2023. 2, 3 +[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv Preprint, 2023. 2 +[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 2, 5 +[59] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv Preprint, 2023. 2, 3 +[60] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. International Conference on Computer Vision, 2023. 1 +[61] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. arXiv Preprint, 2023. 2, 3 +[62] Yue Wu and Qiang Ji. Facial landmark detection: A literature survey. International Journal of Computer Vision, 2019. 2 +[63] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. arXiv Preprint, 2023. 2, 3 +[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3 +[65] Lumin Xu, Sheng Jin, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3 +[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6 +[68] Yanping Zhang, Qiaokang Liang, Kunlin Zou, Zhengwei Li, Wei Sun, and Yaonan Wang. Self-supervised part segmentation via motion imitation. Image and Vision Computing, 2022. 6 + +[69] Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 2023. 2 \ No newline at end of file diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/images.zip b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03a509eee9c4b13c2cdcbff5f0ebb0fca61a01a3 --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:593c16e241b45607428df1a956b6dcf21b4aa83e7c31723bd3bf622cc8efb708 +size 610429 diff --git a/2024/Unsupervised Keypoints from Pretrained Diffusion Models/layout.json b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..20d64111f1b2d7a0949340dbcf8f791e21012c9e --- /dev/null +++ b/2024/Unsupervised Keypoints from Pretrained Diffusion Models/layout.json @@ -0,0 +1,9977 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "text", + "content": "Unsupervised Keypoints from Pretrained Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 141, + 501, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 141, + 501, + 200 + ], + "spans": [ + { + "bbox": [ + 91, + 141, + 501, + 200 + ], + "type": "text", + "content": "Eric Hedlin1 Gopal Sharma1 Shweta Mahajan1,2 Xingzhe He1 Hossam Isack3 Abhishek Kar3 Helge Rhodin1 Andrea Tagliaasacchi4,5,6 Kwang Moo Yi1 1 University of British Columbia 2 Vector Institute for AI 3 Google Research 4 Google DeepMind 5 Simon Fraser University 6 University of Toronto" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 201, + 388, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 201, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 201, + 201, + 388, + 213 + ], + "type": "text", + "content": "https://stablekeypoints.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 219, + 538, + 395 + ], + "blocks": [ + { + "bbox": [ + 53, + 219, + 538, + 395 + ], + "lines": [ + { + "bbox": [ + 53, + 219, + 538, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 538, + 395 + ], + "type": "image", + "image_path": "3cb4f3fa1d3077ec8e473cca9fcb28f10889daff91cb054896eb7458d1e116ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 398, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 398, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 546, + 431 + ], + "type": "text", + "content": "Figure 1. Teaser - we propose an unsupervised method to learn keypoints based on optimizing text embeddings of latent diffusion models [44]. Our method is motivated by the fact that random text tokens already respond roughly consistently to semantically similar regions. By promoting localization we obtain unsupervised keypoints that outperform the state-of-the-art." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 484, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 289, + 700 + ], + "type": "text", + "content": "Unsupervised learning of keypoints and landmarks has seen significant progress with the help of modern neural network architectures, but performance is yet to match the supervised counterpart, making their practicability questionable. We leverage the emergent knowledge within text-to-image diffusion models, towards more robust unsupervised keypoints. Our core idea is to find text embeddings that would cause the generative model to consistently attend to compact regions in images (i.e. keypoints). To do so, we simply optimize the text embedding such that the cross-attention maps within the denoising network are localized as Gaussians with small standard deviations. We validate our performance on multiple datasets: the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human3.6m datasets. We achieve significantly improved accuracy, sometimes even outperforming supervised ones, particularly for data that is non-aligned and less curated. Our code is publicly available at the project page." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 442, + 386, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 386, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 386, + 455 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 468, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 672 + ], + "type": "text", + "content": "Keypoints or landmarks have played a critical role in computer vision for various tasks including image matching [31], 3D reconstruction [18], and motion tracking [32, 60]. Similarly to many other areas of computer vision, research has quickly adopted supervised learning to tackle this problem [3, 27]. However, labeling is tedious and sometimes even ambiguous—for example, it is difficult to consistently decide which keypoints on a human face are the \"most important\". Researchers have therefore been investigating unsupervised approaches [12, 13, 19, 30, 55, 67]. These are typically implemented as autoencoders paired with handcrafted intermediate layers, or losses that enforce spatial locality and equivariance of keypoint locations under deformation. However, as we will show later, these methods struggle with non-preprocessed data, and their performance is heavily reliant on knowing the ground truth location of objects, clearly limiting their practical applicability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "To enhance the learning of unsupervised keypoints, we draw inspiration from the demonstrated success of scaling up datasets [52]. For example, in natural language pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22820" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 542, + 170 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 542, + 170 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 542, + 170 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 542, + 170 + ], + "type": "image", + "image_path": "030a330169ab53748481597408ce0732f89700674297f49a5a01c43ffb13c6cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "type": "text", + "content": "Figure 2. Example attention maps - we show example attention maps for a selected learned keypoint for the CUB-200-2011 dataset, on the CUB-aligned subset. As shown, our keypoint attention map responds consistently across varying images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 289, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 289, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 289, + 320 + ], + "type": "text", + "content": "cessing performance has recently improved to a great extent thanks to large models and data [7, 39, 57]. Similarly, in computer vision, the performance of text-to-image models [43, 44, 46] has drastically improved thanks to the availability of extra large datasets [48]. However, unsupervised keypoint learning typically assumes class-specific datasets, e.g., animals that have a shared skeleton that connects keypoints, and these datasets are small in scale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 503 + ], + "type": "text", + "content": "Rather than collecting larger domain-specific datasets, we instead propose to leverage the knowledge stored within large generative models, such as Stable Diffusion [44]. This has been shown to be very effective across a number of tasks [1, 2, 4, 8, 15, 24, 26, 33, 35, 41, 54, 56, 59, 61, 63, 64, 66], but, to the best of our knowledge, it has not yet found application for the task of keypoint learning. Our main idea is to localize \"important\" keypoints by finding text embeddings that consistently correspond to a distinct location in images of a certain object class. This idea is rooted in the observation that, even with random text embeddings, the attention maps for various images roughly correspond to regions that are semantically similar; see Fig. 1. Therefore, text embeddings carry semantic meaning, which could be used to relate collections of images to each other; see Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "type": "text", + "content": "We find embeddings that are specific to certain locations by enforcing localized attention maps. In more detail, we propose to find (i.e., optimize) a set of tokens in a text embedding that locally responds in the Stable Diffusion cross-attention layers. We enforce locality by maximizing the similarity of the attention responses of each token to a single-mode Gaussian distribution. Thanks to the way the cross-attention layers are constructed within Stable Diffusion, this simple objective also prevents the different tokens from attending to the same locations in an image, a common degenerate solution that typically requires explicit workarounds [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "We evaluate our method on established benchmarks: CelebA [28], CUB-200-2011 [58], Tai-Chi-HD [50], DeepFashion [29], and Human3.6m [17]. Our approach yields results on par with state-of-the-art methods for well-curated and aligned datasets, while notably enhancing performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "content": "for in-the-wild setups, particularly with unaligned data, sometimes even surpassing fully supervised baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 262, + 392, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 392, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 392, + 274 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 282, + 547, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 547, + 331 + ], + "type": "text", + "content": "Below we review the literature of finding keypoints in unsupervised and supervised fashion, along with work that exploits large pre-trained models like stable-diffusion for lower-level computer vision tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 337, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 540 + ], + "type": "text", + "content": "Learning keypoints with supervision. Pose estimation and landmark estimation are fundamental problems in computer vision. They naturally arise in various tasks, including human [69] and animal pose estimation [22], hand [5] and face landmark estimation [62], and object pose tracking [34]. Many fully supervised methods find different ways to induce some prior within the model to better capture the task at hand, such as using part affinity fields [3], temporal consistency for video data [45], spacial relationships [65], and geometry constraints [21] among others. While fully supervised methods have excelled in categories with abundant labeled data, such as human pose estimation, their major drawback is the insatiable need for large and high-quality datasets [22, 37, 42]. The scalability of gathering such extensive and meticulously annotated data for every conceivable object category remains a significant drawback [22, 37, 42]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "content": "Learning keypoints via self-supervision. The amount of unlabeled data far exceeds that of labeled data, so unsupervised keypoint estimation methods attempt to take advantage of this. Self-supervised keypoint detection often relies on tracking how keypoints move with image changes and uses various constraints for known transformations [16, 19, 30, 49, 55, 67], but these methods can struggle with background modeling [49, 67] and pose variations [16]. One can also rely on image reconstruction to learn keypoints. Some methods use GANs to generate images from keypoints [12, 14], but this often results in training instability. Alternatively, auto-encoders can be also used [13, 67], but these require training from scratch on each dataset. Our method neither suffers from GAN" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22821" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "training instability, nor requires dataset fine-tuning. Finally, there exist self-supervised methods that exploit skeletal representations [13, 20, 40]. However, many of these approaches generally require known keypoint connectivity and video data [20, 40], and often face limitations in background handling and generalizability to objects within the same class [13, 20, 40]. Our method has no object-specific priors, and generalizes well due to the large dataset used by the pre-trained diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 289, + 389 + ], + "type": "text", + "content": "Diffusion models for image understanding. Recently, large image diffusion models have reached impressive image generation quality [43, 44, 46]. These models learn priors for real images within the latent space of the diffusion model, and provide a useful initialization for many downstream tasks such as image correspondence [15, 33, 54, 66], object detection [4], semantic segmentation [2, 24, 56, 59, 61, 63, 64], and image classification [1, 8]. Interestingly, without requiring any retraining, these models demonstrate an innate ability to understand 3D spatial configurations [26, 35, 41]. Recent work in each of these areas has shown the emergent power of these large models, most of them using the model without any modifications or extra supervision required. More relevant to our work, Mokady et al. [36] found that the pre-trained Stable Diffusion [44] model's cross-attention maps connect text tokens to semantically relevant areas in images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 394, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 289, + 574 + ], + "type": "text", + "content": "Correspondences via diffusion models. Among works that re-purpose diffusion models, of high relevance, is the effectiveness of diffusion models in correspondence estimation tasks [15, 33, 54, 66]. Hedlin et al. [15] optimizes the attention map for a specific point in a source image and finds the corresponding activation in a target image. However, this method requires a query to be provided in the source image. While our method shares the same inspiration of utilizing attention maps, critically, rather than optimizing the embedding given a single image, we optimize an embedding given a dataset of images from a given object class. In other words, our method discovers on its own, where to focus, rather than relying on user input. Our task is therefore changed from image matching between two images, to semantic matching across all images within the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 585, + 104, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 104, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 104, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "To identify a set of representative keypoints across a dataset of images, we formulate our approach in an unsupervised framework leveraging conditional diffusion models; see Fig. 3. In particular, we utilize the cross-attention maps between the text embeddings and the image features, derived from the latent diffusion model [44], and force them to consistently concentrate their activation on highly localized regions within the images. While Hedlin et al. [15] employed a similar mechanism (i.e., given a set of keypoint locations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "in one image, identify correspondences in another image), in this work we seek to identify semantic correspondences across all images within a class-specific dataset (e.g., human faces), without any given knowledge on what and where to focus. We show that this is possible simply by enforcing locality and equivariance to transformations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 144, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 217 + ], + "type": "text", + "content": "Let us start by quickly reviewing the fundamentals of diffusion models and formalizing the attention maps that we will utilize within these models (Sec. 3.1). We then detail the objectives used to learn the text embeddings that represent keypoints (Sec. 3.2) and discuss important implementation details (Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 225, + 503, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 225, + 503, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 503, + 239 + ], + "type": "text", + "content": "3.1. Attention maps in diffusion networks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": "Diffusion models are a class of generative models that approximate the data distribution by denoising a base (typically Gaussian) distribution [38]. A latent diffusion model operates on a latent representation " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " rather than the image itself, with an encoder that maps an image " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " into a latent " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": ", and a decoder that maps " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": ". These models define a forward diffusion process, where the latent representation " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " is gradually transformed into Gaussian noise over a series of " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " time steps. The inverse process, over a sequence denoising steps " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "t = 1,\\dots,T" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " predicts the latent noise " + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,t)" + }, + { + "bbox": [ + 304, + 244, + 545, + 388 + ], + "type": "text", + "content": " which was gradually added in each iteration in order to recover the original (latent) signal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "text", + "content": "In our work, we are interested in conditional diffusion models, and the explicit attentional relationship between the condition (i.e. text) and the generated outcome (i.e. image) that these models learn. Typically, diffusion models are made conditional on some text " + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "text", + "content": ", by providing an embedding " + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = \\tau_{\\theta}(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "text", + "content": " from a text encoder " + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 304, + 388, + 545, + 473 + ], + "type": "text", + "content": " to the denoiser. They are then trained to optimize" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 483, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 483, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 336, + 483, + 545, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathbf {z}, t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}, t, \\mathbf {e}) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "24ba89a0bb4dbf8f49f438ee396b94171b6a90fc8131174b644d2fa5ed0caa9c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "text", + "content": "where the denoiser " + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathbf{e})" + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "text", + "content": " is typically implemented by a transformer architecture [38] involving a combination of self-attention and cross-attention layers. Of our interest here is the cross-attention layers that relate " + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 304, + 512, + 545, + 571 + ], + "type": "text", + "content": ", which we now formalize." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": "Specifically, in the transformer part of the model, denote " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\Phi_l^c (\\cdot)" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\Psi_l^c (\\cdot)" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": "-th head and the " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": "-th linear layers of the U-Net. We calculate the query as " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l^c = \\Phi_l^c (\\mathbf{z}_{t = 1})\\in \\mathbb{R}^{(H\\times W)\\times D_l}" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": ", and the key from the language embedding " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l^c = \\Psi_l^c (\\mathbf{e})\\in \\mathbb{R}^{N\\times D_l}" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " is the number of tokens, " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " the number of heads in the transformer attention layer, " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " are image height and width at that specific layer in U-Net, and " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " the dimensionality of the layer. Given query and key, the cross-attention map " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_l\\in \\mathbb{R}^{(H\\times W)\\times N}" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " is then computed via softmax along the " + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 573, + 545, + 693 + ], + "type": "text", + "content": " di" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "text", + "content": "1We choose " + }, + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 317, + 702, + 527, + 714 + ], + "type": "text", + "content": " steps via hyper-parameter tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22822" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 532, + 222 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 532, + 222 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 532, + 222 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 532, + 222 + ], + "type": "image", + "image_path": "005224cd5e3204ccb776ddb962aebd704a81c8241103c6fe3c7660c71aebf81d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 277 + ], + "type": "text", + "content": "Figure 3. Overview - we pass a randomly initialized text embedding into Stable Diffusion [44] and extract the attention maps. We then optimize the text embedding to have localized attention maps, by supervising them to become a single-mode Gaussian distribution, drawn at the location of their maxima. We also enforce attention maps to be transformation equivariant to small affine transformations on images. We repeat this process over a set of training images, which after optimization provides a set of " + }, + { + "bbox": [ + 46, + 233, + 547, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 233, + 547, + 277 + ], + "type": "text", + "content": " keypoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 298, + 266, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 266, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 266, + 310 + ], + "type": "text", + "content": "mension, and average pooling along the " + }, + { + "bbox": [ + 47, + 298, + 266, + 310 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 298, + 266, + 310 + ], + "type": "text", + "content": " dimension:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 319, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 319, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 70, + 319, + 287, + 340 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) = \\mathbb {E} _ {c} \\left[ \\operatorname {s o f t m a x} _ {n} \\left(\\mathbf {Q} _ {l} ^ {c} \\cdot \\mathbf {K} _ {l} / \\sqrt {D _ {l}}\\right) \\right]. \\tag {2}", + "image_path": "ce2de829197cb7b530f14462d3dd6a62f28fe762bf96f67ab78aa5ff4f64c1e5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "type": "text", + "content": "As various layers of the U-Net exhibit distinct levels of semantic understanding, following Hedlin et al. [15], we collect this information by average pooling across a selection of layers:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 405, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 405, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 80, + 405, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {M}} = \\mathbb {E} _ {l = 7.. 1 0} \\left[ \\mathbf {M} _ {l} (\\mathbf {e}, \\mathbf {X}) \\right] \\in \\mathbb {R} ^ {(H \\times W) \\times N}. \\tag {3}", + "image_path": "f5144001ee8c438cbf91bc8076b7e28f0b49acaad9a804e75e345df9e8a0ac99.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "content": "In what follows, to lighten the notation, we drop the attention mask arguments " + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "inline_equation", + "content": "(\\mathbf{e},\\mathbf{X})" + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "content": " and write the attention map for the n-th token as " + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}_n" + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 275, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 275, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 275, + 491 + ], + "type": "text", + "content": "3.2. Optimizing to find the keypoint embeddings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 497, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 544 + ], + "type": "text", + "content": "To obtain a text embedding that can be used to locate keypoints, for each of them, we simply optimize for two objectives that respectively encourages localization and equivariance to geometric transformations. We thus write" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 555, + 287, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 555, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 104, + 555, + 287, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {l o c a l i z e}} + \\lambda_ {\\text {e q u i v}} \\mathcal {L} _ {\\text {e q u i v}}, \\tag {4}", + "image_path": "fb83166d73b8cadd8bb46161b38d83d4c312e816d3132ed73c113ce1bf686195.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "text", + "content": "where we apply " + }, + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{equiv}} = 10" + }, + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "text", + "content": " to balance the two losses to be in a similar operating range. Equivariance is enforced in the typical form of learning to be invariant to transformations. We first quickly detail " + }, + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{equiv}}" + }, + { + "bbox": [ + 47, + 578, + 287, + 637 + ], + "type": "text", + "content": ", and then discuss how we enforce localization, which is the core of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "text", + "content": "Equivariance - " + }, + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{equiv}}" + }, + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "text", + "content": ". To ensure our model's attention mechanism remains consistent across different geometric transformations " + }, + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 643, + 287, + 690 + ], + "type": "text", + "content": " of the input, we use the typical equivariance loss [25]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e q u i v}} = \\mathbb {E} _ {n} \\| \\mathcal {T} ^ {- 1} \\left(\\mathbf {M} _ {n} (\\mathbf {e}, \\mathcal {T} (\\mathbf {X})) - \\mathbf {M} _ {n} (\\mathbf {e}, \\mathbf {X}) \\right\\| ^ {2} \\tag {5}", + "image_path": "8a68f6d1aee75ce558c7ed15e86cd887aacee6db749b73694a1ab1ed8245e108.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "content": " we simply utilize minor affine transformations. We use random rotations between " + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\pm 15" + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "content": " degrees, translations between " + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\pm 0.25 \\times W" + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "content": ", and scaling between " + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "100 - 120\\%" + }, + { + "bbox": [ + 305, + 298, + 545, + 346 + ], + "type": "text", + "content": " of the original image size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "content": "Encouraging localization - " + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{localize}}" + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "content": ". We encourage localization by forcing " + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}_n" + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "content": " to be a single-mode Gaussian distribution located at its maximum. In more details, denoting the Gaussian image that shares the same maximum as " + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}_n" + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_n" + }, + { + "bbox": [ + 305, + 351, + 545, + 411 + ], + "type": "text", + "content": ", we write" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 367, + 417, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 417, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 367, + 417, + 545, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o c a l i z e}} = \\mathbb {E} _ {n} \\| \\tilde {\\mathbf {M}} _ {n} - \\mathbf {G} _ {n} \\| ^ {2}. \\tag {6}", + "image_path": "6a8004b80127ef0ab30534105c9917cfbfeb1c6e169c5b8aeadf58b5d3ccfe87.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "content": "To create the Gaussian images " + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_n" + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "content": ", we first identify the spatial location exhibiting the maximal response within the heatmap corresponding to each token " + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "content": " by taking the argmax:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 486, + 545, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 486, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 369, + 486, + 545, + 505 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {n} = \\underset {w, h} {\\arg \\max } \\tilde {\\mathbf {M}} _ {n} [ h, w ]. \\tag {7}", + "image_path": "22372ce7c755cc19ce7d0e0e5d49914d325526ea613e30357f8de881fdd34e90.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 509, + 497, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 497, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 497, + 521 + ], + "type": "text", + "content": "We then generate a Gaussian image; see Fig. 3:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 527, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 527, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 347, + 527, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {n} = \\exp \\left(- \\frac {\\| \\mathbf {X Y} _ {\\mathrm {c o o r d}} - \\boldsymbol {\\mu} _ {n} \\| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {8}", + "image_path": "a37dedf76817e00139ac73eb06565371f02b1eab7f61af9e2bed60f099abe642.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 559, + 506, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 506, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 559, + 506, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{Y}_{\\mathrm{coord}}" + }, + { + "bbox": [ + 306, + 559, + 506, + 572 + ], + "type": "text", + "content": " is a tensor of image coordinates." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "content": "Promoting mutual exclusivity. It is important to note that while " + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{localize}}" + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "content": " in (6) at first glance seem to only encourage localization, it also enforces " + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}_n" + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "content": " to be mutually exclusive for different " + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "content": " because of the softmax operation in (2). Should multiple embeddings become similar, their attention responses in (2) " + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l^c\\cdot \\mathbf{K}_l" + }, + { + "bbox": [ + 305, + 577, + 545, + 685 + ], + "type": "text", + "content": " will become similar, resulting in the softmax of the attention map being a flat response (i.e. deviating from a Gaussian shape). In other words, (6) naturally enforces exclusivity with the help of (2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Stabilizing optimization by working with a subset. We noticed in our experiments that attention maps " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " for some" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22823" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "tokens can be 'spread' for some images, e.g., due to occlusions, destabilizing optimization. We thus opt for a simple solution of looking into the top-K tokens that are local. Specifically, we apply our losses over " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "n \\in \\mathcal{N}(\\kappa)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", which returns the " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\kappa \\in \\mathbb{N}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " entries with the most spatially localized heatmap responses2, as measured by KL divergence:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 152, + 287, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 152, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 80, + 152, + 287, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {N} (\\kappa) = \\operatorname {A r g T o p} _ {\\kappa} \\left\\{- \\mathrm {K L} \\left(\\mathbf {G} _ {n}, \\tilde {\\mathbf {M}} _ {n}\\right) \\right\\} _ {n = 1} ^ {N}. \\tag {9}", + "image_path": "2852f56df6454ae327aaf9666f49d70e88c8d80f4149bcd4a585c2815aa654bc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 240 + ], + "type": "text", + "content": "Final keypoints. While " + }, + { + "bbox": [ + 46, + 181, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{localize}}" + }, + { + "bbox": [ + 46, + 181, + 287, + 240 + ], + "type": "text", + "content": " naturally enforces exclusivity, it does not guarantee a complete coverage of the object. Thus, after we finish optimizing, we refine the set of keypoints through furthest point sampling using the training images. Specifically, for each image we write:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 250, + 287, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 250, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 103, + 250, + 287, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {K} = \\operatorname {F P S} _ {K} (\\{\\boldsymbol {\\mu} _ {i} \\mid i \\in \\mathcal {N} (\\kappa) \\}), \\tag {10}", + "image_path": "2de5b618b747d57468222ab544e9f5f482130945aed7f4315ddffcf0635d1f49.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": " is the desired number of keypoints " + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "inline_equation", + "content": "K < \\kappa" + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": ". Then, as the set " + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": " differs in each image, we simply choose " + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": " to tokens that appeared most frequently in " + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 272, + 287, + 320 + ], + "type": "text", + "content": " within the training image set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 328, + 178, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 178, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 178, + 340 + ], + "type": "text", + "content": "3.3. Implementation details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 346, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 407 + ], + "type": "text", + "content": "Test-time ensembling. At inference time, as in common literature [3, 10, 15], rather than employing the attention map of the original image, we average the attention maps across multiple augmentations (we use the same transformations for test time augmentation as in (5)) of the image:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 415, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 415, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 96, + 415, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {M}} \\leftarrow \\sum_ {i} \\mathcal {T} _ {i} ^ {- 1} \\left(\\mathbf {M} \\left(\\mathbf {e}, \\mathcal {T} _ {i} (\\mathbf {X})\\right)\\right). \\tag {11}", + "image_path": "7ed7c080960ec43136c83c963e28cfb228424b42e7a40146374e601163ea446f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "content": "Upsampling attention maps. The attention maps in (2) are typically of low resolution. Specifically, as we use Stable Diffusion 1.5 [44], depending on the layer we extract the attention maps from, they are either " + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "content": ". We thus opt to upsample the query " + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "content": " via bicubic interpolation to achieve a standard resolution of " + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 450, + 287, + 568 + ], + "type": "text", + "content": ". We have experimented with other upsampling techniques such as the commonly used bilinear sampling or a learned upsampler that is trained alongside, but a simple bicubic upsample was shown to be effective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 580, + 99, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 99, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 99, + 593 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 600, + 161, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 600, + 161, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 161, + 613 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "text", + "content": "We evaluate our method on five standard datasets for unsupervised keypoint evaluation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": "CelebA dataset [28]: A dataset of 202,599 facial images of celebrities. We evaluate both the aligned and nonaligned cases following the standard protocol of omitting images with faces occupying less than " + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": " of the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "text", + "content": "The standard metric for this dataset is to measure the average " + }, + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 72, + 544, + 95 + ], + "type": "text", + "content": " error normalized by the inter-ocular distance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 96, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "text", + "content": "- CUB-200-2011 dataset [58]: This dataset consists of 11,788 bird images. We use both the aligned (CUB-aligned) and non-aligned (CUB-all) variants. For the nonaligned variants, we further look at CUB-001, CUB-002, and CUB-003, which are specific bird subcategories. Notably, these subsets contain only 30 images each—we only use these 30 for training. We follow the standard protocol [6, 30] and normalize the images to be of " + }, + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "text", + "content": ". The standard metric for this dataset is the mean " + }, + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 306, + 96, + 545, + 226 + ], + "type": "text", + "content": " error, normalized by the dimension of the images after normalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "text", + "content": "- Tai-Chi-HD dataset [50]: This dataset contains 3049 training videos and 285 test videos of people performing Tai-Chi, which shows more diverse poses compared to the other datasets, and is the most challenging among the human pose-centric datasets that we use. We follow Siarohin et al. [51] and use 500 images for testing and 300 images for training. The standard metric for this dataset is to measure the accumulated " + }, + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "text", + "content": " error, with the images standardized to " + }, + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 306, + 227, + 545, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 335, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 441 + ], + "type": "text", + "content": "- DeepFashion dataset [29]: This dataset contains 53k images of fashion models, mostly standing with a white background. We follow Lorenz et al. [30] and only keep full body images. This leaves 10,604 images for training and 1,179 images for testing. Also following the baselines, we use keypoints generated by AlphaPose [11] as ground truth. The standard metric for this dataset is the percentage of correct keypoints (PCK) with a 6-pixel threshold." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "text", + "content": "- Human 3.6M dataset [17]: This dataset is of humans performing various actions, comprised of 3.6 million images. We follow the standard protocol [67] and focus on six activities: direction, discussion, posing, waiting, greeting, and walking. We utilize subjects 1, 5, 6, 7, 8, and 9 for training, while subject 11 is reserved for testing. This division yields a training dataset comprising 796,648 images and a testing dataset containing 87,975 images. The background for this dataset is also simple, and often masked out with ground-truth masks for evaluation. This dataset is also typically heavily pre-processed and aligned when used for unsupervised keypoint evaluation. We experiment with the standard pre-processing [30, 67] and also a relaxed version of our own. To relax the alignment, we crop a square bounding box such that the margin from the bounding box to the person is 100 pixels, which on average corresponds to the person's height being " + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "inline_equation", + "content": "2/3" + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "text", + "content": " of the crop. We further add a uniform random translation up to 100 pixels (same as the margin) to remove the central bias. Example crops are visualized in Fig. 4f. The standard metric for this dataset is the " + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "text", + "content": " error after normalizing the image resolution to " + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 306, + 443, + 545, + 705 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 261, + 713 + ], + "type": "text", + "content": "2We empirically found that using " + }, + { + "bbox": [ + 58, + 702, + 261, + 713 + ], + "type": "inline_equation", + "content": "\\kappa = 25" + }, + { + "bbox": [ + 58, + 702, + 261, + 713 + ], + "type": "text", + "content": " works best in general." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22824" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 195 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 195 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 195 + ], + "type": "table", + "html": "
MethodAligned (K=10)↓Wild (K=4) ↓Wild (K=8) ↓
Thewlis et al. [55]7.95-31.30
Zhang et al. [67]3.46-40.82
LatentKeypointGAN [12]5.8525.8121.90
Lorenz et al. [30]3.2415.4911.41
IMM [19]3.1919.428.74
LatentKeypointGAN-tuned [12]3.3112.105.63
Autolink [13]3.927.725.66
Autolink † [13]3.546.115.24
Our method3.605.244.35
", + "image_path": "dab41fa904933c4edf9f54abdc57b85ec32d8a75a353d8d7a3b54d4f0f139ba7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "text", + "content": "Note that each dataset comes with its own metric. To make results more comparable across the human pose datasets, we report both their original metrics as well as the " + }, + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "text", + "content": " error when normalizing the image resolution to " + }, + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 270, + 287, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 382 + ], + "type": "text", + "content": "Regressing human-annotated landmarks. To evaluate the quality of unsupervised keypoints, one must relate them with human-annotated landmarks. As in prior research [55], we use linear regression (without bias) to relate between unsupervised keypoints and human-annotated landmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "content": "Number of keypoints and hyperparameters. For each method, we use the standard number of unsupervised keypoints defined for each evaluation protocol—we denote them in our Tables. We use the same hyperparameter for all our experiments as introduced in Sec. 3.2, except for the number of optimization iterations. We optimize the embeddings for 10k iterations, except for the human pose datasets, for which we optimize 500 iterations. To find the number of optimization rounds we use a " + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "content": " validation subset from the training data. While we observed our results on the validation subset to improve consistently for most datasets, we found 10k to give a reasonable optimization time of two hours on an RTX 3090. For the human pose dataset, we found optimization to have converged already at 500 iterations on our validation split." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 167, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 167, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 167, + 588 + ], + "type": "text", + "content": "4.2. Experimental results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Quantitative results – Tabs. 1 to 3. We report our results for each dataset in Tabs. 1 to 3. As shown, except for the case when data is heavily processed and aligned (CelebA aligned in Tab. 1, CUB-aligned in Tab. 2, and Human 3.6M in Tab. 3), our method significantly outperforms the state of the art. The most visible gains are for the Tai-Chi-HD dataset, the most challenging among human pose datasets, and on CUB unaligned datasets. For the CUB dataset and the Tai-Chi-HD datasets, we outperform even those that have been supervised with silhouettes or saliency maps." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 46, + 204, + 287, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 248 + ], + "type": "text", + "content": "Table 1. Quantitave results for the CelebA dataset - we report results with the standard metrics. Our method performs best for non-aligned cases and is comparable to the state of the art for the aligned case. " + }, + { + "bbox": [ + 46, + 204, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 204, + 287, + 248 + ], + "type": "text", + "content": " symbol represents the thickness-tuned variant." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 182 + ], + "type": "table", + "html": "
MethodSupervisionCUB-aligned (K=10) ↓CUB-001 (K=4) ↓CUB-002 (K=4) ↓CUB-003 (K=4) ↓CUB-all (K=4) ↓
SCOPS [16]GT silhouette-18.317.717.012.6
Choudhury et al. [6]GT silhouette-11.315.010.69.2
DFF [9]testing dataset-22.421.622.0-
SCOPS [16]saliency maps-18.518.821.1-
Lorenz et al. [30]unsupervised3.91----
ULD [55, 67]unsupervised-30.129.428.2-
Zhang et al. [67]unsupervised5.3626.927.627.122.4
LatentKeypointGAN [12]unsupervised5.2122.629.121.214.7
GANSeg [14]unsupervised3.2322.122.321.512.1
Autolink [13]unsupervised4.1520.620.319.711.6
Autolink † [13]unsupervised3.5120.219.218.511.3
Our methodunsupervised5.0610.511.110.35.4
", + "image_path": "45800b2fb463fa78a80eae26c3583145553f34efd240af4e2a47f6959e4189a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 273, + 545, + 414 + ], + "blocks": [ + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "type": "text", + "content": "Table 2. Quantitave results for the CUB-200-2011 dataset - we report results with the standard metrics. Except for the CUB-aligned case, our method performs nearly twice better than the compared methods, even outperforming Choudhury et al. [6], which is supervised with ground-truth silhouettes. " + }, + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 191, + 545, + 257 + ], + "type": "text", + "content": " represents the thickness-tuned variant." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 273, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 414 + ], + "type": "table", + "html": "
MethodSupervisionHuman 3.6M (K=16) \\( {\\ell }_{2} \\) standard / unaligned \\( \\downarrow \\)DeepFashion (K=16) PCK† / Rel. \\( {\\ell }_{2} \\downarrow \\)Tai-Chi-HD (K=10) Cum \\( {\\ell }_{2} \\downarrow \\) / Rel. \\( {\\ell }_{2} \\downarrow \\)
Newell et al. [20]paired gt2.16 / ---
DFF [9]testing dataset--494.48 / 14.78
SCOPS [16]saliency maps--411.38 / 12.29
Jakab et al. [20]video*2.73 / ---
Siarohin et al. [51]videos--389.78 / 11.65
Zhang et al. [68]videos--343.67 / 10.27
Zhang et al. [67]videos4.14 / ---
Schmidtke et al. [47]video*3.31 / ---
Sun et al. [53]videos2.53 / ---
Thewis et al. [55]unsupervised7.51 / ---
Zhang et al. [67]unsupervised4.91 / ---
LatentKeypointGAN [12]unsupervised-49%437.69 / 13.08
Lorenz et al. [30]unsupervised2.79 / -57%-
GANSeg [14]unsupervised-59%417.17 / 12.47
autolink [13]unsupervised2.81 / 7.5965%337.50 / 10.08
autolink † [13]unsupervised2.76 / -66%316.10 / 9.45
Our methodunsupervised4.45 / 5.7770%/6.46234.89 / 7.02
", + "image_path": "2729725696cca38ae30dc9049c01befc3f9987e9c35a38649aadc935a0220d34.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "type": "text", + "content": "Table 3. Quantitative results for human pose datasets - We report results for the Human 3.6M dataset, Deep Fashion dataset, and the Challenging Tai-Chi-HD datasets. We report both standard metrics for each dataset and the relative " + }, + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "type": "text", + "content": " error after normalizing images to " + }, + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 304, + 422, + 545, + 533 + ], + "type": "text", + "content": ". Our method, except for the Human 3.6M dataset that is heavily pre-processed, outperforms all baselines. This includes, for the challenging Tai-Chi-HD dataset, supervised ones. * denotes additional supervision (Jakab et al. [20] uses unpaired ground truth, and Schmidtke et al. [47] use the T-pose). The † symbol represents the thickness-tuned variant for Autolink." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "We note that our primary focus is on unaligned cases, as we argue that they represent more how keypoints would be used in real-world applications—most real-world datasets are unaligned except for specific classes of objects. Moreover, methods focusing on aligned settings use strong locational priors, and as shown by their results in the unaligned setup—CelebA in the wild, non-aligned cases of Human 3.6M and CUB-200-2011, and Tai-Chi-HD—may perform significantly worse once this alignment prior is broken. Given that the performance of our method, even in the aligned case, is not too far off from methods that utilize alignment, we suspect a more in-depth tuning of our method may make our method outperform these methods," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22825" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 293, + 121 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 293, + 121 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 293, + 121 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 293, + 121 + ], + "type": "image", + "image_path": "cfac4e55d86b17a0d932f67f6c95dd867bc2368608c3104b10fde8e9a2774792.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 122, + 218, + 132 + ], + "lines": [ + { + "bbox": [ + 123, + 122, + 218, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 122, + 218, + 132 + ], + "type": "text", + "content": "(a) CelebA dataset keypoints" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 133, + 293, + 183 + ], + "blocks": [ + { + "bbox": [ + 49, + 133, + 293, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 133, + 293, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 133, + 293, + 183 + ], + "type": "image", + "image_path": "13ad7f932b8010928b28ab4573b299cd0af97687060db3d59eb7e93e85f34264.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 185, + 225, + 194 + ], + "lines": [ + { + "bbox": [ + 116, + 185, + 225, + 194 + ], + "spans": [ + { + "bbox": [ + 116, + 185, + 225, + 194 + ], + "type": "text", + "content": "(c) Tai-Chi-HD dataset keypoints" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 195, + 293, + 244 + ], + "blocks": [ + { + "bbox": [ + 49, + 195, + 293, + 244 + ], + "lines": [ + { + "bbox": [ + 49, + 195, + 293, + 244 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 293, + 244 + ], + "type": "image", + "image_path": "a19a4af91ac82dd089039f3f1165af4a1ef579dec65c28fd286b0beea47f2954.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 247, + 227, + 257 + ], + "lines": [ + { + "bbox": [ + 114, + 247, + 227, + 257 + ], + "spans": [ + { + "bbox": [ + 114, + 247, + 227, + 257 + ], + "type": "text", + "content": "(e) Human 3.6M dataset keypoints" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 71, + 545, + 121 + ], + "blocks": [ + { + "bbox": [ + 302, + 71, + 545, + 121 + ], + "lines": [ + { + "bbox": [ + 302, + 71, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 545, + 121 + ], + "type": "image", + "image_path": "775c9a914965358e6990af01426d09bacf9e0b22448aee63c631276642544812.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 122, + 483, + 132 + ], + "lines": [ + { + "bbox": [ + 362, + 122, + 483, + 132 + ], + "spans": [ + { + "bbox": [ + 362, + 122, + 483, + 132 + ], + "type": "text", + "content": "(b) CUB-200-2011 dataset keypoints" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 302, + 133, + 545, + 183 + ], + "blocks": [ + { + "bbox": [ + 302, + 133, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 302, + 133, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 302, + 133, + 545, + 183 + ], + "type": "image", + "image_path": "d47aa7dd9374c394e3881732a00422cf7b4acb23986668ceeb0f575f8b3e7c67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 302, + 195, + 545, + 244 + ], + "blocks": [ + { + "bbox": [ + 367, + 185, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 367, + 185, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 367, + 185, + 479, + 195 + ], + "type": "text", + "content": "(d) DeepFashion dataset keypoints" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 195, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 302, + 195, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 302, + 195, + 545, + 244 + ], + "type": "image", + "image_path": "b773bfeef0bda086fd21d31a6a9bb9250a2fbde1f585d9927cb0e65b64bc49fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 247, + 496, + 257 + ], + "lines": [ + { + "bbox": [ + 349, + 247, + 496, + 257 + ], + "spans": [ + { + "bbox": [ + 349, + 247, + 496, + 257 + ], + "type": "text", + "content": "(f) Unaligned Human 3.6M dataset keypoints" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 266, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 310 + ], + "type": "text", + "content": "Figure 4. Qualitative examples of unsupervised keypoints – we show our learned keypoints for the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human 3.6M datasets (both for cropped and masked as well as our relaxed version). Note how our keypoints are consistent despite the variability. Our method significantly outperforms other baselines, especially for the challenging Tai-Chi-HD dataset and the CUB subsets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 330, + 285, + 416 + ], + "blocks": [ + { + "bbox": [ + 50, + 330, + 285, + 416 + ], + "lines": [ + { + "bbox": [ + 50, + 330, + 285, + 416 + ], + "spans": [ + { + "bbox": [ + 50, + 330, + 285, + 416 + ], + "type": "table", + "html": "
VariantNormalized ℓ2
Full (Our method)5.4
Without test time ensembling5.6
Without furthest point sampling6.4
Without upsampling the query Q8.0
Without equivariance22.2
", + "image_path": "e37beba569ae00c50929123cf446078c67414af5850fca7e6172e920e5c0e209.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 425, + 287, + 458 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 458 + ], + "type": "text", + "content": "Table 4. Ablation results - we report the effect of each of our design choices can be seen on the CUB-all dataset. All components contribute to the final performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 479, + 179, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 179, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 179, + 490 + ], + "type": "text", + "content": "but we leave this as future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 491, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 575 + ], + "type": "text", + "content": "Finally, also note that for CUB-001, CUB-002, and CUB-003, these datasets are small. These datasets are nonaligned, have a large variability between individual images, and only contain 30 images each in the training set. Our method, just from 30 images, successfully identifies keypoints. These results highlight the potential of leveraging emergent (prior) knowledge within Stable Diffusion [44]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": "Qualitative results - Fig. 4. We provide example visualizations of our unsupervised keypoints in Fig. 4. As shown, our method discovers keypoints that are consistently localized across the dataset, despite the wide appearance variety." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 635, + 139, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 139, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 139, + 647 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "We perform an ablation study for various design choices of our method on the CUB-all dataset. We report the performance of our method with different components disabled in Tab. 4. As shown, all components contribute to the final performance. Test-time ensembling enhances performance," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "content": "but the computation cost linearly scales. We choose, ten augmentations, which provide a good compromise between computation time and accuracy. To remove furthest point sampling we set " + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\kappa = K" + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "content": ", which then makes furthest point sampling select all samples. While this causes points to be more grouped, it still provides reasonable performance. To remove upsampling we instead upsample " + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "content": " to the size of the target image, effectively having the attention map build at lower resolutions, sometimes as low as " + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "content": ". This results in significant degradation in performance. " + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{equiv}}" + }, + { + "bbox": [ + 304, + 332, + 546, + 475 + ], + "type": "text", + "content": " is essential, as without it, the tokens can 'cheat' and simply opt to learn fixed positions on the image." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 481, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 589 + ], + "type": "text", + "content": "Number of training images.. Inspired by our results for the small subsets of CUB-200-2011 dataset, we investigate the impact that the number of images that we use to find keypoints has on our results. We thus optimized our keypoints only with 100 images for CelebA non-aligned setup. Surprisingly, we achieve " + }, + { + "bbox": [ + 304, + 481, + 545, + 589 + ], + "type": "inline_equation", + "content": "5.33K = 8" + }, + { + "bbox": [ + 304, + 481, + 545, + 589 + ], + "type": "text", + "content": ", which is comparable to the state of the art. This demonstrates once more how our method is able to leverage information that is already learned in Stable Diffusion [44] to find keypoints." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 598, + 398, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 398, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 398, + 609 + ], + "type": "text", + "content": "4.4. Generalization" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "We further test the generalization capacity of our learned keypoints. As they are effectively text embeddings, we can simply apply them to any image, including those completely outside of the training domain. We quantitatively evaluate our method and the previous best-performing method Autolink [13]. We find that even in these generalization experiments, our keypoints reach performance comparable to data-specific keypoints. Applying Tai-Chi-HD tokens to un" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22826" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 112, + 132 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 112, + 132 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 112, + 132 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 112, + 132 + ], + "type": "image", + "image_path": "0ba9fff91d38fc99eb008a3134a950e3f5e39abd88f3b1e5f4c6c8ab923e95a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 134, + 252, + 144 + ], + "lines": [ + { + "bbox": [ + 94, + 134, + 252, + 144 + ], + "spans": [ + { + "bbox": [ + 94, + 134, + 252, + 144 + ], + "type": "text", + "content": "(a) Applying Tai-Chi-HD tokens to Human 3.6M" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 70, + 174, + 132 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 174, + 132 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 174, + 132 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 174, + 132 + ], + "type": "image", + "image_path": "ecf61c3715f5b6152c3764e1de86434edbb8072e2b8e28d4871e009674dde050.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 174, + 70, + 235, + 132 + ], + "blocks": [ + { + "bbox": [ + 174, + 70, + 235, + 132 + ], + "lines": [ + { + "bbox": [ + 174, + 70, + 235, + 132 + ], + "spans": [ + { + "bbox": [ + 174, + 70, + 235, + 132 + ], + "type": "image", + "image_path": "de0dd9cadffe29e8baa849b8e81201bd8fb53b92e47e97e084378bd8d4750473.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 71, + 295, + 132 + ], + "blocks": [ + { + "bbox": [ + 235, + 71, + 295, + 132 + ], + "lines": [ + { + "bbox": [ + 235, + 71, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 235, + 71, + 295, + 132 + ], + "type": "image", + "image_path": "1305e77d530ed04f30a880baa9c103b549b23b3f961d898a7a697c721edc09d9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 146, + 111, + 206 + ], + "blocks": [ + { + "bbox": [ + 51, + 146, + 111, + 206 + ], + "lines": [ + { + "bbox": [ + 51, + 146, + 111, + 206 + ], + "spans": [ + { + "bbox": [ + 51, + 146, + 111, + 206 + ], + "type": "image", + "image_path": "af0296a3506232df22930460365fc59f08435495653c10a7149e6e62f1a98538.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 209, + 243, + 220 + ], + "lines": [ + { + "bbox": [ + 102, + 209, + 243, + 220 + ], + "spans": [ + { + "bbox": [ + 102, + 209, + 243, + 220 + ], + "type": "text", + "content": "(c) Applying CelebA tokens to Tai-Chi-HD" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 111, + 146, + 173, + 206 + ], + "blocks": [ + { + "bbox": [ + 111, + 146, + 173, + 206 + ], + "lines": [ + { + "bbox": [ + 111, + 146, + 173, + 206 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 173, + 206 + ], + "type": "image", + "image_path": "21936bdc746586d3a09f0775637bdbe1a4b2bab990c1f073f2ed566a9d722cdd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 547, + 252 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 547, + 252 + ], + "type": "text", + "content": "Figure 5. Generalization - we apply our learned text tokens (keypoints) to images from other datasets, including those that are of completely different domains. Our tokens generalize well for data of similar type, and surprisingly well even for some extreme cases." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 173, + 146, + 233, + 206 + ], + "blocks": [ + { + "bbox": [ + 173, + 146, + 233, + 206 + ], + "lines": [ + { + "bbox": [ + 173, + 146, + 233, + 206 + ], + "spans": [ + { + "bbox": [ + 173, + 146, + 233, + 206 + ], + "type": "image", + "image_path": "5f0808c04c992684bb235584fdecf88ab03992734f44284785c5a5f8a7a47242.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 233, + 146, + 294, + 206 + ], + "blocks": [ + { + "bbox": [ + 233, + 146, + 294, + 206 + ], + "lines": [ + { + "bbox": [ + 233, + 146, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 233, + 146, + 294, + 206 + ], + "type": "image", + "image_path": "f23345f7add49711f23a50d802a8750218690a6c68b7dfdd2309ebd6493810f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 359, + 132 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 359, + 132 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 359, + 132 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 359, + 132 + ], + "type": "image", + "image_path": "27b8503d3f55eeb79d519fb91bda66a64a3469ce48f1b85cc6de5982d32225ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 134, + 503, + 144 + ], + "lines": [ + { + "bbox": [ + 337, + 134, + 503, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 134, + 503, + 144 + ], + "type": "text", + "content": "(b) Applying CUB-200-2011 tokens to Tai-Chi-HD" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 419, + 132 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "type": "image", + "image_path": "e01427bd5156d67fab12cc8e2809a1854f6ce5245d01f91d0623d25e6f2ef48e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 481, + 132 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 481, + 132 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 481, + 132 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 481, + 132 + ], + "type": "image", + "image_path": "1a3648aa12019cd974626f865eed21a9314692809bff295b58154556010def7a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 481, + 72, + 541, + 132 + ], + "blocks": [ + { + "bbox": [ + 481, + 72, + 541, + 132 + ], + "lines": [ + { + "bbox": [ + 481, + 72, + 541, + 132 + ], + "spans": [ + { + "bbox": [ + 481, + 72, + 541, + 132 + ], + "type": "image", + "image_path": "096ccc605c6dbec7da30f24be5f2d91c54dc932c82c8119ad9e93f3f057c3f88.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 299, + 146, + 359, + 206 + ], + "blocks": [ + { + "bbox": [ + 299, + 146, + 359, + 206 + ], + "lines": [ + { + "bbox": [ + 299, + 146, + 359, + 206 + ], + "spans": [ + { + "bbox": [ + 299, + 146, + 359, + 206 + ], + "type": "image", + "image_path": "27892a36cc3be66b89859e44332cc3dacc8e0a798c08666ccb1f965ae7861afc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 209, + 495, + 220 + ], + "lines": [ + { + "bbox": [ + 344, + 209, + 495, + 220 + ], + "spans": [ + { + "bbox": [ + 344, + 209, + 495, + 220 + ], + "type": "text", + "content": "(d) Applying CelebA tokens to CUB-200-2011" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 360, + 146, + 420, + 206 + ], + "blocks": [ + { + "bbox": [ + 360, + 146, + 420, + 206 + ], + "lines": [ + { + "bbox": [ + 360, + 146, + 420, + 206 + ], + "spans": [ + { + "bbox": [ + 360, + 146, + 420, + 206 + ], + "type": "image", + "image_path": "a2dde74df91c1a29216fe061ba4e03245b85aa5268c6f25fbe2c67da2750d6b2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 421, + 146, + 481, + 206 + ], + "blocks": [ + { + "bbox": [ + 421, + 146, + 481, + 206 + ], + "lines": [ + { + "bbox": [ + 421, + 146, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 421, + 146, + 481, + 206 + ], + "type": "image", + "image_path": "ffae4282d36896cc5d7ea4a1a0e22db8e4c7a9efa558d43e45707a4d9c4abe5b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 482, + 146, + 542, + 206 + ], + "blocks": [ + { + "bbox": [ + 482, + 146, + 542, + 206 + ], + "lines": [ + { + "bbox": [ + 482, + 146, + 542, + 206 + ], + "spans": [ + { + "bbox": [ + 482, + 146, + 542, + 206 + ], + "type": "image", + "image_path": "48d59cb3d0a9089accf108ccd388c5adaa4c2178b91be90031aae28cbcb2df17.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "text", + "content": "aligned Human 3.6M achieves state-of-the-art performance despite using fewer keypoints " + }, + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 10" + }, + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 16)" + }, + { + "bbox": [ + 46, + 273, + 286, + 367 + ], + "type": "text", + "content": ". We also perform on par with the previous state of the art when we apply CUB-200-2011 tokens to Tai-Chi-HD—a case where the dataset gap is not only about the appearance but also beyond classes. Despite the drastic gap, our keypoints perform extremely well, leveraging the generalization power of large pre-trained diffusion models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 369, + 286, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 286, + 632 + ], + "type": "text", + "content": "We show qualitative examples in Fig. 5. As shown, even when applied to different datasets, they look reasonable. For example, in Fig. 5a, when applying Tai-Chi-HD tokens to Human 3.6M, the tokens respond to the same locations on the human body as in Tai-Chi-HD. Surprising was when we applied CUB-200-2011 tokens to Tai-Chi-HD in Fig. 5b—they still responded to the body of the human being, reasonably consistently, although these tokens were trained to respond to birds. Of note are tokens two and six, which correspond to the front and back of the bird heads in Fig. 4—they also reply to the front and back of human heads. Applying CelebA tokens to Tai-Chi-HD in Fig. 5c also shows interesting outcomes, as tokens generally respond to human faces, despite the scale being drastically different between the two datasets. Finally, applying CelebA tokens to the CUB-200-2011 dataset in Fig. 5d shows mixed results—when it is 'successful' it focuses also on the faces of the bird, when it fails, it fails completely. These results hint that the keypoints (tokens) we have learned carry semantic meanings, as expected. We note that none of the baselines that we compare against are able to generalize beyond the dataset they were trained for." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 645, + 124, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 124, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 124, + 657 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We have proposed a novel method to find unsupervised keypoints using pre-trained text-to-image diffusion models. Given a set of images of a certain object, we propose to optimize the text embeddings (tokens) such that the cross" + } + ] + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 308, + 270, + 544, + 316 + ], + "blocks": [ + { + "bbox": [ + 308, + 270, + 544, + 316 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 544, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 544, + 316 + ], + "type": "table", + "html": "
Tai-Chi-HDCUB-200-2011CelebACelebA
→unaligned→Tai-Chi-HD→Tai-Chi-HD→CUB-200-2011
Human3.6m(K=10)↓(K=8)↓(K=8)↓
(K=10)↓Cum ℓ2 ↓ / Rel. ℓ2 ↓Cum ℓ2 ↓ / Rel. ℓ2 ↓
Ours4.88317.94 / 9.50- / 8.618.60
Autolink [13]16.92535.61 / 16.00- / 28.222.56
", + "image_path": "eb6bcbe188c6104540a9d1f10ae3fb1778af1f8ef0a56716270f71237850dc51.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 323, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 389 + ], + "type": "text", + "content": "Table 5. Generalization - we quantitatively evaluate the performance of our keypoints on other datasets. Our Tai-Chi-HD keypoints applied to the unaligned Human3.6m setting reach state-of-the-art performance. Interestingly, our CUB keypoints applied to Tai-Chi-HD are on par with the previous state of the art, despite the differences between these datasets." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 410, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 506 + ], + "type": "text", + "content": "attention maps within diffusion models become localized as Gaussians with a small standard deviation. By doing so, we find text tokens that can be used to extract keypoints by extracting the maxima of the attention maps. We have shown that our method, on multiple datasets, under the challenging un-aligned setup, significantly outperforms the state of the art. We have further demonstrated that these tokens are also generalizable." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 517, + 415, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 415, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 415, + 530 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 536, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 536, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 545, + 583 + ], + "type": "text", + "content": "The authors would like to thank Cristina Vasconcelos for her constructive feedback during the preparation of this manuscript. Additionally, we extend our gratitude to David Fleet for his approval and support of this work." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 583, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 583, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 583, + 545, + 650 + ], + "type": "text", + "content": "This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant, NSERC Collaborative Research and Development Grant, Google, Digital Research Alliance of Canada, and Advanced Research Computing at the University of British Columbia." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "text", + "content": "[1] Shekoofeh Azizi, Simon Kornblith, Chitwan Sahara, Mohammad Norouzi, and David J Fleet. Synthetic data from" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22827" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "diffusion models improves imagenet classification. Transactions on Machine Learning Research, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "text", + "content": "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 287, + 185 + ], + "type": "text", + "content": "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "type": "text", + "content": "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection. Proceedings of the IEEE International Conference on Computer Vision, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 287, + 274 + ], + "type": "text", + "content": "[5] Weiya Chen, Chenchen Yu, Chenyu Tu, Zehua Lyu, Jing Tang, Shiqi Ou, Yan Fu, and Zhidong Xue. A survey on hand pose estimation with wearable sensors and computer- vision-based methods. Sensors, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 319 + ], + "type": "text", + "content": "[6] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 2021. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 321, + 287, + 376 + ], + "type": "text", + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv Preprint, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 377, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 377, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 377, + 287, + 409 + ], + "type": "text", + "content": "[8] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero-shot classifiers. International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 453 + ], + "type": "text", + "content": "[9] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[11] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "text", + "content": "[12] Xingzhe He, Bastian Wandt, and Helge Rhodin. Latentkeypointgan: Controlling gans via latent keypoints. International Conference on Learning Representations, 2021. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[13] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Xingzhe He, Bastian Wandt, and Helge Rhodin. Ganseg: Learning to segment by unsupervised hierarchical image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[16] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Conference on Computer Vision and Pattern Recognition, 2019. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "text", + "content": "[17] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[18] Marwa Jabberi, Ali Wali, Bidyut Baran Chaudhuri, and Adel M Alimi. 68 landmarks are efficient for 3d face alignment: what about more? 3d face alignment method applied to face recognition. Multimedia Tools and Applications, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in Neural Information Processing Systems, 2018. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "text", + "content": "[20] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[21] You-Yi Jau, Rui Zhu, Hao Su, and Manmohan Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In International Conference on Intelligent Robots and Systems, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[22] Le Jiang, Caleb Lee, Divyang Teotia, and Sarah Ostadabbas. Animal pose estimation: A closer look at the state-of-the-art, existing gaps and opportunities. Computer Vision and Image Understanding, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[23] Yuhe Jin, Weiwei Sun, Jan Hosang, Eduard Trulls, and Kwang Moo Yi. Tusk: Task-agnostic unsupervised keypoints. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "text", + "content": "[24] Aliasghar Khani, Saeid Asgari Taghanaki, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. Slime: Segment like me. arXiv Preprint, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "text", + "content": "[25] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 571, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 634 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "text", + "content": "[27] Wu Liu, Qian Bao, Yu Sun, and Tao Mei. Recent advances of monocular 2d and 3d human pose estimation: A deep learning perspective. ACM Computing Surveys, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 2, 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22828" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[30] Dominik Lorenz, Leonard Bereska, Timo Milbich, and Bjorn Ommer. Unsupervised part-based disentangling of object shape and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "text", + "content": "[31] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 2004. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "type": "text", + "content": "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv Preprint, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "text", + "content": "[33] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 346 + ], + "type": "text", + "content": "[34] Giorgia Marullo, Leonardo Tanzi, Pietro Piazzolla, and Enrico Vezzetti. 6d object position estimation from 2d images: a literature review. Multimedia Tools and Applications, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "text", + "content": "[35] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[37] Aditya Nandy, Chenru Duan, and Heather J Kulik. Audacity of huge: overcoming challenges of data scarcity and data quality for machine learning in computational materials discovery. Current Opinion in Chemical Engineering, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[38] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 216, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 216, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 216, + 548 + ], + "type": "text", + "content": "[39] OpenAI. Gpt-4 technical report, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 613 + ], + "type": "text", + "content": "[40] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 647 + ], + "type": "text", + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[42] Linhao Qu, Siyu Liu, Xiaoyu Liu, Manning Wang, and Zhijian Song. Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv Preprint, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[45] Helena Russello, Rik van der Tol, and Gert Kootstra. T-leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and Electronics in Agriculture, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 208, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 274 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[47] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 397 + ], + "type": "text", + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[49] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "type": "text", + "content": "[50] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "text", + "content": "[51] Aliaksandr Siarohin, Subhankar Roy, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Motion-supervised co-part segmentation. In International Conference on Pattern Recognition, 2021. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[52] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 547, + 644 + ], + "type": "text", + "content": "[53] Jennifer J Sun, Serim Ryou, Roni H Goldshmid, Brandon Weissboud, John O Dabiri, David J Anderson, Ann Kennedy, Yisong Yue, and Pietro Perona. Self-supervised keypoint discovery in behavioral videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 690 + ], + "type": "text", + "content": "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[55] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22829" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "embeddings. In Proceedings of the IEEE International Conference on Computer Vision, 2017. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "text", + "content": "[56] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv Preprint, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 195 + ], + "type": "text", + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv Preprint, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 286, + 241 + ], + "type": "text", + "content": "[58] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 285 + ], + "type": "text", + "content": "[59] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv Preprint, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 286, + 331 + ], + "type": "text", + "content": "[60] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 286, + 376 + ], + "type": "text", + "content": "[61] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. arXiv Preprint, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 286, + 399 + ], + "type": "text", + "content": "[62] Yue Wu and Qiang Ji. Facial landmark detection: A literature survey. International Journal of Computer Vision, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 286, + 443 + ], + "type": "text", + "content": "[63] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. arXiv Preprint, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 500 + ], + "type": "text", + "content": "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 555 + ], + "type": "text", + "content": "[65] Lumin Xu, Sheng Jin, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Zoomnas: searching for whole-body human pose estimation in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 286, + 612 + ], + "type": "text", + "content": "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 668 + ], + "type": "text", + "content": "[67] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[68] Yanping Zhang, Qiaokang Liang, Kunlin Zou, Zhengwei Li, Wei Sun, and Yaonan Wang. Self-supervised part segmentation via motion imitation. Image and Vision Computing, 2022. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 118 + ], + "type": "text", + "content": "[69] Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22830" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_content_list.json b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f982363499543d4b1ad51777bcbc3dfd14b99621 --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_content_list.json @@ -0,0 +1,1791 @@ +[ + { + "type": "text", + "text": "Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos", + "text_level": 1, + "bbox": [ + 86, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonhard Sommer", + "bbox": [ + 148, + 179, + 313, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artur Jesslen", + "bbox": [ + 359, + 180, + 477, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eddy Ilg", + "bbox": [ + 524, + 180, + 607, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adam Kortylewski", + "bbox": [ + 653, + 180, + 815, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Freiburg", + "bbox": [ + 295, + 215, + 480, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Saarland University", + "bbox": [ + 501, + 215, + 666, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 310, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild on the Pascal3D+ and ObjectNet3D datasets.", + "bbox": [ + 75, + 300, + 472, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 207, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Category-level object pose estimation is a fundamentally important task in computer vision and robotics with a multitude of real-world applications, e.g. for training 3D generative models on real data and for robots that need to grasp and manipulate objects. However, defining and determining the pose of an object is a task that is far from easy. Current approaches achieve high performance, but they require large amounts of annotated training data to generalize successfully [1, 26, 28, 33, 38], or additional inputs during inference, such as CAD Models [7, 14], 3D Shapes [3, 32, 37] or RGB-D [8]. However, all of these are either time-consuming to obtain or not available in practice at all.", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This motivates the development of methods for learning category-level 3D pose estimators in a fully unsupervised fashion. While doing so from images in the wild seems infeasible, object-centric video data [19] offers a more accessible alternative. Such videos can be easily captured using consumer-grade cameras and makes it possible to leverage coarse 3D reconstructions during training, providing a practical and cost-effective method for collecting data. Therefore, we propose the new task of learning a single-image category-level 3D pose estimator from casually captured object-centric videos without any human labels or other supervision. In practice, we leverage CO3D [19] as training data and show that our proposed model is able to generalize and predict accurate poses in the wild for Pascal3D+ [29] and ObjectNet3D [31].", + "bbox": [ + 496, + 268, + 893, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the challenging task of learning category-level 3D pose in an unsupervised fashion from casually captured object-centric videos. In particular, we propose a two-step pipeline (Figure 1). The first step extracts DINOv2 [18] features from the images and reconstructs a coarse 3D mesh from the video with off-the-shelf methods [4, 11]. Building on this input, we introduce a novel 3D alignment procedure, where a key contribution is a novel 3D cyclical distance in terms of geometry and appearance that enables the robust alignment of shape reconstructions even under severe noise and variations in the object topology. As a result, we can align all objects from the object-centric training videos into a canonical coordinate frame without supervision.", + "bbox": [ + 496, + 500, + 893, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a second step, we leverage the canonical poses and 3D meshes obtained from the first step to train a category-level neural mesh [9, 16, 26, 36] in an unsupervised manner. In particular, we represent objects using a prototypical 3D mesh with surface features to capture the geometry and neural appearance of an object category and train a neural network backbone to predict, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. Finally, the object 3D pose is solved using a pose fitting algorithm based on the estimated correspondence pairs.", + "bbox": [ + 496, + 700, + 893, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We demonstrate that our method outperforms all baselines by a large margin at the unsupervised alignment of object-centric videos on the CO3D [19] dataset. Moreover,", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3cc4c7f81ffbb91fa36c268284df84f9a7641346611f9de0b7448095e243257.jpg", + "image_caption": [ + "Figure 1. Illustration of our approach for the unsupervised learning of category-level 3D pose. Our method starts from unaligned object-centric videos of an object category (left) and aligns these into a canonical coordinate frame in a self-supervised manner using a prototypical 3D mesh and self-supervised transformer features (center). Using the aligned videos, we train a neural network backbone to predict 2D-3D correspondences from a single image to enable 3D object pose estimation in the wild (right)." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 297, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b26b63ffc31e99dfec2a9179cb1ba8a3ad41c2546e2004f987ad9f6bdb4151c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 89, + 612, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76a5e14bd63acda2f6f88be6ea9bf423816a19b80a84723817b31c86a58de509.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 90, + 887, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our model provides faithful and robust predictions in-the-wild on Pascal3D+ [29] and ObjectNet3D [31] despite being trained from raw object-centric videos only.", + "bbox": [ + 75, + 367, + 468, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 431, + 218, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Category-Level 3D Pose Estimation. Traditional methods to determine object poses were to label keypoints in images and train supervised methods to predict them [23, 38], or to use only pose labels and directly predict them by casting the pose estimation problem as bin classification by discretizing the pose space [21]. More recent methods utilize 3D meshes of objects or object categories. NeMo [26] uses 3D meshes with neural features that are rendered-and-compared to feature maps from a CNN to obtain pose estimates. [13] predicts a single embedding vector per image and shows that superior performance can be obtained by simply retrieving the closest training sample. In contrast to the above methods, our approach is fully unsupervised.", + "bbox": [ + 75, + 458, + 468, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-Shot and Zero-Shot Pose Prediction. [33] proposes to train a supervised pose estimator across many categories and shows that their approach can generalize well to similar but unseen objects. Many works implement zero-shot pose estimation by conditioning the model on the 3D shape of the unseen object [3, 32, 34, 37] or by leveraging renderings of CAD models [7, 14]. Other methods use few-shot learning [22] or zero-shot learning with DINOv2 [18]. In contrast to the above, our method does not require any annotated dataset, 3D shapes or CAD models.", + "bbox": [ + 75, + 671, + 467, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose Alignment. The work from Goodwin et al. [5] (ZSP) is most close to the first step of our method, as it aligns the poses of two object-centric videos in a fully unsupervised fashion. Similar to our work, they use DINO [2] to obtain semantic correspondences. They perform first a", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coarse alignment by matching one image from the source video to one of many images from the reference video. Then they leverage cyclical distances to select few promising correspondences in the two images, and finally leverage respective depth maps to align both images using least squares. Goodwin et al. extend their work in [6] (UCD+) to match many images from the source video to many of the reference video. By finding a consensus over these many to many alignments with a single transformation they demonstrate improved performance. The first step of our work is similar to these works by that it also leverages DINO features and cyclical distances. However, our work adds a geometric distance to perform the alignment directly in 3D and introduces weighted correspondences to enable the necessary robust regression of the SE3 transformations from noisy and inaccurate geometries, which leads to significant improvements in the alignment accuracy. Note also that these previous approaches used RGB-D inputs, while our method works on images directly.", + "bbox": [ + 496, + 367, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surface Embeddings and Neural Mesh Models. Recent work uses known poses and approximate object geometries to learn features in 3D space to uniquely identify parts of objects. [17] first used known mesh templates of deformable objects to train a network that predicts surface embeddings from the images. NeMo [26] presents a generative model trained with contrastive learning. [27] presents an extension through replacing vertices by Gaussian ellipsoids and using volume rendering. Similar to our work, many recent works leverage pre-trained vision transformers DINO [2] and DINOv2 [18] to unproject image features onto depth maps [5, 6, 37]. In contrast to [17, 26, 27], our approach does not require any pose annotation, while also going beyond [5, 6] by enabling 3D pose estimation in the wild from a single image, and not requiring RGB-D images and CAD models as input [37].", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a078b93488844856d028f8741441aa9417a231f14280e0398d570866f231b33.jpg", + "image_caption": [ + "Figure 2. Illustration of a neural mesh. It consists of a 3D mesh with one or several neural features per vertex. These features encode patch-level features from a feature extractor. In our unsupervised alignment method, we capture the viewpoint-dependent features from DINOv2. For pose estimation, we learn a category-level neural mesh with viewpoint-invariant features." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 472, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 368, + 169, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe our approach for learning category-level 3D pose estimation without supervision from object-centric videos. Our method proceeds in a two-step approach. First, we align object instances across videos in an unsupervised manner to bring them into a canonical reference frame (Section 3.2). Given the aligned videos, our model learns to establish dense correspondences between images and a reconstructed 3D template mesh by predicting a feature vector for each pixel in a 2D image that corresponds to a visible vertex in the template mesh (Section 3.3). Finally, we describe how our model can efficiently estimate object poses from in-the-wild data using the predicted correspondences via render-and-compare.", + "bbox": [ + 75, + 393, + 472, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Meshes with Surface Features", + "text_level": 1, + "bbox": [ + 76, + 599, + 344, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In both steps of our approach, the video alignment and the representation learning, we represent objects as neural meshes, i.e. meshes with surface features to capture the geometry and appearance of an object instance or category [9, 16, 26, 36]. In particular, the geometric representation is a triangular mesh, where we denote the set of vertices as $\\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|V|}$ . The appearance is represented by storing one or multiple appearance features $\\mathrm{F} = \\{\\{f_i^k\\in \\mathbb{R}^D\\}_{k = 1}^{|F|}\\}_{i = 1}^{|V|}$ at each mesh vertex. Together, the geometry and appearance define a neural mesh as $S = \\{V,F\\}$ .", + "bbox": [ + 75, + 622, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Self-supervised Alignment of Objects", + "text_level": 1, + "bbox": [ + 76, + 786, + 398, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to align the camera poses of multiple object-centric videos into a common coordinate frame. To achieve this, we represent each object-centric video as a neural mesh with self-supervised surface features. In particular, we utilize off-the-shelf structure-from motion [20] to obtain a coarse object shape reconstruction for each video.", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the reconstructed shapes cover the whole object, as the object-centric videos move in a full circle around the object. We post-process the reconstructed point cloud to clean it and generate a watertight mesh, for which we provide details in the supplementary material. Subsequently, we project the reconstructed coarse meshes into the feature map $\\Psi(I)$ that is obtained from a self-supervised transformer backbone [18]. We collect from every video a set of feature vectors for every vertex $v_i$ that describe the appearance of a local patch (Figure 2). Thus the number of features per vertex depends on the number of images in which the vertex is visible. As feature extractor $\\Psi$ , we use a self-supervised vision transformer [18] which has shown emerging correspondence matching abilities.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finding geometric and appearance correspondences. Given the mesh vertices of the source object instance V and the corresponding aggregated features F, we aim to align them to the reference counterparts $\\bar{\\mathbf{V}}$ and $\\bar{\\mathbf{F}}$ . In practice, we select a reference video at random from the set of all available videos and align the remaining videos to the reference. More precisely, we aim to optimize the transformation T, which is composed of rotation, translation and scale, under which the transformed source vertices and corresponding features yield the minimal distance to the reference counterparts with respect to geometry as well as appearance. Formally, our optimization problem optimizes", + "bbox": [ + 496, + 301, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathrm {T}} \\mathcal {D} (S, \\bar {S}, \\mathrm {T}) = \\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) + \\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 492, + 890, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the similarity between two videos given a transformation $T$ , which combines a geometric distance between the mesh geometries $\\mathcal{D}_{\\mathrm{geo}}(\\mathrm{V},\\bar{\\mathrm{V}})$ and an appearance distance $\\mathcal{D}_{\\mathrm{app}}(S,\\bar{S})$ between surface features.", + "bbox": [ + 496, + 523, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming that the object instances shapes contain a negligible variance and no symmetries, a suitable geometric distance is the Chamfer Distance defined as", + "bbox": [ + 496, + 584, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} | | _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 636, + 890, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\chi (\\mathrm{v}_i)$ is the vertex index of the Euclidean nearest neighbor of vertex $\\mathbf{V}_i$ in the respective other set of vertices", + "bbox": [ + 496, + 679, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\chi \\left(\\mathrm {v} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m i n} | | \\mathrm {T v} _ {i} - \\mathrm {v} _ {j} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\mathrm {V} \\\\ j \\in 1 \\dots | \\bar {\\mathrm {V}} | \\\\ \\operatorname {a r g m i n} | | \\mathrm {T v} _ {j} - \\mathrm {v} _ {i} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 719, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, as the 3D object is rather coarse and noisy, and the alignment can be ambiguous due to symmetries or shape differences among objects, we optimize each vertex to also be geometrically close to its nearest neighbor in feature space using the appearance distance", + "bbox": [ + 496, + 785, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i}, \\mathrm {f} _ {i})} | | _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22789", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the nearest neighbor in feature space $\\psi (\\mathbf{v})$ defined as", + "bbox": [ + 76, + 90, + 460, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\underset {j \\in 1 \\dots | \\bar {\\mathrm {V}} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\mathrm {V} \\\\ \\underset {j \\in 1 \\dots | \\mathrm {V} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 118, + 467, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the self-supervised vertex features are view-dependent, the appearance distance computes the minimum feature distance across all views to select the nearest neighbor.", + "bbox": [ + 76, + 188, + 467, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weighting Correspondences. An open challenge for the alignment of casually captured object-centric videos is that the estimated correspondence pairs between videos can be unreliable. For example, due to errors in the shape reconstruction or significant topology changes among different object instances, such as one bicycle having support wheels whereas the other does not. In these cases the correspondences in the geometry and feature space are ill-defined which leads to unreliable correspondence estimates. To account for such unreliable correspondences, we introduce a weight factor for each correspondence pair that estimates its quality. At the core of the correspondence weighting, we introduce a 3D cyclical distance among the vertices of two neural meshes that is inspired by 2D cyclical distances [5] for correspondence estimation, and is defined as", + "bbox": [ + 75, + 234, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi \\left(\\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right)} \\right| \\right| _ {2}, \\text {w i t h} j = \\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 472, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The nested structure of our 3D cyclical distance first computes $j$ as the index of the nearest neighbor of vertex $\\mathrm{v}_i$ in the feature space, and in turn computes the nearest neighbor of $\\mathrm{v}_j$ as $\\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)}$ . Notably, $\\mathrm{d}_{\\text{cycle}}(\\mathrm{v}_i, \\mathrm{f}_i) = 0$ if the nearest neighbour maps back to the original vertex $\\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)} = \\mathrm{v}_i$ and hence the correspondence is reliable. Building on this 3D cyclical distance, we define the validity criteria for each pair of vertices as the sum of cyclical distances of the correspondence pair", + "bbox": [ + 75, + 500, + 468, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}, \\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right) = - \\frac {\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i} , \\mathrm {f} _ {i}\\right) + \\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {j} , \\mathrm {f} _ {j}\\right)}{2 \\tau (\\mathrm {D} (\\mathrm {V}) + \\mathrm {D} (\\bar {\\mathrm {V}}))}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 647, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{D}(\\cdot)$ is the diameter of a neural mesh given as $\\mathrm{D(V)} = \\max_{\\mathrm{v}_i,\\mathrm{v}_j\\in \\mathrm{V}}||\\mathrm{v}_i - \\mathrm{v}_j||_2$", + "bbox": [ + 76, + 691, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the final weight factor for a correspondence pair we use the softmax normalization $\\sigma_{\\tau}(i,j) = \\mathrm{Softmax}_{\\tau}(\\rho (\\mathrm{v}_i,\\mathrm{f}_i,\\mathrm{v}_j,\\mathrm{f}_j))$ across all feature and gemoetric correspondences. For the softmax normalization, we introduce the temperature $\\tau$ , which enables us to steer between taking into account fewer high quality correspondences or more low quality ones (see Section 4.4). Together with the weighting we formulate the weighted geometric distance as", + "bbox": [ + 75, + 723, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {g e o}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\chi (\\mathrm {v} _ {i})) \\| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and likewise the weighted appearance distance as", + "bbox": [ + 500, + 90, + 826, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {a p p}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\psi (\\mathrm {v} _ {i})) | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i})} | | _ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 119, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our final distance measure to compare two neural meshes is computed as", + "bbox": [ + 498, + 166, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (S, \\bar {S}) = (1 - \\alpha) \\mathcal {D} _ {\\text {g e o}} ^ {*} (S, \\bar {S}) + \\alpha \\mathcal {D} _ {\\text {a p p}} ^ {*} (S, \\bar {S}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 209, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To find an approximately optimal solution, we use a RANSAC strategy, where we randomly choose four vertices on the source surface mesh. Together with their nearest neighbors in the feature space on the reference surface mesh, we estimate a single transformation using the Umeyama method [24].", + "bbox": [ + 498, + 242, + 890, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D Pose Estimation In-the-Wild", + "text_level": 1, + "bbox": [ + 500, + 345, + 779, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to perform 3D pose estimation in in-the-wild images. To achieve this, we generalize our approach from the multi-view setting used to align object-centric videos, towards 3D pose inference from a single image. Our model uses a feature extractor $\\Psi_w(I) = F \\in \\mathbb{R}^{D \\times H \\times W}$ to obtain image features from input image $I$ , where $w$ denotes the parameters of the backbone. The backbone output is a feature map $F$ with feature vectors $f_i \\in \\mathbb{R}^D$ at positions $i$ on a 2D lattice. For training, we use the aligned object-centric videos (Section 3.2) to train the weights $w$ of the feature extractor $\\Psi_w$ such that it predicts dense correspondences between image pixels and the 3D neural mesh template. Specifically, we relate the features of an image $\\Psi_w(I)$ extracted by a backbone feature extractor to the vertex and background features by Von-Mises-Fisher (vMF) probability distributions [12]. In particular, we model the likelihood of generating the feature at an image pixel $f_i$ from corresponding vertex feature $f_r$ as $P(f_i | f_r) = c_p(\\kappa) e^{\\kappa f_i \\cdot f_r}$ , where $f_r$ is the mean of each vMF kernel, $\\kappa$ is the corresponding concentration parameter, and $c_p$ is the normalization constant ( $\\|f_i\\| = 1$ , $\\|f_r\\| = 1$ ). We also model the likelihood of generating the feature $f_i$ from background feature as $P(f_i | \\beta) = c_p(\\kappa) e^{\\kappa f_i \\cdot \\beta}$ for $\\beta \\in \\mathbb{R}^D$ .", + "bbox": [ + 496, + 369, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When learning the models, as described next, we will learn the vertex features $\\{f_r\\}$ , the background feature $\\beta \\in \\mathbb{R}^D$ , and the parameters $w$ of the neural network backbone $\\Psi_w$ . We emphasize that our model requires that the backbone must be able to extract features that are invariant to the viewpoint of the object to ensure that $f_i \\cdot f_r$ is large irrespective of the viewpoint.", + "bbox": [ + 496, + 718, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning viewpoint-invariant vertex features. For training our model, we use the visible vertex features and their corresponding image features $\\mathcal{P} = \\{(f_r,f_i)\\}$ . Further, we use image features randomly sampled from the background $\\mathcal{B} = \\{f_i\\}$ . As optimization objective, we use the", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22790", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cross-entropy loss", + "bbox": [ + 76, + 92, + 200, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r a i n}} = - \\sum_ {(f _ {r}, f _ {i}) \\in \\mathcal {P}} \\log \\left(\\frac {P (f _ {i} | f _ {r})}{\\sum P (f _ {i} | f _ {r}) + \\sum P (f _ {i} | \\beta)}\\right) \\\\ - \\sum_ {f _ {i} \\in \\mathcal {B}} \\log \\left(\\frac {P \\left(f _ {i} \\mid \\beta\\right)}{\\sum P \\left(f _ {i} \\mid f _ {r}\\right) + \\sum P \\left(f _ {i} \\mid \\beta\\right)}\\right). \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 116, + 468, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D pose inference. We use the mesh with the vertex features $\\{f_r\\}$ , the background feature $\\beta$ and the trained backbone $\\Psi_w$ to estimate the camera pose $\\alpha$ via render-and-compare. At each optimization step, we render a feature map $\\{\\bar{f}_i(\\alpha)\\}$ under pose $\\alpha$ and compare it with the encoder's feature map $F = \\{f_i\\}$ . Determined by the rendering, each feature map consists of foreground features $F_{front}$ and background features $F_{back} = F \\setminus F_{front}$ . Thereupon, we maximize the joint likelihood for all image features under the assumption of independence, given as", + "bbox": [ + 76, + 224, + 468, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (F | \\alpha , \\{f _ {r} \\}, \\beta) =\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 387, + 419, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\prod_ {f _ {i} \\in F _ {\\text {f r o n t}}} \\max _ {f ^ {\\prime} \\in \\{\\bar {f} _ {i} (\\alpha), \\beta \\}} P (f _ {i} | f ^ {\\prime}) \\prod_ {f _ {i} \\in F _ {\\text {b a c k}}} P (f _ {i} | \\beta). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 407, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note, by allowing foreground image features to be generated by the background feature, we also account for clutter.", + "bbox": [ + 76, + 452, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We estimate the pose by first finding the best initialization of the object pose $\\alpha$ by computing the joint likelihood (Eq.12) for a set of pre-defined poses via template matching and choosing the one with the highest likelihood. Subsequently, we iteratively update our initial pose using a differentiable renderer to obtain the final pose prediction $\\hat{\\alpha}$ .", + "bbox": [ + 75, + 481, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 585, + 209, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we discuss our experimental setup (Section 4.1), present baselines and results for unsupervised alignment of object-centric videos (Section 4.2) and 3D pose estimation in-the-wild (Section 4.3). Additionally, we perform ablations of key model components in Section 4.4.", + "bbox": [ + 75, + 611, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 696, + 266, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset for alignment. To evaluate the unsupervised alignment of object-centric videos, we use the recently released Common Objects in 3D (CO3D) dataset [19] that provides images of multiple object categories, with a large amount of intra-category instance variation, and with varied object viewpoints. It contains 1.5 million frames, capturing objects from 50 categories, across nearly 19k scenes. For each object instance, CO3D provides approximately $100 - 200$ frames promising a $360^{\\circ}$ viewpoint sweep with handheld cameras. CO3D supplements these videos with relative camera poses and estimated object point clouds using Structure-from-Motion [20].", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find that the unfiltered videos of CO3D are not ideal for our purpose. In particular, we find that videos with little viewpoint variation lead to inferior structure-from-motion results. Also, videos that are not focusing on the object's center in 3D or are taken too close to it, contain little information for correspondence learning. Therefore, we filter the videos accordingly, targeting 50 videos per category. For multiple categories we end up with less than 50 videos namely, \"remote\" 17, \"mouse\" 15, \"tv\" 16, \"toilet\" 7, \"toybus\" 41, \"hairdryer\" 28, \"couch\" 49, and \"cellphone\" 23. With our simple filters, we end up aiming for 50 videos per category. More precise details for the filtering procedure are appended in the supplementary. As labels, we use the ground truth pose annotations provided by ZSP [5], that cover ten object instances of twenty different categories.", + "bbox": [ + 496, + 90, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets for 3D pose estimation in-the-wild. We evaluate on two common datasets PASCAL3D+ [30] and ObjectNet3D [31]. While PASCAL3D+ provides poses for the 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers pose annotations for over 100 categories. The object-centric video dataset CO3D covers 50 categories from the MS-COCO [15] dataset. We find 23 common categories across ObjectNet3D and CO3D, even tolerating the gap between a toybus in CO3D and a real one in PASCAL3D+ and ObjectNet3D. We believe that this non-negligible gap could be bridged by exploiting the multiple viewpoint knowledge of the same object instance. Overall we validate on PASCAL3D+ with 6233 images, using the same validation set as [25], and on ObjectNet3D on 12039 images. Following [38], we center all objects.", + "bbox": [ + 496, + 323, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. In our alignment step, we use $\\tau = 100$ and $\\alpha = 0.2$ . Further, we leverage as self-supervised ViT the publicly available small version of DINOv2 [18] with 21M parameters and a patch size of 14. At the input we use a resolution $448 \\times 448$ ending up with a $32 \\times 32$ feature map, where each feature yields 384 dimensions. In our second step, we use the same ViT as backbone and freeze its parameters. Further, we add on top three ResNet blocks with an upsampling step preceding the final block. Ending up with a $64 \\times 64$ feature map, where each feature has 128 dimensions. We optimize the cross-entropy loss for 10 epochs with Adam [10]. In one epoch, we make use of all filtered videos. The training for each category-level representation takes less than an hour on a single NVIDIA GeForce RTX 2080.", + "bbox": [ + 496, + 556, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that the quality of our alignment method and the subsequent representation learning can vary depending on the chosen reference video. Therefore, we randomly choose five reference videos per category and report the mean performance and the standard deviation across all results.", + "bbox": [ + 496, + 787, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the $30^{\\circ}$ accuracy for pose estimation where the angle error for an estimated rotation $\\mathbf{R}_{pred}$ and a ground", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e5c58287bdc8a714e9bbe5d09e99eabcc25f3d0fda9e577b1f32be809cdbc391.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Per CategoryAll Categories (20)
backpackcarchairkeyboardlaptopmotorcycleAcc. 30°Acc. 15°
TEASER++ [35]1.05.09.08.06.01.03.81.1
ZSP [5]44.065.047.069.085.085.049.428.4
UCD+ [6]67.086.076.076.0100.0100.069.854.6
Ours85.6100.098.982.2100.0100.077.061.6
±10.5±0.0±3.5±10.7±0.0±0.0±11.9±15.9
", + "bbox": [ + 125, + 88, + 846, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Unsupervised alignment evaluation on the CO3D dataset across 20 categories. The reported metric is the ${30}^{ \\circ }$ accuracy if not stated otherwise. The mean is computed across all 20 categories. We see that our method substantially outperforms the state of the art.", + "bbox": [ + 75, + 209, + 893, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d9b2eb42ca8c0b84d3cc7360f142f2044d7c4583a56d72fe90d07e5338e6790.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of two unsupervised alignment methods. The first row shows the alignment of our proposed method. The second row shows the alignment using ZSP [5]. For both methods we use the 5th object instance from left as reference. We see that our proposed method is more accurate compared with ZSP. Especially for cars ZSP often confuses back and front." + ], + "image_footnote": [], + "bbox": [ + 76, + 252, + 895, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "truth rotation $\\mathrm{R}_{pred}$ is given as", + "bbox": [ + 76, + 551, + 282, + 566 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\left(\\mathrm {R} _ {p r e d}, \\mathrm {R} _ {g t}\\right) = \\operatorname {a r c c o s} \\left(\\frac {1}{2} \\operatorname {t r} \\left(\\mathrm {R} _ {p r e d} ^ {\\mathrm {T}} \\mathrm {R} _ {g t}\\right) - 1\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 573, + 468, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that for the current state of unsupervised pose estimation, achieving $30^{\\circ}$ precision remains unsolved.", + "bbox": [ + 76, + 612, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Unsupervised Alignment", + "text_level": 1, + "bbox": [ + 76, + 650, + 305, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the evaluation protocol of ZSP [5] and measure the alignment of one object instance to the nine remaining ones of the same category that are labelled. Additionally, we report the standard deviation across the chosen reference object instances. The quantitative results in Table 1 show that our proposed method significantly improves the state of the art by $7.2\\%$ from $69.8\\%$ to $77.0\\%$ . Our alignment algorithm can more efficiently use the video frames compared to ZSP [5], which only compares a single RGB-D frame from the source video with many RGB-D frames of the reference video. We note that ZSP uses DINOv1 features in contrast to our method, which uses DINOv2. Therefore, we provide an ablation of our method with respect to different feature extractors in the supplementary. One reason for our model to outperform $\\mathrm{UCD + }$ [6], an extension of", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ZSP, is likely that our optimization does exploit the object geometry extensively, whereas others are using it only for refinement. A qualitative comparison of our method against ZSP is depicted in Figure 3. It shows that our alignments are highly accurate despite a large variability in the object instances. We note that at the time of writing, there is no source code publicly available to compare with UCD+.", + "bbox": [ + 496, + 551, + 893, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. In-the-Wild 3D Pose Estimation", + "text_level": 1, + "bbox": [ + 498, + 665, + 782, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we are not aware of any unsupervised method learning pose estimation from videos, we compare our pose estimation method against two supervised methods [27, 38] and ZSP. We provide ZSP with ten uniformly-distributed images of the same reference video that our method uses. Further, we provide ZSP with depth annotations using the category-level CAD models and pose annotations in the PASCAL3D+ and ObjectNet3D data. Despite our method not requiring any depth information, it outperforms ZSP by a large-margin on both PASCAL3D+, see Table 2, and on ObjectNet3D, see Table 3. Qualitative results are depicted in Figure 4. We find that ZSP is highly compute intensive, requiring 10.92 seconds per sample on average, while our proposed method takes only 0.22 seconds on average.", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22792", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da94dbf0317209e8752a94fc1fcc88d67ac3d5c63662b084491758c435eaeac2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 176, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6457e57a978f98377072015ef41987f4cb4fc49df2d67000ca281b67cc0ced99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 90, + 279, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8c3a3f0f45fec3db1768b2994baa7856b8749db0d2f5bab36545766080f37dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 380, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ec5bf4e05badb3942a151afc3e8cf0a40bfff810abaa68bd837b9474b324e85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 90, + 482, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54fed3bdcc5db28eafe45010dac46e097d8a8a5e26f483334a4ba7ecbae85fe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 90, + 584, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eb96cf3df0035265a7b06577bf0907bc09c7b7d1c91aa6bad36061349072c32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 90, + 684, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c537b59d6b78307d6a65921861b9b4e1d5a23faa4d6703ad5d67988e4fd4aa8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 90, + 787, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9527fe4e4ec2cc3237ee078b1c97bda3513a063ae11fa65563b6c45582aaa0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 888, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/127efd52c9091f7ebc18af2986ffa1b9ea175b3c7d7c24f06e732db1be4285da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 167, + 176, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9473087bb87c53d579268e71bc74e771cc0905785e23175c66556a3f28e2310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 167, + 279, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46b735850508d56f7733296b6bfaaf466ba547bc9ea2359ec5b058ee195c8e1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 167, + 380, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0a426b2e4e90b6e11cdd70aafcc13477ba36f70bc8e9ce2c299d9c1bb78108c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 167, + 482, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/997ca964013342ec44bdfcd0546d7ee3b80bf14e6052940ce76a51095ef06af6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 167, + 584, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6888b98e86cb938403a550fab636e2233b2ac229ae7496014b47d048017220ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 167, + 684, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/670c4c8f0b4edba2d62abba3e89e44710ad3a385bab939c95bb05e6437e267b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 167, + 787, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5c65a8be0ffa9d8772b821d4a3d28511c9f3665652227e71d9c0c1cbd112f97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 167, + 888, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab1e5de0bca90902f74bc020686352642d8106d84d663e1b9805c78412d7f252.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of our method (top) and ZSP (bottom) at category-level 3D pose prediction in the wild on samples from PASCAL3D+ and ObjectNet3D (we randomly selected the samples to demonstrate the diversity of the results). For both methods, we overlay our coarse mesh reconstruction in the predicted 3D pose." + ], + "image_footnote": [], + "bbox": [ + 80, + 244, + 176, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a10c90cd36249a4e9a7a992535bb4713c65ef7f6ed84a94504f1a9186bd9a7aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 244, + 279, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7a8d97c76e04fa8089c1a862f360ec37852aa673bd4fade7d65fd17890804ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 244, + 380, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19efa5764ce397c3085c7101a5ef66b73321314bc974a794c0e2e8d939aff650.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 244, + 482, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/498bf511aeaca7479cc82d2cecca00a3c6a87a17f0f045ec273d1cb68e731e1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 244, + 584, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/438bc6e48d6653ffd70f1eb7d5dfe85f370c72142ede424f026d1fbc9f759f7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 244, + 684, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89be11eb1f2fb3c63e8f94719a37d212b0cb96a1b58fd2a077c60bf7fda1cd15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 244, + 787, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b264be3f4c36326eae96aec7ba14ec3cb654bd9c4aa69bfed92eccabe6c10270.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 244, + 888, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c203ba66501ea395bd32eb54a69b5ff40a223cbfc73d4f88aefe40914261cc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodbicyclebuscarchairmotorcyclecouchtvMean
SupervisedStarMap [38]83.294.490.075.468.879.885.882.49
VoGE [27]82.698.199.090.587.594.983.990.93
UnsupervisedZSP61.721.461.642.643.152.939.046.0
±14.7±8.1±11.4±11.3±22.1±16.0±36.2±17.1
Ours58.479.398.251.967.076.653.169.2
±5.0±11.8±1.0±10.5±8.9±13.7±20.4±10.2
", + "bbox": [ + 99, + 388, + 870, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. 3D Pose Estimation in-the-wild on 7 categories of PASCAL3D+. Top two rows show supervised methods (as upper bound) while bottom two rows show unsupervised methods. The reported metric is $30^{\\circ}$ accuracy. The mean is averaged over all 7 categories. Our method shows superior performance over ZSP, which requires depth annotations.", + "bbox": [ + 75, + 516, + 892, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Categorical discussion. We observe that our method performs better for categories with only small topology changes and deformations, (e.g. car, microwave, couch) compared to categories with large intra-class variability (e.g. chair). Further, we recognize, that our method even generalizes well from a toybus to a real bus. Besides that, we analyze, that categories with less available videos (e.g. remote, TV, toilet) on average achieve lower performance.", + "bbox": [ + 75, + 585, + 468, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation", + "text_level": 1, + "bbox": [ + 76, + 723, + 181, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unsupervised alignment. Using the ground-truth annotations of our five references, we measure the effect of both parameters introduced in the alignment method. Namely, the appearance distance weight $\\alpha$ and the cyclical distance temperature $\\tau$ . We remark that for the distance between two meshes with surface features, the appearance weight trades-off feature correspondences versus Euclidean correspondences. Where an appearance weight of $\\alpha = 0$ means that the distance depends solely on the Euclidean correspondences. Contrarily, an appearance weight of $\\alpha = 1$ results", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d32767268aec32c4ff5b68f848e74cd0d77d05f3f69318f37292d52f03e0fd34.jpg", + "image_caption": [ + "Figure 5. We report the $30^{\\circ}$ accuracy of our alignment method for different choices of our appearance weight $\\alpha$ and our cyclical distance temperature $\\tau$ resulting in different distances between two meshes with surface features. We see that the maximum accuracy of $75.2\\%$ is reached for $\\alpha = 0.2$ and $\\tau = 100$ ." + ], + "image_footnote": [], + "bbox": [ + 500, + 580, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in solely depending on feature correspondences. Further, our cyclical distance temperature weights each correspondence, implicitly trading-off many low-quality correspondences versus few high-quality ones. Intuitively, increas", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22793", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/663ad3c88364403685d0f475ff6fdd5895dfbe0f1a24d521a83f9a0479716636.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
phonem'waveb'packbenchcuph'dryerlaptopmouseremotetoaster
ZSP46.450.523.150.833.021.760.528.841.628.8
± 8.4± 44.3± 14.9± 17.0± 18.7± 13.3± 8.9± 10.1± 18.1± 8.16
Ours54.680.318.062.138.214.153.344.754.460.6
± 3.4± 21.7± 12.0± 7.1± 21.6± 5.8± 9.7± 9.9± 4.4± 2.2
toiletb'cyclebuscarchaircouchk'boardm'cyclesuitcasetvMean
ZSP56.358.630.560.336.855.546.850.325.837.942.2
± 13.9± 10.4± 10.6± 9.1± 10.6± 16.1± 14.5± 20.7± 14.3± 35.1± 15.9
Ours39.657.878.398.152.276.626.969.015.553.252.4
± 12.7± 5.1± 12.1± 0.9± 9.6± 12.2± 8.4± 9.1± 5.5± 22.0± 9.8
", + "bbox": [ + 86, + 87, + 883, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b49b5f4427e2534dd459a8ad828b8a837f5d3de473da26bd7766da5949160360.jpg", + "table_caption": [ + "Table 3. 3D Pose Evaluation on 10 categories of ObjectNet3D. The mean is averaged over 20 categories. Metric is the ${30}^{ \\circ }$ accuracy. Despite not requiring any depth information, our method significantly outperforms ZSP." + ], + "table_footnote": [], + "table_body": "
MethodAcc. 30°Acc. 15°
PASCAL3D+Regression66.025.8
Ours69.241.3
ObjectNet3DRegression44.516.4
Ours52.425.5
", + "bbox": [ + 91, + 306, + 455, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Average $30^{\\circ}$ and $15^{\\circ}$ accuracies on PASCAL3D+ and ObjectNet3D for using directly neural network regression.", + "bbox": [ + 75, + 397, + 470, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the value of $\\tau$ results in averaging over more correspondences, while decreasing $\\tau$ results in taking only the correspondences with high validity into account. In Figure 5, we see that both parameters yield a significant impact on the $30^{\\circ}$ accuracy. With an optimum for $\\alpha = 0.2$ and $\\tau = 100$ . Intuitively, this means that taking many correspondences into account is more beneficial. Additionally, the Euclidean correspondences are weighted four times as much as the feature correspondences. Besides that, the ablation shows that while many correspondences are essential for using solely Euclidean correspondences, the opposite is true when using solely feature correspondences.", + "bbox": [ + 75, + 454, + 468, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D pose estimation in-the-wild. Following the alignment, the in-the-wild 3D pose estimation task can also be solved using neural network regression with the 6D rotation representation proposed in [39]. However, we observe that the results are worse than our 3D template learning method combined with render-and-compare, see Table 4.", + "bbox": [ + 75, + 637, + 468, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 739, + 204, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a model which substantially outperforms existing applicable baselines for the task of unsupervised category-level 3D pose estimation in-the-wild. However, our proposed method does not yet reach the performance of fully supervised baselines. One advancement we aspire is to relax the rigidity constraint of our shape model. Therefore, we plan to leverage the aligned reconstructions and introduce a parameterized model for the shape. A deformable shape would yield the potential to improve the correspond-", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dence learning as well as the subsequent matching of features at inference. Moreover, we see a future research direction in enabling the model to learn from a continuous stream of data, instead of building on a set of pre-recorded videos. This would even better reflect the complex real-world scenarios of embodied agents.", + "bbox": [ + 496, + 309, + 893, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 425, + 619, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a highly challenging (but realistic) task: unsupervised category-level 3D pose estimation from object-centric videos. In our proposed task, a model is required to align object-centric videos of instances of an object category without having any pose-labelled data. Subsequently, the model learns a 3D representation from the aligned videos to perform 3D category-level pose estimation in the wild. Our task defines a complex real-world problem which requires both semantic and geometric understanding of objects, and we demonstrate that existing baselines cannot solve the task. We further proposed a novel method for unsupervised learning of category-level 3D pose estimation that follows a two-step process: 1) A multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching. 2) Learning dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. The results showed that our proposed method achieves large improvements over all baselines, and we hope that our work will pave the ground for future advances in this important research direction.", + "bbox": [ + 496, + 454, + 893, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 825, + 679, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adam Kortylewski acknowledges support for his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No. 468670075.", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22794", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shuichi Akizuki and Manabu Hashimoto. Asm-net: Category-level pose and shape estimation using parametric deformation. In Proceedings of the British Machine Vision Conference, pages 1-13, 2021. 1", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 2", + "[3] Kai Chen and Qi Dou. Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2773-2782, 2021. 1, 2", + "[4] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 1", + "[5] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 2, 4, 5, 6", + "[6] Walter Goodwin, Ioannis Havoutis, and Ingmar Posner. You only look at one: Category-level object representations for pose estimation from a single example, 2023. 2, 6", + "[7] Alexander Grabner, Peter M Roth, and Vincent Lepetit. 3d pose estimation and 3d model retrieval for objects in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3022-3031, 2018. 1, 2", + "[8] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Inferring 3d object pose in rgb-d images. arXiv preprint arXiv:1502.04652, 2015. 1", + "[9] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3303-3312, 2021. 1, 3", + "[10] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[11] D Kirkpatrick and Raimund Seidel. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4):551-559, 1983. 1", + "[12] Adam Kortylewski, Ju He, Qing Liu, and Alan L Yuille. Compositional convolutional neural networks: A deep architecture with innate robustness to partial occlusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8940-8949, 2020. 4", + "[13] Georgios Kouros, Shubham Shrivastava, Cedric Picron, Sushruth Nagesh, Punarjay Chakravarty, and Tinne Tuyte-laars. Category-level pose retrieval with contrastive features learnt with occlusion augmentation. arXiv preprint arXiv:2208.06195, 2022. 2", + "[14] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 1, 2", + "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[16] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 1, 3", + "[17] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 2", + "[18] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5", + "[19] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 1, 5", + "[20] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3, 5", + "[21] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE international conference on computer vision, pages 2686-2694, 2015. 2", + "[22] Hung-Yu Tseng, Shalini De Mello, Jonathan Tremblay, Sifei Liu, Stan Birchfield, Ming-Hsuan Yang, and Jan Kautz. Few-shot viewpoint estimation. arXiv preprint arXiv:1905.04957, 2019. 2", + "[23] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2", + "[24] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 4", + "[25] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12645-12654, 2020. 5", + "[26] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 1, 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22795", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 2, 6, 7", + "[28] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 1", + "[29] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 1, 2", + "[30] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 5", + "[31] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 1, 2, 5", + "[32] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. BMVC, 2019. 1, 2", + "[33] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In 2021 International Conference on 3D Vision (3DV), pages 74–84. IEEE, 2021. 1, 2", + "[34] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3090-3106, 2022. 2", + "[35] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Transactions on Robotics, 37(2):314-333, 2020. 6", + "[36] Yanjie Ze and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems, 35:27469-27483, 2022. 1, 3", + "[37] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In The Eleventh International Conference on Learning Representations, 2022. 1, 2", + "[38] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 1, 2, 5, 6, 7", + "[39] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "computer vision and pattern recognition, pages 5745-5753, 2019.8", + "bbox": [ + 535, + 92, + 890, + 118 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22796", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_model.json b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..04b76e170c1980815a971e8085a83ae10b1d089e --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.18, + 0.314, + 0.198 + ], + "angle": 0, + "content": "Leonhard Sommer" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.181, + 0.478, + 0.198 + ], + "angle": 0, + "content": "Artur Jesslen" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.181, + 0.609, + 0.199 + ], + "angle": 0, + "content": "Eddy Ilg" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.181, + 0.816, + 0.199 + ], + "angle": 0, + "content": "Adam Kortylewski" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.216, + 0.482, + 0.235 + ], + "angle": 0, + "content": "1University of Freiburg" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.216, + 0.668, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Saarland University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.312, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.664 + ], + "angle": 0, + "content": "Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild on the Pascal3D+ and ObjectNet3D datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.208, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Category-level object pose estimation is a fundamentally important task in computer vision and robotics with a multitude of real-world applications, e.g. for training 3D generative models on real data and for robots that need to grasp and manipulate objects. However, defining and determining the pose of an object is a task that is far from easy. Current approaches achieve high performance, but they require large amounts of annotated training data to generalize successfully [1, 26, 28, 33, 38], or additional inputs during inference, such as CAD Models [7, 14], 3D Shapes [3, 32, 37] or RGB-D [8]. However, all of these are either time-consuming to obtain or not available in practice at all." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.497 + ], + "angle": 0, + "content": "This motivates the development of methods for learning category-level 3D pose estimators in a fully unsupervised fashion. While doing so from images in the wild seems infeasible, object-centric video data [19] offers a more accessible alternative. Such videos can be easily captured using consumer-grade cameras and makes it possible to leverage coarse 3D reconstructions during training, providing a practical and cost-effective method for collecting data. Therefore, we propose the new task of learning a single-image category-level 3D pose estimator from casually captured object-centric videos without any human labels or other supervision. In practice, we leverage CO3D [19] as training data and show that our proposed model is able to generalize and predict accurate poses in the wild for Pascal3D+ [29] and ObjectNet3D [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.698 + ], + "angle": 0, + "content": "We address the challenging task of learning category-level 3D pose in an unsupervised fashion from casually captured object-centric videos. In particular, we propose a two-step pipeline (Figure 1). The first step extracts DINOv2 [18] features from the images and reconstructs a coarse 3D mesh from the video with off-the-shelf methods [4, 11]. Building on this input, we introduce a novel 3D alignment procedure, where a key contribution is a novel 3D cyclical distance in terms of geometry and appearance that enables the robust alignment of shape reconstructions even under severe noise and variations in the object topology. As a result, we can align all objects from the object-centric training videos into a canonical coordinate frame without supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.853 + ], + "angle": 0, + "content": "In a second step, we leverage the canonical poses and 3D meshes obtained from the first step to train a category-level neural mesh [9, 16, 26, 36] in an unsupervised manner. In particular, we represent objects using a prototypical 3D mesh with surface features to capture the geometry and neural appearance of an object category and train a neural network backbone to predict, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. Finally, the object 3D pose is solved using a pose fitting algorithm based on the estimated correspondence pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We demonstrate that our method outperforms all baselines by a large margin at the unsupervised alignment of object-centric videos on the CO3D [19] dataset. Moreover," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22787" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.298, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.09, + 0.613, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.091, + 0.888, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 1. Illustration of our approach for the unsupervised learning of category-level 3D pose. Our method starts from unaligned object-centric videos of an object category (left) and aligns these into a canonical coordinate frame in a self-supervised manner using a prototypical 3D mesh and self-supervised transformer features (center). Using the aligned videos, we train a neural network backbone to predict 2D-3D correspondences from a single image to enable 3D object pose estimation in the wild (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.414 + ], + "angle": 0, + "content": "our model provides faithful and robust predictions in-the-wild on Pascal3D+ [29] and ObjectNet3D [31] despite being trained from raw object-centric videos only." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.219, + 0.448 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.67 + ], + "angle": 0, + "content": "Supervised Category-Level 3D Pose Estimation. Traditional methods to determine object poses were to label keypoints in images and train supervised methods to predict them [23, 38], or to use only pose labels and directly predict them by casting the pose estimation problem as bin classification by discretizing the pose space [21]. More recent methods utilize 3D meshes of objects or object categories. NeMo [26] uses 3D meshes with neural features that are rendered-and-compared to feature maps from a CNN to obtain pose estimates. [13] predicts a single embedding vector per image and shows that superior performance can be obtained by simply retrieving the closest training sample. In contrast to the above methods, our approach is fully unsupervised." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Few-Shot and Zero-Shot Pose Prediction. [33] proposes to train a supervised pose estimator across many categories and shows that their approach can generalize well to similar but unseen objects. Many works implement zero-shot pose estimation by conditioning the model on the 3D shape of the unseen object [3, 32, 34, 37] or by leveraging renderings of CAD models [7, 14]. Other methods use few-shot learning [22] or zero-shot learning with DINOv2 [18]. In contrast to the above, our method does not require any annotated dataset, 3D shapes or CAD models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Pose Alignment. The work from Goodwin et al. [5] (ZSP) is most close to the first step of our method, as it aligns the poses of two object-centric videos in a fully unsupervised fashion. Similar to our work, they use DINO [2] to obtain semantic correspondences. They perform first a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.655 + ], + "angle": 0, + "content": "coarse alignment by matching one image from the source video to one of many images from the reference video. Then they leverage cyclical distances to select few promising correspondences in the two images, and finally leverage respective depth maps to align both images using least squares. Goodwin et al. extend their work in [6] (UCD+) to match many images from the source video to many of the reference video. By finding a consensus over these many to many alignments with a single transformation they demonstrate improved performance. The first step of our work is similar to these works by that it also leverages DINO features and cyclical distances. However, our work adds a geometric distance to perform the alignment directly in 3D and introduces weighted correspondences to enable the necessary robust regression of the SE3 transformations from noisy and inaccurate geometries, which leads to significant improvements in the alignment accuracy. Note also that these previous approaches used RGB-D inputs, while our method works on images directly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Surface Embeddings and Neural Mesh Models. Recent work uses known poses and approximate object geometries to learn features in 3D space to uniquely identify parts of objects. [17] first used known mesh templates of deformable objects to train a network that predicts surface embeddings from the images. NeMo [26] presents a generative model trained with contrastive learning. [27] presents an extension through replacing vertices by Gaussian ellipsoids and using volume rendering. Similar to our work, many recent works leverage pre-trained vision transformers DINO [2] and DINOv2 [18] to unproject image features onto depth maps [5, 6, 37]. In contrast to [17, 26, 27], our approach does not require any pose annotation, while also going beyond [5, 6] by enabling 3D pose estimation in the wild from a single image, and not requiring RGB-D images and CAD models as input [37]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22788" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.473, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Illustration of a neural mesh. It consists of a 3D mesh with one or several neural features per vertex. These features encode patch-level features from a feature extractor. In our unsupervised alignment method, we capture the viewpoint-dependent features from DINOv2. For pose estimation, we learn a category-level neural mesh with viewpoint-invariant features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.369, + 0.17, + 0.385 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.473, + 0.593 + ], + "angle": 0, + "content": "In this section, we describe our approach for learning category-level 3D pose estimation without supervision from object-centric videos. Our method proceeds in a two-step approach. First, we align object instances across videos in an unsupervised manner to bring them into a canonical reference frame (Section 3.2). Given the aligned videos, our model learns to establish dense correspondences between images and a reconstructed 3D template mesh by predicting a feature vector for each pixel in a 2D image that corresponds to a visible vertex in the template mesh (Section 3.3). Finally, we describe how our model can efficiently estimate object poses from in-the-wild data using the predicted correspondences via render-and-compare." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.345, + 0.615 + ], + "angle": 0, + "content": "3.1. Meshes with Surface Features" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In both steps of our approach, the video alignment and the representation learning, we represent objects as neural meshes, i.e. meshes with surface features to capture the geometry and appearance of an object instance or category [9, 16, 26, 36]. In particular, the geometric representation is a triangular mesh, where we denote the set of vertices as \\( \\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|V|} \\). The appearance is represented by storing one or multiple appearance features \\( \\mathrm{F} = \\{\\{f_i^k\\in \\mathbb{R}^D\\}_{k = 1}^{|F|}\\}_{i = 1}^{|V|} \\) at each mesh vertex. Together, the geometry and appearance define a neural mesh as \\( S = \\{V,F\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.4, + 0.804 + ], + "angle": 0, + "content": "3.2. Self-supervised Alignment of Objects" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our goal is to align the camera poses of multiple object-centric videos into a common coordinate frame. To achieve this, we represent each object-centric video as a neural mesh with self-supervised surface features. In particular, we utilize off-the-shelf structure-from motion [20] to obtain a coarse object shape reconstruction for each video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Note that the reconstructed shapes cover the whole object, as the object-centric videos move in a full circle around the object. We post-process the reconstructed point cloud to clean it and generate a watertight mesh, for which we provide details in the supplementary material. Subsequently, we project the reconstructed coarse meshes into the feature map \\(\\Psi(I)\\) that is obtained from a self-supervised transformer backbone [18]. We collect from every video a set of feature vectors for every vertex \\(v_i\\) that describe the appearance of a local patch (Figure 2). Thus the number of features per vertex depends on the number of images in which the vertex is visible. As feature extractor \\(\\Psi\\), we use a self-supervised vision transformer [18] which has shown emerging correspondence matching abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Finding geometric and appearance correspondences. Given the mesh vertices of the source object instance V and the corresponding aggregated features F, we aim to align them to the reference counterparts \\(\\bar{\\mathbf{V}}\\) and \\(\\bar{\\mathbf{F}}\\). In practice, we select a reference video at random from the set of all available videos and align the remaining videos to the reference. More precisely, we aim to optimize the transformation T, which is composed of rotation, translation and scale, under which the transformed source vertices and corresponding features yield the minimal distance to the reference counterparts with respect to geometry as well as appearance. Formally, our optimization problem optimizes" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.493, + 0.892, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathrm {T}} \\mathcal {D} (S, \\bar {S}, \\mathrm {T}) = \\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) + \\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.585 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the similarity between two videos given a transformation \\(T\\), which combines a geometric distance between the mesh geometries \\(\\mathcal{D}_{\\mathrm{geo}}(\\mathrm{V},\\bar{\\mathrm{V}})\\) and an appearance distance \\(\\mathcal{D}_{\\mathrm{app}}(S,\\bar{S})\\) between surface features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Assuming that the object instances shapes contain a negligible variance and no symmetries, a suitable geometric distance is the Chamfer Distance defined as" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.637, + 0.892, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} | | _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.711 + ], + "angle": 0, + "content": "where \\(\\chi (\\mathrm{v}_i)\\) is the vertex index of the Euclidean nearest neighbor of vertex \\(\\mathbf{V}_i\\) in the respective other set of vertices" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.72, + 0.892, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\chi \\left(\\mathrm {v} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m i n} | | \\mathrm {T v} _ {i} - \\mathrm {v} _ {j} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\mathrm {V} \\\\ j \\in 1 \\dots | \\bar {\\mathrm {V}} | \\\\ \\operatorname {a r g m i n} | | \\mathrm {T v} _ {j} - \\mathrm {v} _ {i} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.862 + ], + "angle": 0, + "content": "However, as the 3D object is rather coarse and noisy, and the alignment can be ambiguous due to symmetries or shape differences among objects, we optimize each vertex to also be geometrically close to its nearest neighbor in feature space using the appearance distance" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.87, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i}, \\mathrm {f} _ {i})} | | _ {2}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22789" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.462, + 0.108 + ], + "angle": 0, + "content": "with the nearest neighbor in feature space \\(\\psi (\\mathbf{v})\\) defined as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.119, + 0.468, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\underset {j \\in 1 \\dots | \\bar {\\mathrm {V}} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\mathrm {V} \\\\ \\underset {j \\in 1 \\dots | \\mathrm {V} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.189, + 0.468, + 0.234 + ], + "angle": 0, + "content": "As the self-supervised vertex features are view-dependent, the appearance distance computes the minimum feature distance across all views to select the nearest neighbor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Weighting Correspondences. An open challenge for the alignment of casually captured object-centric videos is that the estimated correspondence pairs between videos can be unreliable. For example, due to errors in the shape reconstruction or significant topology changes among different object instances, such as one bicycle having support wheels whereas the other does not. In these cases the correspondences in the geometry and feature space are ill-defined which leads to unreliable correspondence estimates. To account for such unreliable correspondences, we introduce a weight factor for each correspondence pair that estimates its quality. At the core of the correspondence weighting, we introduce a 3D cyclical distance among the vertices of two neural meshes that is inspired by 2D cyclical distances [5] for correspondence estimation, and is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.473, + 0.469, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi \\left(\\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right)} \\right| \\right| _ {2}, \\text {w i t h} j = \\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.638 + ], + "angle": 0, + "content": "The nested structure of our 3D cyclical distance first computes \\( j \\) as the index of the nearest neighbor of vertex \\( \\mathrm{v}_i \\) in the feature space, and in turn computes the nearest neighbor of \\( \\mathrm{v}_j \\) as \\( \\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)} \\). Notably, \\( \\mathrm{d}_{\\text{cycle}}(\\mathrm{v}_i, \\mathrm{f}_i) = 0 \\) if the nearest neighbour maps back to the original vertex \\( \\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)} = \\mathrm{v}_i \\) and hence the correspondence is reliable. Building on this 3D cyclical distance, we define the validity criteria for each pair of vertices as the sum of cyclical distances of the correspondence pair" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.648, + 0.469, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}, \\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right) = - \\frac {\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i} , \\mathrm {f} _ {i}\\right) + \\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {j} , \\mathrm {f} _ {j}\\right)}{2 \\tau (\\mathrm {D} (\\mathrm {V}) + \\mathrm {D} (\\bar {\\mathrm {V}}))}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.692, + 0.469, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathrm{D}(\\cdot)\\) is the diameter of a neural mesh given as \\(\\mathrm{D(V)} = \\max_{\\mathrm{v}_i,\\mathrm{v}_j\\in \\mathrm{V}}||\\mathrm{v}_i - \\mathrm{v}_j||_2\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.469, + 0.859 + ], + "angle": 0, + "content": "To obtain the final weight factor for a correspondence pair we use the softmax normalization \\(\\sigma_{\\tau}(i,j) = \\mathrm{Softmax}_{\\tau}(\\rho (\\mathrm{v}_i,\\mathrm{f}_i,\\mathrm{v}_j,\\mathrm{f}_j))\\) across all feature and gemoetric correspondences. For the softmax normalization, we introduce the temperature \\(\\tau\\) , which enables us to steer between taking into account fewer high quality correspondences or more low quality ones (see Section 4.4). Together with the weighting we formulate the weighted geometric distance as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.87, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {g e o}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\chi (\\mathrm {v} _ {i})) \\| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.827, + 0.107 + ], + "angle": 0, + "content": "and likewise the weighted appearance distance as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.121, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {a p p}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\psi (\\mathrm {v} _ {i})) | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i})} | | _ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.167, + 0.892, + 0.197 + ], + "angle": 0, + "content": "Our final distance measure to compare two neural meshes is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.21, + 0.892, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (S, \\bar {S}) = (1 - \\alpha) \\mathcal {D} _ {\\text {g e o}} ^ {*} (S, \\bar {S}) + \\alpha \\mathcal {D} _ {\\text {a p p}} ^ {*} (S, \\bar {S}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.333 + ], + "angle": 0, + "content": "To find an approximately optimal solution, we use a RANSAC strategy, where we randomly choose four vertices on the source surface mesh. Together with their nearest neighbors in the feature space on the reference surface mesh, we estimate a single transformation using the Umeyama method [24]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.346, + 0.78, + 0.361 + ], + "angle": 0, + "content": "3.3. 3D Pose Estimation In-the-Wild" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Our goal is to perform 3D pose estimation in in-the-wild images. To achieve this, we generalize our approach from the multi-view setting used to align object-centric videos, towards 3D pose inference from a single image. Our model uses a feature extractor \\(\\Psi_w(I) = F \\in \\mathbb{R}^{D \\times H \\times W}\\) to obtain image features from input image \\(I\\), where \\(w\\) denotes the parameters of the backbone. The backbone output is a feature map \\(F\\) with feature vectors \\(f_i \\in \\mathbb{R}^D\\) at positions \\(i\\) on a 2D lattice. For training, we use the aligned object-centric videos (Section 3.2) to train the weights \\(w\\) of the feature extractor \\(\\Psi_w\\) such that it predicts dense correspondences between image pixels and the 3D neural mesh template. Specifically, we relate the features of an image \\(\\Psi_w(I)\\) extracted by a backbone feature extractor to the vertex and background features by Von-Mises-Fisher (vMF) probability distributions [12]. In particular, we model the likelihood of generating the feature at an image pixel \\(f_i\\) from corresponding vertex feature \\(f_r\\) as \\(P(f_i | f_r) = c_p(\\kappa) e^{\\kappa f_i \\cdot f_r}\\), where \\(f_r\\) is the mean of each vMF kernel, \\(\\kappa\\) is the corresponding concentration parameter, and \\(c_p\\) is the normalization constant (\\(\\|f_i\\| = 1\\), \\(\\|f_r\\| = 1\\)). We also model the likelihood of generating the feature \\(f_i\\) from background feature as \\(P(f_i | \\beta) = c_p(\\kappa) e^{\\kappa f_i \\cdot \\beta}\\) for \\(\\beta \\in \\mathbb{R}^D\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.825 + ], + "angle": 0, + "content": "When learning the models, as described next, we will learn the vertex features \\(\\{f_r\\}\\), the background feature \\(\\beta \\in \\mathbb{R}^D\\), and the parameters \\(w\\) of the neural network backbone \\(\\Psi_w\\). We emphasize that our model requires that the backbone must be able to extract features that are invariant to the viewpoint of the object to ensure that \\(f_i \\cdot f_r\\) is large irrespective of the viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learning viewpoint-invariant vertex features. For training our model, we use the visible vertex features and their corresponding image features \\(\\mathcal{P} = \\{(f_r,f_i)\\}\\). Further, we use image features randomly sampled from the background \\(\\mathcal{B} = \\{f_i\\}\\). As optimization objective, we use the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22790" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.093, + 0.202, + 0.107 + ], + "angle": 0, + "content": "cross-entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.117, + 0.469, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r a i n}} = - \\sum_ {(f _ {r}, f _ {i}) \\in \\mathcal {P}} \\log \\left(\\frac {P (f _ {i} | f _ {r})}{\\sum P (f _ {i} | f _ {r}) + \\sum P (f _ {i} | \\beta)}\\right) \\\\ - \\sum_ {f _ {i} \\in \\mathcal {B}} \\log \\left(\\frac {P \\left(f _ {i} \\mid \\beta\\right)}{\\sum P \\left(f _ {i} \\mid f _ {r}\\right) + \\sum P \\left(f _ {i} \\mid \\beta\\right)}\\right). \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.225, + 0.47, + 0.378 + ], + "angle": 0, + "content": "3D pose inference. We use the mesh with the vertex features \\(\\{f_r\\}\\), the background feature \\(\\beta\\) and the trained backbone \\(\\Psi_w\\) to estimate the camera pose \\(\\alpha\\) via render-and-compare. At each optimization step, we render a feature map \\(\\{\\bar{f}_i(\\alpha)\\}\\) under pose \\(\\alpha\\) and compare it with the encoder's feature map \\(F = \\{f_i\\}\\). Determined by the rendering, each feature map consists of foreground features \\(F_{front}\\) and background features \\(F_{back} = F \\setminus F_{front}\\). Thereupon, we maximize the joint likelihood for all image features under the assumption of independence, given as" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.388, + 0.42, + 0.404 + ], + "angle": 0, + "content": "\\[\nP (F | \\alpha , \\{f _ {r} \\}, \\beta) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.408, + 0.469, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\prod_ {f _ {i} \\in F _ {\\text {f r o n t}}} \\max _ {f ^ {\\prime} \\in \\{\\bar {f} _ {i} (\\alpha), \\beta \\}} P (f _ {i} | f ^ {\\prime}) \\prod_ {f _ {i} \\in F _ {\\text {b a c k}}} P (f _ {i} | \\beta). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.453, + 0.469, + 0.481 + ], + "angle": 0, + "content": "Note, by allowing foreground image features to be generated by the background feature, we also account for clutter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.573 + ], + "angle": 0, + "content": "We estimate the pose by first finding the best initialization of the object pose \\(\\alpha\\) by computing the joint likelihood (Eq.12) for a set of pre-defined poses via template matching and choosing the one with the highest likelihood. Subsequently, we iteratively update our initial pose using a differentiable renderer to obtain the final pose prediction \\(\\hat{\\alpha}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.21, + 0.604 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.687 + ], + "angle": 0, + "content": "In this section, we discuss our experimental setup (Section 4.1), present baselines and results for unsupervised alignment of object-centric videos (Section 4.2) and 3D pose estimation in-the-wild (Section 4.3). Additionally, we perform ablations of key model components in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.267, + 0.714 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Dataset for alignment. To evaluate the unsupervised alignment of object-centric videos, we use the recently released Common Objects in 3D (CO3D) dataset [19] that provides images of multiple object categories, with a large amount of intra-category instance variation, and with varied object viewpoints. It contains 1.5 million frames, capturing objects from 50 categories, across nearly 19k scenes. For each object instance, CO3D provides approximately \\(100 - 200\\) frames promising a \\(360^{\\circ}\\) viewpoint sweep with handheld cameras. CO3D supplements these videos with relative camera poses and estimated object point clouds using Structure-from-Motion [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.319 + ], + "angle": 0, + "content": "We find that the unfiltered videos of CO3D are not ideal for our purpose. In particular, we find that videos with little viewpoint variation lead to inferior structure-from-motion results. Also, videos that are not focusing on the object's center in 3D or are taken too close to it, contain little information for correspondence learning. Therefore, we filter the videos accordingly, targeting 50 videos per category. For multiple categories we end up with less than 50 videos namely, \"remote\" 17, \"mouse\" 15, \"tv\" 16, \"toilet\" 7, \"toybus\" 41, \"hairdryer\" 28, \"couch\" 49, and \"cellphone\" 23. With our simple filters, we end up aiming for 50 videos per category. More precise details for the filtering procedure are appended in the supplementary. As labels, we use the ground truth pose annotations provided by ZSP [5], that cover ten object instances of twenty different categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Datasets for 3D pose estimation in-the-wild. We evaluate on two common datasets PASCAL3D+ [30] and ObjectNet3D [31]. While PASCAL3D+ provides poses for the 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers pose annotations for over 100 categories. The object-centric video dataset CO3D covers 50 categories from the MS-COCO [15] dataset. We find 23 common categories across ObjectNet3D and CO3D, even tolerating the gap between a toybus in CO3D and a real one in PASCAL3D+ and ObjectNet3D. We believe that this non-negligible gap could be bridged by exploiting the multiple viewpoint knowledge of the same object instance. Overall we validate on PASCAL3D+ with 6233 images, using the same validation set as [25], and on ObjectNet3D on 12039 images. Following [38], we center all objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.783 + ], + "angle": 0, + "content": "Implementation details. In our alignment step, we use \\(\\tau = 100\\) and \\(\\alpha = 0.2\\). Further, we leverage as self-supervised ViT the publicly available small version of DINOv2 [18] with 21M parameters and a patch size of 14. At the input we use a resolution \\(448 \\times 448\\) ending up with a \\(32 \\times 32\\) feature map, where each feature yields 384 dimensions. In our second step, we use the same ViT as backbone and freeze its parameters. Further, we add on top three ResNet blocks with an upsampling step preceding the final block. Ending up with a \\(64 \\times 64\\) feature map, where each feature has 128 dimensions. We optimize the cross-entropy loss for 10 epochs with Adam [10]. In one epoch, we make use of all filtered videos. The training for each category-level representation takes less than an hour on a single NVIDIA GeForce RTX 2080." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.865 + ], + "angle": 0, + "content": "We note that the quality of our alignment method and the subsequent representation learning can vary depending on the chosen reference video. Therefore, we randomly choose five reference videos per category and report the mean performance and the standard deviation across all results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We report the \\(30^{\\circ}\\) accuracy for pose estimation where the angle error for an estimated rotation \\(\\mathbf{R}_{pred}\\) and a ground" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22791" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.847, + 0.201 + ], + "angle": 0, + "content": "
Per CategoryAll Categories (20)
backpackcarchairkeyboardlaptopmotorcycleAcc. 30°Acc. 15°
TEASER++ [35]1.05.09.08.06.01.03.81.1
ZSP [5]44.065.047.069.085.085.049.428.4
UCD+ [6]67.086.076.076.0100.0100.069.854.6
Ours85.6100.098.982.2100.0100.077.061.6
±10.5±0.0±3.5±10.7±0.0±0.0±11.9±15.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.24 + ], + "angle": 0, + "content": "Table 1. Unsupervised alignment evaluation on the CO3D dataset across 20 categories. The reported metric is the \\( {30}^{ \\circ } \\) accuracy if not stated otherwise. The mean is computed across all 20 categories. We see that our method substantially outperforms the state of the art." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.253, + 0.896, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.483, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of two unsupervised alignment methods. The first row shows the alignment of our proposed method. The second row shows the alignment using ZSP [5]. For both methods we use the 5th object instance from left as reference. We see that our proposed method is more accurate compared with ZSP. Especially for cars ZSP often confuses back and front." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.552, + 0.284, + 0.568 + ], + "angle": 0, + "content": "truth rotation \\(\\mathrm{R}_{pred}\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.574, + 0.47, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\Delta \\left(\\mathrm {R} _ {p r e d}, \\mathrm {R} _ {g t}\\right) = \\operatorname {a r c c o s} \\left(\\frac {1}{2} \\operatorname {t r} \\left(\\mathrm {R} _ {p r e d} ^ {\\mathrm {T}} \\mathrm {R} _ {g t}\\right) - 1\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.613, + 0.47, + 0.644 + ], + "angle": 0, + "content": "We note that for the current state of unsupervised pose estimation, achieving \\(30^{\\circ}\\) precision remains unsolved." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.307, + 0.668 + ], + "angle": 0, + "content": "4.2. Unsupervised Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We follow the evaluation protocol of ZSP [5] and measure the alignment of one object instance to the nine remaining ones of the same category that are labelled. Additionally, we report the standard deviation across the chosen reference object instances. The quantitative results in Table 1 show that our proposed method significantly improves the state of the art by \\(7.2\\%\\) from \\(69.8\\%\\) to \\(77.0\\%\\). Our alignment algorithm can more efficiently use the video frames compared to ZSP [5], which only compares a single RGB-D frame from the source video with many RGB-D frames of the reference video. We note that ZSP uses DINOv1 features in contrast to our method, which uses DINOv2. Therefore, we provide an ablation of our method with respect to different feature extractors in the supplementary. One reason for our model to outperform \\(\\mathrm{UCD + }\\) [6], an extension of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.895, + 0.658 + ], + "angle": 0, + "content": "ZSP, is likely that our optimization does exploit the object geometry extensively, whereas others are using it only for refinement. A qualitative comparison of our method against ZSP is depicted in Figure 3. It shows that our alignments are highly accurate despite a large variability in the object instances. We note that at the time of writing, there is no source code publicly available to compare with UCD+." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.783, + 0.682 + ], + "angle": 0, + "content": "4.3. In-the-Wild 3D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As we are not aware of any unsupervised method learning pose estimation from videos, we compare our pose estimation method against two supervised methods [27, 38] and ZSP. We provide ZSP with ten uniformly-distributed images of the same reference video that our method uses. Further, we provide ZSP with depth annotations using the category-level CAD models and pose annotations in the PASCAL3D+ and ObjectNet3D data. Despite our method not requiring any depth information, it outperforms ZSP by a large-margin on both PASCAL3D+, see Table 2, and on ObjectNet3D, see Table 3. Qualitative results are depicted in Figure 4. We find that ZSP is highly compute intensive, requiring 10.92 seconds per sample on average, while our proposed method takes only 0.22 seconds on average." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22792" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.178, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.091, + 0.28, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.091, + 0.382, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.091, + 0.483, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.585, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.091, + 0.686, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.091, + 0.788, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.091, + 0.89, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.168, + 0.178, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.168, + 0.28, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.168, + 0.381, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.168, + 0.483, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.168, + 0.585, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.168, + 0.686, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.168, + 0.788, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.168, + 0.89, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.245, + 0.178, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.245, + 0.28, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.245, + 0.381, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.245, + 0.483, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.245, + 0.585, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.245, + 0.686, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.245, + 0.788, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.245, + 0.89, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.893, + 0.378 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of our method (top) and ZSP (bottom) at category-level 3D pose prediction in the wild on samples from PASCAL3D+ and ObjectNet3D (we randomly selected the samples to demonstrate the diversity of the results). For both methods, we overlay our coarse mesh reconstruction in the predicted 3D pose." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.389, + 0.871, + 0.507 + ], + "angle": 0, + "content": "
MethodbicyclebuscarchairmotorcyclecouchtvMean
SupervisedStarMap [38]83.294.490.075.468.879.885.882.49
VoGE [27]82.698.199.090.587.594.983.990.93
UnsupervisedZSP61.721.461.642.643.152.939.046.0
±14.7±8.1±11.4±11.3±22.1±16.0±36.2±17.1
Ours58.479.398.251.967.076.653.169.2
±5.0±11.8±1.0±10.5±8.9±13.7±20.4±10.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.517, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Table 2. 3D Pose Estimation in-the-wild on 7 categories of PASCAL3D+. Top two rows show supervised methods (as upper bound) while bottom two rows show unsupervised methods. The reported metric is \\(30^{\\circ}\\) accuracy. The mean is averaged over all 7 categories. Our method shows superior performance over ZSP, which requires depth annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.707 + ], + "angle": 0, + "content": "Categorical discussion. We observe that our method performs better for categories with only small topology changes and deformations, (e.g. car, microwave, couch) compared to categories with large intra-class variability (e.g. chair). Further, we recognize, that our method even generalizes well from a toybus to a real bus. Besides that, we analyze, that categories with less available videos (e.g. remote, TV, toilet) on average achieve lower performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.182, + 0.739 + ], + "angle": 0, + "content": "4.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Unsupervised alignment. Using the ground-truth annotations of our five references, we measure the effect of both parameters introduced in the alignment method. Namely, the appearance distance weight \\(\\alpha\\) and the cyclical distance temperature \\(\\tau\\). We remark that for the distance between two meshes with surface features, the appearance weight trades-off feature correspondences versus Euclidean correspondences. Where an appearance weight of \\(\\alpha = 0\\) means that the distance depends solely on the Euclidean correspondences. Contrarily, an appearance weight of \\(\\alpha = 1\\) results" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.582, + 0.891, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.743, + 0.893, + 0.813 + ], + "angle": 0, + "content": "Figure 5. We report the \\(30^{\\circ}\\) accuracy of our alignment method for different choices of our appearance weight \\(\\alpha\\) and our cyclical distance temperature \\(\\tau\\) resulting in different distances between two meshes with surface features. We see that the maximum accuracy of \\(75.2\\%\\) is reached for \\(\\alpha = 0.2\\) and \\(\\tau = 100\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "in solely depending on feature correspondences. Further, our cyclical distance temperature weights each correspondence, implicitly trading-off many low-quality correspondences versus few high-quality ones. Intuitively, increas" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22793" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.088, + 0.885, + 0.246 + ], + "angle": 0, + "content": "
phonem'waveb'packbenchcuph'dryerlaptopmouseremotetoaster
ZSP46.450.523.150.833.021.760.528.841.628.8
± 8.4± 44.3± 14.9± 17.0± 18.7± 13.3± 8.9± 10.1± 18.1± 8.16
Ours54.680.318.062.138.214.153.344.754.460.6
± 3.4± 21.7± 12.0± 7.1± 21.6± 5.8± 9.7± 9.9± 4.4± 2.2
toiletb'cyclebuscarchaircouchk'boardm'cyclesuitcasetvMean
ZSP56.358.630.560.336.855.546.850.325.837.942.2
± 13.9± 10.4± 10.6± 9.1± 10.6± 16.1± 14.5± 20.7± 14.3± 35.1± 15.9
Ours39.657.878.398.152.276.626.969.015.553.252.4
± 12.7± 5.1± 12.1± 0.9± 9.6± 12.2± 8.4± 9.1± 5.5± 22.0± 9.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.893, + 0.285 + ], + "angle": 0, + "content": "Table 3. 3D Pose Evaluation on 10 categories of ObjectNet3D. The mean is averaged over 20 categories. Metric is the \\( {30}^{ \\circ } \\) accuracy. Despite not requiring any depth information, our method significantly outperforms ZSP." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.308, + 0.456, + 0.389 + ], + "angle": 0, + "content": "
MethodAcc. 30°Acc. 15°
PASCAL3D+Regression66.025.8
Ours69.241.3
ObjectNet3DRegression44.516.4
Ours52.425.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.398, + 0.472, + 0.429 + ], + "angle": 0, + "content": "Table 4. Average \\(30^{\\circ}\\) and \\(15^{\\circ}\\) accuracies on PASCAL3D+ and ObjectNet3D for using directly neural network regression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.637 + ], + "angle": 0, + "content": "ing the value of \\(\\tau\\) results in averaging over more correspondences, while decreasing \\(\\tau\\) results in taking only the correspondences with high validity into account. In Figure 5, we see that both parameters yield a significant impact on the \\(30^{\\circ}\\) accuracy. With an optimum for \\(\\alpha = 0.2\\) and \\(\\tau = 100\\). Intuitively, this means that taking many correspondences into account is more beneficial. Additionally, the Euclidean correspondences are weighted four times as much as the feature correspondences. Besides that, the ablation shows that while many correspondences are essential for using solely Euclidean correspondences, the opposite is true when using solely feature correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.73 + ], + "angle": 0, + "content": "3D pose estimation in-the-wild. Following the alignment, the in-the-wild 3D pose estimation task can also be solved using neural network regression with the 6D rotation representation proposed in [39]. However, we observe that the results are worse than our 3D template learning method combined with render-and-compare, see Table 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.205, + 0.755 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We have proposed a model which substantially outperforms existing applicable baselines for the task of unsupervised category-level 3D pose estimation in-the-wild. However, our proposed method does not yet reach the performance of fully supervised baselines. One advancement we aspire is to relax the rigidity constraint of our shape model. Therefore, we plan to leverage the aligned reconstructions and introduce a parameterized model for the shape. A deformable shape would yield the potential to improve the correspond-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.403 + ], + "angle": 0, + "content": "dence learning as well as the subsequent matching of features at inference. Moreover, we see a future research direction in enabling the model to learn from a continuous stream of data, instead of building on a set of pre-recorded videos. This would even better reflect the complex real-world scenarios of embodied agents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.426, + 0.62, + 0.442 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.803 + ], + "angle": 0, + "content": "In this paper, we have proposed a highly challenging (but realistic) task: unsupervised category-level 3D pose estimation from object-centric videos. In our proposed task, a model is required to align object-centric videos of instances of an object category without having any pose-labelled data. Subsequently, the model learns a 3D representation from the aligned videos to perform 3D category-level pose estimation in the wild. Our task defines a complex real-world problem which requires both semantic and geometric understanding of objects, and we demonstrate that existing baselines cannot solve the task. We further proposed a novel method for unsupervised learning of category-level 3D pose estimation that follows a two-step process: 1) A multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching. 2) Learning dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. The results showed that our proposed method achieves large improvements over all baselines, and we hope that our work will pave the ground for future advances in this important research direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.827, + 0.681, + 0.844 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Adam Kortylewski acknowledges support for his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No. 468670075." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22794" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Shuichi Akizuki and Manabu Hashimoto. Asm-net: Category-level pose and shape estimation using parametric deformation. In Proceedings of the British Machine Vision Conference, pages 1-13, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[3] Kai Chen and Qi Dou. Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2773-2782, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.414 + ], + "angle": 0, + "content": "[5] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.458 + ], + "angle": 0, + "content": "[6] Walter Goodwin, Ioannis Havoutis, and Ingmar Posner. You only look at one: Category-level object representations for pose estimation from a single example, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[7] Alexander Grabner, Peter M Roth, and Vincent Lepetit. 3d pose estimation and 3d model retrieval for objects in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3022-3031, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[8] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Inferring 3d object pose in rgb-d images. arXiv preprint arXiv:1502.04652, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[9] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3303-3312, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[10] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[11] D Kirkpatrick and Raimund Seidel. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4):551-559, 1983. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[12] Adam Kortylewski, Ju He, Qing Liu, and Alan L Yuille. Compositional convolutional neural networks: A deep architecture with innate robustness to partial occlusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8940-8949, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[13] Georgios Kouros, Shubham Shrivastava, Cedric Picron, Sushruth Nagesh, Punarjay Chakravarty, and Tinne Tuyte-laars. Category-level pose retrieval with contrastive features learnt with occlusion augmentation. arXiv preprint arXiv:2208.06195, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.387, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[19] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE international conference on computer vision, pages 2686-2694, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[22] Hung-Yu Tseng, Shalini De Mello, Jonathan Tremblay, Sifei Liu, Stan Birchfield, Ming-Hsuan Yang, and Jan Kautz. Few-shot viewpoint estimation. arXiv preprint arXiv:1905.04957, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[24] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[25] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12645-12654, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22795" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[28] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[29] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[30] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[31] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[32] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. BMVC, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[33] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In 2021 International Conference on 3D Vision (3DV), pages 74–84. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[34] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3090-3106, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[35] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Transactions on Robotics, 37(2):314-333, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[36] Yanjie Ze and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems, 35:27469-27483, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[37] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In The Eleventh International Conference on Learning Representations, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[38] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 5745-5753, 2019.8" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22796" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_origin.pdf b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b203d48cc5c4dc5d3274752f2f508d8f93d16f6 --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/3103a10e-d4f1-46bf-b544-2a9a7eded8b5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa091b4f123035bd629da2b0ebc780515ddd1963558d0e46394b40ea6a9f048e +size 9877280 diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/full.md b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..775116c16e353337fa1390dc71948fe8d6043a79 --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/full.md @@ -0,0 +1,343 @@ +# Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos + +Leonhard Sommer + +Artur Jesslen + +Eddy Ilg + +Adam Kortylewski + +1University of Freiburg + +$^{2}$ Saarland University + +# Abstract + +Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild on the Pascal3D+ and ObjectNet3D datasets. + +# 1. Introduction + +Category-level object pose estimation is a fundamentally important task in computer vision and robotics with a multitude of real-world applications, e.g. for training 3D generative models on real data and for robots that need to grasp and manipulate objects. However, defining and determining the pose of an object is a task that is far from easy. Current approaches achieve high performance, but they require large amounts of annotated training data to generalize successfully [1, 26, 28, 33, 38], or additional inputs during inference, such as CAD Models [7, 14], 3D Shapes [3, 32, 37] or RGB-D [8]. However, all of these are either time-consuming to obtain or not available in practice at all. + +This motivates the development of methods for learning category-level 3D pose estimators in a fully unsupervised fashion. While doing so from images in the wild seems infeasible, object-centric video data [19] offers a more accessible alternative. Such videos can be easily captured using consumer-grade cameras and makes it possible to leverage coarse 3D reconstructions during training, providing a practical and cost-effective method for collecting data. Therefore, we propose the new task of learning a single-image category-level 3D pose estimator from casually captured object-centric videos without any human labels or other supervision. In practice, we leverage CO3D [19] as training data and show that our proposed model is able to generalize and predict accurate poses in the wild for Pascal3D+ [29] and ObjectNet3D [31]. + +We address the challenging task of learning category-level 3D pose in an unsupervised fashion from casually captured object-centric videos. In particular, we propose a two-step pipeline (Figure 1). The first step extracts DINOv2 [18] features from the images and reconstructs a coarse 3D mesh from the video with off-the-shelf methods [4, 11]. Building on this input, we introduce a novel 3D alignment procedure, where a key contribution is a novel 3D cyclical distance in terms of geometry and appearance that enables the robust alignment of shape reconstructions even under severe noise and variations in the object topology. As a result, we can align all objects from the object-centric training videos into a canonical coordinate frame without supervision. + +In a second step, we leverage the canonical poses and 3D meshes obtained from the first step to train a category-level neural mesh [9, 16, 26, 36] in an unsupervised manner. In particular, we represent objects using a prototypical 3D mesh with surface features to capture the geometry and neural appearance of an object category and train a neural network backbone to predict, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. Finally, the object 3D pose is solved using a pose fitting algorithm based on the estimated correspondence pairs. + +We demonstrate that our method outperforms all baselines by a large margin at the unsupervised alignment of object-centric videos on the CO3D [19] dataset. Moreover, + +![](images/c3cc4c7f81ffbb91fa36c268284df84f9a7641346611f9de0b7448095e243257.jpg) +Figure 1. Illustration of our approach for the unsupervised learning of category-level 3D pose. Our method starts from unaligned object-centric videos of an object category (left) and aligns these into a canonical coordinate frame in a self-supervised manner using a prototypical 3D mesh and self-supervised transformer features (center). Using the aligned videos, we train a neural network backbone to predict 2D-3D correspondences from a single image to enable 3D object pose estimation in the wild (right). + +![](images/b26b63ffc31e99dfec2a9179cb1ba8a3ad41c2546e2004f987ad9f6bdb4151c3.jpg) + +![](images/76a5e14bd63acda2f6f88be6ea9bf423816a19b80a84723817b31c86a58de509.jpg) + +our model provides faithful and robust predictions in-the-wild on Pascal3D+ [29] and ObjectNet3D [31] despite being trained from raw object-centric videos only. + +# 2. Related Work + +Supervised Category-Level 3D Pose Estimation. Traditional methods to determine object poses were to label keypoints in images and train supervised methods to predict them [23, 38], or to use only pose labels and directly predict them by casting the pose estimation problem as bin classification by discretizing the pose space [21]. More recent methods utilize 3D meshes of objects or object categories. NeMo [26] uses 3D meshes with neural features that are rendered-and-compared to feature maps from a CNN to obtain pose estimates. [13] predicts a single embedding vector per image and shows that superior performance can be obtained by simply retrieving the closest training sample. In contrast to the above methods, our approach is fully unsupervised. + +Few-Shot and Zero-Shot Pose Prediction. [33] proposes to train a supervised pose estimator across many categories and shows that their approach can generalize well to similar but unseen objects. Many works implement zero-shot pose estimation by conditioning the model on the 3D shape of the unseen object [3, 32, 34, 37] or by leveraging renderings of CAD models [7, 14]. Other methods use few-shot learning [22] or zero-shot learning with DINOv2 [18]. In contrast to the above, our method does not require any annotated dataset, 3D shapes or CAD models. + +Pose Alignment. The work from Goodwin et al. [5] (ZSP) is most close to the first step of our method, as it aligns the poses of two object-centric videos in a fully unsupervised fashion. Similar to our work, they use DINO [2] to obtain semantic correspondences. They perform first a + +coarse alignment by matching one image from the source video to one of many images from the reference video. Then they leverage cyclical distances to select few promising correspondences in the two images, and finally leverage respective depth maps to align both images using least squares. Goodwin et al. extend their work in [6] (UCD+) to match many images from the source video to many of the reference video. By finding a consensus over these many to many alignments with a single transformation they demonstrate improved performance. The first step of our work is similar to these works by that it also leverages DINO features and cyclical distances. However, our work adds a geometric distance to perform the alignment directly in 3D and introduces weighted correspondences to enable the necessary robust regression of the SE3 transformations from noisy and inaccurate geometries, which leads to significant improvements in the alignment accuracy. Note also that these previous approaches used RGB-D inputs, while our method works on images directly. + +Surface Embeddings and Neural Mesh Models. Recent work uses known poses and approximate object geometries to learn features in 3D space to uniquely identify parts of objects. [17] first used known mesh templates of deformable objects to train a network that predicts surface embeddings from the images. NeMo [26] presents a generative model trained with contrastive learning. [27] presents an extension through replacing vertices by Gaussian ellipsoids and using volume rendering. Similar to our work, many recent works leverage pre-trained vision transformers DINO [2] and DINOv2 [18] to unproject image features onto depth maps [5, 6, 37]. In contrast to [17, 26, 27], our approach does not require any pose annotation, while also going beyond [5, 6] by enabling 3D pose estimation in the wild from a single image, and not requiring RGB-D images and CAD models as input [37]. + +![](images/4a078b93488844856d028f8741441aa9417a231f14280e0398d570866f231b33.jpg) +Figure 2. Illustration of a neural mesh. It consists of a 3D mesh with one or several neural features per vertex. These features encode patch-level features from a feature extractor. In our unsupervised alignment method, we capture the viewpoint-dependent features from DINOv2. For pose estimation, we learn a category-level neural mesh with viewpoint-invariant features. + +# 3. Method + +In this section, we describe our approach for learning category-level 3D pose estimation without supervision from object-centric videos. Our method proceeds in a two-step approach. First, we align object instances across videos in an unsupervised manner to bring them into a canonical reference frame (Section 3.2). Given the aligned videos, our model learns to establish dense correspondences between images and a reconstructed 3D template mesh by predicting a feature vector for each pixel in a 2D image that corresponds to a visible vertex in the template mesh (Section 3.3). Finally, we describe how our model can efficiently estimate object poses from in-the-wild data using the predicted correspondences via render-and-compare. + +# 3.1. Meshes with Surface Features + +In both steps of our approach, the video alignment and the representation learning, we represent objects as neural meshes, i.e. meshes with surface features to capture the geometry and appearance of an object instance or category [9, 16, 26, 36]. In particular, the geometric representation is a triangular mesh, where we denote the set of vertices as $\mathrm{V} = \{\mathrm{v}_i\in \mathbb{R}^3\}_{i = 1}^{|V|}$ . The appearance is represented by storing one or multiple appearance features $\mathrm{F} = \{\{f_i^k\in \mathbb{R}^D\}_{k = 1}^{|F|}\}_{i = 1}^{|V|}$ at each mesh vertex. Together, the geometry and appearance define a neural mesh as $S = \{V,F\}$ . + +# 3.2. Self-supervised Alignment of Objects + +Our goal is to align the camera poses of multiple object-centric videos into a common coordinate frame. To achieve this, we represent each object-centric video as a neural mesh with self-supervised surface features. In particular, we utilize off-the-shelf structure-from motion [20] to obtain a coarse object shape reconstruction for each video. + +Note that the reconstructed shapes cover the whole object, as the object-centric videos move in a full circle around the object. We post-process the reconstructed point cloud to clean it and generate a watertight mesh, for which we provide details in the supplementary material. Subsequently, we project the reconstructed coarse meshes into the feature map $\Psi(I)$ that is obtained from a self-supervised transformer backbone [18]. We collect from every video a set of feature vectors for every vertex $v_i$ that describe the appearance of a local patch (Figure 2). Thus the number of features per vertex depends on the number of images in which the vertex is visible. As feature extractor $\Psi$ , we use a self-supervised vision transformer [18] which has shown emerging correspondence matching abilities. + +Finding geometric and appearance correspondences. Given the mesh vertices of the source object instance V and the corresponding aggregated features F, we aim to align them to the reference counterparts $\bar{\mathbf{V}}$ and $\bar{\mathbf{F}}$ . In practice, we select a reference video at random from the set of all available videos and align the remaining videos to the reference. More precisely, we aim to optimize the transformation T, which is composed of rotation, translation and scale, under which the transformed source vertices and corresponding features yield the minimal distance to the reference counterparts with respect to geometry as well as appearance. Formally, our optimization problem optimizes + +$$ +\min _ {\mathrm {T}} \mathcal {D} (S, \bar {S}, \mathrm {T}) = \mathcal {D} _ {\mathrm {g e o}} (\mathrm {V}, \bar {\mathrm {V}}) + \mathcal {D} _ {\mathrm {a p p}} (S, \bar {S}), \tag {1} +$$ + +where $\mathcal{D}$ is the similarity between two videos given a transformation $T$ , which combines a geometric distance between the mesh geometries $\mathcal{D}_{\mathrm{geo}}(\mathrm{V},\bar{\mathrm{V}})$ and an appearance distance $\mathcal{D}_{\mathrm{app}}(S,\bar{S})$ between surface features. + +Assuming that the object instances shapes contain a negligible variance and no symmetries, a suitable geometric distance is the Chamfer Distance defined as + +$$ +\mathcal {D} _ {\mathrm {g e o}} (\mathrm {V}, \bar {\mathrm {V}}) = \sum_ {\mathrm {v} _ {i} \in \mathrm {V} \cup \bar {\mathrm {V}}} | | \mathrm {v} _ {i} - \mathrm {v} _ {\chi (\mathrm {v} _ {i})} | | _ {2}, \tag {2} +$$ + +where $\chi (\mathrm{v}_i)$ is the vertex index of the Euclidean nearest neighbor of vertex $\mathbf{V}_i$ in the respective other set of vertices + +$$ +\chi \left(\mathrm {v} _ {i}\right) = \left\{ \begin{array}{l l} \operatorname {a r g m i n} | | \mathrm {T v} _ {i} - \mathrm {v} _ {j} | | _ {2}, & \text {f o r} \mathrm {v} _ {i} \in \mathrm {V} \\ j \in 1 \dots | \bar {\mathrm {V}} | \\ \operatorname {a r g m i n} | | \mathrm {T v} _ {j} - \mathrm {v} _ {i} | | _ {2}, & \text {f o r} \mathrm {v} _ {i} \in \bar {\mathrm {V}}. \end{array} \right. \tag {3} +$$ + +However, as the 3D object is rather coarse and noisy, and the alignment can be ambiguous due to symmetries or shape differences among objects, we optimize each vertex to also be geometrically close to its nearest neighbor in feature space using the appearance distance + +$$ +\mathcal {D} _ {\mathrm {a p p}} (S, \bar {S}) = \sum_ {\mathrm {v} _ {i} \in \mathrm {V} \cup \bar {\mathrm {V}}} | | \mathrm {v} _ {i} - \mathrm {v} _ {\psi (\mathrm {v} _ {i}, \mathrm {f} _ {i})} | | _ {2}, \tag {4} +$$ + +with the nearest neighbor in feature space $\psi (\mathbf{v})$ defined as + +$$ +\psi \left(\mathrm {v} _ {i}, \mathrm {f} _ {i}\right) = \left\{ \begin{array}{l l} \underset {j \in 1 \dots | \bar {\mathrm {V}} |} {\operatorname {a r g m i n}} \min _ {k, l} \left| \left| \mathrm {f} _ {j} ^ {k} - \mathrm {f} _ {i} ^ {l} \right| \right|, & \text {f o r v} _ {i} \in \mathrm {V} \\ \underset {j \in 1 \dots | \mathrm {V} |} {\operatorname {a r g m i n}} \min _ {k, l} \left| \left| \mathrm {f} _ {j} ^ {k} - \mathrm {f} _ {i} ^ {l} \right| \right|, & \text {f o r v} _ {i} \in \bar {\mathrm {V}}. \end{array} \right. \tag {5} +$$ + +As the self-supervised vertex features are view-dependent, the appearance distance computes the minimum feature distance across all views to select the nearest neighbor. + +Weighting Correspondences. An open challenge for the alignment of casually captured object-centric videos is that the estimated correspondence pairs between videos can be unreliable. For example, due to errors in the shape reconstruction or significant topology changes among different object instances, such as one bicycle having support wheels whereas the other does not. In these cases the correspondences in the geometry and feature space are ill-defined which leads to unreliable correspondence estimates. To account for such unreliable correspondences, we introduce a weight factor for each correspondence pair that estimates its quality. At the core of the correspondence weighting, we introduce a 3D cyclical distance among the vertices of two neural meshes that is inspired by 2D cyclical distances [5] for correspondence estimation, and is defined as + +$$ +\mathrm {d} _ {\text {c y c l e}} \left(\mathrm {v} _ {i}, \mathrm {f} _ {i}\right) = \left| \left| \mathrm {v} _ {i} - \mathrm {v} _ {\psi \left(\mathrm {v} _ {j}, \mathrm {f} _ {j}\right)} \right| \right| _ {2}, \text {w i t h} j = \psi \left(\mathrm {v} _ {i}, \mathrm {f} _ {i}\right). \tag {6} +$$ + +The nested structure of our 3D cyclical distance first computes $j$ as the index of the nearest neighbor of vertex $\mathrm{v}_i$ in the feature space, and in turn computes the nearest neighbor of $\mathrm{v}_j$ as $\mathrm{v}_{\psi(\mathrm{v}_j, \mathrm{f}_j)}$ . Notably, $\mathrm{d}_{\text{cycle}}(\mathrm{v}_i, \mathrm{f}_i) = 0$ if the nearest neighbour maps back to the original vertex $\mathrm{v}_{\psi(\mathrm{v}_j, \mathrm{f}_j)} = \mathrm{v}_i$ and hence the correspondence is reliable. Building on this 3D cyclical distance, we define the validity criteria for each pair of vertices as the sum of cyclical distances of the correspondence pair + +$$ +\rho \left(\mathrm {v} _ {i}, \mathrm {f} _ {i}, \mathrm {v} _ {j}, \mathrm {f} _ {j}\right) = - \frac {\mathrm {d} _ {\text {c y c l e}} \left(\mathrm {v} _ {i} , \mathrm {f} _ {i}\right) + \mathrm {d} _ {\text {c y c l e}} \left(\mathrm {v} _ {j} , \mathrm {f} _ {j}\right)}{2 \tau (\mathrm {D} (\mathrm {V}) + \mathrm {D} (\bar {\mathrm {V}}))}, \tag {7} +$$ + +where $\mathrm{D}(\cdot)$ is the diameter of a neural mesh given as $\mathrm{D(V)} = \max_{\mathrm{v}_i,\mathrm{v}_j\in \mathrm{V}}||\mathrm{v}_i - \mathrm{v}_j||_2$ + +To obtain the final weight factor for a correspondence pair we use the softmax normalization $\sigma_{\tau}(i,j) = \mathrm{Softmax}_{\tau}(\rho (\mathrm{v}_i,\mathrm{f}_i,\mathrm{v}_j,\mathrm{f}_j))$ across all feature and gemoetric correspondences. For the softmax normalization, we introduce the temperature $\tau$ , which enables us to steer between taking into account fewer high quality correspondences or more low quality ones (see Section 4.4). Together with the weighting we formulate the weighted geometric distance as + +$$ +\mathcal {D} _ {\mathrm {g e o}} ^ {*} (S, \bar {S}) = \sum_ {\mathrm {v} _ {i} \in \mathrm {V} \cup \bar {\mathrm {V}}} \sigma (i, \chi (\mathrm {v} _ {i})) \| \mathrm {v} _ {i} - \mathrm {v} _ {\chi (\mathrm {v} _ {i})} \| _ {2}, \tag {8} +$$ + +and likewise the weighted appearance distance as + +$$ +\mathcal {D} _ {\mathrm {a p p}} ^ {*} (S, \bar {S}) = \sum_ {\mathrm {v} _ {i} \in \mathrm {V} \cup \bar {\mathrm {V}}} \sigma (i, \psi (\mathrm {v} _ {i})) | | \mathrm {v} _ {i} - \mathrm {v} _ {\psi (\mathrm {v} _ {i})} | | _ {2}. \tag {9} +$$ + +Our final distance measure to compare two neural meshes is computed as + +$$ +\mathcal {L} (S, \bar {S}) = (1 - \alpha) \mathcal {D} _ {\text {g e o}} ^ {*} (S, \bar {S}) + \alpha \mathcal {D} _ {\text {a p p}} ^ {*} (S, \bar {S}). \tag {10} +$$ + +To find an approximately optimal solution, we use a RANSAC strategy, where we randomly choose four vertices on the source surface mesh. Together with their nearest neighbors in the feature space on the reference surface mesh, we estimate a single transformation using the Umeyama method [24]. + +# 3.3. 3D Pose Estimation In-the-Wild + +Our goal is to perform 3D pose estimation in in-the-wild images. To achieve this, we generalize our approach from the multi-view setting used to align object-centric videos, towards 3D pose inference from a single image. Our model uses a feature extractor $\Psi_w(I) = F \in \mathbb{R}^{D \times H \times W}$ to obtain image features from input image $I$ , where $w$ denotes the parameters of the backbone. The backbone output is a feature map $F$ with feature vectors $f_i \in \mathbb{R}^D$ at positions $i$ on a 2D lattice. For training, we use the aligned object-centric videos (Section 3.2) to train the weights $w$ of the feature extractor $\Psi_w$ such that it predicts dense correspondences between image pixels and the 3D neural mesh template. Specifically, we relate the features of an image $\Psi_w(I)$ extracted by a backbone feature extractor to the vertex and background features by Von-Mises-Fisher (vMF) probability distributions [12]. In particular, we model the likelihood of generating the feature at an image pixel $f_i$ from corresponding vertex feature $f_r$ as $P(f_i | f_r) = c_p(\kappa) e^{\kappa f_i \cdot f_r}$ , where $f_r$ is the mean of each vMF kernel, $\kappa$ is the corresponding concentration parameter, and $c_p$ is the normalization constant ( $\|f_i\| = 1$ , $\|f_r\| = 1$ ). We also model the likelihood of generating the feature $f_i$ from background feature as $P(f_i | \beta) = c_p(\kappa) e^{\kappa f_i \cdot \beta}$ for $\beta \in \mathbb{R}^D$ . + +When learning the models, as described next, we will learn the vertex features $\{f_r\}$ , the background feature $\beta \in \mathbb{R}^D$ , and the parameters $w$ of the neural network backbone $\Psi_w$ . We emphasize that our model requires that the backbone must be able to extract features that are invariant to the viewpoint of the object to ensure that $f_i \cdot f_r$ is large irrespective of the viewpoint. + +Learning viewpoint-invariant vertex features. For training our model, we use the visible vertex features and their corresponding image features $\mathcal{P} = \{(f_r,f_i)\}$ . Further, we use image features randomly sampled from the background $\mathcal{B} = \{f_i\}$ . As optimization objective, we use the + +cross-entropy loss + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t r a i n}} = - \sum_ {(f _ {r}, f _ {i}) \in \mathcal {P}} \log \left(\frac {P (f _ {i} | f _ {r})}{\sum P (f _ {i} | f _ {r}) + \sum P (f _ {i} | \beta)}\right) \\ - \sum_ {f _ {i} \in \mathcal {B}} \log \left(\frac {P \left(f _ {i} \mid \beta\right)}{\sum P \left(f _ {i} \mid f _ {r}\right) + \sum P \left(f _ {i} \mid \beta\right)}\right). \tag {11} \\ \end{array} +$$ + +3D pose inference. We use the mesh with the vertex features $\{f_r\}$ , the background feature $\beta$ and the trained backbone $\Psi_w$ to estimate the camera pose $\alpha$ via render-and-compare. At each optimization step, we render a feature map $\{\bar{f}_i(\alpha)\}$ under pose $\alpha$ and compare it with the encoder's feature map $F = \{f_i\}$ . Determined by the rendering, each feature map consists of foreground features $F_{front}$ and background features $F_{back} = F \setminus F_{front}$ . Thereupon, we maximize the joint likelihood for all image features under the assumption of independence, given as + +$$ +P (F | \alpha , \{f _ {r} \}, \beta) = +$$ + +$$ +\prod_ {f _ {i} \in F _ {\text {f r o n t}}} \max _ {f ^ {\prime} \in \{\bar {f} _ {i} (\alpha), \beta \}} P (f _ {i} | f ^ {\prime}) \prod_ {f _ {i} \in F _ {\text {b a c k}}} P (f _ {i} | \beta). \tag {12} +$$ + +Note, by allowing foreground image features to be generated by the background feature, we also account for clutter. + +We estimate the pose by first finding the best initialization of the object pose $\alpha$ by computing the joint likelihood (Eq.12) for a set of pre-defined poses via template matching and choosing the one with the highest likelihood. Subsequently, we iteratively update our initial pose using a differentiable renderer to obtain the final pose prediction $\hat{\alpha}$ . + +# 4. Experiments + +In this section, we discuss our experimental setup (Section 4.1), present baselines and results for unsupervised alignment of object-centric videos (Section 4.2) and 3D pose estimation in-the-wild (Section 4.3). Additionally, we perform ablations of key model components in Section 4.4. + +# 4.1. Experimental Setup + +Dataset for alignment. To evaluate the unsupervised alignment of object-centric videos, we use the recently released Common Objects in 3D (CO3D) dataset [19] that provides images of multiple object categories, with a large amount of intra-category instance variation, and with varied object viewpoints. It contains 1.5 million frames, capturing objects from 50 categories, across nearly 19k scenes. For each object instance, CO3D provides approximately $100 - 200$ frames promising a $360^{\circ}$ viewpoint sweep with handheld cameras. CO3D supplements these videos with relative camera poses and estimated object point clouds using Structure-from-Motion [20]. + +We find that the unfiltered videos of CO3D are not ideal for our purpose. In particular, we find that videos with little viewpoint variation lead to inferior structure-from-motion results. Also, videos that are not focusing on the object's center in 3D or are taken too close to it, contain little information for correspondence learning. Therefore, we filter the videos accordingly, targeting 50 videos per category. For multiple categories we end up with less than 50 videos namely, "remote" 17, "mouse" 15, "tv" 16, "toilet" 7, "toybus" 41, "hairdryer" 28, "couch" 49, and "cellphone" 23. With our simple filters, we end up aiming for 50 videos per category. More precise details for the filtering procedure are appended in the supplementary. As labels, we use the ground truth pose annotations provided by ZSP [5], that cover ten object instances of twenty different categories. + +Datasets for 3D pose estimation in-the-wild. We evaluate on two common datasets PASCAL3D+ [30] and ObjectNet3D [31]. While PASCAL3D+ provides poses for the 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers pose annotations for over 100 categories. The object-centric video dataset CO3D covers 50 categories from the MS-COCO [15] dataset. We find 23 common categories across ObjectNet3D and CO3D, even tolerating the gap between a toybus in CO3D and a real one in PASCAL3D+ and ObjectNet3D. We believe that this non-negligible gap could be bridged by exploiting the multiple viewpoint knowledge of the same object instance. Overall we validate on PASCAL3D+ with 6233 images, using the same validation set as [25], and on ObjectNet3D on 12039 images. Following [38], we center all objects. + +Implementation details. In our alignment step, we use $\tau = 100$ and $\alpha = 0.2$ . Further, we leverage as self-supervised ViT the publicly available small version of DINOv2 [18] with 21M parameters and a patch size of 14. At the input we use a resolution $448 \times 448$ ending up with a $32 \times 32$ feature map, where each feature yields 384 dimensions. In our second step, we use the same ViT as backbone and freeze its parameters. Further, we add on top three ResNet blocks with an upsampling step preceding the final block. Ending up with a $64 \times 64$ feature map, where each feature has 128 dimensions. We optimize the cross-entropy loss for 10 epochs with Adam [10]. In one epoch, we make use of all filtered videos. The training for each category-level representation takes less than an hour on a single NVIDIA GeForce RTX 2080. + +We note that the quality of our alignment method and the subsequent representation learning can vary depending on the chosen reference video. Therefore, we randomly choose five reference videos per category and report the mean performance and the standard deviation across all results. + +We report the $30^{\circ}$ accuracy for pose estimation where the angle error for an estimated rotation $\mathbf{R}_{pred}$ and a ground + +
Per CategoryAll Categories (20)
backpackcarchairkeyboardlaptopmotorcycleAcc. 30°Acc. 15°
TEASER++ [35]1.05.09.08.06.01.03.81.1
ZSP [5]44.065.047.069.085.085.049.428.4
UCD+ [6]67.086.076.076.0100.0100.069.854.6
Ours85.6100.098.982.2100.0100.077.061.6
±10.5±0.0±3.5±10.7±0.0±0.0±11.9±15.9
+ +Table 1. Unsupervised alignment evaluation on the CO3D dataset across 20 categories. The reported metric is the ${30}^{ \circ }$ accuracy if not stated otherwise. The mean is computed across all 20 categories. We see that our method substantially outperforms the state of the art. + +![](images/6d9b2eb42ca8c0b84d3cc7360f142f2044d7c4583a56d72fe90d07e5338e6790.jpg) +Figure 3. Qualitative comparison of two unsupervised alignment methods. The first row shows the alignment of our proposed method. The second row shows the alignment using ZSP [5]. For both methods we use the 5th object instance from left as reference. We see that our proposed method is more accurate compared with ZSP. Especially for cars ZSP often confuses back and front. + +truth rotation $\mathrm{R}_{pred}$ is given as + +$$ +\Delta \left(\mathrm {R} _ {p r e d}, \mathrm {R} _ {g t}\right) = \operatorname {a r c c o s} \left(\frac {1}{2} \operatorname {t r} \left(\mathrm {R} _ {p r e d} ^ {\mathrm {T}} \mathrm {R} _ {g t}\right) - 1\right). \tag {13} +$$ + +We note that for the current state of unsupervised pose estimation, achieving $30^{\circ}$ precision remains unsolved. + +# 4.2. Unsupervised Alignment + +We follow the evaluation protocol of ZSP [5] and measure the alignment of one object instance to the nine remaining ones of the same category that are labelled. Additionally, we report the standard deviation across the chosen reference object instances. The quantitative results in Table 1 show that our proposed method significantly improves the state of the art by $7.2\%$ from $69.8\%$ to $77.0\%$ . Our alignment algorithm can more efficiently use the video frames compared to ZSP [5], which only compares a single RGB-D frame from the source video with many RGB-D frames of the reference video. We note that ZSP uses DINOv1 features in contrast to our method, which uses DINOv2. Therefore, we provide an ablation of our method with respect to different feature extractors in the supplementary. One reason for our model to outperform $\mathrm{UCD + }$ [6], an extension of + +ZSP, is likely that our optimization does exploit the object geometry extensively, whereas others are using it only for refinement. A qualitative comparison of our method against ZSP is depicted in Figure 3. It shows that our alignments are highly accurate despite a large variability in the object instances. We note that at the time of writing, there is no source code publicly available to compare with UCD+. + +# 4.3. In-the-Wild 3D Pose Estimation + +As we are not aware of any unsupervised method learning pose estimation from videos, we compare our pose estimation method against two supervised methods [27, 38] and ZSP. We provide ZSP with ten uniformly-distributed images of the same reference video that our method uses. Further, we provide ZSP with depth annotations using the category-level CAD models and pose annotations in the PASCAL3D+ and ObjectNet3D data. Despite our method not requiring any depth information, it outperforms ZSP by a large-margin on both PASCAL3D+, see Table 2, and on ObjectNet3D, see Table 3. Qualitative results are depicted in Figure 4. We find that ZSP is highly compute intensive, requiring 10.92 seconds per sample on average, while our proposed method takes only 0.22 seconds on average. + +![](images/da94dbf0317209e8752a94fc1fcc88d67ac3d5c63662b084491758c435eaeac2.jpg) + +![](images/6457e57a978f98377072015ef41987f4cb4fc49df2d67000ca281b67cc0ced99.jpg) + +![](images/a8c3a3f0f45fec3db1768b2994baa7856b8749db0d2f5bab36545766080f37dd.jpg) + +![](images/2ec5bf4e05badb3942a151afc3e8cf0a40bfff810abaa68bd837b9474b324e85.jpg) + +![](images/54fed3bdcc5db28eafe45010dac46e097d8a8a5e26f483334a4ba7ecbae85fe4.jpg) + +![](images/5eb96cf3df0035265a7b06577bf0907bc09c7b7d1c91aa6bad36061349072c32.jpg) + +![](images/c537b59d6b78307d6a65921861b9b4e1d5a23faa4d6703ad5d67988e4fd4aa8b.jpg) + +![](images/e9527fe4e4ec2cc3237ee078b1c97bda3513a063ae11fa65563b6c45582aaa0a.jpg) + +![](images/127efd52c9091f7ebc18af2986ffa1b9ea175b3c7d7c24f06e732db1be4285da.jpg) + +![](images/a9473087bb87c53d579268e71bc74e771cc0905785e23175c66556a3f28e2310.jpg) + +![](images/46b735850508d56f7733296b6bfaaf466ba547bc9ea2359ec5b058ee195c8e1f.jpg) + +![](images/e0a426b2e4e90b6e11cdd70aafcc13477ba36f70bc8e9ce2c299d9c1bb78108c.jpg) + +![](images/997ca964013342ec44bdfcd0546d7ee3b80bf14e6052940ce76a51095ef06af6.jpg) + +![](images/6888b98e86cb938403a550fab636e2233b2ac229ae7496014b47d048017220ad.jpg) + +![](images/670c4c8f0b4edba2d62abba3e89e44710ad3a385bab939c95bb05e6437e267b0.jpg) + +![](images/d5c65a8be0ffa9d8772b821d4a3d28511c9f3665652227e71d9c0c1cbd112f97.jpg) + +![](images/ab1e5de0bca90902f74bc020686352642d8106d84d663e1b9805c78412d7f252.jpg) +Figure 4. Qualitative comparison of our method (top) and ZSP (bottom) at category-level 3D pose prediction in the wild on samples from PASCAL3D+ and ObjectNet3D (we randomly selected the samples to demonstrate the diversity of the results). For both methods, we overlay our coarse mesh reconstruction in the predicted 3D pose. + +![](images/a10c90cd36249a4e9a7a992535bb4713c65ef7f6ed84a94504f1a9186bd9a7aa.jpg) + +![](images/b7a8d97c76e04fa8089c1a862f360ec37852aa673bd4fade7d65fd17890804ec.jpg) + +![](images/19efa5764ce397c3085c7101a5ef66b73321314bc974a794c0e2e8d939aff650.jpg) + +![](images/498bf511aeaca7479cc82d2cecca00a3c6a87a17f0f045ec273d1cb68e731e1d.jpg) + +![](images/438bc6e48d6653ffd70f1eb7d5dfe85f370c72142ede424f026d1fbc9f759f7b.jpg) + +![](images/89be11eb1f2fb3c63e8f94719a37d212b0cb96a1b58fd2a077c60bf7fda1cd15.jpg) + +![](images/b264be3f4c36326eae96aec7ba14ec3cb654bd9c4aa69bfed92eccabe6c10270.jpg) + +
MethodbicyclebuscarchairmotorcyclecouchtvMean
SupervisedStarMap [38]83.294.490.075.468.879.885.882.49
VoGE [27]82.698.199.090.587.594.983.990.93
UnsupervisedZSP61.721.461.642.643.152.939.046.0
±14.7±8.1±11.4±11.3±22.1±16.0±36.2±17.1
Ours58.479.398.251.967.076.653.169.2
±5.0±11.8±1.0±10.5±8.9±13.7±20.4±10.2
+ +Table 2. 3D Pose Estimation in-the-wild on 7 categories of PASCAL3D+. Top two rows show supervised methods (as upper bound) while bottom two rows show unsupervised methods. The reported metric is $30^{\circ}$ accuracy. The mean is averaged over all 7 categories. Our method shows superior performance over ZSP, which requires depth annotations. + +Categorical discussion. We observe that our method performs better for categories with only small topology changes and deformations, (e.g. car, microwave, couch) compared to categories with large intra-class variability (e.g. chair). Further, we recognize, that our method even generalizes well from a toybus to a real bus. Besides that, we analyze, that categories with less available videos (e.g. remote, TV, toilet) on average achieve lower performance. + +# 4.4. Ablation + +Unsupervised alignment. Using the ground-truth annotations of our five references, we measure the effect of both parameters introduced in the alignment method. Namely, the appearance distance weight $\alpha$ and the cyclical distance temperature $\tau$ . We remark that for the distance between two meshes with surface features, the appearance weight trades-off feature correspondences versus Euclidean correspondences. Where an appearance weight of $\alpha = 0$ means that the distance depends solely on the Euclidean correspondences. Contrarily, an appearance weight of $\alpha = 1$ results + +![](images/d32767268aec32c4ff5b68f848e74cd0d77d05f3f69318f37292d52f03e0fd34.jpg) +Figure 5. We report the $30^{\circ}$ accuracy of our alignment method for different choices of our appearance weight $\alpha$ and our cyclical distance temperature $\tau$ resulting in different distances between two meshes with surface features. We see that the maximum accuracy of $75.2\%$ is reached for $\alpha = 0.2$ and $\tau = 100$ . + +in solely depending on feature correspondences. Further, our cyclical distance temperature weights each correspondence, implicitly trading-off many low-quality correspondences versus few high-quality ones. Intuitively, increas + +
phonem'waveb'packbenchcuph'dryerlaptopmouseremotetoaster
ZSP46.450.523.150.833.021.760.528.841.628.8
± 8.4± 44.3± 14.9± 17.0± 18.7± 13.3± 8.9± 10.1± 18.1± 8.16
Ours54.680.318.062.138.214.153.344.754.460.6
± 3.4± 21.7± 12.0± 7.1± 21.6± 5.8± 9.7± 9.9± 4.4± 2.2
toiletb'cyclebuscarchaircouchk'boardm'cyclesuitcasetvMean
ZSP56.358.630.560.336.855.546.850.325.837.942.2
± 13.9± 10.4± 10.6± 9.1± 10.6± 16.1± 14.5± 20.7± 14.3± 35.1± 15.9
Ours39.657.878.398.152.276.626.969.015.553.252.4
± 12.7± 5.1± 12.1± 0.9± 9.6± 12.2± 8.4± 9.1± 5.5± 22.0± 9.8
+ +Table 3. 3D Pose Evaluation on 10 categories of ObjectNet3D. The mean is averaged over 20 categories. Metric is the ${30}^{ \circ }$ accuracy. Despite not requiring any depth information, our method significantly outperforms ZSP. + +
MethodAcc. 30°Acc. 15°
PASCAL3D+Regression66.025.8
Ours69.241.3
ObjectNet3DRegression44.516.4
Ours52.425.5
+ +Table 4. Average $30^{\circ}$ and $15^{\circ}$ accuracies on PASCAL3D+ and ObjectNet3D for using directly neural network regression. + +ing the value of $\tau$ results in averaging over more correspondences, while decreasing $\tau$ results in taking only the correspondences with high validity into account. In Figure 5, we see that both parameters yield a significant impact on the $30^{\circ}$ accuracy. With an optimum for $\alpha = 0.2$ and $\tau = 100$ . Intuitively, this means that taking many correspondences into account is more beneficial. Additionally, the Euclidean correspondences are weighted four times as much as the feature correspondences. Besides that, the ablation shows that while many correspondences are essential for using solely Euclidean correspondences, the opposite is true when using solely feature correspondences. + +3D pose estimation in-the-wild. Following the alignment, the in-the-wild 3D pose estimation task can also be solved using neural network regression with the 6D rotation representation proposed in [39]. However, we observe that the results are worse than our 3D template learning method combined with render-and-compare, see Table 4. + +# 4.5. Limitations + +We have proposed a model which substantially outperforms existing applicable baselines for the task of unsupervised category-level 3D pose estimation in-the-wild. However, our proposed method does not yet reach the performance of fully supervised baselines. One advancement we aspire is to relax the rigidity constraint of our shape model. Therefore, we plan to leverage the aligned reconstructions and introduce a parameterized model for the shape. A deformable shape would yield the potential to improve the correspond- + +dence learning as well as the subsequent matching of features at inference. Moreover, we see a future research direction in enabling the model to learn from a continuous stream of data, instead of building on a set of pre-recorded videos. This would even better reflect the complex real-world scenarios of embodied agents. + +# 5. Conclusion + +In this paper, we have proposed a highly challenging (but realistic) task: unsupervised category-level 3D pose estimation from object-centric videos. In our proposed task, a model is required to align object-centric videos of instances of an object category without having any pose-labelled data. Subsequently, the model learns a 3D representation from the aligned videos to perform 3D category-level pose estimation in the wild. Our task defines a complex real-world problem which requires both semantic and geometric understanding of objects, and we demonstrate that existing baselines cannot solve the task. We further proposed a novel method for unsupervised learning of category-level 3D pose estimation that follows a two-step process: 1) A multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching. 2) Learning dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. The results showed that our proposed method achieves large improvements over all baselines, and we hope that our work will pave the ground for future advances in this important research direction. + +# 6. Acknowledgement + +Adam Kortylewski acknowledges support for his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No. 468670075. + +# References + +[1] Shuichi Akizuki and Manabu Hashimoto. Asm-net: Category-level pose and shape estimation using parametric deformation. In Proceedings of the British Machine Vision Conference, pages 1-13, 2021. 1 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 2 +[3] Kai Chen and Qi Dou. Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2773-2782, 2021. 1, 2 +[4] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 1 +[5] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 2, 4, 5, 6 +[6] Walter Goodwin, Ioannis Havoutis, and Ingmar Posner. You only look at one: Category-level object representations for pose estimation from a single example, 2023. 2, 6 +[7] Alexander Grabner, Peter M Roth, and Vincent Lepetit. 3d pose estimation and 3d model retrieval for objects in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3022-3031, 2018. 1, 2 +[8] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Inferring 3d object pose in rgb-d images. arXiv preprint arXiv:1502.04652, 2015. 1 +[9] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3303-3312, 2021. 1, 3 +[10] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[11] D Kirkpatrick and Raimund Seidel. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4):551-559, 1983. 1 +[12] Adam Kortylewski, Ju He, Qing Liu, and Alan L Yuille. Compositional convolutional neural networks: A deep architecture with innate robustness to partial occlusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8940-8949, 2020. 4 +[13] Georgios Kouros, Shubham Shrivastava, Cedric Picron, Sushruth Nagesh, Punarjay Chakravarty, and Tinne Tuyte-laars. Category-level pose retrieval with contrastive features learnt with occlusion augmentation. arXiv preprint arXiv:2208.06195, 2022. 2 +[14] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d + +pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 1, 2 +[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[16] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 1, 3 +[17] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 2 +[18] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5 +[19] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 1, 5 +[20] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3, 5 +[21] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE international conference on computer vision, pages 2686-2694, 2015. 2 +[22] Hung-Yu Tseng, Shalini De Mello, Jonathan Tremblay, Sifei Liu, Stan Birchfield, Ming-Hsuan Yang, and Jan Kautz. Few-shot viewpoint estimation. arXiv preprint arXiv:1905.04957, 2019. 2 +[23] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2 +[24] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 4 +[25] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12645-12654, 2020. 5 +[26] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 1, 2, 3 + +[27] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 2, 6, 7 +[28] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 1 +[29] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 1, 2 +[30] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 5 +[31] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 1, 2, 5 +[32] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. BMVC, 2019. 1, 2 +[33] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In 2021 International Conference on 3D Vision (3DV), pages 74–84. IEEE, 2021. 1, 2 +[34] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3090-3106, 2022. 2 +[35] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Transactions on Robotics, 37(2):314-333, 2020. 6 +[36] Yanjie Ze and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems, 35:27469-27483, 2022. 1, 3 +[37] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In The Eleventh International Conference on Learning Representations, 2022. 1, 2 +[38] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 1, 2, 5, 6, 7 +[39] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 5745-5753, 2019.8 \ No newline at end of file diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/images.zip b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..744fdfae62c4a2e3e9c75623bd059f42260906ab --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3e887d2810178a7880bd4516b6370c0dd6f638c6accf206cae6d5496549f196 +size 703889 diff --git a/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/layout.json b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbe734ca76fdcbec0e99a75b9b0737876fb08165 --- /dev/null +++ b/2024/Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos/layout.json @@ -0,0 +1,8847 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "text", + "content": "Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 192, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 192, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 192, + 156 + ], + "type": "text", + "content": "Leonhard Sommer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 143, + 292, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 143, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 220, + 143, + 292, + 156 + ], + "type": "text", + "content": "Artur Jesslen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 143, + 372, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 143, + 372, + 157 + ], + "spans": [ + { + "bbox": [ + 321, + 143, + 372, + 157 + ], + "type": "text", + "content": "Eddy Ilg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 400, + 143, + 499, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 143, + 499, + 157 + ], + "spans": [ + { + "bbox": [ + 400, + 143, + 499, + 157 + ], + "type": "text", + "content": "Adam Kortylewski" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 171, + 294, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 171, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 181, + 171, + 294, + 186 + ], + "type": "text", + "content": "1University of Freiburg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 171, + 408, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 171, + 408, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 307, + 171, + 408, + 186 + ], + "type": "text", + "content": "Saarland University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 190, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 190, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 190, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 238, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 525 + ], + "type": "text", + "content": "Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild on the Pascal3D+ and ObjectNet3D datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 549, + 127, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 127, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 127, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Category-level object pose estimation is a fundamentally important task in computer vision and robotics with a multitude of real-world applications, e.g. for training 3D generative models on real data and for robots that need to grasp and manipulate objects. However, defining and determining the pose of an object is a task that is far from easy. Current approaches achieve high performance, but they require large amounts of annotated training data to generalize successfully [1, 26, 28, 33, 38], or additional inputs during inference, such as CAD Models [7, 14], 3D Shapes [3, 32, 37] or RGB-D [8]. However, all of these are either time-consuming to obtain or not available in practice at all." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 393 + ], + "type": "text", + "content": "This motivates the development of methods for learning category-level 3D pose estimators in a fully unsupervised fashion. While doing so from images in the wild seems infeasible, object-centric video data [19] offers a more accessible alternative. Such videos can be easily captured using consumer-grade cameras and makes it possible to leverage coarse 3D reconstructions during training, providing a practical and cost-effective method for collecting data. Therefore, we propose the new task of learning a single-image category-level 3D pose estimator from casually captured object-centric videos without any human labels or other supervision. In practice, we leverage CO3D [19] as training data and show that our proposed model is able to generalize and predict accurate poses in the wild for Pascal3D+ [29] and ObjectNet3D [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 396, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 552 + ], + "type": "text", + "content": "We address the challenging task of learning category-level 3D pose in an unsupervised fashion from casually captured object-centric videos. In particular, we propose a two-step pipeline (Figure 1). The first step extracts DINOv2 [18] features from the images and reconstructs a coarse 3D mesh from the video with off-the-shelf methods [4, 11]. Building on this input, we introduce a novel 3D alignment procedure, where a key contribution is a novel 3D cyclical distance in terms of geometry and appearance that enables the robust alignment of shape reconstructions even under severe noise and variations in the object topology. As a result, we can align all objects from the object-centric training videos into a canonical coordinate frame without supervision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "content": "In a second step, we leverage the canonical poses and 3D meshes obtained from the first step to train a category-level neural mesh [9, 16, 26, 36] in an unsupervised manner. In particular, we represent objects using a prototypical 3D mesh with surface features to capture the geometry and neural appearance of an object category and train a neural network backbone to predict, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. Finally, the object 3D pose is solved using a pose fitting algorithm based on the estimated correspondence pairs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "We demonstrate that our method outperforms all baselines by a large margin at the unsupervised alignment of object-centric videos on the CO3D [19] dataset. Moreover," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22787" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 182, + 216 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 182, + 216 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 182, + 216 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 182, + 216 + ], + "type": "image", + "image_path": "c3cc4c7f81ffbb91fa36c268284df84f9a7641346611f9de0b7448095e243257.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": "Figure 1. Illustration of our approach for the unsupervised learning of category-level 3D pose. Our method starts from unaligned object-centric videos of an object category (left) and aligns these into a canonical coordinate frame in a self-supervised manner using a prototypical 3D mesh and self-supervised transformer features (center). Using the aligned videos, we train a neural network backbone to predict 2D-3D correspondences from a single image to enable 3D object pose estimation in the wild (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 71, + 375, + 216 + ], + "blocks": [ + { + "bbox": [ + 184, + 71, + 375, + 216 + ], + "lines": [ + { + "bbox": [ + 184, + 71, + 375, + 216 + ], + "spans": [ + { + "bbox": [ + 184, + 71, + 375, + 216 + ], + "type": "image", + "image_path": "b26b63ffc31e99dfec2a9179cb1ba8a3ad41c2546e2004f987ad9f6bdb4151c3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 543, + 215 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 543, + 215 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 543, + 215 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 543, + 215 + ], + "type": "image", + "image_path": "76a5e14bd63acda2f6f88be6ea9bf423816a19b80a84723817b31c86a58de509.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "content": "our model provides faithful and robust predictions in-the-wild on Pascal3D+ [29] and ObjectNet3D [31] despite being trained from raw object-centric videos only." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 342, + 134, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 134, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 134, + 354 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 363, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 530 + ], + "type": "text", + "content": "Supervised Category-Level 3D Pose Estimation. Traditional methods to determine object poses were to label keypoints in images and train supervised methods to predict them [23, 38], or to use only pose labels and directly predict them by casting the pose estimation problem as bin classification by discretizing the pose space [21]. More recent methods utilize 3D meshes of objects or object categories. NeMo [26] uses 3D meshes with neural features that are rendered-and-compared to feature maps from a CNN to obtain pose estimates. [13] predicts a single embedding vector per image and shows that superior performance can be obtained by simply retrieving the closest training sample. In contrast to the above methods, our approach is fully unsupervised." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 286, + 652 + ], + "type": "text", + "content": "Few-Shot and Zero-Shot Pose Prediction. [33] proposes to train a supervised pose estimator across many categories and shows that their approach can generalize well to similar but unseen objects. Many works implement zero-shot pose estimation by conditioning the model on the 3D shape of the unseen object [3, 32, 34, 37] or by leveraging renderings of CAD models [7, 14]. Other methods use few-shot learning [22] or zero-shot learning with DINOv2 [18]. In contrast to the above, our method does not require any annotated dataset, 3D shapes or CAD models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "Pose Alignment. The work from Goodwin et al. [5] (ZSP) is most close to the first step of our method, as it aligns the poses of two object-centric videos in a fully unsupervised fashion. Similar to our work, they use DINO [2] to obtain semantic correspondences. They perform first a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 291, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 518 + ], + "type": "text", + "content": "coarse alignment by matching one image from the source video to one of many images from the reference video. Then they leverage cyclical distances to select few promising correspondences in the two images, and finally leverage respective depth maps to align both images using least squares. Goodwin et al. extend their work in [6] (UCD+) to match many images from the source video to many of the reference video. By finding a consensus over these many to many alignments with a single transformation they demonstrate improved performance. The first step of our work is similar to these works by that it also leverages DINO features and cyclical distances. However, our work adds a geometric distance to perform the alignment directly in 3D and introduces weighted correspondences to enable the necessary robust regression of the SE3 transformations from noisy and inaccurate geometries, which leads to significant improvements in the alignment accuracy. Note also that these previous approaches used RGB-D inputs, while our method works on images directly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Surface Embeddings and Neural Mesh Models. Recent work uses known poses and approximate object geometries to learn features in 3D space to uniquely identify parts of objects. [17] first used known mesh templates of deformable objects to train a network that predicts surface embeddings from the images. NeMo [26] presents a generative model trained with contrastive learning. [27] presents an extension through replacing vertices by Gaussian ellipsoids and using volume rendering. Similar to our work, many recent works leverage pre-trained vision transformers DINO [2] and DINOv2 [18] to unproject image features onto depth maps [5, 6, 37]. In contrast to [17, 26, 27], our approach does not require any pose annotation, while also going beyond [5, 6] by enabling 3D pose estimation in the wild from a single image, and not requiring RGB-D images and CAD models as input [37]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22788" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 289, + 199 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 289, + 199 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 289, + 199 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 289, + 199 + ], + "type": "image", + "image_path": "4a078b93488844856d028f8741441aa9417a231f14280e0398d570866f231b33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "type": "text", + "content": "Figure 2. Illustration of a neural mesh. It consists of a 3D mesh with one or several neural features per vertex. These features encode patch-level features from a feature extractor. In our unsupervised alignment method, we capture the viewpoint-dependent features from DINOv2. For pose estimation, we learn a category-level neural mesh with viewpoint-invariant features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 292, + 104, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 104, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 104, + 304 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 469 + ], + "type": "text", + "content": "In this section, we describe our approach for learning category-level 3D pose estimation without supervision from object-centric videos. Our method proceeds in a two-step approach. First, we align object instances across videos in an unsupervised manner to bring them into a canonical reference frame (Section 3.2). Given the aligned videos, our model learns to establish dense correspondences between images and a reconstructed 3D template mesh by predicting a feature vector for each pixel in a 2D image that corresponds to a visible vertex in the template mesh (Section 3.3). Finally, we describe how our model can efficiently estimate object poses from in-the-wild data using the predicted correspondences via render-and-compare." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 211, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 211, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 211, + 487 + ], + "type": "text", + "content": "3.1. Meshes with Surface Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "text", + "content": "In both steps of our approach, the video alignment and the representation learning, we represent objects as neural meshes, i.e. meshes with surface features to capture the geometry and appearance of an object instance or category [9, 16, 26, 36]. In particular, the geometric representation is a triangular mesh, where we denote the set of vertices as " + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|V|}" + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "text", + "content": ". The appearance is represented by storing one or multiple appearance features " + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{F} = \\{\\{f_i^k\\in \\mathbb{R}^D\\}_{k = 1}^{|F|}\\}_{i = 1}^{|V|}" + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "text", + "content": " at each mesh vertex. Together, the geometry and appearance define a neural mesh as " + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "inline_equation", + "content": "S = \\{V,F\\}" + }, + { + "bbox": [ + 46, + 493, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 244, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 244, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 244, + 636 + ], + "type": "text", + "content": "3.2. Self-supervised Alignment of Objects" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Our goal is to align the camera poses of multiple object-centric videos into a common coordinate frame. To achieve this, we represent each object-centric video as a neural mesh with self-supervised surface features. In particular, we utilize off-the-shelf structure-from motion [20] to obtain a coarse object shape reconstruction for each video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "Note that the reconstructed shapes cover the whole object, as the object-centric videos move in a full circle around the object. We post-process the reconstructed point cloud to clean it and generate a watertight mesh, for which we provide details in the supplementary material. Subsequently, we project the reconstructed coarse meshes into the feature map " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\Psi(I)" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": " that is obtained from a self-supervised transformer backbone [18]. We collect from every video a set of feature vectors for every vertex " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": " that describe the appearance of a local patch (Figure 2). Thus the number of features per vertex depends on the number of images in which the vertex is visible. As feature extractor " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": ", we use a self-supervised vision transformer [18] which has shown emerging correspondence matching abilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "text", + "content": "Finding geometric and appearance correspondences. Given the mesh vertices of the source object instance V and the corresponding aggregated features F, we aim to align them to the reference counterparts " + }, + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{V}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 239, + 545, + 384 + ], + "type": "text", + "content": ". In practice, we select a reference video at random from the set of all available videos and align the remaining videos to the reference. More precisely, we aim to optimize the transformation T, which is composed of rotation, translation and scale, under which the transformed source vertices and corresponding features yield the minimal distance to the reference counterparts with respect to geometry as well as appearance. Formally, our optimization problem optimizes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 390, + 545, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 390, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 334, + 390, + 545, + 409 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathrm {T}} \\mathcal {D} (S, \\bar {S}, \\mathrm {T}) = \\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) + \\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}), \\tag {1}", + "image_path": "aac8740780a21ae99d629ca5a89f1596efbcdbbedd0afd8ec40a0fceb42ac7f4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "content": " is the similarity between two videos given a transformation " + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "content": ", which combines a geometric distance between the mesh geometries " + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{geo}}(\\mathrm{V},\\bar{\\mathrm{V}})" + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "content": " and an appearance distance " + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{app}}(S,\\bar{S})" + }, + { + "bbox": [ + 304, + 415, + 545, + 463 + ], + "type": "text", + "content": " between surface features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 463, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 498 + ], + "type": "text", + "content": "Assuming that the object instances shapes contain a negligible variance and no symmetries, a suitable geometric distance is the Chamfer Distance defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 504, + 545, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 504, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 345, + 504, + 545, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {g e o}} (\\mathrm {V}, \\bar {\\mathrm {V}}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} | | _ {2}, \\tag {2}", + "image_path": "a9c431c96190f8cdb97a412d320dfd7aebf6cd5ebcf1d4f6484a291e04ecf82a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\chi (\\mathrm{v}_i)" + }, + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "text", + "content": " is the vertex index of the Euclidean nearest neighbor of vertex " + }, + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 304, + 538, + 545, + 563 + ], + "type": "text", + "content": " in the respective other set of vertices" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 570, + 545, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 570, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 328, + 570, + 545, + 616 + ], + "type": "interline_equation", + "content": "\\chi \\left(\\mathrm {v} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m i n} | | \\mathrm {T v} _ {i} - \\mathrm {v} _ {j} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\mathrm {V} \\\\ j \\in 1 \\dots | \\bar {\\mathrm {V}} | \\\\ \\operatorname {a r g m i n} | | \\mathrm {T v} _ {j} - \\mathrm {v} _ {i} | | _ {2}, & \\text {f o r} \\mathrm {v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {3}", + "image_path": "e4a0c915e7f0efa734ac68f92c950da6ef0701a8b24155e56d022a256ac3c750.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 622, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 682 + ], + "type": "text", + "content": "However, as the 3D object is rather coarse and noisy, and the alignment can be ambiguous due to symmetries or shape differences among objects, we optimize each vertex to also be geometrically close to its nearest neighbor in feature space using the appearance distance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 342, + 689, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 342, + 689, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {a p p}} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i}, \\mathrm {f} _ {i})} | | _ {2}, \\tag {4}", + "image_path": "6702400e82f4ac9e59ab65262777a2d0336bcd6a99faf5d569656c2dda9fa63d.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22789" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": "with the nearest neighbor in feature space " + }, + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "inline_equation", + "content": "\\psi (\\mathbf{v})" + }, + { + "bbox": [ + 47, + 72, + 282, + 85 + ], + "type": "text", + "content": " defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 94, + 286, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 286, + 148 + ], + "type": "interline_equation", + "content": "\\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\underset {j \\in 1 \\dots | \\bar {\\mathrm {V}} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\mathrm {V} \\\\ \\underset {j \\in 1 \\dots | \\mathrm {V} |} {\\operatorname {a r g m i n}} \\min _ {k, l} \\left| \\left| \\mathrm {f} _ {j} ^ {k} - \\mathrm {f} _ {i} ^ {l} \\right| \\right|, & \\text {f o r v} _ {i} \\in \\bar {\\mathrm {V}}. \\end{array} \\right. \\tag {5}", + "image_path": "4cc24f9d15ffe73bf6489262f1f1aa012f74c498750981d3663b9d57378606b8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 149, + 286, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 286, + 185 + ], + "type": "text", + "content": "As the self-supervised vertex features are view-dependent, the appearance distance computes the minimum feature distance across all views to select the nearest neighbor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 365 + ], + "type": "text", + "content": "Weighting Correspondences. An open challenge for the alignment of casually captured object-centric videos is that the estimated correspondence pairs between videos can be unreliable. For example, due to errors in the shape reconstruction or significant topology changes among different object instances, such as one bicycle having support wheels whereas the other does not. In these cases the correspondences in the geometry and feature space are ill-defined which leads to unreliable correspondence estimates. To account for such unreliable correspondences, we introduce a weight factor for each correspondence pair that estimates its quality. At the core of the correspondence weighting, we introduce a 3D cyclical distance among the vertices of two neural meshes that is inspired by 2D cyclical distances [5] for correspondence estimation, and is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 374, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 374, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 52, + 374, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right) = \\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi \\left(\\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right)} \\right| \\right| _ {2}, \\text {w i t h} j = \\psi \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}\\right). \\tag {6}", + "image_path": "0633527d2812281b3edf984e057fe5783cbe24e197189c8e055a13be47f0f5c8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": "The nested structure of our 3D cyclical distance first computes " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " as the index of the nearest neighbor of vertex " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_i" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " in the feature space, and in turn computes the nearest neighbor of " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_j" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)}" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": ". Notably, " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{d}_{\\text{cycle}}(\\mathrm{v}_i, \\mathrm{f}_i) = 0" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " if the nearest neighbour maps back to the original vertex " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_{\\psi(\\mathrm{v}_j, \\mathrm{f}_j)} = \\mathrm{v}_i" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " and hence the correspondence is reliable. Building on this 3D cyclical distance, we define the validity criteria for each pair of vertices as the sum of cyclical distances of the correspondence pair" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 513, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 513, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 513, + 287, + 540 + ], + "type": "interline_equation", + "content": "\\rho \\left(\\mathrm {v} _ {i}, \\mathrm {f} _ {i}, \\mathrm {v} _ {j}, \\mathrm {f} _ {j}\\right) = - \\frac {\\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {i} , \\mathrm {f} _ {i}\\right) + \\mathrm {d} _ {\\text {c y c l e}} \\left(\\mathrm {v} _ {j} , \\mathrm {f} _ {j}\\right)}{2 \\tau (\\mathrm {D} (\\mathrm {V}) + \\mathrm {D} (\\bar {\\mathrm {V}}))}, \\tag {7}", + "image_path": "369d3132d94df9b01c78ecf9ae91faab048edbd53a5b32efbbfee33372a29c2a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(\\cdot)" + }, + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "type": "text", + "content": " is the diameter of a neural mesh given as " + }, + { + "bbox": [ + 47, + 548, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{D(V)} = \\max_{\\mathrm{v}_i,\\mathrm{v}_j\\in \\mathrm{V}}||\\mathrm{v}_i - \\mathrm{v}_j||_2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "text", + "content": "To obtain the final weight factor for a correspondence pair we use the softmax normalization " + }, + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\sigma_{\\tau}(i,j) = \\mathrm{Softmax}_{\\tau}(\\rho (\\mathrm{v}_i,\\mathrm{f}_i,\\mathrm{v}_j,\\mathrm{f}_j))" + }, + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "text", + "content": " across all feature and gemoetric correspondences. For the softmax normalization, we introduce the temperature " + }, + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 573, + 287, + 680 + ], + "type": "text", + "content": " , which enables us to steer between taking into account fewer high quality correspondences or more low quality ones (see Section 4.4). Together with the weighting we formulate the weighted geometric distance as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {g e o}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\chi (\\mathrm {v} _ {i})) \\| \\mathrm {v} _ {i} - \\mathrm {v} _ {\\chi (\\mathrm {v} _ {i})} \\| _ {2}, \\tag {8}", + "image_path": "198dbe4e9706b81c986c90148a18d98cfa703c4a010a3c91fff89729e4b82e5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 506, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 506, + 84 + ], + "type": "text", + "content": "and likewise the weighted appearance distance as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 95, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {a p p}} ^ {*} (S, \\bar {S}) = \\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} \\cup \\bar {\\mathrm {V}}} \\sigma (i, \\psi (\\mathrm {v} _ {i})) | | \\mathrm {v} _ {i} - \\mathrm {v} _ {\\psi (\\mathrm {v} _ {i})} | | _ {2}. \\tag {9}", + "image_path": "c8a9ed20d23e3064ea23d02c193ad3689e8eead83ec20ee39100665f9a6f1ca9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 132, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 132, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 545, + 156 + ], + "type": "text", + "content": "Our final distance measure to compare two neural meshes is computed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 166, + 545, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 166, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 166, + 545, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (S, \\bar {S}) = (1 - \\alpha) \\mathcal {D} _ {\\text {g e o}} ^ {*} (S, \\bar {S}) + \\alpha \\mathcal {D} _ {\\text {a p p}} ^ {*} (S, \\bar {S}). \\tag {10}", + "image_path": "58233680179a135b0da97497d1287c3b9f8d2376978c293d11bd299a788c54db.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 192, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 263 + ], + "type": "text", + "content": "To find an approximately optimal solution, we use a RANSAC strategy, where we randomly choose four vertices on the source surface mesh. Together with their nearest neighbors in the feature space on the reference surface mesh, we estimate a single transformation using the Umeyama method [24]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 274, + 477, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 477, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 477, + 285 + ], + "type": "text", + "content": "3.3. 3D Pose Estimation In-the-Wild" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": "Our goal is to perform 3D pose estimation in in-the-wild images. To achieve this, we generalize our approach from the multi-view setting used to align object-centric videos, towards 3D pose inference from a single image. Our model uses a feature extractor " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\Psi_w(I) = F \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " to obtain image features from input image " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " denotes the parameters of the backbone. The backbone output is a feature map " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " with feature vectors " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " at positions " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " on a 2D lattice. For training, we use the aligned object-centric videos (Section 3.2) to train the weights " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " of the feature extractor " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\Psi_w" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " such that it predicts dense correspondences between image pixels and the 3D neural mesh template. Specifically, we relate the features of an image " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\Psi_w(I)" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " extracted by a backbone feature extractor to the vertex and background features by Von-Mises-Fisher (vMF) probability distributions [12]. In particular, we model the likelihood of generating the feature at an image pixel " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " from corresponding vertex feature " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_r" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "P(f_i | f_r) = c_p(\\kappa) e^{\\kappa f_i \\cdot f_r}" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_r" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " is the mean of each vMF kernel, " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " is the corresponding concentration parameter, and " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "c_p" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " is the normalization constant (" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\|f_i\\| = 1" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\|f_r\\| = 1" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": "). We also model the likelihood of generating the feature " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " from background feature as " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "P(f_i | \\beta) = c_p(\\kappa) e^{\\kappa f_i \\cdot \\beta}" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 293, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": "When learning the models, as described next, we will learn the vertex features " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\{f_r\\}" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": ", the background feature " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": ", and the parameters " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": " of the neural network backbone " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Psi_w" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": ". We emphasize that our model requires that the backbone must be able to extract features that are invariant to the viewpoint of the object to ensure that " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "f_i \\cdot f_r" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": " is large irrespective of the viewpoint." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Learning viewpoint-invariant vertex features. For training our model, we use the visible vertex features and their corresponding image features " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(f_r,f_i)\\}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Further, we use image features randomly sampled from the background " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{f_i\\}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". As optimization objective, we use the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22790" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 123, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 123, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 123, + 84 + ], + "type": "text", + "content": "cross-entropy loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 92, + 287, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 169 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t r a i n}} = - \\sum_ {(f _ {r}, f _ {i}) \\in \\mathcal {P}} \\log \\left(\\frac {P (f _ {i} | f _ {r})}{\\sum P (f _ {i} | f _ {r}) + \\sum P (f _ {i} | \\beta)}\\right) \\\\ - \\sum_ {f _ {i} \\in \\mathcal {B}} \\log \\left(\\frac {P \\left(f _ {i} \\mid \\beta\\right)}{\\sum P \\left(f _ {i} \\mid f _ {r}\\right) + \\sum P \\left(f _ {i} \\mid \\beta\\right)}\\right). \\tag {11} \\\\ \\end{array}", + "image_path": "20d14f0d50edc497a36d94ac87a6406a5067570bfdb09e371fc4dbe8e6c05773.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": "3D pose inference. We use the mesh with the vertex features " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\{f_r\\}" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": ", the background feature " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " and the trained backbone " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\Psi_w" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " to estimate the camera pose " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " via render-and-compare. At each optimization step, we render a feature map " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\{\\bar{f}_i(\\alpha)\\}" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " under pose " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " and compare it with the encoder's feature map " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "F = \\{f_i\\}" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": ". Determined by the rendering, each feature map consists of foreground features " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "F_{front}" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": " and background features " + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "inline_equation", + "content": "F_{back} = F \\setminus F_{front}" + }, + { + "bbox": [ + 47, + 178, + 287, + 299 + ], + "type": "text", + "content": ". Thereupon, we maximize the joint likelihood for all image features under the assumption of independence, given as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 307, + 257, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 307, + 257, + 319 + ], + "spans": [ + { + "bbox": [ + 176, + 307, + 257, + 319 + ], + "type": "interline_equation", + "content": "P (F | \\alpha , \\{f _ {r} \\}, \\beta) =", + "image_path": "ec2391c816b016116f4c013294af71a0168af0321dbc41ba2dbc9d5b29bcac25.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 323, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\prod_ {f _ {i} \\in F _ {\\text {f r o n t}}} \\max _ {f ^ {\\prime} \\in \\{\\bar {f} _ {i} (\\alpha), \\beta \\}} P (f _ {i} | f ^ {\\prime}) \\prod_ {f _ {i} \\in F _ {\\text {b a c k}}} P (f _ {i} | \\beta). \\tag {12}", + "image_path": "baf3ef0dd2a8ef65c1086140f29ac079d48e433c0032beadc4b23842124422ea.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 287, + 380 + ], + "type": "text", + "content": "Note, by allowing foreground image features to be generated by the background feature, we also account for clutter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": "We estimate the pose by first finding the best initialization of the object pose " + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": " by computing the joint likelihood (Eq.12) for a set of pre-defined poses via template matching and choosing the one with the highest likelihood. Subsequently, we iteratively update our initial pose using a differentiable renderer to obtain the final pose prediction " + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 464, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 128, + 478 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "text", + "content": "In this section, we discuss our experimental setup (Section 4.1), present baselines and results for unsupervised alignment of object-centric videos (Section 4.2) and 3D pose estimation in-the-wild (Section 4.3). Additionally, we perform ablations of key model components in Section 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Dataset for alignment. To evaluate the unsupervised alignment of object-centric videos, we use the recently released Common Objects in 3D (CO3D) dataset [19] that provides images of multiple object categories, with a large amount of intra-category instance variation, and with varied object viewpoints. It contains 1.5 million frames, capturing objects from 50 categories, across nearly 19k scenes. For each object instance, CO3D provides approximately " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "100 - 200" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " frames promising a " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " viewpoint sweep with handheld cameras. CO3D supplements these videos with relative camera poses and estimated object point clouds using Structure-from-Motion [20]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": "We find that the unfiltered videos of CO3D are not ideal for our purpose. In particular, we find that videos with little viewpoint variation lead to inferior structure-from-motion results. Also, videos that are not focusing on the object's center in 3D or are taken too close to it, contain little information for correspondence learning. Therefore, we filter the videos accordingly, targeting 50 videos per category. For multiple categories we end up with less than 50 videos namely, \"remote\" 17, \"mouse\" 15, \"tv\" 16, \"toilet\" 7, \"toybus\" 41, \"hairdryer\" 28, \"couch\" 49, and \"cellphone\" 23. With our simple filters, we end up aiming for 50 videos per category. More precise details for the filtering procedure are appended in the supplementary. As labels, we use the ground truth pose annotations provided by ZSP [5], that cover ten object instances of twenty different categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 256, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 437 + ], + "type": "text", + "content": "Datasets for 3D pose estimation in-the-wild. We evaluate on two common datasets PASCAL3D+ [30] and ObjectNet3D [31]. While PASCAL3D+ provides poses for the 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers pose annotations for over 100 categories. The object-centric video dataset CO3D covers 50 categories from the MS-COCO [15] dataset. We find 23 common categories across ObjectNet3D and CO3D, even tolerating the gap between a toybus in CO3D and a real one in PASCAL3D+ and ObjectNet3D. We believe that this non-negligible gap could be bridged by exploiting the multiple viewpoint knowledge of the same object instance. Overall we validate on PASCAL3D+ with 6233 images, using the same validation set as [25], and on ObjectNet3D on 12039 images. Following [38], we center all objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": "Implementation details. In our alignment step, we use " + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\tau = 100" + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": ". Further, we leverage as self-supervised ViT the publicly available small version of DINOv2 [18] with 21M parameters and a patch size of 14. At the input we use a resolution " + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": " ending up with a " + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": " feature map, where each feature yields 384 dimensions. In our second step, we use the same ViT as backbone and freeze its parameters. Further, we add on top three ResNet blocks with an upsampling step preceding the final block. Ending up with a " + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 441, + 545, + 620 + ], + "type": "text", + "content": " feature map, where each feature has 128 dimensions. We optimize the cross-entropy loss for 10 epochs with Adam [10]. In one epoch, we make use of all filtered videos. The training for each category-level representation takes less than an hour on a single NVIDIA GeForce RTX 2080." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": "We note that the quality of our alignment method and the subsequent representation learning can vary depending on the chosen reference video. Therefore, we randomly choose five reference videos per category and report the mean performance and the standard deviation across all results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We report the " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " accuracy for pose estimation where the angle error for an estimated rotation " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{pred}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " and a ground" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22791" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 518, + 159 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 518, + 159 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 518, + 159 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 518, + 159 + ], + "type": "table", + "html": "
Per CategoryAll Categories (20)
backpackcarchairkeyboardlaptopmotorcycleAcc. 30°Acc. 15°
TEASER++ [35]1.05.09.08.06.01.03.81.1
ZSP [5]44.065.047.069.085.085.049.428.4
UCD+ [6]67.086.076.076.0100.0100.069.854.6
Ours85.6100.098.982.2100.0100.077.061.6
±10.5±0.0±3.5±10.7±0.0±0.0±11.9±15.9
", + "image_path": "e5c58287bdc8a714e9bbe5d09e99eabcc25f3d0fda9e577b1f32be809cdbc391.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "type": "text", + "content": "Table 1. Unsupervised alignment evaluation on the CO3D dataset across 20 categories. The reported metric is the " + }, + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "type": "inline_equation", + "content": "{30}^{ \\circ }" + }, + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "type": "text", + "content": " accuracy if not stated otherwise. The mean is computed across all 20 categories. We see that our method substantially outperforms the state of the art." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 200, + 548, + 373 + ], + "blocks": [ + { + "bbox": [ + 47, + 200, + 548, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 200, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 548, + 373 + ], + "type": "image", + "image_path": "6d9b2eb42ca8c0b84d3cc7360f142f2044d7c4583a56d72fe90d07e5338e6790.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 382, + 547, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 382, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 547, + 417 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of two unsupervised alignment methods. The first row shows the alignment of our proposed method. The second row shows the alignment using ZSP [5]. For both methods we use the 5th object instance from left as reference. We see that our proposed method is more accurate compared with ZSP. Especially for cars ZSP often confuses back and front." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 173, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 173, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 173, + 449 + ], + "type": "text", + "content": "truth rotation " + }, + { + "bbox": [ + 47, + 437, + 173, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{R}_{pred}" + }, + { + "bbox": [ + 47, + 437, + 173, + 449 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 454, + 287, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 287, + 482 + ], + "type": "interline_equation", + "content": "\\Delta \\left(\\mathrm {R} _ {p r e d}, \\mathrm {R} _ {g t}\\right) = \\operatorname {a r c c o s} \\left(\\frac {1}{2} \\operatorname {t r} \\left(\\mathrm {R} _ {p r e d} ^ {\\mathrm {T}} \\mathrm {R} _ {g t}\\right) - 1\\right). \\tag {13}", + "image_path": "624bbe34af1335bdb4e3396a3b9c3301921b72c770a39806d16ed95507062e51.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 485, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 287, + 510 + ], + "type": "text", + "content": "We note that for the current state of unsupervised pose estimation, achieving " + }, + { + "bbox": [ + 47, + 485, + 287, + 510 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 47, + 485, + 287, + 510 + ], + "type": "text", + "content": " precision remains unsolved." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 187, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 187, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 187, + 529 + ], + "type": "text", + "content": "4.2. Unsupervised Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "We follow the evaluation protocol of ZSP [5] and measure the alignment of one object instance to the nine remaining ones of the same category that are labelled. Additionally, we report the standard deviation across the chosen reference object instances. The quantitative results in Table 1 show that our proposed method significantly improves the state of the art by " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "69.8\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "77.0\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ". Our alignment algorithm can more efficiently use the video frames compared to ZSP [5], which only compares a single RGB-D frame from the source video with many RGB-D frames of the reference video. We note that ZSP uses DINOv1 features in contrast to our method, which uses DINOv2. Therefore, we provide an ablation of our method with respect to different feature extractors in the supplementary. One reason for our model to outperform " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{UCD + }" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " [6], an extension of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 521 + ], + "type": "text", + "content": "ZSP, is likely that our optimization does exploit the object geometry extensively, whereas others are using it only for refinement. A qualitative comparison of our method against ZSP is depicted in Figure 3. It shows that our alignments are highly accurate despite a large variability in the object instances. We note that at the time of writing, there is no source code publicly available to compare with UCD+." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 527, + 479, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 479, + 540 + ], + "type": "text", + "content": "4.3. In-the-Wild 3D Pose Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "As we are not aware of any unsupervised method learning pose estimation from videos, we compare our pose estimation method against two supervised methods [27, 38] and ZSP. We provide ZSP with ten uniformly-distributed images of the same reference video that our method uses. Further, we provide ZSP with depth annotations using the category-level CAD models and pose annotations in the PASCAL3D+ and ObjectNet3D data. Despite our method not requiring any depth information, it outperforms ZSP by a large-margin on both PASCAL3D+, see Table 2, and on ObjectNet3D, see Table 3. Qualitative results are depicted in Figure 4. We find that ZSP is highly compute intensive, requiring 10.92 seconds per sample on average, while our proposed method takes only 0.22 seconds on average." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22792" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 108, + 132 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 108, + 132 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 108, + 132 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 108, + 132 + ], + "type": "image", + "image_path": "da94dbf0317209e8752a94fc1fcc88d67ac3d5c63662b084491758c435eaeac2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 72, + 171, + 132 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 171, + 132 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 171, + 132 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 171, + 132 + ], + "type": "image", + "image_path": "6457e57a978f98377072015ef41987f4cb4fc49df2d67000ca281b67cc0ced99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 233, + 132 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 233, + 132 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 233, + 132 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 233, + 132 + ], + "type": "image", + "image_path": "a8c3a3f0f45fec3db1768b2994baa7856b8749db0d2f5bab36545766080f37dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 72, + 295, + 132 + ], + "blocks": [ + { + "bbox": [ + 235, + 72, + 295, + 132 + ], + "lines": [ + { + "bbox": [ + 235, + 72, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 235, + 72, + 295, + 132 + ], + "type": "image", + "image_path": "2ec5bf4e05badb3942a151afc3e8cf0a40bfff810abaa68bd837b9474b324e85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 358, + 132 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 358, + 132 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 358, + 132 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 358, + 132 + ], + "type": "image", + "image_path": "54fed3bdcc5db28eafe45010dac46e097d8a8a5e26f483334a4ba7ecbae85fe4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 419, + 132 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 419, + 132 + ], + "type": "image", + "image_path": "5eb96cf3df0035265a7b06577bf0907bc09c7b7d1c91aa6bad36061349072c32.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 422, + 72, + 482, + 132 + ], + "blocks": [ + { + "bbox": [ + 422, + 72, + 482, + 132 + ], + "lines": [ + { + "bbox": [ + 422, + 72, + 482, + 132 + ], + "spans": [ + { + "bbox": [ + 422, + 72, + 482, + 132 + ], + "type": "image", + "image_path": "c537b59d6b78307d6a65921861b9b4e1d5a23faa4d6703ad5d67988e4fd4aa8b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 544, + 132 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 544, + 132 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 544, + 132 + ], + "type": "image", + "image_path": "e9527fe4e4ec2cc3237ee078b1c97bda3513a063ae11fa65563b6c45582aaa0a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 49, + 133, + 108, + 193 + ], + "blocks": [ + { + "bbox": [ + 49, + 133, + 108, + 193 + ], + "lines": [ + { + "bbox": [ + 49, + 133, + 108, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 133, + 108, + 193 + ], + "type": "image", + "image_path": "127efd52c9091f7ebc18af2986ffa1b9ea175b3c7d7c24f06e732db1be4285da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 111, + 133, + 171, + 193 + ], + "blocks": [ + { + "bbox": [ + 111, + 133, + 171, + 193 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 171, + 193 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 171, + 193 + ], + "type": "image", + "image_path": "a9473087bb87c53d579268e71bc74e771cc0905785e23175c66556a3f28e2310.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 173, + 133, + 233, + 193 + ], + "blocks": [ + { + "bbox": [ + 173, + 133, + 233, + 193 + ], + "lines": [ + { + "bbox": [ + 173, + 133, + 233, + 193 + ], + "spans": [ + { + "bbox": [ + 173, + 133, + 233, + 193 + ], + "type": "image", + "image_path": "46b735850508d56f7733296b6bfaaf466ba547bc9ea2359ec5b058ee195c8e1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 133, + 295, + 193 + ], + "blocks": [ + { + "bbox": [ + 235, + 133, + 295, + 193 + ], + "lines": [ + { + "bbox": [ + 235, + 133, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 235, + 133, + 295, + 193 + ], + "type": "image", + "image_path": "e0a426b2e4e90b6e11cdd70aafcc13477ba36f70bc8e9ce2c299d9c1bb78108c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 133, + 358, + 193 + ], + "blocks": [ + { + "bbox": [ + 298, + 133, + 358, + 193 + ], + "lines": [ + { + "bbox": [ + 298, + 133, + 358, + 193 + ], + "spans": [ + { + "bbox": [ + 298, + 133, + 358, + 193 + ], + "type": "image", + "image_path": "997ca964013342ec44bdfcd0546d7ee3b80bf14e6052940ce76a51095ef06af6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 359, + 133, + 419, + 193 + ], + "blocks": [ + { + "bbox": [ + 359, + 133, + 419, + 193 + ], + "lines": [ + { + "bbox": [ + 359, + 133, + 419, + 193 + ], + "spans": [ + { + "bbox": [ + 359, + 133, + 419, + 193 + ], + "type": "image", + "image_path": "6888b98e86cb938403a550fab636e2233b2ac229ae7496014b47d048017220ad.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 422, + 133, + 482, + 193 + ], + "blocks": [ + { + "bbox": [ + 422, + 133, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 422, + 133, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 422, + 133, + 482, + 193 + ], + "type": "image", + "image_path": "670c4c8f0b4edba2d62abba3e89e44710ad3a385bab939c95bb05e6437e267b0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 484, + 133, + 544, + 193 + ], + "blocks": [ + { + "bbox": [ + 484, + 133, + 544, + 193 + ], + "lines": [ + { + "bbox": [ + 484, + 133, + 544, + 193 + ], + "spans": [ + { + "bbox": [ + 484, + 133, + 544, + 193 + ], + "type": "image", + "image_path": "d5c65a8be0ffa9d8772b821d4a3d28511c9f3665652227e71d9c0c1cbd112f97.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 49, + 194, + 108, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 194, + 108, + 255 + ], + "lines": [ + { + "bbox": [ + 49, + 194, + 108, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 108, + 255 + ], + "type": "image", + "image_path": "ab1e5de0bca90902f74bc020686352642d8106d84d663e1b9805c78412d7f252.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of our method (top) and ZSP (bottom) at category-level 3D pose prediction in the wild on samples from PASCAL3D+ and ObjectNet3D (we randomly selected the samples to demonstrate the diversity of the results). For both methods, we overlay our coarse mesh reconstruction in the predicted 3D pose." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 111, + 194, + 171, + 255 + ], + "blocks": [ + { + "bbox": [ + 111, + 194, + 171, + 255 + ], + "lines": [ + { + "bbox": [ + 111, + 194, + 171, + 255 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 171, + 255 + ], + "type": "image", + "image_path": "a10c90cd36249a4e9a7a992535bb4713c65ef7f6ed84a94504f1a9186bd9a7aa.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 173, + 194, + 233, + 255 + ], + "blocks": [ + { + "bbox": [ + 173, + 194, + 233, + 255 + ], + "lines": [ + { + "bbox": [ + 173, + 194, + 233, + 255 + ], + "spans": [ + { + "bbox": [ + 173, + 194, + 233, + 255 + ], + "type": "image", + "image_path": "b7a8d97c76e04fa8089c1a862f360ec37852aa673bd4fade7d65fd17890804ec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 235, + 194, + 295, + 255 + ], + "blocks": [ + { + "bbox": [ + 235, + 194, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 235, + 194, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 235, + 194, + 295, + 255 + ], + "type": "image", + "image_path": "19efa5764ce397c3085c7101a5ef66b73321314bc974a794c0e2e8d939aff650.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 298, + 194, + 358, + 255 + ], + "blocks": [ + { + "bbox": [ + 298, + 194, + 358, + 255 + ], + "lines": [ + { + "bbox": [ + 298, + 194, + 358, + 255 + ], + "spans": [ + { + "bbox": [ + 298, + 194, + 358, + 255 + ], + "type": "image", + "image_path": "498bf511aeaca7479cc82d2cecca00a3c6a87a17f0f045ec273d1cb68e731e1d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 359, + 194, + 419, + 255 + ], + "blocks": [ + { + "bbox": [ + 359, + 194, + 419, + 255 + ], + "lines": [ + { + "bbox": [ + 359, + 194, + 419, + 255 + ], + "spans": [ + { + "bbox": [ + 359, + 194, + 419, + 255 + ], + "type": "image", + "image_path": "438bc6e48d6653ffd70f1eb7d5dfe85f370c72142ede424f026d1fbc9f759f7b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 422, + 194, + 482, + 255 + ], + "blocks": [ + { + "bbox": [ + 422, + 194, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 422, + 194, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 422, + 194, + 482, + 255 + ], + "type": "image", + "image_path": "89be11eb1f2fb3c63e8f94719a37d212b0cb96a1b58fd2a077c60bf7fda1cd15.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 484, + 194, + 544, + 255 + ], + "blocks": [ + { + "bbox": [ + 484, + 194, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 484, + 194, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 484, + 194, + 544, + 255 + ], + "type": "image", + "image_path": "b264be3f4c36326eae96aec7ba14ec3cb654bd9c4aa69bfed92eccabe6c10270.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 61, + 308, + 533, + 401 + ], + "blocks": [ + { + "bbox": [ + 61, + 308, + 533, + 401 + ], + "lines": [ + { + "bbox": [ + 61, + 308, + 533, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 308, + 533, + 401 + ], + "type": "table", + "html": "
MethodbicyclebuscarchairmotorcyclecouchtvMean
SupervisedStarMap [38]83.294.490.075.468.879.885.882.49
VoGE [27]82.698.199.090.587.594.983.990.93
UnsupervisedZSP61.721.461.642.643.152.939.046.0
±14.7±8.1±11.4±11.3±22.1±16.0±36.2±17.1
Ours58.479.398.251.967.076.653.169.2
±5.0±11.8±1.0±10.5±8.9±13.7±20.4±10.2
", + "image_path": "4c203ba66501ea395bd32eb54a69b5ff40a223cbfc73d4f88aefe40914261cc6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 409, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 409, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 546, + 444 + ], + "type": "text", + "content": "Table 2. 3D Pose Estimation in-the-wild on 7 categories of PASCAL3D+. Top two rows show supervised methods (as upper bound) while bottom two rows show unsupervised methods. The reported metric is " + }, + { + "bbox": [ + 46, + 409, + 546, + 444 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 409, + 546, + 444 + ], + "type": "text", + "content": " accuracy. The mean is averaged over all 7 categories. Our method shows superior performance over ZSP, which requires depth annotations." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "type": "text", + "content": "Categorical discussion. We observe that our method performs better for categories with only small topology changes and deformations, (e.g. car, microwave, couch) compared to categories with large intra-class variability (e.g. chair). Further, we recognize, that our method even generalizes well from a toybus to a real bus. Besides that, we analyze, that categories with less available videos (e.g. remote, TV, toilet) on average achieve lower performance." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 573, + 111, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 111, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 111, + 585 + ], + "type": "text", + "content": "4.4. Ablation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Unsupervised alignment. Using the ground-truth annotations of our five references, we measure the effect of both parameters introduced in the alignment method. Namely, the appearance distance weight " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and the cyclical distance temperature " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ". We remark that for the distance between two meshes with surface features, the appearance weight trades-off feature correspondences versus Euclidean correspondences. Where an appearance weight of " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " means that the distance depends solely on the Euclidean correspondences. Contrarily, an appearance weight of " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " results" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 306, + 460, + 545, + 580 + ], + "blocks": [ + { + "bbox": [ + 306, + 460, + 545, + 580 + ], + "lines": [ + { + "bbox": [ + 306, + 460, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 545, + 580 + ], + "type": "image", + "image_path": "d32767268aec32c4ff5b68f848e74cd0d77d05f3f69318f37292d52f03e0fd34.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "lines": [ + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": "Figure 5. We report the " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": " accuracy of our alignment method for different choices of our appearance weight " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": " and our cyclical distance temperature " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": " resulting in different distances between two meshes with surface features. We see that the maximum accuracy of " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "75.2\\%" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": " is reached for " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "inline_equation", + "content": "\\tau = 100" + }, + { + "bbox": [ + 304, + 588, + 546, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "in solely depending on feature correspondences. Further, our cyclical distance temperature weights each correspondence, implicitly trading-off many low-quality correspondences versus few high-quality ones. Intuitively, increas" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22793" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 69, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 541, + 194 + ], + "type": "table", + "html": "
phonem'waveb'packbenchcuph'dryerlaptopmouseremotetoaster
ZSP46.450.523.150.833.021.760.528.841.628.8
± 8.4± 44.3± 14.9± 17.0± 18.7± 13.3± 8.9± 10.1± 18.1± 8.16
Ours54.680.318.062.138.214.153.344.754.460.6
± 3.4± 21.7± 12.0± 7.1± 21.6± 5.8± 9.7± 9.9± 4.4± 2.2
toiletb'cyclebuscarchaircouchk'boardm'cyclesuitcasetvMean
ZSP56.358.630.560.336.855.546.850.325.837.942.2
± 13.9± 10.4± 10.6± 9.1± 10.6± 16.1± 14.5± 20.7± 14.3± 35.1± 15.9
Ours39.657.878.398.152.276.626.969.015.553.252.4
± 12.7± 5.1± 12.1± 0.9± 9.6± 12.2± 8.4± 9.1± 5.5± 22.0± 9.8
", + "image_path": "663ad3c88364403685d0f475ff6fdd5895dfbe0f1a24d521a83f9a0479716636.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 243, + 279, + 308 + ], + "blocks": [ + { + "bbox": [ + 46, + 202, + 546, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 546, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 546, + 225 + ], + "type": "text", + "content": "Table 3. 3D Pose Evaluation on 10 categories of ObjectNet3D. The mean is averaged over 20 categories. Metric is the " + }, + { + "bbox": [ + 46, + 202, + 546, + 225 + ], + "type": "inline_equation", + "content": "{30}^{ \\circ }" + }, + { + "bbox": [ + 46, + 202, + 546, + 225 + ], + "type": "text", + "content": " accuracy. Despite not requiring any depth information, our method significantly outperforms ZSP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 243, + 279, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 243, + 279, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 279, + 308 + ], + "type": "table", + "html": "
MethodAcc. 30°Acc. 15°
PASCAL3D+Regression66.025.8
Ours69.241.3
ObjectNet3DRegression44.516.4
Ours52.425.5
", + "image_path": "b49b5f4427e2534dd459a8ad828b8a837f5d3de473da26bd7766da5949160360.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "type": "text", + "content": "Table 4. Average " + }, + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 46, + 315, + 288, + 339 + ], + "type": "text", + "content": " accuracies on PASCAL3D+ and ObjectNet3D for using directly neural network regression." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": "ing the value of " + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": " results in averaging over more correspondences, while decreasing " + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": " results in taking only the correspondences with high validity into account. In Figure 5, we see that both parameters yield a significant impact on the " + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": " accuracy. With an optimum for " + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\tau = 100" + }, + { + "bbox": [ + 46, + 360, + 287, + 504 + ], + "type": "text", + "content": ". Intuitively, this means that taking many correspondences into account is more beneficial. Additionally, the Euclidean correspondences are weighted four times as much as the feature correspondences. Besides that, the ablation shows that while many correspondences are essential for using solely Euclidean correspondences, the opposite is true when using solely feature correspondences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 505, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 578 + ], + "type": "text", + "content": "3D pose estimation in-the-wild. Following the alignment, the in-the-wild 3D pose estimation task can also be solved using neural network regression with the 6D rotation representation proposed in [39]. However, we observe that the results are worse than our 3D template learning method combined with render-and-compare, see Table 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We have proposed a model which substantially outperforms existing applicable baselines for the task of unsupervised category-level 3D pose estimation in-the-wild. However, our proposed method does not yet reach the performance of fully supervised baselines. One advancement we aspire is to relax the rigidity constraint of our shape model. Therefore, we plan to leverage the aligned reconstructions and introduce a parameterized model for the shape. A deformable shape would yield the potential to improve the correspond-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 245, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 319 + ], + "type": "text", + "content": "dence learning as well as the subsequent matching of features at inference. Moreover, we see a future research direction in enabling the model to learn from a continuous stream of data, instead of building on a set of pre-recorded videos. This would even better reflect the complex real-world scenarios of embodied agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 337, + 379, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 379, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 379, + 350 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 635 + ], + "type": "text", + "content": "In this paper, we have proposed a highly challenging (but realistic) task: unsupervised category-level 3D pose estimation from object-centric videos. In our proposed task, a model is required to align object-centric videos of instances of an object category without having any pose-labelled data. Subsequently, the model learns a 3D representation from the aligned videos to perform 3D category-level pose estimation in the wild. Our task defines a complex real-world problem which requires both semantic and geometric understanding of objects, and we demonstrate that existing baselines cannot solve the task. We further proposed a novel method for unsupervised learning of category-level 3D pose estimation that follows a two-step process: 1) A multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching. 2) Learning dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. The results showed that our proposed method achieves large improvements over all baselines, and we hope that our work will pave the ground for future advances in this important research direction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 654, + 416, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 416, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 416, + 668 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Adam Kortylewski acknowledges support for his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No. 468670075." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22794" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Shuichi Akizuki and Manabu Hashimoto. Asm-net: Category-level pose and shape estimation using parametric deformation. In Proceedings of the British Machine Vision Conference, pages 1-13, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "text", + "content": "[3] Kai Chen and Qi Dou. Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2773-2782, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "text", + "content": "[4] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 362 + ], + "type": "text", + "content": "[6] Walter Goodwin, Ioannis Havoutis, and Ingmar Posner. You only look at one: Category-level object representations for pose estimation from a single example, 2023. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "text", + "content": "[7] Alexander Grabner, Peter M Roth, and Vincent Lepetit. 3d pose estimation and 3d model retrieval for objects in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3022-3031, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 441 + ], + "type": "text", + "content": "[8] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Inferring 3d object pose in rgb-d images. arXiv preprint arXiv:1502.04652, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 497 + ], + "type": "text", + "content": "[9] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3303-3312, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "text", + "content": "[10] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "text", + "content": "[11] D Kirkpatrick and Raimund Seidel. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4):551-559, 1983. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "type": "text", + "content": "[12] Adam Kortylewski, Ju He, Qing Liu, and Alan L Yuille. Compositional convolutional neural networks: A deep architecture with innate robustness to partial occlusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8940-8949, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[13] Georgios Kouros, Shubham Shrivastava, Cedric Picron, Sushruth Nagesh, Punarjay Chakravarty, and Tinne Tuyte-laars. Category-level pose retrieval with contrastive features learnt with occlusion augmentation. arXiv preprint arXiv:2208.06195, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[14] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[16] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[17] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "text", + "content": "[18] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 306, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 306, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 306, + 545, + 371 + ], + "type": "text", + "content": "[19] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 1, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "type": "text", + "content": "[20] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[21] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE international conference on computer vision, pages 2686-2694, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "text", + "content": "[22] Hung-Yu Tseng, Shalini De Mello, Jonathan Tremblay, Sifei Liu, Stan Birchfield, Ming-Hsuan Yang, and Jan Kautz. Few-shot viewpoint estimation. arXiv preprint arXiv:1905.04957, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 536, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 545, + 569 + ], + "type": "text", + "content": "[23] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[24] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[25] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12645-12654, 2020. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 1, 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22795" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. In The Eleventh International Conference on Learning Representations, 2022. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "text", + "content": "[28] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[29] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[30] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "text", + "content": "[31] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "type": "text", + "content": "[32] Yang Xiao, Xuchong Qiu, Pierre-Alain Langlois, Mathieu Aubry, and Renaud Marlet. Pose from shape: Deep pose estimation for arbitrary 3d objects. BMVC, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[33] Yang Xiao, Yuming Du, and Renaud Marlet. Posecontrast: Class-agnostic object viewpoint estimation in the wild with pose-aware contrastive learning. In 2021 International Conference on 3D Vision (3DV), pages 74–84. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[34] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3090-3106, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 532 + ], + "type": "text", + "content": "[35] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Transactions on Robotics, 37(2):314-333, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[36] Yanjie Ze and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Neural Information Processing Systems, 35:27469-27483, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[37] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. In The Eleventh International Conference on Learning Representations, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[38] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[39] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 5745-5753, 2019.8" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22796" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_content_list.json b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebdc5183f66a50e4af9e74778aa626ea9ca9e871 --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "Unsupervised Occupancy Learning from Sparse Point Cloud", + "text_level": 1, + "bbox": [ + 179, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amine Ousafi Adnane Boukhayma \nInria, Univ. Rennes, CNRS, IRISA, M2S, France", + "bbox": [ + 287, + 181, + 676, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from 3D point clouds in the absence of ground truth supervision remains a very challenging task. In this paper, we propose a method to infer occupancy fields instead of SDFs as they are easier to learn from sparse inputs. We leverage a margin-based uncertainty measure to differentiably sample from the decision boundary of the occupancy function and supervise the sampled boundary points using the input point cloud. We further stabilise the optimization process at the early stages of the training by biasing the occupancy function towards minimal entropy fields while maximizing its entropy at the input point cloud. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve implicit shape inference with respect to baselines and the state-of-the-art using synthetic and real data.", + "bbox": [ + 75, + 282, + 472, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing full 3D shape from limited and corrupted data is a long standing problem. In this regard, one of the data type instances that has received increasing attention and investigation from computer vision, graphics and machine learning is point clouds. This interest emanates primarily from the ubiquity of this light, although topologically incomplete 3D representation, either as acquired e.g. through industrial and commodity depth sensors, or as an intermediate representation within computational photogrammetry [66, 67] pipelines for instance. While classical optimization methods such as Poisson Reconstruction [36] or Moving Least Squares [30] can be effective with dense, clean point sets and accurate normal pre-estimations, recent deep learning-based alternatives provide more robust predictions, particularly for noisy and sparse inputs, eliminating the need for", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "normal data in many cases.", + "bbox": [ + 500, + 253, + 681, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several methods rely on priors learned from large fully labeled data such as the synthetic dataset ShapeNet [14]. However, this strategy entails computationally expensive trainings, and the resulting models can still be prone to out-of-distribution generalization issues, as pointed in [15, 55], whether caused by change in the input size or domain shift. For instance, Table 2 shows that our unsupervised method outperforms supervised generalizable models when testing on data that is sparser and different in nature from their training corpus. Hence, it is important to design learning frameworks that can lead to robust reconstruction under such extreme constraints.", + "bbox": [ + 496, + 271, + 893, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural implicit representations (INR) were established recently as a powerful representation for 3D shape, generally taking the form of coordinate based MLPs. They have been applied to model various shape presentations, the most popular ones of which remain arguably signed distance functions (SDF) [58] and binary occupancy fields [52] for watertight shapes. For the task of point cloud reconstruction, there is a wide literature on learning SDFs from dense labels, as well as from point clouds solely (unsupervised), either in the generalizable setting or in instance specific optimization. However, apart from its applications in the supervised generalizable setting [8, 20, 60], it seems that occupancy learning from point sets has not been as widely considered in other scenarios. In particular, to the best of our knowledge, using occupancy fields to learn shapes from sparse noisy unoriented point clouds (i.e. unsupervisedly) has not been explored thoroughly. This is even the more intriguing as it seems to be easier to learn a binary classification task as opposed to regressing a continuous field that must additionally respect specific properties [28].", + "bbox": [ + 496, + 457, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concordantly, we introduce occupancy in this work for learning shape from a sparse point cloud. While other representations such as SDF can readily incorporate surface sample supervision directly (e.g. [28]), among other strategies, it is not entirely straightforward to define such supervision for occupancy fields. Hence, we propose a novel strategy for our training that combines a loss applied near the boundary decision, harnessing the input point samples, and an accompanying regularization. For the first loss, we encourage", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21729", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the decision boundary of our occupancy to align with the point cloud. We achieve this by supervising an uncertainty sampling mechanism for our occupancy field with the point cloud samples. We approximate this sampling through root-finding of a margin function associated to the field. As regularization, we propose to minimize the uncertainty of our field, while maximizing it near the decision boundary, via the entropy of our occupancy.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments under several real and synthetic benchmarks for object, non-rigid and scene level shape reconstruction, our results show that our method outperforms the existing state-of-the-art literature in reconstruction from sparse point cloud, using standard metrics and through superior visual results. Our ablation studies validate our design choices and showcase the importance of the components of the supervision scheme that we propose.", + "bbox": [ + 75, + 212, + 470, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 345, + 218, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shape Representations in Deep Learning Shapes can be depicted in deep learning through either intrinsic or extrinsic representations. Intrinsic representations focus on discriminating the shape itself. When explicitly implemented, for instance, using tetrahedral or polygonal meshes [35, 79] or point clouds [25], the output topology is predefined, thereby restricting the range of shapes that can be generated. Among other intrinsic representations, 2D patches [23, 29, 83] may introduce discontinuities, while the simplicity of shape primitives like cuboids [78, 93], planes [43], and Gaussians [26] constrains their expressiveness. On the other hand, extrinsic shape representations model the entire space containing the scene or object of interest. Voxel grids [86, 87] are the most popular, serving as a direct extension of 2D pixels to the 3D domain. However, their capacity is constrained by the memory cost associated with cubic resolution. Sparse representations, such as octrees [64, 76, 81], can mitigate this issue to some extent.", + "bbox": [ + 75, + 371, + 468, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Neural Shape Representations Implicit Neural Shape Representations have recently emerged as a significant approach for modeling extrinsic shape and radiance fields (e.g. [13, 33, 53, 80, 88]). These representations address many of the limitations associated with classical representations by showcasing the ability to represent shapes with arbitrary topologies at virtually infinite resolution. Typically, they are parameterized with MLPs, which map spatial locations or features to properties such as occupancy [52], signed [58], or unsigned [21, 91] distances relative to the target shape. The level-set derived from these MLPs can be visualized through techniques like ray marching [31] or tessellated into an explicit shape using methods like Marching Cubes [47]. Another notable line of research involves the development of hybrid implicit/explicit representations [19, 22, 57, 89], primarily based on differentiable space partitioning. To simultaneously represent", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "collections of shapes, implicit neural models necessitate conditioning mechanisms. These mechanisms encompass features and latent code concatenation, batch normalization, hypernetworks [17, 70, 72, 73, 82], and gradient-based meta-learning [54, 71]. Concatenation-based conditioning was initially implemented using single global latent codes [18, 52, 58] and has been further refined with the incorporation of local features [20, 24, 27, 34, 39, 60, 74, 77].", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction from Point Clouds Classical approaches to reconstruction include combinatorical methods where the shape is defined based on input point clouds through space partitioning, employing techniques such as alpha shapes [6], Voronoi diagrams [1], or triangulation [11, 45, 63]. Alternatively, the input samples can contribute to defining an implicit function, with its zero level set representing the target shape. This is achieved through global smoothing priors [40, 84, 85], such as radial basis functions [10] and Gaussian kernel fitting [65], or local smoothing priors like moving least squares [30, 37, 46, 51]. Another approach involves solving a boundary-conditioned Poisson equation [36]. Recent literature suggests parameterizing these implicit functions with deep neural networks and learning their parameters through gradient descent, either in a supervised or unsupervised manner.", + "bbox": [ + 496, + 214, + 892, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Implicit Neural Reconstruction In supervised methods, there is an assumption of having labeled training data, typically in the form of dense samples containing ground truth shape information. Auto-decoding techniques [12, 34, 39, 58, 77] necessitate test-time optimization to adapt to a new point cloud. On the other hand, encoder-decoder methods allow for swift feed-forward inference. Initially introduced for this purpose, Pooling-based set encoders [18, 27, 52] like PointNet [62] have been found to underfit the context. State-of-the-art performance is achieved by convolutional encoders, utilizing local features defined either in explicit volumes and planes [20, 41, 60, 61] or solely at the input points [8, 56]. Peng et al. [61] proposed a differentiable Poisson solving layer that efficiently converts predicted normals into an indicator function grid. However, its applicability is limited to small scenes due to the cubic memory requirement in grid resolution. The work in [32, 85] suggests constructing generalizable models with implicit decoder functions as a kernel regression. Despite these advancements, many of the generalizable methods still face challenges related to generalization.", + "bbox": [ + 496, + 459, + 892, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Implicit Neural Reconstruction For unsupervised approaches, a neural network is typically fitted to the input point cloud without additional information. Convergence improvements can be achieved through regularization techniques, such as the spatial gradient constraint based on the Eikonal equation introduced by Gropp et al. [28], a spatial divergence constraint as described in [5], and", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21730", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lipschitz regularization on the network [44]. Periodic activations were introduced in [72]. Lipman [42] learns a function that converges to occupancy, while its log transform converges to a distance function. Atzmon et al. [2] learn an SDF from unsigned distances, further supervising the spatial gradient of the function with normals [3]. Ma et al. [48] express the nearest point on the surface as a function of the neural signed distance and its gradient. They also utilize self-supervised local priors to handle very sparse inputs [49] and enhance generalization [50]. [38] guides the implicit field learning with an Octree based labelling. [9] predicts occupancy fields by learning whether a dropped needle goes across the surface or no. [15] learns a surface parameterization leveraged to provide additional coarse surface supervision to the shape network. In [84], infinitely wide shallow MLPs are learned as random feature kernels using points and their normals. However, most of the methods mentioned above encounter challenges when dealing with sparse and noisy input, primarily due to the lack of supervision. Differently from this literature, we propose here to apply occupancy to unsupervised sparse point cloud neural fitting for the first time to the best of our knowledge.", + "bbox": [ + 75, + 90, + 472, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 436, + 171, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sparse noisy unoriented input point cloud $\\mathcal{P} \\subset \\mathbb{R}^{3 \\times N}$ , our goal is to recover a 3D watertight shape surface $S$ that best explains this observation, i.e. points from $\\mathcal{P}$ approximating noisy samples from $S$ .", + "bbox": [ + 75, + 462, + 468, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the proven success of implicit neural representations, we propose to address this problem by fitting an implicit neural shape function to the scarce observation, which amounts to learning an MLP mapping coordinates to shape attributes. In this context, dubbed unsupervised, signed distance fields are the representation of choice in the community so far (e.g. [15, 16, 38, 48, 61, 84]). Oddly enough, binary occupancy fields were not explored extensively yet as a representation for this context to the best of our knowledge, even though in theory, it seems easier to learn occupancy than SDFs. In fact, it is safe to assume that binary classification is an easier task than continuous regression in general. Besides, we can see intuitively that occupancy is only the sign part of the SDF information. Additionally, occupancy fields do not seem to stipulate any strong structural requirement, as it is the case for distance fields, whose gradient ought to satisfy the Eikonal constraint [28] for instance. Hence, we propose here to learn a binary occupancy field.", + "bbox": [ + 75, + 523, + 472, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When labeled samples inside and outside the shape are available, such a neural occupancy field could be learned in this supervised setting via a standard cross-entropy classification loss (e.g. [52]). However when only surface points are available, as in our case $(\\mathcal{P})$ , it is not entirely clear how a successful learning can be achieved using such occupancy", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decision boundary samples alone. We derive in the following a methodology to approach this task.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning Occupancy Through Margin Uncertainty Sampling", + "text_level": 1, + "bbox": [ + 498, + 128, + 890, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are interested in learning the following posterior $P(y|\\mathbf{x},\\theta)$ , where $\\mathbf{x}$ represents query points in $\\mathbb{R}^3$ , and label $y \\in \\{0,1\\}$ stands for binary shape occupancy. $P$ is parameterized with a Softmax activated MLP $\\theta$ similarly to [8]. Since we do not have points with their occupancy labels, all we can learn with is the input point cloud samples.", + "bbox": [ + 498, + 167, + 893, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us consider the following margin function:", + "bbox": [ + 517, + 258, + 828, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {\\theta} (\\mathbf {x}) = P (y = 1 | \\mathbf {x}, \\theta) - P (y = 0 | \\mathbf {x}, \\theta). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 281, + 890, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Borrowing from active learning terminology, finding the roots of this expression is akin to a form of uncertainty sampling from our occupancy field (margin sampling [68]). Intuitively, the smaller the margin of a point the more uncertain its shape prediction. In this particular case, the margin function zeroes correspond to the samples with upmost level of uncertainty.", + "bbox": [ + 496, + 306, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that our input point cloud points are samples from the surface, i.e. the ground truth occupancy function decision boundary, we propose to learn our occupancy field by supervising uncertainty sampling with the point cloud, i.e. minimizing the distance between uncertain samples and their nearest point cloud samples. We define our uncertainty sampling as root finding of our margin function. We Initialize this root finding near the target surface. We hypothesis that training our occupancy field by making its uncertain samples coincide with the surface through minimal amount of steps of a root finding algorithm will encourage it to converge quickly and decisively.", + "bbox": [ + 496, + 412, + 892, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by generating a pool of query points near $\\mathcal{P}$ , i.e. near the ground-truth surface, by sampling around the points, i.e. $\\{\\mathbf{q} \\sim \\mathcal{N}(\\mathbf{p}, \\sigma_{\\mathbf{p}}\\mathbf{I}_3)\\}$ . Here, $\\sigma_{\\mathbf{p}}$ is chosen as the maximal euclidean distance to the $K$ nearest points to $\\mathbf{q}$ in $\\mathcal{P}$ (as in [28]). We subsequently recompute the nearest point $\\mathbf{p}$ in $\\mathcal{P}$ to each sample $\\mathbf{q}$ , thus forming the following set of training pairs:", + "bbox": [ + 496, + 593, + 893, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} := \\left\\{\\left(\\mathbf {q}, \\mathbf {p}\\right), \\mathbf {p} = _ {\\mathbf {v} \\in \\mathcal {P}} \\| \\mathbf {v} - \\mathbf {q} \\| _ {2} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 707, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pair $(\\mathbf{q},\\mathbf{p})$ in $\\mathcal{Q}$ , we train by bounding one step of Newton-Raphson [90] root finding on our occupancy margin, initialized at $\\mathbf{q}$ , to result in an uncertain sample near $\\mathbf{p}$ in mean squared error terms.", + "bbox": [ + 496, + 731, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us recall that a Generalized Newton [4] (Theorem 1, with $m = 1$ and $n = 3$ ) iteration updates query point $\\mathbf{q}$ accordingly: $\\mathbf{q} - \\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}U_{\\theta}(\\mathbf{q})$ , where $\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}$ is the Moore-Penrose pseudoinverse of Jacobian $\\nabla U_{\\theta}(\\mathbf{q})$ . As $U_{\\theta}$ is a scalar function $(\\mathbb{R}^{3}\\to \\mathbb{R})$ , $\\nabla U_{\\theta}$ is a row 3-vector (we can use it or its transpose interchangeably), and the pseudoinverse writes: $\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger} = \\frac{\\nabla U_{\\theta}(\\mathbf{q})}{||\\nabla U_{\\theta}(\\mathbf{q})||_2^2}$ .", + "bbox": [ + 496, + 792, + 893, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21731", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In conclusion, our margin uncertainty sampling loss can be expressed as follows:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s a m p}} (\\theta , \\mathcal {Q}) = \\underset {(\\mathbf {q}, \\mathbf {p}) \\sim \\mathcal {Q}} {\\mathbb {E}} | | \\mathbf {q} - U _ {\\theta} (\\mathbf {q}) \\cdot \\frac {\\nabla U _ {\\theta} (\\mathbf {q})}{| | \\nabla U _ {\\theta} (\\mathbf {q}) | | _ {2} ^ {2}} - \\mathbf {p} | | _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 128, + 468, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where spatial gradient $\\nabla U_{\\theta}$ can be computed efficiently through automatic differentiation (e.g. PyTorch [59]).", + "bbox": [ + 76, + 176, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This loss offers two primary advantages. Firstly, it serves as a form of supervision for the occupancy of the query points. The sign of the margin function differs on opposite sides of the decision boundary. Consequently, based on this sign, a Newton-Raphson iteration will move points along or against the direction of its gradient resulting in an occupancy field whose decision boundary is aligned with the input point cloud. Additionally, this loss operates as a smoothness constraint as Newton-Raphson root finding relies on approximating the function at the roots by its first-order approximation on the initial points $\\mathbf{q}$ , hence smoothing the margin function around the input points.", + "bbox": [ + 75, + 207, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Entropy Based Regularization", + "text_level": 1, + "bbox": [ + 76, + 396, + 344, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure better convergence of our training, we use the MLP initialization in [2], while adapting it such that the last layer predicts binary occupancy for an $r$ -radius sphere.", + "bbox": [ + 76, + 419, + 468, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to ease and improve our learning, we suggest to urge our occupancy field to have minimal uncertainty almost everywhere in space. We use the entropy [69] as an uncertainty measure here. However, this is the opposite behavior that we expect from our network to display at the decision boundary i.e. near the ground-truth surface. Hence, we propose to maximize the uncertainty (i.e. entropy) conversely at the input point cloud $\\mathcal{P}$ simultaneously.", + "bbox": [ + 75, + 465, + 468, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a random variable $X$ , Shannon entropy [69] is defined as follows:", + "bbox": [ + 76, + 585, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {H} (X) = - \\mathbb {E} \\log (P (X)). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 626, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create a set $\\Omega \\subset \\mathbb{R}^3$ of query points $\\mathbf{x}$ by sampling space uniformly within a bounding box of input point cloud $\\mathcal{P}$ . As a measure of uncertainty we consider the entropy of the occupancy distribution conditioned on the observed spatial location:", + "bbox": [ + 76, + 652, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e n t r}} (\\theta , \\Omega , \\mathcal {P}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\Omega} \\mathbb {H} (y | \\mathbf {x}, \\theta) - \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\mathbb {H} (y | \\mathbf {p}, \\theta). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 738, + 468, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This entropy polarization loss acts both as an initialization of the field and a regularization mid-learning, as we empirically found that it performs best when its weight ( $\\lambda$ in Equation 6) is reduced progressively in the final loss in the midst of training.", + "bbox": [ + 76, + 765, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally we can train our model using the following combined empirical risk minimization:", + "bbox": [ + 76, + 840, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s a m p}} (\\theta , \\mathcal {Q}) + \\lambda \\mathcal {L} _ {\\text {e n t r}} (\\theta , \\Omega , \\mathcal {P}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 881, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd5079e6c96045da732480b6e22c9bd93973f7b80a5dd35f2107de61f7983e88.jpg", + "image_caption": [ + "Figure 1. Illustration of our training. Our method learns a neural binary occupancy field without off-surface labels. It uses the combination of a margin uncertainty sampling loss near the surface (Green), maximizing entropy at the input point cloud samples (Red), and minimizing entropy everywhere else (Blue)." + ], + "image_footnote": [], + "bbox": [ + 519, + 85, + 870, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 shows a summary of our training procedure. We note that once an implicit occupancy field is learned, an explicit triangle mesh $\\hat{S}$ can be obtained with it through the Marching Cubes algorithm [47].", + "bbox": [ + 498, + 412, + 890, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The training procedure of our method.", + "bbox": [ + 500, + 487, + 816, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Point cloud $\\mathcal{P}$ , learning rate $\\alpha$ , number of iterations $N_{\\mathrm{it}}$ , number of near surface queries $N_{\\mathcal{Q}}$ , number of uniform queries $N_{\\Omega}$ , number of batch input points $N_{\\mathcal{P}}$ , number of uniform queries $N_{\\Omega}$ , reg. loss weight $\\lambda$ .", + "bbox": [ + 500, + 503, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Optimal weights $\\theta^{*}$", + "bbox": [ + 500, + 559, + 681, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute local st. dev. $\\{\\sigma_{\\mathbf{p}} = \\max_{\\mathbf{v}\\in K\\mathrm{mn}(\\mathbf{p},\\mathbf{P})}||\\mathbf{v} - \\mathbf{p}||_2\\}$", + "bbox": [ + 514, + 573, + 867, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generate sets $\\mathcal{Q}$ (Sec.3.1) and $\\varOmega$ (Sec.3.2).", + "bbox": [ + 514, + 588, + 769, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize $\\lambda$ (Sec.4).", + "bbox": [ + 514, + 601, + 632, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $N_{\\mathrm{it}}$ times do", + "bbox": [ + 514, + 614, + 612, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample a batch $\\mathcal{Q}_b$ of size $N_{\\mathcal{Q}}$ from $\\mathcal{Q}$ .", + "bbox": [ + 537, + 628, + 771, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample a batch $\\varOmega_{b}$ of size $N_{\\varOmega}$ from, $\\varOmega_{b}$", + "bbox": [ + 537, + 642, + 782, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample a batch $\\mathcal{P}_b$ of size $N_{\\mathcal{P}}$ from $\\mathcal{P}$ .", + "bbox": [ + 537, + 657, + 769, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute losses $\\mathcal{L}_{\\mathrm{samp}}(\\theta ,\\mathcal{Q}_b)$ (Eq.3).", + "bbox": [ + 537, + 671, + 764, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute losses $\\mathcal{L}_{\\mathrm{entr}}(\\theta, \\Omega_b, \\mathcal{P}_b)$ (Eq.5).", + "bbox": [ + 537, + 685, + 779, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\theta \\gets \\theta -\\alpha \\nabla_{\\theta}\\left(\\mathcal{L}_{\\mathrm{samp}} + \\lambda \\mathcal{L}_{\\mathrm{entr}}\\right)$", + "bbox": [ + 537, + 699, + 728, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Update $\\lambda$ (Sec.4).", + "bbox": [ + 537, + 712, + 643, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 516, + 726, + 563, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 768, + 720, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set number of queries as $N_{\\mathcal{Q}} = 1000k$ and $N_{\\Omega} = 10k$ . Our MLP follows the architecture in [48], and similarly to the latter we set $K = 51$ . We train for $N_{it} = 40k$ iterations on a Nvidia RTX A6000 GPU using the Adam optimizer with learning rate $\\alpha = 0.001$ . Our training takes roughly 5 minutes for a 1024 sized input point cloud. The regularization parameter $\\lambda$ (Equation 6) follows an exponential decay", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21732", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dee2fc2b82cc613e812c2f40b2b555b5f535f0bacb130be14c490555225a3d71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CD1CD2NCFS
SPSR [36]2.340.2240.740.50
OG-INR [38]1.360.0510.550.55
N-Pull [48]1.160.0740.840.75
G-Pull [16]1.070.0320.700.74
NTPS [15]1.110.0670.880.74
Ours0.760.0200.880.83
", + "bbox": [ + 116, + 87, + 431, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the form $\\exp -\\kappa t$ where $\\kappa$ is set to $1.84\\times 10^{-2}$", + "bbox": [ + 75, + 267, + 416, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 76, + 303, + 165, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To test the efficacy of our approach, we assess our capacity to infer implicit representations of shapes when presented with sparse and noisy point clouds. We employ datasets from standard reconstruction benchmarks, showcasing a diverse array of challenges associated with implicit shape function inference from sparse inputs. Consistently with existing literature, we evaluate the performance of our method by quantifying the accuracy of 3D explicit shape models obtained post-convergence from our MLPs. It is worth noting that our approach entails fitting an MLP independently for each point cloud, operating without reliance on pre-learned priors or additional training data. We compare quantitatively and qualitatively to the state-of-the-art in our problem setting, i.e. unsupervised reconstruction from unoriented point, including deep learning methods N-Pull [48], NTPS [15], G-Pull [16], OG-INR [38], DiGs [5], NDrop [9], SAP [61]. We show results for NSpline [84] even though it requires normals. We also compare to classical Poisson Reconstruction (SPSR [36]). We note that NTPS is the closest method to ours as it focuses specifically on the sparse input case. For comprehensive evaluation, we also include comparisons with feed-forward generalizable methods, namely POCO [8] and CONet [60], alongside the prior-based optimization method On-Surf [49] dedicated to sparse inputs. Unless stated differently, we use the publicly available official implementations of existing methods.", + "bbox": [ + 73, + 330, + 473, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Metrics", + "text_level": 1, + "bbox": [ + 76, + 739, + 174, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following seminal work, we evaluate our method and the competition w.r.t. the ground truth using standard metrics for the 3D reconstruction task. Namely, the L1 Chamfer Distance (CD1) $(\\times 10^{2})$ , L2 Chamfer Distance (CD2) $(\\times 10^{2})$ , the Hausdorff distance (HD) and the euclidean distance based F-Score (FS) when ground truth points are available, and finally Normal Consistency (NC) when ground truth normals are available. We detail the expressions of these metrics in the supplementary material.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ecd792566d053c4700f3557f02731c0ef734b9bb839afc493256ccbe0d4e7cc7.jpg", + "table_caption": [ + "Table 1. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + ], + "table_footnote": [], + "table_body": "
CD1CD2NCFS
POCO [8]0.3080.0020.9340.981
CONet [60]1.260.0280.8290.599
On-Surf [49]0.5840.0120.9360.915
SPSR [36]0.7510.0280.8710.839
G-Pull [16]0.4950.0050.8870.945
NTPS [15]0.7370.0150.9430.844
Ours0.2600.0020.9520.974
", + "bbox": [ + 529, + 87, + 864, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Faust [7] reconstructions from sparse noisy unoriented point clouds.", + "bbox": [ + 498, + 224, + 893, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Datasets and Input Definitions", + "text_level": 1, + "bbox": [ + 500, + 279, + 769, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ShapeNet [14] consists of various instances of 13 different synthetic 3D object classes. We follow the train/test splits defined in [84]. We generate noisy input point clouds by sampling 1024 points from the meshes and adding Gaussian noise of standard deviation 0.005 following the literature (e.g. [8, 60]). For brevity we show results on classes Tables, Chairs and Lamps. Faust [7] consists of real scans of 10 human body identities in 10 different poses. We sample sets of 1024 points from the scans as inputs. 3D Scene [92] contains large scale complex real world scenes obtained with a handheld commodity range sensor. We follow [15, 34, 48] and sample our input point clouds with a sparse density of $100\\mathrm{per}\\mathrm{m}^2$ , and we report performance similarly for scenes Burghers, Copyroom, Lounge, Stonewall and Totempole. Surface Reconstruction Benchmark (SRB) [83] consists of five object scans, each with different challenges such as complex topology, high level of detail, missing data and varying feature scales. We sample 1024 points from the scans for the sparse input experiment, and we also experiment using the dense inputs.", + "bbox": [ + 496, + 301, + 893, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Object Level Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 616, + 756, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct the reconstruction of ShapeNet [14] objects using sparse and noisy point clouds. Table 1 and Figure 2 present a quantitative and qualitative comparison with other methods. Our approach consistently outperforms the competition across all metrics, as seen in the visual superiority of our reconstructions. We excel in capturing fine structures and details with greater fidelity. While NTPS demonstrates good overall coarse reconstructions, its thin plate spline smoothing prior appears to limit its expressivity. Additionally, OG-INR struggles to achieve satisfactory results in the sparse and noisy regime, despite its effective Octree-based sign field guidance in denser settings.", + "bbox": [ + 496, + 638, + 893, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.4. Real Articulated Shape Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 830, + 836, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We undertake the reconstruction of Faust [7] human shapes using sparse and noisy point clouds. Table 2 and Figure 3 provide a numerical and qualitative comparison with other", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21733", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81eb2ee4c8c9f849f909841468ea9026f2b8119bb29806b877e5643cffc8ae49.jpg", + "image_caption": [ + "Figure 2. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + ], + "image_footnote": [], + "bbox": [ + 200, + 85, + 774, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods. Our approach surpasses the alternatives across all metrics, with visually superior reconstructions, as especially noticeable at the extremities of the body. Analogous to the fine structures in the ShapeNet experiment, these areas pose challenges due to limited input point cloud samples, making shape prediction inherently difficult and ambiguous. NTPS reconstructions, similarly, exhibit coarser and less detailed results on this dataset. It is worth highlighting that our method outperforms not only feed-forward generalizable approaches POCO and CONet but also the prior-based optimization method On-Surf. These methods rely on priors trained on ShapeNet, thereby constraining their generalization ability.", + "bbox": [ + 75, + 529, + 472, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Real Scene Level Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 739, + 364, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In accordance with [15], we present the reconstruction outcomes on the 3D Scene [92] dataset derived from spatially sparse point clouds. Table 3 provides a summary of numerical results, incorporating results for methods NTPS, N-Pull, SAP, NDrop, and NSpline as reported in the state-of-the-art method NTPS. Our method excels in this benchmark, surpassing the competition in most metrics. Notably, our baseline N-Pull exhibits more conspicuous failures in this extensive sparse setup. Figure 5 offers qualitative compar", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "isons against baselines N-Pull and SPSR.", + "bbox": [ + 500, + 529, + 772, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 558, + 663, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss. Our approach centers around aligning the decision boundary of our occupancy function with the shape surface through a loss function based on uncertainty sampling, denoted as $\\mathcal{L}_{\\mathrm{samp}}$ (Equation 3). To facilitate and enhance our learning process, we impose constraints on the network to minimize uncertainty almost everywhere in space while maximizing it at the decision boundary. Entropy serves as our chosen uncertainty measure. However, other measures can be considered. We justify this choice in Table 4. We compare our entropy-based regularization to margin-based regularization (Margin reg), using the absolute value of our margin function $U_{\\theta}$ (Equation 3) as an uncertainty measure instead of entropy. Additionally, we compare it to the sign-agnostic occupancy supervision strategy introduced in [75] denoted as SA-Occ. Although this strategy fails to converge in the absence of priors, we demonstrate its beneficial use as a regularization in place of our entropy-based approach. It is noteworthy that while margin-based regularization (Margin reg) yields better results, it is still outperformed by our entropy-based regularization. Figure 4 shows an example of validation plots, displaying the role of our regularization", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21734", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48f8d5bb93b51cc1571c9892691a6f4c79db3af6966ce1cee4d803c5a64a4529.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BurghersCopyroomLoungeStonewallTotempleMean
CD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NC
SPSR [36]0.1780.2050.8740.2250.2860.8610.2800.3650.8690.3000.4800.8660.5881.6730.8790.3140.6020.870
NDrop [9]0.2000.1140.8250.1680.0630.6960.1560.0500.6630.1500.0810.8150.2030.1390.8440.1750.0890.769
N-Pull [48]0.0640.0080.8980.0490.0050.8280.1330.0380.8470.0600.0050.9100.1780.0240.9080.0970.0160.878
SAP [61]0.1530.1010.8070.0530.0090.7710.1340.0330.8130.0700.0070.8670.4740.3820.7250.1510.1000.797
NSpline [84]0.1350.1230.8910.0560.0230.8550.0630.0390.8270.1240.0910.8970.3780.7680.8920.1510.2090.88
NTPS [15]0.0550.0050.9090.0450.0030.8920.1290.0220.8720.0540.0040.9390.1030.0170.9350.0770.0100.897
Ours0.0220.0010.8710.0410.0120.8120.0210.0010.8700.0280.0030.9310.0260.0010.9360.0270.0030.886
", + "bbox": [ + 81, + 87, + 924, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7eb3af6fe9f64604aa6453410b2577c7ab2ce9c92474e7acb0e6b6217f438e9.jpg", + "image_caption": [ + "Figure 3. Faust [7] reconstructions from sparse unoriented point clouds." + ], + "image_footnote": [], + "bbox": [ + 81, + 239, + 467, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in avoiding overfitting. The baseline in this example is SDF method N-Pull [48].", + "bbox": [ + 75, + 776, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point Cloud Density. We utilize the SRB [83] benchmark to evaluate the performance of our method across varying point cloud densities. Table 5 presents comparative results for sparse (1024) and dense input point clouds. We include results for the competition from OG-INR in the dense setting. Our method outperforms the competition in the sparse", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/349e8e5adbe7a19cd92bf9511ab81e613da5b972f08820a1c4c2f3ae7b630f97.jpg", + "image_caption": [ + "Figure 4. CD1 distance to GT for reconstructions of shape Gargoyle of benchmark SRB [83] from a sparse unoriented point cloud." + ], + "image_footnote": [], + "bbox": [ + 573, + 246, + 794, + 377 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5ef64b149d69b3cec784ba2f43217e7735bdf270145d8b0603bacc8f853c9f00.jpg", + "table_caption": [ + "Table 3. 3D Scene [92] reconstructions from sparse point clouds." + ], + "table_footnote": [], + "table_body": "
CD1NC
N-Pull [48] (baseline)1.100.85
\\( \\mathcal{L}_{\\text{samp}} \\) (Equ.3)1.020.86
\\( \\mathcal{L}_{\\text{samp}} + \\text{Margin reg} \\)0.800.88
\\( \\mathcal{L}_{\\text{samp}} + \\text{SA-Occ [75] reg} \\)0.950.88
Ours0.770.89
", + "bbox": [ + 584, + 433, + 815, + 511 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f432362f6c8f98b84504b7f1388d5266279a292d50069d553914775f6b2d01ac.jpg", + "table_caption": [ + "Table 4. Ablation of our method on the Tables class of ShapeNet [14]." + ], + "table_footnote": [], + "table_body": "
SparseDense
CD1HDCD1HD
SPSR [36]2.2721.11.2522.59
DiGs [5]0.686.050.193.52
OG-INR [38]0.857.100.204.06
NTPS [15]0.737.78--
N-Pull [48]0.588.900.234.46
Ours0.496.040.203.94
", + "bbox": [ + 570, + 564, + 825, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation of point cloud density.", + "bbox": [ + 573, + 675, + 816, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "case and performs comparably to the state-of-the-art in the dense case. The improvement over our baseline is substantial for both sparse and dense inputs, as highlighted visually in Figure 6, where we showcase reconstructions for both sparse and dense cases. Notably, we achieve better topologies in the sparse case and enhanced, more accurate details in the dense case, as indicated by the red boxes. These results underscore the utility and advantages of our contribution, even in the dense setting.", + "bbox": [ + 496, + 718, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Running Time and Performance against Number of Queries and Input Points. Fig.7 shows our best performance over time (minutes) for various input point cloud", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21735", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/265fdd355c746497529ef65089ba3ea07d009c114700d275932561cc9752789f.jpg", + "image_caption": [ + "Figure 5. 3D Scene [92] reconstructions from sparse unoriented point clouds." + ], + "image_footnote": [], + "bbox": [ + 202, + 88, + 772, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e610bebc72aa7673d6d946c2ad20600042bc4d0e9d2156767070b607559f6e2.jpg", + "image_caption": [ + "Figure 6. SRB [83] reconstructions from sparse and dense unoriented inputs." + ], + "image_footnote": [], + "bbox": [ + 76, + 448, + 467, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/596d88e2e444ce3b141f827b8d2ae41b6dbc430fe3a70c11008493f1f0a3b57e.jpg", + "image_caption": [ + "Figure 7. Performance over time (varying input size)." + ], + "image_footnote": [], + "bbox": [ + 81, + 650, + 240, + 743 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b90b9c197cd2cee7ff89d513c475eac763bd097949670ff3fb8436e9e87ff952.jpg", + "image_caption": [ + "Figure 8. Performance over time (varying query set size)." + ], + "image_footnote": [], + "bbox": [ + 289, + 651, + 447, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sizes. Models using larger inputs reach the baseline performance at earlier times. Fig.8 shows our best performance over time while decreasing the total number of query samples. Performance does not deteriorate drastically under less queries. These Experiments were performed using shape Gargoyle of benchmark SRB.", + "bbox": [ + 75, + 792, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to Noise We evaluate here (Tab.6) our method", + "bbox": [ + 76, + 885, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b40e99b7e7cbf44c8f2370fcb95a636a02d931fd3c86f5293153c4a3c7691a40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
noise std. dev.00.0050.025
CD1 ↓0.560.772.16
NC ↑0.930.890.68
", + "bbox": [ + 586, + 448, + 808, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Noise ablation.", + "bbox": [ + 622, + 500, + 767, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with various noise levels on class Table of ShapeNet. Performance remains reasonable even under heavily corrupted inputs, which confirms our qualitative superiority to competition in Fig.2, and our resilience to noise.", + "bbox": [ + 500, + 542, + 890, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations", + "text_level": 1, + "bbox": [ + 500, + 619, + 622, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We notice that excess regularization can cause reconstructions to be overly smooth in some cases, which can affect our NC numbers is some benchmarks. This can be improved by tuning the hyperparameters of the method per benchmark or per object/scene.", + "bbox": [ + 500, + 646, + 890, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 737, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a method for implicit shape reconstruction from sparse, noisy and unoriented point cloud. Our results demonstrate that occupancy fields offer and effective implicit shape representation for this task, and that they can be learned successfully even in the unsupervised setting. Under our carefully designed losses, the occupancy outperforms state-of-the-art methods based on SDF, as-well-as other occupancy based baselines that we show in ablation studies.", + "bbox": [ + 500, + 763, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21736", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. CG, 2001. 2", + "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3, 4", + "[3] Matan Atzmon and Yaron Lipman. Sald: Sign agnostic learning with derivatives. In ICML, 2020. 3", + "[4] Adi Ben-Israel. A newton-raphson method for the solution of systems of equations. In Journal of Mathematical analysis and applications, 1966. 3", + "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In CVPR, 2022. 2, 5, 7", + "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. TVCG, 1999. 2", + "[7] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In CVPR, 2014. 5, 7", + "[8] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2, 3, 5", + "[9] Alexandre Boulch, Pierre-Alain Langlois, Gilles Puy, and Renaud Marlet. Needrop: Self-supervised shape representation from sparse point clouds using needle dropping. In 2021 International Conference on 3D Vision (3DV), pages 940-950. IEEE, 2021. 3, 5, 7", + "[10] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3d objects with radial basis functions. In SIGGRAPH, 2001. 2", + "[11] Frédéric Casals and Joachim Giesen. Effective Computational Geometry for Curves and Surfaces. 2006. 2", + "[12] Rohan Chabra, Jan E Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In ECCV, 2020. 2", + "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[14] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 5, 6, 7", + "[15] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 3, 5", + "[17] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, 2022. 2", + "[18] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2", + "[19] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[20] Julian Chibane and Gerard Pons-Moll. Implicit feature networks for texture completion from partial 3d data. In European Conference on Computer Vision, pages 717-725. Springer, 2020. 1, 2", + "[21] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, 2020. 2", + "[22] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In CVPR, 2020. 2", + "[23] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2", + "[24] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In ECCV, 2020. 2", + "[25] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2", + "[26] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2", + "[27] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In CVPR, 2020. 2", + "[28] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 1, 2, 3", + "[29] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 2", + "[30] Gaël Guennebaud and Markus Gross. Algebraic point set surfaces. In ACM siggraph 2007 papers, pages 23–es. 2007. 1, 2", + "[31] John C Hart. Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer, 1996. 2", + "[32] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 2", + "[33] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 2", + "[34] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 2, 5", + "[35] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2", + "[36] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. TOG, 2013. 1, 2, 5, 7", + "[37] Ravikrishna Kolluri. Provably good moving least squares. TALG, 2008. 2", + "[38] Chamin Hewa Koneputugodage, Yizhak Ben-Shabat, and Stephen Gould. Octree guided unoriented surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16717-16726, 2023. 3, 5, 7", + "[39] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3d shapes with dynamic code clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12840-12850, 2022. 2", + "[40] Siyou Lin, Dong Xiao, Zuoqiang Shi, and Bin Wang. Surface reconstruction from point clouds without normals by parametrizing the gauss formula. ACM Transactions on Graphics, 42(2):1-19, 2022. 2", + "[41] Stefan Lionar, Daniil Emtsev, Dusan Svilarkovic, and Songyou Peng. Dynamic plane convolutional occupancy networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1829-1838, 2021. 2", + "[42] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In ICML, 2021. 3", + "[43] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In CVPR, 2018. 2", + "[44] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization. arXiv preprint arXiv:2202.08345, 2022. 3", + "[45] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In ECCV, 2020. 2", + "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In CVPR, 2021. 2", + "[47] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, 1987. 2, 4", + "[48] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In ICML, 2021. 3, 4, 5, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6315-6325, 2022. 3, 5", + "[50] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6326-6337, 2022. 3", + "[51] Corentin Mercier, Thibault Lescoat, Pierre Roussillon, Tamy Boubekeur, and Jean-Marc Thiery. Moving level-of-detail surfaces. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2", + "[52] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1, 2, 3", + "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2", + "[54] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 2", + "[55] Amine Ousafi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. In NeurIPS, 2023. 1", + "[56] Amine Ousafi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. In 3DV, 2024. 2", + "[57] David Palmer, Dmitriy Smirnov, Stephanie Wang, Albert Chern, and Justin Solomon. Deepcurrents: Learning implicit representations of shapes with boundaries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18665-18675, 2022. 2", + "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2", + "[59] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4", + "[60] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1, 2, 5", + "[61] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 2, 3, 5, 7", + "[62] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21738", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Marie-Julie Rakotosaona, Noam Aigerman, Niloy Mitra, Maks Ovsjanikov, and Paul Guerrero. Differentiable surface triangulation. In SIGGRAPH Asia, 2021. 2", + "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2", + "[65] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In NeurIPS, 2004. 2", + "[66] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[67] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1", + "[68] Burr Settles. Active learning literature survey. 2009. 3", + "[69] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 4", + "[70] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2", + "[71] Vincent Sitzmann, Eric R Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 2", + "[72] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2, 3", + "[73] Vincent Sitzmann, Semon Rezchikov, William T Freeman, Joshua B Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2", + "[74] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2", + "[75] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. Sa-convonet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6504–6513, 2021. 6, 7", + "[76] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 2", + "[77] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. Patchnets: Patch-based generalizable deep implicit 3d shape representations. In ECCV, 2020. 2", + "[78] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 2", + "[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2", + "[81] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. TOG, 2017. 2", + "[82] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. Advances in Neural Information Processing Systems, 34:2810-2822, 2021. 2", + "[83] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In CVPR, 2019. 2, 5, 7, 8", + "[84] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural splines: Fitting 3d surfaces with infinitely-wide neural networks. In CVPR, 2021. 2, 3, 5, 7", + "[85] Francis Williams, Zan Gojcic, Sameh Khamis, Denis Zorin, Joan Bruna, Sanja Fidler, and Or Litany. Neural fields as learnable kernels for 3d reconstruction. In CVPR, 2022. 2", + "[86] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T Freeman, and Joshua B Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In NeurIPS, 2016. 2", + "[87] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2", + "[88] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2", + "[89] Mohsen Yavartanoo, Jaeyoung Chung, Reyhaneh Neshatavar, and Kyoung Mu Lee. 3dias: 3d shape reconstruction with implicit algebraic surfaces. In ICCV, 2021. 2", + "[90] Tjalling J Ypma. Historical development of the newton-raphson method. SIAM review, 37(4):531-551, 1995. 3", + "[91] Junsheng Zhou, Baorui Ma, Liu Yu-Shen, Fang Yi, and Han Zhizhong. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[92] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics (ToG), 32(4):1-8, 2013. 5, 6, 7, 8", + "[93] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3d-prnn: Generating shape primitives with recurrent neural networks. In CVPR, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 771 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21739", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_model.json b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60efa51dcff45a2539ecd8b62efe1583b4f3591b --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.792, + 0.153 + ], + "angle": 0, + "content": "Unsupervised Occupancy Learning from Sparse Point Cloud" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.182, + 0.677, + 0.216 + ], + "angle": 0, + "content": "Amine Ousafi Adnane Boukhayma \nInria, Univ. Rennes, CNRS, IRISA, M2S, France" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.473, + 0.617 + ], + "angle": 0, + "content": "Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from 3D point clouds in the absence of ground truth supervision remains a very challenging task. In this paper, we propose a method to infer occupancy fields instead of SDFs as they are easier to learn from sparse inputs. We leverage a margin-based uncertainty measure to differentiably sample from the decision boundary of the occupancy function and supervise the sampled boundary points using the input point cloud. We further stabilise the optimization process at the early stages of the training by biasing the occupancy function towards minimal entropy fields while maximizing its entropy at the input point cloud. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve implicit shape inference with respect to baselines and the state-of-the-art using synthetic and real data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Capturing full 3D shape from limited and corrupted data is a long standing problem. In this regard, one of the data type instances that has received increasing attention and investigation from computer vision, graphics and machine learning is point clouds. This interest emanates primarily from the ubiquity of this light, although topologically incomplete 3D representation, either as acquired e.g. through industrial and commodity depth sensors, or as an intermediate representation within computational photogrammetry [66, 67] pipelines for instance. While classical optimization methods such as Poisson Reconstruction [36] or Moving Least Squares [30] can be effective with dense, clean point sets and accurate normal pre-estimations, recent deep learning-based alternatives provide more robust predictions, particularly for noisy and sparse inputs, eliminating the need for" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.254, + 0.682, + 0.269 + ], + "angle": 0, + "content": "normal data in many cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.895, + 0.455 + ], + "angle": 0, + "content": "Several methods rely on priors learned from large fully labeled data such as the synthetic dataset ShapeNet [14]. However, this strategy entails computationally expensive trainings, and the resulting models can still be prone to out-of-distribution generalization issues, as pointed in [15, 55], whether caused by change in the input size or domain shift. For instance, Table 2 shows that our unsupervised method outperforms supervised generalizable models when testing on data that is sparser and different in nature from their training corpus. Hence, it is important to design learning frameworks that can lead to robust reconstruction under such extreme constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Neural implicit representations (INR) were established recently as a powerful representation for 3D shape, generally taking the form of coordinate based MLPs. They have been applied to model various shape presentations, the most popular ones of which remain arguably signed distance functions (SDF) [58] and binary occupancy fields [52] for watertight shapes. For the task of point cloud reconstruction, there is a wide literature on learning SDFs from dense labels, as well as from point clouds solely (unsupervised), either in the generalizable setting or in instance specific optimization. However, apart from its applications in the supervised generalizable setting [8, 20, 60], it seems that occupancy learning from point sets has not been as widely considered in other scenarios. In particular, to the best of our knowledge, using occupancy fields to learn shapes from sparse noisy unoriented point clouds (i.e. unsupervisedly) has not been explored thoroughly. This is even the more intriguing as it seems to be easier to learn a binary classification task as opposed to regressing a continuous field that must additionally respect specific properties [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Concordantly, we introduce occupancy in this work for learning shape from a sparse point cloud. While other representations such as SDF can readily incorporate surface sample supervision directly (e.g. [28]), among other strategies, it is not entirely straightforward to define such supervision for occupancy fields. Hence, we propose a novel strategy for our training that combines a loss applied near the boundary decision, harnessing the input point samples, and an accompanying regularization. For the first loss, we encourage" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21729" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "the decision boundary of our occupancy to align with the point cloud. We achieve this by supervising an uncertainty sampling mechanism for our occupancy field with the point cloud samples. We approximate this sampling through root-finding of a margin function associated to the field. As regularization, we propose to minimize the uncertainty of our field, while maximizing it near the decision boundary, via the entropy of our occupancy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.334 + ], + "angle": 0, + "content": "Through extensive experiments under several real and synthetic benchmarks for object, non-rigid and scene level shape reconstruction, our results show that our method outperforms the existing state-of-the-art literature in reconstruction from sparse point cloud, using standard metrics and through superior visual results. Our ablation studies validate our design choices and showcase the importance of the components of the supervision scheme that we propose." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.347, + 0.22, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Shape Representations in Deep Learning Shapes can be depicted in deep learning through either intrinsic or extrinsic representations. Intrinsic representations focus on discriminating the shape itself. When explicitly implemented, for instance, using tetrahedral or polygonal meshes [35, 79] or point clouds [25], the output topology is predefined, thereby restricting the range of shapes that can be generated. Among other intrinsic representations, 2D patches [23, 29, 83] may introduce discontinuities, while the simplicity of shape primitives like cuboids [78, 93], planes [43], and Gaussians [26] constrains their expressiveness. On the other hand, extrinsic shape representations model the entire space containing the scene or object of interest. Voxel grids [86, 87] are the most popular, serving as a direct extension of 2D pixels to the 3D domain. However, their capacity is constrained by the memory cost associated with cubic resolution. Sparse representations, such as octrees [64, 76, 81], can mitigate this issue to some extent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implicit Neural Shape Representations Implicit Neural Shape Representations have recently emerged as a significant approach for modeling extrinsic shape and radiance fields (e.g. [13, 33, 53, 80, 88]). These representations address many of the limitations associated with classical representations by showcasing the ability to represent shapes with arbitrary topologies at virtually infinite resolution. Typically, they are parameterized with MLPs, which map spatial locations or features to properties such as occupancy [52], signed [58], or unsigned [21, 91] distances relative to the target shape. The level-set derived from these MLPs can be visualized through techniques like ray marching [31] or tessellated into an explicit shape using methods like Marching Cubes [47]. Another notable line of research involves the development of hybrid implicit/explicit representations [19, 22, 57, 89], primarily based on differentiable space partitioning. To simultaneously represent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "collections of shapes, implicit neural models necessitate conditioning mechanisms. These mechanisms encompass features and latent code concatenation, batch normalization, hypernetworks [17, 70, 72, 73, 82], and gradient-based meta-learning [54, 71]. Concatenation-based conditioning was initially implemented using single global latent codes [18, 52, 58] and has been further refined with the incorporation of local features [20, 24, 27, 34, 39, 60, 74, 77]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Reconstruction from Point Clouds Classical approaches to reconstruction include combinatorical methods where the shape is defined based on input point clouds through space partitioning, employing techniques such as alpha shapes [6], Voronoi diagrams [1], or triangulation [11, 45, 63]. Alternatively, the input samples can contribute to defining an implicit function, with its zero level set representing the target shape. This is achieved through global smoothing priors [40, 84, 85], such as radial basis functions [10] and Gaussian kernel fitting [65], or local smoothing priors like moving least squares [30, 37, 46, 51]. Another approach involves solving a boundary-conditioned Poisson equation [36]. Recent literature suggests parameterizing these implicit functions with deep neural networks and learning their parameters through gradient descent, either in a supervised or unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.791 + ], + "angle": 0, + "content": "Supervised Implicit Neural Reconstruction In supervised methods, there is an assumption of having labeled training data, typically in the form of dense samples containing ground truth shape information. Auto-decoding techniques [12, 34, 39, 58, 77] necessitate test-time optimization to adapt to a new point cloud. On the other hand, encoder-decoder methods allow for swift feed-forward inference. Initially introduced for this purpose, Pooling-based set encoders [18, 27, 52] like PointNet [62] have been found to underfit the context. State-of-the-art performance is achieved by convolutional encoders, utilizing local features defined either in explicit volumes and planes [20, 41, 60, 61] or solely at the input points [8, 56]. Peng et al. [61] proposed a differentiable Poisson solving layer that efficiently converts predicted normals into an indicator function grid. However, its applicability is limited to small scenes due to the cubic memory requirement in grid resolution. The work in [32, 85] suggests constructing generalizable models with implicit decoder functions as a kernel regression. Despite these advancements, many of the generalizable methods still face challenges related to generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Unsupervised Implicit Neural Reconstruction For unsupervised approaches, a neural network is typically fitted to the input point cloud without additional information. Convergence improvements can be achieved through regularization techniques, such as the spatial gradient constraint based on the Eikonal equation introduced by Gropp et al. [28], a spatial divergence constraint as described in [5], and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.425 + ], + "angle": 0, + "content": "Lipschitz regularization on the network [44]. Periodic activations were introduced in [72]. Lipman [42] learns a function that converges to occupancy, while its log transform converges to a distance function. Atzmon et al. [2] learn an SDF from unsigned distances, further supervising the spatial gradient of the function with normals [3]. Ma et al. [48] express the nearest point on the surface as a function of the neural signed distance and its gradient. They also utilize self-supervised local priors to handle very sparse inputs [49] and enhance generalization [50]. [38] guides the implicit field learning with an Octree based labelling. [9] predicts occupancy fields by learning whether a dropped needle goes across the surface or no. [15] learns a surface parameterization leveraged to provide additional coarse surface supervision to the shape network. In [84], infinitely wide shallow MLPs are learned as random feature kernels using points and their normals. However, most of the methods mentioned above encounter challenges when dealing with sparse and noisy input, primarily due to the lack of supervision. Differently from this literature, we propose here to apply occupancy to unsupervised sparse point cloud neural fitting for the first time to the best of our knowledge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.172, + 0.453 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.523 + ], + "angle": 0, + "content": "Given a sparse noisy unoriented input point cloud \\(\\mathcal{P} \\subset \\mathbb{R}^{3 \\times N}\\), our goal is to recover a 3D watertight shape surface \\(S\\) that best explains this observation, i.e. points from \\(\\mathcal{P}\\) approximating noisy samples from \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Based on the proven success of implicit neural representations, we propose to address this problem by fitting an implicit neural shape function to the scarce observation, which amounts to learning an MLP mapping coordinates to shape attributes. In this context, dubbed unsupervised, signed distance fields are the representation of choice in the community so far (e.g. [15, 16, 38, 48, 61, 84]). Oddly enough, binary occupancy fields were not explored extensively yet as a representation for this context to the best of our knowledge, even though in theory, it seems easier to learn occupancy than SDFs. In fact, it is safe to assume that binary classification is an easier task than continuous regression in general. Besides, we can see intuitively that occupancy is only the sign part of the SDF information. Additionally, occupancy fields do not seem to stipulate any strong structural requirement, as it is the case for distance fields, whose gradient ought to satisfy the Eikonal constraint [28] for instance. Hence, we propose here to learn a binary occupancy field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.904 + ], + "angle": 0, + "content": "When labeled samples inside and outside the shape are available, such a neural occupancy field could be learned in this supervised setting via a standard cross-entropy classification loss (e.g. [52]). However when only surface points are available, as in our case \\((\\mathcal{P})\\), it is not entirely clear how a successful learning can be achieved using such occupancy" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "decision boundary samples alone. We derive in the following a methodology to approach this task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.129, + 0.892, + 0.162 + ], + "angle": 0, + "content": "3.1. Learning Occupancy Through Margin Uncertainty Sampling" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.894, + 0.259 + ], + "angle": 0, + "content": "We are interested in learning the following posterior \\( P(y|\\mathbf{x},\\theta) \\), where \\( \\mathbf{x} \\) represents query points in \\( \\mathbb{R}^3 \\), and label \\( y \\in \\{0,1\\} \\) stands for binary shape occupancy. \\( P \\) is parameterized with a Softmax activated MLP \\( \\theta \\) similarly to [8]. Since we do not have points with their occupancy labels, all we can learn with is the input point cloud samples." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.259, + 0.83, + 0.274 + ], + "angle": 0, + "content": "Let us consider the following margin function:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.282, + 0.892, + 0.299 + ], + "angle": 0, + "content": "\\[\nU _ {\\theta} (\\mathbf {x}) = P (y = 1 | \\mathbf {x}, \\theta) - P (y = 0 | \\mathbf {x}, \\theta). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Borrowing from active learning terminology, finding the roots of this expression is akin to a form of uncertainty sampling from our occupancy field (margin sampling [68]). Intuitively, the smaller the margin of a point the more uncertain its shape prediction. In this particular case, the margin function zeroes correspond to the samples with upmost level of uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Given that our input point cloud points are samples from the surface, i.e. the ground truth occupancy function decision boundary, we propose to learn our occupancy field by supervising uncertainty sampling with the point cloud, i.e. minimizing the distance between uncertain samples and their nearest point cloud samples. We define our uncertainty sampling as root finding of our margin function. We Initialize this root finding near the target surface. We hypothesis that training our occupancy field by making its uncertain samples coincide with the surface through minimal amount of steps of a root finding algorithm will encourage it to converge quickly and decisively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.894, + 0.701 + ], + "angle": 0, + "content": "We start by generating a pool of query points near \\(\\mathcal{P}\\), i.e. near the ground-truth surface, by sampling around the points, i.e. \\(\\{\\mathbf{q} \\sim \\mathcal{N}(\\mathbf{p}, \\sigma_{\\mathbf{p}}\\mathbf{I}_3)\\}\\). Here, \\(\\sigma_{\\mathbf{p}}\\) is chosen as the maximal euclidean distance to the \\(K\\) nearest points to \\(\\mathbf{q}\\) in \\(\\mathcal{P}\\) (as in [28]). We subsequently recompute the nearest point \\(\\mathbf{p}\\) in \\(\\mathcal{P}\\) to each sample \\(\\mathbf{q}\\), thus forming the following set of training pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.708, + 0.892, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} := \\left\\{\\left(\\mathbf {q}, \\mathbf {p}\\right), \\mathbf {p} = _ {\\mathbf {v} \\in \\mathcal {P}} \\| \\mathbf {v} - \\mathbf {q} \\| _ {2} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Given a pair \\((\\mathbf{q},\\mathbf{p})\\) in \\(\\mathcal{Q}\\), we train by bounding one step of Newton-Raphson [90] root finding on our occupancy margin, initialized at \\(\\mathbf{q}\\), to result in an uncertain sample near \\(\\mathbf{p}\\) in mean squared error terms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.894, + 0.905 + ], + "angle": 0, + "content": "Let us recall that a Generalized Newton [4] (Theorem 1, with \\(m = 1\\) and \\(n = 3\\)) iteration updates query point \\(\\mathbf{q}\\) accordingly: \\(\\mathbf{q} - \\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}U_{\\theta}(\\mathbf{q})\\), where \\(\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}\\) is the Moore-Penrose pseudoinverse of Jacobian \\(\\nabla U_{\\theta}(\\mathbf{q})\\). As \\(U_{\\theta}\\) is a scalar function \\((\\mathbb{R}^{3}\\to \\mathbb{R})\\), \\(\\nabla U_{\\theta}\\) is a row 3-vector (we can use it or its transpose interchangeably), and the pseudoinverse writes: \\(\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger} = \\frac{\\nabla U_{\\theta}(\\mathbf{q})}{||\\nabla U_{\\theta}(\\mathbf{q})||_2^2}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "In conclusion, our margin uncertainty sampling loss can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.13, + 0.469, + 0.176 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s a m p}} (\\theta , \\mathcal {Q}) = \\underset {(\\mathbf {q}, \\mathbf {p}) \\sim \\mathcal {Q}} {\\mathbb {E}} | | \\mathbf {q} - U _ {\\theta} (\\mathbf {q}) \\cdot \\frac {\\nabla U _ {\\theta} (\\mathbf {q})}{| | \\nabla U _ {\\theta} (\\mathbf {q}) | | _ {2} ^ {2}} - \\mathbf {p} | | _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.207 + ], + "angle": 0, + "content": "where spatial gradient \\(\\nabla U_{\\theta}\\) can be computed efficiently through automatic differentiation (e.g. PyTorch [59])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.208, + 0.47, + 0.388 + ], + "angle": 0, + "content": "This loss offers two primary advantages. Firstly, it serves as a form of supervision for the occupancy of the query points. The sign of the margin function differs on opposite sides of the decision boundary. Consequently, based on this sign, a Newton-Raphson iteration will move points along or against the direction of its gradient resulting in an occupancy field whose decision boundary is aligned with the input point cloud. Additionally, this loss operates as a smoothness constraint as Newton-Raphson root finding relies on approximating the function at the roots by its first-order approximation on the initial points \\(\\mathbf{q}\\), hence smoothing the margin function around the input points." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.345, + 0.414 + ], + "angle": 0, + "content": "3.2. Entropy Based Regularization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.42, + 0.469, + 0.465 + ], + "angle": 0, + "content": "To ensure better convergence of our training, we use the MLP initialization in [2], while adapting it such that the last layer predicts binary occupancy for an \\( r \\)-radius sphere." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.586 + ], + "angle": 0, + "content": "Furthermore, to ease and improve our learning, we suggest to urge our occupancy field to have minimal uncertainty almost everywhere in space. We use the entropy [69] as an uncertainty measure here. However, this is the opposite behavior that we expect from our network to display at the decision boundary i.e. near the ground-truth surface. Hence, we propose to maximize the uncertainty (i.e. entropy) conversely at the input point cloud \\(\\mathcal{P}\\) simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.586, + 0.469, + 0.615 + ], + "angle": 0, + "content": "Given a random variable \\(X\\), Shannon entropy [69] is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.627, + 0.469, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbb {H} (X) = - \\mathbb {E} \\log (P (X)). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.469, + 0.729 + ], + "angle": 0, + "content": "We create a set \\(\\Omega \\subset \\mathbb{R}^3\\) of query points \\(\\mathbf{x}\\) by sampling space uniformly within a bounding box of input point cloud \\(\\mathcal{P}\\). As a measure of uncertainty we consider the entropy of the occupancy distribution conditioned on the observed spatial location:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.739, + 0.469, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e n t r}} (\\theta , \\Omega , \\mathcal {P}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\Omega} \\mathbb {H} (y | \\mathbf {x}, \\theta) - \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\mathbb {H} (y | \\mathbf {p}, \\theta). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.766, + 0.469, + 0.841 + ], + "angle": 0, + "content": "This entropy polarization loss acts both as an initialization of the field and a regularization mid-learning, as we empirically found that it performs best when its weight (\\(\\lambda\\) in Equation 6) is reduced progressively in the final loss in the midst of training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Finally we can train our model using the following combined empirical risk minimization:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.882, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s a m p}} (\\theta , \\mathcal {Q}) + \\lambda \\mathcal {L} _ {\\text {e n t r}} (\\theta , \\Omega , \\mathcal {P}). \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.086, + 0.872, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Figure 1. Illustration of our training. Our method learns a neural binary occupancy field without off-surface labels. It uses the combination of a margin uncertainty sampling loss near the surface (Green), maximizing entropy at the input point cloud samples (Red), and minimizing entropy everywhere else (Blue)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Algorithm 1 shows a summary of our training procedure. We note that once an implicit occupancy field is learned, an explicit triangle mesh \\(\\hat{S}\\) can be obtained with it through the Marching Cubes algorithm [47]." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.488, + 0.818, + 0.503 + ], + "angle": 0, + "content": "Algorithm 1 The training procedure of our method." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "Input: Point cloud \\(\\mathcal{P}\\), learning rate \\(\\alpha\\), number of iterations \\(N_{\\mathrm{it}}\\), number of near surface queries \\(N_{\\mathcal{Q}}\\), number of uniform queries \\(N_{\\Omega}\\), number of batch input points \\(N_{\\mathcal{P}}\\), number of uniform queries \\(N_{\\Omega}\\), reg. loss weight \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.56, + 0.682, + 0.574 + ], + "angle": 0, + "content": "Output: Optimal weights \\(\\theta^{*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.574, + 0.868, + 0.589 + ], + "angle": 0, + "content": "Compute local st. dev. \\(\\{\\sigma_{\\mathbf{p}} = \\max_{\\mathbf{v}\\in K\\mathrm{mn}(\\mathbf{p},\\mathbf{P})}||\\mathbf{v} - \\mathbf{p}||_2\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.589, + 0.771, + 0.601 + ], + "angle": 0, + "content": "Generate sets \\(\\mathcal{Q}\\) (Sec.3.1) and \\(\\varOmega\\) (Sec.3.2)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.602, + 0.633, + 0.614 + ], + "angle": 0, + "content": "Initialize \\(\\lambda\\) (Sec.4)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.616, + 0.614, + 0.628 + ], + "angle": 0, + "content": "for \\(N_{\\mathrm{it}}\\) times do" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.629, + 0.772, + 0.643 + ], + "angle": 0, + "content": "Sample a batch \\(\\mathcal{Q}_b\\) of size \\(N_{\\mathcal{Q}}\\) from \\(\\mathcal{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.643, + 0.783, + 0.657 + ], + "angle": 0, + "content": "Sample a batch \\(\\varOmega_{b}\\) of size \\(N_{\\varOmega}\\) from, \\(\\varOmega_{b}\\)" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.658, + 0.77, + 0.671 + ], + "angle": 0, + "content": "Sample a batch \\(\\mathcal{P}_b\\) of size \\(N_{\\mathcal{P}}\\) from \\(\\mathcal{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.672, + 0.765, + 0.685 + ], + "angle": 0, + "content": "Compute losses \\(\\mathcal{L}_{\\mathrm{samp}}(\\theta ,\\mathcal{Q}_b)\\) (Eq.3)." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.686, + 0.781, + 0.699 + ], + "angle": 0, + "content": "Compute losses \\(\\mathcal{L}_{\\mathrm{entr}}(\\theta, \\Omega_b, \\mathcal{P}_b)\\) (Eq.5)." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.7, + 0.73, + 0.713 + ], + "angle": 0, + "content": "\\(\\theta \\gets \\theta -\\alpha \\nabla_{\\theta}\\left(\\mathcal{L}_{\\mathrm{samp}} + \\lambda \\mathcal{L}_{\\mathrm{entr}}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.713, + 0.645, + 0.726 + ], + "angle": 0, + "content": "Update \\(\\lambda\\) (Sec.4)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.727, + 0.564, + 0.738 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.721, + 0.787 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We set number of queries as \\( N_{\\mathcal{Q}} = 1000k \\) and \\( N_{\\Omega} = 10k \\). Our MLP follows the architecture in [48], and similarly to the latter we set \\( K = 51 \\). We train for \\( N_{it} = 40k \\) iterations on a Nvidia RTX A6000 GPU using the Adam optimizer with learning rate \\( \\alpha = 0.001 \\). Our training takes roughly 5 minutes for a 1024 sized input point cloud. The regularization parameter \\( \\lambda \\) (Equation 6) follows an exponential decay" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21732" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.088, + 0.432, + 0.2 + ], + "angle": 0, + "content": "
CD1CD2NCFS
SPSR [36]2.340.2240.740.50
OG-INR [38]1.360.0510.550.55
N-Pull [48]1.160.0740.840.75
G-Pull [16]1.070.0320.700.74
NTPS [15]1.110.0670.880.74
Ours0.760.0200.880.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.209, + 0.47, + 0.239 + ], + "angle": 0, + "content": "Table 1. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.418, + 0.285 + ], + "angle": 0, + "content": "of the form \\(\\exp -\\kappa t\\) where \\(\\kappa\\) is set to \\(1.84\\times 10^{-2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.304, + 0.166, + 0.32 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.332, + 0.474, + 0.725 + ], + "angle": 0, + "content": "To test the efficacy of our approach, we assess our capacity to infer implicit representations of shapes when presented with sparse and noisy point clouds. We employ datasets from standard reconstruction benchmarks, showcasing a diverse array of challenges associated with implicit shape function inference from sparse inputs. Consistently with existing literature, we evaluate the performance of our method by quantifying the accuracy of 3D explicit shape models obtained post-convergence from our MLPs. It is worth noting that our approach entails fitting an MLP independently for each point cloud, operating without reliance on pre-learned priors or additional training data. We compare quantitatively and qualitatively to the state-of-the-art in our problem setting, i.e. unsupervised reconstruction from unoriented point, including deep learning methods N-Pull [48], NTPS [15], G-Pull [16], OG-INR [38], DiGs [5], NDrop [9], SAP [61]. We show results for NSpline [84] even though it requires normals. We also compare to classical Poisson Reconstruction (SPSR [36]). We note that NTPS is the closest method to ours as it focuses specifically on the sparse input case. For comprehensive evaluation, we also include comparisons with feed-forward generalizable methods, namely POCO [8] and CONet [60], alongside the prior-based optimization method On-Surf [49] dedicated to sparse inputs. Unless stated differently, we use the publicly available official implementations of existing methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.175, + 0.755 + ], + "angle": 0, + "content": "5.1. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Following seminal work, we evaluate our method and the competition w.r.t. the ground truth using standard metrics for the 3D reconstruction task. Namely, the L1 Chamfer Distance (CD1) \\((\\times 10^{2})\\), L2 Chamfer Distance (CD2) \\((\\times 10^{2})\\), the Hausdorff distance (HD) and the euclidean distance based F-Score (FS) when ground truth points are available, and finally Normal Consistency (NC) when ground truth normals are available. We detail the expressions of these metrics in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.088, + 0.865, + 0.216 + ], + "angle": 0, + "content": "
CD1CD2NCFS
POCO [8]0.3080.0020.9340.981
CONet [60]1.260.0280.8290.599
On-Surf [49]0.5840.0120.9360.915
SPSR [36]0.7510.0280.8710.839
G-Pull [16]0.4950.0050.8870.945
NTPS [15]0.7370.0150.9430.844
Ours0.2600.0020.9520.974
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.225, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Table 2. Faust [7] reconstructions from sparse noisy unoriented point clouds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.28, + 0.771, + 0.296 + ], + "angle": 0, + "content": "5.2. Datasets and Input Definitions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.606 + ], + "angle": 0, + "content": "ShapeNet [14] consists of various instances of 13 different synthetic 3D object classes. We follow the train/test splits defined in [84]. We generate noisy input point clouds by sampling 1024 points from the meshes and adding Gaussian noise of standard deviation 0.005 following the literature (e.g. [8, 60]). For brevity we show results on classes Tables, Chairs and Lamps. Faust [7] consists of real scans of 10 human body identities in 10 different poses. We sample sets of 1024 points from the scans as inputs. 3D Scene [92] contains large scale complex real world scenes obtained with a handheld commodity range sensor. We follow [15, 34, 48] and sample our input point clouds with a sparse density of \\(100\\mathrm{per}\\mathrm{m}^2\\), and we report performance similarly for scenes Burghers, Copyroom, Lounge, Stonewall and Totempole. Surface Reconstruction Benchmark (SRB) [83] consists of five object scans, each with different challenges such as complex topology, high level of detail, missing data and varying feature scales. We sample 1024 points from the scans for the sparse input experiment, and we also experiment using the dense inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.757, + 0.633 + ], + "angle": 0, + "content": "5.3. Object Level Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.894, + 0.822 + ], + "angle": 0, + "content": "We conduct the reconstruction of ShapeNet [14] objects using sparse and noisy point clouds. Table 1 and Figure 2 present a quantitative and qualitative comparison with other methods. Our approach consistently outperforms the competition across all metrics, as seen in the visual superiority of our reconstructions. We excel in capturing fine structures and details with greater fidelity. While NTPS demonstrates good overall coarse reconstructions, its thin plate spline smoothing prior appears to limit its expressivity. Additionally, OG-INR struggles to achieve satisfactory results in the sparse and noisy regime, despite its effective Octree-based sign field guidance in denser settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.837, + 0.848 + ], + "angle": 0, + "content": "5.4. Real Articulated Shape Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We undertake the reconstruction of Faust [7] human shapes using sparse and noisy point clouds. Table 2 and Figure 3 provide a numerical and qualitative comparison with other" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21733" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.086, + 0.775, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.489, + 0.731, + 0.504 + ], + "angle": 0, + "content": "Figure 2. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.727 + ], + "angle": 0, + "content": "methods. Our approach surpasses the alternatives across all metrics, with visually superior reconstructions, as especially noticeable at the extremities of the body. Analogous to the fine structures in the ShapeNet experiment, these areas pose challenges due to limited input point cloud samples, making shape prediction inherently difficult and ambiguous. NTPS reconstructions, similarly, exhibit coarser and less detailed results on this dataset. It is worth highlighting that our method outperforms not only feed-forward generalizable approaches POCO and CONet but also the prior-based optimization method On-Surf. These methods rely on priors trained on ShapeNet, thereby constraining their generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.365, + 0.755 + ], + "angle": 0, + "content": "5.5. Real Scene Level Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In accordance with [15], we present the reconstruction outcomes on the 3D Scene [92] dataset derived from spatially sparse point clouds. Table 3 provides a summary of numerical results, incorporating results for methods NTPS, N-Pull, SAP, NDrop, and NSpline as reported in the state-of-the-art method NTPS. Our method excels in this benchmark, surpassing the competition in most metrics. Notably, our baseline N-Pull exhibits more conspicuous failures in this extensive sparse setup. Figure 5 offers qualitative compar" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.53, + 0.774, + 0.545 + ], + "angle": 0, + "content": "isons against baselines N-Pull and SPSR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.665, + 0.574 + ], + "angle": 0, + "content": "6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Loss. Our approach centers around aligning the decision boundary of our occupancy function with the shape surface through a loss function based on uncertainty sampling, denoted as \\(\\mathcal{L}_{\\mathrm{samp}}\\) (Equation 3). To facilitate and enhance our learning process, we impose constraints on the network to minimize uncertainty almost everywhere in space while maximizing it at the decision boundary. Entropy serves as our chosen uncertainty measure. However, other measures can be considered. We justify this choice in Table 4. We compare our entropy-based regularization to margin-based regularization (Margin reg), using the absolute value of our margin function \\(U_{\\theta}\\) (Equation 3) as an uncertainty measure instead of entropy. Additionally, we compare it to the sign-agnostic occupancy supervision strategy introduced in [75] denoted as SA-Occ. Although this strategy fails to converge in the absence of priors, we demonstrate its beneficial use as a regularization in place of our entropy-based approach. It is noteworthy that while margin-based regularization (Margin reg) yields better results, it is still outperformed by our entropy-based regularization. Figure 4 shows an example of validation plots, displaying the role of our regularization" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21734" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.925, + 0.196 + ], + "angle": 0, + "content": "
BurghersCopyroomLoungeStonewallTotempleMean
CD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NC
SPSR [36]0.1780.2050.8740.2250.2860.8610.2800.3650.8690.3000.4800.8660.5881.6730.8790.3140.6020.870
NDrop [9]0.2000.1140.8250.1680.0630.6960.1560.0500.6630.1500.0810.8150.2030.1390.8440.1750.0890.769
N-Pull [48]0.0640.0080.8980.0490.0050.8280.1330.0380.8470.0600.0050.9100.1780.0240.9080.0970.0160.878
SAP [61]0.1530.1010.8070.0530.0090.7710.1340.0330.8130.0700.0070.8670.4740.3820.7250.1510.1000.797
NSpline [84]0.1350.1230.8910.0560.0230.8550.0630.0390.8270.1240.0910.8970.3780.7680.8920.1510.2090.88
NTPS [15]0.0550.0050.9090.0450.0030.8920.1290.0220.8720.0540.0040.9390.1030.0170.9350.0770.0100.897
Ours0.0220.0010.8710.0410.0120.8120.0210.0010.8700.0280.0030.9310.0260.0010.9360.0270.0030.886
" + }, + { + "type": "table_caption", + "bbox": [ + 0.291, + 0.204, + 0.678, + 0.219 + ], + "angle": 0, + "content": "Table 3. 3D Scene [92] reconstructions from sparse point clouds." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.241, + 0.468, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.717, + 0.471, + 0.745 + ], + "angle": 0, + "content": "Figure 3. Faust [7] reconstructions from sparse unoriented point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.809 + ], + "angle": 0, + "content": "in avoiding overfitting. The baseline in this example is SDF method N-Pull [48]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Point Cloud Density. We utilize the SRB [83] benchmark to evaluate the performance of our method across varying point cloud densities. Table 5 presents comparative results for sparse (1024) and dense input point clouds. We include results for the competition from OG-INR in the dense setting. Our method outperforms the competition in the sparse" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.247, + 0.795, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Figure 4. CD1 distance to GT for reconstructions of shape Gargoyle of benchmark SRB [83] from a sparse unoriented point cloud." + }, + { + "type": "table", + "bbox": [ + 0.585, + 0.434, + 0.816, + 0.512 + ], + "angle": 0, + "content": "
CD1NC
N-Pull [48] (baseline)1.100.85
\\( \\mathcal{L}_{\\text{samp}} \\) (Equ.3)1.020.86
\\( \\mathcal{L}_{\\text{samp}} + \\text{Margin reg} \\)0.800.88
\\( \\mathcal{L}_{\\text{samp}} + \\text{SA-Occ [75] reg} \\)0.950.88
Ours0.770.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Table 4. Ablation of our method on the Tables class of ShapeNet [14]." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.565, + 0.826, + 0.667 + ], + "angle": 0, + "content": "
SparseDense
CD1HDCD1HD
SPSR [36]2.2721.11.2522.59
DiGs [5]0.686.050.193.52
OG-INR [38]0.857.100.204.06
NTPS [15]0.737.78--
N-Pull [48]0.588.900.234.46
Ours0.496.040.203.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.676, + 0.818, + 0.691 + ], + "angle": 0, + "content": "Table 5. Ablation of point cloud density." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.854 + ], + "angle": 0, + "content": "case and performs comparably to the state-of-the-art in the dense case. The improvement over our baseline is substantial for both sparse and dense inputs, as highlighted visually in Figure 6, where we showcase reconstructions for both sparse and dense cases. Notably, we achieve better topologies in the sparse case and enhanced, more accurate details in the dense case, as indicated by the red boxes. These results underscore the utility and advantages of our contribution, even in the dense setting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Running Time and Performance against Number of Queries and Input Points. Fig.7 shows our best performance over time (minutes) for various input point cloud" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.203, + 0.089, + 0.773, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.411, + 0.713, + 0.425 + ], + "angle": 0, + "content": "Figure 5. 3D Scene [92] reconstructions from sparse unoriented point clouds." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.449, + 0.468, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.6, + 0.468, + 0.629 + ], + "angle": 0, + "content": "Figure 6. SRB [83] reconstructions from sparse and dense unoriented inputs." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.651, + 0.241, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.745, + 0.256, + 0.77 + ], + "angle": 0, + "content": "Figure 7. Performance over time (varying input size)." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.652, + 0.449, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.745, + 0.465, + 0.77 + ], + "angle": 0, + "content": "Figure 8. Performance over time (varying query set size)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.468, + 0.884 + ], + "angle": 0, + "content": "sizes. Models using larger inputs reach the baseline performance at earlier times. Fig.8 shows our best performance over time while decreasing the total number of query samples. Performance does not deteriorate drastically under less queries. These Experiments were performed using shape Gargoyle of benchmark SRB." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Robustness to Noise We evaluate here (Tab.6) our method" + }, + { + "type": "table", + "bbox": [ + 0.588, + 0.449, + 0.809, + 0.49 + ], + "angle": 0, + "content": "
noise std. dev.00.0050.025
CD1 ↓0.560.772.16
NC ↑0.930.890.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.624, + 0.5, + 0.768, + 0.513 + ], + "angle": 0, + "content": "Table 6. Noise ablation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.543, + 0.892, + 0.604 + ], + "angle": 0, + "content": "with various noise levels on class Table of ShapeNet. Performance remains reasonable even under heavily corrupted inputs, which confirms our qualitative superiority to competition in Fig.2, and our resilience to noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.623, + 0.636 + ], + "angle": 0, + "content": "7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.647, + 0.892, + 0.722 + ], + "angle": 0, + "content": "We notice that excess regularization can cause reconstructions to be overly smooth in some cases, which can affect our NC numbers is some benchmarks. This can be improved by tuning the hyperparameters of the method per benchmark or per object/scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.619, + 0.754 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We presented a method for implicit shape reconstruction from sparse, noisy and unoriented point cloud. Our results demonstrate that occupancy fields offer and effective implicit shape representation for this task, and that they can be learned successfully even in the unsupervised setting. Under our carefully designed losses, the occupancy outperforms state-of-the-art methods based on SDF, as-well-as other occupancy based baselines that we show in ablation studies." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21736" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. CG, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.186 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.468, + 0.215 + ], + "angle": 0, + "content": "[3] Matan Atzmon and Yaron Lipman. Sald: Sign agnostic learning with derivatives. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.468, + 0.259 + ], + "angle": 0, + "content": "[4] Adi Ben-Israel. A newton-raphson method for the solution of systems of equations. In Journal of Mathematical analysis and applications, 1966. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.468, + 0.315 + ], + "angle": 0, + "content": "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In CVPR, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. TVCG, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[7] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In CVPR, 2014. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[8] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.462, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[9] Alexandre Boulch, Pierre-Alain Langlois, Gilles Puy, and Renaud Marlet. Needrop: Self-supervised shape representation from sparse point clouds using needle dropping. In 2021 International Conference on 3D Vision (3DV), pages 940-950. IEEE, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[10] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3d objects with radial basis functions. In SIGGRAPH, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.468, + 0.617 + ], + "angle": 0, + "content": "[11] Frédéric Casals and Joachim Giesen. Effective Computational Geometry for Curves and Surfaces. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.468, + 0.673 + ], + "angle": 0, + "content": "[12] Rohan Chabra, Jan E Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[14] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[15] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[16] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[17] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Julian Chibane and Gerard Pons-Moll. Implicit feature networks for texture completion from partial 3d data. In European Conference on Computer Vision, pages 717-725. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[21] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[22] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[23] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[25] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[26] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[27] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[28] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[29] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[30] Gaël Guennebaud and Markus Gross. Algebraic point set surfaces. In ACM siggraph 2007 papers, pages 23–es. 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] John C Hart. Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21737" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[33] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[34] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.251 + ], + "angle": 0, + "content": "[35] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[36] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. TOG, 2013. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[37] Ravikrishna Kolluri. Provably good moving least squares. TALG, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.312, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[38] Chamin Hewa Koneputugodage, Yizhak Ben-Shabat, and Stephen Gould. Octree guided unoriented surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16717-16726, 2023. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[39] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3d shapes with dynamic code clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12840-12850, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[40] Siyou Lin, Dong Xiao, Zuoqiang Shi, and Bin Wang. Surface reconstruction from point clouds without normals by parametrizing the gauss formula. ACM Transactions on Graphics, 42(2):1-19, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[41] Stefan Lionar, Daniil Emtsev, Dusan Svilarkovic, and Songyou Peng. Dynamic plane convolutional occupancy networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1829-1838, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[42] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[43] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[44] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization. arXiv preprint arXiv:2202.08345, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[45] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[47] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, 1987. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[48] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In ICML, 2021. 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[49] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6315-6325, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[50] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6326-6337, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[51] Corentin Mercier, Thibault Lescoat, Pierre Roussillon, Tamy Boubekeur, and Jean-Marc Thiery. Moving level-of-detail surfaces. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[52] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.423, + 0.892, + 0.465 + ], + "angle": 0, + "content": "[54] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[55] Amine Ousafi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[56] Amine Ousafi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[57] David Palmer, Dmitriy Smirnov, Stephanie Wang, Albert Chern, and Justin Solomon. Deepcurrents: Learning implicit representations of shapes with boundaries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18665-18675, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[59] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[60] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[61] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21738" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[63] Marie-Julie Rakotosaona, Noam Aigerman, Niloy Mitra, Maks Ovsjanikov, and Paul Guerrero. Differentiable surface triangulation. In SIGGRAPH Asia, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[65] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In NeurIPS, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[66] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[67] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.437, + 0.33 + ], + "angle": 0, + "content": "[68] Burr Settles. Active learning literature survey. 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[69] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.371, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[70] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[71] Vincent Sitzmann, Eric R Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[72] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[73] Vincent Sitzmann, Semon Rezchikov, William T Freeman, Joshua B Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.58, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[74] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.65, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[75] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. Sa-convonet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6504–6513, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[76] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[77] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. Patchnets: Patch-based generalizable deep implicit 3d shape representations. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[78] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[81] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. TOG, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[82] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. Advances in Neural Information Processing Systems, 34:2810-2822, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[83] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In CVPR, 2019. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[84] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural splines: Fitting 3d surfaces with infinitely-wide neural networks. In CVPR, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[85] Francis Williams, Zan Gojcic, Sameh Khamis, Denis Zorin, Joan Bruna, Sanja Fidler, and Or Litany. Neural fields as learnable kernels for 3d reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[86] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T Freeman, and Joshua B Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[87] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[88] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[89] Mohsen Yavartanoo, Jaeyoung Chung, Reyhaneh Neshatavar, and Kyoung Mu Lee. 3dias: 3d shape reconstruction with implicit algebraic surfaces. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[90] Tjalling J Ypma. Historical development of the newton-raphson method. SIAM review, 37(4):531-551, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[91] Junsheng Zhou, Baorui Ma, Liu Yu-Shen, Fang Yi, and Han Zhizhong. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[92] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics (ToG), 32(4):1-8, 2013. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[93] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3d-prnn: Generating shape primitives with recurrent neural networks. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21739" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_origin.pdf b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..329b05c6c0a1d0584a5e93560b4e1239f421936c --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/421a640a-85ac-4767-9ea6-13d292c25e4d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fc010613237ba7816439d79247654eac8bb4cfed3c3bc903df922271f8a2ab9 +size 5454671 diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/full.md b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/full.md new file mode 100644 index 0000000000000000000000000000000000000000..988ae418361b796f3ec1ee58dc9d397ac9932b1d --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/full.md @@ -0,0 +1,352 @@ +# Unsupervised Occupancy Learning from Sparse Point Cloud + +Amine Ousafi Adnane Boukhayma +Inria, Univ. Rennes, CNRS, IRISA, M2S, France + +# Abstract + +Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from 3D point clouds in the absence of ground truth supervision remains a very challenging task. In this paper, we propose a method to infer occupancy fields instead of SDFs as they are easier to learn from sparse inputs. We leverage a margin-based uncertainty measure to differentiably sample from the decision boundary of the occupancy function and supervise the sampled boundary points using the input point cloud. We further stabilise the optimization process at the early stages of the training by biasing the occupancy function towards minimal entropy fields while maximizing its entropy at the input point cloud. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve implicit shape inference with respect to baselines and the state-of-the-art using synthetic and real data. + +# 1. Introduction + +Capturing full 3D shape from limited and corrupted data is a long standing problem. In this regard, one of the data type instances that has received increasing attention and investigation from computer vision, graphics and machine learning is point clouds. This interest emanates primarily from the ubiquity of this light, although topologically incomplete 3D representation, either as acquired e.g. through industrial and commodity depth sensors, or as an intermediate representation within computational photogrammetry [66, 67] pipelines for instance. While classical optimization methods such as Poisson Reconstruction [36] or Moving Least Squares [30] can be effective with dense, clean point sets and accurate normal pre-estimations, recent deep learning-based alternatives provide more robust predictions, particularly for noisy and sparse inputs, eliminating the need for + +normal data in many cases. + +Several methods rely on priors learned from large fully labeled data such as the synthetic dataset ShapeNet [14]. However, this strategy entails computationally expensive trainings, and the resulting models can still be prone to out-of-distribution generalization issues, as pointed in [15, 55], whether caused by change in the input size or domain shift. For instance, Table 2 shows that our unsupervised method outperforms supervised generalizable models when testing on data that is sparser and different in nature from their training corpus. Hence, it is important to design learning frameworks that can lead to robust reconstruction under such extreme constraints. + +Neural implicit representations (INR) were established recently as a powerful representation for 3D shape, generally taking the form of coordinate based MLPs. They have been applied to model various shape presentations, the most popular ones of which remain arguably signed distance functions (SDF) [58] and binary occupancy fields [52] for watertight shapes. For the task of point cloud reconstruction, there is a wide literature on learning SDFs from dense labels, as well as from point clouds solely (unsupervised), either in the generalizable setting or in instance specific optimization. However, apart from its applications in the supervised generalizable setting [8, 20, 60], it seems that occupancy learning from point sets has not been as widely considered in other scenarios. In particular, to the best of our knowledge, using occupancy fields to learn shapes from sparse noisy unoriented point clouds (i.e. unsupervisedly) has not been explored thoroughly. This is even the more intriguing as it seems to be easier to learn a binary classification task as opposed to regressing a continuous field that must additionally respect specific properties [28]. + +Concordantly, we introduce occupancy in this work for learning shape from a sparse point cloud. While other representations such as SDF can readily incorporate surface sample supervision directly (e.g. [28]), among other strategies, it is not entirely straightforward to define such supervision for occupancy fields. Hence, we propose a novel strategy for our training that combines a loss applied near the boundary decision, harnessing the input point samples, and an accompanying regularization. For the first loss, we encourage + +the decision boundary of our occupancy to align with the point cloud. We achieve this by supervising an uncertainty sampling mechanism for our occupancy field with the point cloud samples. We approximate this sampling through root-finding of a margin function associated to the field. As regularization, we propose to minimize the uncertainty of our field, while maximizing it near the decision boundary, via the entropy of our occupancy. + +Through extensive experiments under several real and synthetic benchmarks for object, non-rigid and scene level shape reconstruction, our results show that our method outperforms the existing state-of-the-art literature in reconstruction from sparse point cloud, using standard metrics and through superior visual results. Our ablation studies validate our design choices and showcase the importance of the components of the supervision scheme that we propose. + +# 2. Related Work + +Shape Representations in Deep Learning Shapes can be depicted in deep learning through either intrinsic or extrinsic representations. Intrinsic representations focus on discriminating the shape itself. When explicitly implemented, for instance, using tetrahedral or polygonal meshes [35, 79] or point clouds [25], the output topology is predefined, thereby restricting the range of shapes that can be generated. Among other intrinsic representations, 2D patches [23, 29, 83] may introduce discontinuities, while the simplicity of shape primitives like cuboids [78, 93], planes [43], and Gaussians [26] constrains their expressiveness. On the other hand, extrinsic shape representations model the entire space containing the scene or object of interest. Voxel grids [86, 87] are the most popular, serving as a direct extension of 2D pixels to the 3D domain. However, their capacity is constrained by the memory cost associated with cubic resolution. Sparse representations, such as octrees [64, 76, 81], can mitigate this issue to some extent. + +Implicit Neural Shape Representations Implicit Neural Shape Representations have recently emerged as a significant approach for modeling extrinsic shape and radiance fields (e.g. [13, 33, 53, 80, 88]). These representations address many of the limitations associated with classical representations by showcasing the ability to represent shapes with arbitrary topologies at virtually infinite resolution. Typically, they are parameterized with MLPs, which map spatial locations or features to properties such as occupancy [52], signed [58], or unsigned [21, 91] distances relative to the target shape. The level-set derived from these MLPs can be visualized through techniques like ray marching [31] or tessellated into an explicit shape using methods like Marching Cubes [47]. Another notable line of research involves the development of hybrid implicit/explicit representations [19, 22, 57, 89], primarily based on differentiable space partitioning. To simultaneously represent + +collections of shapes, implicit neural models necessitate conditioning mechanisms. These mechanisms encompass features and latent code concatenation, batch normalization, hypernetworks [17, 70, 72, 73, 82], and gradient-based meta-learning [54, 71]. Concatenation-based conditioning was initially implemented using single global latent codes [18, 52, 58] and has been further refined with the incorporation of local features [20, 24, 27, 34, 39, 60, 74, 77]. + +Reconstruction from Point Clouds Classical approaches to reconstruction include combinatorical methods where the shape is defined based on input point clouds through space partitioning, employing techniques such as alpha shapes [6], Voronoi diagrams [1], or triangulation [11, 45, 63]. Alternatively, the input samples can contribute to defining an implicit function, with its zero level set representing the target shape. This is achieved through global smoothing priors [40, 84, 85], such as radial basis functions [10] and Gaussian kernel fitting [65], or local smoothing priors like moving least squares [30, 37, 46, 51]. Another approach involves solving a boundary-conditioned Poisson equation [36]. Recent literature suggests parameterizing these implicit functions with deep neural networks and learning their parameters through gradient descent, either in a supervised or unsupervised manner. + +Supervised Implicit Neural Reconstruction In supervised methods, there is an assumption of having labeled training data, typically in the form of dense samples containing ground truth shape information. Auto-decoding techniques [12, 34, 39, 58, 77] necessitate test-time optimization to adapt to a new point cloud. On the other hand, encoder-decoder methods allow for swift feed-forward inference. Initially introduced for this purpose, Pooling-based set encoders [18, 27, 52] like PointNet [62] have been found to underfit the context. State-of-the-art performance is achieved by convolutional encoders, utilizing local features defined either in explicit volumes and planes [20, 41, 60, 61] or solely at the input points [8, 56]. Peng et al. [61] proposed a differentiable Poisson solving layer that efficiently converts predicted normals into an indicator function grid. However, its applicability is limited to small scenes due to the cubic memory requirement in grid resolution. The work in [32, 85] suggests constructing generalizable models with implicit decoder functions as a kernel regression. Despite these advancements, many of the generalizable methods still face challenges related to generalization. + +Unsupervised Implicit Neural Reconstruction For unsupervised approaches, a neural network is typically fitted to the input point cloud without additional information. Convergence improvements can be achieved through regularization techniques, such as the spatial gradient constraint based on the Eikonal equation introduced by Gropp et al. [28], a spatial divergence constraint as described in [5], and + +Lipschitz regularization on the network [44]. Periodic activations were introduced in [72]. Lipman [42] learns a function that converges to occupancy, while its log transform converges to a distance function. Atzmon et al. [2] learn an SDF from unsigned distances, further supervising the spatial gradient of the function with normals [3]. Ma et al. [48] express the nearest point on the surface as a function of the neural signed distance and its gradient. They also utilize self-supervised local priors to handle very sparse inputs [49] and enhance generalization [50]. [38] guides the implicit field learning with an Octree based labelling. [9] predicts occupancy fields by learning whether a dropped needle goes across the surface or no. [15] learns a surface parameterization leveraged to provide additional coarse surface supervision to the shape network. In [84], infinitely wide shallow MLPs are learned as random feature kernels using points and their normals. However, most of the methods mentioned above encounter challenges when dealing with sparse and noisy input, primarily due to the lack of supervision. Differently from this literature, we propose here to apply occupancy to unsupervised sparse point cloud neural fitting for the first time to the best of our knowledge. + +# 3. Method + +Given a sparse noisy unoriented input point cloud $\mathcal{P} \subset \mathbb{R}^{3 \times N}$ , our goal is to recover a 3D watertight shape surface $S$ that best explains this observation, i.e. points from $\mathcal{P}$ approximating noisy samples from $S$ . + +Based on the proven success of implicit neural representations, we propose to address this problem by fitting an implicit neural shape function to the scarce observation, which amounts to learning an MLP mapping coordinates to shape attributes. In this context, dubbed unsupervised, signed distance fields are the representation of choice in the community so far (e.g. [15, 16, 38, 48, 61, 84]). Oddly enough, binary occupancy fields were not explored extensively yet as a representation for this context to the best of our knowledge, even though in theory, it seems easier to learn occupancy than SDFs. In fact, it is safe to assume that binary classification is an easier task than continuous regression in general. Besides, we can see intuitively that occupancy is only the sign part of the SDF information. Additionally, occupancy fields do not seem to stipulate any strong structural requirement, as it is the case for distance fields, whose gradient ought to satisfy the Eikonal constraint [28] for instance. Hence, we propose here to learn a binary occupancy field. + +When labeled samples inside and outside the shape are available, such a neural occupancy field could be learned in this supervised setting via a standard cross-entropy classification loss (e.g. [52]). However when only surface points are available, as in our case $(\mathcal{P})$ , it is not entirely clear how a successful learning can be achieved using such occupancy + +decision boundary samples alone. We derive in the following a methodology to approach this task. + +# 3.1. Learning Occupancy Through Margin Uncertainty Sampling + +We are interested in learning the following posterior $P(y|\mathbf{x},\theta)$ , where $\mathbf{x}$ represents query points in $\mathbb{R}^3$ , and label $y \in \{0,1\}$ stands for binary shape occupancy. $P$ is parameterized with a Softmax activated MLP $\theta$ similarly to [8]. Since we do not have points with their occupancy labels, all we can learn with is the input point cloud samples. + +Let us consider the following margin function: + +$$ +U _ {\theta} (\mathbf {x}) = P (y = 1 | \mathbf {x}, \theta) - P (y = 0 | \mathbf {x}, \theta). \tag {1} +$$ + +Borrowing from active learning terminology, finding the roots of this expression is akin to a form of uncertainty sampling from our occupancy field (margin sampling [68]). Intuitively, the smaller the margin of a point the more uncertain its shape prediction. In this particular case, the margin function zeroes correspond to the samples with upmost level of uncertainty. + +Given that our input point cloud points are samples from the surface, i.e. the ground truth occupancy function decision boundary, we propose to learn our occupancy field by supervising uncertainty sampling with the point cloud, i.e. minimizing the distance between uncertain samples and their nearest point cloud samples. We define our uncertainty sampling as root finding of our margin function. We Initialize this root finding near the target surface. We hypothesis that training our occupancy field by making its uncertain samples coincide with the surface through minimal amount of steps of a root finding algorithm will encourage it to converge quickly and decisively. + +We start by generating a pool of query points near $\mathcal{P}$ , i.e. near the ground-truth surface, by sampling around the points, i.e. $\{\mathbf{q} \sim \mathcal{N}(\mathbf{p}, \sigma_{\mathbf{p}}\mathbf{I}_3)\}$ . Here, $\sigma_{\mathbf{p}}$ is chosen as the maximal euclidean distance to the $K$ nearest points to $\mathbf{q}$ in $\mathcal{P}$ (as in [28]). We subsequently recompute the nearest point $\mathbf{p}$ in $\mathcal{P}$ to each sample $\mathbf{q}$ , thus forming the following set of training pairs: + +$$ +\mathcal {Q} := \left\{\left(\mathbf {q}, \mathbf {p}\right), \mathbf {p} = _ {\mathbf {v} \in \mathcal {P}} \| \mathbf {v} - \mathbf {q} \| _ {2} \right\}. \tag {2} +$$ + +Given a pair $(\mathbf{q},\mathbf{p})$ in $\mathcal{Q}$ , we train by bounding one step of Newton-Raphson [90] root finding on our occupancy margin, initialized at $\mathbf{q}$ , to result in an uncertain sample near $\mathbf{p}$ in mean squared error terms. + +Let us recall that a Generalized Newton [4] (Theorem 1, with $m = 1$ and $n = 3$ ) iteration updates query point $\mathbf{q}$ accordingly: $\mathbf{q} - \nabla U_{\theta}(\mathbf{q})^{\dagger}U_{\theta}(\mathbf{q})$ , where $\nabla U_{\theta}(\mathbf{q})^{\dagger}$ is the Moore-Penrose pseudoinverse of Jacobian $\nabla U_{\theta}(\mathbf{q})$ . As $U_{\theta}$ is a scalar function $(\mathbb{R}^{3}\to \mathbb{R})$ , $\nabla U_{\theta}$ is a row 3-vector (we can use it or its transpose interchangeably), and the pseudoinverse writes: $\nabla U_{\theta}(\mathbf{q})^{\dagger} = \frac{\nabla U_{\theta}(\mathbf{q})}{||\nabla U_{\theta}(\mathbf{q})||_2^2}$ . + +In conclusion, our margin uncertainty sampling loss can be expressed as follows: + +$$ +\mathcal {L} _ {\mathrm {s a m p}} (\theta , \mathcal {Q}) = \underset {(\mathbf {q}, \mathbf {p}) \sim \mathcal {Q}} {\mathbb {E}} | | \mathbf {q} - U _ {\theta} (\mathbf {q}) \cdot \frac {\nabla U _ {\theta} (\mathbf {q})}{| | \nabla U _ {\theta} (\mathbf {q}) | | _ {2} ^ {2}} - \mathbf {p} | | _ {2} ^ {2}, \tag {3} +$$ + +where spatial gradient $\nabla U_{\theta}$ can be computed efficiently through automatic differentiation (e.g. PyTorch [59]). + +This loss offers two primary advantages. Firstly, it serves as a form of supervision for the occupancy of the query points. The sign of the margin function differs on opposite sides of the decision boundary. Consequently, based on this sign, a Newton-Raphson iteration will move points along or against the direction of its gradient resulting in an occupancy field whose decision boundary is aligned with the input point cloud. Additionally, this loss operates as a smoothness constraint as Newton-Raphson root finding relies on approximating the function at the roots by its first-order approximation on the initial points $\mathbf{q}$ , hence smoothing the margin function around the input points. + +# 3.2. Entropy Based Regularization + +To ensure better convergence of our training, we use the MLP initialization in [2], while adapting it such that the last layer predicts binary occupancy for an $r$ -radius sphere. + +Furthermore, to ease and improve our learning, we suggest to urge our occupancy field to have minimal uncertainty almost everywhere in space. We use the entropy [69] as an uncertainty measure here. However, this is the opposite behavior that we expect from our network to display at the decision boundary i.e. near the ground-truth surface. Hence, we propose to maximize the uncertainty (i.e. entropy) conversely at the input point cloud $\mathcal{P}$ simultaneously. + +Given a random variable $X$ , Shannon entropy [69] is defined as follows: + +$$ +\mathbb {H} (X) = - \mathbb {E} \log (P (X)). \tag {4} +$$ + +We create a set $\Omega \subset \mathbb{R}^3$ of query points $\mathbf{x}$ by sampling space uniformly within a bounding box of input point cloud $\mathcal{P}$ . As a measure of uncertainty we consider the entropy of the occupancy distribution conditioned on the observed spatial location: + +$$ +\mathcal {L} _ {\mathrm {e n t r}} (\theta , \Omega , \mathcal {P}) = \mathbb {E} _ {\mathbf {x} \sim \Omega} \mathbb {H} (y | \mathbf {x}, \theta) - \mathbb {E} _ {\mathbf {p} \sim \mathcal {P}} \mathbb {H} (y | \mathbf {p}, \theta). \tag {5} +$$ + +This entropy polarization loss acts both as an initialization of the field and a regularization mid-learning, as we empirically found that it performs best when its weight ( $\lambda$ in Equation 6) is reduced progressively in the final loss in the midst of training. + +Finally we can train our model using the following combined empirical risk minimization: + +$$ +\min _ {\theta} \mathcal {L} _ {\text {s a m p}} (\theta , \mathcal {Q}) + \lambda \mathcal {L} _ {\text {e n t r}} (\theta , \Omega , \mathcal {P}). \tag {6} +$$ + +![](images/fd5079e6c96045da732480b6e22c9bd93973f7b80a5dd35f2107de61f7983e88.jpg) +Figure 1. Illustration of our training. Our method learns a neural binary occupancy field without off-surface labels. It uses the combination of a margin uncertainty sampling loss near the surface (Green), maximizing entropy at the input point cloud samples (Red), and minimizing entropy everywhere else (Blue). + +Algorithm 1 shows a summary of our training procedure. We note that once an implicit occupancy field is learned, an explicit triangle mesh $\hat{S}$ can be obtained with it through the Marching Cubes algorithm [47]. + +Algorithm 1 The training procedure of our method. + +Input: Point cloud $\mathcal{P}$ , learning rate $\alpha$ , number of iterations $N_{\mathrm{it}}$ , number of near surface queries $N_{\mathcal{Q}}$ , number of uniform queries $N_{\Omega}$ , number of batch input points $N_{\mathcal{P}}$ , number of uniform queries $N_{\Omega}$ , reg. loss weight $\lambda$ . + +Output: Optimal weights $\theta^{*}$ + +Compute local st. dev. $\{\sigma_{\mathbf{p}} = \max_{\mathbf{v}\in K\mathrm{mn}(\mathbf{p},\mathbf{P})}||\mathbf{v} - \mathbf{p}||_2\}$ + +Generate sets $\mathcal{Q}$ (Sec.3.1) and $\varOmega$ (Sec.3.2). + +Initialize $\lambda$ (Sec.4). + +for $N_{\mathrm{it}}$ times do + +Sample a batch $\mathcal{Q}_b$ of size $N_{\mathcal{Q}}$ from $\mathcal{Q}$ . + +Sample a batch $\varOmega_{b}$ of size $N_{\varOmega}$ from, $\varOmega_{b}$ + +Sample a batch $\mathcal{P}_b$ of size $N_{\mathcal{P}}$ from $\mathcal{P}$ . + +Compute losses $\mathcal{L}_{\mathrm{samp}}(\theta ,\mathcal{Q}_b)$ (Eq.3). + +Compute losses $\mathcal{L}_{\mathrm{entr}}(\theta, \Omega_b, \mathcal{P}_b)$ (Eq.5). + +$\theta \gets \theta -\alpha \nabla_{\theta}\left(\mathcal{L}_{\mathrm{samp}} + \lambda \mathcal{L}_{\mathrm{entr}}\right)$ + +Update $\lambda$ (Sec.4). + +end for + +# 4. Implementation Details + +We set number of queries as $N_{\mathcal{Q}} = 1000k$ and $N_{\Omega} = 10k$ . Our MLP follows the architecture in [48], and similarly to the latter we set $K = 51$ . We train for $N_{it} = 40k$ iterations on a Nvidia RTX A6000 GPU using the Adam optimizer with learning rate $\alpha = 0.001$ . Our training takes roughly 5 minutes for a 1024 sized input point cloud. The regularization parameter $\lambda$ (Equation 6) follows an exponential decay + +
CD1CD2NCFS
SPSR [36]2.340.2240.740.50
OG-INR [38]1.360.0510.550.55
N-Pull [48]1.160.0740.840.75
G-Pull [16]1.070.0320.700.74
NTPS [15]1.110.0670.880.74
Ours0.760.0200.880.83
+ +of the form $\exp -\kappa t$ where $\kappa$ is set to $1.84\times 10^{-2}$ + +# 5. Results + +To test the efficacy of our approach, we assess our capacity to infer implicit representations of shapes when presented with sparse and noisy point clouds. We employ datasets from standard reconstruction benchmarks, showcasing a diverse array of challenges associated with implicit shape function inference from sparse inputs. Consistently with existing literature, we evaluate the performance of our method by quantifying the accuracy of 3D explicit shape models obtained post-convergence from our MLPs. It is worth noting that our approach entails fitting an MLP independently for each point cloud, operating without reliance on pre-learned priors or additional training data. We compare quantitatively and qualitatively to the state-of-the-art in our problem setting, i.e. unsupervised reconstruction from unoriented point, including deep learning methods N-Pull [48], NTPS [15], G-Pull [16], OG-INR [38], DiGs [5], NDrop [9], SAP [61]. We show results for NSpline [84] even though it requires normals. We also compare to classical Poisson Reconstruction (SPSR [36]). We note that NTPS is the closest method to ours as it focuses specifically on the sparse input case. For comprehensive evaluation, we also include comparisons with feed-forward generalizable methods, namely POCO [8] and CONet [60], alongside the prior-based optimization method On-Surf [49] dedicated to sparse inputs. Unless stated differently, we use the publicly available official implementations of existing methods. + +# 5.1. Metrics + +Following seminal work, we evaluate our method and the competition w.r.t. the ground truth using standard metrics for the 3D reconstruction task. Namely, the L1 Chamfer Distance (CD1) $(\times 10^{2})$ , L2 Chamfer Distance (CD2) $(\times 10^{2})$ , the Hausdorff distance (HD) and the euclidean distance based F-Score (FS) when ground truth points are available, and finally Normal Consistency (NC) when ground truth normals are available. We detail the expressions of these metrics in the supplementary material. + +Table 1. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds. + +
CD1CD2NCFS
POCO [8]0.3080.0020.9340.981
CONet [60]1.260.0280.8290.599
On-Surf [49]0.5840.0120.9360.915
SPSR [36]0.7510.0280.8710.839
G-Pull [16]0.4950.0050.8870.945
NTPS [15]0.7370.0150.9430.844
Ours0.2600.0020.9520.974
+ +Table 2. Faust [7] reconstructions from sparse noisy unoriented point clouds. + +# 5.2. Datasets and Input Definitions + +ShapeNet [14] consists of various instances of 13 different synthetic 3D object classes. We follow the train/test splits defined in [84]. We generate noisy input point clouds by sampling 1024 points from the meshes and adding Gaussian noise of standard deviation 0.005 following the literature (e.g. [8, 60]). For brevity we show results on classes Tables, Chairs and Lamps. Faust [7] consists of real scans of 10 human body identities in 10 different poses. We sample sets of 1024 points from the scans as inputs. 3D Scene [92] contains large scale complex real world scenes obtained with a handheld commodity range sensor. We follow [15, 34, 48] and sample our input point clouds with a sparse density of $100\mathrm{per}\mathrm{m}^2$ , and we report performance similarly for scenes Burghers, Copyroom, Lounge, Stonewall and Totempole. Surface Reconstruction Benchmark (SRB) [83] consists of five object scans, each with different challenges such as complex topology, high level of detail, missing data and varying feature scales. We sample 1024 points from the scans for the sparse input experiment, and we also experiment using the dense inputs. + +# 5.3. Object Level Reconstruction + +We conduct the reconstruction of ShapeNet [14] objects using sparse and noisy point clouds. Table 1 and Figure 2 present a quantitative and qualitative comparison with other methods. Our approach consistently outperforms the competition across all metrics, as seen in the visual superiority of our reconstructions. We excel in capturing fine structures and details with greater fidelity. While NTPS demonstrates good overall coarse reconstructions, its thin plate spline smoothing prior appears to limit its expressivity. Additionally, OG-INR struggles to achieve satisfactory results in the sparse and noisy regime, despite its effective Octree-based sign field guidance in denser settings. + +# 5.4. Real Articulated Shape Reconstruction + +We undertake the reconstruction of Faust [7] human shapes using sparse and noisy point clouds. Table 2 and Figure 3 provide a numerical and qualitative comparison with other + +![](images/81eb2ee4c8c9f849f909841468ea9026f2b8119bb29806b877e5643cffc8ae49.jpg) +Figure 2. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds. + +methods. Our approach surpasses the alternatives across all metrics, with visually superior reconstructions, as especially noticeable at the extremities of the body. Analogous to the fine structures in the ShapeNet experiment, these areas pose challenges due to limited input point cloud samples, making shape prediction inherently difficult and ambiguous. NTPS reconstructions, similarly, exhibit coarser and less detailed results on this dataset. It is worth highlighting that our method outperforms not only feed-forward generalizable approaches POCO and CONet but also the prior-based optimization method On-Surf. These methods rely on priors trained on ShapeNet, thereby constraining their generalization ability. + +# 5.5. Real Scene Level Reconstruction + +In accordance with [15], we present the reconstruction outcomes on the 3D Scene [92] dataset derived from spatially sparse point clouds. Table 3 provides a summary of numerical results, incorporating results for methods NTPS, N-Pull, SAP, NDrop, and NSpline as reported in the state-of-the-art method NTPS. Our method excels in this benchmark, surpassing the competition in most metrics. Notably, our baseline N-Pull exhibits more conspicuous failures in this extensive sparse setup. Figure 5 offers qualitative compar + +isons against baselines N-Pull and SPSR. + +# 6. Ablation Studies + +Loss. Our approach centers around aligning the decision boundary of our occupancy function with the shape surface through a loss function based on uncertainty sampling, denoted as $\mathcal{L}_{\mathrm{samp}}$ (Equation 3). To facilitate and enhance our learning process, we impose constraints on the network to minimize uncertainty almost everywhere in space while maximizing it at the decision boundary. Entropy serves as our chosen uncertainty measure. However, other measures can be considered. We justify this choice in Table 4. We compare our entropy-based regularization to margin-based regularization (Margin reg), using the absolute value of our margin function $U_{\theta}$ (Equation 3) as an uncertainty measure instead of entropy. Additionally, we compare it to the sign-agnostic occupancy supervision strategy introduced in [75] denoted as SA-Occ. Although this strategy fails to converge in the absence of priors, we demonstrate its beneficial use as a regularization in place of our entropy-based approach. It is noteworthy that while margin-based regularization (Margin reg) yields better results, it is still outperformed by our entropy-based regularization. Figure 4 shows an example of validation plots, displaying the role of our regularization + +
BurghersCopyroomLoungeStonewallTotempleMean
CD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NC
SPSR [36]0.1780.2050.8740.2250.2860.8610.2800.3650.8690.3000.4800.8660.5881.6730.8790.3140.6020.870
NDrop [9]0.2000.1140.8250.1680.0630.6960.1560.0500.6630.1500.0810.8150.2030.1390.8440.1750.0890.769
N-Pull [48]0.0640.0080.8980.0490.0050.8280.1330.0380.8470.0600.0050.9100.1780.0240.9080.0970.0160.878
SAP [61]0.1530.1010.8070.0530.0090.7710.1340.0330.8130.0700.0070.8670.4740.3820.7250.1510.1000.797
NSpline [84]0.1350.1230.8910.0560.0230.8550.0630.0390.8270.1240.0910.8970.3780.7680.8920.1510.2090.88
NTPS [15]0.0550.0050.9090.0450.0030.8920.1290.0220.8720.0540.0040.9390.1030.0170.9350.0770.0100.897
Ours0.0220.0010.8710.0410.0120.8120.0210.0010.8700.0280.0030.9310.0260.0010.9360.0270.0030.886
+ +![](images/a7eb3af6fe9f64604aa6453410b2577c7ab2ce9c92474e7acb0e6b6217f438e9.jpg) +Figure 3. Faust [7] reconstructions from sparse unoriented point clouds. + +in avoiding overfitting. The baseline in this example is SDF method N-Pull [48]. + +Point Cloud Density. We utilize the SRB [83] benchmark to evaluate the performance of our method across varying point cloud densities. Table 5 presents comparative results for sparse (1024) and dense input point clouds. We include results for the competition from OG-INR in the dense setting. Our method outperforms the competition in the sparse + +![](images/349e8e5adbe7a19cd92bf9511ab81e613da5b972f08820a1c4c2f3ae7b630f97.jpg) +Figure 4. CD1 distance to GT for reconstructions of shape Gargoyle of benchmark SRB [83] from a sparse unoriented point cloud. + +Table 3. 3D Scene [92] reconstructions from sparse point clouds. + +
CD1NC
N-Pull [48] (baseline)1.100.85
\( \mathcal{L}_{\text{samp}} \) (Equ.3)1.020.86
\( \mathcal{L}_{\text{samp}} + \text{Margin reg} \)0.800.88
\( \mathcal{L}_{\text{samp}} + \text{SA-Occ [75] reg} \)0.950.88
Ours0.770.89
+ +Table 4. Ablation of our method on the Tables class of ShapeNet [14]. + +
SparseDense
CD1HDCD1HD
SPSR [36]2.2721.11.2522.59
DiGs [5]0.686.050.193.52
OG-INR [38]0.857.100.204.06
NTPS [15]0.737.78--
N-Pull [48]0.588.900.234.46
Ours0.496.040.203.94
+ +Table 5. Ablation of point cloud density. + +case and performs comparably to the state-of-the-art in the dense case. The improvement over our baseline is substantial for both sparse and dense inputs, as highlighted visually in Figure 6, where we showcase reconstructions for both sparse and dense cases. Notably, we achieve better topologies in the sparse case and enhanced, more accurate details in the dense case, as indicated by the red boxes. These results underscore the utility and advantages of our contribution, even in the dense setting. + +Running Time and Performance against Number of Queries and Input Points. Fig.7 shows our best performance over time (minutes) for various input point cloud + +![](images/265fdd355c746497529ef65089ba3ea07d009c114700d275932561cc9752789f.jpg) +Figure 5. 3D Scene [92] reconstructions from sparse unoriented point clouds. + +![](images/2e610bebc72aa7673d6d946c2ad20600042bc4d0e9d2156767070b607559f6e2.jpg) +Figure 6. SRB [83] reconstructions from sparse and dense unoriented inputs. + +![](images/596d88e2e444ce3b141f827b8d2ae41b6dbc430fe3a70c11008493f1f0a3b57e.jpg) +Figure 7. Performance over time (varying input size). + +![](images/b90b9c197cd2cee7ff89d513c475eac763bd097949670ff3fb8436e9e87ff952.jpg) +Figure 8. Performance over time (varying query set size). + +sizes. Models using larger inputs reach the baseline performance at earlier times. Fig.8 shows our best performance over time while decreasing the total number of query samples. Performance does not deteriorate drastically under less queries. These Experiments were performed using shape Gargoyle of benchmark SRB. + +Robustness to Noise We evaluate here (Tab.6) our method + +
noise std. dev.00.0050.025
CD1 ↓0.560.772.16
NC ↑0.930.890.68
+ +Table 6. Noise ablation. + +with various noise levels on class Table of ShapeNet. Performance remains reasonable even under heavily corrupted inputs, which confirms our qualitative superiority to competition in Fig.2, and our resilience to noise. + +# 7. Limitations + +We notice that excess regularization can cause reconstructions to be overly smooth in some cases, which can affect our NC numbers is some benchmarks. This can be improved by tuning the hyperparameters of the method per benchmark or per object/scene. + +# 8. Conclusion + +We presented a method for implicit shape reconstruction from sparse, noisy and unoriented point cloud. Our results demonstrate that occupancy fields offer and effective implicit shape representation for this task, and that they can be learned successfully even in the unsupervised setting. Under our carefully designed losses, the occupancy outperforms state-of-the-art methods based on SDF, as-well-as other occupancy based baselines that we show in ablation studies. + +# References + +[1] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. CG, 2001. 2 +[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3, 4 +[3] Matan Atzmon and Yaron Lipman. Sald: Sign agnostic learning with derivatives. In ICML, 2020. 3 +[4] Adi Ben-Israel. A newton-raphson method for the solution of systems of equations. In Journal of Mathematical analysis and applications, 1966. 3 +[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In CVPR, 2022. 2, 5, 7 +[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. TVCG, 1999. 2 +[7] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In CVPR, 2014. 5, 7 +[8] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2, 3, 5 +[9] Alexandre Boulch, Pierre-Alain Langlois, Gilles Puy, and Renaud Marlet. Needrop: Self-supervised shape representation from sparse point clouds using needle dropping. In 2021 International Conference on 3D Vision (3DV), pages 940-950. IEEE, 2021. 3, 5, 7 +[10] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3d objects with radial basis functions. In SIGGRAPH, 2001. 2 +[11] Frédéric Casals and Joachim Giesen. Effective Computational Geometry for Curves and Surfaces. 2006. 2 +[12] Rohan Chabra, Jan E Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In ECCV, 2020. 2 +[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[14] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 5, 6, 7 +[15] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7 + +[16] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 3, 5 +[17] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, 2022. 2 +[18] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2 +[19] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[20] Julian Chibane and Gerard Pons-Moll. Implicit feature networks for texture completion from partial 3d data. In European Conference on Computer Vision, pages 717-725. Springer, 2020. 1, 2 +[21] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, 2020. 2 +[22] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In CVPR, 2020. 2 +[23] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2 +[24] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In ECCV, 2020. 2 +[25] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2 +[26] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2 +[27] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In CVPR, 2020. 2 +[28] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 1, 2, 3 +[29] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 2 +[30] Gaël Guennebaud and Markus Gross. Algebraic point set surfaces. In ACM siggraph 2007 papers, pages 23–es. 2007. 1, 2 +[31] John C Hart. Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer, 1996. 2 +[32] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 2 +[33] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 2 +[34] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 2, 5 +[35] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2 +[36] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. TOG, 2013. 1, 2, 5, 7 +[37] Ravikrishna Kolluri. Provably good moving least squares. TALG, 2008. 2 +[38] Chamin Hewa Koneputugodage, Yizhak Ben-Shabat, and Stephen Gould. Octree guided unoriented surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16717-16726, 2023. 3, 5, 7 +[39] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3d shapes with dynamic code clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12840-12850, 2022. 2 +[40] Siyou Lin, Dong Xiao, Zuoqiang Shi, and Bin Wang. Surface reconstruction from point clouds without normals by parametrizing the gauss formula. ACM Transactions on Graphics, 42(2):1-19, 2022. 2 +[41] Stefan Lionar, Daniil Emtsev, Dusan Svilarkovic, and Songyou Peng. Dynamic plane convolutional occupancy networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1829-1838, 2021. 2 +[42] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In ICML, 2021. 3 +[43] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In CVPR, 2018. 2 +[44] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization. arXiv preprint arXiv:2202.08345, 2022. 3 +[45] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In ECCV, 2020. 2 +[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In CVPR, 2021. 2 +[47] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, 1987. 2, 4 +[48] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In ICML, 2021. 3, 4, 5, 7 + +[49] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6315-6325, 2022. 3, 5 +[50] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6326-6337, 2022. 3 +[51] Corentin Mercier, Thibault Lescoat, Pierre Roussillon, Tamy Boubekeur, and Jean-Marc Thiery. Moving level-of-detail surfaces. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2 +[52] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1, 2, 3 +[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[54] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 2 +[55] Amine Ousafi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. In NeurIPS, 2023. 1 +[56] Amine Ousafi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. In 3DV, 2024. 2 +[57] David Palmer, Dmitriy Smirnov, Stephanie Wang, Albert Chern, and Justin Solomon. Deepcurrents: Learning implicit representations of shapes with boundaries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18665-18675, 2022. 2 +[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2 +[59] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4 +[60] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1, 2, 5 +[61] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 2, 3, 5, 7 +[62] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2 + +[63] Marie-Julie Rakotosaona, Noam Aigerman, Niloy Mitra, Maks Ovsjanikov, and Paul Guerrero. Differentiable surface triangulation. In SIGGRAPH Asia, 2021. 2 +[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2 +[65] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In NeurIPS, 2004. 2 +[66] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[67] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1 +[68] Burr Settles. Active learning literature survey. 2009. 3 +[69] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 4 +[70] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2 +[71] Vincent Sitzmann, Eric R Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 2 +[72] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2, 3 +[73] Vincent Sitzmann, Semon Rezchikov, William T Freeman, Joshua B Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2 +[74] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2 +[75] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. Sa-convonet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6504–6513, 2021. 6, 7 +[76] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 2 +[77] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. Patchnets: Patch-based generalizable deep implicit 3d shape representations. In ECCV, 2020. 2 +[78] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 2 +[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2 + +[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2 +[81] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. TOG, 2017. 2 +[82] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. Advances in Neural Information Processing Systems, 34:2810-2822, 2021. 2 +[83] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In CVPR, 2019. 2, 5, 7, 8 +[84] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural splines: Fitting 3d surfaces with infinitely-wide neural networks. In CVPR, 2021. 2, 3, 5, 7 +[85] Francis Williams, Zan Gojcic, Sameh Khamis, Denis Zorin, Joan Bruna, Sanja Fidler, and Or Litany. Neural fields as learnable kernels for 3d reconstruction. In CVPR, 2022. 2 +[86] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T Freeman, and Joshua B Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In NeurIPS, 2016. 2 +[87] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2 +[88] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2 +[89] Mohsen Yavartanoo, Jaeyoung Chung, Reyhaneh Neshatavar, and Kyoung Mu Lee. 3dias: 3d shape reconstruction with implicit algebraic surfaces. In ICCV, 2021. 2 +[90] Tjalling J Ypma. Historical development of the newton-raphson method. SIAM review, 37(4):531-551, 1995. 3 +[91] Junsheng Zhou, Baorui Ma, Liu Yu-Shen, Fang Yi, and Han Zhizhong. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[92] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics (ToG), 32(4):1-8, 2013. 5, 6, 7, 8 +[93] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3d-prnn: Generating shape primitives with recurrent neural networks. In CVPR, 2017. 2 \ No newline at end of file diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/images.zip b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..87404e45be6a5883a9af80ab10cb466b70517185 --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce34e7c64446466698130fd6c15852a7cc1a5dfb8cb5f38caac742d9b7d202d4 +size 575340 diff --git a/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/layout.json b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..044f75343d81981829b961ff0aaa1ea40989ceb6 --- /dev/null +++ b/2024/Unsupervised Occupancy Learning from Sparse Point Cloud/layout.json @@ -0,0 +1,9654 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "text", + "content": "Unsupervised Occupancy Learning from Sparse Point Cloud" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 144, + 414, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 144, + 414, + 171 + ], + "spans": [ + { + "bbox": [ + 176, + 144, + 414, + 171 + ], + "type": "text", + "content": "Amine Ousafi Adnane Boukhayma \nInria, Univ. Rennes, CNRS, IRISA, M2S, France" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 289, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 289, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 289, + 488 + ], + "type": "text", + "content": "Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from 3D point clouds in the absence of ground truth supervision remains a very challenging task. In this paper, we propose a method to infer occupancy fields instead of SDFs as they are easier to learn from sparse inputs. We leverage a margin-based uncertainty measure to differentiably sample from the decision boundary of the occupancy function and supervise the sampled boundary points using the input point cloud. We further stabilise the optimization process at the early stages of the training by biasing the occupancy function towards minimal entropy fields while maximizing its entropy at the input point cloud. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve implicit shape inference with respect to baselines and the state-of-the-art using synthetic and real data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Capturing full 3D shape from limited and corrupted data is a long standing problem. In this regard, one of the data type instances that has received increasing attention and investigation from computer vision, graphics and machine learning is point clouds. This interest emanates primarily from the ubiquity of this light, although topologically incomplete 3D representation, either as acquired e.g. through industrial and commodity depth sensors, or as an intermediate representation within computational photogrammetry [66, 67] pipelines for instance. While classical optimization methods such as Poisson Reconstruction [36] or Moving Least Squares [30] can be effective with dense, clean point sets and accurate normal pre-estimations, recent deep learning-based alternatives provide more robust predictions, particularly for noisy and sparse inputs, eliminating the need for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 201, + 417, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 417, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 417, + 213 + ], + "type": "text", + "content": "normal data in many cases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 215, + 547, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 547, + 360 + ], + "type": "text", + "content": "Several methods rely on priors learned from large fully labeled data such as the synthetic dataset ShapeNet [14]. However, this strategy entails computationally expensive trainings, and the resulting models can still be prone to out-of-distribution generalization issues, as pointed in [15, 55], whether caused by change in the input size or domain shift. For instance, Table 2 shows that our unsupervised method outperforms supervised generalizable models when testing on data that is sparser and different in nature from their training corpus. Hence, it is important to design learning frameworks that can lead to robust reconstruction under such extreme constraints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 362, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 547, + 602 + ], + "type": "text", + "content": "Neural implicit representations (INR) were established recently as a powerful representation for 3D shape, generally taking the form of coordinate based MLPs. They have been applied to model various shape presentations, the most popular ones of which remain arguably signed distance functions (SDF) [58] and binary occupancy fields [52] for watertight shapes. For the task of point cloud reconstruction, there is a wide literature on learning SDFs from dense labels, as well as from point clouds solely (unsupervised), either in the generalizable setting or in instance specific optimization. However, apart from its applications in the supervised generalizable setting [8, 20, 60], it seems that occupancy learning from point sets has not been as widely considered in other scenarios. In particular, to the best of our knowledge, using occupancy fields to learn shapes from sparse noisy unoriented point clouds (i.e. unsupervisedly) has not been explored thoroughly. This is even the more intriguing as it seems to be easier to learn a binary classification task as opposed to regressing a continuous field that must additionally respect specific properties [28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Concordantly, we introduce occupancy in this work for learning shape from a sparse point cloud. While other representations such as SDF can readily incorporate surface sample supervision directly (e.g. [28]), among other strategies, it is not entirely straightforward to define such supervision for occupancy fields. Hence, we propose a novel strategy for our training that combines a loss applied near the boundary decision, harnessing the input point samples, and an accompanying regularization. For the first loss, we encourage" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21729" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "the decision boundary of our occupancy to align with the point cloud. We achieve this by supervising an uncertainty sampling mechanism for our occupancy field with the point cloud samples. We approximate this sampling through root-finding of a margin function associated to the field. As regularization, we propose to minimize the uncertainty of our field, while maximizing it near the decision boundary, via the entropy of our occupancy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 264 + ], + "type": "text", + "content": "Through extensive experiments under several real and synthetic benchmarks for object, non-rigid and scene level shape reconstruction, our results show that our method outperforms the existing state-of-the-art literature in reconstruction from sparse point cloud, using standard metrics and through superior visual results. Our ablation studies validate our design choices and showcase the importance of the components of the supervision scheme that we propose." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 274, + 134, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 134, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 134, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "type": "text", + "content": "Shape Representations in Deep Learning Shapes can be depicted in deep learning through either intrinsic or extrinsic representations. Intrinsic representations focus on discriminating the shape itself. When explicitly implemented, for instance, using tetrahedral or polygonal meshes [35, 79] or point clouds [25], the output topology is predefined, thereby restricting the range of shapes that can be generated. Among other intrinsic representations, 2D patches [23, 29, 83] may introduce discontinuities, while the simplicity of shape primitives like cuboids [78, 93], planes [43], and Gaussians [26] constrains their expressiveness. On the other hand, extrinsic shape representations model the entire space containing the scene or object of interest. Voxel grids [86, 87] are the most popular, serving as a direct extension of 2D pixels to the 3D domain. However, their capacity is constrained by the memory cost associated with cubic resolution. Sparse representations, such as octrees [64, 76, 81], can mitigate this issue to some extent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Implicit Neural Shape Representations Implicit Neural Shape Representations have recently emerged as a significant approach for modeling extrinsic shape and radiance fields (e.g. [13, 33, 53, 80, 88]). These representations address many of the limitations associated with classical representations by showcasing the ability to represent shapes with arbitrary topologies at virtually infinite resolution. Typically, they are parameterized with MLPs, which map spatial locations or features to properties such as occupancy [52], signed [58], or unsigned [21, 91] distances relative to the target shape. The level-set derived from these MLPs can be visualized through techniques like ray marching [31] or tessellated into an explicit shape using methods like Marching Cubes [47]. Another notable line of research involves the development of hybrid implicit/explicit representations [19, 22, 57, 89], primarily based on differentiable space partitioning. To simultaneously represent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "collections of shapes, implicit neural models necessitate conditioning mechanisms. These mechanisms encompass features and latent code concatenation, batch normalization, hypernetworks [17, 70, 72, 73, 82], and gradient-based meta-learning [54, 71]. Concatenation-based conditioning was initially implemented using single global latent codes [18, 52, 58] and has been further refined with the incorporation of local features [20, 24, 27, 34, 39, 60, 74, 77]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 170, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 546, + 361 + ], + "type": "text", + "content": "Reconstruction from Point Clouds Classical approaches to reconstruction include combinatorical methods where the shape is defined based on input point clouds through space partitioning, employing techniques such as alpha shapes [6], Voronoi diagrams [1], or triangulation [11, 45, 63]. Alternatively, the input samples can contribute to defining an implicit function, with its zero level set representing the target shape. This is achieved through global smoothing priors [40, 84, 85], such as radial basis functions [10] and Gaussian kernel fitting [65], or local smoothing priors like moving least squares [30, 37, 46, 51]. Another approach involves solving a boundary-conditioned Poisson equation [36]. Recent literature suggests parameterizing these implicit functions with deep neural networks and learning their parameters through gradient descent, either in a supervised or unsupervised manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 364, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 626 + ], + "type": "text", + "content": "Supervised Implicit Neural Reconstruction In supervised methods, there is an assumption of having labeled training data, typically in the form of dense samples containing ground truth shape information. Auto-decoding techniques [12, 34, 39, 58, 77] necessitate test-time optimization to adapt to a new point cloud. On the other hand, encoder-decoder methods allow for swift feed-forward inference. Initially introduced for this purpose, Pooling-based set encoders [18, 27, 52] like PointNet [62] have been found to underfit the context. State-of-the-art performance is achieved by convolutional encoders, utilizing local features defined either in explicit volumes and planes [20, 41, 60, 61] or solely at the input points [8, 56]. Peng et al. [61] proposed a differentiable Poisson solving layer that efficiently converts predicted normals into an indicator function grid. However, its applicability is limited to small scenes due to the cubic memory requirement in grid resolution. The work in [32, 85] suggests constructing generalizable models with implicit decoder functions as a kernel regression. Despite these advancements, many of the generalizable methods still face challenges related to generalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Unsupervised Implicit Neural Reconstruction For unsupervised approaches, a neural network is typically fitted to the input point cloud without additional information. Convergence improvements can be achieved through regularization techniques, such as the spatial gradient constraint based on the Eikonal equation introduced by Gropp et al. [28], a spatial divergence constraint as described in [5], and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21730" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 336 + ], + "type": "text", + "content": "Lipschitz regularization on the network [44]. Periodic activations were introduced in [72]. Lipman [42] learns a function that converges to occupancy, while its log transform converges to a distance function. Atzmon et al. [2] learn an SDF from unsigned distances, further supervising the spatial gradient of the function with normals [3]. Ma et al. [48] express the nearest point on the surface as a function of the neural signed distance and its gradient. They also utilize self-supervised local priors to handle very sparse inputs [49] and enhance generalization [50]. [38] guides the implicit field learning with an Octree based labelling. [9] predicts occupancy fields by learning whether a dropped needle goes across the surface or no. [15] learns a surface parameterization leveraged to provide additional coarse surface supervision to the shape network. In [84], infinitely wide shallow MLPs are learned as random feature kernels using points and their normals. However, most of the methods mentioned above encounter challenges when dealing with sparse and noisy input, primarily due to the lack of supervision. Differently from this literature, we propose here to apply occupancy to unsupervised sparse point cloud neural fitting for the first time to the best of our knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 346, + 105, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 105, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 105, + 358 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "content": "Given a sparse noisy unoriented input point cloud " + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\subset \\mathbb{R}^{3 \\times N}" + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "content": ", our goal is to recover a 3D watertight shape surface " + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "content": " that best explains this observation, i.e. points from " + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "content": " approximating noisy samples from " + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 366, + 287, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 640 + ], + "type": "text", + "content": "Based on the proven success of implicit neural representations, we propose to address this problem by fitting an implicit neural shape function to the scarce observation, which amounts to learning an MLP mapping coordinates to shape attributes. In this context, dubbed unsupervised, signed distance fields are the representation of choice in the community so far (e.g. [15, 16, 38, 48, 61, 84]). Oddly enough, binary occupancy fields were not explored extensively yet as a representation for this context to the best of our knowledge, even though in theory, it seems easier to learn occupancy than SDFs. In fact, it is safe to assume that binary classification is an easier task than continuous regression in general. Besides, we can see intuitively that occupancy is only the sign part of the SDF information. Additionally, occupancy fields do not seem to stipulate any strong structural requirement, as it is the case for distance fields, whose gradient ought to satisfy the Eikonal constraint [28] for instance. Hence, we propose here to learn a binary occupancy field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "When labeled samples inside and outside the shape are available, such a neural occupancy field could be learned in this supervised setting via a standard cross-entropy classification loss (e.g. [52]). However when only surface points are available, as in our case " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": ", it is not entirely clear how a successful learning can be achieved using such occupancy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "decision boundary samples alone. We derive in the following a methodology to approach this task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 102, + 545, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 102, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 305, + 102, + 545, + 128 + ], + "type": "text", + "content": "3.1. Learning Occupancy Through Margin Uncertainty Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": "We are interested in learning the following posterior " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "P(y|\\mathbf{x},\\theta)" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": " represents query points in " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": ", and label " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": " stands for binary shape occupancy. " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": " is parameterized with a Softmax activated MLP " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": " similarly to [8]. Since we do not have points with their occupancy labels, all we can learn with is the input point cloud samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 205, + 507, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 205, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 507, + 217 + ], + "type": "text", + "content": "Let us consider the following margin function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 223, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 223, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 339, + 223, + 545, + 236 + ], + "type": "interline_equation", + "content": "U _ {\\theta} (\\mathbf {x}) = P (y = 1 | \\mathbf {x}, \\theta) - P (y = 0 | \\mathbf {x}, \\theta). \\tag {1}", + "image_path": "5d15a18577b4f987bbb1603df934ff19edc319e40c62a370788256aa58a0ad00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 326 + ], + "type": "text", + "content": "Borrowing from active learning terminology, finding the roots of this expression is akin to a form of uncertainty sampling from our occupancy field (margin sampling [68]). Intuitively, the smaller the margin of a point the more uncertain its shape prediction. In this particular case, the margin function zeroes correspond to the samples with upmost level of uncertainty." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 469 + ], + "type": "text", + "content": "Given that our input point cloud points are samples from the surface, i.e. the ground truth occupancy function decision boundary, we propose to learn our occupancy field by supervising uncertainty sampling with the point cloud, i.e. minimizing the distance between uncertain samples and their nearest point cloud samples. We define our uncertainty sampling as root finding of our margin function. We Initialize this root finding near the target surface. We hypothesis that training our occupancy field by making its uncertain samples coincide with the surface through minimal amount of steps of a root finding algorithm will encourage it to converge quickly and decisively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": "We start by generating a pool of query points near " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": ", i.e. near the ground-truth surface, by sampling around the points, i.e. " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q} \\sim \\mathcal{N}(\\mathbf{p}, \\sigma_{\\mathbf{p}}\\mathbf{I}_3)\\}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " is chosen as the maximal euclidean distance to the " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " nearest points to " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " (as in [28]). We subsequently recompute the nearest point " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " to each sample " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": ", thus forming the following set of training pairs:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 560, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 560, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 353, + 560, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} := \\left\\{\\left(\\mathbf {q}, \\mathbf {p}\\right), \\mathbf {p} = _ {\\mathbf {v} \\in \\mathcal {P}} \\| \\mathbf {v} - \\mathbf {q} \\| _ {2} \\right\\}. \\tag {2}", + "image_path": "f930e61b4e6ca74dd4527b2699832603a9cfb496fce1d46409dbd8b359ff3984.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": "Given a pair " + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "inline_equation", + "content": "(\\mathbf{q},\\mathbf{p})" + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": ", we train by bounding one step of Newton-Raphson [90] root finding on our occupancy margin, initialized at " + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": ", to result in an uncertain sample near " + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": " in mean squared error terms." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": "Let us recall that a Generalized Newton [4] (Theorem 1, with " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": ") iteration updates query point " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": " accordingly: " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{q} - \\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}U_{\\theta}(\\mathbf{q})" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger}" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": " is the Moore-Penrose pseudoinverse of Jacobian " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\nabla U_{\\theta}(\\mathbf{q})" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "U_{\\theta}" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": " is a scalar function " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{3}\\to \\mathbb{R})" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\nabla U_{\\theta}" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": " is a row 3-vector (we can use it or its transpose interchangeably), and the pseudoinverse writes: " + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\nabla U_{\\theta}(\\mathbf{q})^{\\dagger} = \\frac{\\nabla U_{\\theta}(\\mathbf{q})}{||\\nabla U_{\\theta}(\\mathbf{q})||_2^2}" + }, + { + "bbox": [ + 304, + 628, + 547, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21731" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "In conclusion, our margin uncertainty sampling loss can be expressed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 102, + 287, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 102, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 287, + 139 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s a m p}} (\\theta , \\mathcal {Q}) = \\underset {(\\mathbf {q}, \\mathbf {p}) \\sim \\mathcal {Q}} {\\mathbb {E}} | | \\mathbf {q} - U _ {\\theta} (\\mathbf {q}) \\cdot \\frac {\\nabla U _ {\\theta} (\\mathbf {q})}{| | \\nabla U _ {\\theta} (\\mathbf {q}) | | _ {2} ^ {2}} - \\mathbf {p} | | _ {2} ^ {2}, \\tag {3}", + "image_path": "f054ba02147d3241e0484eeea7df230dac00a06bad768df4f117a2928e6523f2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "text", + "content": "where spatial gradient " + }, + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\nabla U_{\\theta}" + }, + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "text", + "content": " can be computed efficiently through automatic differentiation (e.g. PyTorch [59])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 164, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 164, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 287, + 307 + ], + "type": "text", + "content": "This loss offers two primary advantages. Firstly, it serves as a form of supervision for the occupancy of the query points. The sign of the margin function differs on opposite sides of the decision boundary. Consequently, based on this sign, a Newton-Raphson iteration will move points along or against the direction of its gradient resulting in an occupancy field whose decision boundary is aligned with the input point cloud. Additionally, this loss operates as a smoothness constraint as Newton-Raphson root finding relies on approximating the function at the roots by its first-order approximation on the initial points " + }, + { + "bbox": [ + 46, + 164, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 164, + 287, + 307 + ], + "type": "text", + "content": ", hence smoothing the margin function around the input points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 314, + 211, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 211, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 211, + 327 + ], + "type": "text", + "content": "3.2. Entropy Based Regularization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 332, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 287, + 368 + ], + "type": "text", + "content": "To ensure better convergence of our training, we use the MLP initialization in [2], while adapting it such that the last layer predicts binary occupancy for an " + }, + { + "bbox": [ + 47, + 332, + 287, + 368 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 332, + 287, + 368 + ], + "type": "text", + "content": "-radius sphere." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 369, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 464 + ], + "type": "text", + "content": "Furthermore, to ease and improve our learning, we suggest to urge our occupancy field to have minimal uncertainty almost everywhere in space. We use the entropy [69] as an uncertainty measure here. However, this is the opposite behavior that we expect from our network to display at the decision boundary i.e. near the ground-truth surface. Hence, we propose to maximize the uncertainty (i.e. entropy) conversely at the input point cloud " + }, + { + "bbox": [ + 46, + 369, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 369, + 287, + 464 + ], + "type": "text", + "content": " simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 464, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 487 + ], + "type": "text", + "content": "Given a random variable " + }, + { + "bbox": [ + 47, + 464, + 287, + 487 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 464, + 287, + 487 + ], + "type": "text", + "content": ", Shannon entropy [69] is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 496, + 287, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 496, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 114, + 496, + 287, + 509 + ], + "type": "interline_equation", + "content": "\\mathbb {H} (X) = - \\mathbb {E} \\log (P (X)). \\tag {4}", + "image_path": "05b5482943968c8eb693e49751f7dd15cdeb54d6299e719dcf738e7a097b59b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "text", + "content": "We create a set " + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "text", + "content": " of query points " + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "text", + "content": " by sampling space uniformly within a bounding box of input point cloud " + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 517, + 287, + 577 + ], + "type": "text", + "content": ". As a measure of uncertainty we consider the entropy of the occupancy distribution conditioned on the observed spatial location:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 585, + 287, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 585, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 585, + 287, + 599 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e n t r}} (\\theta , \\Omega , \\mathcal {P}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\Omega} \\mathbb {H} (y | \\mathbf {x}, \\theta) - \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\mathbb {H} (y | \\mathbf {p}, \\theta). \\tag {5}", + "image_path": "32630a6301e0045c491823b0714cfc1da0dc5a89789486d8a1717c8a1bd18d30.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 606, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 287, + 666 + ], + "type": "text", + "content": "This entropy polarization loss acts both as an initialization of the field and a regularization mid-learning, as we empirically found that it performs best when its weight (" + }, + { + "bbox": [ + 47, + 606, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 606, + 287, + 666 + ], + "type": "text", + "content": " in Equation 6) is reduced progressively in the final loss in the midst of training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 666, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 689 + ], + "type": "text", + "content": "Finally we can train our model using the following combined empirical risk minimization:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 94, + 698, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 698, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 94, + 698, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} _ {\\text {s a m p}} (\\theta , \\mathcal {Q}) + \\lambda \\mathcal {L} _ {\\text {e n t r}} (\\theta , \\Omega , \\mathcal {P}). \\tag {6}", + "image_path": "048d0a3e56399ef79c7a734335b0ae710f4901b30179dd695283b7a8e0d3b062.jpg" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 68, + 533, + 243 + ], + "blocks": [ + { + "bbox": [ + 318, + 68, + 533, + 243 + ], + "lines": [ + { + "bbox": [ + 318, + 68, + 533, + 243 + ], + "spans": [ + { + "bbox": [ + 318, + 68, + 533, + 243 + ], + "type": "image", + "image_path": "fd5079e6c96045da732480b6e22c9bd93973f7b80a5dd35f2107de61f7983e88.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 251, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 307 + ], + "type": "text", + "content": "Figure 1. Illustration of our training. Our method learns a neural binary occupancy field without off-surface labels. It uses the combination of a margin uncertainty sampling loss near the surface (Green), maximizing entropy at the input point cloud samples (Red), and minimizing entropy everywhere else (Blue)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 327, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 375 + ], + "type": "text", + "content": "Algorithm 1 shows a summary of our training procedure. We note that once an implicit occupancy field is learned, an explicit triangle mesh " + }, + { + "bbox": [ + 305, + 327, + 545, + 375 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 305, + 327, + 545, + 375 + ], + "type": "text", + "content": " can be obtained with it through the Marching Cubes algorithm [47]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 386, + 500, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 500, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 500, + 398 + ], + "type": "text", + "content": "Algorithm 1 The training procedure of our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": "Input: Point cloud " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", number of iterations " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{it}}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", number of near surface queries " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{Q}}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", number of uniform queries " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "N_{\\Omega}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", number of batch input points " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{P}}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", number of uniform queries " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "N_{\\Omega}" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": ", reg. loss weight " + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 399, + 545, + 442 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 443, + 417, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 417, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 417, + 454 + ], + "type": "text", + "content": "Output: Optimal weights " + }, + { + "bbox": [ + 306, + 443, + 417, + 454 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 454, + 531, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 454, + 531, + 466 + ], + "spans": [ + { + "bbox": [ + 315, + 454, + 531, + 466 + ], + "type": "text", + "content": "Compute local st. dev. " + }, + { + "bbox": [ + 315, + 454, + 531, + 466 + ], + "type": "inline_equation", + "content": "\\{\\sigma_{\\mathbf{p}} = \\max_{\\mathbf{v}\\in K\\mathrm{mn}(\\mathbf{p},\\mathbf{P})}||\\mathbf{v} - \\mathbf{p}||_2\\}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "text", + "content": "Generate sets " + }, + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "text", + "content": " (Sec.3.1) and " + }, + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "inline_equation", + "content": "\\varOmega" + }, + { + "bbox": [ + 315, + 466, + 471, + 475 + ], + "type": "text", + "content": " (Sec.3.2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 476, + 387, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 476, + 387, + 486 + ], + "spans": [ + { + "bbox": [ + 315, + 476, + 387, + 486 + ], + "type": "text", + "content": "Initialize " + }, + { + "bbox": [ + 315, + 476, + 387, + 486 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 315, + 476, + 387, + 486 + ], + "type": "text", + "content": " (Sec.4)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 487, + 375, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 487, + 375, + 497 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 375, + 497 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 315, + 487, + 375, + 497 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{it}}" + }, + { + "bbox": [ + 315, + 487, + 375, + 497 + ], + "type": "text", + "content": " times do" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "spans": [ + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "text", + "content": "Sample a batch " + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_b" + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{Q}}" + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 329, + 498, + 472, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "text", + "content": "Sample a batch " + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "inline_equation", + "content": "\\varOmega_{b}" + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "inline_equation", + "content": "N_{\\varOmega}" + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "text", + "content": " from, " + }, + { + "bbox": [ + 329, + 509, + 479, + 520 + ], + "type": "inline_equation", + "content": "\\varOmega_{b}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "spans": [ + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "text", + "content": "Sample a batch " + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_b" + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{P}}" + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 329, + 521, + 471, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 329, + 532, + 468, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 532, + 468, + 542 + ], + "spans": [ + { + "bbox": [ + 329, + 532, + 468, + 542 + ], + "type": "text", + "content": "Compute losses " + }, + { + "bbox": [ + 329, + 532, + 468, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{samp}}(\\theta ,\\mathcal{Q}_b)" + }, + { + "bbox": [ + 329, + 532, + 468, + 542 + ], + "type": "text", + "content": " (Eq.3)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 329, + 543, + 477, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 543, + 477, + 553 + ], + "spans": [ + { + "bbox": [ + 329, + 543, + 477, + 553 + ], + "type": "text", + "content": "Compute losses " + }, + { + "bbox": [ + 329, + 543, + 477, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{entr}}(\\theta, \\Omega_b, \\mathcal{P}_b)" + }, + { + "bbox": [ + 329, + 543, + 477, + 553 + ], + "type": "text", + "content": " (Eq.5)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 329, + 554, + 446, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 554, + 446, + 564 + ], + "spans": [ + { + "bbox": [ + 329, + 554, + 446, + 564 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\alpha \\nabla_{\\theta}\\left(\\mathcal{L}_{\\mathrm{samp}} + \\lambda \\mathcal{L}_{\\mathrm{entr}}\\right)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 329, + 564, + 394, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 564, + 394, + 574 + ], + "spans": [ + { + "bbox": [ + 329, + 564, + 394, + 574 + ], + "type": "text", + "content": "Update " + }, + { + "bbox": [ + 329, + 564, + 394, + 574 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 329, + 564, + 394, + 574 + ], + "type": "text", + "content": " (Sec.4)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 575, + 345, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 575, + 345, + 584 + ], + "spans": [ + { + "bbox": [ + 316, + 575, + 345, + 584 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 609, + 441, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 441, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 441, + 623 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "We set number of queries as " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{Q}} = 1000k" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\Omega} = 10k" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Our MLP follows the architecture in [48], and similarly to the latter we set " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 51" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". We train for " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{it} = 40k" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " iterations on a Nvidia RTX A6000 GPU using the Adam optimizer with learning rate " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.001" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Our training takes roughly 5 minutes for a 1024 sized input point cloud. The regularization parameter " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " (Equation 6) follows an exponential decay" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21732" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 69, + 264, + 158 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 264, + 158 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 264, + 158 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 264, + 158 + ], + "type": "table", + "html": "
CD1CD2NCFS
SPSR [36]2.340.2240.740.50
OG-INR [38]1.360.0510.550.55
N-Pull [48]1.160.0740.840.75
G-Pull [16]1.070.0320.700.74
NTPS [15]1.110.0670.880.74
Ours0.760.0200.880.83
", + "image_path": "dee2fc2b82cc613e812c2f40b2b555b5f535f0bacb130be14c490555225a3d71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "text", + "content": "of the form " + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "inline_equation", + "content": "\\exp -\\kappa t" + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 212, + 255, + 225 + ], + "type": "inline_equation", + "content": "1.84\\times 10^{-2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 101, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 101, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 101, + 253 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 262, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 262, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 45, + 262, + 290, + 574 + ], + "type": "text", + "content": "To test the efficacy of our approach, we assess our capacity to infer implicit representations of shapes when presented with sparse and noisy point clouds. We employ datasets from standard reconstruction benchmarks, showcasing a diverse array of challenges associated with implicit shape function inference from sparse inputs. Consistently with existing literature, we evaluate the performance of our method by quantifying the accuracy of 3D explicit shape models obtained post-convergence from our MLPs. It is worth noting that our approach entails fitting an MLP independently for each point cloud, operating without reliance on pre-learned priors or additional training data. We compare quantitatively and qualitatively to the state-of-the-art in our problem setting, i.e. unsupervised reconstruction from unoriented point, including deep learning methods N-Pull [48], NTPS [15], G-Pull [16], OG-INR [38], DiGs [5], NDrop [9], SAP [61]. We show results for NSpline [84] even though it requires normals. We also compare to classical Poisson Reconstruction (SPSR [36]). We note that NTPS is the closest method to ours as it focuses specifically on the sparse input case. For comprehensive evaluation, we also include comparisons with feed-forward generalizable methods, namely POCO [8] and CONet [60], alongside the prior-based optimization method On-Surf [49] dedicated to sparse inputs. Unless stated differently, we use the publicly available official implementations of existing methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 107, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 107, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 107, + 597 + ], + "type": "text", + "content": "5.1. Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Following seminal work, we evaluate our method and the competition w.r.t. the ground truth using standard metrics for the 3D reconstruction task. Namely, the L1 Chamfer Distance (CD1) " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", L2 Chamfer Distance (CD2) " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", the Hausdorff distance (HD) and the euclidean distance based F-Score (FS) when ground truth points are available, and finally Normal Consistency (NC) when ground truth normals are available. We detail the expressions of these metrics in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 69, + 529, + 171 + ], + "blocks": [ + { + "bbox": [ + 46, + 165, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 287, + 189 + ], + "type": "text", + "content": "Table 1. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 69, + 529, + 171 + ], + "lines": [ + { + "bbox": [ + 324, + 69, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 324, + 69, + 529, + 171 + ], + "type": "table", + "html": "
CD1CD2NCFS
POCO [8]0.3080.0020.9340.981
CONet [60]1.260.0280.8290.599
On-Surf [49]0.5840.0120.9360.915
SPSR [36]0.7510.0280.8710.839
G-Pull [16]0.4950.0050.8870.945
NTPS [15]0.7370.0150.9430.844
Ours0.2600.0020.9520.974
", + "image_path": "ecd792566d053c4700f3557f02731c0ef734b9bb839afc493256ccbe0d4e7cc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 178, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 547, + 200 + ], + "type": "text", + "content": "Table 2. Faust [7] reconstructions from sparse noisy unoriented point clouds." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 221, + 471, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 221, + 471, + 234 + ], + "spans": [ + { + "bbox": [ + 306, + 221, + 471, + 234 + ], + "type": "text", + "content": "5.2. Datasets and Input Definitions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 479 + ], + "type": "text", + "content": "ShapeNet [14] consists of various instances of 13 different synthetic 3D object classes. We follow the train/test splits defined in [84]. We generate noisy input point clouds by sampling 1024 points from the meshes and adding Gaussian noise of standard deviation 0.005 following the literature (e.g. [8, 60]). For brevity we show results on classes Tables, Chairs and Lamps. Faust [7] consists of real scans of 10 human body identities in 10 different poses. We sample sets of 1024 points from the scans as inputs. 3D Scene [92] contains large scale complex real world scenes obtained with a handheld commodity range sensor. We follow [15, 34, 48] and sample our input point clouds with a sparse density of " + }, + { + "bbox": [ + 304, + 239, + 547, + 479 + ], + "type": "inline_equation", + "content": "100\\mathrm{per}\\mathrm{m}^2" + }, + { + "bbox": [ + 304, + 239, + 547, + 479 + ], + "type": "text", + "content": ", and we report performance similarly for scenes Burghers, Copyroom, Lounge, Stonewall and Totempole. Surface Reconstruction Benchmark (SRB) [83] consists of five object scans, each with different challenges such as complex topology, high level of detail, missing data and varying feature scales. We sample 1024 points from the scans for the sparse input experiment, and we also experiment using the dense inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 488, + 463, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 463, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 463, + 501 + ], + "type": "text", + "content": "5.3. Object Level Reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "content": "We conduct the reconstruction of ShapeNet [14] objects using sparse and noisy point clouds. Table 1 and Figure 2 present a quantitative and qualitative comparison with other methods. Our approach consistently outperforms the competition across all metrics, as seen in the visual superiority of our reconstructions. We excel in capturing fine structures and details with greater fidelity. While NTPS demonstrates good overall coarse reconstructions, its thin plate spline smoothing prior appears to limit its expressivity. Additionally, OG-INR struggles to achieve satisfactory results in the sparse and noisy regime, despite its effective Octree-based sign field guidance in denser settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "type": "text", + "content": "5.4. Real Articulated Shape Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "We undertake the reconstruction of Faust [7] human shapes using sparse and noisy point clouds. Table 2 and Figure 3 provide a numerical and qualitative comparison with other" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21733" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 68, + 474, + 380 + ], + "blocks": [ + { + "bbox": [ + 123, + 68, + 474, + 380 + ], + "lines": [ + { + "bbox": [ + 123, + 68, + 474, + 380 + ], + "spans": [ + { + "bbox": [ + 123, + 68, + 474, + 380 + ], + "type": "image", + "image_path": "81eb2ee4c8c9f849f909841468ea9026f2b8119bb29806b877e5643cffc8ae49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 387, + 447, + 399 + ], + "lines": [ + { + "bbox": [ + 144, + 387, + 447, + 399 + ], + "spans": [ + { + "bbox": [ + 144, + 387, + 447, + 399 + ], + "type": "text", + "content": "Figure 2. ShapeNet [14] reconstructions from sparse noisy unoriented point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 419, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 575 + ], + "type": "text", + "content": "methods. Our approach surpasses the alternatives across all metrics, with visually superior reconstructions, as especially noticeable at the extremities of the body. Analogous to the fine structures in the ShapeNet experiment, these areas pose challenges due to limited input point cloud samples, making shape prediction inherently difficult and ambiguous. NTPS reconstructions, similarly, exhibit coarser and less detailed results on this dataset. It is worth highlighting that our method outperforms not only feed-forward generalizable approaches POCO and CONet but also the prior-based optimization method On-Surf. These methods rely on priors trained on ShapeNet, thereby constraining their generalization ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 223, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 223, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 223, + 597 + ], + "type": "text", + "content": "5.5. Real Scene Level Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "In accordance with [15], we present the reconstruction outcomes on the 3D Scene [92] dataset derived from spatially sparse point clouds. Table 3 provides a summary of numerical results, incorporating results for methods NTPS, N-Pull, SAP, NDrop, and NSpline as reported in the state-of-the-art method NTPS. Our method excels in this benchmark, surpassing the competition in most metrics. Notably, our baseline N-Pull exhibits more conspicuous failures in this extensive sparse setup. Figure 5 offers qualitative compar" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 419, + 473, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 473, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 473, + 431 + ], + "type": "text", + "content": "isons against baselines N-Pull and SPSR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 442, + 406, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 406, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 406, + 454 + ], + "type": "text", + "content": "6. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Loss. Our approach centers around aligning the decision boundary of our occupancy function with the shape surface through a loss function based on uncertainty sampling, denoted as " + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{samp}}" + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": " (Equation 3). To facilitate and enhance our learning process, we impose constraints on the network to minimize uncertainty almost everywhere in space while maximizing it at the decision boundary. Entropy serves as our chosen uncertainty measure. However, other measures can be considered. We justify this choice in Table 4. We compare our entropy-based regularization to margin-based regularization (Margin reg), using the absolute value of our margin function " + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "U_{\\theta}" + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": " (Equation 3) as an uncertainty measure instead of entropy. Additionally, we compare it to the sign-agnostic occupancy supervision strategy introduced in [75] denoted as SA-Occ. Although this strategy fails to converge in the absence of priors, we demonstrate its beneficial use as a regularization in place of our entropy-based approach. It is noteworthy that while margin-based regularization (Margin reg) yields better results, it is still outperformed by our entropy-based regularization. Figure 4 shows an example of validation plots, displaying the role of our regularization" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21734" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 566, + 155 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 566, + 155 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 566, + 155 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 566, + 155 + ], + "type": "table", + "html": "
BurghersCopyroomLoungeStonewallTotempleMean
CD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NCCD1CD2NC
SPSR [36]0.1780.2050.8740.2250.2860.8610.2800.3650.8690.3000.4800.8660.5881.6730.8790.3140.6020.870
NDrop [9]0.2000.1140.8250.1680.0630.6960.1560.0500.6630.1500.0810.8150.2030.1390.8440.1750.0890.769
N-Pull [48]0.0640.0080.8980.0490.0050.8280.1330.0380.8470.0600.0050.9100.1780.0240.9080.0970.0160.878
SAP [61]0.1530.1010.8070.0530.0090.7710.1340.0330.8130.0700.0070.8670.4740.3820.7250.1510.1000.797
NSpline [84]0.1350.1230.8910.0560.0230.8550.0630.0390.8270.1240.0910.8970.3780.7680.8920.1510.2090.88
NTPS [15]0.0550.0050.9090.0450.0030.8920.1290.0220.8720.0540.0040.9390.1030.0170.9350.0770.0100.897
Ours0.0220.0010.8710.0410.0120.8120.0210.0010.8700.0280.0030.9310.0260.0010.9360.0270.0030.886
", + "image_path": "48f8d5bb93b51cc1571c9892691a6f4c79db3af6966ce1cee4d803c5a64a4529.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 190, + 286, + 561 + ], + "blocks": [ + { + "bbox": [ + 50, + 190, + 286, + 561 + ], + "lines": [ + { + "bbox": [ + 50, + 190, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 50, + 190, + 286, + 561 + ], + "type": "image", + "image_path": "a7eb3af6fe9f64604aa6453410b2577c7ab2ce9c92474e7acb0e6b6217f438e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 567, + 288, + 590 + ], + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 590 + ], + "type": "text", + "content": "Figure 3. Faust [7] reconstructions from sparse unoriented point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 615, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 640 + ], + "type": "text", + "content": "in avoiding overfitting. The baseline in this example is SDF method N-Pull [48]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Point Cloud Density. We utilize the SRB [83] benchmark to evaluate the performance of our method across varying point cloud densities. Table 5 presents comparative results for sparse (1024) and dense input point clouds. We include results for the competition from OG-INR in the dense setting. Our method outperforms the competition in the sparse" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 351, + 195, + 486, + 299 + ], + "blocks": [ + { + "bbox": [ + 351, + 195, + 486, + 299 + ], + "lines": [ + { + "bbox": [ + 351, + 195, + 486, + 299 + ], + "spans": [ + { + "bbox": [ + 351, + 195, + 486, + 299 + ], + "type": "image", + "image_path": "349e8e5adbe7a19cd92bf9511ab81e613da5b972f08820a1c4c2f3ae7b630f97.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 299, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 331 + ], + "type": "text", + "content": "Figure 4. CD1 distance to GT for reconstructions of shape Gargoyle of benchmark SRB [83] from a sparse unoriented point cloud." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 358, + 343, + 499, + 405 + ], + "blocks": [ + { + "bbox": [ + 178, + 161, + 414, + 173 + ], + "lines": [ + { + "bbox": [ + 178, + 161, + 414, + 173 + ], + "spans": [ + { + "bbox": [ + 178, + 161, + 414, + 173 + ], + "type": "text", + "content": "Table 3. 3D Scene [92] reconstructions from sparse point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 358, + 343, + 499, + 405 + ], + "lines": [ + { + "bbox": [ + 358, + 343, + 499, + 405 + ], + "spans": [ + { + "bbox": [ + 358, + 343, + 499, + 405 + ], + "type": "table", + "html": "
CD1NC
N-Pull [48] (baseline)1.100.85
\\( \\mathcal{L}_{\\text{samp}} \\) (Equ.3)1.020.86
\\( \\mathcal{L}_{\\text{samp}} + \\text{Margin reg} \\)0.800.88
\\( \\mathcal{L}_{\\text{samp}} + \\text{SA-Occ [75] reg} \\)0.950.88
Ours0.770.89
", + "image_path": "5ef64b149d69b3cec784ba2f43217e7735bdf270145d8b0603bacc8f853c9f00.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 349, + 447, + 505, + 528 + ], + "blocks": [ + { + "bbox": [ + 305, + 413, + 545, + 435 + ], + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 435 + ], + "type": "text", + "content": "Table 4. Ablation of our method on the Tables class of ShapeNet [14]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 447, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 349, + 447, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 349, + 447, + 505, + 528 + ], + "type": "table", + "html": "
SparseDense
CD1HDCD1HD
SPSR [36]2.2721.11.2522.59
DiGs [5]0.686.050.193.52
OG-INR [38]0.857.100.204.06
NTPS [15]0.737.78--
N-Pull [48]0.588.900.234.46
Ours0.496.040.203.94
", + "image_path": "f432362f6c8f98b84504b7f1388d5266279a292d50069d553914775f6b2d01ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 535, + 500, + 547 + ], + "lines": [ + { + "bbox": [ + 351, + 535, + 500, + 547 + ], + "spans": [ + { + "bbox": [ + 351, + 535, + 500, + 547 + ], + "type": "text", + "content": "Table 5. Ablation of point cloud density." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": "case and performs comparably to the state-of-the-art in the dense case. The improvement over our baseline is substantial for both sparse and dense inputs, as highlighted visually in Figure 6, where we showcase reconstructions for both sparse and dense cases. Notably, we achieve better topologies in the sparse case and enhanced, more accurate details in the dense case, as indicated by the red boxes. These results underscore the utility and advantages of our contribution, even in the dense setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Running Time and Performance against Number of Queries and Input Points. Fig.7 shows our best performance over time (minutes) for various input point cloud" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21735" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 70, + 473, + 318 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 473, + 318 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 473, + 318 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 473, + 318 + ], + "type": "image", + "image_path": "265fdd355c746497529ef65089ba3ea07d009c114700d275932561cc9752789f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 325, + 436, + 336 + ], + "lines": [ + { + "bbox": [ + 155, + 325, + 436, + 336 + ], + "spans": [ + { + "bbox": [ + 155, + 325, + 436, + 336 + ], + "type": "text", + "content": "Figure 5. 3D Scene [92] reconstructions from sparse unoriented point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 355, + 286, + 468 + ], + "blocks": [ + { + "bbox": [ + 47, + 355, + 286, + 468 + ], + "lines": [ + { + "bbox": [ + 47, + 355, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 286, + 468 + ], + "type": "image", + "image_path": "2e610bebc72aa7673d6d946c2ad20600042bc4d0e9d2156767070b607559f6e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 475, + 286, + 498 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 286, + 498 + ], + "type": "text", + "content": "Figure 6. SRB [83] reconstructions from sparse and dense unoriented inputs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 515, + 147, + 589 + ], + "blocks": [ + { + "bbox": [ + 50, + 515, + 147, + 589 + ], + "lines": [ + { + "bbox": [ + 50, + 515, + 147, + 589 + ], + "spans": [ + { + "bbox": [ + 50, + 515, + 147, + 589 + ], + "type": "image", + "image_path": "596d88e2e444ce3b141f827b8d2ae41b6dbc430fe3a70c11008493f1f0a3b57e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 590, + 156, + 609 + ], + "lines": [ + { + "bbox": [ + 47, + 590, + 156, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 156, + 609 + ], + "type": "text", + "content": "Figure 7. Performance over time (varying input size)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 177, + 516, + 274, + 589 + ], + "blocks": [ + { + "bbox": [ + 177, + 516, + 274, + 589 + ], + "lines": [ + { + "bbox": [ + 177, + 516, + 274, + 589 + ], + "spans": [ + { + "bbox": [ + 177, + 516, + 274, + 589 + ], + "type": "image", + "image_path": "b90b9c197cd2cee7ff89d513c475eac763bd097949670ff3fb8436e9e87ff952.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 590, + 284, + 609 + ], + "lines": [ + { + "bbox": [ + 175, + 590, + 284, + 609 + ], + "spans": [ + { + "bbox": [ + 175, + 590, + 284, + 609 + ], + "type": "text", + "content": "Figure 8. Performance over time (varying query set size)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": "sizes. Models using larger inputs reach the baseline performance at earlier times. Fig.8 shows our best performance over time while decreasing the total number of query samples. Performance does not deteriorate drastically under less queries. These Experiments were performed using shape Gargoyle of benchmark SRB." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "content": "Robustness to Noise We evaluate here (Tab.6) our method" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 359, + 355, + 495, + 388 + ], + "blocks": [ + { + "bbox": [ + 359, + 355, + 495, + 388 + ], + "lines": [ + { + "bbox": [ + 359, + 355, + 495, + 388 + ], + "spans": [ + { + "bbox": [ + 359, + 355, + 495, + 388 + ], + "type": "table", + "html": "
noise std. dev.00.0050.025
CD1 ↓0.560.772.16
NC ↑0.930.890.68
", + "image_path": "b40e99b7e7cbf44c8f2370fcb95a636a02d931fd3c86f5293153c4a3c7691a40.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 396, + 470, + 406 + ], + "lines": [ + { + "bbox": [ + 381, + 396, + 470, + 406 + ], + "spans": [ + { + "bbox": [ + 381, + 396, + 470, + 406 + ], + "type": "text", + "content": "Table 6. Noise ablation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 430, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 545, + 478 + ], + "type": "text", + "content": "with various noise levels on class Table of ShapeNet. Performance remains reasonable even under heavily corrupted inputs, which confirms our qualitative superiority to competition in Fig.2, and our resilience to noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 381, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 381, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 381, + 503 + ], + "type": "text", + "content": "7. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 512, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 545, + 571 + ], + "type": "text", + "content": "We notice that excess regularization can cause reconstructions to be overly smooth in some cases, which can affect our NC numbers is some benchmarks. This can be improved by tuning the hyperparameters of the method per benchmark or per object/scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 584, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 378, + 597 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 545, + 712 + ], + "type": "text", + "content": "We presented a method for implicit shape reconstruction from sparse, noisy and unoriented point cloud. Our results demonstrate that occupancy fields offer and effective implicit shape representation for this task, and that they can be learned successfully even in the unsupervised setting. Under our carefully designed losses, the occupancy outperforms state-of-the-art methods based on SDF, as-well-as other occupancy based baselines that we show in ablation studies." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21736" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. CG, 2001. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 147 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 170 + ], + "type": "text", + "content": "[3] Matan Atzmon and Yaron Lipman. Sald: Sign agnostic learning with derivatives. In ICML, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 172, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 286, + 205 + ], + "type": "text", + "content": "[4] Adi Ben-Israel. A newton-raphson method for the solution of systems of equations. In Journal of Mathematical analysis and applications, 1966. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 286, + 249 + ], + "type": "text", + "content": "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In CVPR, 2022. 2, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "type": "text", + "content": "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. TVCG, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 318 + ], + "type": "text", + "content": "[7] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In CVPR, 2014. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 320, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 286, + 363 + ], + "type": "text", + "content": "[8] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 365, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 286, + 419 + ], + "type": "text", + "content": "[9] Alexandre Boulch, Pierre-Alain Langlois, Gilles Puy, and Renaud Marlet. Needrop: Self-supervised shape representation from sparse point clouds using needle dropping. In 2021 International Conference on 3D Vision (3DV), pages 940-950. IEEE, 2021. 3, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "type": "text", + "content": "[10] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3d objects with radial basis functions. In SIGGRAPH, 2001. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 286, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 286, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 286, + 488 + ], + "type": "text", + "content": "[11] Frédéric Casals and Joachim Giesen. Effective Computational Geometry for Curves and Surfaces. 2006. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 286, + 533 + ], + "type": "text", + "content": "[12] Rohan Chabra, Jan E Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 535, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 286, + 601 + ], + "type": "text", + "content": "[13] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 656 + ], + "type": "text", + "content": "[14] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[15] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[16] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[17] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "text", + "content": "[18] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[19] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "text", + "content": "[20] Julian Chibane and Gerard Pons-Moll. Implicit feature networks for texture completion from partial 3d data. In European Conference on Computer Vision, pages 717-725. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "text", + "content": "[21] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "text", + "content": "[22] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In CVPR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[23] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[24] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In ECCV, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "type": "text", + "content": "[25] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[26] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "type": "text", + "content": "[27] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In CVPR, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "text", + "content": "[28] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "type": "text", + "content": "[29] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "text", + "content": "[30] Gaël Guennebaud and Markus Gross. Algebraic point set surfaces. In ACM siggraph 2007 papers, pages 23–es. 2007. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[31] John C Hart. Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer, 1996. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21737" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[33] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[34] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 176, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 198 + ], + "type": "text", + "content": "[35] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 200, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 287, + 221 + ], + "type": "text", + "content": "[36] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. TOG, 2013. 1, 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "type": "text", + "content": "[37] Ravikrishna Kolluri. Provably good moving least squares. TALG, 2008. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "type": "text", + "content": "[38] Chamin Hewa Koneputugodage, Yizhak Ben-Shabat, and Stephen Gould. Octree guided unoriented surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16717-16726, 2023. 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "type": "text", + "content": "[39] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3d shapes with dynamic code clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12840-12850, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "text", + "content": "[40] Siyou Lin, Dong Xiao, Zuoqiang Shi, and Bin Wang. Surface reconstruction from point clouds without normals by parametrizing the gauss formula. ACM Transactions on Graphics, 42(2):1-19, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 459 + ], + "type": "text", + "content": "[41] Stefan Lionar, Daniil Emtsev, Dusan Svilarkovic, and Songyou Peng. Dynamic plane convolutional occupancy networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1829-1838, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 461, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 287, + 483 + ], + "type": "text", + "content": "[42] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In ICML, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "text", + "content": "[43] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In CVPR, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[44] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization. arXiv preprint arXiv:2202.08345, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "text", + "content": "[45] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In ECCV, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[47] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, 1987. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[48] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In ICML, 2021. 3, 4, 5, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[49] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6315-6325, 2022. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[50] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6326-6337, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[51] Corentin Mercier, Thibault Lescoat, Pierre Roussillon, Tamy Boubekeur, and Jean-Marc Thiery. Moving level-of-detail surfaces. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 287 + ], + "type": "text", + "content": "[52] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "type": "text", + "content": "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 335, + 545, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 368 + ], + "type": "text", + "content": "[54] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "text", + "content": "[55] Amine Ousafi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 405, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 426 + ], + "type": "text", + "content": "[56] Amine Ousafi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. In 3DV, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 483 + ], + "type": "text", + "content": "[57] David Palmer, Dmitriy Smirnov, Stephanie Wang, Albert Chern, and Justin Solomon. Deepcurrents: Learning implicit representations of shapes with boundaries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18665-18675, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 485, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 529 + ], + "type": "text", + "content": "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[59] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[60] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1, 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[61] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[62] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21738" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[63] Marie-Julie Rakotosaona, Noam Aigerman, Niloy Mitra, Maks Ovsjanikov, and Paul Guerrero. Differentiable surface triangulation. In SIGGRAPH Asia, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 287, + 139 + ], + "type": "text", + "content": "[64] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 171 + ], + "type": "text", + "content": "[65] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In NeurIPS, 2004. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 287, + 205 + ], + "type": "text", + "content": "[66] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 206, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 249 + ], + "type": "text", + "content": "[67] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 250, + 267, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 267, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 267, + 261 + ], + "type": "text", + "content": "[68] Burr Settles. Active learning literature survey. 2009. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 293 + ], + "type": "text", + "content": "[69] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 293, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 293, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 293, + 287, + 337 + ], + "type": "text", + "content": "[70] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "type": "text", + "content": "[71] Vincent Sitzmann, Eric R Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 287, + 414 + ], + "type": "text", + "content": "[72] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "type": "text", + "content": "[73] Vincent Sitzmann, Semon Rezchikov, William T Freeman, Joshua B Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 459, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 459, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 459, + 287, + 514 + ], + "type": "text", + "content": "[74] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "type": "text", + "content": "[75] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. Sa-convonet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6504–6513, 2021. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 570, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 287, + 603 + ], + "type": "text", + "content": "[76] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 646 + ], + "type": "text", + "content": "[77] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. Patchnets: Patch-based generalizable deep implicit 3d shape representations. In ECCV, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 680 + ], + "type": "text", + "content": "[78] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 611 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[81] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. TOG, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 205 + ], + "type": "text", + "content": "[82] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. Advances in Neural Information Processing Systems, 34:2810-2822, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 206, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 240 + ], + "type": "text", + "content": "[83] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In CVPR, 2019. 2, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[84] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural splines: Fitting 3d surfaces with infinitely-wide neural networks. In CVPR, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "type": "text", + "content": "[85] Francis Williams, Zan Gojcic, Sameh Khamis, Denis Zorin, Joan Bruna, Sanja Fidler, and Or Litany. Neural fields as learnable kernels for 3d reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "type": "text", + "content": "[86] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T Freeman, and Joshua B Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[87] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "type": "text", + "content": "[88] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 432, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 466 + ], + "type": "text", + "content": "[89] Mohsen Yavartanoo, Jaeyoung Chung, Reyhaneh Neshatavar, and Kyoung Mu Lee. 3dias: 3d shape reconstruction with implicit algebraic surfaces. In ICCV, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 466, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 489 + ], + "type": "text", + "content": "[90] Tjalling J Ypma. Historical development of the newton-raphson method. SIAM review, 37(4):531-551, 1995. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "text", + "content": "[91] Junsheng Zhou, Baorui Ma, Liu Yu-Shen, Fang Yi, and Han Zhizhong. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "text", + "content": "[92] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics (ToG), 32(4):1-8, 2013. 5, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[93] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3d-prnn: Generating shape primitives with recurrent neural networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21739" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_content_list.json b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d648c5e79fe2c07435fd033373d2417d5682e80 --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_content_list.json @@ -0,0 +1,1845 @@ +[ + { + "type": "text", + "text": "Unsupervised Salient Instance Detection", + "text_level": 1, + "bbox": [ + 281, + 130, + 689, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Tian $^{1}$ Ke Xu $^{2,\\dagger}$ Rynson Lau $^{2,\\dagger}$", + "bbox": [ + 333, + 179, + 635, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Huawei Technologies $^{2}$ City University of Hong Kong", + "bbox": [ + 297, + 200, + 673, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xin.tian.831, kkangwing}@gmail.com, Rynson.Lau@cityu.edu.hk", + "bbox": [ + 222, + 219, + 746, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† joint corresponding authors", + "bbox": [ + 397, + 236, + 570, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 275, + 312, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The significant amount of manual efforts in annotating pixel-level labels has triggered the advancement of unsupervised saliency learning. However, without supervision signals, state-of-the-art methods can only infer region-level saliency. In this paper, we propose to explore the unsupervised salient instance detection (USID) problem, for a more fine-grained visual understanding. Our key observation is that self-supervised transformer features may exhibit local similarities as well as different levels of contrast to other regions, which provide informative cues to identify salient instances. Hence, we propose SCoCo, a novel network that models saliency coherence and contrast for USID. SCoCo includes two novel modules: (1) a global background adaptation (GBA) module with a scene-level contrastive loss to extract salient regions from the scene by searching the adaptive \"saliency threshold\" in the self-supervised transformer features, and (2) a locality-aware similarity (LAS) module with an instance-level contrastive loss to group salient regions into instances by modeling the in-region saliency coherence and cross-region saliency contrasts. Extensive experiments show that SCoCo outperforms state-of-the-art weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods.", + "bbox": [ + 75, + 309, + 473, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Salient Instance Detection (SID) aims to detect instance-level salient objects. It can provide a fine-grained visual understanding of the scene, and is therefore able to facilitate a wide range of applications, e.g., image captioning [21], scene text recognition [20], medical image analysis [19], and others [1, 4, 43, 48, 91]. However, existing SID methods either require pixel-level labels [25, 74] or image-level labels [62] as supervisions, which are expensive to prepare. Despite the advancements in unsupervised saliency detection [32, 44, 57, 72, 79, 81, 87, 88], none of the existing methods have demonstrated the capability to learn instance-level saliency, which essentially requires fine-grained dif", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/749713cb83f1b6bb3bb5394ea778f4943b7dc4ec04775556f55e4b1b11e7e66d.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 509, + 273, + 581, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/480dcb5c837848e086cd1f085018e00232c7fdbd43e4e4e4a54083a922ee24f4.jpg", + "image_caption": [ + "(b) FOUND [57]" + ], + "image_footnote": [], + "bbox": [ + 583, + 273, + 656, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ea38f01bb960d52fc92c00555d4cc2e9b6e5a23b90e75aa4cbc1c17825de35e.jpg", + "image_caption": [ + "(c) SelfMask [55]" + ], + "image_footnote": [], + "bbox": [ + 658, + 273, + 733, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53ddf3eae9959d73c0d1281b0f3a97f7ca3bbf4f694d4949a5aac9aa3c303521.jpg", + "image_caption": [ + "(d) Ours", + "Figure 1. Existing unsupervised salient object detection methods, FOUND [57] (b) and SelfMask [55] (c), cannot differentiate salient instances (rows 1 and 2), but tend to detect strong semantic instances, e.g., the woman in row 2. In contrast, our method (d) correctly detects all salient instances by modeling saliency coherence and different levels of saliency contrast." + ], + "image_footnote": [], + "bbox": [ + 735, + 273, + 808, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/697767f4816ac2c15169534c711550fc3c370169b02f646918398e99fa3e2ead.jpg", + "image_caption": [ + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 810, + 273, + 883, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferentiation not only between salient and non-salient regions but also between different salient regions (Figure 1). In this paper, we propose to study this unexplored task, i.e., unsupervised salient instance detection (USID).", + "bbox": [ + 498, + 492, + 890, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our key observation of this problem is built upon the advanced self-supervised transformer representations. We observe that their features exhibit saliency coherence (high similarities) within local regions and may show varying degrees of saliency contrast (lower similarities and dissimilarities) between different regions. In Figure 2, we visualize global saliency information within the self-supervised transformer features (b-d), and the saliency coherence and contrast between three feature pixels and the rest of the image (e-g), where each of these pixels is randomly selected from a salient instance. While the transformer features themselves may exhibit some degree of global saliency contrast (b-d), the computed pixel-based correlations show that each selected pixel tends to have high saliency coherence with pixels of the same instance but different levels of saliency contrast to pixels of other instances and the background (e-g). This observation inspires us to mine regional saliency coherence and contrast information for USID.", + "bbox": [ + 496, + 553, + 892, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach, called SCoCo, to model Saliency Coherence and Contrast information from deep self-supervised features to address the USID problem. Specifically, SCoCo first models scene-level saliency contrast to determine the \"saliency thresh", + "bbox": [ + 498, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2702", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a849b1a076fdd06f7fb82676c01b2a862ca2b4fec1d0697cffaed2ff4d40dce7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 120, + 98, + 222, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cde78aa92b16fdab70119187ee7cb9baa6d9414b8cfaf0938c77dabbfd70fb6c.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 223, + 98, + 326, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ccfb52e2ea8b07b7e3a119fa00d7815df00501e4ebedffbc6a5189105f947cf3.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 328, + 98, + 431, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e8ad4739a5d955d6d45bea70645daa869f9f8a1de36540354ee7e1e80aec1940.jpg", + "image_caption": [ + "(d)", + "High" + ], + "image_footnote": [], + "bbox": [ + 434, + 98, + 535, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c57dba02e7ffe8ddd961ac83d0d759a88e4830fb4387cbfa64d63386348065af.jpg", + "image_caption": [ + "(e)", + "← Pixel-based Saliency Contrast Degree →" + ], + "image_footnote": [], + "bbox": [ + 537, + 98, + 640, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88a1d0ce9fdfd94df4b79c17aeeca1ae27e31753d6557332fc4938b28b7f489c.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 642, + 98, + 745, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aab17e91de9f16d9bfaa3a7d8bcda67d7f55aa4730e034a70a62d4ea8b772323.jpg", + "image_caption": [ + "(g)", + "High" + ], + "image_footnote": [], + "bbox": [ + 746, + 98, + 849, + 203 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/368200a641c5a62059d7fb4166b41c715b67a218e86063e23636a6cd6c5fad3e.jpg", + "image_caption": [ + "Global Saliency Contrast Degree $\\rightarrow$", + "Figure 2. Given an input image (a), we first visualize its corresponding self-supervised transformer features [8] (b), the pixel correlations to the [CLS] token (c), and the pixel correlations over the global scene (d). We then randomly sample three pixels (purple dots), each for one instance, and visualize the saliency contrast of the transformer features between the sampled pixel and all other pixels of the image by computing feature point correlations (e-g). We observe that while the self-supervised transformer features may show the global saliency contrast to some degree, pixels from the same instance tend to be more coherent, while pixels from different regions tend to have different levels of contrast. This inspires us to mine such regional saliency coherence and contrast for USID." + ], + "image_footnote": [], + "bbox": [ + 127, + 215, + 222, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85f5b9f7b7d35ad3929348ee40b1b82ee268eef7c2fc61c231ffbed97e639dc2.jpg", + "image_caption": [ + "一" + ], + "image_footnote": [], + "bbox": [ + 498, + 215, + 532, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "old,\" which differentiates salient regions with respect to diverse background contents. We implement this via a novel global background adaptation (GBA) module with a scene-level contrastive loss, which learns to iteratively pull the saliency regions out of the complex background. SCoCo then models the within-region saliency coherence and cross-region saliency contrast information for grouping pixels into salient instances. We implement this by proposing a novel locality-aware similarity (LAS) module with an instance-level contrastive loss, which initially focuses on modeling the saliency coherence within individual regions and then shifts to model different levels of saliency contrast between different regions.", + "bbox": [ + 75, + 330, + 468, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, this paper has three main contributions:", + "bbox": [ + 96, + 527, + 459, + 541 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our knowledge, we propose the first USID approach, named SCoCo, to learn to detect salient instances without any annotated labels, by modeling the intra-region coherence and inter-region contrast information from self-supervised transformer features.", + "- SCoCo has two novel modules: a GBA module with a scene-level contrastive loss to find the saliency thresholds of different image contents, and a LAS module with an instance-level contrastive loss to detect salient instances.", + "- Extensive experiments show that SCoCo outperforms existing weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods." + ], + "bbox": [ + 76, + 542, + 468, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 746, + 225, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Salient Instance Detection (SID)", + "text_level": 1, + "bbox": [ + 76, + 771, + 357, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SID aims to individualize visually prominent objects in the input image. Existing deep SID methods can be classified into two categories based on their learning labels: fully-supervised and weakly-supervised SID methods.", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-supervised SID methods rely on costly pixel-wise annotated masks for training supervision. Li et al. [25] propose to apply MAP [80] on instance proposals, which are", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generated based on binary saliency and contour maps, to select and segment salient instances. Fan et al. [14] devise a single-stage salient instance detection method, with a RoIMasking layer that utilizes both local and surrounding contexts of each instance for detection. Pei et al. [45] propose a spectral clustering method to obtain salient instances by exploring pixel masks and subitizing information of salient objects, and later a transformer-based method [46] to detect salient instances in a single stage without NMS for post-processing. Liu et al. [37] propose an interleaved execution strategy to incorporate global context and object contour information jointly to detect salient instances. RGB-D information has also been explored in CalibNet et al. [47], with a dual-branch cross-modal calibration method.", + "bbox": [ + 496, + 330, + 890, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised SID methods typically use bounding boxes and image-level labels, e.g., classes and subitizing as supervisions. Zhang et al. [80] propose a Maximum A Posteriori (MAP)-based subset optimization formulation to select a compact set of salient instances from the redundant box-level proposals, followed by CRF to help segment pixel-wise instances from box-level predictions. Tian et al. [61, 62] propose a triple-branch network to exploit salient boundaries, centroids, and regions jointly from class and subitizing labels for salient instance detection.", + "bbox": [ + 496, + 542, + 890, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their success, existing SID methods require human efforts to prepare different types of labels, which are expensive. In contrast, we propose to model the saliency coherence and contrast information based on self-supervised transformer features, for USID.", + "bbox": [ + 496, + 696, + 890, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Salient Object Detection (SOD)", + "text_level": 1, + "bbox": [ + 500, + 784, + 774, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SOD is an important task to scene understanding with a lot of methods proposed. Deep SOD methods can be divided into three groups w.r.t. the supervision signals: fully-supervised, weakly-supervised, and unsupervised.", + "bbox": [ + 496, + 808, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-supervised SOD methods mainly incorporate four types of deep techniques to learn salient object represent-", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Input image", + "bbox": [ + 143, + 88, + 204, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Sel", + "bbox": [ + 207, + 88, + 225, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "f-sup", + "bbox": [ + 227, + 88, + 246, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "ervis", + "bbox": [ + 250, + 88, + 269, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "ed tran", + "bbox": [ + 272, + 88, + 302, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "sform", + "bbox": [ + 303, + 88, + 326, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "her fear", + "bbox": [ + 328, + 88, + 354, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "tures", + "bbox": [ + 356, + 88, + 377, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "with", + "bbox": [ + 379, + 88, + 403, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "differ", + "bbox": [ + 406, + 88, + 431, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "nt visi", + "bbox": [ + 433, + 88, + 460, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "alizational", + "bbox": [ + 462, + 88, + 493, + 97 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Saliency contrasts for three random pixels", + "bbox": [ + 591, + 88, + 790, + 97 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2703", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tations from pixel-level annotated masks. First, deep feature fusion [34, 39, 68, 73, 83] is used to aggregate multilevel context information including low-level stimulus and high-level semantics for detecting salient objects. Second, attention mechanisms [30, 35, 36, 58, 84, 85] are used to reweight multi-scale features and enhance context learning to help the model focus on the salient regions and suppress noise of the background regions. Third, boundary enhancement [9, 33, 52, 86] is often applied to improve the localization and segmentation results by penalizing the errors of boundary-surrounding pixels. Last, recurrent mechanisms [22, 29, 60, 66, 84] may also be reused to iterate an identical network to predict and refine the saliency regions in a coarse-to-fine manner.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-supervised SOD methods have explored class labels, image captions, scribbles, bounding boxes, and points as supervisions. Wang et al. [67] propose the first weakly-supervised SOD method, which leverages image-level class labels to localize salient regions and then refines those predicted regions with CRF. Following [67], Li et al. [26] propose a pseudo saliency annotation updating scheme to refine the spatial coherence of predicted salient regions. Piao et al. [50] propose the multi-filter directive network to distill clean saliency maps from multiple class-based pseudo labels. Zeng et al. [78] devise a multi-source weak supervision framework to jointly utilize category labels, captions, and a set of unlabelled images for training. Based on scribble annotations, WSSA [82] and AGGM [77] are proposed to detect salient objects by exploring local consistency and structure information of objects. Liu et al. [40] propose to use bounding box supervisions to help address the over/under detection problems. Gao et al. [15] design an adaptive masked flood filling algorithm to detect salient objects from point-wise labels.", + "bbox": [ + 75, + 306, + 472, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised SOD methods can be classified into four categories according to the features and technologies used. First, conventional methods are typically based on handcrafted features [2, 5, 38, 49, 76, 92], e.g., local and global contrast [10, 38, 49], boundary priors [27, 76, 92], and spatial frequency [2], for detecting salient objects. These methods may not handle complex scenes as their low-level features have limited representation capacities. Second, a series of methods [32, 44, 72, 79, 81, 87, 88] try to distill clean labels from pseudo labels, which are produced by some of the aforementioned conventional methods, for learning SOD models. Specifically, they use distillation strategies like iterative refinement [44], causal debiasing [32], and mining textures around boundaries [89] to reduce the noise coming from the pseudo labels. Third, instead of learning from noisy pseudo labels, Yan et al. [75] propose a domain adaptive SOD method that learns from synthetic but clean saliency data. Fourth, FOUND [57] and SelfMask [55] also detect salient objects based on self-supervised deep", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "features. However, our SCoCo differs from them in both tasks and proposed techniques. First, they cannot predict instance-level saliency information, as shown in row 1 of Figure 1(b,c). Second, they only model the binary contrast between salient and non-salient regions; they are unable to suppress non-salient regions/objects with uncertain contrast degrees. For example, in row 2 of Figure 1(c,d), FOUND and SelfMask focus mainly on the most prominent woman, neglecting the umbrella, as the woman has much stronger semantics than the umbrella. In contrast, our SCoCo (Figure 1(d)) can correctly individualize salient instances by modeling intra-region saliency coherence and different levels of saliency contrast between regions.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, unlike the above SOD methods, which do not consider instance-level object saliency, our work aims to detect salient instances but without the need to use annotations for supervision.", + "bbox": [ + 496, + 287, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Unsupervised Object Detection (UOD)", + "text_level": 1, + "bbox": [ + 498, + 359, + 831, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UOD aims to localize objects without using any human annotations. Early methods [11, 63-65] explore inter-image similarities (e.g., clustering, ranking) to discover objects for an image collection. Recently, several methods learn to detect objects based on self-supervised deep features. LOST [56] and TokenCut [71] propose to segment one object per image based on the pixel graph of self-supervised deep features. FreeSOLO [69] and CutLER [70] learn to detect multiple objects by self-retraining [69] and MaskCut [70].", + "bbox": [ + 496, + 383, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these methods tend to detect any objects (salient, non-salient), our method learns to identify salient instances.", + "bbox": [ + 498, + 518, + 893, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 565, + 635, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning an unsupervised salient instance detector is challenging, as there are no explicit labels/knowledge to teach the detector to identify salient regions at the instance level. In this work, we observe some interesting cues in the self-supervised transformer features - they exhibit saliency coherence within local regions but saliency contrast of diverse degrees between different regions. If a detector can model such correlations well, it may be able to identify salient regions and group them into separate salient instances. Here, we propose the SCoCo method to mine regional saliency coherence and contrast information for USID.", + "bbox": [ + 496, + 584, + 893, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 shows the overview of our SCoCo. Given an input image, we first apply a self-supervised transformer (with frozen weights) to generate the corresponding deep features. We then feed these deep features separately to a novel global background adaptation (GBA) module (Section 3.1) and a novel locality-aware similarity (LAS) module (Section 3.2). The GBA module with its scene-level contrastive loss aims to find the suitable 'saliency thresholds' to distinguish salient regions from diverse background contents to produce a scene-level saliency map. Meanwhile,", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2704", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5e3e9751d07c12cb514c909341476bae32d44d3a3cfff2392b65f8bed21ee6af.jpg", + "image_caption": [ + "Figure 3. The overview of our unsupervised pipeline. Given an image, we first obtain its features $F$ from a self-supervised transformer (whose weights are frozen). We then use $F$ to produce pseudo scene- and instance-level supervisions $\\mathrm{P_{S+}}$ and $\\mathrm{P_{I+}^n}$ ( $1 \\leq n \\leq N$ ) by computing feature similarities, and feed $F$ to our novel GBA and LAS modules. The GBA module learns with a scene-level contrastive loss $\\mathcal{L}_S$ to predict a saliency map $\\mathrm{P_s}$ , while the LAS module learns with an instance-level contrastive loss $\\mathcal{L}_{\\mathcal{I}}$ to model saliency coherence and contrast features. The output features of both GBA and LAS modules are combined in the salient instance detection head to produce the final salient instance map $\\mathrm{P_I}$ ." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 851, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the LAS module works with an instance-level contrastive loss to learn intra-region saliency coherence and interregion saliency contrast features. These features are then fed into a salient instance detection head to produce salient instances, with the guidance of the scene-level saliency map produced by the GBA module.", + "bbox": [ + 75, + 277, + 472, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Global Background Adaptation (GBA) Module", + "text_level": 1, + "bbox": [ + 76, + 378, + 470, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of the GBA module is to adaptively dig saliency information from diverse background contents. As Figure 2(b-d) shows, it is challenging to find a fixed saliency threshold to select salient regions from different scenes with diverse complexities. To address this issue, we propose the GBA module, which has learnable saliency threshold functions working in a bifurcation-and-combination manner, to detect the fine-grained adaptive saliency thresholds. Specifically, as shown in Figure 4, the GBA module is designed to have two bifurcated branches at the first, with each of them equipped with its own learnable saliency threshold function, to handle the potential intra-scene saliency-background ambiguities separately. The two bifurcated branches are then combined with another saliency threshold-based branch to produce the scene-level saliency map.", + "bbox": [ + 75, + 402, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our GBA module differs from previous unsupervised salient object detection methods [55, 57] in the way of detecting saliency thresholds. Existing USOD methods typically rely on stacking convolutional operations to detect salient objects. However, such operations (even with dynamic convolution kernels [24, 59, 90]) are content-agnostic as convolution filters are shared across all the feature pixels. As a result, these methods often fail to detect the salient regions buried in diverse background contents.", + "bbox": [ + 75, + 631, + 468, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Structure. Given features $F \\in \\mathbb{R}^{H \\times W \\times C}$ , we first process it with two parallel branches, each consisting of a group of $1 \\times 1$ and $3 \\times 3$ conv operators, and a learnable saliency threshold function $f(x)$ . We define $f(x)$ as:", + "bbox": [ + 75, + 766, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = \\left(p _ {1} - p _ {2}\\right) x \\cdot \\sigma \\left(\\theta \\left(p _ {1} - p _ {2}\\right) x\\right) + p _ {2} x, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 840, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(x)$ is a smooth version of Swish rectifying function [42, 53], $p_1$ and $p_2$ are learnable parameters that can jointly", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "determine the upper and lower bounds of the function, $\\sigma$ is the Sigmoid function, and the learnable parameter $\\theta$ controls the \"selection degree\" of salient objects with respect to the background. Each $\\theta$ is learned using two fully connected (FC) layers upon a global image vector generated by the global average pooling (GAP) layer. We then use another branch to combine the former two branches, with another set of convolutional operations for feature transformation and learnable saliency threshold function to unify the separately modeled saliency thresholds, to produce the scene-level saliency map.", + "bbox": [ + 496, + 277, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-level Contrastive Loss. To boost the saliency learning of the GBA module, we introduce a scene-level contrastive loss $\\mathcal{L}_S$ . Since the image background is expected to have low saliency degrees, we first derive the scene-level pseudo saliency supervision $\\mathrm{P}_{\\mathrm{S}+}$ through inverting the identified background regions with low saliency degrees. However, we note that such pseudo-saliency maps may contain a significant number of noisy pixels, particularly those false positive background pixels with medium saliency degrees, that may hinder learning efficacy. To address this problem, we leverage the pixels with high saliency degrees in our instance-level saliency predictions $\\mathrm{P_I}$ (introduced in Section 3.2) to help classify those uncertain pixels into salient and non-salient ones. To this end, we take $\\mathrm{P_I}$ as positive samples and utilize scene-level saliency in different images as negative samples. $\\mathcal{L}_S$ can then be formulated as:", + "bbox": [ + 496, + 444, + 892, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {S}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau}}{e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau} + \\sum_ {k = 1 , k \\neq i} ^ {N} e ^ {\\left. \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {k}\\right) / \\tau \\right.}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 694, + 942, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{N}$ is the batch size, $\\mathrm{P}_{\\mathrm{S}}$ is the scene-level saliency prediction, $\\mathrm{d}$ measures the mean square error (MSE), and $\\tau$ is a temperature parameter.", + "bbox": [ + 496, + 750, + 893, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Locality-aware Similarity (LAS) Module", + "text_level": 1, + "bbox": [ + 500, + 801, + 846, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The LAS module aims to capture regional saliency coherence information from the self-supervised deep features. This information can then be used to help learn different levels of saliency contrast using an instance-level contrastive loss. We note that the latest saliency instance de", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2705", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2949c6faab73fb952a69682528e9af672190b3016d27a9d458ed8393d7c115c2.jpg", + "image_caption": [ + "Figure 4. The bifurcation-to-combination learning structure of the global background adaptation (GBA) module. The pink lines indicate the learning flows of learnable saliency threshold functions for extracting salient objects from different background contents." + ], + "image_footnote": [], + "bbox": [ + 135, + 87, + 410, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tection methods [37, 46] typically leverage the object detector head [6, 17] (a combination of convolutions and self-attention blocks) to classify and group salient instance pixels. However, although convolutional filters are good at detecting repeated patterns, they may fail in discerning pixel similarities. The self-attention mechanisms tend to model the interactions of pixel tokens but do not compare region-level saliency contrasts.", + "bbox": [ + 75, + 401, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, our LAS module first models intra-region feature coherence by computing channel-wise correlations with respect to regions surrounding every pixel location, and then feeds regional features to an instance-level contrastive loss to learn regional saliency contrasts for discriminating salient instances. The overview of LAS module is depicted in Figure 5.", + "bbox": [ + 75, + 522, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Structure. Given features $F \\in \\mathbb{R}^{H \\times W \\times C}$ , we downsample it to $F' \\in \\mathbb{R}^{H \\times W \\times C'}$ with a linear layer. We then compute cosine similarity between a positional feature $x \\in [1, H] \\times [1, W]$ and the surrounding region of $x$ in $F'$ , and collect them into a similarity feature tensor $S \\in \\mathbb{R}^{H \\times W \\times M \\times M \\times C'}$ . The formulation of $S$ is:", + "bbox": [ + 76, + 627, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (x, m) = \\frac {F _ {(x)} ^ {\\prime} \\cdot F _ {(x + m)} ^ {\\prime}}{\\| F _ {(x)} ^ {\\prime} \\| \\cdot \\| F _ {(x + m)} ^ {\\prime} \\|}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 728, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m \\in [-M + 1, M - 1] \\times [-M + 1, M - 1]$ , and $M \\times M$ is the region size. Eq. 3 aims to encode the local saliency coherence of each pixel $x$ into $S$ . We further apply convolutional filters on $S$ to compact regional features. The spatial size of $S$ is progressively aggregated from $M \\times M$ to $1 \\times 1$ with cascaded $3 \\times 3$ convolutional filters (without padding) to obtain $S'$ with the same spatial size as the input feature $F'$ . After each convolutional filter, we also", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "insert the batch normalization and rectifying linear unit layers. Finally, we combine $F'$ and $S'$ to produce the final features $\\hat{F}'$ with saliency coherence information, which further facilitates the saliency contrast modeling for differentiating salient instances.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance-level Contrastive Loss. We propose the instance-level contrastive loss $\\mathcal{L}_{\\mathcal{I}}$ to derive the saliency contrast information from $\\hat{F}^{\\prime}$ for salient instance discrimination, as the binary cross entropy (BCE) loss adopted by existing unsupervised SOD methods [56, 69, 70] cannot distinguish different salient instances.", + "bbox": [ + 496, + 167, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the self-supervised transformer features, we extract the key embeddings $F_{key}$ , and feed the self-affinity matrix of $F_{key}$ to Normalized Cut [54, 71] to find highly coherent salient pixels as pseudo salient instance supervision $\\mathrm{P}_{\\mathrm{I}+}$ in an iterative manner. We find one salient instance $\\mathbf{P}_{\\mathrm{I}+}^{k}$ in each iteration $k \\in [1, K]$ , and the iteration ends if the $\\mathbf{P}_{\\mathrm{I}+}^{K+1}$ is located in the low-saliency background. Note that $\\mathrm{P}_{I}$ is used in $\\mathcal{L}_S$ (Eq. 2) for penalizing pixels with uncertain saliency degrees. Meanwhile, we can measure the spatial discrepancy between predicted instances $\\mathbf{P}_{\\mathrm{I}}^{L}$ and $\\mathbf{P}_{\\mathrm{I}+}^{K}$ to formulate $\\mathcal{L}_{\\mathcal{I}}$ , where $L$ in the number of instances. To reduce the influence of noise that may be contained in pseudo labels, we propose to sample point vectors from $\\hat{F}'$ to compute $\\mathcal{L}_{\\mathcal{I}}$ , where $\\hat{F}', \\mathbf{P}_{\\mathrm{I}}^{L}$ , and $\\mathbf{P}_{\\mathrm{I}+}^{K}$ are spatially aligned for the sampling. By sampling a small subset of point vectors, we make the model less likely to be influenced by extreme noise that may be present in the full pseudo label set. Specifically, given $\\hat{F}'$ , we randomly sample $T(t \\in [1, T])$ triplets of point vectors $\\hat{F}_{\\mathrm{I}}^{l(t)}, \\hat{F}_{\\mathrm{I}+}^{l(t)}$ , and $\\hat{F}_{\\mathrm{I}-}^{l(t)}$ from $\\hat{F}'$ with respect to each predicted instance $\\mathbf{P}_{\\mathrm{I}}^{l}$ , its positive supervision $\\mathbf{P}_{\\mathrm{I}+}^{k}$ (IoU $(\\mathbf{P}_{\\mathrm{I}}^{l}, \\mathbf{P}_{\\mathrm{I}+}^{k}) \\geq 0.7$ ), and negative supervision $\\mathbf{P}_{\\mathrm{I}-}^{f}$ ( $f \\neq k, f \\in [1, K]$ ), respectively. $\\mathcal{L}_{\\mathcal{I}}$ can then be formulated as:", + "bbox": [ + 496, + 260, + 892, + 618 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {I}} = - \\frac {1}{L \\times T ^ {2}} \\sum_ {l = 1} ^ {L} \\sum_ {t = 1} ^ {T} \\sum_ {t = 1} ^ {T} \\log \\left(g \\left(\\hat {F} ^ {\\prime}, l, t\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 630, + 890, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g(\\cdot)$ is a contrastive estimation function, which is formulated as:", + "bbox": [ + 498, + 686, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(\\hat {F} ^ {\\prime}, l, t\\right) = \\frac {e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau}}{e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau} + \\sum_ {t = 1} ^ {T} e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} -} ^ {l (t)} \\right\\rangle / \\tau}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 728, + 908, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $sim\\langle \\cdot \\rangle$ is the cosine similarity function to compute the distance between two point vectors $\\hat{F}_{\\mathrm{I}}^{l(t)}$ and $\\hat{F}_{\\mathrm{I}+}^{\\prime l(t)}$ , and $l(t)$ represents the $t_{th}$ point vector of the $l_{th}$ instance. Each vector is extracted from $\\hat{F}^{\\prime}$ according to its spatial location. Our instance-level contrastive loss (Eq. 4) helps the network learn the amplified saliency contrast of different instances while keeping the saliency coherence within each instance by maintaining their self-similarities.", + "bbox": [ + 496, + 782, + 890, + 909 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2706", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b8f9f112f8f2f426b9a78a337766e43d6bdc56072be9c3bae94672b76204e51f.jpg", + "image_caption": [ + "Figure 5. The structure of locality-aware similarity (LAS) module." + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 816, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 232, + 209, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 257, + 284, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use DINO [8] to initialize the self-supervised transformer (i.e., ViT [13]) of our SCoCo network. Note that neither DINO nor our SCoCo uses any labeled data (e.g., class labels) for supervision. The input to both GBA and LAS modules, i.e., $F$ , is the key features of multi-head attention layers in the last feature block of DINO-based ViT. In the GBA module, we use the mean attention value of $F$ as the low saliency degree per image to find its scene-level saliency region $\\mathrm{P_{S+}}$ . In the LAS module, we sample $T = 32$ triplets of point vectors to formulate the loss $\\mathcal{L}_{\\mathcal{I}}$ . $\\tau$ in both $\\mathcal{L}_S$ and $\\mathcal{L}_{\\mathcal{I}}$ is set to 0.1. We adopt the head of Mask-RCNN [17] as our salient instance detection head. During training, each input image has a resolution of $1024 \\times 1024$ , and is augmented using largescale jittering [16]. The batch size $N$ is set to 4 and we use AdamW $(\\beta_1 = 0.9, \\beta_2 = 0.999)$ with step-wise learning rate decay for model optimization. We first train our SCoCo network for 10 epochs using the pseudo supervision derived from DINO. We then update the pseudo labels using SCoCo. Last, we train SCoCo with the updated pseudo labels for another epoch. SCoCo is trained on two 3090ti GPU cards.", + "bbox": [ + 75, + 280, + 472, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. We conduct our experiments on three SID datasets, including ILSO-1K [25], ILSO-2K [25], and SIS [46]. The (training/validation/test) sets in the ILSO-1K [25], ILSO-2K [25], and SIS [46] contain (500/200/300), (1000/400/600), and (7030/2100/1170) images, respectively. For each dataset, we only use the unlabeled images of the training set to train our SCoCo network, and we evaluate the performance on the corresponding test set. We adopt the mean Average Precision (mAP) metric for evaluation, following previous SID approaches.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 76, + 771, + 184, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we propose the first unsupervised method to detect salient instances, to evaluate our method comprehensively, we design several baselines from three aspects, inspired by related areas, as follows:", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- First, we note that existing unsupervised object detection (OD) methods, i.e., FreeSOLO [69], and CutLER [70], detect objects in a saliency-agnostic way. We retrain their", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models with our pseudo salient instance labels for comparison. Their models are initialized with the model parameters trained on their pseudo labels extracted from large-scale datasets, e.g., COCO [31] and ImageNet [23].", + "bbox": [ + 509, + 234, + 890, + 294 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Second, VitDet [28] is the best-performing method among fully-supervised object detection methods that are based on the ViT [13] backbone. Since we also use ViT as our backbone, we retrain VitDet with our salient instance pseudo labels for comparison.", + "- Third, we adapt existing USOD methods, i.e., SelfMask [55] and FOUND [57], to detect salient instances. To derive instance-level saliency predictions, we compute IoU between the region-level saliency predictions from USOD methods and the instance predictions from CutlER [70] (a state-of-the-art unsupervised object detection method)." + ], + "bbox": [ + 500, + 295, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 500, + 469, + 638, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method to 19 state-of-the-art SID methods, including seven fully-supervised methods: MSRNet-V1 [25], MSRNet-V2 [25], S4Net [14], SCG [37], MDNN [45], RDPNet [74], and OQTR [46]; six weakly-supervised methods: MAP [80], NLDF [41], DeepMask [21], PRM+D [12], IRN [3], and WSID-Net [62]; and six unsupervised baselines: FreeSOLO [69], CutLER [70], VitDet [28] initialized with the classification backbone, VitDet[28] initialized with the DINO-based backbone, SelfMask [55], and FOUND [57].", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Table 1 shows the quantitative results. First, we can see that our method outperforms all carefully designed unsupervised baselines (the bottom group). This shows that simply retraining unsupervised methods with our pseudo labels and/or filtering out nonsalient objects predicted by unsupervised object detectors is not effective in deriving accurate salient instance predictions. Second, we can also see that our method surpasses all existing weakly-supervised methods (middle group), while achieving comparable performances to the fully-supervised methods (first group). Existing weak supervision signals (i.e., object-level pixel masks, bounding boxes, classes, and subitizing labels) cannot provide instance-level supervision, which limits the performance of existing weakly-supervised methods. In contrast, even without supervision, SCoCo can learn to derive instance-level saliency information by modeling saliency coherence and contrast information based on", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2707", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f44b16ece7493a942518842b9ca03100b644f552c0482bb77dde470001aef606.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with 7 fully-supervised methods, 6 weakly-supervised methods, and 6 unsupervised baselines mentioned in Section 4.2. Column 1 shows the supervision and annotation types of the methods listed in Column 2. Column 3 shows their original tasks (SID: salient instance detection, SOD: salient object detection, OD: object detection, SIS: semantic instance segmentation). Column 4 shows backbone initialization choices, where Cls denotes that their backbones were pre-trained on class labels of ImageNet, and DINO is a self-supervised backbone pretraining method. - denotes missing results as codes/results are not publicly available. For each supervision category, best results are marked in bold, while second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
Supervision type & Annotation typeMethodOriginal taskInit.Datasets and Metrics (mAP IoU ↑)
ILSO-1K [25]ILSO-2K [25]SIS [46]
mAP50mAP70mAP50mAP70mAP50mAP70
Full SupervisionMSRNet-V1 [25]SIDCls65.3%52.1%----
MSRNet-V2 [25]SIDCls85.1%74.7%78.3%66.5%--
[Images are labeled with instance-level pixel masks.]S4Net [14]SIDCls82.2%59.6%73.1%52.9%86.7%63.6%
SCG [37]SIDCls88.8%78.5%79.8%68.9%83.4%71.2%
MDNN [45]SIDCls84.9%67.8%76.1%63.6%84.6%67.4%
RDPNet [74]SIDCls88.9%73.8%80.7%67.2%82.0%69.4%
OQTR [46]SIDCls89.2%81.0%81.4%72.0%88.1%81.7%
Weak SupervisionMAPΔ [80]SIDCls56.6%24.8%51.6%30.3%58.4%22.7%
[Images are labeled with region-level pixel masks*, bounding boxesΔ, classes°, or subitizing⊗]NLDF* [41]SODCls45.5%24.5%43.8%25.2%46.4%24.2%
DeepMaskΔ [51]ODCls37.1%20.5%35.4%18.4%36.4%21.3%
PRM+D⊗[12]SISCls49.6%31.2%43.7%29.8%52.9%34.2%
IRN° [3]SISCls57.1%37.4%50.2%38.4%60.1%40.8%
WSID-Net⊗[62]SIDCls68.3%51.7%59.4%44.6%68.4%47.1%
No SupervisionFreeSOLOΦ [69]ODCls56.8%43.7%50.7%35.8%55.4%43.1%
FreeSOLOΦ [70]ODDINO53.2%40.2%46.9%32.6%51.8%40.3%
[Million-levelhundred thousand-levelΦ, or thousand-levelunlabeled images are used for training.]CutLERΦ [70]ODCls68.6%54.8%63.2%48.3%66.9%54.4%
CutLERΦ [70]ODDINO65.4%52.1%59.8%44.9%62.4%51.5%
VitDet [28]SISCls55.6%43.6%51.9%35.8%54.4%44.1%
VitDet [28]SISDINO55.7%43.4%52.4%37.0%56.5%44.8%
SelfMask [55]SODDINO65.5%51.1%58.9%44.0%64.8%52.1%
FOUND [57]SODDINO66.2%51.3%61.4%45.9%67.1%53.8%
Ours [■]SIDDINO70.7%58.3%64.8%50.4%71.4%55.7%
", + "bbox": [ + 109, + 176, + 857, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "just the self-supervised Transformer features.", + "bbox": [ + 75, + 484, + 375, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Figure 6 compares the visual results of our method and 10 top performing methods from the three groups of methods in Table 1. We can see that the unsupervised baselines (columns 8 to 11) often fail to detect the complete salient instances and produce messy predictions. For the weakly-supervised methods (columns 5 to 7), MAP, which is trained with bounding box labels, may fail to segment complete objects even with the CRF post-processing. IRN and WSID-Net, which are trained with different class labels, may struggle to determine the boundaries between salient instances of the same class, e.g., the sausages in row 3. Fully-supervised methods (columns 2 to 4) tend to be trained to predict as many salient instances as possible. However, they may predict multiple instances from a salient instance of self-contrast (e.g., rows 1, 2, 5 and 6) and sometimes miss salient instances of solid colors (e.g., rows 4 and 6). In comparison, our method can successfully detect salient instances and delineate their boundaries, with the help of learning intra-instance saliency coherence and cross-instance saliency contrast.", + "bbox": [ + 75, + 501, + 470, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 814, + 230, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study of SCoCo. We first study the effectiveness of the proposed GBA and LAS modules, and the loss functions $\\mathcal{L}_S$ and $\\mathcal{L}_{\\mathcal{I}}$ used in them. The $\\mathrm{mAP}_{50}$ results in Table 2 show that solely applying either scene-level (GBA +", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70f36373b38642d50bfafb649431d88aa69823f09bd564bb05adce14f6c61c5c.jpg", + "table_caption": [ + "Table 2. Ablation study of the GBA and LAS modules, and the corresponding losses, $\\mathcal{L}_{\\mathcal{S}}$ and $\\mathcal{L}_{\\mathcal{I}}$ , on the SIS [46] test set using mAP metrics. Best performances are marked in bold." + ], + "table_footnote": [], + "table_body": "
MethodGBALSLASL2mAP50mAP70
SCoCo✓64.7%52.6%
✓✓67.0%53.5%
✓68.4%54.1%
67.4%53.1%
✓✓69.0%54.4%
✓✓✓✓71.4%55.7%
", + "bbox": [ + 529, + 530, + 864, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathcal{S}}$ or instance-level $(\\mathrm{LAS} + \\mathcal{L}_{\\mathcal{I}})$ saliency contrast learning may produce a performance decrease of $2 - 3\\%$ , and the performance drops by around $4\\%$ if we use the GBA or LAS module only. Combining all of them produces the best performance, which verifies the necessities of our designs.", + "bbox": [ + 498, + 636, + 890, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of the GBA Module. Figure 7 visualizes the scene-level salinocy map learned by the GBA module and the corresponding SID result. We can see that without the GBA module, the model tends to be distracted by high-contrast background regions, such as the shadow, and identifies them as salient regions. In contrast, using the GBA module can successfully suppress such noise by learning proper saliency thresholds.", + "bbox": [ + 496, + 715, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of the LAS Module. In Figure 8, we investigate the influence of our LAS module by visualizing feature correlations to points (purplish dots) in individual instances. We can see that the two pigs have non-uniform white and", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2708", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7b7ab1386600035a140e92d3d389693b8610706e0536dc591e6629882c4754a.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with existing state-of-the-art methods and baselines. Each color represents a unique salient instance." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 885, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65b184711d9a46e22a67642fc2df6ac78a774c90cc5f98398037e1837df3db2a.jpg", + "image_caption": [ + "Figure 7. Visualization of the GBA module." + ], + "image_footnote": [], + "bbox": [ + 109, + 381, + 439, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e11b02fad18bcf18a0c48c52c704079f9ad09b9562605f1dc00448619de10c0.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 460, + 248, + 541 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc9f992aa265ac132147baefa8cc3b10cd1bb92e62f9c258a35ffa2c75bb3370.jpg", + "image_caption": [ + "(b) Pixel-based saliency contrast w/o LAS" + ], + "image_footnote": [], + "bbox": [ + 251, + 459, + 447, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18e1ef80f001548192bba9c96d2b4a900d623ec60df669fa47ab00efaabb631c.jpg", + "image_caption": [ + "(c) Pixel-based saliency contrast w/ LAS", + "Figure 8. Visualization of the LAS module." + ], + "image_footnote": [], + "bbox": [ + 253, + 508, + 447, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "black colors on their bodies, causing our model to learn incomplete saliency coherence for each pig and produce overdetection results without the LAS module (b). With the LAS module, our model can separate the two pigs successfully by learning saliency contrast between instances while maintaining intra-instance saliency coherence.", + "bbox": [ + 75, + 589, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Backbone Initialization. Finally, we study the alternatives of self-supervised methods for initializing our ViT backbone. As shown in Table 3, in comparison to MoCo-v3 [18] and SWaV [7], we choose DINO [8] to initialize our backbone as it can provide better performances from its learned feature similarities. In addition, we select ViT-S as our backbone since ViT-B can only introduce minor improvements with larger parameters.", + "bbox": [ + 75, + 680, + 468, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 194, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have studied a new task, unsupervised salient instance detection. We have observed that self-supervised transformer features exhibit intra-region saliency coherence and different levels of inter-region", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5dba184c3f7d0f3400e3a2807ada324812f8049484022e73f957edb6603f917a.jpg", + "table_caption": [ + "Table 3. Quantitative analysis of backbone initialization choices. Best performances are marked in bold." + ], + "table_footnote": [], + "table_body": "
MethodInit.Arch.\\( \\mathbf{{mAP}}_{50} \\)\\( {\\mathbf{{mAP}}}_{70} \\)
SCoCoMoCo-v3 [18]ViT-S70.5%54.9%
SWaV [7]ViT-S70.2%54.3%
DINO [8]ViT-S71.4%55.7%
DINO [8]ViT-B71.9%55.6%
", + "bbox": [ + 542, + 417, + 849, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "saliency contrast, which can help differentiate salient instances. With this observation, we have proposed SCoCo, with two novel designs, to model such regional saliency coherence and contrast information. First, we propose a global background adaptation (GBA) module and a scene-level contrastive loss to find saliency thresholds for salient objects w.r.t. diverse background contents. Second, we propose a locality-aware similarity (LAS) module and an instance-level contrastive loss to model in-region saliency coherence and cross-region saliency contrasts for detecting salient instances. Extensive experiments have verified the effectiveness of our method against SOTA methods.", + "bbox": [ + 496, + 498, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method does have limitations. We note that SCoCo may not be able to detect small salient instances as it learns from high-level, low-resolution features. As shown in Figure 9, the person standing on the speedboat is too small to be detected by SCoCo. As a future work, it would be interesting to study multi-granularity salient feature similarity and contrast modeling to detect multi-scale salient instances.", + "bbox": [ + 496, + 680, + 890, + 784 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a98f5c76c139e8c0f39129323bba345224258fb95fe6535e3db6143fd57022ca.jpg", + "image_caption": [ + "Figure 9. SCoCo may fail to detect small salient instances." + ], + "image_footnote": [], + "bbox": [ + 506, + 797, + 880, + 873 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2709", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Junfeng He, Yossi Gandelsman, Inbar Mosseri, David E Jacobs, Kai Kohlhoff, Yael Pritch, and Michael Rubinstein. Deep saliency prior for reducing visual distraction. In CVPR, 2022. 1", + "[2] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 3", + "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 6, 7, 8", + "[4] Codruta Orniana Ancuti, Cosmin Ancuti, and Phillipe Bekaert. Enhancing by saliency-guided decolorization. In CVPR, 2011. 1", + "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVM, 2019. 3", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 2020. 8", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 6, 8", + "[9] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE TIP, 2020. 3", + "[10] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 2014. 3", + "[11] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In CVPR, 2015. 3", + "[12] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In CVPR, 2019. 6, 7", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6", + "[14] Ruochen Fan, Ming-Ming Cheng, Qibin Hou, Tai-Jiang Mu, Jingdong Wang, and Shi-Min Hu. S4net: Single stage salient-instance segmentation. In CVPR, 2019. 2, 6, 7, 8", + "[15] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI, 2022. 3", + "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 6" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 5, 6", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 8", + "[19] Brian Hu, Bhavan Vasu, and Anthony Hoogs. X-mir: Explainable medical image retrieval. In WACV, 2022. 1", + "[20] Lai Jiang, Yifei Li, Shengxi Li, Mai Xu, Se Lei, Yichen Guo, and Bo Huang. Does text attract attention on e-commerce images: A novel saliency prediction dataset and method. In CVPR, 2022. 1", + "[21] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. 1, 6", + "[22] Yun Yi Ke and Takahiro Tsubono. Recursive contour-saliency blending network for accurate salient object detection. In WACV, 2022. 3", + "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 6", + "[24] Duo Li, Jie Hu, Changhu Wang, Xiangtai Li, Qi She, Lei Zhu, Tong Zhang, and Qifeng Chen. Involution: Inverting the inheritance of convolution for visual recognition. In CVPR, 2021. 4", + "[25] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, 2017. 1, 2, 6, 7", + "[26] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In AAAI, 2018. 3", + "[27] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In ICCV, 2013. 3", + "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 6, 7", + "[29] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In CVPR, 2015. 3", + "[30] Zijian Liang, Pengjie Wang, Ke Xu, Pingping Zhang, and Rynson WH Lau. Weakly-supervised salient object detection on light fields. IEEE TIP, 2022. 3", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[32] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In AAAI, 2022. 1, 3", + "[33] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In CVPR, 2019. 3", + "[34] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016. 3", + "[35] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, 2018. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2710", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, 2021. 3", + "[37] Nian Liu, Wangbo Zhao, Ling Shao, and Junwei Han. Scg: Saliency and contour guided salient instance segmentation. IEEE TIP, 2021. 2, 5, 6, 7", + "[38] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE TPAMI, 2010. 3", + "[39] Yudong Liu, Yongtao Wang, Siwei Wang, TingTing Liang, Qijie Zhao, Zhi Tang, and Haibin Ling. Cbnet: A novel composite backbone network architecture for object detection. In AAAI, 2020. 3", + "[40] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE TIP, 2021. 3", + "[41] Zhiming Luo, Akshaya Mishra, Andrew Achkar, Justin Eichel, Shaozi Li, and Pierre-Marc Jodoin. Non-local deep features for salient object detection. In CVPR, 2017. 6, 7", + "[42] Ningning Ma, Xiangyu Zhang, Ming Liu, and Jian Sun. Activate or not: Learning customized activation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8042, 2021. 4", + "[43] S Mahdi H Miangoleh, Zoya Bylinskii, Eric Kee, Eli Shechtman, and Yaqiz Aksoy. Realistic saliency guided image enhancement. In CVPR, 2023. 1", + "[44] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. NeurIPS, 2019. 1, 3", + "[45] Jialun Pei, He Tang, Chao Liu, and Chuanbo Chen. Salient instance segmentation via subitizing and clustering. Neurocomputing, 2020. 2, 6, 7", + "[46] Jialun Pei, Tianyang Cheng, He Tang, and Chuanbo Chen. Transformer-based efficient salient instance segmentation networks with orientative query. IEEE TMM, 2022. 2, 5, 6, 7, 8", + "[47] Jialun Pei, Tao Jiang, He Tang, Nian Liu, Yueming Jin, Deng-Ping Fan, and Pheng-Ann Heng. Calibnet: Dual-branch cross-modal calibration for rgb-d salient instance segmentation. arXiv preprint arXiv:2307.08098, 2023. 2", + "[48] Guohao Peng, Jun Zhang, Heshan Li, and Danwei Wang. Attentional pyramid pooling of salient visual residuals for place recognition. In ICCV, 2021. 1", + "[49] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, 2012. 3", + "[50] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In ICCV, 2021. 3", + "[51] Pedro OO Pinheiro, Ronan Collobert, and Piotr Dólar. Learning to segment object candidates. In NeurIPS, 2015. 7", + "[52] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, 2019. 3", + "[53] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5", + "[55] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPR Workshop, 2022. 1, 3, 4, 6, 7, 8", + "[56] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3, 5", + "[57] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecý, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In CVPR, 2023. 1, 3, 4, 6, 7, 8", + "[58] Avishek Siris, Jianbo Jiao, Gary KL Tam, Xianghua Xie, and Rynson WH Lau. Scene context-aware salient object detection. In ICCV, 2021. 3", + "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, 2019. 4", + "[60] Chang Tang, Xinzhong Zhu, Xinwang Liu, and Pichao Wang. Salient object detection via recurrently aggregating spatial attention weighted cross-level deep features. In ICME, 2019. 3", + "[61] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Weakly-supervised salient instance detection. In BMVC, 2020. 2", + "[62] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Learning to detect instance-level salient objects using complementary image labels. IJCV, 2022. 1, 2, 6, 7, 8", + "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In CVPR, 2019. 3", + "[64] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In ECCV, 2020.", + "[65] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. NeurIPS, 2021. 3", + "[66] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang, and Xiang Ruan. Saliency detection with recurrent fully convolutional networks. In ECCV, 2016. 3", + "[67] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 3", + "[68] Xiang Wang, Huimin Ma, Xiaozhi Chen, and Shaodi You. Edge preserving and multi-scale contextual neural network for salient object detection. TIP, 2017. 3", + "[69] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In CVPR, 2022. 3, 5, 6, 7, 8", + "[70] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 3, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2711", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufreydaz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv:2209.00383, 2022. 3, 5", + "[72] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In CVPR, 2022. 1, 3", + "[73] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, 2020. 3", + "[74] Yu-Huan Wu, Yun Liu, Le Zhang, Wang Gao, and Ming-Ming Cheng. Regularized densely-connected pyramid network for salient instance segmentation. IEEE TIP, 2021. 1, 6, 7, 8", + "[75] Pengxiang Yan, Ziyi Wu, Mengmeng Liu, Kun Zeng, Liang Lin, and Guanbin Li. Unsupervised domain adaptive salient object detection through uncertainty-aware pseudolabel learning. In AAAI, 2022. 3", + "[76] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 3", + "[77] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In AAAI, 2021. 3", + "[78] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In CVPR, 2019. 3", + "[79] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In ICCV, 2017. 1, 3", + "[80] Jianming Zhang, Stan Sclaroff, Zhe Lin, Xiaohui Shen, Brian Price, and Radomir Mech. Unconstrained salient object detection via proposal subset optimization. In CVPR, 2016. 2, 6, 7, 8", + "[81] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In CVPR, 2018. 1, 3", + "[82] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In CVPR, 2020. 3", + "[83] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, 2018. 3", + "[84] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, 2018. 3", + "[85] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, 2019. 3", + "[86] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In CVPR, 2020. 3", + "[87] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE TCSVT, 2022. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 1, 3", + "[89] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 3", + "[90] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In CVPR, 2021. 4", + "[91] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In ICCV, 2021. 1", + "[92] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In CVPR, 2014. 3" + ], + "bbox": [ + 501, + 92, + 890, + 304 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2712", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_model.json b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da7a0da6740fca67294aaa26c67e956028eef40c --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.282, + 0.131, + 0.69, + 0.152 + ], + "angle": 0, + "content": "Unsupervised Salient Instance Detection" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.18, + 0.636, + 0.199 + ], + "angle": 0, + "content": "Xin Tian\\(^{1}\\) Ke Xu\\(^{2,\\dagger}\\) Rynson Lau\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.201, + 0.674, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Huawei Technologies \\(^{2}\\)City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.22, + 0.748, + 0.234 + ], + "angle": 0, + "content": "{xin.tian.831, kkangwing}@gmail.com, Rynson.Lau@cityu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.237, + 0.571, + 0.251 + ], + "angle": 0, + "content": "† joint corresponding authors" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.276, + 0.313, + 0.292 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.474, + 0.673 + ], + "angle": 0, + "content": "The significant amount of manual efforts in annotating pixel-level labels has triggered the advancement of unsupervised saliency learning. However, without supervision signals, state-of-the-art methods can only infer region-level saliency. In this paper, we propose to explore the unsupervised salient instance detection (USID) problem, for a more fine-grained visual understanding. Our key observation is that self-supervised transformer features may exhibit local similarities as well as different levels of contrast to other regions, which provide informative cues to identify salient instances. Hence, we propose SCoCo, a novel network that models saliency coherence and contrast for USID. SCoCo includes two novel modules: (1) a global background adaptation (GBA) module with a scene-level contrastive loss to extract salient regions from the scene by searching the adaptive \"saliency threshold\" in the self-supervised transformer features, and (2) a locality-aware similarity (LAS) module with an instance-level contrastive loss to group salient regions into instances by modeling the in-region saliency coherence and cross-region saliency contrasts. Extensive experiments show that SCoCo outperforms state-of-the-art weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Salient Instance Detection (SID) aims to detect instance-level salient objects. It can provide a fine-grained visual understanding of the scene, and is therefore able to facilitate a wide range of applications, e.g., image captioning [21], scene text recognition [20], medical image analysis [19], and others [1, 4, 43, 48, 91]. However, existing SID methods either require pixel-level labels [25, 74] or image-level labels [62] as supervisions, which are expensive to prepare. Despite the advancements in unsupervised saliency detection [32, 44, 57, 72, 79, 81, 87, 88], none of the existing methods have demonstrated the capability to learn instance-level saliency, which essentially requires fine-grained dif" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.275, + 0.583, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.394, + 0.56, + 0.405 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.275, + 0.658, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.394, + 0.655, + 0.404 + ], + "angle": 0, + "content": "(b) FOUND [57]" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.275, + 0.735, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.394, + 0.745, + 0.404 + ], + "angle": 0, + "content": "(c) SelfMask [55]" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.275, + 0.809, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.394, + 0.803, + 0.404 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.275, + 0.885, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.394, + 0.867, + 0.404 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.405, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Existing unsupervised salient object detection methods, FOUND [57] (b) and SelfMask [55] (c), cannot differentiate salient instances (rows 1 and 2), but tend to detect strong semantic instances, e.g., the woman in row 2. In contrast, our method (d) correctly detects all salient instances by modeling saliency coherence and different levels of saliency contrast." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.553 + ], + "angle": 0, + "content": "ferentiation not only between salient and non-salient regions but also between different salient regions (Figure 1). In this paper, we propose to study this unexplored task, i.e., unsupervised salient instance detection (USID)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Our key observation of this problem is built upon the advanced self-supervised transformer representations. We observe that their features exhibit saliency coherence (high similarities) within local regions and may show varying degrees of saliency contrast (lower similarities and dissimilarities) between different regions. In Figure 2, we visualize global saliency information within the self-supervised transformer features (b-d), and the saliency coherence and contrast between three feature pixels and the rest of the image (e-g), where each of these pixels is randomly selected from a salient instance. While the transformer features themselves may exhibit some degree of global saliency contrast (b-d), the computed pixel-based correlations show that each selected pixel tends to have high saliency coherence with pixels of the same instance but different levels of saliency contrast to pixels of other instances and the background (e-g). This observation inspires us to mine regional saliency coherence and contrast information for USID." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach, called SCoCo, to model Saliency Coherence and Contrast information from deep self-supervised features to address the USID problem. Specifically, SCoCo first models scene-level saliency contrast to determine the \"saliency thresh" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2702" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.145, + 0.089, + 0.205, + 0.098 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "header", + "bbox": [ + 0.208, + 0.089, + 0.226, + 0.098 + ], + "angle": 0, + "content": "Sel" + }, + { + "type": "header", + "bbox": [ + 0.228, + 0.089, + 0.248, + 0.098 + ], + "angle": 0, + "content": "f-sup" + }, + { + "type": "header", + "bbox": [ + 0.25, + 0.089, + 0.271, + 0.098 + ], + "angle": 0, + "content": "ervis" + }, + { + "type": "header", + "bbox": [ + 0.273, + 0.089, + 0.303, + 0.098 + ], + "angle": 0, + "content": "ed tran" + }, + { + "type": "header", + "bbox": [ + 0.305, + 0.089, + 0.327, + 0.098 + ], + "angle": 0, + "content": "sform" + }, + { + "type": "header", + "bbox": [ + 0.329, + 0.089, + 0.355, + 0.098 + ], + "angle": 0, + "content": "her fear" + }, + { + "type": "header", + "bbox": [ + 0.357, + 0.089, + 0.378, + 0.098 + ], + "angle": 0, + "content": "tures" + }, + { + "type": "header", + "bbox": [ + 0.38, + 0.089, + 0.405, + 0.098 + ], + "angle": 0, + "content": "with" + }, + { + "type": "header", + "bbox": [ + 0.407, + 0.089, + 0.432, + 0.098 + ], + "angle": 0, + "content": "differ" + }, + { + "type": "header", + "bbox": [ + 0.434, + 0.089, + 0.461, + 0.098 + ], + "angle": 0, + "content": "nt visi" + }, + { + "type": "header", + "bbox": [ + 0.463, + 0.089, + 0.495, + 0.098 + ], + "angle": 0, + "content": "alizational" + }, + { + "type": "header", + "bbox": [ + 0.592, + 0.089, + 0.792, + 0.098 + ], + "angle": 0, + "content": "Saliency contrasts for three random pixels" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.099, + 0.223, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.206, + 0.178, + 0.215 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.099, + 0.328, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.206, + 0.283, + 0.215 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.099, + 0.433, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.206, + 0.388, + 0.215 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.099, + 0.537, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.206, + 0.493, + 0.215 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.099, + 0.641, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.206, + 0.598, + 0.215 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.099, + 0.746, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.206, + 0.702, + 0.215 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.099, + 0.851, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.206, + 0.809, + 0.215 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.217, + 0.223, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.227, + 0.391, + 0.239 + ], + "angle": 0, + "content": "Global Saliency Contrast Degree \\(\\rightarrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.227, + 0.461, + 0.237 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.217, + 0.534, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.228, + 0.562, + 0.237 + ], + "angle": 0, + "content": "一" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.228, + 0.776, + 0.239 + ], + "angle": 0, + "content": "← Pixel-based Saliency Contrast Degree →" + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.228, + 0.836, + 0.238 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.894, + 0.323 + ], + "angle": 0, + "content": "Figure 2. Given an input image (a), we first visualize its corresponding self-supervised transformer features [8] (b), the pixel correlations to the [CLS] token (c), and the pixel correlations over the global scene (d). We then randomly sample three pixels (purple dots), each for one instance, and visualize the saliency contrast of the transformer features between the sampled pixel and all other pixels of the image by computing feature point correlations (e-g). We observe that while the self-supervised transformer features may show the global saliency contrast to some degree, pixels from the same instance tend to be more coherent, while pixels from different regions tend to have different levels of contrast. This inspires us to mine such regional saliency coherence and contrast for USID." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.527 + ], + "angle": 0, + "content": "old,\" which differentiates salient regions with respect to diverse background contents. We implement this via a novel global background adaptation (GBA) module with a scene-level contrastive loss, which learns to iteratively pull the saliency regions out of the complex background. SCoCo then models the within-region saliency coherence and cross-region saliency contrast information for grouping pixels into salient instances. We implement this by proposing a novel locality-aware similarity (LAS) module with an instance-level contrastive loss, which initially focuses on modeling the saliency coherence within individual regions and then shifts to model different levels of saliency contrast between different regions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.528, + 0.46, + 0.542 + ], + "angle": 0, + "content": "To summarize, this paper has three main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.543, + 0.469, + 0.617 + ], + "angle": 0, + "content": "- To our knowledge, we propose the first USID approach, named SCoCo, to learn to detect salient instances without any annotated labels, by modeling the intra-region coherence and inter-region contrast information from self-supervised transformer features." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.618, + 0.469, + 0.677 + ], + "angle": 0, + "content": "- SCoCo has two novel modules: a GBA module with a scene-level contrastive loss to find the saliency thresholds of different image contents, and a LAS module with an instance-level contrastive loss to detect salient instances." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.679, + 0.469, + 0.739 + ], + "angle": 0, + "content": "- Extensive experiments show that SCoCo outperforms existing weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.543, + 0.469, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.227, + 0.763 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.359, + 0.787 + ], + "angle": 0, + "content": "2.1. Salient Instance Detection (SID)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.855 + ], + "angle": 0, + "content": "SID aims to individualize visually prominent objects in the input image. Existing deep SID methods can be classified into two categories based on their learning labels: fully-supervised and weakly-supervised SID methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Fully-supervised SID methods rely on costly pixel-wise annotated masks for training supervision. Li et al. [25] propose to apply MAP [80] on instance proposals, which are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.542 + ], + "angle": 0, + "content": "generated based on binary saliency and contour maps, to select and segment salient instances. Fan et al. [14] devise a single-stage salient instance detection method, with a RoIMasking layer that utilizes both local and surrounding contexts of each instance for detection. Pei et al. [45] propose a spectral clustering method to obtain salient instances by exploring pixel masks and subitizing information of salient objects, and later a transformer-based method [46] to detect salient instances in a single stage without NMS for post-processing. Liu et al. [37] propose an interleaved execution strategy to incorporate global context and object contour information jointly to detect salient instances. RGB-D information has also been explored in CalibNet et al. [47], with a dual-branch cross-modal calibration method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Weakly-supervised SID methods typically use bounding boxes and image-level labels, e.g., classes and subitizing as supervisions. Zhang et al. [80] propose a Maximum A Posteriori (MAP)-based subset optimization formulation to select a compact set of salient instances from the redundant box-level proposals, followed by CRF to help segment pixel-wise instances from box-level predictions. Tian et al. [61, 62] propose a triple-branch network to exploit salient boundaries, centroids, and regions jointly from class and subitizing labels for salient instance detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Despite their success, existing SID methods require human efforts to prepare different types of labels, which are expensive. In contrast, we propose to model the saliency coherence and contrast information based on self-supervised transformer features, for USID." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.776, + 0.801 + ], + "angle": 0, + "content": "2.2. Salient Object Detection (SOD)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.87 + ], + "angle": 0, + "content": "SOD is an important task to scene understanding with a lot of methods proposed. Deep SOD methods can be divided into three groups w.r.t. the supervision signals: fully-supervised, weakly-supervised, and unsupervised." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fully-supervised SOD methods mainly incorporate four types of deep techniques to learn salient object represent-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2703" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "tations from pixel-level annotated masks. First, deep feature fusion [34, 39, 68, 73, 83] is used to aggregate multilevel context information including low-level stimulus and high-level semantics for detecting salient objects. Second, attention mechanisms [30, 35, 36, 58, 84, 85] are used to reweight multi-scale features and enhance context learning to help the model focus on the salient regions and suppress noise of the background regions. Third, boundary enhancement [9, 33, 52, 86] is often applied to improve the localization and segmentation results by penalizing the errors of boundary-surrounding pixels. Last, recurrent mechanisms [22, 29, 60, 66, 84] may also be reused to iterate an identical network to predict and refine the saliency regions in a coarse-to-fine manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Weakly-supervised SOD methods have explored class labels, image captions, scribbles, bounding boxes, and points as supervisions. Wang et al. [67] propose the first weakly-supervised SOD method, which leverages image-level class labels to localize salient regions and then refines those predicted regions with CRF. Following [67], Li et al. [26] propose a pseudo saliency annotation updating scheme to refine the spatial coherence of predicted salient regions. Piao et al. [50] propose the multi-filter directive network to distill clean saliency maps from multiple class-based pseudo labels. Zeng et al. [78] devise a multi-source weak supervision framework to jointly utilize category labels, captions, and a set of unlabelled images for training. Based on scribble annotations, WSSA [82] and AGGM [77] are proposed to detect salient objects by exploring local consistency and structure information of objects. Liu et al. [40] propose to use bounding box supervisions to help address the over/under detection problems. Gao et al. [15] design an adaptive masked flood filling algorithm to detect salient objects from point-wise labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Unsupervised SOD methods can be classified into four categories according to the features and technologies used. First, conventional methods are typically based on handcrafted features [2, 5, 38, 49, 76, 92], e.g., local and global contrast [10, 38, 49], boundary priors [27, 76, 92], and spatial frequency [2], for detecting salient objects. These methods may not handle complex scenes as their low-level features have limited representation capacities. Second, a series of methods [32, 44, 72, 79, 81, 87, 88] try to distill clean labels from pseudo labels, which are produced by some of the aforementioned conventional methods, for learning SOD models. Specifically, they use distillation strategies like iterative refinement [44], causal debiasing [32], and mining textures around boundaries [89] to reduce the noise coming from the pseudo labels. Third, instead of learning from noisy pseudo labels, Yan et al. [75] propose a domain adaptive SOD method that learns from synthetic but clean saliency data. Fourth, FOUND [57] and SelfMask [55] also detect salient objects based on self-supervised deep" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "features. However, our SCoCo differs from them in both tasks and proposed techniques. First, they cannot predict instance-level saliency information, as shown in row 1 of Figure 1(b,c). Second, they only model the binary contrast between salient and non-salient regions; they are unable to suppress non-salient regions/objects with uncertain contrast degrees. For example, in row 2 of Figure 1(c,d), FOUND and SelfMask focus mainly on the most prominent woman, neglecting the umbrella, as the woman has much stronger semantics than the umbrella. In contrast, our SCoCo (Figure 1(d)) can correctly individualize salient instances by modeling intra-region saliency coherence and different levels of saliency contrast between regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.35 + ], + "angle": 0, + "content": "In summary, unlike the above SOD methods, which do not consider instance-level object saliency, our work aims to detect salient instances but without the need to use annotations for supervision." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.36, + 0.833, + 0.377 + ], + "angle": 0, + "content": "2.3. Unsupervised Object Detection (UOD)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.519 + ], + "angle": 0, + "content": "UOD aims to localize objects without using any human annotations. Early methods [11, 63-65] explore inter-image similarities (e.g., clustering, ranking) to discover objects for an image collection. Recently, several methods learn to detect objects based on self-supervised deep features. LOST [56] and TokenCut [71] propose to segment one object per image based on the pixel graph of self-supervised deep features. FreeSOLO [69] and CutLER [70] learn to detect multiple objects by self-retraining [69] and MaskCut [70]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.894, + 0.551 + ], + "angle": 0, + "content": "While these methods tend to detect any objects (salient, non-salient), our method learns to identify salient instances." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.566, + 0.637, + 0.583 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.749 + ], + "angle": 0, + "content": "Learning an unsupervised salient instance detector is challenging, as there are no explicit labels/knowledge to teach the detector to identify salient regions at the instance level. In this work, we observe some interesting cues in the self-supervised transformer features - they exhibit saliency coherence within local regions but saliency contrast of diverse degrees between different regions. If a detector can model such correlations well, it may be able to identify salient regions and group them into separate salient instances. Here, we propose the SCoCo method to mine regional saliency coherence and contrast information for USID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Figure 3 shows the overview of our SCoCo. Given an input image, we first apply a self-supervised transformer (with frozen weights) to generate the corresponding deep features. We then feed these deep features separately to a novel global background adaptation (GBA) module (Section 3.1) and a novel locality-aware similarity (LAS) module (Section 3.2). The GBA module with its scene-level contrastive loss aims to find the suitable 'saliency thresholds' to distinguish salient regions from diverse background contents to produce a scene-level saliency map. Meanwhile," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "2704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.852, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.179, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Figure 3. The overview of our unsupervised pipeline. Given an image, we first obtain its features \\( F \\) from a self-supervised transformer (whose weights are frozen). We then use \\( F \\) to produce pseudo scene- and instance-level supervisions \\( \\mathrm{P_{S+}} \\) and \\( \\mathrm{P_{I+}^n} \\) (\\( 1 \\leq n \\leq N \\)) by computing feature similarities, and feed \\( F \\) to our novel GBA and LAS modules. The GBA module learns with a scene-level contrastive loss \\( \\mathcal{L}_S \\) to predict a saliency map \\( \\mathrm{P_s} \\), while the LAS module learns with an instance-level contrastive loss \\( \\mathcal{L}_{\\mathcal{I}} \\) to model saliency coherence and contrast features. The output features of both GBA and LAS modules are combined in the salient instance detection head to produce the final salient instance map \\( \\mathrm{P_I} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.473, + 0.37 + ], + "angle": 0, + "content": "the LAS module works with an instance-level contrastive loss to learn intra-region saliency coherence and interregion saliency contrast features. These features are then fed into a salient instance detection head to produce salient instances, with the guidance of the scene-level saliency map produced by the GBA module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.471, + 0.397 + ], + "angle": 0, + "content": "3.1. Global Background Adaptation (GBA) Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.631 + ], + "angle": 0, + "content": "The goal of the GBA module is to adaptively dig saliency information from diverse background contents. As Figure 2(b-d) shows, it is challenging to find a fixed saliency threshold to select salient regions from different scenes with diverse complexities. To address this issue, we propose the GBA module, which has learnable saliency threshold functions working in a bifurcation-and-combination manner, to detect the fine-grained adaptive saliency thresholds. Specifically, as shown in Figure 4, the GBA module is designed to have two bifurcated branches at the first, with each of them equipped with its own learnable saliency threshold function, to handle the potential intra-scene saliency-background ambiguities separately. The two bifurcated branches are then combined with another saliency threshold-based branch to produce the scene-level saliency map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.767 + ], + "angle": 0, + "content": "Our GBA module differs from previous unsupervised salient object detection methods [55, 57] in the way of detecting saliency thresholds. Existing USOD methods typically rely on stacking convolutional operations to detect salient objects. However, such operations (even with dynamic convolution kernels [24, 59, 90]) are content-agnostic as convolution filters are shared across all the feature pixels. As a result, these methods often fail to detect the salient regions buried in diverse background contents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Network Structure. Given features \\( F \\in \\mathbb{R}^{H \\times W \\times C} \\), we first process it with two parallel branches, each consisting of a group of \\( 1 \\times 1 \\) and \\( 3 \\times 3 \\) conv operators, and a learnable saliency threshold function \\( f(x) \\). We define \\( f(x) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.842, + 0.47, + 0.859 + ], + "angle": 0, + "content": "\\[\nf (x) = \\left(p _ {1} - p _ {2}\\right) x \\cdot \\sigma \\left(\\theta \\left(p _ {1} - p _ {2}\\right) x\\right) + p _ {2} x, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\( f(x) \\) is a smooth version of Swish rectifying function [42, 53], \\( p_1 \\) and \\( p_2 \\) are learnable parameters that can jointly" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.445 + ], + "angle": 0, + "content": "determine the upper and lower bounds of the function, \\(\\sigma\\) is the Sigmoid function, and the learnable parameter \\(\\theta\\) controls the \"selection degree\" of salient objects with respect to the background. Each \\(\\theta\\) is learned using two fully connected (FC) layers upon a global image vector generated by the global average pooling (GAP) layer. We then use another branch to combine the former two branches, with another set of convolutional operations for feature transformation and learnable saliency threshold function to unify the separately modeled saliency thresholds, to produce the scene-level saliency map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.893, + 0.687 + ], + "angle": 0, + "content": "Scene-level Contrastive Loss. To boost the saliency learning of the GBA module, we introduce a scene-level contrastive loss \\(\\mathcal{L}_S\\). Since the image background is expected to have low saliency degrees, we first derive the scene-level pseudo saliency supervision \\(\\mathrm{P}_{\\mathrm{S}+}\\) through inverting the identified background regions with low saliency degrees. However, we note that such pseudo-saliency maps may contain a significant number of noisy pixels, particularly those false positive background pixels with medium saliency degrees, that may hinder learning efficacy. To address this problem, we leverage the pixels with high saliency degrees in our instance-level saliency predictions \\(\\mathrm{P_I}\\) (introduced in Section 3.2) to help classify those uncertain pixels into salient and non-salient ones. To this end, we take \\(\\mathrm{P_I}\\) as positive samples and utilize scene-level saliency in different images as negative samples. \\(\\mathcal{L}_S\\) can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.695, + 0.943, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {S}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau}}{e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau} + \\sum_ {k = 1 , k \\neq i} ^ {N} e ^ {\\left. \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {k}\\right) / \\tau \\right.}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.795 + ], + "angle": 0, + "content": "where \\( \\mathrm{N} \\) is the batch size, \\( \\mathrm{P}_{\\mathrm{S}} \\) is the scene-level saliency prediction, \\( \\mathrm{d} \\) measures the mean square error (MSE), and \\( \\tau \\) is a temperature parameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.848, + 0.818 + ], + "angle": 0, + "content": "3.2. Locality-aware Similarity (LAS) Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The LAS module aims to capture regional saliency coherence information from the self-supervised deep features. This information can then be used to help learn different levels of saliency contrast using an instance-level contrastive loss. We note that the latest saliency instance de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2705" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.088, + 0.411, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.473, + 0.386 + ], + "angle": 0, + "content": "Figure 4. The bifurcation-to-combination learning structure of the global background adaptation (GBA) module. The pink lines indicate the learning flows of learnable saliency threshold functions for extracting salient objects from different background contents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.522 + ], + "angle": 0, + "content": "tection methods [37, 46] typically leverage the object detector head [6, 17] (a combination of convolutions and self-attention blocks) to classify and group salient instance pixels. However, although convolutional filters are good at detecting repeated patterns, they may fail in discerning pixel similarities. The self-attention mechanisms tend to model the interactions of pixel tokens but do not compare region-level saliency contrasts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.629 + ], + "angle": 0, + "content": "On the other hand, our LAS module first models intra-region feature coherence by computing channel-wise correlations with respect to regions surrounding every pixel location, and then feeds regional features to an instance-level contrastive loss to learn regional saliency contrasts for discriminating salient instances. The overview of LAS module is depicted in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Network Structure. Given features \\( F \\in \\mathbb{R}^{H \\times W \\times C} \\), we downsample it to \\( F' \\in \\mathbb{R}^{H \\times W \\times C'} \\) with a linear layer. We then compute cosine similarity between a positional feature \\( x \\in [1, H] \\times [1, W] \\) and the surrounding region of \\( x \\) in \\( F' \\), and collect them into a similarity feature tensor \\( S \\in \\mathbb{R}^{H \\times W \\times M \\times M \\times C'} \\). The formulation of \\( S \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.729, + 0.47, + 0.769 + ], + "angle": 0, + "content": "\\[\nS (x, m) = \\frac {F _ {(x)} ^ {\\prime} \\cdot F _ {(x + m)} ^ {\\prime}}{\\| F _ {(x)} ^ {\\prime} \\| \\cdot \\| F _ {(x + m)} ^ {\\prime} \\|}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "where \\(m \\in [-M + 1, M - 1] \\times [-M + 1, M - 1]\\), and \\(M \\times M\\) is the region size. Eq. 3 aims to encode the local saliency coherence of each pixel \\(x\\) into \\(S\\). We further apply convolutional filters on \\(S\\) to compact regional features. The spatial size of \\(S\\) is progressively aggregated from \\(M \\times M\\) to \\(1 \\times 1\\) with cascaded \\(3 \\times 3\\) convolutional filters (without padding) to obtain \\(S'\\) with the same spatial size as the input feature \\(F'\\). After each convolutional filter, we also" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "insert the batch normalization and rectifying linear unit layers. Finally, we combine \\( F' \\) and \\( S' \\) to produce the final features \\( \\hat{F}' \\) with saliency coherence information, which further facilitates the saliency contrast modeling for differentiating salient instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Instance-level Contrastive Loss. We propose the instance-level contrastive loss \\(\\mathcal{L}_{\\mathcal{I}}\\) to derive the saliency contrast information from \\(\\hat{F}^{\\prime}\\) for salient instance discrimination, as the binary cross entropy (BCE) loss adopted by existing unsupervised SOD methods [56, 69, 70] cannot distinguish different salient instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Given the self-supervised transformer features, we extract the key embeddings \\( F_{key} \\), and feed the self-affinity matrix of \\( F_{key} \\) to Normalized Cut [54, 71] to find highly coherent salient pixels as pseudo salient instance supervision \\( \\mathrm{P}_{\\mathrm{I}+} \\) in an iterative manner. We find one salient instance \\( \\mathbf{P}_{\\mathrm{I}+}^{k} \\) in each iteration \\( k \\in [1, K] \\), and the iteration ends if the \\( \\mathbf{P}_{\\mathrm{I}+}^{K+1} \\) is located in the low-saliency background. Note that \\( \\mathrm{P}_{I} \\) is used in \\( \\mathcal{L}_S \\) (Eq. 2) for penalizing pixels with uncertain saliency degrees. Meanwhile, we can measure the spatial discrepancy between predicted instances \\( \\mathbf{P}_{\\mathrm{I}}^{L} \\) and \\( \\mathbf{P}_{\\mathrm{I}+}^{K} \\) to formulate \\( \\mathcal{L}_{\\mathcal{I}} \\), where \\( L \\) in the number of instances. To reduce the influence of noise that may be contained in pseudo labels, we propose to sample point vectors from \\( \\hat{F}' \\) to compute \\( \\mathcal{L}_{\\mathcal{I}} \\), where \\( \\hat{F}', \\mathbf{P}_{\\mathrm{I}}^{L} \\), and \\( \\mathbf{P}_{\\mathrm{I}+}^{K} \\) are spatially aligned for the sampling. By sampling a small subset of point vectors, we make the model less likely to be influenced by extreme noise that may be present in the full pseudo label set. Specifically, given \\( \\hat{F}' \\), we randomly sample \\( T(t \\in [1, T]) \\) triplets of point vectors \\( \\hat{F}_{\\mathrm{I}}^{l(t)}, \\hat{F}_{\\mathrm{I}+}^{l(t)} \\), and \\( \\hat{F}_{\\mathrm{I}-}^{l(t)} \\) from \\( \\hat{F}' \\) with respect to each predicted instance \\( \\mathbf{P}_{\\mathrm{I}}^{l} \\), its positive supervision \\( \\mathbf{P}_{\\mathrm{I}+}^{k} \\) (IoU \\( (\\mathbf{P}_{\\mathrm{I}}^{l}, \\mathbf{P}_{\\mathrm{I}+}^{k}) \\geq 0.7 \\)), and negative supervision \\( \\mathbf{P}_{\\mathrm{I}-}^{f} \\) (\\( f \\neq k, f \\in [1, K] \\)), respectively. \\( \\mathcal{L}_{\\mathcal{I}} \\) can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.631, + 0.892, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {I}} = - \\frac {1}{L \\times T ^ {2}} \\sum_ {l = 1} ^ {L} \\sum_ {t = 1} ^ {T} \\sum_ {t = 1} ^ {T} \\log \\left(g \\left(\\hat {F} ^ {\\prime}, l, t\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.892, + 0.716 + ], + "angle": 0, + "content": "where \\(g(\\cdot)\\) is a contrastive estimation function, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.729, + 0.91, + 0.784 + ], + "angle": 0, + "content": "\\[\ng \\left(\\hat {F} ^ {\\prime}, l, t\\right) = \\frac {e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau}}{e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau} + \\sum_ {t = 1} ^ {T} e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} -} ^ {l (t)} \\right\\rangle / \\tau}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.91 + ], + "angle": 0, + "content": "where \\(sim\\langle \\cdot \\rangle\\) is the cosine similarity function to compute the distance between two point vectors \\(\\hat{F}_{\\mathrm{I}}^{l(t)}\\) and \\(\\hat{F}_{\\mathrm{I}+}^{\\prime l(t)}\\), and \\(l(t)\\) represents the \\(t_{th}\\) point vector of the \\(l_{th}\\) instance. Each vector is extracted from \\(\\hat{F}^{\\prime}\\) according to its spatial location. Our instance-level contrastive loss (Eq. 4) helps the network learn the amplified saliency contrast of different instances while keeping the saliency coherence within each instance by maintaining their self-similarities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2706" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.086, + 0.817, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.208, + 0.684, + 0.223 + ], + "angle": 0, + "content": "Figure 5. The structure of locality-aware similarity (LAS) module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.233, + 0.21, + 0.251 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.258, + 0.285, + 0.274 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.473, + 0.612 + ], + "angle": 0, + "content": "Implementation Details. We use DINO [8] to initialize the self-supervised transformer (i.e., ViT [13]) of our SCoCo network. Note that neither DINO nor our SCoCo uses any labeled data (e.g., class labels) for supervision. The input to both GBA and LAS modules, i.e., \\( F \\), is the key features of multi-head attention layers in the last feature block of DINO-based ViT. In the GBA module, we use the mean attention value of \\( F \\) as the low saliency degree per image to find its scene-level saliency region \\( \\mathrm{P_{S+}} \\). In the LAS module, we sample \\( T = 32 \\) triplets of point vectors to formulate the loss \\( \\mathcal{L}_{\\mathcal{I}} \\). \\( \\tau \\) in both \\( \\mathcal{L}_S \\) and \\( \\mathcal{L}_{\\mathcal{I}} \\) is set to 0.1. We adopt the head of Mask-RCNN [17] as our salient instance detection head. During training, each input image has a resolution of \\( 1024 \\times 1024 \\), and is augmented using largescale jittering [16]. The batch size \\( N \\) is set to 4 and we use AdamW \\( (\\beta_1 = 0.9, \\beta_2 = 0.999) \\) with step-wise learning rate decay for model optimization. We first train our SCoCo network for 10 epochs using the pseudo supervision derived from DINO. We then update the pseudo labels using SCoCo. Last, we train SCoCo with the updated pseudo labels for another epoch. SCoCo is trained on two 3090ti GPU cards." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. We conduct our experiments on three SID datasets, including ILSO-1K [25], ILSO-2K [25], and SIS [46]. The (training/validation/test) sets in the ILSO-1K [25], ILSO-2K [25], and SIS [46] contain (500/200/300), (1000/400/600), and (7030/2100/1170) images, respectively. For each dataset, we only use the unlabeled images of the training set to train our SCoCo network, and we evaluate the performance on the corresponding test set. We adopt the mean Average Precision (mAP) metric for evaluation, following previous SID approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.186, + 0.787 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Since we propose the first unsupervised method to detect salient instances, to evaluate our method comprehensively, we design several baselines from three aspects, inspired by related areas, as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- First, we note that existing unsupervised object detection (OD) methods, i.e., FreeSOLO [69], and CutLER [70], detect objects in a saliency-agnostic way. We retrain their" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.235, + 0.892, + 0.295 + ], + "angle": 0, + "content": "models with our pseudo salient instance labels for comparison. Their models are initialized with the model parameters trained on their pseudo labels extracted from large-scale datasets, e.g., COCO [31] and ImageNet [23]." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.296, + 0.892, + 0.37 + ], + "angle": 0, + "content": "- Second, VitDet [28] is the best-performing method among fully-supervised object detection methods that are based on the ViT [13] backbone. Since we also use ViT as our backbone, we retrain VitDet with our salient instance pseudo labels for comparison." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.371, + 0.892, + 0.461 + ], + "angle": 0, + "content": "- Third, we adapt existing USOD methods, i.e., SelfMask [55] and FOUND [57], to detect salient instances. To derive instance-level saliency predictions, we compute IoU between the region-level saliency predictions from USOD methods and the instance predictions from CutlER [70] (a state-of-the-art unsupervised object detection method)." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.296, + 0.892, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.64, + 0.485 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "We compare our method to 19 state-of-the-art SID methods, including seven fully-supervised methods: MSRNet-V1 [25], MSRNet-V2 [25], S4Net [14], SCG [37], MDNN [45], RDPNet [74], and OQTR [46]; six weakly-supervised methods: MAP [80], NLDF [41], DeepMask [21], PRM+D [12], IRN [3], and WSID-Net [62]; and six unsupervised baselines: FreeSOLO [69], CutLER [70], VitDet [28] initialized with the classification backbone, VitDet[28] initialized with the DINO-based backbone, SelfMask [55], and FOUND [57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparisons. Table 1 shows the quantitative results. First, we can see that our method outperforms all carefully designed unsupervised baselines (the bottom group). This shows that simply retraining unsupervised methods with our pseudo labels and/or filtering out nonsalient objects predicted by unsupervised object detectors is not effective in deriving accurate salient instance predictions. Second, we can also see that our method surpasses all existing weakly-supervised methods (middle group), while achieving comparable performances to the fully-supervised methods (first group). Existing weak supervision signals (i.e., object-level pixel masks, bounding boxes, classes, and subitizing labels) cannot provide instance-level supervision, which limits the performance of existing weakly-supervised methods. In contrast, even without supervision, SCoCo can learn to derive instance-level saliency information by modeling saliency coherence and contrast information based on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2707" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.174 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with 7 fully-supervised methods, 6 weakly-supervised methods, and 6 unsupervised baselines mentioned in Section 4.2. Column 1 shows the supervision and annotation types of the methods listed in Column 2. Column 3 shows their original tasks (SID: salient instance detection, SOD: salient object detection, OD: object detection, SIS: semantic instance segmentation). Column 4 shows backbone initialization choices, where Cls denotes that their backbones were pre-trained on class labels of ImageNet, and DINO is a self-supervised backbone pretraining method. - denotes missing results as codes/results are not publicly available. For each supervision category, best results are marked in bold, while second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.178, + 0.859, + 0.466 + ], + "angle": 0, + "content": "
Supervision type & Annotation typeMethodOriginal taskInit.Datasets and Metrics (mAP IoU ↑)
ILSO-1K [25]ILSO-2K [25]SIS [46]
mAP50mAP70mAP50mAP70mAP50mAP70
Full SupervisionMSRNet-V1 [25]SIDCls65.3%52.1%----
MSRNet-V2 [25]SIDCls85.1%74.7%78.3%66.5%--
[Images are labeled with instance-level pixel masks.]S4Net [14]SIDCls82.2%59.6%73.1%52.9%86.7%63.6%
SCG [37]SIDCls88.8%78.5%79.8%68.9%83.4%71.2%
MDNN [45]SIDCls84.9%67.8%76.1%63.6%84.6%67.4%
RDPNet [74]SIDCls88.9%73.8%80.7%67.2%82.0%69.4%
OQTR [46]SIDCls89.2%81.0%81.4%72.0%88.1%81.7%
Weak SupervisionMAPΔ [80]SIDCls56.6%24.8%51.6%30.3%58.4%22.7%
[Images are labeled with region-level pixel masks*, bounding boxesΔ, classes°, or subitizing⊗]NLDF* [41]SODCls45.5%24.5%43.8%25.2%46.4%24.2%
DeepMaskΔ [51]ODCls37.1%20.5%35.4%18.4%36.4%21.3%
PRM+D⊗[12]SISCls49.6%31.2%43.7%29.8%52.9%34.2%
IRN° [3]SISCls57.1%37.4%50.2%38.4%60.1%40.8%
WSID-Net⊗[62]SIDCls68.3%51.7%59.4%44.6%68.4%47.1%
No SupervisionFreeSOLOΦ [69]ODCls56.8%43.7%50.7%35.8%55.4%43.1%
FreeSOLOΦ [70]ODDINO53.2%40.2%46.9%32.6%51.8%40.3%
[Million-levelhundred thousand-levelΦ, or thousand-levelunlabeled images are used for training.]CutLERΦ [70]ODCls68.6%54.8%63.2%48.3%66.9%54.4%
CutLERΦ [70]ODDINO65.4%52.1%59.8%44.9%62.4%51.5%
VitDet [28]SISCls55.6%43.6%51.9%35.8%54.4%44.1%
VitDet [28]SISDINO55.7%43.4%52.4%37.0%56.5%44.8%
SelfMask [55]SODDINO65.5%51.1%58.9%44.0%64.8%52.1%
FOUND [57]SODDINO66.2%51.3%61.4%45.9%67.1%53.8%
Ours [■]SIDDINO70.7%58.3%64.8%50.4%71.4%55.7%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.377, + 0.5 + ], + "angle": 0, + "content": "just the self-supervised Transformer features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.471, + 0.804 + ], + "angle": 0, + "content": "Qualitative Comparison. Figure 6 compares the visual results of our method and 10 top performing methods from the three groups of methods in Table 1. We can see that the unsupervised baselines (columns 8 to 11) often fail to detect the complete salient instances and produce messy predictions. For the weakly-supervised methods (columns 5 to 7), MAP, which is trained with bounding box labels, may fail to segment complete objects even with the CRF post-processing. IRN and WSID-Net, which are trained with different class labels, may struggle to determine the boundaries between salient instances of the same class, e.g., the sausages in row 3. Fully-supervised methods (columns 2 to 4) tend to be trained to predict as many salient instances as possible. However, they may predict multiple instances from a salient instance of self-contrast (e.g., rows 1, 2, 5 and 6) and sometimes miss salient instances of solid colors (e.g., rows 4 and 6). In comparison, our method can successfully detect salient instances and delineate their boundaries, with the help of learning intra-instance saliency coherence and cross-instance saliency contrast." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.231, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ablation Study of SCoCo. We first study the effectiveness of the proposed GBA and LAS modules, and the loss functions \\(\\mathcal{L}_S\\) and \\(\\mathcal{L}_{\\mathcal{I}}\\) used in them. The \\(\\mathrm{mAP}_{50}\\) results in Table 2 show that solely applying either scene-level (GBA +" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Table 2. Ablation study of the GBA and LAS modules, and the corresponding losses, \\(\\mathcal{L}_{\\mathcal{S}}\\) and \\(\\mathcal{L}_{\\mathcal{I}}\\), on the SIS [46] test set using mAP metrics. Best performances are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.531, + 0.865, + 0.621 + ], + "angle": 0, + "content": "
MethodGBALSLASL2mAP50mAP70
SCoCo✓64.7%52.6%
✓✓67.0%53.5%
✓68.4%54.1%
67.4%53.1%
✓✓69.0%54.4%
✓✓✓✓71.4%55.7%
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathcal{S}}\\) or instance-level \\((\\mathrm{LAS} + \\mathcal{L}_{\\mathcal{I}})\\) saliency contrast learning may produce a performance decrease of \\(2 - 3\\%\\), and the performance drops by around \\(4\\%\\) if we use the GBA or LAS module only. Combining all of them produces the best performance, which verifies the necessities of our designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Visualization of the GBA Module. Figure 7 visualizes the scene-level salinocy map learned by the GBA module and the corresponding SID result. We can see that without the GBA module, the model tends to be distracted by high-contrast background regions, such as the shadow, and identifies them as salient regions. In contrast, using the GBA module can successfully suppress such noise by learning proper saliency thresholds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Visualization of the LAS Module. In Figure 8, we investigate the influence of our LAS module by visualizing feature correlations to points (purplish dots) in individual instances. We can see that the two pigs have non-uniform white and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2708" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.087, + 0.887, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.355, + 0.88, + 0.37 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with existing state-of-the-art methods and baselines. Each color represents a unique salient instance." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.382, + 0.44, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.443, + 0.404, + 0.457 + ], + "angle": 0, + "content": "Figure 7. Visualization of the GBA module." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.462, + 0.249, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.545, + 0.213, + 0.556 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.46, + 0.449, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.497, + 0.448, + 0.508 + ], + "angle": 0, + "content": "(b) Pixel-based saliency contrast w/o LAS" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.51, + 0.449, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.545, + 0.445, + 0.556 + ], + "angle": 0, + "content": "(c) Pixel-based saliency contrast w/ LAS" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.564, + 0.402, + 0.577 + ], + "angle": 0, + "content": "Figure 8. Visualization of the LAS module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.469, + 0.68 + ], + "angle": 0, + "content": "black colors on their bodies, causing our model to learn incomplete saliency coherence for each pig and produce overdetection results without the LAS module (b). With the LAS module, our model can separate the two pigs successfully by learning saliency contrast between instances while maintaining intra-instance saliency coherence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.801 + ], + "angle": 0, + "content": "Backbone Initialization. Finally, we study the alternatives of self-supervised methods for initializing our ViT backbone. As shown in Table 3, in comparison to MoCo-v3 [18] and SWaV [7], we choose DINO [8] to initialize our backbone as it can provide better performances from its learned feature similarities. In addition, we select ViT-S as our backbone since ViT-B can only introduce minor improvements with larger parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.196, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have studied a new task, unsupervised salient instance detection. We have observed that self-supervised transformer features exhibit intra-region saliency coherence and different levels of inter-region" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.383, + 0.89, + 0.41 + ], + "angle": 0, + "content": "Table 3. Quantitative analysis of backbone initialization choices. Best performances are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.418, + 0.851, + 0.487 + ], + "angle": 0, + "content": "
MethodInit.Arch.\\( \\mathbf{{mAP}}_{50} \\)\\( {\\mathbf{{mAP}}}_{70} \\)
SCoCoMoCo-v3 [18]ViT-S70.5%54.9%
SWaV [7]ViT-S70.2%54.3%
DINO [8]ViT-S71.4%55.7%
DINO [8]ViT-B71.9%55.6%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.68 + ], + "angle": 0, + "content": "saliency contrast, which can help differentiate salient instances. With this observation, we have proposed SCoCo, with two novel designs, to model such regional saliency coherence and contrast information. First, we propose a global background adaptation (GBA) module and a scene-level contrastive loss to find saliency thresholds for salient objects w.r.t. diverse background contents. Second, we propose a locality-aware similarity (LAS) module and an instance-level contrastive loss to model in-region saliency coherence and cross-region saliency contrasts for detecting salient instances. Extensive experiments have verified the effectiveness of our method against SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Our method does have limitations. We note that SCoCo may not be able to detect small salient instances as it learns from high-level, low-resolution features. As shown in Figure 9, the person standing on the speedboat is too small to be detected by SCoCo. As a future work, it would be interesting to study multi-granularity salient feature similarity and contrast modeling to detect multi-scale salient instances." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.798, + 0.882, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.878, + 0.87, + 0.892 + ], + "angle": 0, + "content": "Figure 9. SCoCo may fail to detect small salient instances." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2709" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Junfeng He, Yossi Gandelsman, Inbar Mosseri, David E Jacobs, Kai Kohlhoff, Yael Pritch, and Michael Rubinstein. Deep saliency prior for reducing visual distraction. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Codruta Orniana Ancuti, Cosmin Ancuti, and Phillipe Bekaert. Enhancing by saliency-guided decolorization. In CVPR, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVM, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.469, + 0.368 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[9] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE TIP, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[10] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[11] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[12] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[14] Ruochen Fan, Ming-Ming Cheng, Qibin Hou, Tai-Jiang Mu, Jingdong Wang, and Shi-Min Hu. S4net: Single stage salient-instance segmentation. In CVPR, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[15] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.193 + ], + "angle": 0, + "content": "[19] Brian Hu, Bhavan Vasu, and Anthony Hoogs. X-mir: Explainable medical image retrieval. In WACV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[20] Lai Jiang, Yifei Li, Shengxi Li, Mai Xu, Se Lei, Yichen Guo, and Bo Huang. Does text attract attention on e-commerce images: A novel saliency prediction dataset and method. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[21] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[22] Yun Yi Ke and Takahiro Tsubono. Recursive contour-saliency blending network for accurate salient object detection. In WACV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[24] Duo Li, Jie Hu, Changhu Wang, Xiangtai Li, Qi She, Lei Zhu, Tong Zhang, and Qifeng Chen. Involution: Inverting the inheritance of convolution for visual recognition. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[25] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[26] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In AAAI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[27] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In ICCV, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[29] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[30] Zijian Liang, Pengjie Wang, Ke Xu, Pingping Zhang, and Rynson WH Lau. Weakly-supervised salient object detection on light fields. IEEE TIP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[32] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In AAAI, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[33] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[34] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "[36] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[37] Nian Liu, Wangbo Zhao, Ling Shao, and Junwei Han. Scg: Saliency and contour guided salient instance segmentation. IEEE TIP, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[38] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE TPAMI, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[39] Yudong Liu, Yongtao Wang, Siwei Wang, TingTing Liang, Qijie Zhao, Zhi Tang, and Haibin Ling. Cbnet: A novel composite backbone network architecture for object detection. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[40] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE TIP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[41] Zhiming Luo, Akshaya Mishra, Andrew Achkar, Justin Eichel, Shaozi Li, and Pierre-Marc Jodoin. Non-local deep features for salient object detection. In CVPR, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[42] Ningning Ma, Xiangyu Zhang, Ming Liu, and Jian Sun. Activate or not: Learning customized activation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8042, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[43] S Mahdi H Miangoleh, Zoya Bylinskii, Eric Kee, Eli Shechtman, and Yaqiz Aksoy. Realistic saliency guided image enhancement. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[44] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. NeurIPS, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[45] Jialun Pei, He Tang, Chao Liu, and Chuanbo Chen. Salient instance segmentation via subitizing and clustering. Neurocomputing, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[46] Jialun Pei, Tianyang Cheng, He Tang, and Chuanbo Chen. Transformer-based efficient salient instance segmentation networks with orientative query. IEEE TMM, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[47] Jialun Pei, Tao Jiang, He Tang, Nian Liu, Yueming Jin, Deng-Ping Fan, and Pheng-Ann Heng. Calibnet: Dual-branch cross-modal calibration for rgb-d salient instance segmentation. arXiv preprint arXiv:2307.08098, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[48] Guohao Peng, Jun Zhang, Heshan Li, and Danwei Wang. Attentional pyramid pooling of salient visual residuals for place recognition. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[49] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[50] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[51] Pedro OO Pinheiro, Ronan Collobert, and Piotr Dólar. Learning to segment object candidates. In NeurIPS, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[52] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[53] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[54] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[55] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPR Workshop, 2022. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[56] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[57] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecý, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In CVPR, 2023. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[58] Avishek Siris, Jianbo Jiao, Gary KL Tam, Xianghua Xie, and Rynson WH Lau. Scene context-aware salient object detection. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[60] Chang Tang, Xinzhong Zhu, Xinwang Liu, and Pichao Wang. Salient object detection via recurrently aggregating spatial attention weighted cross-level deep features. In ICME, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.425, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[61] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Weakly-supervised salient instance detection. In BMVC, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[62] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Learning to detect instance-level salient objects using complementary image labels. IJCV, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.57, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[64] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[65] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[66] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang, and Xiang Ruan. Saliency detection with recurrent fully convolutional networks. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[67] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[68] Xiang Wang, Huimin Ma, Xiaozhi Chen, and Shaodi You. Edge preserving and multi-scale contextual neural network for salient object detection. TIP, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[69] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In CVPR, 2022. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[70] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2711" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[71] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufreydaz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv:2209.00383, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[72] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[73] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.47, + 0.306 + ], + "angle": 0, + "content": "[74] Yu-Huan Wu, Yun Liu, Le Zhang, Wang Gao, and Ming-Ming Cheng. Regularized densely-connected pyramid network for salient instance segmentation. IEEE TIP, 2021. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.47, + 0.364 + ], + "angle": 0, + "content": "[75] Pengxiang Yan, Ziyi Wu, Mengmeng Liu, Kun Zeng, Liang Lin, and Guanbin Li. Unsupervised domain adaptive salient object detection through uncertainty-aware pseudolabel learning. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[76] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[77] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[78] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[79] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In ICCV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.47, + 0.595 + ], + "angle": 0, + "content": "[80] Jianming Zhang, Stan Sclaroff, Zhe Lin, Xiaohui Shen, Brian Price, and Radomir Mech. Unconstrained salient object detection via proposal subset optimization. In CVPR, 2016. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.598, + 0.47, + 0.64 + ], + "angle": 0, + "content": "[81] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In CVPR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[82] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.726 + ], + "angle": 0, + "content": "[83] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[84] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[85] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[86] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[87] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE TCSVT, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[88] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[89] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[90] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[91] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[92] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In CVPR, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2712" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_origin.pdf b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16f6207dfdee6bc6b8fe981bc488eedd7e7027f6 --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/d3668a6f-65c6-4577-bc44-ffa98f5eddad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7185c89114927925742defef386eeada93c7b5e3690a861aa5c5cf3e25d87ca +size 579106 diff --git a/2024/Unsupervised Salient Instance Detection/full.md b/2024/Unsupervised Salient Instance Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e68941a8e0746216bf90e7f47e070eba45489d0 --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/full.md @@ -0,0 +1,371 @@ +# Unsupervised Salient Instance Detection + +Xin Tian $^{1}$ Ke Xu $^{2,\dagger}$ Rynson Lau $^{2,\dagger}$ + +$^{1}$ Huawei Technologies $^{2}$ City University of Hong Kong + +{xin.tian.831, kkangwing}@gmail.com, Rynson.Lau@cityu.edu.hk + +† joint corresponding authors + +# Abstract + +The significant amount of manual efforts in annotating pixel-level labels has triggered the advancement of unsupervised saliency learning. However, without supervision signals, state-of-the-art methods can only infer region-level saliency. In this paper, we propose to explore the unsupervised salient instance detection (USID) problem, for a more fine-grained visual understanding. Our key observation is that self-supervised transformer features may exhibit local similarities as well as different levels of contrast to other regions, which provide informative cues to identify salient instances. Hence, we propose SCoCo, a novel network that models saliency coherence and contrast for USID. SCoCo includes two novel modules: (1) a global background adaptation (GBA) module with a scene-level contrastive loss to extract salient regions from the scene by searching the adaptive "saliency threshold" in the self-supervised transformer features, and (2) a locality-aware similarity (LAS) module with an instance-level contrastive loss to group salient regions into instances by modeling the in-region saliency coherence and cross-region saliency contrasts. Extensive experiments show that SCoCo outperforms state-of-the-art weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods. + +# 1. Introduction + +Salient Instance Detection (SID) aims to detect instance-level salient objects. It can provide a fine-grained visual understanding of the scene, and is therefore able to facilitate a wide range of applications, e.g., image captioning [21], scene text recognition [20], medical image analysis [19], and others [1, 4, 43, 48, 91]. However, existing SID methods either require pixel-level labels [25, 74] or image-level labels [62] as supervisions, which are expensive to prepare. Despite the advancements in unsupervised saliency detection [32, 44, 57, 72, 79, 81, 87, 88], none of the existing methods have demonstrated the capability to learn instance-level saliency, which essentially requires fine-grained dif + +![](images/749713cb83f1b6bb3bb5394ea778f4943b7dc4ec04775556f55e4b1b11e7e66d.jpg) +(a) Image + +![](images/480dcb5c837848e086cd1f085018e00232c7fdbd43e4e4e4a54083a922ee24f4.jpg) +(b) FOUND [57] + +![](images/9ea38f01bb960d52fc92c00555d4cc2e9b6e5a23b90e75aa4cbc1c17825de35e.jpg) +(c) SelfMask [55] + +![](images/53ddf3eae9959d73c0d1281b0f3a97f7ca3bbf4f694d4949a5aac9aa3c303521.jpg) +(d) Ours +Figure 1. Existing unsupervised salient object detection methods, FOUND [57] (b) and SelfMask [55] (c), cannot differentiate salient instances (rows 1 and 2), but tend to detect strong semantic instances, e.g., the woman in row 2. In contrast, our method (d) correctly detects all salient instances by modeling saliency coherence and different levels of saliency contrast. + +![](images/697767f4816ac2c15169534c711550fc3c370169b02f646918398e99fa3e2ead.jpg) +(e) GT + +ferentiation not only between salient and non-salient regions but also between different salient regions (Figure 1). In this paper, we propose to study this unexplored task, i.e., unsupervised salient instance detection (USID). + +Our key observation of this problem is built upon the advanced self-supervised transformer representations. We observe that their features exhibit saliency coherence (high similarities) within local regions and may show varying degrees of saliency contrast (lower similarities and dissimilarities) between different regions. In Figure 2, we visualize global saliency information within the self-supervised transformer features (b-d), and the saliency coherence and contrast between three feature pixels and the rest of the image (e-g), where each of these pixels is randomly selected from a salient instance. While the transformer features themselves may exhibit some degree of global saliency contrast (b-d), the computed pixel-based correlations show that each selected pixel tends to have high saliency coherence with pixels of the same instance but different levels of saliency contrast to pixels of other instances and the background (e-g). This observation inspires us to mine regional saliency coherence and contrast information for USID. + +In this paper, we propose a novel approach, called SCoCo, to model Saliency Coherence and Contrast information from deep self-supervised features to address the USID problem. Specifically, SCoCo first models scene-level saliency contrast to determine the "saliency thresh + +![](images/a849b1a076fdd06f7fb82676c01b2a862ca2b4fec1d0697cffaed2ff4d40dce7.jpg) +(a) + +![](images/cde78aa92b16fdab70119187ee7cb9baa6d9414b8cfaf0938c77dabbfd70fb6c.jpg) +(b) + +![](images/ccfb52e2ea8b07b7e3a119fa00d7815df00501e4ebedffbc6a5189105f947cf3.jpg) +(c) + +![](images/e8ad4739a5d955d6d45bea70645daa869f9f8a1de36540354ee7e1e80aec1940.jpg) +(d) +High + +![](images/c57dba02e7ffe8ddd961ac83d0d759a88e4830fb4387cbfa64d63386348065af.jpg) +(e) +← Pixel-based Saliency Contrast Degree → + +![](images/88a1d0ce9fdfd94df4b79c17aeeca1ae27e31753d6557332fc4938b28b7f489c.jpg) +(f) + +![](images/aab17e91de9f16d9bfaa3a7d8bcda67d7f55aa4730e034a70a62d4ea8b772323.jpg) +(g) +High + +![](images/368200a641c5a62059d7fb4166b41c715b67a218e86063e23636a6cd6c5fad3e.jpg) +Global Saliency Contrast Degree $\rightarrow$ +Figure 2. Given an input image (a), we first visualize its corresponding self-supervised transformer features [8] (b), the pixel correlations to the [CLS] token (c), and the pixel correlations over the global scene (d). We then randomly sample three pixels (purple dots), each for one instance, and visualize the saliency contrast of the transformer features between the sampled pixel and all other pixels of the image by computing feature point correlations (e-g). We observe that while the self-supervised transformer features may show the global saliency contrast to some degree, pixels from the same instance tend to be more coherent, while pixels from different regions tend to have different levels of contrast. This inspires us to mine such regional saliency coherence and contrast for USID. + +![](images/85f5b9f7b7d35ad3929348ee40b1b82ee268eef7c2fc61c231ffbed97e639dc2.jpg) +一 + +old," which differentiates salient regions with respect to diverse background contents. We implement this via a novel global background adaptation (GBA) module with a scene-level contrastive loss, which learns to iteratively pull the saliency regions out of the complex background. SCoCo then models the within-region saliency coherence and cross-region saliency contrast information for grouping pixels into salient instances. We implement this by proposing a novel locality-aware similarity (LAS) module with an instance-level contrastive loss, which initially focuses on modeling the saliency coherence within individual regions and then shifts to model different levels of saliency contrast between different regions. + +To summarize, this paper has three main contributions: + +- To our knowledge, we propose the first USID approach, named SCoCo, to learn to detect salient instances without any annotated labels, by modeling the intra-region coherence and inter-region contrast information from self-supervised transformer features. +- SCoCo has two novel modules: a GBA module with a scene-level contrastive loss to find the saliency thresholds of different image contents, and a LAS module with an instance-level contrastive loss to detect salient instances. +- Extensive experiments show that SCoCo outperforms existing weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods. + +# 2. Related Works + +# 2.1. Salient Instance Detection (SID) + +SID aims to individualize visually prominent objects in the input image. Existing deep SID methods can be classified into two categories based on their learning labels: fully-supervised and weakly-supervised SID methods. + +Fully-supervised SID methods rely on costly pixel-wise annotated masks for training supervision. Li et al. [25] propose to apply MAP [80] on instance proposals, which are + +generated based on binary saliency and contour maps, to select and segment salient instances. Fan et al. [14] devise a single-stage salient instance detection method, with a RoIMasking layer that utilizes both local and surrounding contexts of each instance for detection. Pei et al. [45] propose a spectral clustering method to obtain salient instances by exploring pixel masks and subitizing information of salient objects, and later a transformer-based method [46] to detect salient instances in a single stage without NMS for post-processing. Liu et al. [37] propose an interleaved execution strategy to incorporate global context and object contour information jointly to detect salient instances. RGB-D information has also been explored in CalibNet et al. [47], with a dual-branch cross-modal calibration method. + +Weakly-supervised SID methods typically use bounding boxes and image-level labels, e.g., classes and subitizing as supervisions. Zhang et al. [80] propose a Maximum A Posteriori (MAP)-based subset optimization formulation to select a compact set of salient instances from the redundant box-level proposals, followed by CRF to help segment pixel-wise instances from box-level predictions. Tian et al. [61, 62] propose a triple-branch network to exploit salient boundaries, centroids, and regions jointly from class and subitizing labels for salient instance detection. + +Despite their success, existing SID methods require human efforts to prepare different types of labels, which are expensive. In contrast, we propose to model the saliency coherence and contrast information based on self-supervised transformer features, for USID. + +# 2.2. Salient Object Detection (SOD) + +SOD is an important task to scene understanding with a lot of methods proposed. Deep SOD methods can be divided into three groups w.r.t. the supervision signals: fully-supervised, weakly-supervised, and unsupervised. + +Fully-supervised SOD methods mainly incorporate four types of deep techniques to learn salient object represent- + +tations from pixel-level annotated masks. First, deep feature fusion [34, 39, 68, 73, 83] is used to aggregate multilevel context information including low-level stimulus and high-level semantics for detecting salient objects. Second, attention mechanisms [30, 35, 36, 58, 84, 85] are used to reweight multi-scale features and enhance context learning to help the model focus on the salient regions and suppress noise of the background regions. Third, boundary enhancement [9, 33, 52, 86] is often applied to improve the localization and segmentation results by penalizing the errors of boundary-surrounding pixels. Last, recurrent mechanisms [22, 29, 60, 66, 84] may also be reused to iterate an identical network to predict and refine the saliency regions in a coarse-to-fine manner. + +Weakly-supervised SOD methods have explored class labels, image captions, scribbles, bounding boxes, and points as supervisions. Wang et al. [67] propose the first weakly-supervised SOD method, which leverages image-level class labels to localize salient regions and then refines those predicted regions with CRF. Following [67], Li et al. [26] propose a pseudo saliency annotation updating scheme to refine the spatial coherence of predicted salient regions. Piao et al. [50] propose the multi-filter directive network to distill clean saliency maps from multiple class-based pseudo labels. Zeng et al. [78] devise a multi-source weak supervision framework to jointly utilize category labels, captions, and a set of unlabelled images for training. Based on scribble annotations, WSSA [82] and AGGM [77] are proposed to detect salient objects by exploring local consistency and structure information of objects. Liu et al. [40] propose to use bounding box supervisions to help address the over/under detection problems. Gao et al. [15] design an adaptive masked flood filling algorithm to detect salient objects from point-wise labels. + +Unsupervised SOD methods can be classified into four categories according to the features and technologies used. First, conventional methods are typically based on handcrafted features [2, 5, 38, 49, 76, 92], e.g., local and global contrast [10, 38, 49], boundary priors [27, 76, 92], and spatial frequency [2], for detecting salient objects. These methods may not handle complex scenes as their low-level features have limited representation capacities. Second, a series of methods [32, 44, 72, 79, 81, 87, 88] try to distill clean labels from pseudo labels, which are produced by some of the aforementioned conventional methods, for learning SOD models. Specifically, they use distillation strategies like iterative refinement [44], causal debiasing [32], and mining textures around boundaries [89] to reduce the noise coming from the pseudo labels. Third, instead of learning from noisy pseudo labels, Yan et al. [75] propose a domain adaptive SOD method that learns from synthetic but clean saliency data. Fourth, FOUND [57] and SelfMask [55] also detect salient objects based on self-supervised deep + +features. However, our SCoCo differs from them in both tasks and proposed techniques. First, they cannot predict instance-level saliency information, as shown in row 1 of Figure 1(b,c). Second, they only model the binary contrast between salient and non-salient regions; they are unable to suppress non-salient regions/objects with uncertain contrast degrees. For example, in row 2 of Figure 1(c,d), FOUND and SelfMask focus mainly on the most prominent woman, neglecting the umbrella, as the woman has much stronger semantics than the umbrella. In contrast, our SCoCo (Figure 1(d)) can correctly individualize salient instances by modeling intra-region saliency coherence and different levels of saliency contrast between regions. + +In summary, unlike the above SOD methods, which do not consider instance-level object saliency, our work aims to detect salient instances but without the need to use annotations for supervision. + +# 2.3. Unsupervised Object Detection (UOD) + +UOD aims to localize objects without using any human annotations. Early methods [11, 63-65] explore inter-image similarities (e.g., clustering, ranking) to discover objects for an image collection. Recently, several methods learn to detect objects based on self-supervised deep features. LOST [56] and TokenCut [71] propose to segment one object per image based on the pixel graph of self-supervised deep features. FreeSOLO [69] and CutLER [70] learn to detect multiple objects by self-retraining [69] and MaskCut [70]. + +While these methods tend to detect any objects (salient, non-salient), our method learns to identify salient instances. + +# 3. Methodology + +Learning an unsupervised salient instance detector is challenging, as there are no explicit labels/knowledge to teach the detector to identify salient regions at the instance level. In this work, we observe some interesting cues in the self-supervised transformer features - they exhibit saliency coherence within local regions but saliency contrast of diverse degrees between different regions. If a detector can model such correlations well, it may be able to identify salient regions and group them into separate salient instances. Here, we propose the SCoCo method to mine regional saliency coherence and contrast information for USID. + +Figure 3 shows the overview of our SCoCo. Given an input image, we first apply a self-supervised transformer (with frozen weights) to generate the corresponding deep features. We then feed these deep features separately to a novel global background adaptation (GBA) module (Section 3.1) and a novel locality-aware similarity (LAS) module (Section 3.2). The GBA module with its scene-level contrastive loss aims to find the suitable 'saliency thresholds' to distinguish salient regions from diverse background contents to produce a scene-level saliency map. Meanwhile, + +![](images/5e3e9751d07c12cb514c909341476bae32d44d3a3cfff2392b65f8bed21ee6af.jpg) +Figure 3. The overview of our unsupervised pipeline. Given an image, we first obtain its features $F$ from a self-supervised transformer (whose weights are frozen). We then use $F$ to produce pseudo scene- and instance-level supervisions $\mathrm{P_{S+}}$ and $\mathrm{P_{I+}^n}$ ( $1 \leq n \leq N$ ) by computing feature similarities, and feed $F$ to our novel GBA and LAS modules. The GBA module learns with a scene-level contrastive loss $\mathcal{L}_S$ to predict a saliency map $\mathrm{P_s}$ , while the LAS module learns with an instance-level contrastive loss $\mathcal{L}_{\mathcal{I}}$ to model saliency coherence and contrast features. The output features of both GBA and LAS modules are combined in the salient instance detection head to produce the final salient instance map $\mathrm{P_I}$ . + +the LAS module works with an instance-level contrastive loss to learn intra-region saliency coherence and interregion saliency contrast features. These features are then fed into a salient instance detection head to produce salient instances, with the guidance of the scene-level saliency map produced by the GBA module. + +# 3.1. Global Background Adaptation (GBA) Module + +The goal of the GBA module is to adaptively dig saliency information from diverse background contents. As Figure 2(b-d) shows, it is challenging to find a fixed saliency threshold to select salient regions from different scenes with diverse complexities. To address this issue, we propose the GBA module, which has learnable saliency threshold functions working in a bifurcation-and-combination manner, to detect the fine-grained adaptive saliency thresholds. Specifically, as shown in Figure 4, the GBA module is designed to have two bifurcated branches at the first, with each of them equipped with its own learnable saliency threshold function, to handle the potential intra-scene saliency-background ambiguities separately. The two bifurcated branches are then combined with another saliency threshold-based branch to produce the scene-level saliency map. + +Our GBA module differs from previous unsupervised salient object detection methods [55, 57] in the way of detecting saliency thresholds. Existing USOD methods typically rely on stacking convolutional operations to detect salient objects. However, such operations (even with dynamic convolution kernels [24, 59, 90]) are content-agnostic as convolution filters are shared across all the feature pixels. As a result, these methods often fail to detect the salient regions buried in diverse background contents. + +Network Structure. Given features $F \in \mathbb{R}^{H \times W \times C}$ , we first process it with two parallel branches, each consisting of a group of $1 \times 1$ and $3 \times 3$ conv operators, and a learnable saliency threshold function $f(x)$ . We define $f(x)$ as: + +$$ +f (x) = \left(p _ {1} - p _ {2}\right) x \cdot \sigma \left(\theta \left(p _ {1} - p _ {2}\right) x\right) + p _ {2} x, \tag {1} +$$ + +where $f(x)$ is a smooth version of Swish rectifying function [42, 53], $p_1$ and $p_2$ are learnable parameters that can jointly + +determine the upper and lower bounds of the function, $\sigma$ is the Sigmoid function, and the learnable parameter $\theta$ controls the "selection degree" of salient objects with respect to the background. Each $\theta$ is learned using two fully connected (FC) layers upon a global image vector generated by the global average pooling (GAP) layer. We then use another branch to combine the former two branches, with another set of convolutional operations for feature transformation and learnable saliency threshold function to unify the separately modeled saliency thresholds, to produce the scene-level saliency map. + +Scene-level Contrastive Loss. To boost the saliency learning of the GBA module, we introduce a scene-level contrastive loss $\mathcal{L}_S$ . Since the image background is expected to have low saliency degrees, we first derive the scene-level pseudo saliency supervision $\mathrm{P}_{\mathrm{S}+}$ through inverting the identified background regions with low saliency degrees. However, we note that such pseudo-saliency maps may contain a significant number of noisy pixels, particularly those false positive background pixels with medium saliency degrees, that may hinder learning efficacy. To address this problem, we leverage the pixels with high saliency degrees in our instance-level saliency predictions $\mathrm{P_I}$ (introduced in Section 3.2) to help classify those uncertain pixels into salient and non-salient ones. To this end, we take $\mathrm{P_I}$ as positive samples and utilize scene-level saliency in different images as negative samples. $\mathcal{L}_S$ can then be formulated as: + +$$ +\mathcal {L} _ {\mathcal {S}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {e ^ {\left(\mathrm {d} \left(\mathrm {P} _ {\mathrm {s}} ^ {i} , \mathrm {P} _ {\mathrm {s} +} ^ {i}\right) + \mathrm {d} \left(\mathrm {P} _ {\mathrm {s}} ^ {i} , \mathrm {P} _ {\mathrm {I}} ^ {i}\right)\right) / \tau}}{e ^ {\left(\mathrm {d} \left(\mathrm {P} _ {\mathrm {s}} ^ {i} , \mathrm {P} _ {\mathrm {s} +} ^ {i}\right) + \mathrm {d} \left(\mathrm {P} _ {\mathrm {s}} ^ {i} , \mathrm {P} _ {\mathrm {I}} ^ {i}\right)\right) / \tau} + \sum_ {k = 1 , k \neq i} ^ {N} e ^ {\left. \mathrm {d} \left(\mathrm {P} _ {\mathrm {s}} ^ {i} , \mathrm {P} _ {\mathrm {s} +} ^ {k}\right) / \tau \right.}}, \tag {2} +$$ + +where $\mathrm{N}$ is the batch size, $\mathrm{P}_{\mathrm{S}}$ is the scene-level saliency prediction, $\mathrm{d}$ measures the mean square error (MSE), and $\tau$ is a temperature parameter. + +# 3.2. Locality-aware Similarity (LAS) Module + +The LAS module aims to capture regional saliency coherence information from the self-supervised deep features. This information can then be used to help learn different levels of saliency contrast using an instance-level contrastive loss. We note that the latest saliency instance de + +![](images/2949c6faab73fb952a69682528e9af672190b3016d27a9d458ed8393d7c115c2.jpg) +Figure 4. The bifurcation-to-combination learning structure of the global background adaptation (GBA) module. The pink lines indicate the learning flows of learnable saliency threshold functions for extracting salient objects from different background contents. + +tection methods [37, 46] typically leverage the object detector head [6, 17] (a combination of convolutions and self-attention blocks) to classify and group salient instance pixels. However, although convolutional filters are good at detecting repeated patterns, they may fail in discerning pixel similarities. The self-attention mechanisms tend to model the interactions of pixel tokens but do not compare region-level saliency contrasts. + +On the other hand, our LAS module first models intra-region feature coherence by computing channel-wise correlations with respect to regions surrounding every pixel location, and then feeds regional features to an instance-level contrastive loss to learn regional saliency contrasts for discriminating salient instances. The overview of LAS module is depicted in Figure 5. + +Network Structure. Given features $F \in \mathbb{R}^{H \times W \times C}$ , we downsample it to $F' \in \mathbb{R}^{H \times W \times C'}$ with a linear layer. We then compute cosine similarity between a positional feature $x \in [1, H] \times [1, W]$ and the surrounding region of $x$ in $F'$ , and collect them into a similarity feature tensor $S \in \mathbb{R}^{H \times W \times M \times M \times C'}$ . The formulation of $S$ is: + +$$ +S (x, m) = \frac {F _ {(x)} ^ {\prime} \cdot F _ {(x + m)} ^ {\prime}}{\| F _ {(x)} ^ {\prime} \| \cdot \| F _ {(x + m)} ^ {\prime} \|}, \tag {3} +$$ + +where $m \in [-M + 1, M - 1] \times [-M + 1, M - 1]$ , and $M \times M$ is the region size. Eq. 3 aims to encode the local saliency coherence of each pixel $x$ into $S$ . We further apply convolutional filters on $S$ to compact regional features. The spatial size of $S$ is progressively aggregated from $M \times M$ to $1 \times 1$ with cascaded $3 \times 3$ convolutional filters (without padding) to obtain $S'$ with the same spatial size as the input feature $F'$ . After each convolutional filter, we also + +insert the batch normalization and rectifying linear unit layers. Finally, we combine $F'$ and $S'$ to produce the final features $\hat{F}'$ with saliency coherence information, which further facilitates the saliency contrast modeling for differentiating salient instances. + +Instance-level Contrastive Loss. We propose the instance-level contrastive loss $\mathcal{L}_{\mathcal{I}}$ to derive the saliency contrast information from $\hat{F}^{\prime}$ for salient instance discrimination, as the binary cross entropy (BCE) loss adopted by existing unsupervised SOD methods [56, 69, 70] cannot distinguish different salient instances. + +Given the self-supervised transformer features, we extract the key embeddings $F_{key}$ , and feed the self-affinity matrix of $F_{key}$ to Normalized Cut [54, 71] to find highly coherent salient pixels as pseudo salient instance supervision $\mathrm{P}_{\mathrm{I}+}$ in an iterative manner. We find one salient instance $\mathbf{P}_{\mathrm{I}+}^{k}$ in each iteration $k \in [1, K]$ , and the iteration ends if the $\mathbf{P}_{\mathrm{I}+}^{K+1}$ is located in the low-saliency background. Note that $\mathrm{P}_{I}$ is used in $\mathcal{L}_S$ (Eq. 2) for penalizing pixels with uncertain saliency degrees. Meanwhile, we can measure the spatial discrepancy between predicted instances $\mathbf{P}_{\mathrm{I}}^{L}$ and $\mathbf{P}_{\mathrm{I}+}^{K}$ to formulate $\mathcal{L}_{\mathcal{I}}$ , where $L$ in the number of instances. To reduce the influence of noise that may be contained in pseudo labels, we propose to sample point vectors from $\hat{F}'$ to compute $\mathcal{L}_{\mathcal{I}}$ , where $\hat{F}', \mathbf{P}_{\mathrm{I}}^{L}$ , and $\mathbf{P}_{\mathrm{I}+}^{K}$ are spatially aligned for the sampling. By sampling a small subset of point vectors, we make the model less likely to be influenced by extreme noise that may be present in the full pseudo label set. Specifically, given $\hat{F}'$ , we randomly sample $T(t \in [1, T])$ triplets of point vectors $\hat{F}_{\mathrm{I}}^{l(t)}, \hat{F}_{\mathrm{I}+}^{l(t)}$ , and $\hat{F}_{\mathrm{I}-}^{l(t)}$ from $\hat{F}'$ with respect to each predicted instance $\mathbf{P}_{\mathrm{I}}^{l}$ , its positive supervision $\mathbf{P}_{\mathrm{I}+}^{k}$ (IoU $(\mathbf{P}_{\mathrm{I}}^{l}, \mathbf{P}_{\mathrm{I}+}^{k}) \geq 0.7$ ), and negative supervision $\mathbf{P}_{\mathrm{I}-}^{f}$ ( $f \neq k, f \in [1, K]$ ), respectively. $\mathcal{L}_{\mathcal{I}}$ can then be formulated as: + +$$ +\mathcal {L} _ {\mathcal {I}} = - \frac {1}{L \times T ^ {2}} \sum_ {l = 1} ^ {L} \sum_ {t = 1} ^ {T} \sum_ {t = 1} ^ {T} \log \left(g \left(\hat {F} ^ {\prime}, l, t\right)\right), \tag {4} +$$ + +where $g(\cdot)$ is a contrastive estimation function, which is formulated as: + +$$ +g \left(\hat {F} ^ {\prime}, l, t\right) = \frac {e ^ {s i m \left\langle \hat {F} _ {\mathrm {I}} ^ {l (t)} , \hat {F} _ {\mathrm {I} +} ^ {l (t)} \right\rangle / \tau}}{e ^ {s i m \left\langle \hat {F} _ {\mathrm {I}} ^ {l (t)} , \hat {F} _ {\mathrm {I} +} ^ {l (t)} \right\rangle / \tau} + \sum_ {t = 1} ^ {T} e ^ {s i m \left\langle \hat {F} _ {\mathrm {I}} ^ {l (t)} , \hat {F} _ {\mathrm {I} -} ^ {l (t)} \right\rangle / \tau}}, \tag {5} +$$ + +where $sim\langle \cdot \rangle$ is the cosine similarity function to compute the distance between two point vectors $\hat{F}_{\mathrm{I}}^{l(t)}$ and $\hat{F}_{\mathrm{I}+}^{\prime l(t)}$ , and $l(t)$ represents the $t_{th}$ point vector of the $l_{th}$ instance. Each vector is extracted from $\hat{F}^{\prime}$ according to its spatial location. Our instance-level contrastive loss (Eq. 4) helps the network learn the amplified saliency contrast of different instances while keeping the saliency coherence within each instance by maintaining their self-similarities. + +![](images/b8f9f112f8f2f426b9a78a337766e43d6bdc56072be9c3bae94672b76204e51f.jpg) +Figure 5. The structure of locality-aware similarity (LAS) module. + +# 4. Experiments + +# 4.1. Experimental Settings + +Implementation Details. We use DINO [8] to initialize the self-supervised transformer (i.e., ViT [13]) of our SCoCo network. Note that neither DINO nor our SCoCo uses any labeled data (e.g., class labels) for supervision. The input to both GBA and LAS modules, i.e., $F$ , is the key features of multi-head attention layers in the last feature block of DINO-based ViT. In the GBA module, we use the mean attention value of $F$ as the low saliency degree per image to find its scene-level saliency region $\mathrm{P_{S+}}$ . In the LAS module, we sample $T = 32$ triplets of point vectors to formulate the loss $\mathcal{L}_{\mathcal{I}}$ . $\tau$ in both $\mathcal{L}_S$ and $\mathcal{L}_{\mathcal{I}}$ is set to 0.1. We adopt the head of Mask-RCNN [17] as our salient instance detection head. During training, each input image has a resolution of $1024 \times 1024$ , and is augmented using largescale jittering [16]. The batch size $N$ is set to 4 and we use AdamW $(\beta_1 = 0.9, \beta_2 = 0.999)$ with step-wise learning rate decay for model optimization. We first train our SCoCo network for 10 epochs using the pseudo supervision derived from DINO. We then update the pseudo labels using SCoCo. Last, we train SCoCo with the updated pseudo labels for another epoch. SCoCo is trained on two 3090ti GPU cards. + +Datasets and Evaluation Metrics. We conduct our experiments on three SID datasets, including ILSO-1K [25], ILSO-2K [25], and SIS [46]. The (training/validation/test) sets in the ILSO-1K [25], ILSO-2K [25], and SIS [46] contain (500/200/300), (1000/400/600), and (7030/2100/1170) images, respectively. For each dataset, we only use the unlabeled images of the training set to train our SCoCo network, and we evaluate the performance on the corresponding test set. We adopt the mean Average Precision (mAP) metric for evaluation, following previous SID approaches. + +# 4.2. Baselines + +Since we propose the first unsupervised method to detect salient instances, to evaluate our method comprehensively, we design several baselines from three aspects, inspired by related areas, as follows: + +- First, we note that existing unsupervised object detection (OD) methods, i.e., FreeSOLO [69], and CutLER [70], detect objects in a saliency-agnostic way. We retrain their + +models with our pseudo salient instance labels for comparison. Their models are initialized with the model parameters trained on their pseudo labels extracted from large-scale datasets, e.g., COCO [31] and ImageNet [23]. + +- Second, VitDet [28] is the best-performing method among fully-supervised object detection methods that are based on the ViT [13] backbone. Since we also use ViT as our backbone, we retrain VitDet with our salient instance pseudo labels for comparison. +- Third, we adapt existing USOD methods, i.e., SelfMask [55] and FOUND [57], to detect salient instances. To derive instance-level saliency predictions, we compute IoU between the region-level saliency predictions from USOD methods and the instance predictions from CutlER [70] (a state-of-the-art unsupervised object detection method). + +# 4.3. Main Results + +We compare our method to 19 state-of-the-art SID methods, including seven fully-supervised methods: MSRNet-V1 [25], MSRNet-V2 [25], S4Net [14], SCG [37], MDNN [45], RDPNet [74], and OQTR [46]; six weakly-supervised methods: MAP [80], NLDF [41], DeepMask [21], PRM+D [12], IRN [3], and WSID-Net [62]; and six unsupervised baselines: FreeSOLO [69], CutLER [70], VitDet [28] initialized with the classification backbone, VitDet[28] initialized with the DINO-based backbone, SelfMask [55], and FOUND [57]. + +Quantitative Comparisons. Table 1 shows the quantitative results. First, we can see that our method outperforms all carefully designed unsupervised baselines (the bottom group). This shows that simply retraining unsupervised methods with our pseudo labels and/or filtering out nonsalient objects predicted by unsupervised object detectors is not effective in deriving accurate salient instance predictions. Second, we can also see that our method surpasses all existing weakly-supervised methods (middle group), while achieving comparable performances to the fully-supervised methods (first group). Existing weak supervision signals (i.e., object-level pixel masks, bounding boxes, classes, and subitizing labels) cannot provide instance-level supervision, which limits the performance of existing weakly-supervised methods. In contrast, even without supervision, SCoCo can learn to derive instance-level saliency information by modeling saliency coherence and contrast information based on + +Table 1. Quantitative comparison with 7 fully-supervised methods, 6 weakly-supervised methods, and 6 unsupervised baselines mentioned in Section 4.2. Column 1 shows the supervision and annotation types of the methods listed in Column 2. Column 3 shows their original tasks (SID: salient instance detection, SOD: salient object detection, OD: object detection, SIS: semantic instance segmentation). Column 4 shows backbone initialization choices, where Cls denotes that their backbones were pre-trained on class labels of ImageNet, and DINO is a self-supervised backbone pretraining method. - denotes missing results as codes/results are not publicly available. For each supervision category, best results are marked in bold, while second-best results are underlined. + +
Supervision type & Annotation typeMethodOriginal taskInit.Datasets and Metrics (mAP IoU ↑)
ILSO-1K [25]ILSO-2K [25]SIS [46]
mAP50mAP70mAP50mAP70mAP50mAP70
Full SupervisionMSRNet-V1 [25]SIDCls65.3%52.1%----
MSRNet-V2 [25]SIDCls85.1%74.7%78.3%66.5%--
[Images are labeled with instance-level pixel masks.]S4Net [14]SIDCls82.2%59.6%73.1%52.9%86.7%63.6%
SCG [37]SIDCls88.8%78.5%79.8%68.9%83.4%71.2%
MDNN [45]SIDCls84.9%67.8%76.1%63.6%84.6%67.4%
RDPNet [74]SIDCls88.9%73.8%80.7%67.2%82.0%69.4%
OQTR [46]SIDCls89.2%81.0%81.4%72.0%88.1%81.7%
Weak SupervisionMAPΔ [80]SIDCls56.6%24.8%51.6%30.3%58.4%22.7%
[Images are labeled with region-level pixel masks*, bounding boxesΔ, classes°, or subitizing⊗]NLDF* [41]SODCls45.5%24.5%43.8%25.2%46.4%24.2%
DeepMaskΔ [51]ODCls37.1%20.5%35.4%18.4%36.4%21.3%
PRM+D⊗[12]SISCls49.6%31.2%43.7%29.8%52.9%34.2%
IRN° [3]SISCls57.1%37.4%50.2%38.4%60.1%40.8%
WSID-Net⊗[62]SIDCls68.3%51.7%59.4%44.6%68.4%47.1%
No SupervisionFreeSOLOΦ [69]ODCls56.8%43.7%50.7%35.8%55.4%43.1%
FreeSOLOΦ [70]ODDINO53.2%40.2%46.9%32.6%51.8%40.3%
[Million-levelhundred thousand-levelΦ, or thousand-levelunlabeled images are used for training.]CutLERΦ [70]ODCls68.6%54.8%63.2%48.3%66.9%54.4%
CutLERΦ [70]ODDINO65.4%52.1%59.8%44.9%62.4%51.5%
VitDet [28]SISCls55.6%43.6%51.9%35.8%54.4%44.1%
VitDet [28]SISDINO55.7%43.4%52.4%37.0%56.5%44.8%
SelfMask [55]SODDINO65.5%51.1%58.9%44.0%64.8%52.1%
FOUND [57]SODDINO66.2%51.3%61.4%45.9%67.1%53.8%
Ours [■]SIDDINO70.7%58.3%64.8%50.4%71.4%55.7%
+ +just the self-supervised Transformer features. + +Qualitative Comparison. Figure 6 compares the visual results of our method and 10 top performing methods from the three groups of methods in Table 1. We can see that the unsupervised baselines (columns 8 to 11) often fail to detect the complete salient instances and produce messy predictions. For the weakly-supervised methods (columns 5 to 7), MAP, which is trained with bounding box labels, may fail to segment complete objects even with the CRF post-processing. IRN and WSID-Net, which are trained with different class labels, may struggle to determine the boundaries between salient instances of the same class, e.g., the sausages in row 3. Fully-supervised methods (columns 2 to 4) tend to be trained to predict as many salient instances as possible. However, they may predict multiple instances from a salient instance of self-contrast (e.g., rows 1, 2, 5 and 6) and sometimes miss salient instances of solid colors (e.g., rows 4 and 6). In comparison, our method can successfully detect salient instances and delineate their boundaries, with the help of learning intra-instance saliency coherence and cross-instance saliency contrast. + +# 4.4. Ablation Study + +Ablation Study of SCoCo. We first study the effectiveness of the proposed GBA and LAS modules, and the loss functions $\mathcal{L}_S$ and $\mathcal{L}_{\mathcal{I}}$ used in them. The $\mathrm{mAP}_{50}$ results in Table 2 show that solely applying either scene-level (GBA + + +Table 2. Ablation study of the GBA and LAS modules, and the corresponding losses, $\mathcal{L}_{\mathcal{S}}$ and $\mathcal{L}_{\mathcal{I}}$ , on the SIS [46] test set using mAP metrics. Best performances are marked in bold. + +
MethodGBALSLASL2mAP50mAP70
SCoCo✓64.7%52.6%
✓✓67.0%53.5%
✓68.4%54.1%
67.4%53.1%
✓✓69.0%54.4%
✓✓✓✓71.4%55.7%
+ +$\mathcal{L}_{\mathcal{S}}$ or instance-level $(\mathrm{LAS} + \mathcal{L}_{\mathcal{I}})$ saliency contrast learning may produce a performance decrease of $2 - 3\%$ , and the performance drops by around $4\%$ if we use the GBA or LAS module only. Combining all of them produces the best performance, which verifies the necessities of our designs. + +Visualization of the GBA Module. Figure 7 visualizes the scene-level salinocy map learned by the GBA module and the corresponding SID result. We can see that without the GBA module, the model tends to be distracted by high-contrast background regions, such as the shadow, and identifies them as salient regions. In contrast, using the GBA module can successfully suppress such noise by learning proper saliency thresholds. + +Visualization of the LAS Module. In Figure 8, we investigate the influence of our LAS module by visualizing feature correlations to points (purplish dots) in individual instances. We can see that the two pigs have non-uniform white and + +![](images/e7b7ab1386600035a140e92d3d389693b8610706e0536dc591e6629882c4754a.jpg) +Figure 6. Qualitative comparison with existing state-of-the-art methods and baselines. Each color represents a unique salient instance. + +![](images/65b184711d9a46e22a67642fc2df6ac78a774c90cc5f98398037e1837df3db2a.jpg) +Figure 7. Visualization of the GBA module. + +![](images/7e11b02fad18bcf18a0c48c52c704079f9ad09b9562605f1dc00448619de10c0.jpg) +(a) Input Image + +![](images/cc9f992aa265ac132147baefa8cc3b10cd1bb92e62f9c258a35ffa2c75bb3370.jpg) +(b) Pixel-based saliency contrast w/o LAS + +![](images/18e1ef80f001548192bba9c96d2b4a900d623ec60df669fa47ab00efaabb631c.jpg) +(c) Pixel-based saliency contrast w/ LAS +Figure 8. Visualization of the LAS module. + +black colors on their bodies, causing our model to learn incomplete saliency coherence for each pig and produce overdetection results without the LAS module (b). With the LAS module, our model can separate the two pigs successfully by learning saliency contrast between instances while maintaining intra-instance saliency coherence. + +Backbone Initialization. Finally, we study the alternatives of self-supervised methods for initializing our ViT backbone. As shown in Table 3, in comparison to MoCo-v3 [18] and SWaV [7], we choose DINO [8] to initialize our backbone as it can provide better performances from its learned feature similarities. In addition, we select ViT-S as our backbone since ViT-B can only introduce minor improvements with larger parameters. + +# 5. Conclusion + +In this paper, we have studied a new task, unsupervised salient instance detection. We have observed that self-supervised transformer features exhibit intra-region saliency coherence and different levels of inter-region + +Table 3. Quantitative analysis of backbone initialization choices. Best performances are marked in bold. + +
MethodInit.Arch.\( \mathbf{{mAP}}_{50} \)\( {\mathbf{{mAP}}}_{70} \)
SCoCoMoCo-v3 [18]ViT-S70.5%54.9%
SWaV [7]ViT-S70.2%54.3%
DINO [8]ViT-S71.4%55.7%
DINO [8]ViT-B71.9%55.6%
+ +saliency contrast, which can help differentiate salient instances. With this observation, we have proposed SCoCo, with two novel designs, to model such regional saliency coherence and contrast information. First, we propose a global background adaptation (GBA) module and a scene-level contrastive loss to find saliency thresholds for salient objects w.r.t. diverse background contents. Second, we propose a locality-aware similarity (LAS) module and an instance-level contrastive loss to model in-region saliency coherence and cross-region saliency contrasts for detecting salient instances. Extensive experiments have verified the effectiveness of our method against SOTA methods. + +Our method does have limitations. We note that SCoCo may not be able to detect small salient instances as it learns from high-level, low-resolution features. As shown in Figure 9, the person standing on the speedboat is too small to be detected by SCoCo. As a future work, it would be interesting to study multi-granularity salient feature similarity and contrast modeling to detect multi-scale salient instances. + +![](images/a98f5c76c139e8c0f39129323bba345224258fb95fe6535e3db6143fd57022ca.jpg) +Figure 9. SCoCo may fail to detect small salient instances. + +# References + +[1] Kfir Aberman, Junfeng He, Yossi Gandelsman, Inbar Mosseri, David E Jacobs, Kai Kohlhoff, Yael Pritch, and Michael Rubinstein. Deep saliency prior for reducing visual distraction. In CVPR, 2022. 1 +[2] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 3 +[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 6, 7, 8 +[4] Codruta Orniana Ancuti, Cosmin Ancuti, and Phillipe Bekaert. Enhancing by saliency-guided decolorization. In CVPR, 2011. 1 +[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVM, 2019. 3 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 2020. 8 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 6, 8 +[9] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE TIP, 2020. 3 +[10] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 2014. 3 +[11] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In CVPR, 2015. 3 +[12] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In CVPR, 2019. 6, 7 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6 +[14] Ruochen Fan, Ming-Ming Cheng, Qibin Hou, Tai-Jiang Mu, Jingdong Wang, and Shi-Min Hu. S4net: Single stage salient-instance segmentation. In CVPR, 2019. 2, 6, 7, 8 +[15] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI, 2022. 3 +[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 6 + +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 5, 6 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 8 +[19] Brian Hu, Bhavan Vasu, and Anthony Hoogs. X-mir: Explainable medical image retrieval. In WACV, 2022. 1 +[20] Lai Jiang, Yifei Li, Shengxi Li, Mai Xu, Se Lei, Yichen Guo, and Bo Huang. Does text attract attention on e-commerce images: A novel saliency prediction dataset and method. In CVPR, 2022. 1 +[21] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. 1, 6 +[22] Yun Yi Ke and Takahiro Tsubono. Recursive contour-saliency blending network for accurate salient object detection. In WACV, 2022. 3 +[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 6 +[24] Duo Li, Jie Hu, Changhu Wang, Xiangtai Li, Qi She, Lei Zhu, Tong Zhang, and Qifeng Chen. Involution: Inverting the inheritance of convolution for visual recognition. In CVPR, 2021. 4 +[25] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, 2017. 1, 2, 6, 7 +[26] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In AAAI, 2018. 3 +[27] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In ICCV, 2013. 3 +[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 6, 7 +[29] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In CVPR, 2015. 3 +[30] Zijian Liang, Pengjie Wang, Ke Xu, Pingping Zhang, and Rynson WH Lau. Weakly-supervised salient object detection on light fields. IEEE TIP, 2022. 3 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[32] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In AAAI, 2022. 1, 3 +[33] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In CVPR, 2019. 3 +[34] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016. 3 +[35] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, 2018. 3 + +[36] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, 2021. 3 +[37] Nian Liu, Wangbo Zhao, Ling Shao, and Junwei Han. Scg: Saliency and contour guided salient instance segmentation. IEEE TIP, 2021. 2, 5, 6, 7 +[38] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE TPAMI, 2010. 3 +[39] Yudong Liu, Yongtao Wang, Siwei Wang, TingTing Liang, Qijie Zhao, Zhi Tang, and Haibin Ling. Cbnet: A novel composite backbone network architecture for object detection. In AAAI, 2020. 3 +[40] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE TIP, 2021. 3 +[41] Zhiming Luo, Akshaya Mishra, Andrew Achkar, Justin Eichel, Shaozi Li, and Pierre-Marc Jodoin. Non-local deep features for salient object detection. In CVPR, 2017. 6, 7 +[42] Ningning Ma, Xiangyu Zhang, Ming Liu, and Jian Sun. Activate or not: Learning customized activation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8042, 2021. 4 +[43] S Mahdi H Miangoleh, Zoya Bylinskii, Eric Kee, Eli Shechtman, and Yaqiz Aksoy. Realistic saliency guided image enhancement. In CVPR, 2023. 1 +[44] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. NeurIPS, 2019. 1, 3 +[45] Jialun Pei, He Tang, Chao Liu, and Chuanbo Chen. Salient instance segmentation via subitizing and clustering. Neurocomputing, 2020. 2, 6, 7 +[46] Jialun Pei, Tianyang Cheng, He Tang, and Chuanbo Chen. Transformer-based efficient salient instance segmentation networks with orientative query. IEEE TMM, 2022. 2, 5, 6, 7, 8 +[47] Jialun Pei, Tao Jiang, He Tang, Nian Liu, Yueming Jin, Deng-Ping Fan, and Pheng-Ann Heng. Calibnet: Dual-branch cross-modal calibration for rgb-d salient instance segmentation. arXiv preprint arXiv:2307.08098, 2023. 2 +[48] Guohao Peng, Jun Zhang, Heshan Li, and Danwei Wang. Attentional pyramid pooling of salient visual residuals for place recognition. In ICCV, 2021. 1 +[49] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, 2012. 3 +[50] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In ICCV, 2021. 3 +[51] Pedro OO Pinheiro, Ronan Collobert, and Piotr Dólar. Learning to segment object candidates. In NeurIPS, 2015. 7 +[52] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, 2019. 3 +[53] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 4 + +[54] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5 +[55] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPR Workshop, 2022. 1, 3, 4, 6, 7, 8 +[56] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3, 5 +[57] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecý, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In CVPR, 2023. 1, 3, 4, 6, 7, 8 +[58] Avishek Siris, Jianbo Jiao, Gary KL Tam, Xianghua Xie, and Rynson WH Lau. Scene context-aware salient object detection. In ICCV, 2021. 3 +[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, 2019. 4 +[60] Chang Tang, Xinzhong Zhu, Xinwang Liu, and Pichao Wang. Salient object detection via recurrently aggregating spatial attention weighted cross-level deep features. In ICME, 2019. 3 +[61] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Weakly-supervised salient instance detection. In BMVC, 2020. 2 +[62] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Learning to detect instance-level salient objects using complementary image labels. IJCV, 2022. 1, 2, 6, 7, 8 +[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In CVPR, 2019. 3 +[64] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In ECCV, 2020. +[65] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. NeurIPS, 2021. 3 +[66] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang, and Xiang Ruan. Saliency detection with recurrent fully convolutional networks. In ECCV, 2016. 3 +[67] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 3 +[68] Xiang Wang, Huimin Ma, Xiaozhi Chen, and Shaodi You. Edge preserving and multi-scale contextual neural network for salient object detection. TIP, 2017. 3 +[69] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In CVPR, 2022. 3, 5, 6, 7, 8 +[70] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 3, 5, 6, 7, 8 + +[71] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufreydaz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv:2209.00383, 2022. 3, 5 +[72] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In CVPR, 2022. 1, 3 +[73] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, 2020. 3 +[74] Yu-Huan Wu, Yun Liu, Le Zhang, Wang Gao, and Ming-Ming Cheng. Regularized densely-connected pyramid network for salient instance segmentation. IEEE TIP, 2021. 1, 6, 7, 8 +[75] Pengxiang Yan, Ziyi Wu, Mengmeng Liu, Kun Zeng, Liang Lin, and Guanbin Li. Unsupervised domain adaptive salient object detection through uncertainty-aware pseudolabel learning. In AAAI, 2022. 3 +[76] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 3 +[77] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In AAAI, 2021. 3 +[78] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In CVPR, 2019. 3 +[79] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In ICCV, 2017. 1, 3 +[80] Jianming Zhang, Stan Sclaroff, Zhe Lin, Xiaohui Shen, Brian Price, and Radomir Mech. Unconstrained salient object detection via proposal subset optimization. In CVPR, 2016. 2, 6, 7, 8 +[81] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In CVPR, 2018. 1, 3 +[82] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In CVPR, 2020. 3 +[83] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, 2018. 3 +[84] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, 2018. 3 +[85] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, 2019. 3 +[86] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In CVPR, 2020. 3 +[87] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE TCSVT, 2022. 1, 3 + +[88] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 1, 3 +[89] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 3 +[90] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In CVPR, 2021. 4 +[91] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In ICCV, 2021. 1 +[92] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In CVPR, 2014. 3 \ No newline at end of file diff --git a/2024/Unsupervised Salient Instance Detection/images.zip b/2024/Unsupervised Salient Instance Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0f20dc589e80e4539db53016e3cd826cea6a85c --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3dc29fa1338b96f7b51c5347e6af3451c57786157cb10a6914283c443946ca +size 545176 diff --git a/2024/Unsupervised Salient Instance Detection/layout.json b/2024/Unsupervised Salient Instance Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39c6e30951657f8391706b2fd4267ba60af6c0f9 --- /dev/null +++ b/2024/Unsupervised Salient Instance Detection/layout.json @@ -0,0 +1,11577 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 172, + 103, + 422, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 103, + 422, + 120 + ], + "spans": [ + { + "bbox": [ + 172, + 103, + 422, + 120 + ], + "type": "text", + "content": "Unsupervised Salient Instance Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "spans": [ + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "text", + "content": "Xin Tian" + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "text", + "content": " Ke Xu" + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "text", + "content": " Rynson Lau" + }, + { + "bbox": [ + 204, + 142, + 389, + 157 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "spans": [ + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "type": "text", + "content": "Huawei Technologies " + }, + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 182, + 159, + 412, + 171 + ], + "type": "text", + "content": "City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 174, + 457, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 174, + 457, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 174, + 457, + 185 + ], + "type": "text", + "content": "{xin.tian.831, kkangwing}@gmail.com, Rynson.Lau@cityu.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 187, + 349, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 187, + 349, + 198 + ], + "spans": [ + { + "bbox": [ + 243, + 187, + 349, + 198 + ], + "type": "text", + "content": "† joint corresponding authors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 218, + 191, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 218, + 191, + 231 + ], + "spans": [ + { + "bbox": [ + 143, + 218, + 191, + 231 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 245, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 290, + 533 + ], + "type": "text", + "content": "The significant amount of manual efforts in annotating pixel-level labels has triggered the advancement of unsupervised saliency learning. However, without supervision signals, state-of-the-art methods can only infer region-level saliency. In this paper, we propose to explore the unsupervised salient instance detection (USID) problem, for a more fine-grained visual understanding. Our key observation is that self-supervised transformer features may exhibit local similarities as well as different levels of contrast to other regions, which provide informative cues to identify salient instances. Hence, we propose SCoCo, a novel network that models saliency coherence and contrast for USID. SCoCo includes two novel modules: (1) a global background adaptation (GBA) module with a scene-level contrastive loss to extract salient regions from the scene by searching the adaptive \"saliency threshold\" in the self-supervised transformer features, and (2) a locality-aware similarity (LAS) module with an instance-level contrastive loss to group salient regions into instances by modeling the in-region saliency coherence and cross-region saliency contrasts. Extensive experiments show that SCoCo outperforms state-of-the-art weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Salient Instance Detection (SID) aims to detect instance-level salient objects. It can provide a fine-grained visual understanding of the scene, and is therefore able to facilitate a wide range of applications, e.g., image captioning [21], scene text recognition [20], medical image analysis [19], and others [1, 4, 43, 48, 91]. However, existing SID methods either require pixel-level labels [25, 74] or image-level labels [62] as supervisions, which are expensive to prepare. Despite the advancements in unsupervised saliency detection [32, 44, 57, 72, 79, 81, 87, 88], none of the existing methods have demonstrated the capability to learn instance-level saliency, which essentially requires fine-grained dif" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 217, + 356, + 311 + ], + "blocks": [ + { + "bbox": [ + 312, + 217, + 356, + 311 + ], + "lines": [ + { + "bbox": [ + 312, + 217, + 356, + 311 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 356, + 311 + ], + "type": "image", + "image_path": "749713cb83f1b6bb3bb5394ea778f4943b7dc4ec04775556f55e4b1b11e7e66d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 312, + 342, + 320 + ], + "lines": [ + { + "bbox": [ + 314, + 312, + 342, + 320 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 342, + 320 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 357, + 217, + 402, + 311 + ], + "blocks": [ + { + "bbox": [ + 357, + 217, + 402, + 311 + ], + "lines": [ + { + "bbox": [ + 357, + 217, + 402, + 311 + ], + "spans": [ + { + "bbox": [ + 357, + 217, + 402, + 311 + ], + "type": "image", + "image_path": "480dcb5c837848e086cd1f085018e00232c7fdbd43e4e4e4a54083a922ee24f4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 312, + 400, + 319 + ], + "lines": [ + { + "bbox": [ + 351, + 312, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 351, + 312, + 400, + 319 + ], + "type": "text", + "content": "(b) FOUND [57]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 403, + 217, + 449, + 311 + ], + "blocks": [ + { + "bbox": [ + 403, + 217, + 449, + 311 + ], + "lines": [ + { + "bbox": [ + 403, + 217, + 449, + 311 + ], + "spans": [ + { + "bbox": [ + 403, + 217, + 449, + 311 + ], + "type": "image", + "image_path": "9ea38f01bb960d52fc92c00555d4cc2e9b6e5a23b90e75aa4cbc1c17825de35e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 312, + 455, + 319 + ], + "lines": [ + { + "bbox": [ + 404, + 312, + 455, + 319 + ], + "spans": [ + { + "bbox": [ + 404, + 312, + 455, + 319 + ], + "type": "text", + "content": "(c) SelfMask [55]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 450, + 217, + 495, + 311 + ], + "blocks": [ + { + "bbox": [ + 450, + 217, + 495, + 311 + ], + "lines": [ + { + "bbox": [ + 450, + 217, + 495, + 311 + ], + "spans": [ + { + "bbox": [ + 450, + 217, + 495, + 311 + ], + "type": "image", + "image_path": "53ddf3eae9959d73c0d1281b0f3a97f7ca3bbf4f694d4949a5aac9aa3c303521.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 312, + 491, + 319 + ], + "lines": [ + { + "bbox": [ + 466, + 312, + 491, + 319 + ], + "spans": [ + { + "bbox": [ + 466, + 312, + 491, + 319 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 320, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 320, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 545, + 386 + ], + "type": "text", + "content": "Figure 1. Existing unsupervised salient object detection methods, FOUND [57] (b) and SelfMask [55] (c), cannot differentiate salient instances (rows 1 and 2), but tend to detect strong semantic instances, e.g., the woman in row 2. In contrast, our method (d) correctly detects all salient instances by modeling saliency coherence and different levels of saliency contrast." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 496, + 217, + 541, + 311 + ], + "blocks": [ + { + "bbox": [ + 496, + 217, + 541, + 311 + ], + "lines": [ + { + "bbox": [ + 496, + 217, + 541, + 311 + ], + "spans": [ + { + "bbox": [ + 496, + 217, + 541, + 311 + ], + "type": "image", + "image_path": "697767f4816ac2c15169534c711550fc3c370169b02f646918398e99fa3e2ead.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 312, + 530, + 319 + ], + "lines": [ + { + "bbox": [ + 511, + 312, + 530, + 319 + ], + "spans": [ + { + "bbox": [ + 511, + 312, + 530, + 319 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 390, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 437 + ], + "type": "text", + "content": "ferentiation not only between salient and non-salient regions but also between different salient regions (Figure 1). In this paper, we propose to study this unexplored task, i.e., unsupervised salient instance detection (USID)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 438, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 653 + ], + "type": "text", + "content": "Our key observation of this problem is built upon the advanced self-supervised transformer representations. We observe that their features exhibit saliency coherence (high similarities) within local regions and may show varying degrees of saliency contrast (lower similarities and dissimilarities) between different regions. In Figure 2, we visualize global saliency information within the self-supervised transformer features (b-d), and the saliency coherence and contrast between three feature pixels and the rest of the image (e-g), where each of these pixels is randomly selected from a salient instance. While the transformer features themselves may exhibit some degree of global saliency contrast (b-d), the computed pixel-based correlations show that each selected pixel tends to have high saliency coherence with pixels of the same instance but different levels of saliency contrast to pixels of other instances and the background (e-g). This observation inspires us to mine regional saliency coherence and contrast information for USID." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel approach, called SCoCo, to model Saliency Coherence and Contrast information from deep self-supervised features to address the USID problem. Specifically, SCoCo first models scene-level saliency contrast to determine the \"saliency thresh" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2702" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 78, + 136, + 161 + ], + "blocks": [ + { + "bbox": [ + 74, + 78, + 136, + 161 + ], + "lines": [ + { + "bbox": [ + 74, + 78, + 136, + 161 + ], + "spans": [ + { + "bbox": [ + 74, + 78, + 136, + 161 + ], + "type": "image", + "image_path": "a849b1a076fdd06f7fb82676c01b2a862ca2b4fec1d0697cffaed2ff4d40dce7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 163, + 108, + 170 + ], + "lines": [ + { + "bbox": [ + 99, + 163, + 108, + 170 + ], + "spans": [ + { + "bbox": [ + 99, + 163, + 108, + 170 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 137, + 78, + 200, + 161 + ], + "blocks": [ + { + "bbox": [ + 137, + 78, + 200, + 161 + ], + "lines": [ + { + "bbox": [ + 137, + 78, + 200, + 161 + ], + "spans": [ + { + "bbox": [ + 137, + 78, + 200, + 161 + ], + "type": "image", + "image_path": "cde78aa92b16fdab70119187ee7cb9baa6d9414b8cfaf0938c77dabbfd70fb6c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 163, + 173, + 170 + ], + "lines": [ + { + "bbox": [ + 163, + 163, + 173, + 170 + ], + "spans": [ + { + "bbox": [ + 163, + 163, + 173, + 170 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 201, + 78, + 264, + 161 + ], + "blocks": [ + { + "bbox": [ + 201, + 78, + 264, + 161 + ], + "lines": [ + { + "bbox": [ + 201, + 78, + 264, + 161 + ], + "spans": [ + { + "bbox": [ + 201, + 78, + 264, + 161 + ], + "type": "image", + "image_path": "ccfb52e2ea8b07b7e3a119fa00d7815df00501e4ebedffbc6a5189105f947cf3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 163, + 237, + 170 + ], + "lines": [ + { + "bbox": [ + 228, + 163, + 237, + 170 + ], + "spans": [ + { + "bbox": [ + 228, + 163, + 237, + 170 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 266, + 78, + 328, + 161 + ], + "blocks": [ + { + "bbox": [ + 266, + 78, + 328, + 161 + ], + "lines": [ + { + "bbox": [ + 266, + 78, + 328, + 161 + ], + "spans": [ + { + "bbox": [ + 266, + 78, + 328, + 161 + ], + "type": "image", + "image_path": "e8ad4739a5d955d6d45bea70645daa869f9f8a1de36540354ee7e1e80aec1940.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 163, + 301, + 170 + ], + "lines": [ + { + "bbox": [ + 292, + 163, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 292, + 163, + 301, + 170 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 266, + 179, + 282, + 187 + ], + "lines": [ + { + "bbox": [ + 266, + 179, + 282, + 187 + ], + "spans": [ + { + "bbox": [ + 266, + 179, + 282, + 187 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 329, + 78, + 392, + 161 + ], + "blocks": [ + { + "bbox": [ + 329, + 78, + 392, + 161 + ], + "lines": [ + { + "bbox": [ + 329, + 78, + 392, + 161 + ], + "spans": [ + { + "bbox": [ + 329, + 78, + 392, + 161 + ], + "type": "image", + "image_path": "c57dba02e7ffe8ddd961ac83d0d759a88e4830fb4387cbfa64d63386348065af.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 163, + 365, + 170 + ], + "lines": [ + { + "bbox": [ + 356, + 163, + 365, + 170 + ], + "spans": [ + { + "bbox": [ + 356, + 163, + 365, + 170 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 180, + 474, + 189 + ], + "lines": [ + { + "bbox": [ + 347, + 180, + 474, + 189 + ], + "spans": [ + { + "bbox": [ + 347, + 180, + 474, + 189 + ], + "type": "text", + "content": "← Pixel-based Saliency Contrast Degree →" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 393, + 78, + 456, + 161 + ], + "blocks": [ + { + "bbox": [ + 393, + 78, + 456, + 161 + ], + "lines": [ + { + "bbox": [ + 393, + 78, + 456, + 161 + ], + "spans": [ + { + "bbox": [ + 393, + 78, + 456, + 161 + ], + "type": "image", + "image_path": "88a1d0ce9fdfd94df4b79c17aeeca1ae27e31753d6557332fc4938b28b7f489c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 163, + 429, + 170 + ], + "lines": [ + { + "bbox": [ + 421, + 163, + 429, + 170 + ], + "spans": [ + { + "bbox": [ + 421, + 163, + 429, + 170 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 457, + 78, + 520, + 161 + ], + "blocks": [ + { + "bbox": [ + 457, + 78, + 520, + 161 + ], + "lines": [ + { + "bbox": [ + 457, + 78, + 520, + 161 + ], + "spans": [ + { + "bbox": [ + 457, + 78, + 520, + 161 + ], + "type": "image", + "image_path": "aab17e91de9f16d9bfaa3a7d8bcda67d7f55aa4730e034a70a62d4ea8b772323.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 163, + 495, + 170 + ], + "lines": [ + { + "bbox": [ + 485, + 163, + 495, + 170 + ], + "spans": [ + { + "bbox": [ + 485, + 163, + 495, + 170 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 477, + 180, + 511, + 188 + ], + "lines": [ + { + "bbox": [ + 477, + 180, + 511, + 188 + ], + "spans": [ + { + "bbox": [ + 477, + 180, + 511, + 188 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 78, + 171, + 136, + 188 + ], + "blocks": [ + { + "bbox": [ + 78, + 171, + 136, + 188 + ], + "lines": [ + { + "bbox": [ + 78, + 171, + 136, + 188 + ], + "spans": [ + { + "bbox": [ + 78, + 171, + 136, + 188 + ], + "type": "image", + "image_path": "368200a641c5a62059d7fb4166b41c715b67a218e86063e23636a6cd6c5fad3e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 179, + 239, + 189 + ], + "lines": [ + { + "bbox": [ + 137, + 179, + 239, + 189 + ], + "spans": [ + { + "bbox": [ + 137, + 179, + 239, + 189 + ], + "type": "text", + "content": "Global Saliency Contrast Degree " + }, + { + "bbox": [ + 137, + 179, + 239, + 189 + ], + "type": "inline_equation", + "content": "\\rightarrow" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 190, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. Given an input image (a), we first visualize its corresponding self-supervised transformer features [8] (b), the pixel correlations to the [CLS] token (c), and the pixel correlations over the global scene (d). We then randomly sample three pixels (purple dots), each for one instance, and visualize the saliency contrast of the transformer features between the sampled pixel and all other pixels of the image by computing feature point correlations (e-g). We observe that while the self-supervised transformer features may show the global saliency contrast to some degree, pixels from the same instance tend to be more coherent, while pixels from different regions tend to have different levels of contrast. This inspires us to mine such regional saliency coherence and contrast for USID." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 305, + 171, + 326, + 179 + ], + "blocks": [ + { + "bbox": [ + 305, + 171, + 326, + 179 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 326, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 326, + 179 + ], + "type": "image", + "image_path": "85f5b9f7b7d35ad3929348ee40b1b82ee268eef7c2fc61c231ffbed97e639dc2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 180, + 343, + 187 + ], + "lines": [ + { + "bbox": [ + 328, + 180, + 343, + 187 + ], + "spans": [ + { + "bbox": [ + 328, + 180, + 343, + 187 + ], + "type": "text", + "content": "一" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 262, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 417 + ], + "type": "text", + "content": "old,\" which differentiates salient regions with respect to diverse background contents. We implement this via a novel global background adaptation (GBA) module with a scene-level contrastive loss, which learns to iteratively pull the saliency regions out of the complex background. SCoCo then models the within-region saliency coherence and cross-region saliency contrast information for grouping pixels into salient instances. We implement this by proposing a novel locality-aware similarity (LAS) module with an instance-level contrastive loss, which initially focuses on modeling the saliency coherence within individual regions and then shifts to model different levels of saliency contrast between different regions." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 59, + 418, + 281, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 418, + 281, + 429 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 281, + 429 + ], + "type": "text", + "content": "To summarize, this paper has three main contributions:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 47, + 430, + 287, + 585 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 488 + ], + "type": "text", + "content": "- To our knowledge, we propose the first USID approach, named SCoCo, to learn to detect salient instances without any annotated labels, by modeling the intra-region coherence and inter-region contrast information from self-supervised transformer features." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 287, + 536 + ], + "type": "text", + "content": "- SCoCo has two novel modules: a GBA module with a scene-level contrastive loss to find the saliency thresholds of different image contents, and a LAS module with an instance-level contrastive loss to detect salient instances." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 47, + 537, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 287, + 585 + ], + "type": "text", + "content": "- Extensive experiments show that SCoCo outperforms existing weakly-supervised SID methods and carefully designed unsupervised baselines, and has comparable performances to fully-supervised SID methods." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 591, + 138, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 138, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 138, + 604 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 47, + 611, + 219, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 219, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 219, + 623 + ], + "type": "text", + "content": "2.1. Salient Instance Detection (SID)" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "SID aims to individualize visually prominent objects in the input image. Existing deep SID methods can be classified into two categories based on their learning labels: fully-supervised and weakly-supervised SID methods." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Fully-supervised SID methods rely on costly pixel-wise annotated masks for training supervision. Li et al. [25] propose to apply MAP [80] on instance proposals, which are" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 262, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 429 + ], + "type": "text", + "content": "generated based on binary saliency and contour maps, to select and segment salient instances. Fan et al. [14] devise a single-stage salient instance detection method, with a RoIMasking layer that utilizes both local and surrounding contexts of each instance for detection. Pei et al. [45] propose a spectral clustering method to obtain salient instances by exploring pixel masks and subitizing information of salient objects, and later a transformer-based method [46] to detect salient instances in a single stage without NMS for post-processing. Liu et al. [37] propose an interleaved execution strategy to incorporate global context and object contour information jointly to detect salient instances. RGB-D information has also been explored in CalibNet et al. [47], with a dual-branch cross-modal calibration method." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 430, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 550 + ], + "type": "text", + "content": "Weakly-supervised SID methods typically use bounding boxes and image-level labels, e.g., classes and subitizing as supervisions. Zhang et al. [80] propose a Maximum A Posteriori (MAP)-based subset optimization formulation to select a compact set of salient instances from the redundant box-level proposals, followed by CRF to help segment pixel-wise instances from box-level predictions. Tian et al. [61, 62] propose a triple-branch network to exploit salient boundaries, centroids, and regions jointly from class and subitizing labels for salient instance detection." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 552, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 610 + ], + "type": "text", + "content": "Despite their success, existing SID methods require human efforts to prepare different types of labels, which are expensive. In contrast, we propose to model the saliency coherence and contrast information based on self-supervised transformer features, for USID." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 306, + 621, + 474, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 474, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 474, + 634 + ], + "type": "text", + "content": "2.2. Salient Object Detection (SOD)" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 640, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 689 + ], + "type": "text", + "content": "SOD is an important task to scene understanding with a lot of methods proposed. Deep SOD methods can be divided into three groups w.r.t. the supervision signals: fully-supervised, weakly-supervised, and unsupervised." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Fully-supervised SOD methods mainly incorporate four types of deep techniques to learn salient object represent-" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 88, + 70, + 125, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 70, + 125, + 77 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 125, + 77 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 70, + 138, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 70, + 138, + 77 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 138, + 77 + ], + "type": "text", + "content": "Sel" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 70, + 151, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 70, + 151, + 77 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 151, + 77 + ], + "type": "text", + "content": "f-sup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 70, + 165, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 70, + 165, + 77 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 165, + 77 + ], + "type": "text", + "content": "ervis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 70, + 185, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 70, + 185, + 77 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 185, + 77 + ], + "type": "text", + "content": "ed tran" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 70, + 200, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 70, + 200, + 77 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 200, + 77 + ], + "type": "text", + "content": "sform" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 70, + 217, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 70, + 217, + 77 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 217, + 77 + ], + "type": "text", + "content": "her fear" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 218, + 70, + 231, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 70, + 231, + 77 + ], + "spans": [ + { + "bbox": [ + 218, + 70, + 231, + 77 + ], + "type": "text", + "content": "tures" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 232, + 70, + 247, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 70, + 247, + 77 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 247, + 77 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 70, + 264, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 70, + 264, + 77 + ], + "spans": [ + { + "bbox": [ + 249, + 70, + 264, + 77 + ], + "type": "text", + "content": "differ" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 265, + 70, + 282, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 70, + 282, + 77 + ], + "spans": [ + { + "bbox": [ + 265, + 70, + 282, + 77 + ], + "type": "text", + "content": "nt visi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 283, + 70, + 302, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 70, + 302, + 77 + ], + "spans": [ + { + "bbox": [ + 283, + 70, + 302, + 77 + ], + "type": "text", + "content": "alizational" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 70, + 484, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 70, + 484, + 77 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 484, + 77 + ], + "type": "text", + "content": "Saliency contrasts for three random pixels" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2703" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "tations from pixel-level annotated masks. First, deep feature fusion [34, 39, 68, 73, 83] is used to aggregate multilevel context information including low-level stimulus and high-level semantics for detecting salient objects. Second, attention mechanisms [30, 35, 36, 58, 84, 85] are used to reweight multi-scale features and enhance context learning to help the model focus on the salient regions and suppress noise of the background regions. Third, boundary enhancement [9, 33, 52, 86] is often applied to improve the localization and segmentation results by penalizing the errors of boundary-surrounding pixels. Last, recurrent mechanisms [22, 29, 60, 66, 84] may also be reused to iterate an identical network to predict and refine the saliency regions in a coarse-to-fine manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 483 + ], + "type": "text", + "content": "Weakly-supervised SOD methods have explored class labels, image captions, scribbles, bounding boxes, and points as supervisions. Wang et al. [67] propose the first weakly-supervised SOD method, which leverages image-level class labels to localize salient regions and then refines those predicted regions with CRF. Following [67], Li et al. [26] propose a pseudo saliency annotation updating scheme to refine the spatial coherence of predicted salient regions. Piao et al. [50] propose the multi-filter directive network to distill clean saliency maps from multiple class-based pseudo labels. Zeng et al. [78] devise a multi-source weak supervision framework to jointly utilize category labels, captions, and a set of unlabelled images for training. Based on scribble annotations, WSSA [82] and AGGM [77] are proposed to detect salient objects by exploring local consistency and structure information of objects. Liu et al. [40] propose to use bounding box supervisions to help address the over/under detection problems. Gao et al. [15] design an adaptive masked flood filling algorithm to detect salient objects from point-wise labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Unsupervised SOD methods can be classified into four categories according to the features and technologies used. First, conventional methods are typically based on handcrafted features [2, 5, 38, 49, 76, 92], e.g., local and global contrast [10, 38, 49], boundary priors [27, 76, 92], and spatial frequency [2], for detecting salient objects. These methods may not handle complex scenes as their low-level features have limited representation capacities. Second, a series of methods [32, 44, 72, 79, 81, 87, 88] try to distill clean labels from pseudo labels, which are produced by some of the aforementioned conventional methods, for learning SOD models. Specifically, they use distillation strategies like iterative refinement [44], causal debiasing [32], and mining textures around boundaries [89] to reduce the noise coming from the pseudo labels. Third, instead of learning from noisy pseudo labels, Yan et al. [75] propose a domain adaptive SOD method that learns from synthetic but clean saliency data. Fourth, FOUND [57] and SelfMask [55] also detect salient objects based on self-supervised deep" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "features. However, our SCoCo differs from them in both tasks and proposed techniques. First, they cannot predict instance-level saliency information, as shown in row 1 of Figure 1(b,c). Second, they only model the binary contrast between salient and non-salient regions; they are unable to suppress non-salient regions/objects with uncertain contrast degrees. For example, in row 2 of Figure 1(c,d), FOUND and SelfMask focus mainly on the most prominent woman, neglecting the umbrella, as the woman has much stronger semantics than the umbrella. In contrast, our SCoCo (Figure 1(d)) can correctly individualize salient instances by modeling intra-region saliency coherence and different levels of saliency contrast between regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 277 + ], + "type": "text", + "content": "In summary, unlike the above SOD methods, which do not consider instance-level object saliency, our work aims to detect salient instances but without the need to use annotations for supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 285, + 509, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 509, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 509, + 298 + ], + "type": "text", + "content": "2.3. Unsupervised Object Detection (UOD)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 304, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 411 + ], + "type": "text", + "content": "UOD aims to localize objects without using any human annotations. Early methods [11, 63-65] explore inter-image similarities (e.g., clustering, ranking) to discover objects for an image collection. Recently, several methods learn to detect objects based on self-supervised deep features. LOST [56] and TokenCut [71] propose to segment one object per image based on the pixel graph of self-supervised deep features. FreeSOLO [69] and CutLER [70] learn to detect multiple objects by self-retraining [69] and MaskCut [70]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 411, + 547, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 436 + ], + "type": "text", + "content": "While these methods tend to detect any objects (salient, non-salient), our method learns to identify salient instances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 448, + 389, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 389, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 389, + 461 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 463, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 593 + ], + "type": "text", + "content": "Learning an unsupervised salient instance detector is challenging, as there are no explicit labels/knowledge to teach the detector to identify salient regions at the instance level. In this work, we observe some interesting cues in the self-supervised transformer features - they exhibit saliency coherence within local regions but saliency contrast of diverse degrees between different regions. If a detector can model such correlations well, it may be able to identify salient regions and group them into separate salient instances. Here, we propose the SCoCo method to mine regional saliency coherence and contrast information for USID." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Figure 3 shows the overview of our SCoCo. Given an input image, we first apply a self-supervised transformer (with frozen weights) to generate the corresponding deep features. We then feed these deep features separately to a novel global background adaptation (GBA) module (Section 3.1) and a novel locality-aware similarity (LAS) module (Section 3.2). The GBA module with its scene-level contrastive loss aims to find the suitable 'saliency thresholds' to distinguish salient regions from diverse background contents to produce a scene-level saliency map. Meanwhile," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "2704" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 521, + 139 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 521, + 139 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 521, + 139 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 521, + 139 + ], + "type": "image", + "image_path": "5e3e9751d07c12cb514c909341476bae32d44d3a3cfff2392b65f8bed21ee6af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": "Figure 3. The overview of our unsupervised pipeline. Given an image, we first obtain its features " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " from a self-supervised transformer (whose weights are frozen). We then use " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " to produce pseudo scene- and instance-level supervisions " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{S+}}" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{I+}^n}" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "1 \\leq n \\leq N" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": ") by computing feature similarities, and feed " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " to our novel GBA and LAS modules. The GBA module learns with a scene-level contrastive loss " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " to predict a saliency map " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{P_s}" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": ", while the LAS module learns with an instance-level contrastive loss " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": " to model saliency coherence and contrast features. The output features of both GBA and LAS modules are combined in the salient instance detection head to produce the final salient instance map " + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{P_I}" + }, + { + "bbox": [ + 46, + 141, + 547, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 289, + 293 + ], + "type": "text", + "content": "the LAS module works with an instance-level contrastive loss to learn intra-region saliency coherence and interregion saliency contrast features. These features are then fed into a salient instance detection head to produce salient instances, with the guidance of the scene-level saliency map produced by the GBA module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 288, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 288, + 314 + ], + "type": "text", + "content": "3.1. Global Background Adaptation (GBA) Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 499 + ], + "type": "text", + "content": "The goal of the GBA module is to adaptively dig saliency information from diverse background contents. As Figure 2(b-d) shows, it is challenging to find a fixed saliency threshold to select salient regions from different scenes with diverse complexities. To address this issue, we propose the GBA module, which has learnable saliency threshold functions working in a bifurcation-and-combination manner, to detect the fine-grained adaptive saliency thresholds. Specifically, as shown in Figure 4, the GBA module is designed to have two bifurcated branches at the first, with each of them equipped with its own learnable saliency threshold function, to handle the potential intra-scene saliency-background ambiguities separately. The two bifurcated branches are then combined with another saliency threshold-based branch to produce the scene-level saliency map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 287, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 607 + ], + "type": "text", + "content": "Our GBA module differs from previous unsupervised salient object detection methods [55, 57] in the way of detecting saliency thresholds. Existing USOD methods typically rely on stacking convolutional operations to detect salient objects. However, such operations (even with dynamic convolution kernels [24, 59, 90]) are content-agnostic as convolution filters are shared across all the feature pixels. As a result, these methods often fail to detect the salient regions buried in diverse background contents." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": "Network Structure. Given features " + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": ", we first process it with two parallel branches, each consisting of a group of " + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": " conv operators, and a learnable saliency threshold function " + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 607, + 287, + 657 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 666, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 666, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 76, + 666, + 287, + 680 + ], + "type": "interline_equation", + "content": "f (x) = \\left(p _ {1} - p _ {2}\\right) x \\cdot \\sigma \\left(\\theta \\left(p _ {1} - p _ {2}\\right) x\\right) + p _ {2} x, \\tag {1}", + "image_path": "0993da1bda7ab01add2614dc18429ab651d548680ddc53192dfbb1e445180819.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is a smooth version of Swish rectifying function [42, 53], " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " are learnable parameters that can jointly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "content": "determine the upper and lower bounds of the function, " + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "content": " is the Sigmoid function, and the learnable parameter " + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "content": " controls the \"selection degree\" of salient objects with respect to the background. Each " + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "content": " is learned using two fully connected (FC) layers upon a global image vector generated by the global average pooling (GAP) layer. We then use another branch to combine the former two branches, with another set of convolutional operations for feature transformation and learnable saliency threshold function to unify the separately modeled saliency thresholds, to produce the scene-level saliency map." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": "Scene-level Contrastive Loss. To boost the saliency learning of the GBA module, we introduce a scene-level contrastive loss " + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": ". Since the image background is expected to have low saliency degrees, we first derive the scene-level pseudo saliency supervision " + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{\\mathrm{S}+}" + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": " through inverting the identified background regions with low saliency degrees. However, we note that such pseudo-saliency maps may contain a significant number of noisy pixels, particularly those false positive background pixels with medium saliency degrees, that may hinder learning efficacy. To address this problem, we leverage the pixels with high saliency degrees in our instance-level saliency predictions " + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{P_I}" + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": " (introduced in Section 3.2) to help classify those uncertain pixels into salient and non-salient ones. To this end, we take " + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{P_I}" + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": " as positive samples and utilize scene-level saliency in different images as negative samples. " + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 304, + 352, + 546, + 544 + ], + "type": "text", + "content": " can then be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 550, + 577, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 577, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 577, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {S}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau}}{e ^ {\\left(\\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {i}\\right) + \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {I}} ^ {i}\\right)\\right) / \\tau} + \\sum_ {k = 1 , k \\neq i} ^ {N} e ^ {\\left. \\mathrm {d} \\left(\\mathrm {P} _ {\\mathrm {s}} ^ {i} , \\mathrm {P} _ {\\mathrm {s} +} ^ {k}\\right) / \\tau \\right.}}, \\tag {2}", + "image_path": "e106039b413ab5e252d17cecb95f03db1c79650380e94488a80e4ea127f34fc0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{N}" + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{\\mathrm{S}}" + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "content": " is the scene-level saliency prediction, " + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{d}" + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "content": " measures the mean square error (MSE), and " + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 594, + 547, + 629 + ], + "type": "text", + "content": " is a temperature parameter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 518, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 518, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 518, + 647 + ], + "type": "text", + "content": "3.2. Locality-aware Similarity (LAS) Module" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "The LAS module aims to capture regional saliency coherence information from the self-supervised deep features. This information can then be used to help learn different levels of saliency contrast using an instance-level contrastive loss. We note that the latest saliency instance de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2705" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 69, + 251, + 256 + ], + "blocks": [ + { + "bbox": [ + 83, + 69, + 251, + 256 + ], + "lines": [ + { + "bbox": [ + 83, + 69, + 251, + 256 + ], + "spans": [ + { + "bbox": [ + 83, + 69, + 251, + 256 + ], + "type": "image", + "image_path": "2949c6faab73fb952a69682528e9af672190b3016d27a9d458ed8393d7c115c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 289, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 289, + 305 + ], + "type": "text", + "content": "Figure 4. The bifurcation-to-combination learning structure of the global background adaptation (GBA) module. The pink lines indicate the learning flows of learnable saliency threshold functions for extracting salient objects from different background contents." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": "tection methods [37, 46] typically leverage the object detector head [6, 17] (a combination of convolutions and self-attention blocks) to classify and group salient instance pixels. However, although convolutional filters are good at detecting repeated patterns, they may fail in discerning pixel similarities. The self-attention mechanisms tend to model the interactions of pixel tokens but do not compare region-level saliency contrasts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": "On the other hand, our LAS module first models intra-region feature coherence by computing channel-wise correlations with respect to regions surrounding every pixel location, and then feeds regional features to an instance-level contrastive loss to learn regional saliency contrasts for discriminating salient instances. The overview of LAS module is depicted in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": "Network Structure. Given features " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": ", we downsample it to " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "F' \\in \\mathbb{R}^{H \\times W \\times C'}" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": " with a linear layer. We then compute cosine similarity between a positional feature " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "x \\in [1, H] \\times [1, W]" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": " and the surrounding region of " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": ", and collect them into a similarity feature tensor " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{H \\times W \\times M \\times M \\times C'}" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": ". The formulation of " + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 497, + 287, + 570 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 577, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 101, + 577, + 287, + 609 + ], + "type": "interline_equation", + "content": "S (x, m) = \\frac {F _ {(x)} ^ {\\prime} \\cdot F _ {(x + m)} ^ {\\prime}}{\\| F _ {(x)} ^ {\\prime} \\| \\cdot \\| F _ {(x + m)} ^ {\\prime} \\|}, \\tag {3}", + "image_path": "725d848214ad4f695bc6a5008e6bec2c8a6c49d1df15aaf78e98bc2fc06e9273.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "m \\in [-M + 1, M - 1] \\times [-M + 1, M - 1]" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " is the region size. Eq. 3 aims to encode the local saliency coherence of each pixel " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ". We further apply convolutional filters on " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " to compact regional features. The spatial size of " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " is progressively aggregated from " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " with cascaded " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " convolutional filters (without padding) to obtain " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " with the same spatial size as the input feature " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ". After each convolutional filter, we also" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "insert the batch normalization and rectifying linear unit layers. Finally, we combine " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " to produce the final features " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\hat{F}'" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " with saliency coherence information, which further facilitates the saliency contrast modeling for differentiating salient instances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "text", + "content": "Instance-level Contrastive Loss. We propose the instance-level contrastive loss " + }, + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "text", + "content": " to derive the saliency contrast information from " + }, + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\hat{F}^{\\prime}" + }, + { + "bbox": [ + 304, + 133, + 545, + 204 + ], + "type": "text", + "content": " for salient instance discrimination, as the binary cross entropy (BCE) loss adopted by existing unsupervised SOD methods [56, 69, 70] cannot distinguish different salient instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": "Given the self-supervised transformer features, we extract the key embeddings " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", and feed the self-affinity matrix of " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " to Normalized Cut [54, 71] to find highly coherent salient pixels as pseudo salient instance supervision " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{\\mathrm{I}+}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " in an iterative manner. We find one salient instance " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}+}^{k}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " in each iteration " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "k \\in [1, K]" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", and the iteration ends if the " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}+}^{K+1}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " is located in the low-saliency background. Note that " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{I}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " is used in " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " (Eq. 2) for penalizing pixels with uncertain saliency degrees. Meanwhile, we can measure the spatial discrepancy between predicted instances " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}}^{L}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}+}^{K}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " to formulate " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " in the number of instances. To reduce the influence of noise that may be contained in pseudo labels, we propose to sample point vectors from " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}'" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}', \\mathbf{P}_{\\mathrm{I}}^{L}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}+}^{K}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " are spatially aligned for the sampling. By sampling a small subset of point vectors, we make the model less likely to be influenced by extreme noise that may be present in the full pseudo label set. Specifically, given " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}'" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", we randomly sample " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "T(t \\in [1, T])" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " triplets of point vectors " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{I}}^{l(t)}, \\hat{F}_{\\mathrm{I}+}^{l(t)}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{I}-}^{l(t)}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\hat{F}'" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " with respect to each predicted instance " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}}^{l}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": ", its positive supervision " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}+}^{k}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " (IoU " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "(\\mathbf{P}_{\\mathrm{I}}^{l}, \\mathbf{P}_{\\mathrm{I}+}^{k}) \\geq 0.7" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": "), and negative supervision " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{I}-}^{f}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "f \\neq k, f \\in [1, K]" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": "), respectively. " + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 490 + ], + "type": "text", + "content": " can then be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 499, + 545, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 499, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 334, + 499, + 545, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {I}} = - \\frac {1}{L \\times T ^ {2}} \\sum_ {l = 1} ^ {L} \\sum_ {t = 1} ^ {T} \\sum_ {t = 1} ^ {T} \\log \\left(g \\left(\\hat {F} ^ {\\prime}, l, t\\right)\\right), \\tag {4}", + "image_path": "f3a28086cc18dca3fa954d0fe111fe227dc395a7600381699f316eb331c8f3d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "text", + "content": " is a contrastive estimation function, which is formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 577, + 556, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 556, + 620 + ], + "type": "interline_equation", + "content": "g \\left(\\hat {F} ^ {\\prime}, l, t\\right) = \\frac {e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau}}{e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} +} ^ {l (t)} \\right\\rangle / \\tau} + \\sum_ {t = 1} ^ {T} e ^ {s i m \\left\\langle \\hat {F} _ {\\mathrm {I}} ^ {l (t)} , \\hat {F} _ {\\mathrm {I} -} ^ {l (t)} \\right\\rangle / \\tau}}, \\tag {5}", + "image_path": "8f45b14297aea90716a459a948047428eaf681fe4358b661a838f7334ed786d5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "sim\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " is the cosine similarity function to compute the distance between two point vectors " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{I}}^{l(t)}" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{I}+}^{\\prime l(t)}" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "l(t)" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "t_{th}" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " point vector of the " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "l_{th}" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " instance. Each vector is extracted from " + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "inline_equation", + "content": "\\hat{F}^{\\prime}" + }, + { + "bbox": [ + 304, + 620, + 545, + 720 + ], + "type": "text", + "content": " according to its spatial location. Our instance-level contrastive loss (Eq. 4) helps the network learn the amplified saliency contrast of different instances while keeping the saliency coherence within each instance by maintaining their self-similarities." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2706" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 500, + 157 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 500, + 157 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 500, + 157 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 500, + 157 + ], + "type": "image", + "image_path": "b8f9f112f8f2f426b9a78a337766e43d6bdc56072be9c3bae94672b76204e51f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 164, + 418, + 176 + ], + "lines": [ + { + "bbox": [ + 175, + 164, + 418, + 176 + ], + "spans": [ + { + "bbox": [ + 175, + 164, + 418, + 176 + ], + "type": "text", + "content": "Figure 5. The structure of locality-aware similarity (LAS) module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 184, + 128, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 128, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 128, + 198 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 174, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 174, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 174, + 217 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": "Implementation Details. We use DINO [8] to initialize the self-supervised transformer (i.e., ViT [13]) of our SCoCo network. Note that neither DINO nor our SCoCo uses any labeled data (e.g., class labels) for supervision. The input to both GBA and LAS modules, i.e., " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": ", is the key features of multi-head attention layers in the last feature block of DINO-based ViT. In the GBA module, we use the mean attention value of " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " as the low saliency degree per image to find its scene-level saliency region " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{S+}}" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": ". In the LAS module, we sample " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " triplets of point vectors to formulate the loss " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " in both " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " is set to 0.1. We adopt the head of Mask-RCNN [17] as our salient instance detection head. During training, each input image has a resolution of " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": ", and is augmented using largescale jittering [16]. The batch size " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " is set to 4 and we use AdamW " + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "inline_equation", + "content": "(\\beta_1 = 0.9, \\beta_2 = 0.999)" + }, + { + "bbox": [ + 46, + 222, + 289, + 484 + ], + "type": "text", + "content": " with step-wise learning rate decay for model optimization. We first train our SCoCo network for 10 epochs using the pseudo supervision derived from DINO. We then update the pseudo labels using SCoCo. Last, we train SCoCo with the updated pseudo labels for another epoch. SCoCo is trained on two 3090ti GPU cards." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. We conduct our experiments on three SID datasets, including ILSO-1K [25], ILSO-2K [25], and SIS [46]. The (training/validation/test) sets in the ILSO-1K [25], ILSO-2K [25], and SIS [46] contain (500/200/300), (1000/400/600), and (7030/2100/1170) images, respectively. For each dataset, we only use the unlabeled images of the training set to train our SCoCo network, and we evaluate the performance on the corresponding test set. We adopt the mean Average Precision (mAP) metric for evaluation, following previous SID approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 113, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 113, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 113, + 623 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "Since we propose the first unsupervised method to detect salient instances, to evaluate our method comprehensively, we design several baselines from three aspects, inspired by related areas, as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- First, we note that existing unsupervised object detection (OD) methods, i.e., FreeSOLO [69], and CutLER [70], detect objects in a saliency-agnostic way. We retrain their" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 186, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 233 + ], + "type": "text", + "content": "models with our pseudo salient instance labels for comparison. Their models are initialized with the model parameters trained on their pseudo labels extracted from large-scale datasets, e.g., COCO [31] and ImageNet [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 234, + 545, + 365 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 234, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 234, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 545, + 293 + ], + "type": "text", + "content": "- Second, VitDet [28] is the best-performing method among fully-supervised object detection methods that are based on the ViT [13] backbone. Since we also use ViT as our backbone, we retrain VitDet with our salient instance pseudo labels for comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 293, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 365 + ], + "type": "text", + "content": "- Third, we adapt existing USOD methods, i.e., SelfMask [55] and FOUND [57], to detect salient instances. To derive instance-level saliency predictions, we compute IoU between the region-level saliency predictions from USOD methods and the instance predictions from CutlER [70] (a state-of-the-art unsupervised object detection method)." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 372, + 391, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 391, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 391, + 384 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "We compare our method to 19 state-of-the-art SID methods, including seven fully-supervised methods: MSRNet-V1 [25], MSRNet-V2 [25], S4Net [14], SCG [37], MDNN [45], RDPNet [74], and OQTR [46]; six weakly-supervised methods: MAP [80], NLDF [41], DeepMask [21], PRM+D [12], IRN [3], and WSID-Net [62]; and six unsupervised baselines: FreeSOLO [69], CutLER [70], VitDet [28] initialized with the classification backbone, VitDet[28] initialized with the DINO-based backbone, SelfMask [55], and FOUND [57]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Quantitative Comparisons. Table 1 shows the quantitative results. First, we can see that our method outperforms all carefully designed unsupervised baselines (the bottom group). This shows that simply retraining unsupervised methods with our pseudo labels and/or filtering out nonsalient objects predicted by unsupervised object detectors is not effective in deriving accurate salient instance predictions. Second, we can also see that our method surpasses all existing weakly-supervised methods (middle group), while achieving comparable performances to the fully-supervised methods (first group). Existing weak supervision signals (i.e., object-level pixel masks, bounding boxes, classes, and subitizing labels) cannot provide instance-level supervision, which limits the performance of existing weakly-supervised methods. In contrast, even without supervision, SCoCo can learn to derive instance-level saliency information by modeling saliency coherence and contrast information based on" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2707" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 140, + 525, + 369 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with 7 fully-supervised methods, 6 weakly-supervised methods, and 6 unsupervised baselines mentioned in Section 4.2. Column 1 shows the supervision and annotation types of the methods listed in Column 2. Column 3 shows their original tasks (SID: salient instance detection, SOD: salient object detection, OD: object detection, SIS: semantic instance segmentation). Column 4 shows backbone initialization choices, where Cls denotes that their backbones were pre-trained on class labels of ImageNet, and DINO is a self-supervised backbone pretraining method. - denotes missing results as codes/results are not publicly available. For each supervision category, best results are marked in bold, while second-best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 140, + 525, + 369 + ], + "lines": [ + { + "bbox": [ + 67, + 140, + 525, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 140, + 525, + 369 + ], + "type": "table", + "html": "
Supervision type & Annotation typeMethodOriginal taskInit.Datasets and Metrics (mAP IoU ↑)
ILSO-1K [25]ILSO-2K [25]SIS [46]
mAP50mAP70mAP50mAP70mAP50mAP70
Full SupervisionMSRNet-V1 [25]SIDCls65.3%52.1%----
MSRNet-V2 [25]SIDCls85.1%74.7%78.3%66.5%--
[Images are labeled with instance-level pixel masks.]S4Net [14]SIDCls82.2%59.6%73.1%52.9%86.7%63.6%
SCG [37]SIDCls88.8%78.5%79.8%68.9%83.4%71.2%
MDNN [45]SIDCls84.9%67.8%76.1%63.6%84.6%67.4%
RDPNet [74]SIDCls88.9%73.8%80.7%67.2%82.0%69.4%
OQTR [46]SIDCls89.2%81.0%81.4%72.0%88.1%81.7%
Weak SupervisionMAPΔ [80]SIDCls56.6%24.8%51.6%30.3%58.4%22.7%
[Images are labeled with region-level pixel masks*, bounding boxesΔ, classes°, or subitizing⊗]NLDF* [41]SODCls45.5%24.5%43.8%25.2%46.4%24.2%
DeepMaskΔ [51]ODCls37.1%20.5%35.4%18.4%36.4%21.3%
PRM+D⊗[12]SISCls49.6%31.2%43.7%29.8%52.9%34.2%
IRN° [3]SISCls57.1%37.4%50.2%38.4%60.1%40.8%
WSID-Net⊗[62]SIDCls68.3%51.7%59.4%44.6%68.4%47.1%
No SupervisionFreeSOLOΦ [69]ODCls56.8%43.7%50.7%35.8%55.4%43.1%
FreeSOLOΦ [70]ODDINO53.2%40.2%46.9%32.6%51.8%40.3%
[Million-levelhundred thousand-levelΦ, or thousand-levelunlabeled images are used for training.]CutLERΦ [70]ODCls68.6%54.8%63.2%48.3%66.9%54.4%
CutLERΦ [70]ODDINO65.4%52.1%59.8%44.9%62.4%51.5%
VitDet [28]SISCls55.6%43.6%51.9%35.8%54.4%44.1%
VitDet [28]SISDINO55.7%43.4%52.4%37.0%56.5%44.8%
SelfMask [55]SODDINO65.5%51.1%58.9%44.0%64.8%52.1%
FOUND [57]SODDINO66.2%51.3%61.4%45.9%67.1%53.8%
Ours [■]SIDDINO70.7%58.3%64.8%50.4%71.4%55.7%
", + "image_path": "f44b16ece7493a942518842b9ca03100b644f552c0482bb77dde470001aef606.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 230, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 230, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 230, + 396 + ], + "type": "text", + "content": "just the self-supervised Transformer features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 288, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 288, + 636 + ], + "type": "text", + "content": "Qualitative Comparison. Figure 6 compares the visual results of our method and 10 top performing methods from the three groups of methods in Table 1. We can see that the unsupervised baselines (columns 8 to 11) often fail to detect the complete salient instances and produce messy predictions. For the weakly-supervised methods (columns 5 to 7), MAP, which is trained with bounding box labels, may fail to segment complete objects even with the CRF post-processing. IRN and WSID-Net, which are trained with different class labels, may struggle to determine the boundaries between salient instances of the same class, e.g., the sausages in row 3. Fully-supervised methods (columns 2 to 4) tend to be trained to predict as many salient instances as possible. However, they may predict multiple instances from a salient instance of self-contrast (e.g., rows 1, 2, 5 and 6) and sometimes miss salient instances of solid colors (e.g., rows 4 and 6). In comparison, our method can successfully detect salient instances and delineate their boundaries, with the help of learning intra-instance saliency coherence and cross-instance saliency contrast." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Ablation Study of SCoCo. We first study the effectiveness of the proposed GBA and LAS modules, and the loss functions " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " used in them. The " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " results in Table 2 show that solely applying either scene-level (GBA +" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 420, + 529, + 491 + ], + "blocks": [ + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "text", + "content": "Table 2. Ablation study of the GBA and LAS modules, and the corresponding losses, " + }, + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{S}}" + }, + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{I}}" + }, + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "text", + "content": ", on the SIS [46] test set using mAP metrics. Best performances are marked in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 420, + 529, + 491 + ], + "lines": [ + { + "bbox": [ + 324, + 420, + 529, + 491 + ], + "spans": [ + { + "bbox": [ + 324, + 420, + 529, + 491 + ], + "type": "table", + "html": "
MethodGBALSLASL2mAP50mAP70
SCoCo✓64.7%52.6%
✓✓67.0%53.5%
✓68.4%54.1%
67.4%53.1%
✓✓69.0%54.4%
✓✓✓✓71.4%55.7%
", + "image_path": "70f36373b38642d50bfafb649431d88aa69823f09bd564bb05adce14f6c61c5c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{S}}" + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "text", + "content": " or instance-level " + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "inline_equation", + "content": "(\\mathrm{LAS} + \\mathcal{L}_{\\mathcal{I}})" + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "text", + "content": " saliency contrast learning may produce a performance decrease of " + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "inline_equation", + "content": "2 - 3\\%" + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "text", + "content": ", and the performance drops by around " + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 305, + 504, + 545, + 564 + ], + "type": "text", + "content": " if we use the GBA or LAS module only. Combining all of them produces the best performance, which verifies the necessities of our designs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 662 + ], + "type": "text", + "content": "Visualization of the GBA Module. Figure 7 visualizes the scene-level salinocy map learned by the GBA module and the corresponding SID result. We can see that without the GBA module, the model tends to be distracted by high-contrast background regions, such as the shadow, and identifies them as salient regions. In contrast, using the GBA module can successfully suppress such noise by learning proper saliency thresholds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "Visualization of the LAS Module. In Figure 8, we investigate the influence of our LAS module by visualizing feature correlations to points (purplish dots) in individual instances. We can see that the two pigs have non-uniform white and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2708" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 542, + 274 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 542, + 274 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 542, + 274 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 542, + 274 + ], + "type": "image", + "image_path": "e7b7ab1386600035a140e92d3d389693b8610706e0536dc591e6629882c4754a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 281, + 538, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 538, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 538, + 293 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with existing state-of-the-art methods and baselines. Each color represents a unique salient instance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 302, + 269, + 347 + ], + "blocks": [ + { + "bbox": [ + 67, + 302, + 269, + 347 + ], + "lines": [ + { + "bbox": [ + 67, + 302, + 269, + 347 + ], + "spans": [ + { + "bbox": [ + 67, + 302, + 269, + 347 + ], + "type": "image", + "image_path": "65b184711d9a46e22a67642fc2df6ac78a774c90cc5f98398037e1837df3db2a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 350, + 247, + 361 + ], + "lines": [ + { + "bbox": [ + 86, + 350, + 247, + 361 + ], + "spans": [ + { + "bbox": [ + 86, + 350, + 247, + 361 + ], + "type": "text", + "content": "Figure 7. Visualization of the GBA module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 365, + 152, + 429 + ], + "blocks": [ + { + "bbox": [ + 57, + 365, + 152, + 429 + ], + "lines": [ + { + "bbox": [ + 57, + 365, + 152, + 429 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 152, + 429 + ], + "type": "image", + "image_path": "7e11b02fad18bcf18a0c48c52c704079f9ad09b9562605f1dc00448619de10c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 431, + 130, + 440 + ], + "lines": [ + { + "bbox": [ + 79, + 431, + 130, + 440 + ], + "spans": [ + { + "bbox": [ + 79, + 431, + 130, + 440 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 154, + 364, + 274, + 392 + ], + "blocks": [ + { + "bbox": [ + 154, + 364, + 274, + 392 + ], + "lines": [ + { + "bbox": [ + 154, + 364, + 274, + 392 + ], + "spans": [ + { + "bbox": [ + 154, + 364, + 274, + 392 + ], + "type": "image", + "image_path": "cc9f992aa265ac132147baefa8cc3b10cd1bb92e62f9c258a35ffa2c75bb3370.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 393, + 274, + 402 + ], + "lines": [ + { + "bbox": [ + 154, + 393, + 274, + 402 + ], + "spans": [ + { + "bbox": [ + 154, + 393, + 274, + 402 + ], + "type": "text", + "content": "(b) Pixel-based saliency contrast w/o LAS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 155, + 403, + 274, + 430 + ], + "blocks": [ + { + "bbox": [ + 155, + 403, + 274, + 430 + ], + "lines": [ + { + "bbox": [ + 155, + 403, + 274, + 430 + ], + "spans": [ + { + "bbox": [ + 155, + 403, + 274, + 430 + ], + "type": "image", + "image_path": "18e1ef80f001548192bba9c96d2b4a900d623ec60df669fa47ab00efaabb631c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 431, + 272, + 440 + ], + "lines": [ + { + "bbox": [ + 156, + 431, + 272, + 440 + ], + "spans": [ + { + "bbox": [ + 156, + 431, + 272, + 440 + ], + "type": "text", + "content": "(c) Pixel-based saliency contrast w/ LAS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 446, + 246, + 456 + ], + "lines": [ + { + "bbox": [ + 87, + 446, + 246, + 456 + ], + "spans": [ + { + "bbox": [ + 87, + 446, + 246, + 456 + ], + "type": "text", + "content": "Figure 8. Visualization of the LAS module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "type": "text", + "content": "black colors on their bodies, causing our model to learn incomplete saliency coherence for each pig and produce overdetection results without the LAS module (b). With the LAS module, our model can separate the two pigs successfully by learning saliency contrast between instances while maintaining intra-instance saliency coherence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 539, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 634 + ], + "type": "text", + "content": "Backbone Initialization. Finally, we study the alternatives of self-supervised methods for initializing our ViT backbone. As shown in Table 3, in comparison to MoCo-v3 [18] and SWaV [7], we choose DINO [8] to initialize our backbone as it can provide better performances from its learned feature similarities. In addition, we select ViT-S as our backbone since ViT-B can only introduce minor improvements with larger parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we have studied a new task, unsupervised salient instance detection. We have observed that self-supervised transformer features exhibit intra-region saliency coherence and different levels of inter-region" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 332, + 331, + 520, + 385 + ], + "blocks": [ + { + "bbox": [ + 305, + 303, + 544, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 544, + 324 + ], + "type": "text", + "content": "Table 3. Quantitative analysis of backbone initialization choices. Best performances are marked in bold." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 331, + 520, + 385 + ], + "lines": [ + { + "bbox": [ + 332, + 331, + 520, + 385 + ], + "spans": [ + { + "bbox": [ + 332, + 331, + 520, + 385 + ], + "type": "table", + "html": "
MethodInit.Arch.\\( \\mathbf{{mAP}}_{50} \\)\\( {\\mathbf{{mAP}}}_{70} \\)
SCoCoMoCo-v3 [18]ViT-S70.5%54.9%
SWaV [7]ViT-S70.2%54.3%
DINO [8]ViT-S71.4%55.7%
DINO [8]ViT-B71.9%55.6%
", + "image_path": "5dba184c3f7d0f3400e3a2807ada324812f8049484022e73f957edb6603f917a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 395, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 538 + ], + "type": "text", + "content": "saliency contrast, which can help differentiate salient instances. With this observation, we have proposed SCoCo, with two novel designs, to model such regional saliency coherence and contrast information. First, we propose a global background adaptation (GBA) module and a scene-level contrastive loss to find saliency thresholds for salient objects w.r.t. diverse background contents. Second, we propose a locality-aware similarity (LAS) module and an instance-level contrastive loss to model in-region saliency coherence and cross-region saliency contrasts for detecting salient instances. Extensive experiments have verified the effectiveness of our method against SOTA methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 539, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 621 + ], + "type": "text", + "content": "Our method does have limitations. We note that SCoCo may not be able to detect small salient instances as it learns from high-level, low-resolution features. As shown in Figure 9, the person standing on the speedboat is too small to be detected by SCoCo. As a future work, it would be interesting to study multi-granularity salient feature similarity and contrast modeling to detect multi-scale salient instances." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 632, + 539, + 692 + ], + "blocks": [ + { + "bbox": [ + 310, + 632, + 539, + 692 + ], + "lines": [ + { + "bbox": [ + 310, + 632, + 539, + 692 + ], + "spans": [ + { + "bbox": [ + 310, + 632, + 539, + 692 + ], + "type": "image", + "image_path": "a98f5c76c139e8c0f39129323bba345224258fb95fe6535e3db6143fd57022ca.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 695, + 532, + 706 + ], + "lines": [ + { + "bbox": [ + 318, + 695, + 532, + 706 + ], + "spans": [ + { + "bbox": [ + 318, + 695, + 532, + 706 + ], + "type": "text", + "content": "Figure 9. SCoCo may fail to detect small salient instances." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2709" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Kfir Aberman, Junfeng He, Yossi Gandelsman, Inbar Mosseri, David E Jacobs, Kai Kohlhoff, Yael Pritch, and Michael Rubinstein. Deep saliency prior for reducing visual distraction. In CVPR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "text", + "content": "[2] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In CVPR, 2009. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In CVPR, 2019. 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "text", + "content": "[4] Codruta Orniana Ancuti, Cosmin Ancuti, and Phillipe Bekaert. Enhancing by saliency-guided decolorization. In CVPR, 2011. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 258 + ], + "type": "text", + "content": "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVM, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 379 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[9] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE TIP, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 424, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 424, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 287, + 457 + ], + "type": "text", + "content": "[10] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 459, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 501 + ], + "type": "text", + "content": "[11] Minsu Cho, Suha Kwak, Cordelia Schmid, and Jean Ponce. Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals. In CVPR, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 536 + ], + "type": "text", + "content": "[12] Hisham Cholakkal, Guolei Sun, Fahad Shahbaz Khan, and Ling Shao. Object counting and instance segmentation with image-level supervision. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[14] Ruochen Fan, Ming-Ming Cheng, Qibin Hou, Tai-Jiang Mu, Jingdong Wang, and Shi-Min Hu. S4net: Single stage salient-instance segmentation. In CVPR, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[15] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "type": "text", + "content": "[19] Brian Hu, Bhavan Vasu, and Anthony Hoogs. X-mir: Explainable medical image retrieval. In WACV, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "text", + "content": "[20] Lai Jiang, Yifei Li, Shengxi Li, Mai Xu, Se Lei, Yichen Guo, and Bo Huang. Does text attract attention on e-commerce images: A novel saliency prediction dataset and method. In CVPR, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "type": "text", + "content": "[21] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "type": "text", + "content": "[22] Yun Yi Ke and Takahiro Tsubono. Recursive contour-saliency blending network for accurate salient object detection. In WACV, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "text", + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "type": "text", + "content": "[24] Duo Li, Jie Hu, Changhu Wang, Xiangtai Li, Qi She, Lei Zhu, Tong Zhang, and Qifeng Chen. Involution: Inverting the inheritance of convolution for visual recognition. In CVPR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "text", + "content": "[25] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 381, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 404 + ], + "type": "text", + "content": "[26] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In AAAI, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 406, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 437 + ], + "type": "text", + "content": "[27] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In ICCV, 2013. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "type": "text", + "content": "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 474, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 496 + ], + "type": "text", + "content": "[29] Ming Liang and Xiaolin Hu. Recurrent convolutional neural network for object recognition. In CVPR, 2015. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 530 + ], + "type": "text", + "content": "[30] Zijian Liang, Pengjie Wang, Ke Xu, Pingping Zhang, and Rynson WH Lau. Weakly-supervised salient object detection on light fields. IEEE TIP, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[32] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In AAAI, 2022. 1, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "text", + "content": "[33] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[34] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[35] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, 2018. 3" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2710" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[36] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[37] Nian Liu, Wangbo Zhao, Ling Shao, and Junwei Han. Scg: Saliency and contour guided salient instance segmentation. IEEE TIP, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "text", + "content": "[38] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE TPAMI, 2010. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "text", + "content": "[39] Yudong Liu, Yongtao Wang, Siwei Wang, TingTing Liang, Qijie Zhao, Zhi Tang, and Haibin Ling. Cbnet: A novel composite backbone network architecture for object detection. In AAAI, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "text", + "content": "[40] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE TIP, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "text", + "content": "[41] Zhiming Luo, Akshaya Mishra, Andrew Achkar, Justin Eichel, Shaozi Li, and Pierre-Marc Jodoin. Non-local deep features for salient object detection. In CVPR, 2017. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "text", + "content": "[42] Ningning Ma, Xiangyu Zhang, Ming Liu, and Jian Sun. Activate or not: Learning customized activation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8042, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[43] S Mahdi H Miangoleh, Zoya Bylinskii, Eric Kee, Eli Shechtman, and Yaqiz Aksoy. Realistic saliency guided image enhancement. In CVPR, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 349, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 393 + ], + "type": "text", + "content": "[44] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. NeurIPS, 2019. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[45] Jialun Pei, He Tang, Chao Liu, and Chuanbo Chen. Salient instance segmentation via subitizing and clustering. Neurocomputing, 2020. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[46] Jialun Pei, Tianyang Cheng, He Tang, and Chuanbo Chen. Transformer-based efficient salient instance segmentation networks with orientative query. IEEE TMM, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[47] Jialun Pei, Tao Jiang, He Tang, Nian Liu, Yueming Jin, Deng-Ping Fan, and Pheng-Ann Heng. Calibnet: Dual-branch cross-modal calibration for rgb-d salient instance segmentation. arXiv preprint arXiv:2307.08098, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "text", + "content": "[48] Guohao Peng, Jun Zhang, Heshan Li, and Danwei Wang. Attentional pyramid pooling of salient visual residuals for place recognition. In ICCV, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[49] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, 2012. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[50] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In ICCV, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[51] Pedro OO Pinheiro, Ronan Collobert, and Piotr Dólar. Learning to segment object candidates. In NeurIPS, 2015. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[52] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[53] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[54] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE TPAMI, 2000. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[55] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPR Workshop, 2022. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "text", + "content": "[56] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "text", + "content": "[57] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecý, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In CVPR, 2023. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[58] Avishek Siris, Jianbo Jiao, Gary KL Tam, Xianghua Xie, and Rynson WH Lau. Scene context-aware salient object detection. In ICCV, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "text", + "content": "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "text", + "content": "[60] Chang Tang, Xinzhong Zhu, Xinwang Liu, and Pichao Wang. Salient object detection via recurrently aggregating spatial attention weighted cross-level deep features. In ICME, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 336, + 545, + 369 + ], + "type": "text", + "content": "[61] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Weakly-supervised salient instance detection. In BMVC, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "type": "text", + "content": "[62] Xin Tian, Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Learning to detect instance-level salient objects using complementary image labels. IJCV, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 449 + ], + "type": "text", + "content": "[63] Huy V Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In CVPR, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 451, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 451, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 451, + 545, + 483 + ], + "type": "text", + "content": "[64] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In ECCV, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "text", + "content": "[65] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "type": "text", + "content": "[66] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang, and Xiang Ruan. Saliency detection with recurrent fully convolutional networks. In ECCV, 2016. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "[67] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 633 + ], + "type": "text", + "content": "[68] Xiang Wang, Huimin Ma, Xiaozhi Chen, and Shaodi You. Edge preserving and multi-scale contextual neural network for salient object detection. TIP, 2017. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[69] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In CVPR, 2022. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[70] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2711" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[71] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufreydaz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv:2209.00383, 2022. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[72] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 196 + ], + "type": "text", + "content": "[73] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 242 + ], + "type": "text", + "content": "[74] Yu-Huan Wu, Yun Liu, Le Zhang, Wang Gao, and Ming-Ming Cheng. Regularized densely-connected pyramid network for salient instance segmentation. IEEE TIP, 2021. 1, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "text", + "content": "[75] Pengxiang Yan, Ziyi Wu, Mengmeng Liu, Kun Zeng, Liang Lin, and Guanbin Li. Unsupervised domain adaptive salient object detection through uncertainty-aware pseudolabel learning. In AAAI, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 289, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 323 + ], + "type": "text", + "content": "[76] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 324, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 357 + ], + "type": "text", + "content": "[77] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In AAAI, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "type": "text", + "content": "[78] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "type": "text", + "content": "[79] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In ICCV, 2017. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 471 + ], + "type": "text", + "content": "[80] Jianming Zhang, Stan Sclaroff, Zhe Lin, Xiaohui Shen, Brian Price, and Radomir Mech. Unconstrained salient object detection via proposal subset optimization. In CVPR, 2016. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "type": "text", + "content": "[81] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In CVPR, 2018. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "type": "text", + "content": "[82] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In CVPR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "type": "text", + "content": "[83] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "text", + "content": "[84] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 611, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 633 + ], + "type": "text", + "content": "[85] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[86] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In CVPR, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[87] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE TCSVT, 2022. 1, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 241 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[88] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[89] Huajun Zhou, Bo Qiao, Lingxiao Yang, Jianhuang Lai, and Xiaohua Xie. Texture-guided saliency distilling for unsupervised salient object detection. In CVPR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[90] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In CVPR, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[91] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In ICCV, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[92] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In CVPR, 2014. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2712" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_content_list.json b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..33112a0026fb166e1c14024cd743f0dbe238b660 --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_content_list.json @@ -0,0 +1,1307 @@ +[ + { + "type": "text", + "text": "Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling", + "text_level": 1, + "bbox": [ + 189, + 130, + 781, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leon Sick Ulm University", + "bbox": [ + 155, + 203, + 281, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dominik Engel Ulm University", + "bbox": [ + 328, + 203, + 452, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pedro Hermosilla TU Vienna", + "bbox": [ + 500, + 203, + 640, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Timo Ropinski \nUlm University", + "bbox": [ + 687, + 203, + 812, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditionally, training neural networks to perform semantic segmentation requires expensive human-made annotations. But more recently, advances in the field of unsupervised learning have made significant progress on this issue and towards closing the gap to supervised algorithms. To achieve this, semantic knowledge is distilled by learning to correlate randomly sampled features from images across an entire dataset. In this work, we build upon these advances by incorporating information about the structure of the scene into the training process through the use of depth information. We achieve this by (1) learning depth-feature correlation by spatially correlating the feature maps with the depth maps to induce knowledge about the structure of the scene and (2) exploiting farthest-point sampling to more effectively select relevant features by utilizing 3D sampling techniques on depth information of the scene. Finally, we demonstrate the effectiveness of our technical contributions through extensive experimentation and present significant improvements in performance across multiple benchmark datasets.", + "bbox": [ + 76, + 324, + 473, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 661, + 209, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation plays a critical role in many of today's vision systems in a multitude of domains. These include, among others, autonomous driving [15], medical applications [33, 39], and many more [9, 27, 41, 45]. Until recently, the main body of research in this area was focused on supervised models that require a large amount pixel-level annotations for training. Not only is sourcing this image data often a labor intensive process, but also annotating the large datasets required for good performance comes at a high price. Several benchmark datasets report their annotation times. For example, the MS COCO dataset [28] required more than $28\\mathrm{K}$ hours of human annotation for around $164\\mathrm{K}$ images, and annotating a single image in the Cityscapes dataset [11] took 1.5 hours on average.", + "bbox": [ + 76, + 688, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These costs have triggered the advent of unsupervised semantic segmentation [10, 16, 20, 35], which aims to remove the need for labeled training data in order to train segmentation models. Recently, work by Hamilton et al. [16] has accelerated the progress towards removing the need for labels to achieve good results on semantic segmentation tasks. Their model, STEGO, uses a DINO-pretrained [7] Vision Transformer (ViT) [13] to extract features that are then distilled across the entire dataset to learn semantically relevant features, using a contrastive learning approach. The to-be-distilled features are sampled randomly from feature maps produced from the same image, k-NN matched images as well as other negative images. Seong et al. [35] build on this process by trying to identify features that are most relevant to the model by discovering hidden positives. Their work exposes an inefficiency of random sampling in STEGO as hidden positives sampling leads to significant improvements. However, both approaches only operate on the pixel space and therefore fail to take into account the spatial layout of the scene. Not only do we humans perceive the world in 3D, but also previous work [5, 8, 18, 37, 38, 44] has shown that supervised semantic segmentation can benefit greatly from spatial information during training. Inspired by these observations, we propose to incorporate spatial information in the form of depth maps into the STEGO training process. Depth is considered a product of vision and does not provide a labeled training signal. To obtain depth information for the benchmark image datasets in our evaluations, we make use of an off-the-shelf zero-shot monocular depth estimator to obtain spatial information of the scene. This allows us to incorporate the depth information without the need for human annotations or sensor ground truth.", + "bbox": [ + 496, + 291, + 893, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With our method, $DepthG$ , we propose to (1) guide the model to learn a rough spatial layout of the scene, since we hypothesize this will aid the network in differentiating objects much better. We achieve this by extending the contrastive process to the spatial dimension: We do not limit the model to learning only Feature-Feature Correlations, but also Depth-Feature Correlations. Through this process, the model is guided towards pulling apart the features with high", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3637", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40bcafca6e9638f4a9deefdaae77ca1abd41bf8d4c9de28e7095e01c1291f39b.jpg", + "image_caption": [ + "Figure 1. Guiding the feature space for unsupervised segmentation with depth information. Our intuition behind the proposed approach is simple: For locations in the 3D space with a low distance, we guide the model to map their features closer together. Vice versa, the features are learned to be drawn apart in feature space if their distance in the metric space is large." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distances in 3D space, as well as mapping them closer together if their distance is low depth space. Figure 1 visualizes this process.", + "bbox": [ + 75, + 554, + 468, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the information about the spatial layout of the scene present, we furthermore propose to (2) spatially inform our feature sampling process by utilizing Farthest-Point Sampling (FPS) [14, 31] on the depth map, which equally samples scenes in 3D. We show these techniques in combination make our approach to unsupervised segmentation highly effective, since in our evaluations across multiple benchmark datasets, we demonstrate state-of-the-art performance. We include a short video presenting our method as part of the supplementary materials. To the best of our knowledge, we are the first to propose a mechanism to incorporate 3D knowledge of the scene into unsupervised learning for 2D images without encoding depth maps as part of the network input. This design aspect of our method alleviates the risk of the model developing an input dependency. Therefore, our approach does not rely on depth information during inference and is not affected by availability or quality of such depth information.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unsupervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 114, + 825, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [10, 16, 20, 35] have attempted to tackle semantic segmentation without the use of human annotations. Ji et al. [20] propose IIC, a method that aims to maximize the mutual information between different augmented versions of an image. PiCIE, published by Cho et al. [10], introduces an inductive bias based on the invariance to photometric transformations and equivariance to geometric manipulations. DINO [7] often serves as a critical component to unsupervised segmentation algorithms, since the self-supervised pre-trained ViT can produce semantically relevant features. Recent work by Seitzer et al. [34] builds upon this ability by training a model with slot attention [30] to reconstruct the feature maps produced by DINO from the different slots. The features of their object-centric model are clustered with k-means [29] where each slot is associated with a cluster. In their 2021 work, Hamilton et al. [16] have also built upon DINO features by introducing a feature distillation process with features from the same image, k-NN retrieved examples as well as random other images from the dataset. Their learned representations are finally clustered and refined with a CRF [22] for semantic segmentation. While STEGO's feature selection process is random, Seong et al. [35] introduce a more effective sampling strategy by discovering hidden positives. During training, they form task-agnostic and task-specific feature pools. For an anchor feature, they then compute the maximum similarity to any of the pool features and sample locations in the image with greater similarity than the determined value. A more detailed introduction to STEGO is provided in Section 3.1.", + "bbox": [ + 496, + 138, + 893, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Depth For Semantic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 589, + 802, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous research [5, 17, 18, 38, 40, 43] has sought to incorporate depth for semantic segmentation in different settings. Wang et al. [38] propose to use depth for adapting segmentation models to new data domains. Their method adds depth estimation as an auxiliary task to strengthen the prediction of segmentation tasks. Furthermore, they approximate the pixel-wise adaption difficulty from source to target domain through the use of depth decoders. Work by Hoyer et al. [18] explores three further strategies of how depth can be useful for segmentation. First, they propose using a shared backbone to share learning features for segmentation and self-supervised depth estimation, similar to Wang et al. [38]. Second, they use depth maps to introduce a data augmentation that is informed by the structure of the scene. And lastly, they detail the integration of depth into an active learning loop as part of a student-teacher setup. Another work by Hou et al. [17] incorporates depth information into a pre-training algorithm with the aim to learn better representations for semantic segmentation. In their", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3638", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work, Mask3D, they propose to incorporate depth in the form of a 3D prior by formulating a reconstruction task that operates on masked RGB and depth patches, enabling them to learn more useful features for semantic segmentation.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 167, + 166, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we detail our proposed method for guiding unsupervised segmentation with depth information. An overview of our approach is presented in Figure 2.", + "bbox": [ + 76, + 193, + 468, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 250, + 205, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach builds upon work by Hamilton et al. [16]. In their work, each image is 5-cropped and k-NN correspondences between these images are calculated using the DINO ViT [7]. Generally, STEGO uses a feature extractor $\\mathcal{F}:\\mathbb{R}^{3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}\\to \\mathbb{R}^{C\\times H\\times W}$ with input image height $H_{\\mathrm{in}}$ and width $W_{\\mathrm{in}}$ , to calculate a feature map $f\\in \\mathbb{R}^{C\\times H\\times W}$ with height $H$ , width $W$ and feature dimension $C$ from the input image. These features are then further encoded by a segmentation head $\\mathcal{S}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{Q\\times H\\times W}$ to calculate the code space $s\\in \\mathbb{R}^{Q\\times H\\times W}$ with code dimension $Q$ . With the goal of forming compact clusters and amplifying the correlation of the learned features, let $f_{i}\\coloneqq \\mathcal{F}(x_{i})$ and $f_{j}\\coloneqq \\mathcal{F}(x_{j})$ be feature maps for a given input pair of $x_{i}$ and $x_{j}$ , which are then used to calculate $s_i\\coloneqq S(f_i)$ and $s_j\\coloneqq S(f_j)$ from the segmentation head $S$ . In practice, STEGO samples $N^2$ vectors from the feature map during training. Hamilton et al. [16] introduce the concept of constructing the feature correspondence tensor $\\pmb {F}\\in \\mathbb{R}^{H\\times W\\times H\\times W}$ as follows:", + "bbox": [ + 75, + 273, + 472, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {h w, u v} = \\frac {f _ {i} ^ {h w} \\cdot f _ {j} ^ {u v}}{\\| f _ {i} ^ {h w} \\| \\| f _ {j} ^ {u v} \\|} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 570, + 468, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\cdot$ denotes the dot product. Using the same formula we obtain $S$ using $s_i, s_j$ . Consequently, the feature correlation loss is defined as:", + "bbox": [ + 75, + 619, + 468, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {C o r r}} := - \\sum_ {h w, u v} \\left(\\boldsymbol {F} _ {h w, u v} - b\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 675, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b$ is a scalar bias hyperparameter. Empirical evaluations from STEGO have shown that applying spatial centering to the feature correlation loss along with zero-clamping further improves performance [16]. These correlations are calculated for two crops from the same image $(\\mathcal{L}_{\\mathrm{self}})$ and one from a different but similar image, determined by the k-NN correspondence pre-processing $(\\mathcal{L}_{\\mathrm{knn}})$ . Finally, negative images are sampled randomly $(\\mathcal{L}_{\\mathrm{random}})$ . The final loss is a weighted sum of the different losses where each of them has their individual weight $\\lambda_{i}$ :", + "bbox": [ + 75, + 720, + 468, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S T E G O}} = \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k n n}} \\mathcal {L} _ {\\text {k n n}} + \\lambda_ {\\text {r a n d o m}} \\mathcal {L} _ {\\text {r a n d o m}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 885, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After training, the inferred feature maps for a test image are clustered using k-means and refined with a conditional random field (CRF) [22].", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Depth Map Generation", + "text_level": 1, + "bbox": [ + 500, + 143, + 714, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since in many cases, depth information about the scene is not readily available, we make use of recent progress in the field of monocular depth estimation [1-3, 25, 32] to obtain depth maps from RGB images. Recently, methods from this field have made significant advances in zero-shot depth estimation, i.e. predicting depth values for scenes from data domains not seen during training. This property makes them especially suitable for our method, since it enables us to obtain high-quality depth predictions for a wide variety of data domains without ever re-training the depth network. This property also limits the computational cost for our method. For our method, we experiment with different state-of-the-art monocular depth estimators, detailed in Section 5, and found ZoeDepth [3] to perform best in our evaluations. Given a cropped RGB image $x_{i}$ , we use the monocular depth estimator $\\mathcal{M}$ together with average pooling to predict depth $d_{i}\\in [0,1]^{H\\times W}$ at feature resolution:", + "bbox": [ + 496, + 167, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} = \\operatorname {p o o l} \\left(\\mathcal {M} \\left(x _ {i}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 431, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The average pooling operation is used to match the dimensions of the feature map, which is required to sample nonoverlapping locations at the patch resolution.", + "bbox": [ + 496, + 454, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Depth-Feature Correlation Loss", + "text_level": 1, + "bbox": [ + 500, + 507, + 782, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With our Depth-Feature Correlation loss, we aim to enforce spatial consistency in the feature map by transferring the distances from the depth map to the latent feature space.", + "bbox": [ + 496, + 530, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrastive learning, the network is incentivized to decrease the distance in feature space for similar instances, therefore learning to map their latent representations closer together. Likewise, different instances are drawn further apart in feature distance. We assume the same concept to be true in 3D space: The spatial distance between two points from the same depth plateau is smaller, while the distance between a point in the foreground and one in the background is larger. Since, in both spaces, the concept of measuring difference is represented by the distance between two points, we propose to align them through our concept of Depth-Feature Correlation: For large distances in the 3D space, we encourage the network to produce vectors that are further apart, and vice versa. With this, we induce the model with knowledge about the spatial structure of the scene, enabling it to better differentiate between objects. To achieve this we construct the depth correspondence tensor similar to the feature correspondence from Equation 1. The depth correspondence tensor $D$ is computed from the depths of two different image crops as follows:", + "bbox": [ + 496, + 575, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {D} _ {h w, u v} = d _ {i} ^ {h w} d _ {j} ^ {u v}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3639", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4814db22c9dc63ad43748ca08fbc5988ddb67f17e0bed4334757c1e413bd4902.jpg", + "image_caption": [ + "Figure 2. Overview of the DepthG training process. After 5-cropping the image, each crop is encoded by the DINO-pretrained ViT $\\mathcal{F}$ to output a feature map. Using farthest point sampling (FPS), we sample the 3D space equally and convert the coordinates to select samples in the feature map. The sampled features are further transformed by the segmentation head $S$ . For both feature maps, the correlation tensor is computed. Following, we sample the depth map at the coordinates obtained by FPS and compute a correlation tensor in the same fashion. Finally, we compute our Depth-Feature Correlation loss and combine it with the feature distillation loss from STEGO. We guide the model to learn depth-feature correlation for crops of the same image, while the feature distillation loss is also applied to k-NN-selected and random images." + ], + "image_footnote": [ + "$D_{hw,uv}$" + ], + "bbox": [ + 80, + 85, + 545, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(h,w)$ and $(u,v)$ represent the pixel positions in the depth maps $d_{i}$ and $d_{j}$ respectively. Together with the zero clamping, our Depth-Feature Correlation loss is defined as:", + "bbox": [ + 75, + 436, + 470, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D e p t h G}} = - \\sum_ {h w, u v} \\left(\\boldsymbol {D} _ {h w, u v} - b _ {\\text {D e p t h G}}\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 493, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{hw,uv}$ represents the depth correlation tensor, $b_{\\mathrm{DepthG}}$ is the bias for our loss term, and $S_{hw,uv}$ represents the feature correlation tensor computed from the output features of the segmentation head $\\mathcal{S}$ . By also using zero-clamping, we limit erroneous learning signals that aim to draw apart instances of the same class if they have large spatial differences. With this, we extend the STEGO loss so it can be formulated as follows:", + "bbox": [ + 75, + 540, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S T E G O + D e p t h G}} = \\mathcal {L} _ {\\text {S T E G O}} + \\lambda_ {\\text {D e p t h G}} \\mathcal {L} _ {\\text {D e p t h G}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 672, + 468, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with Depth-Feature Correlation weight $\\lambda_{\\mathrm{DepthG}}$ . By inducing depth knowledge during training without encoding the depth maps as part of the model input, our model can predict spatially informed segmentations on RGB images with its distilled knowledge and does not rely on depth to be available at test time.", + "bbox": [ + 75, + 700, + 468, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Depth-Guided Feature Sampling", + "text_level": 1, + "bbox": [ + 76, + 801, + 366, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also aim to make the feature sampling process informed by the spatial layout of the scene. To perform sampling in the 3D space, we transform the downsampled depth map $d(x_{i})$ into a point cloud with points $\\{p_1,p_2,\\dots ,p_n\\}$ . On this point cloud, we apply farthest point sampling", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(FPS) [14], in an iterative fashion by always selecting the next point $p_k$ as the point with the maximum distance in 3D space with respect to the already sampled points $\\{p_1, p_2, \\dots, p_{k-1}\\}$ . After having sampled $N^2$ points, we end up with a set of samples $\\{p_1, p_2, \\dots, p_{N^2}\\}$ which are consequently converted to 2D sampling indices for the feature maps $f$ and $g$ . In contrast to the data-agnostic random sampling applied in STEGO, our feature selection process takes into account the geometry of the input scene and covers the spatial structure more equally. In our ablations in Section 5, we show this scene coverage from FPS of the depth space further increases the effectiveness of our Depth Feature Correlation loss, due to the increased diversity in selected 3D locations. We show a visual comparison between random and farthest point sampling in Figure 5.", + "bbox": [ + 496, + 436, + 893, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Guidance Scheduling", + "text_level": 1, + "bbox": [ + 500, + 678, + 700, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our Depth-Feature Correlation loss is effective at enriching the model's learning process with spatial information of the scene, we aim to alleviate the danger of it interfering with the learning of feature correlations during model training. We hypothesize that our model most greatly benefits from depth information in the beginning of training when its only knowledge is encoded in the features maps by the frozen ViT backbone. To give it a head start, we increase the weight of our Depth-Feature Correlation loss in the beginning and gradually decrease its influence during training. Vice versa, the distillation process in the feature space will increasingly emphasised as the training progresses. In this way, the network builds upon the already learned rough spa", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3640", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2489f9a943cdf638050e6fdad377dc99fbf9e6cff4f93949ca8f4551da53e33.jpg", + "image_caption": [ + "Figure 3. Local Hidden Positives. We visualize the use of depth and attention maps for local hidden positives. For this visualization, we sample the respective propagation maps at the yellow patch in the center of the crops. We observe the depth map to have sharper borders and more consistent propagation values. We experiment with both propagation strategies in Section 5." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tial structure of the scene achieved through our depth guidance. Therefore, we implement an exponential decay of the weight $\\lambda_{\\mathrm{DepthG}}$ and bias $b_{\\mathrm{DepthG}}$ of our loss component. We ablate on the use of guidance scheduling in the appendix.", + "bbox": [ + 75, + 342, + 470, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Local Hidden Positives In 3D Space", + "text_level": 1, + "bbox": [ + 76, + 414, + 385, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further explore the combination of our method with Hidden Positives [35], which also builds upon STEGO. As we demonstrate as part of our ablation on the individual influence of our contributions in Section 5, our feature sampling method, implemented through farthest point sampling, is integral to the effectiveness of DepthG. Therefore, we decide not to replace it with the global hidden positives sampling from Seong et al. [35]. Instead, we implement a depth-informed variant of local hidden positives, 3D-LHP, where the loss for an individual patch is propagated to eight neighboring patches proportionally to their attention values obtained from the feature extractor. We modify this strategy by instead propagating the learning signal to the closest patches in 3D space proportional to their relative distances, using the point cloud described in Section 3.4. As visualized in Figure 3, we observe that our depth-based propagation map has sharper and more consistent surfaces. To propagate the learning signal to the selected patches, we follow Hidden Positives and mix the features coming from the segmentation head $S$ in proportion to their propagation values (point distances). The mixed representations are then fed through an additional projection head $\\mathcal{P}$ . We then calculate $\\mathcal{L}_{\\mathrm{STEGO+DepthG}}$ again for the produced output features and combine it with the loss from the segmentation head output.", + "bbox": [ + 75, + 438, + 472, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 816, + 209, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Models. We conduct experiments on the COCO-Stuff [4], Cityscapes [11] and Potsdam-3 datasets. COCO-Stuff contains a wide variety of real-world scenes. In our evaluation, we follow [16, 20, 35] and provide re", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sults on the coarse class split, COCO-Stuff 27. In contrast, Cityscapes contains traffic scenes from 50 cities from a driver-like viewpoint. Lastly, the Potsdam-3 dataset is composed of aerial, top-down images of the city of Potsdam. We use the DINO [7] backbones ViT-Small (ViT-S) and ViT-Base (ViT-B), which were pre-trained in a self-supervised manner on ImageNet-1k [12]. We choose the models with a patch size of $8 \\times 8$ , since they have been shown to perform best for semantic segmentation due to the higher resolution of the resulting feature maps [16, 35]. In the result tables, we refer to DepthG models trained with our DepthFeature Correlation loss and FPS as Ours. Models that utilize 3D-LHP propagation with depth maps in addition are displayed as Ours w/ 3D-LHP. We compare our approach against competing methods which were never trained with human annotations, i.e. human labels or language supervision, neither in the feature extractor nor the segmentation training.", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. Similar to STEGO and related work [16, 35], we evaluate our models in the unsupervised, clustering-based setting, as well as linear probing. Since the output of our model is a pixel-level map of features and not class labels, these features are consequently clustered. Following, the pseudo-labeled clusters are aligned with the ground truth labels through Hungarian matching [23, 24] across the entire validation dataset. To perform linear probing, an additional linear layer is added on top of the model and trained with cross-entropy loss to learn classification of the features.", + "bbox": [ + 496, + 398, + 893, + 566 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/86303acc50b22f95b1a7afb0cb943bc2bc074cf870c395bdf9ab47569c3b3f59.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingUnsupervisedLinear
MethodModelAcc.mIoUAcc.mIoU
IIC [20]R18+FPN21.86.744.58.4
PiCIE [10]R18+FPN48.113.854.213.9
PiCIE+H [10]R18+FPN50.014.454.814.8
DINO [7]ViT-S/828.711.368.633.9
ACSeg [26]ViT-S/816.4---
TransFGU [42]ViT-S/817.552.7--
STEGO + HP [35]ViT-S/857.224.675.642.7
STEGO [16]ViT-S/848.324.574.438.3
+ OursViT-S/856.325.673.738.9
+ Ours w/ 3D-LHPViT-S/855.126.773.937.8
DINO [7,16]ViT-B/830.59.666.829.4
DINOSAUR [34]*ViT-B/844.924.0--
STEGO [16]ViT-B/856.928.276.141.0
+ OursViT-B/858.629.075.541.6
", + "bbox": [ + 503, + 580, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Evaluation on COCO-Stuff 27. We report results on COCO-Stuff with 27 high-level classes. Overall, our method outperforms STEGO and HP on unsupervised segmentation with the ViT-B/8, while showing competitive performance for the ViT-S/8. *Results obtained without post-processing optimization.", + "bbox": [ + 496, + 818, + 893, + 891 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3641", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. COCO-Stuff", + "text_level": 1, + "bbox": [ + 76, + 90, + 215, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present our evaluation on COCO-Stuff 27 in Table 1. For the ViT-S/8, our experiments show that Ours is able to improve upon STEGO in most metrics, with improved unsupervised accuracy by $+8.0\\%$ and unsupervised mIoU increased by $+1.1\\%$ . Ours w/ 3D-LHP further increases this mIoU delta to $+1.8\\%$ , highlighting the effectiveness of our 3D-information propagation strategy in combination with DepthG. When comparing our approach to Hidden Positives, a method with more computational overhead, for the ViT-S/8, we show competitive performance for unsupervised accuracy and outperform their approach by $+1.0\\%$ on unsupervised mIoU with Ours and $+1.7\\%$ with Ours w/ 3D-LHP. When using the DINO ViT-B/8 encoder, our approach again outperforms STEGO, as well as all other presented methods on unsupervised metrics. Most notably, we are able to increase the unsupervised mIoU by $+0.8\\%$ .", + "bbox": [ + 75, + 113, + 472, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Cityscapes", + "text_level": 1, + "bbox": [ + 76, + 367, + 197, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate our approach on the Cityscapes dataset [11], consisting of various scenes from 50 different cities. We follow the training setting from STEGO and, contrary to all other datasets, do not sample point-wise but use the full feature map for our learning process along with the full depth map. As can be seen in Table 2, our method significantly outperforms STEGO as well as Hidden Positives. For unsupervised mIoU, while Hidden Positives decreased performance compared to STEGO, we observe that our approach to achieves a $+2.1\\%$ increase. Similarly, we report state-of-the-art performance in accuracy, building upon Hidden Positives' already impressive improvements over STEGO and outperforming it by $+2.1\\%$ .", + "bbox": [ + 75, + 391, + 472, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Potsdam", + "text_level": 1, + "bbox": [ + 76, + 599, + 181, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is further evaluated on the Potsdam-3 dataset, containing aerial images of the German city of Potsdam. Contrary to the other benchmarks, which contain images in a first-person perspective, Potsdam-3 contains only birds-eye-view images, a perspective that is considered out-of-distribution for the monocular depth estimator. Despite this inherent limitation of our approach for aerial data, we are able to demonstrate relatively commendable performance in Table 3 by improving STEGO's performance and reporting state-of-the-art performance for the ViT-S backbone. In contrast, Hidden Positives [35] use a ViT-B/8, and with roughly twice the parameters as ours, reach an accuracy of $82.4\\%$ . We present a visual overview of the predicted Potsdam depth maps in the appendix.", + "bbox": [ + 75, + 623, + 472, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 261, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present qualitative results of our method in Figure 4 and compare with segmentation maps from STEGO. On", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6b3fe1628976c2526f1cd33b4f58949724bd637e4c2190c3cc6de27d16e5b242.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModelU. AccU. mIoU
IIC [20]R18+FPN47.96.4
PiCIE [10]R18+FPN65.612.3
DINO [7]ViT-B/843.611.8
STEGO + HP [35]ViT-B/879.518.4
STEGO [16]ViT-B/873.221.0
+ OursViT-B/881.623.1
", + "bbox": [ + 509, + 88, + 883, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b742b0cb1400e48a1963120a8e8d9702289c8042ff57bd2d89fd8f6ccda5cc6c.jpg", + "table_caption": [ + "Table 2. Results on Cityscapes. We report unsupervised accuracy and mIoU on Cityscapes. Our method outperforms both STEGO variants by substantial margins. Notably, our method is the first to improve upon unsupervised mIoU." + ], + "table_footnote": [], + "table_body": "
MethodModelU. Acc.
CC [19]VGG1163.9
DeepCluster [6]VGG1141.7
IIC [20]VGG1165.1
DINO [7,21]ViT-S/871.3
STEGO [16,21]ViT-S/877.0
+ OursViT-S/880.4
+ Ours w/ 3D-LHPViT-S/880.4
", + "bbox": [ + 542, + 294, + 849, + 439 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b44384b9bf35062d5580ca68aeae15e70f520dd7e875d6fd932fe99ee6b79657.jpg", + "table_caption": [ + "Table 3. Results on Potsdam. We report unsupervised accuracy on the Potsdam dataset. Our method is able to improve upon STEGO. We hypothesize that with a zero-shot depth estimator more suitable for aerial images, the results for our method could further improve." + ], + "table_footnote": [], + "table_body": "
MethodU. mIoU.U. Acc
STEGO [16]24.548.3
+ Depth-Feature Correlation (1)24.751.2
+ FPS (2)24.649.1
+ 3D-LHP (3)25.248.5
+ Ours (1, 2)25.656.3
+ Ours w/ 3D-LHP (1, 2, 3)26.755.1
", + "bbox": [ + 501, + 527, + 893, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Effect of our contributions. We compare our individual contributions and the combination of all contributions.", + "bbox": [ + 498, + 667, + 890, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multiple occasions, our depth guidance reduces erroneous predictions from the model caused by visual irritations in the pixel space. In the example of the boy with the baseball bat in Figure 4a, false classifications from STEGO are caused by shadows on the ground. Our model is able to correct this. Furthermore, it goes beyond the noisy label and also correctly classifies the glimpse of a plant that can be seen through a hole in the background. This is an indication that our model does not overfit to the depth map, since this visual cue is only observable in the pixel space, but not the depth map.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3642", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59e641edfed74e60fae016a3268151762a0cad18f01b708fda77b3a1348329b6.jpg", + "image_caption": [ + "(a) COCO-Stuff" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 478, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/652e6f12fc2a495b4f2261e12088cab65129cb38427224d8bcb0e4c49ee744fe.jpg", + "image_caption": [ + "(b) Cityscapes", + "Figure 4. Qualitative results. We show qualitative differences for plain STEGO compared to STEGO with our depth guidance, using ViT-S models for COCO and ViT-B for Cityscapes. Where STEGO struggles to differentiate instances, our model is able to correct this and successfully separates them for segmentation. In the case of the building in (a), our method alleviates visual irritations from the pixel space and corrects the segmentation of the building. In (b), our model is able to better handle visual inconsistencies from shadows." + ], + "image_footnote": [], + "bbox": [ + 493, + 85, + 893, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablations", + "text_level": 1, + "bbox": [ + 76, + 448, + 183, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Individual Influence. We investigate the effect of our technical contributions on training our model with a ViT-S/8 backbone on COCO-Stuff 27. Our observations in Table 4 show that our Depth-Feature Correlation loss itself already improves the performance of STEGO. This improvement is further increased through the use of FPS, which enables us to sample the depth space more meaningfully and therefore encourages more diversity in the depth correlation tensor $D_{hw,uv}$ . Intuitively, this sampling diversity significantly amplifies our Depth-Feature Correlation for aligning the feature space with the depth space. We provide a visual comparison to random sampling in Figure 5 and additional illustrations in the appendix. FPS retrieves more diverse locations and specifically selects locations with depth discontinuities. This naturally benefits the Depth-Feature Correlation to learn sharper edges in this area for the output segmentation. Adding local hidden positives with depth maps further increases the unsupervised mIoU, while slightly lowering the accuracy.", + "bbox": [ + 75, + 473, + 472, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Source Of Depth Maps. We investigate the effect of different monocular depth estimators to generate the depth maps used to train our model. In our experiments, we consider three options: The previously mentioned ZoeDepth [3], Kick Back & Relax (KBR) [36] which uses self-supervision to learn depth from Slow-TV videos, as well as MiDaS [32], the base model to ZoeDepth. Empirically, ZoeDepth produces the most accurate zero-shot", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "monodepth results across indoor and outdoor datasets, followed by MiDaS and KBR [3,36]. We provide a qualitative comparison in the appendix. To evaluate the influence of the depth map quality on the performance of our model, we first generate depth maps for COCO-Stuff 27 using each of the introduced models. We then train our model with a ViT-S/8 backbone and report the results in Table 5. We observe that depth maps from ZoeDepth work best with our method, while the model trained with MiDaS depth has an edge over the KBR counterpart.", + "bbox": [ + 496, + 449, + 893, + 602 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4f3c4efd1e84904ac46fe56ea15f26a851f17e279e0d2b0c10efd522e629a07.jpg", + "image_caption": [ + "Figure 5. Random vs. Farthest Point Sampling. We observe that random sampling can miss entire structures like trees in the first top and the plane in the bottom row. In contrast, our method meaningfully samples the depth space and selects locations across the different structures and at depth edges. We show further illustrations of FPS in the appendix." + ], + "image_footnote": [], + "bbox": [ + 500, + 617, + 890, + 795 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3643", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d330e6a3c0ef13c3e80d3f32b1e61f8fa42269c0f7e0760153deb52c1c5d1c8f.jpg", + "image_caption": [ + "Figure 6. Failure cases. We show cases where our model fails to correctly segment and classify the scene. The top row is a prime example where the difference in depth is correctly distilled, though the model fails to correctly classify the snow region." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Signal Propagation With Local Hidden Positives. We ablate the implementation of LHP along with different propagation strategies. As described in Section 3.6, our method takes advantage of the depth information of the scene to propagate the learning signal to patches which are nearby in 3D space. We further implement utilizing the attention map from the DINO backbone and propagate proportionally to their values i.e., the approach utilized in Hidden Positives [35]. While, in Table 6, we show that our approach benefits from signal propagation with 3D-LHP, applying LHP with Attention leads to lower performance. We speculate 3D-LHP is more suitable for our approach since, for a given location, the signal from our Depth-Feature Correlation loss is calculated w.r.t. the depth at this sample. Propagating to locations with the same depth does not corrupt this signal, though this can happen with LHP (Attention), since it does not consider depth for propagation.", + "bbox": [ + 75, + 368, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Cost. Our method only leads to an insignificant increase in runtime versus the baseline STEGO model, since we solely guide the loss as well as the feature sampling and only for Ours w/ 3D-LHP add an additional small segmentation head. In contrast, the competing method Hidden Positives [35] relies on a computationally more expensive process to select features and introduces an additional segmentation head to fill their task-specific feature pool. To keep our computational overhead low, we make use of a pre-trained monocular depth estimation network with impressive zero-shot capabilities. We consider task specific training of the depth estimator not a necessity, since the model has zero-shot capabilities that generalize well to different scenes and domains. Therefore, in our experiments on a diverse array of scenes, we do not re-train the depth estimator, and consider the additional computational cost for generating the depth maps to be negligible.", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6b7920799ce5b47643586229418c9f94f593c3c240f8aac2a7aea2df72c67db7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrained withU. Acc.U. mIoU
ZoeDepth [3]Sensor Depth56.325.6
MiDaS [32]Sensor Depth53.025.0
KBR [36]Self-Supervision50.623.1
", + "bbox": [ + 501, + 88, + 890, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fed6098ada020a7f52e33fefa9261dadda94ad1990863270907beecd9da6ce9d.jpg", + "table_caption": [ + "Table 5. Different depth map sources. We experiment with different monocular depth estimators which were trained with either sensor depth or self-supervision. Overall, the model trained with depth maps from ZoeDepth performs best on COCO-Stuff 27." + ], + "table_footnote": [], + "table_body": "
LHPPropagation StrategyU. Acc.U. mIoU
X-56.325.6
✓3D-LHP (Depth)55.126.7
✓LHP (Attention)52.624.5
", + "bbox": [ + 516, + 250, + 874, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. LHP Ablation. We compare the use of depth and attention for local hidden positives. We find that using depth improves unsupervised mIoU, while we find that using attention does not improve our method.", + "bbox": [ + 498, + 337, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 416, + 622, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we have demonstrated our method's effectiveness for many real-world cases, our method's applicability is limited in settings unsuitable for depth estimation, such as slices of CT scans and other medical data domains. Furthermore, the experiments on Potsdam-3 have shown, our method can improve unsupervised semantic segmentation despite suboptimal viewing perspectives for the monocular depth estimator. We assume the scenario of aerial images represents a rare case where our method would profit from a domain-specific monocular depth estimator. We also present failure cases of our model in Figure 6.", + "bbox": [ + 496, + 444, + 890, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion & Future Work", + "text_level": 1, + "bbox": [ + 500, + 631, + 754, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have presented a novel method to induce spatial knowledge of the scene into our model for unsupervised semantic segmentation. We have proposed to correlate the feature space with the depth space and use the 3D information to more meaningfully sample features in a spatially informed way. Furthermore, we have demonstrated that these contributions produce state-of-the-art performance on many real-world datasets and thus foster the progress in unsupervised segmentation. The applicability of our approach for other tasks is further to be explored, since we hypothesize it can be useful beyond unsupervised segmentation as part of other contrastive processes. We consider this to be a promising direction for future work. Furthermore, it remains to be investigated what kind of information could be useful in domains where depth is not an obviously meaningful signal, like medical data.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3644", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3", + "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3", + "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 7, 8", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5", + "[5] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1129-1139, 2022. 1, 2", + "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6", + "[8] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2624-2632, 2019. 1", + "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1", + "[10] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 2, 5, 6", + "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 5, 6", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1", + "[14] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 4", + "[15] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1", + "[16] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6", + "[17] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2", + "[18] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11130-11140, 2021. 1, 2", + "[19] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H Adelson. Learning visual groups from co-occurrences in space and time. arXiv preprint arXiv:1511.06811, 2015. 6", + "[20] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9865-9874, 2019. 1, 2, 5, 6", + "[21] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of stego for safe unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3788-3797, 2023. 6", + "[22] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011, 2, 3", + "[23] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[24] Harold W Kuhn. Variants of the hungarian method for assignment problems. Naval research logistics quarterly, 3(4):253-258, 1956. 5", + "[25] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 3", + "[26] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 5", + "[27] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1", + "[29] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2):129-137, 1982. 2", + "[30] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2", + "[32] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 3, 7, 8", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1", + "[34] M Seitzer, M Horn, A Zadaianchuk, D Zietlow, T Xiao, C Simon-Gabriel, T He, Z Zhang, B Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In International Conference on Learning Representations (ICLR), 2023. 2, 5", + "[35] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19540-19549, 2023. 1, 2, 5, 6, 8", + "[36] Jaime Spencer, Chris Russell, Simon Hadfield, and Richard Bowden. Kick back & relax: Learning to reconstruct the world by watching slowtv. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 7, 8", + "[37] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In Proceedings of the" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019. 1", + "[38] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8515-8525, 2021. 1, 2", + "[39] Risheng Wang, Tao Lei, Ruixia Cui, Bingtao Zhang, Hongying Meng, and Asoke K Nandi. Medical image segmentation using deep learning: A survey. IET Image Processing, 16(5):1243-1267, 2022. 1", + "[40] Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. Multimodal token fusion for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12186-12195, 2022. 2", + "[41] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1", + "[42] Zhaoyuan Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, and Rong Jin. Transfgu: a top-down approach to fine-grained unsupervised semantic segmentation. In European conference on computer vision, pages 73-89. Springer, 2022. 5", + "[43] Xiaowen Ying and Mooi Choo Chuah. Uctnet: Uncertainty-aware cross-modal transformer network for indoor rgb-d semantic segmentation. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2", + "[44] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 1", + "[45] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 681 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3646", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_model.json b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_model.json new file mode 100644 index 0000000000000000000000000000000000000000..26be9993f553909098985d436a54c180c633d729 --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_model.json @@ -0,0 +1,1837 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.782, + 0.177 + ], + "angle": 0, + "content": "Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.204, + 0.282, + 0.24 + ], + "angle": 0, + "content": "Leon Sick Ulm University" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.204, + 0.454, + 0.24 + ], + "angle": 0, + "content": "Dominik Engel Ulm University" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.204, + 0.641, + 0.238 + ], + "angle": 0, + "content": "Pedro Hermosilla TU Vienna" + }, + { + "type": "text", + "bbox": [ + 0.689, + 0.204, + 0.813, + 0.24 + ], + "angle": 0, + "content": "Timo Ropinski \nUlm University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.325, + 0.474, + 0.628 + ], + "angle": 0, + "content": "Traditionally, training neural networks to perform semantic segmentation requires expensive human-made annotations. But more recently, advances in the field of unsupervised learning have made significant progress on this issue and towards closing the gap to supervised algorithms. To achieve this, semantic knowledge is distilled by learning to correlate randomly sampled features from images across an entire dataset. In this work, we build upon these advances by incorporating information about the structure of the scene into the training process through the use of depth information. We achieve this by (1) learning depth-feature correlation by spatially correlating the feature maps with the depth maps to induce knowledge about the structure of the scene and (2) exploiting farthest-point sampling to more effectively select relevant features by utilizing 3D sampling techniques on depth information of the scene. Finally, we demonstrate the effectiveness of our technical contributions through extensive experimentation and present significant improvements in performance across multiple benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.662, + 0.21, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Semantic segmentation plays a critical role in many of today's vision systems in a multitude of domains. These include, among others, autonomous driving [15], medical applications [33, 39], and many more [9, 27, 41, 45]. Until recently, the main body of research in this area was focused on supervised models that require a large amount pixel-level annotations for training. Not only is sourcing this image data often a labor intensive process, but also annotating the large datasets required for good performance comes at a high price. Several benchmark datasets report their annotation times. For example, the MS COCO dataset [28] required more than \\(28\\mathrm{K}\\) hours of human annotation for around \\(164\\mathrm{K}\\) images, and annotating a single image in the Cityscapes dataset [11] took 1.5 hours on average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.776 + ], + "angle": 0, + "content": "These costs have triggered the advent of unsupervised semantic segmentation [10, 16, 20, 35], which aims to remove the need for labeled training data in order to train segmentation models. Recently, work by Hamilton et al. [16] has accelerated the progress towards removing the need for labels to achieve good results on semantic segmentation tasks. Their model, STEGO, uses a DINO-pretrained [7] Vision Transformer (ViT) [13] to extract features that are then distilled across the entire dataset to learn semantically relevant features, using a contrastive learning approach. The to-be-distilled features are sampled randomly from feature maps produced from the same image, k-NN matched images as well as other negative images. Seong et al. [35] build on this process by trying to identify features that are most relevant to the model by discovering hidden positives. Their work exposes an inefficiency of random sampling in STEGO as hidden positives sampling leads to significant improvements. However, both approaches only operate on the pixel space and therefore fail to take into account the spatial layout of the scene. Not only do we humans perceive the world in 3D, but also previous work [5, 8, 18, 37, 38, 44] has shown that supervised semantic segmentation can benefit greatly from spatial information during training. Inspired by these observations, we propose to incorporate spatial information in the form of depth maps into the STEGO training process. Depth is considered a product of vision and does not provide a labeled training signal. To obtain depth information for the benchmark image datasets in our evaluations, we make use of an off-the-shelf zero-shot monocular depth estimator to obtain spatial information of the scene. This allows us to incorporate the depth information without the need for human annotations or sensor ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "With our method, \\(DepthG\\), we propose to (1) guide the model to learn a rough spatial layout of the scene, since we hypothesize this will aid the network in differentiating objects much better. We achieve this by extending the contrastive process to the spatial dimension: We do not limit the model to learning only Feature-Feature Correlations, but also Depth-Feature Correlations. Through this process, the model is guided towards pulling apart the features with high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.473, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.472, + 0.471 + ], + "angle": 0, + "content": "Figure 1. Guiding the feature space for unsupervised segmentation with depth information. Our intuition behind the proposed approach is simple: For locations in the 3D space with a low distance, we guide the model to map their features closer together. Vice versa, the features are learned to be drawn apart in feature space if their distance in the metric space is large." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.603 + ], + "angle": 0, + "content": "distances in 3D space, as well as mapping them closer together if their distance is low depth space. Figure 1 visualizes this process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "With the information about the spatial layout of the scene present, we furthermore propose to (2) spatially inform our feature sampling process by utilizing Farthest-Point Sampling (FPS) [14, 31] on the depth map, which equally samples scenes in 3D. We show these techniques in combination make our approach to unsupervised segmentation highly effective, since in our evaluations across multiple benchmark datasets, we demonstrate state-of-the-art performance. We include a short video presenting our method as part of the supplementary materials. To the best of our knowledge, we are the first to propose a mechanism to incorporate 3D knowledge of the scene into unsupervised learning for 2D images without encoding depth maps as part of the network input. This design aspect of our method alleviates the risk of the model developing an input dependency. Therefore, our approach does not rely on depth information during inference and is not affected by availability or quality of such depth information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.826, + 0.132 + ], + "angle": 0, + "content": "2.1. Unsupervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.895, + 0.58 + ], + "angle": 0, + "content": "Recent works [10, 16, 20, 35] have attempted to tackle semantic segmentation without the use of human annotations. Ji et al. [20] propose IIC, a method that aims to maximize the mutual information between different augmented versions of an image. PiCIE, published by Cho et al. [10], introduces an inductive bias based on the invariance to photometric transformations and equivariance to geometric manipulations. DINO [7] often serves as a critical component to unsupervised segmentation algorithms, since the self-supervised pre-trained ViT can produce semantically relevant features. Recent work by Seitzer et al. [34] builds upon this ability by training a model with slot attention [30] to reconstruct the feature maps produced by DINO from the different slots. The features of their object-centric model are clustered with k-means [29] where each slot is associated with a cluster. In their 2021 work, Hamilton et al. [16] have also built upon DINO features by introducing a feature distillation process with features from the same image, k-NN retrieved examples as well as random other images from the dataset. Their learned representations are finally clustered and refined with a CRF [22] for semantic segmentation. While STEGO's feature selection process is random, Seong et al. [35] introduce a more effective sampling strategy by discovering hidden positives. During training, they form task-agnostic and task-specific feature pools. For an anchor feature, they then compute the maximum similarity to any of the pool features and sample locations in the image with greater similarity than the determined value. A more detailed introduction to STEGO is provided in Section 3.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.803, + 0.607 + ], + "angle": 0, + "content": "2.2. Depth For Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Previous research [5, 17, 18, 38, 40, 43] has sought to incorporate depth for semantic segmentation in different settings. Wang et al. [38] propose to use depth for adapting segmentation models to new data domains. Their method adds depth estimation as an auxiliary task to strengthen the prediction of segmentation tasks. Furthermore, they approximate the pixel-wise adaption difficulty from source to target domain through the use of depth decoders. Work by Hoyer et al. [18] explores three further strategies of how depth can be useful for segmentation. First, they propose using a shared backbone to share learning features for segmentation and self-supervised depth estimation, similar to Wang et al. [38]. Second, they use depth maps to introduce a data augmentation that is informed by the structure of the scene. And lastly, they detail the integration of depth into an active learning loop as part of a student-teacher setup. Another work by Hou et al. [17] incorporates depth information into a pre-training algorithm with the aim to learn better representations for semantic segmentation. In their" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3638" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "work, Mask3D, they propose to incorporate depth in the form of a 3D prior by formulating a reconstruction task that operates on masked RGB and depth patches, enabling them to learn more useful features for semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.168, + 0.168, + 0.183 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.194, + 0.469, + 0.24 + ], + "angle": 0, + "content": "In the following, we detail our proposed method for guiding unsupervised segmentation with depth information. An overview of our approach is presented in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.25, + 0.206, + 0.266 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.56 + ], + "angle": 0, + "content": "Our approach builds upon work by Hamilton et al. [16]. In their work, each image is 5-cropped and k-NN correspondences between these images are calculated using the DINO ViT [7]. Generally, STEGO uses a feature extractor \\(\\mathcal{F}:\\mathbb{R}^{3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}\\to \\mathbb{R}^{C\\times H\\times W}\\) with input image height \\(H_{\\mathrm{in}}\\) and width \\(W_{\\mathrm{in}}\\), to calculate a feature map \\(f\\in \\mathbb{R}^{C\\times H\\times W}\\) with height \\(H\\), width \\(W\\) and feature dimension \\(C\\) from the input image. These features are then further encoded by a segmentation head \\(\\mathcal{S}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{Q\\times H\\times W}\\) to calculate the code space \\(s\\in \\mathbb{R}^{Q\\times H\\times W}\\) with code dimension \\(Q\\). With the goal of forming compact clusters and amplifying the correlation of the learned features, let \\(f_{i}\\coloneqq \\mathcal{F}(x_{i})\\) and \\(f_{j}\\coloneqq \\mathcal{F}(x_{j})\\) be feature maps for a given input pair of \\(x_{i}\\) and \\(x_{j}\\), which are then used to calculate \\(s_i\\coloneqq S(f_i)\\) and \\(s_j\\coloneqq S(f_j)\\) from the segmentation head \\(S\\). In practice, STEGO samples \\(N^2\\) vectors from the feature map during training. Hamilton et al. [16] introduce the concept of constructing the feature correspondence tensor \\(\\pmb {F}\\in \\mathbb{R}^{H\\times W\\times H\\times W}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.571, + 0.469, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {h w, u v} = \\frac {f _ {i} ^ {h w} \\cdot f _ {j} ^ {u v}}{\\| f _ {i} ^ {h w} \\| \\| f _ {j} ^ {u v} \\|} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.665 + ], + "angle": 0, + "content": "where \\(\\cdot\\) denotes the dot product. Using the same formula we obtain \\(S\\) using \\(s_i, s_j\\). Consequently, the feature correlation loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.676, + 0.469, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {C o r r}} := - \\sum_ {h w, u v} \\left(\\boldsymbol {F} _ {h w, u v} - b\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.872 + ], + "angle": 0, + "content": "where \\( b \\) is a scalar bias hyperparameter. Empirical evaluations from STEGO have shown that applying spatial centering to the feature correlation loss along with zero-clamping further improves performance [16]. These correlations are calculated for two crops from the same image \\( (\\mathcal{L}_{\\mathrm{self}}) \\) and one from a different but similar image, determined by the k-NN correspondence pre-processing \\( (\\mathcal{L}_{\\mathrm{knn}}) \\). Finally, negative images are sampled randomly \\( (\\mathcal{L}_{\\mathrm{random}}) \\). The final loss is a weighted sum of the different losses where each of them has their individual weight \\( \\lambda_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S T E G O}} = \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k n n}} \\mathcal {L} _ {\\text {k n n}} + \\lambda_ {\\text {r a n d o m}} \\mathcal {L} _ {\\text {r a n d o m}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "After training, the inferred feature maps for a test image are clustered using k-means and refined with a conditional random field (CRF) [22]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.145, + 0.715, + 0.16 + ], + "angle": 0, + "content": "3.2. Depth Map Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Since in many cases, depth information about the scene is not readily available, we make use of recent progress in the field of monocular depth estimation [1-3, 25, 32] to obtain depth maps from RGB images. Recently, methods from this field have made significant advances in zero-shot depth estimation, i.e. predicting depth values for scenes from data domains not seen during training. This property makes them especially suitable for our method, since it enables us to obtain high-quality depth predictions for a wide variety of data domains without ever re-training the depth network. This property also limits the computational cost for our method. For our method, we experiment with different state-of-the-art monocular depth estimators, detailed in Section 5, and found ZoeDepth [3] to perform best in our evaluations. Given a cropped RGB image \\( x_{i} \\), we use the monocular depth estimator \\( \\mathcal{M} \\) together with average pooling to predict depth \\( d_{i}\\in [0,1]^{H\\times W} \\) at feature resolution:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.432, + 0.891, + 0.448 + ], + "angle": 0, + "content": "\\[\nd _ {i} = \\operatorname {p o o l} \\left(\\mathcal {M} \\left(x _ {i}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.501 + ], + "angle": 0, + "content": "The average pooling operation is used to match the dimensions of the feature map, which is required to sample nonoverlapping locations at the patch resolution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.783, + 0.523 + ], + "angle": 0, + "content": "3.3. Depth-Feature Correlation Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.576 + ], + "angle": 0, + "content": "With our Depth-Feature Correlation loss, we aim to enforce spatial consistency in the feature map by transferring the distances from the depth map to the latent feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.879 + ], + "angle": 0, + "content": "In contrastive learning, the network is incentivized to decrease the distance in feature space for similar instances, therefore learning to map their latent representations closer together. Likewise, different instances are drawn further apart in feature distance. We assume the same concept to be true in 3D space: The spatial distance between two points from the same depth plateau is smaller, while the distance between a point in the foreground and one in the background is larger. Since, in both spaces, the concept of measuring difference is represented by the distance between two points, we propose to align them through our concept of Depth-Feature Correlation: For large distances in the 3D space, we encourage the network to produce vectors that are further apart, and vice versa. With this, we induce the model with knowledge about the spatial structure of the scene, enabling it to better differentiate between objects. To achieve this we construct the depth correspondence tensor similar to the feature correspondence from Equation 1. The depth correspondence tensor \\( D \\) is computed from the depths of two different image crops as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.884, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {D} _ {h w, u v} = d _ {i} ^ {h w} d _ {j} ^ {u v}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.087, + 0.546, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.698, + 0.171, + 0.744, + 0.186 + ], + "angle": 0, + "content": "\\(D_{hw,uv}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.895, + 0.421 + ], + "angle": 0, + "content": "Figure 2. Overview of the DepthG training process. After 5-cropping the image, each crop is encoded by the DINO-pretrained ViT \\(\\mathcal{F}\\) to output a feature map. Using farthest point sampling (FPS), we sample the 3D space equally and convert the coordinates to select samples in the feature map. The sampled features are further transformed by the segmentation head \\(S\\). For both feature maps, the correlation tensor is computed. Following, we sample the depth map at the coordinates obtained by FPS and compute a correlation tensor in the same fashion. Finally, we compute our Depth-Feature Correlation loss and combine it with the feature distillation loss from STEGO. We guide the model to learn depth-feature correlation for crops of the same image, while the feature distillation loss is also applied to k-NN-selected and random images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.471, + 0.482 + ], + "angle": 0, + "content": "where \\((h,w)\\) and \\((u,v)\\) represent the pixel positions in the depth maps \\(d_{i}\\) and \\(d_{j}\\) respectively. Together with the zero clamping, our Depth-Feature Correlation loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.494, + 0.47, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D e p t h G}} = - \\sum_ {h w, u v} \\left(\\boldsymbol {D} _ {h w, u v} - b _ {\\text {D e p t h G}}\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.541, + 0.47, + 0.661 + ], + "angle": 0, + "content": "where \\(D_{hw,uv}\\) represents the depth correlation tensor, \\(b_{\\mathrm{DepthG}}\\) is the bias for our loss term, and \\(S_{hw,uv}\\) represents the feature correlation tensor computed from the output features of the segmentation head \\(\\mathcal{S}\\). By also using zero-clamping, we limit erroneous learning signals that aim to draw apart instances of the same class if they have large spatial differences. With this, we extend the STEGO loss so it can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.674, + 0.47, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S T E G O + D e p t h G}} = \\mathcal {L} _ {\\text {S T E G O}} + \\lambda_ {\\text {D e p t h G}} \\mathcal {L} _ {\\text {D e p t h G}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.792 + ], + "angle": 0, + "content": "with Depth-Feature Correlation weight \\(\\lambda_{\\mathrm{DepthG}}\\). By inducing depth knowledge during training without encoding the depth maps as part of the model input, our model can predict spatially informed segmentations on RGB images with its distilled knowledge and does not rely on depth to be available at test time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.367, + 0.819 + ], + "angle": 0, + "content": "3.4. Depth-Guided Feature Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also aim to make the feature sampling process informed by the spatial layout of the scene. To perform sampling in the 3D space, we transform the downsampled depth map \\( d(x_{i}) \\) into a point cloud with points \\( \\{p_1,p_2,\\dots ,p_n\\} \\). On this point cloud, we apply farthest point sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.895, + 0.665 + ], + "angle": 0, + "content": "(FPS) [14], in an iterative fashion by always selecting the next point \\( p_k \\) as the point with the maximum distance in 3D space with respect to the already sampled points \\( \\{p_1, p_2, \\dots, p_{k-1}\\} \\). After having sampled \\( N^2 \\) points, we end up with a set of samples \\( \\{p_1, p_2, \\dots, p_{N^2}\\} \\) which are consequently converted to 2D sampling indices for the feature maps \\( f \\) and \\( g \\). In contrast to the data-agnostic random sampling applied in STEGO, our feature selection process takes into account the geometry of the input scene and covers the spatial structure more equally. In our ablations in Section 5, we show this scene coverage from FPS of the depth space further increases the effectiveness of our Depth Feature Correlation loss, due to the increased diversity in selected 3D locations. We show a visual comparison between random and farthest point sampling in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.702, + 0.696 + ], + "angle": 0, + "content": "3.5. Guidance Scheduling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "While our Depth-Feature Correlation loss is effective at enriching the model's learning process with spatial information of the scene, we aim to alleviate the danger of it interfering with the learning of feature correlations during model training. We hypothesize that our model most greatly benefits from depth information in the beginning of training when its only knowledge is encoded in the features maps by the frozen ViT backbone. To give it a head start, we increase the weight of our Depth-Feature Correlation loss in the beginning and gradually decrease its influence during training. Vice versa, the distillation process in the feature space will increasingly emphasised as the training progresses. In this way, the network builds upon the already learned rough spa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Local Hidden Positives. We visualize the use of depth and attention maps for local hidden positives. For this visualization, we sample the respective propagation maps at the yellow patch in the center of the crops. We observe the depth map to have sharper borders and more consistent propagation values. We experiment with both propagation strategies in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.471, + 0.406 + ], + "angle": 0, + "content": "tial structure of the scene achieved through our depth guidance. Therefore, we implement an exponential decay of the weight \\(\\lambda_{\\mathrm{DepthG}}\\) and bias \\(b_{\\mathrm{DepthG}}\\) of our loss component. We ablate on the use of guidance scheduling in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.387, + 0.433 + ], + "angle": 0, + "content": "3.6. Local Hidden Positives In 3D Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.804 + ], + "angle": 0, + "content": "We further explore the combination of our method with Hidden Positives [35], which also builds upon STEGO. As we demonstrate as part of our ablation on the individual influence of our contributions in Section 5, our feature sampling method, implemented through farthest point sampling, is integral to the effectiveness of DepthG. Therefore, we decide not to replace it with the global hidden positives sampling from Seong et al. [35]. Instead, we implement a depth-informed variant of local hidden positives, 3D-LHP, where the loss for an individual patch is propagated to eight neighboring patches proportionally to their attention values obtained from the feature extractor. We modify this strategy by instead propagating the learning signal to the closest patches in 3D space proportional to their relative distances, using the point cloud described in Section 3.4. As visualized in Figure 3, we observe that our depth-based propagation map has sharper and more consistent surfaces. To propagate the learning signal to the selected patches, we follow Hidden Positives and mix the features coming from the segmentation head \\( S \\) in proportion to their propagation values (point distances). The mixed representations are then fed through an additional projection head \\( \\mathcal{P} \\). We then calculate \\( \\mathcal{L}_{\\mathrm{STEGO+DepthG}} \\) again for the produced output features and combine it with the loss from the segmentation head output." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.21, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets and Models. We conduct experiments on the COCO-Stuff [4], Cityscapes [11] and Potsdam-3 datasets. COCO-Stuff contains a wide variety of real-world scenes. In our evaluation, we follow [16, 20, 35] and provide re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.365 + ], + "angle": 0, + "content": "sults on the coarse class split, COCO-Stuff 27. In contrast, Cityscapes contains traffic scenes from 50 cities from a driver-like viewpoint. Lastly, the Potsdam-3 dataset is composed of aerial, top-down images of the city of Potsdam. We use the DINO [7] backbones ViT-Small (ViT-S) and ViT-Base (ViT-B), which were pre-trained in a self-supervised manner on ImageNet-1k [12]. We choose the models with a patch size of \\(8 \\times 8\\), since they have been shown to perform best for semantic segmentation due to the higher resolution of the resulting feature maps [16, 35]. In the result tables, we refer to DepthG models trained with our DepthFeature Correlation loss and FPS as Ours. Models that utilize 3D-LHP propagation with depth maps in addition are displayed as Ours w/ 3D-LHP. We compare our approach against competing methods which were never trained with human annotations, i.e. human labels or language supervision, neither in the feature extractor nor the segmentation training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Evaluation Protocols. Similar to STEGO and related work [16, 35], we evaluate our models in the unsupervised, clustering-based setting, as well as linear probing. Since the output of our model is a pixel-level map of features and not class labels, these features are consequently clustered. Following, the pseudo-labeled clusters are aligned with the ground truth labels through Hungarian matching [23, 24] across the entire validation dataset. To perform linear probing, an additional linear layer is added on top of the model and trained with cross-entropy loss to learn classification of the features." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.81 + ], + "angle": 0, + "content": "
SettingUnsupervisedLinear
MethodModelAcc.mIoUAcc.mIoU
IIC [20]R18+FPN21.86.744.58.4
PiCIE [10]R18+FPN48.113.854.213.9
PiCIE+H [10]R18+FPN50.014.454.814.8
DINO [7]ViT-S/828.711.368.633.9
ACSeg [26]ViT-S/816.4---
TransFGU [42]ViT-S/817.552.7--
STEGO + HP [35]ViT-S/857.224.675.642.7
STEGO [16]ViT-S/848.324.574.438.3
+ OursViT-S/856.325.673.738.9
+ Ours w/ 3D-LHPViT-S/855.126.773.937.8
DINO [7,16]ViT-B/830.59.666.829.4
DINOSAUR [34]*ViT-B/844.924.0--
STEGO [16]ViT-B/856.928.276.141.0
+ OursViT-B/858.629.075.541.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.819, + 0.895, + 0.892 + ], + "angle": 0, + "content": "Table 1. Evaluation on COCO-Stuff 27. We report results on COCO-Stuff with 27 high-level classes. Overall, our method outperforms STEGO and HP on unsupervised segmentation with the ViT-B/8, while showing competitive performance for the ViT-S/8. *Results obtained without post-processing optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "3641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.216, + 0.105 + ], + "angle": 0, + "content": "4.1. COCO-Stuff" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.357 + ], + "angle": 0, + "content": "We present our evaluation on COCO-Stuff 27 in Table 1. For the ViT-S/8, our experiments show that Ours is able to improve upon STEGO in most metrics, with improved unsupervised accuracy by \\(+8.0\\%\\) and unsupervised mIoU increased by \\(+1.1\\%\\). Ours w/ 3D-LHP further increases this mIoU delta to \\(+1.8\\%\\), highlighting the effectiveness of our 3D-information propagation strategy in combination with DepthG. When comparing our approach to Hidden Positives, a method with more computational overhead, for the ViT-S/8, we show competitive performance for unsupervised accuracy and outperform their approach by \\(+1.0\\%\\) on unsupervised mIoU with Ours and \\(+1.7\\%\\) with Ours w/ 3D-LHP. When using the DINO ViT-B/8 encoder, our approach again outperforms STEGO, as well as all other presented methods on unsupervised metrics. Most notably, we are able to increase the unsupervised mIoU by \\(+0.8\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.198, + 0.385 + ], + "angle": 0, + "content": "4.2. Cityscapes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.59 + ], + "angle": 0, + "content": "We further evaluate our approach on the Cityscapes dataset [11], consisting of various scenes from 50 different cities. We follow the training setting from STEGO and, contrary to all other datasets, do not sample point-wise but use the full feature map for our learning process along with the full depth map. As can be seen in Table 2, our method significantly outperforms STEGO as well as Hidden Positives. For unsupervised mIoU, while Hidden Positives decreased performance compared to STEGO, we observe that our approach to achieves a \\(+2.1\\%\\) increase. Similarly, we report state-of-the-art performance in accuracy, building upon Hidden Positives' already impressive improvements over STEGO and outperforming it by \\(+2.1\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.182, + 0.615 + ], + "angle": 0, + "content": "4.3. Potsdam" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.473, + 0.837 + ], + "angle": 0, + "content": "Our model is further evaluated on the Potsdam-3 dataset, containing aerial images of the German city of Potsdam. Contrary to the other benchmarks, which contain images in a first-person perspective, Potsdam-3 contains only birds-eye-view images, a perspective that is considered out-of-distribution for the monocular depth estimator. Despite this inherent limitation of our approach for aerial data, we are able to demonstrate relatively commendable performance in Table 3 by improving STEGO's performance and reporting state-of-the-art performance for the ViT-S backbone. In contrast, Hidden Positives [35] use a ViT-B/8, and with roughly twice the parameters as ours, reach an accuracy of \\(82.4\\%\\). We present a visual overview of the predicted Potsdam depth maps in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.262, + 0.863 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We present qualitative results of our method in Figure 4 and compare with segmentation maps from STEGO. On" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.089, + 0.885, + 0.219 + ], + "angle": 0, + "content": "
MethodModelU. AccU. mIoU
IIC [20]R18+FPN47.96.4
PiCIE [10]R18+FPN65.612.3
DINO [7]ViT-B/843.611.8
STEGO + HP [35]ViT-B/879.518.4
STEGO [16]ViT-B/873.221.0
+ OursViT-B/881.623.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.228, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Table 2. Results on Cityscapes. We report unsupervised accuracy and mIoU on Cityscapes. Our method outperforms both STEGO variants by substantial margins. Notably, our method is the first to improve upon unsupervised mIoU." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.295, + 0.851, + 0.44 + ], + "angle": 0, + "content": "
MethodModelU. Acc.
CC [19]VGG1163.9
DeepCluster [6]VGG1141.7
IIC [20]VGG1165.1
DINO [7,21]ViT-S/871.3
STEGO [16,21]ViT-S/877.0
+ OursViT-S/880.4
+ Ours w/ 3D-LHPViT-S/880.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.45, + 0.895, + 0.521 + ], + "angle": 0, + "content": "Table 3. Results on Potsdam. We report unsupervised accuracy on the Potsdam dataset. Our method is able to improve upon STEGO. We hypothesize that with a zero-shot depth estimator more suitable for aerial images, the results for our method could further improve." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.528, + 0.894, + 0.657 + ], + "angle": 0, + "content": "
MethodU. mIoU.U. Acc
STEGO [16]24.548.3
+ Depth-Feature Correlation (1)24.751.2
+ FPS (2)24.649.1
+ 3D-LHP (3)25.248.5
+ Ours (1, 2)25.656.3
+ Ours w/ 3D-LHP (1, 2, 3)26.755.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.668, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Table 4. Effect of our contributions. We compare our individual contributions and the combination of all contributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "multiple occasions, our depth guidance reduces erroneous predictions from the model caused by visual irritations in the pixel space. In the example of the boy with the baseball bat in Figure 4a, false classifications from STEGO are caused by shadows on the ground. Our model is able to correct this. Furthermore, it goes beyond the noisy label and also correctly classifies the glimpse of a plant that can be seen through a hole in the background. This is an indication that our model does not overfit to the depth map, since this visual cue is only observable in the pixel space, but not the depth map." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.479, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.355, + 0.321, + 0.366 + ], + "angle": 0, + "content": "(a) COCO-Stuff" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.087, + 0.895, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.355, + 0.733, + 0.367 + ], + "angle": 0, + "content": "(b) Cityscapes" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.895, + 0.437 + ], + "angle": 0, + "content": "Figure 4. Qualitative results. We show qualitative differences for plain STEGO compared to STEGO with our depth guidance, using ViT-S models for COCO and ViT-B for Cityscapes. Where STEGO struggles to differentiate instances, our model is able to correct this and successfully separates them for segmentation. In the case of the building in (a), our method alleviates visual irritations from the pixel space and corrects the segmentation of the building. In (b), our model is able to better handle visual inconsistencies from shadows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.184, + 0.465 + ], + "angle": 0, + "content": "5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Individual Influence. We investigate the effect of our technical contributions on training our model with a ViT-S/8 backbone on COCO-Stuff 27. Our observations in Table 4 show that our Depth-Feature Correlation loss itself already improves the performance of STEGO. This improvement is further increased through the use of FPS, which enables us to sample the depth space more meaningfully and therefore encourages more diversity in the depth correlation tensor \\( D_{hw,uv} \\). Intuitively, this sampling diversity significantly amplifies our Depth-Feature Correlation for aligning the feature space with the depth space. We provide a visual comparison to random sampling in Figure 5 and additional illustrations in the appendix. FPS retrieves more diverse locations and specifically selects locations with depth discontinuities. This naturally benefits the Depth-Feature Correlation to learn sharper edges in this area for the output segmentation. Adding local hidden positives with depth maps further increases the unsupervised mIoU, while slightly lowering the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Source Of Depth Maps. We investigate the effect of different monocular depth estimators to generate the depth maps used to train our model. In our experiments, we consider three options: The previously mentioned ZoeDepth [3], Kick Back & Relax (KBR) [36] which uses self-supervision to learn depth from Slow-TV videos, as well as MiDaS [32], the base model to ZoeDepth. Empirically, ZoeDepth produces the most accurate zero-shot" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.603 + ], + "angle": 0, + "content": "monodepth results across indoor and outdoor datasets, followed by MiDaS and KBR [3,36]. We provide a qualitative comparison in the appendix. To evaluate the influence of the depth map quality on the performance of our model, we first generate depth maps for COCO-Stuff 27 using each of the introduced models. We then train our model with a ViT-S/8 backbone and report the results in Table 5. We observe that depth maps from ZoeDepth work best with our method, while the model trained with MiDaS depth has an edge over the KBR counterpart." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.618, + 0.892, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.806, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Figure 5. Random vs. Farthest Point Sampling. We observe that random sampling can miss entire structures like trees in the first top and the plane in the bottom row. In contrast, our method meaningfully samples the depth space and selects locations across the different structures and at depth edges. We show further illustrations of FPS in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Figure 6. Failure cases. We show cases where our model fails to correctly segment and classify the scene. The top row is a prime example where the difference in depth is correctly distilled, though the model fails to correctly classify the snow region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Signal Propagation With Local Hidden Positives. We ablate the implementation of LHP along with different propagation strategies. As described in Section 3.6, our method takes advantage of the depth information of the scene to propagate the learning signal to patches which are nearby in 3D space. We further implement utilizing the attention map from the DINO backbone and propagate proportionally to their values i.e., the approach utilized in Hidden Positives [35]. While, in Table 6, we show that our approach benefits from signal propagation with 3D-LHP, applying LHP with Attention leads to lower performance. We speculate 3D-LHP is more suitable for our approach since, for a given location, the signal from our Depth-Feature Correlation loss is calculated w.r.t. the depth at this sample. Propagating to locations with the same depth does not corrupt this signal, though this can happen with LHP (Attention), since it does not consider depth for propagation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Computational Cost. Our method only leads to an insignificant increase in runtime versus the baseline STEGO model, since we solely guide the loss as well as the feature sampling and only for Ours w/ 3D-LHP add an additional small segmentation head. In contrast, the competing method Hidden Positives [35] relies on a computationally more expensive process to select features and introduces an additional segmentation head to fill their task-specific feature pool. To keep our computational overhead low, we make use of a pre-trained monocular depth estimation network with impressive zero-shot capabilities. We consider task specific training of the depth estimator not a necessity, since the model has zero-shot capabilities that generalize well to different scenes and domains. Therefore, in our experiments on a diverse array of scenes, we do not re-train the depth estimator, and consider the additional computational cost for generating the depth maps to be negligible." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.166 + ], + "angle": 0, + "content": "
MethodTrained withU. Acc.U. mIoU
ZoeDepth [3]Sensor Depth56.325.6
MiDaS [32]Sensor Depth53.025.0
KBR [36]Self-Supervision50.623.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 5. Different depth map sources. We experiment with different monocular depth estimators which were trained with either sensor depth or self-supervision. Overall, the model trained with depth maps from ZoeDepth performs best on COCO-Stuff 27." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.251, + 0.875, + 0.328 + ], + "angle": 0, + "content": "
LHPPropagation StrategyU. Acc.U. mIoU
X-56.325.6
✓3D-LHP (Depth)55.126.7
✓LHP (Attention)52.624.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.338, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Table 6. LHP Ablation. We compare the use of depth and attention for local hidden positives. We find that using depth improves unsupervised mIoU, while we find that using attention does not improve our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.623, + 0.433 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.611 + ], + "angle": 0, + "content": "While we have demonstrated our method's effectiveness for many real-world cases, our method's applicability is limited in settings unsuitable for depth estimation, such as slices of CT scans and other medical data domains. Furthermore, the experiments on Potsdam-3 have shown, our method can improve unsupervised semantic segmentation despite suboptimal viewing perspectives for the monocular depth estimator. We assume the scenario of aerial images represents a rare case where our method would profit from a domain-specific monocular depth estimator. We also present failure cases of our model in Figure 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.755, + 0.647 + ], + "angle": 0, + "content": "7. Conclusion & Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we have presented a novel method to induce spatial knowledge of the scene into our model for unsupervised semantic segmentation. We have proposed to correlate the feature space with the depth space and use the 3D information to more meaningfully sample features in a spatially informed way. Furthermore, we have demonstrated that these contributions produce state-of-the-art performance on many real-world datasets and thus foster the progress in unsupervised segmentation. The applicability of our approach for other tasks is further to be explored, since we hypothesize it can be useful beyond unsupervised segmentation as part of other contrastive processes. We consider this to be a promising direction for future work. Furthermore, it remains to be investigated what kind of information could be useful in domains where depth is not an obviously meaningful signal, like medical data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3644" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.469, + 0.229 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[5] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1129-1139, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[8] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2624-2632, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[10] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[14] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[15] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[16] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.399, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[17] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[18] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11130-11140, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[19] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H Adelson. Learning visual groups from co-occurrences in space and time. arXiv preprint arXiv:1511.06811, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[20] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9865-9874, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[21] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of stego for safe unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3788-3797, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[22] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Harold W Kuhn. Variants of the hungarian method for assignment problems. Naval research logistics quarterly, 3(4):253-258, 1956. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3645" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[26] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[27] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[29] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2):129-137, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[30] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.471, + 0.504 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.471, + 0.575 + ], + "angle": 0, + "content": "[32] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[34] M Seitzer, M Horn, A Zadaianchuk, D Zietlow, T Xiao, C Simon-Gabriel, T He, Z Zhang, B Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In International Conference on Learning Representations (ICLR), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[35] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19540-19549, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[36] Jaime Spencer, Chris Russell, Simon Hadfield, and Richard Bowden. Kick back & relax: Learning to reconstruct the world by watching slowtv. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[37] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[38] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8515-8525, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[39] Risheng Wang, Tao Lei, Ruixia Cui, Bingtao Zhang, Hongying Meng, and Asoke K Nandi. Medical image segmentation using deep learning: A survey. IET Image Processing, 16(5):1243-1267, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[40] Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. Multimodal token fusion for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12186-12195, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[41] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[42] Zhaoyuan Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, and Rong Jin. Transfgu: a top-down approach to fine-grained unsupervised semantic segmentation. In European conference on computer vision, pages 73-89. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[43] Xiaowen Ying and Mooi Choo Chuah. Uctnet: Uncertainty-aware cross-modal transformer network for indoor rgb-d semantic segmentation. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[44] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[45] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3646" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_origin.pdf b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25c0194ca1cb5269f9f302d97d84a3664581a301 --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/490ac818-ae06-4da1-a305-2db07efc0513_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7618e16b28548fdce5c1d3b0318670f844aad04dacda4ea211a2f047e255f1dc +size 3671246 diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/full.md b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e6da1840b65b6e21fbb837afbe58e5b7e243683 --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/full.md @@ -0,0 +1,260 @@ +# Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling + +Leon Sick Ulm University + +Dominik Engel Ulm University + +Pedro Hermosilla TU Vienna + +Timo Ropinski +Ulm University + +# Abstract + +Traditionally, training neural networks to perform semantic segmentation requires expensive human-made annotations. But more recently, advances in the field of unsupervised learning have made significant progress on this issue and towards closing the gap to supervised algorithms. To achieve this, semantic knowledge is distilled by learning to correlate randomly sampled features from images across an entire dataset. In this work, we build upon these advances by incorporating information about the structure of the scene into the training process through the use of depth information. We achieve this by (1) learning depth-feature correlation by spatially correlating the feature maps with the depth maps to induce knowledge about the structure of the scene and (2) exploiting farthest-point sampling to more effectively select relevant features by utilizing 3D sampling techniques on depth information of the scene. Finally, we demonstrate the effectiveness of our technical contributions through extensive experimentation and present significant improvements in performance across multiple benchmark datasets. + +# 1. Introduction + +Semantic segmentation plays a critical role in many of today's vision systems in a multitude of domains. These include, among others, autonomous driving [15], medical applications [33, 39], and many more [9, 27, 41, 45]. Until recently, the main body of research in this area was focused on supervised models that require a large amount pixel-level annotations for training. Not only is sourcing this image data often a labor intensive process, but also annotating the large datasets required for good performance comes at a high price. Several benchmark datasets report their annotation times. For example, the MS COCO dataset [28] required more than $28\mathrm{K}$ hours of human annotation for around $164\mathrm{K}$ images, and annotating a single image in the Cityscapes dataset [11] took 1.5 hours on average. + +These costs have triggered the advent of unsupervised semantic segmentation [10, 16, 20, 35], which aims to remove the need for labeled training data in order to train segmentation models. Recently, work by Hamilton et al. [16] has accelerated the progress towards removing the need for labels to achieve good results on semantic segmentation tasks. Their model, STEGO, uses a DINO-pretrained [7] Vision Transformer (ViT) [13] to extract features that are then distilled across the entire dataset to learn semantically relevant features, using a contrastive learning approach. The to-be-distilled features are sampled randomly from feature maps produced from the same image, k-NN matched images as well as other negative images. Seong et al. [35] build on this process by trying to identify features that are most relevant to the model by discovering hidden positives. Their work exposes an inefficiency of random sampling in STEGO as hidden positives sampling leads to significant improvements. However, both approaches only operate on the pixel space and therefore fail to take into account the spatial layout of the scene. Not only do we humans perceive the world in 3D, but also previous work [5, 8, 18, 37, 38, 44] has shown that supervised semantic segmentation can benefit greatly from spatial information during training. Inspired by these observations, we propose to incorporate spatial information in the form of depth maps into the STEGO training process. Depth is considered a product of vision and does not provide a labeled training signal. To obtain depth information for the benchmark image datasets in our evaluations, we make use of an off-the-shelf zero-shot monocular depth estimator to obtain spatial information of the scene. This allows us to incorporate the depth information without the need for human annotations or sensor ground truth. + +With our method, $DepthG$ , we propose to (1) guide the model to learn a rough spatial layout of the scene, since we hypothesize this will aid the network in differentiating objects much better. We achieve this by extending the contrastive process to the spatial dimension: We do not limit the model to learning only Feature-Feature Correlations, but also Depth-Feature Correlations. Through this process, the model is guided towards pulling apart the features with high + +![](images/40bcafca6e9638f4a9deefdaae77ca1abd41bf8d4c9de28e7095e01c1291f39b.jpg) +Figure 1. Guiding the feature space for unsupervised segmentation with depth information. Our intuition behind the proposed approach is simple: For locations in the 3D space with a low distance, we guide the model to map their features closer together. Vice versa, the features are learned to be drawn apart in feature space if their distance in the metric space is large. + +distances in 3D space, as well as mapping them closer together if their distance is low depth space. Figure 1 visualizes this process. + +With the information about the spatial layout of the scene present, we furthermore propose to (2) spatially inform our feature sampling process by utilizing Farthest-Point Sampling (FPS) [14, 31] on the depth map, which equally samples scenes in 3D. We show these techniques in combination make our approach to unsupervised segmentation highly effective, since in our evaluations across multiple benchmark datasets, we demonstrate state-of-the-art performance. We include a short video presenting our method as part of the supplementary materials. To the best of our knowledge, we are the first to propose a mechanism to incorporate 3D knowledge of the scene into unsupervised learning for 2D images without encoding depth maps as part of the network input. This design aspect of our method alleviates the risk of the model developing an input dependency. Therefore, our approach does not rely on depth information during inference and is not affected by availability or quality of such depth information. + +# 2. Related Work + +# 2.1. Unsupervised Semantic Segmentation + +Recent works [10, 16, 20, 35] have attempted to tackle semantic segmentation without the use of human annotations. Ji et al. [20] propose IIC, a method that aims to maximize the mutual information between different augmented versions of an image. PiCIE, published by Cho et al. [10], introduces an inductive bias based on the invariance to photometric transformations and equivariance to geometric manipulations. DINO [7] often serves as a critical component to unsupervised segmentation algorithms, since the self-supervised pre-trained ViT can produce semantically relevant features. Recent work by Seitzer et al. [34] builds upon this ability by training a model with slot attention [30] to reconstruct the feature maps produced by DINO from the different slots. The features of their object-centric model are clustered with k-means [29] where each slot is associated with a cluster. In their 2021 work, Hamilton et al. [16] have also built upon DINO features by introducing a feature distillation process with features from the same image, k-NN retrieved examples as well as random other images from the dataset. Their learned representations are finally clustered and refined with a CRF [22] for semantic segmentation. While STEGO's feature selection process is random, Seong et al. [35] introduce a more effective sampling strategy by discovering hidden positives. During training, they form task-agnostic and task-specific feature pools. For an anchor feature, they then compute the maximum similarity to any of the pool features and sample locations in the image with greater similarity than the determined value. A more detailed introduction to STEGO is provided in Section 3.1. + +# 2.2. Depth For Semantic Segmentation + +Previous research [5, 17, 18, 38, 40, 43] has sought to incorporate depth for semantic segmentation in different settings. Wang et al. [38] propose to use depth for adapting segmentation models to new data domains. Their method adds depth estimation as an auxiliary task to strengthen the prediction of segmentation tasks. Furthermore, they approximate the pixel-wise adaption difficulty from source to target domain through the use of depth decoders. Work by Hoyer et al. [18] explores three further strategies of how depth can be useful for segmentation. First, they propose using a shared backbone to share learning features for segmentation and self-supervised depth estimation, similar to Wang et al. [38]. Second, they use depth maps to introduce a data augmentation that is informed by the structure of the scene. And lastly, they detail the integration of depth into an active learning loop as part of a student-teacher setup. Another work by Hou et al. [17] incorporates depth information into a pre-training algorithm with the aim to learn better representations for semantic segmentation. In their + +work, Mask3D, they propose to incorporate depth in the form of a 3D prior by formulating a reconstruction task that operates on masked RGB and depth patches, enabling them to learn more useful features for semantic segmentation. + +# 3. Method + +In the following, we detail our proposed method for guiding unsupervised segmentation with depth information. An overview of our approach is presented in Figure 2. + +# 3.1. Preliminary + +Our approach builds upon work by Hamilton et al. [16]. In their work, each image is 5-cropped and k-NN correspondences between these images are calculated using the DINO ViT [7]. Generally, STEGO uses a feature extractor $\mathcal{F}:\mathbb{R}^{3\times H_{\mathrm{in}}\times W_{\mathrm{in}}}\to \mathbb{R}^{C\times H\times W}$ with input image height $H_{\mathrm{in}}$ and width $W_{\mathrm{in}}$ , to calculate a feature map $f\in \mathbb{R}^{C\times H\times W}$ with height $H$ , width $W$ and feature dimension $C$ from the input image. These features are then further encoded by a segmentation head $\mathcal{S}:\mathbb{R}^{C\times H\times W}\to \mathbb{R}^{Q\times H\times W}$ to calculate the code space $s\in \mathbb{R}^{Q\times H\times W}$ with code dimension $Q$ . With the goal of forming compact clusters and amplifying the correlation of the learned features, let $f_{i}\coloneqq \mathcal{F}(x_{i})$ and $f_{j}\coloneqq \mathcal{F}(x_{j})$ be feature maps for a given input pair of $x_{i}$ and $x_{j}$ , which are then used to calculate $s_i\coloneqq S(f_i)$ and $s_j\coloneqq S(f_j)$ from the segmentation head $S$ . In practice, STEGO samples $N^2$ vectors from the feature map during training. Hamilton et al. [16] introduce the concept of constructing the feature correspondence tensor $\pmb {F}\in \mathbb{R}^{H\times W\times H\times W}$ as follows: + +$$ +\boldsymbol {F} _ {h w, u v} = \frac {f _ {i} ^ {h w} \cdot f _ {j} ^ {u v}}{\| f _ {i} ^ {h w} \| \| f _ {j} ^ {u v} \|} \tag {1} +$$ + +where $\cdot$ denotes the dot product. Using the same formula we obtain $S$ using $s_i, s_j$ . Consequently, the feature correlation loss is defined as: + +$$ +\mathcal {L} _ {\text {C o r r}} := - \sum_ {h w, u v} \left(\boldsymbol {F} _ {h w, u v} - b\right) \max \left(\boldsymbol {S} _ {h w, u v}, 0\right) \tag {2} +$$ + +where $b$ is a scalar bias hyperparameter. Empirical evaluations from STEGO have shown that applying spatial centering to the feature correlation loss along with zero-clamping further improves performance [16]. These correlations are calculated for two crops from the same image $(\mathcal{L}_{\mathrm{self}})$ and one from a different but similar image, determined by the k-NN correspondence pre-processing $(\mathcal{L}_{\mathrm{knn}})$ . Finally, negative images are sampled randomly $(\mathcal{L}_{\mathrm{random}})$ . The final loss is a weighted sum of the different losses where each of them has their individual weight $\lambda_{i}$ : + +$$ +\mathcal {L} _ {\text {S T E G O}} = \lambda_ {\text {s e l f}} \mathcal {L} _ {\text {s e l f}} + \lambda_ {\text {k n n}} \mathcal {L} _ {\text {k n n}} + \lambda_ {\text {r a n d o m}} \mathcal {L} _ {\text {r a n d o m}} \tag {3} +$$ + +After training, the inferred feature maps for a test image are clustered using k-means and refined with a conditional random field (CRF) [22]. + +# 3.2. Depth Map Generation + +Since in many cases, depth information about the scene is not readily available, we make use of recent progress in the field of monocular depth estimation [1-3, 25, 32] to obtain depth maps from RGB images. Recently, methods from this field have made significant advances in zero-shot depth estimation, i.e. predicting depth values for scenes from data domains not seen during training. This property makes them especially suitable for our method, since it enables us to obtain high-quality depth predictions for a wide variety of data domains without ever re-training the depth network. This property also limits the computational cost for our method. For our method, we experiment with different state-of-the-art monocular depth estimators, detailed in Section 5, and found ZoeDepth [3] to perform best in our evaluations. Given a cropped RGB image $x_{i}$ , we use the monocular depth estimator $\mathcal{M}$ together with average pooling to predict depth $d_{i}\in [0,1]^{H\times W}$ at feature resolution: + +$$ +d _ {i} = \operatorname {p o o l} \left(\mathcal {M} \left(x _ {i}\right)\right) \tag {4} +$$ + +The average pooling operation is used to match the dimensions of the feature map, which is required to sample nonoverlapping locations at the patch resolution. + +# 3.3. Depth-Feature Correlation Loss + +With our Depth-Feature Correlation loss, we aim to enforce spatial consistency in the feature map by transferring the distances from the depth map to the latent feature space. + +In contrastive learning, the network is incentivized to decrease the distance in feature space for similar instances, therefore learning to map their latent representations closer together. Likewise, different instances are drawn further apart in feature distance. We assume the same concept to be true in 3D space: The spatial distance between two points from the same depth plateau is smaller, while the distance between a point in the foreground and one in the background is larger. Since, in both spaces, the concept of measuring difference is represented by the distance between two points, we propose to align them through our concept of Depth-Feature Correlation: For large distances in the 3D space, we encourage the network to produce vectors that are further apart, and vice versa. With this, we induce the model with knowledge about the spatial structure of the scene, enabling it to better differentiate between objects. To achieve this we construct the depth correspondence tensor similar to the feature correspondence from Equation 1. The depth correspondence tensor $D$ is computed from the depths of two different image crops as follows: + +$$ +\boldsymbol {D} _ {h w, u v} = d _ {i} ^ {h w} d _ {j} ^ {u v}, \tag {5} +$$ + +Figure 2. Overview of the DepthG training process. After 5-cropping the image, each crop is encoded by the DINO-pretrained ViT $\mathcal{F}$ to output a feature map. Using farthest point sampling (FPS), we sample the 3D space equally and convert the coordinates to select samples in the feature map. The sampled features are further transformed by the segmentation head $S$ . For both feature maps, the correlation tensor is computed. Following, we sample the depth map at the coordinates obtained by FPS and compute a correlation tensor in the same fashion. Finally, we compute our Depth-Feature Correlation loss and combine it with the feature distillation loss from STEGO. We guide the model to learn depth-feature correlation for crops of the same image, while the feature distillation loss is also applied to k-NN-selected and random images. +![](images/4814db22c9dc63ad43748ca08fbc5988ddb67f17e0bed4334757c1e413bd4902.jpg) +$D_{hw,uv}$ + +where $(h,w)$ and $(u,v)$ represent the pixel positions in the depth maps $d_{i}$ and $d_{j}$ respectively. Together with the zero clamping, our Depth-Feature Correlation loss is defined as: + +$$ +\mathcal {L} _ {\text {D e p t h G}} = - \sum_ {h w, u v} \left(\boldsymbol {D} _ {h w, u v} - b _ {\text {D e p t h G}}\right) \max \left(\boldsymbol {S} _ {h w, u v}, 0\right) \tag {6} +$$ + +where $D_{hw,uv}$ represents the depth correlation tensor, $b_{\mathrm{DepthG}}$ is the bias for our loss term, and $S_{hw,uv}$ represents the feature correlation tensor computed from the output features of the segmentation head $\mathcal{S}$ . By also using zero-clamping, we limit erroneous learning signals that aim to draw apart instances of the same class if they have large spatial differences. With this, we extend the STEGO loss so it can be formulated as follows: + +$$ +\mathcal {L} _ {\text {S T E G O + D e p t h G}} = \mathcal {L} _ {\text {S T E G O}} + \lambda_ {\text {D e p t h G}} \mathcal {L} _ {\text {D e p t h G}} \tag {7} +$$ + +with Depth-Feature Correlation weight $\lambda_{\mathrm{DepthG}}$ . By inducing depth knowledge during training without encoding the depth maps as part of the model input, our model can predict spatially informed segmentations on RGB images with its distilled knowledge and does not rely on depth to be available at test time. + +# 3.4. Depth-Guided Feature Sampling + +We also aim to make the feature sampling process informed by the spatial layout of the scene. To perform sampling in the 3D space, we transform the downsampled depth map $d(x_{i})$ into a point cloud with points $\{p_1,p_2,\dots ,p_n\}$ . On this point cloud, we apply farthest point sampling + +(FPS) [14], in an iterative fashion by always selecting the next point $p_k$ as the point with the maximum distance in 3D space with respect to the already sampled points $\{p_1, p_2, \dots, p_{k-1}\}$ . After having sampled $N^2$ points, we end up with a set of samples $\{p_1, p_2, \dots, p_{N^2}\}$ which are consequently converted to 2D sampling indices for the feature maps $f$ and $g$ . In contrast to the data-agnostic random sampling applied in STEGO, our feature selection process takes into account the geometry of the input scene and covers the spatial structure more equally. In our ablations in Section 5, we show this scene coverage from FPS of the depth space further increases the effectiveness of our Depth Feature Correlation loss, due to the increased diversity in selected 3D locations. We show a visual comparison between random and farthest point sampling in Figure 5. + +# 3.5. Guidance Scheduling + +While our Depth-Feature Correlation loss is effective at enriching the model's learning process with spatial information of the scene, we aim to alleviate the danger of it interfering with the learning of feature correlations during model training. We hypothesize that our model most greatly benefits from depth information in the beginning of training when its only knowledge is encoded in the features maps by the frozen ViT backbone. To give it a head start, we increase the weight of our Depth-Feature Correlation loss in the beginning and gradually decrease its influence during training. Vice versa, the distillation process in the feature space will increasingly emphasised as the training progresses. In this way, the network builds upon the already learned rough spa + +![](images/b2489f9a943cdf638050e6fdad377dc99fbf9e6cff4f93949ca8f4551da53e33.jpg) +Figure 3. Local Hidden Positives. We visualize the use of depth and attention maps for local hidden positives. For this visualization, we sample the respective propagation maps at the yellow patch in the center of the crops. We observe the depth map to have sharper borders and more consistent propagation values. We experiment with both propagation strategies in Section 5. + +tial structure of the scene achieved through our depth guidance. Therefore, we implement an exponential decay of the weight $\lambda_{\mathrm{DepthG}}$ and bias $b_{\mathrm{DepthG}}$ of our loss component. We ablate on the use of guidance scheduling in the appendix. + +# 3.6. Local Hidden Positives In 3D Space + +We further explore the combination of our method with Hidden Positives [35], which also builds upon STEGO. As we demonstrate as part of our ablation on the individual influence of our contributions in Section 5, our feature sampling method, implemented through farthest point sampling, is integral to the effectiveness of DepthG. Therefore, we decide not to replace it with the global hidden positives sampling from Seong et al. [35]. Instead, we implement a depth-informed variant of local hidden positives, 3D-LHP, where the loss for an individual patch is propagated to eight neighboring patches proportionally to their attention values obtained from the feature extractor. We modify this strategy by instead propagating the learning signal to the closest patches in 3D space proportional to their relative distances, using the point cloud described in Section 3.4. As visualized in Figure 3, we observe that our depth-based propagation map has sharper and more consistent surfaces. To propagate the learning signal to the selected patches, we follow Hidden Positives and mix the features coming from the segmentation head $S$ in proportion to their propagation values (point distances). The mixed representations are then fed through an additional projection head $\mathcal{P}$ . We then calculate $\mathcal{L}_{\mathrm{STEGO+DepthG}}$ again for the produced output features and combine it with the loss from the segmentation head output. + +# 4. Experiments + +Datasets and Models. We conduct experiments on the COCO-Stuff [4], Cityscapes [11] and Potsdam-3 datasets. COCO-Stuff contains a wide variety of real-world scenes. In our evaluation, we follow [16, 20, 35] and provide re + +sults on the coarse class split, COCO-Stuff 27. In contrast, Cityscapes contains traffic scenes from 50 cities from a driver-like viewpoint. Lastly, the Potsdam-3 dataset is composed of aerial, top-down images of the city of Potsdam. We use the DINO [7] backbones ViT-Small (ViT-S) and ViT-Base (ViT-B), which were pre-trained in a self-supervised manner on ImageNet-1k [12]. We choose the models with a patch size of $8 \times 8$ , since they have been shown to perform best for semantic segmentation due to the higher resolution of the resulting feature maps [16, 35]. In the result tables, we refer to DepthG models trained with our DepthFeature Correlation loss and FPS as Ours. Models that utilize 3D-LHP propagation with depth maps in addition are displayed as Ours w/ 3D-LHP. We compare our approach against competing methods which were never trained with human annotations, i.e. human labels or language supervision, neither in the feature extractor nor the segmentation training. + +Evaluation Protocols. Similar to STEGO and related work [16, 35], we evaluate our models in the unsupervised, clustering-based setting, as well as linear probing. Since the output of our model is a pixel-level map of features and not class labels, these features are consequently clustered. Following, the pseudo-labeled clusters are aligned with the ground truth labels through Hungarian matching [23, 24] across the entire validation dataset. To perform linear probing, an additional linear layer is added on top of the model and trained with cross-entropy loss to learn classification of the features. + +
SettingUnsupervisedLinear
MethodModelAcc.mIoUAcc.mIoU
IIC [20]R18+FPN21.86.744.58.4
PiCIE [10]R18+FPN48.113.854.213.9
PiCIE+H [10]R18+FPN50.014.454.814.8
DINO [7]ViT-S/828.711.368.633.9
ACSeg [26]ViT-S/816.4---
TransFGU [42]ViT-S/817.552.7--
STEGO + HP [35]ViT-S/857.224.675.642.7
STEGO [16]ViT-S/848.324.574.438.3
+ OursViT-S/856.325.673.738.9
+ Ours w/ 3D-LHPViT-S/855.126.773.937.8
DINO [7,16]ViT-B/830.59.666.829.4
DINOSAUR [34]*ViT-B/844.924.0--
STEGO [16]ViT-B/856.928.276.141.0
+ OursViT-B/858.629.075.541.6
+ +Table 1. Evaluation on COCO-Stuff 27. We report results on COCO-Stuff with 27 high-level classes. Overall, our method outperforms STEGO and HP on unsupervised segmentation with the ViT-B/8, while showing competitive performance for the ViT-S/8. *Results obtained without post-processing optimization. + +# 4.1. COCO-Stuff + +We present our evaluation on COCO-Stuff 27 in Table 1. For the ViT-S/8, our experiments show that Ours is able to improve upon STEGO in most metrics, with improved unsupervised accuracy by $+8.0\%$ and unsupervised mIoU increased by $+1.1\%$ . Ours w/ 3D-LHP further increases this mIoU delta to $+1.8\%$ , highlighting the effectiveness of our 3D-information propagation strategy in combination with DepthG. When comparing our approach to Hidden Positives, a method with more computational overhead, for the ViT-S/8, we show competitive performance for unsupervised accuracy and outperform their approach by $+1.0\%$ on unsupervised mIoU with Ours and $+1.7\%$ with Ours w/ 3D-LHP. When using the DINO ViT-B/8 encoder, our approach again outperforms STEGO, as well as all other presented methods on unsupervised metrics. Most notably, we are able to increase the unsupervised mIoU by $+0.8\%$ . + +# 4.2. Cityscapes + +We further evaluate our approach on the Cityscapes dataset [11], consisting of various scenes from 50 different cities. We follow the training setting from STEGO and, contrary to all other datasets, do not sample point-wise but use the full feature map for our learning process along with the full depth map. As can be seen in Table 2, our method significantly outperforms STEGO as well as Hidden Positives. For unsupervised mIoU, while Hidden Positives decreased performance compared to STEGO, we observe that our approach to achieves a $+2.1\%$ increase. Similarly, we report state-of-the-art performance in accuracy, building upon Hidden Positives' already impressive improvements over STEGO and outperforming it by $+2.1\%$ . + +# 4.3. Potsdam + +Our model is further evaluated on the Potsdam-3 dataset, containing aerial images of the German city of Potsdam. Contrary to the other benchmarks, which contain images in a first-person perspective, Potsdam-3 contains only birds-eye-view images, a perspective that is considered out-of-distribution for the monocular depth estimator. Despite this inherent limitation of our approach for aerial data, we are able to demonstrate relatively commendable performance in Table 3 by improving STEGO's performance and reporting state-of-the-art performance for the ViT-S backbone. In contrast, Hidden Positives [35] use a ViT-B/8, and with roughly twice the parameters as ours, reach an accuracy of $82.4\%$ . We present a visual overview of the predicted Potsdam depth maps in the appendix. + +# 4.4. Qualitative Results + +We present qualitative results of our method in Figure 4 and compare with segmentation maps from STEGO. On + +
MethodModelU. AccU. mIoU
IIC [20]R18+FPN47.96.4
PiCIE [10]R18+FPN65.612.3
DINO [7]ViT-B/843.611.8
STEGO + HP [35]ViT-B/879.518.4
STEGO [16]ViT-B/873.221.0
+ OursViT-B/881.623.1
+ +Table 2. Results on Cityscapes. We report unsupervised accuracy and mIoU on Cityscapes. Our method outperforms both STEGO variants by substantial margins. Notably, our method is the first to improve upon unsupervised mIoU. + +
MethodModelU. Acc.
CC [19]VGG1163.9
DeepCluster [6]VGG1141.7
IIC [20]VGG1165.1
DINO [7,21]ViT-S/871.3
STEGO [16,21]ViT-S/877.0
+ OursViT-S/880.4
+ Ours w/ 3D-LHPViT-S/880.4
+ +Table 3. Results on Potsdam. We report unsupervised accuracy on the Potsdam dataset. Our method is able to improve upon STEGO. We hypothesize that with a zero-shot depth estimator more suitable for aerial images, the results for our method could further improve. + +
MethodU. mIoU.U. Acc
STEGO [16]24.548.3
+ Depth-Feature Correlation (1)24.751.2
+ FPS (2)24.649.1
+ 3D-LHP (3)25.248.5
+ Ours (1, 2)25.656.3
+ Ours w/ 3D-LHP (1, 2, 3)26.755.1
+ +Table 4. Effect of our contributions. We compare our individual contributions and the combination of all contributions. + +multiple occasions, our depth guidance reduces erroneous predictions from the model caused by visual irritations in the pixel space. In the example of the boy with the baseball bat in Figure 4a, false classifications from STEGO are caused by shadows on the ground. Our model is able to correct this. Furthermore, it goes beyond the noisy label and also correctly classifies the glimpse of a plant that can be seen through a hole in the background. This is an indication that our model does not overfit to the depth map, since this visual cue is only observable in the pixel space, but not the depth map. + +![](images/59e641edfed74e60fae016a3268151762a0cad18f01b708fda77b3a1348329b6.jpg) +(a) COCO-Stuff + +![](images/652e6f12fc2a495b4f2261e12088cab65129cb38427224d8bcb0e4c49ee744fe.jpg) +(b) Cityscapes +Figure 4. Qualitative results. We show qualitative differences for plain STEGO compared to STEGO with our depth guidance, using ViT-S models for COCO and ViT-B for Cityscapes. Where STEGO struggles to differentiate instances, our model is able to correct this and successfully separates them for segmentation. In the case of the building in (a), our method alleviates visual irritations from the pixel space and corrects the segmentation of the building. In (b), our model is able to better handle visual inconsistencies from shadows. + +# 5. Ablations + +Individual Influence. We investigate the effect of our technical contributions on training our model with a ViT-S/8 backbone on COCO-Stuff 27. Our observations in Table 4 show that our Depth-Feature Correlation loss itself already improves the performance of STEGO. This improvement is further increased through the use of FPS, which enables us to sample the depth space more meaningfully and therefore encourages more diversity in the depth correlation tensor $D_{hw,uv}$ . Intuitively, this sampling diversity significantly amplifies our Depth-Feature Correlation for aligning the feature space with the depth space. We provide a visual comparison to random sampling in Figure 5 and additional illustrations in the appendix. FPS retrieves more diverse locations and specifically selects locations with depth discontinuities. This naturally benefits the Depth-Feature Correlation to learn sharper edges in this area for the output segmentation. Adding local hidden positives with depth maps further increases the unsupervised mIoU, while slightly lowering the accuracy. + +Source Of Depth Maps. We investigate the effect of different monocular depth estimators to generate the depth maps used to train our model. In our experiments, we consider three options: The previously mentioned ZoeDepth [3], Kick Back & Relax (KBR) [36] which uses self-supervision to learn depth from Slow-TV videos, as well as MiDaS [32], the base model to ZoeDepth. Empirically, ZoeDepth produces the most accurate zero-shot + +monodepth results across indoor and outdoor datasets, followed by MiDaS and KBR [3,36]. We provide a qualitative comparison in the appendix. To evaluate the influence of the depth map quality on the performance of our model, we first generate depth maps for COCO-Stuff 27 using each of the introduced models. We then train our model with a ViT-S/8 backbone and report the results in Table 5. We observe that depth maps from ZoeDepth work best with our method, while the model trained with MiDaS depth has an edge over the KBR counterpart. + +![](images/c4f3c4efd1e84904ac46fe56ea15f26a851f17e279e0d2b0c10efd522e629a07.jpg) +Figure 5. Random vs. Farthest Point Sampling. We observe that random sampling can miss entire structures like trees in the first top and the plane in the bottom row. In contrast, our method meaningfully samples the depth space and selects locations across the different structures and at depth edges. We show further illustrations of FPS in the appendix. + +![](images/d330e6a3c0ef13c3e80d3f32b1e61f8fa42269c0f7e0760153deb52c1c5d1c8f.jpg) +Figure 6. Failure cases. We show cases where our model fails to correctly segment and classify the scene. The top row is a prime example where the difference in depth is correctly distilled, though the model fails to correctly classify the snow region. + +Signal Propagation With Local Hidden Positives. We ablate the implementation of LHP along with different propagation strategies. As described in Section 3.6, our method takes advantage of the depth information of the scene to propagate the learning signal to patches which are nearby in 3D space. We further implement utilizing the attention map from the DINO backbone and propagate proportionally to their values i.e., the approach utilized in Hidden Positives [35]. While, in Table 6, we show that our approach benefits from signal propagation with 3D-LHP, applying LHP with Attention leads to lower performance. We speculate 3D-LHP is more suitable for our approach since, for a given location, the signal from our Depth-Feature Correlation loss is calculated w.r.t. the depth at this sample. Propagating to locations with the same depth does not corrupt this signal, though this can happen with LHP (Attention), since it does not consider depth for propagation. + +Computational Cost. Our method only leads to an insignificant increase in runtime versus the baseline STEGO model, since we solely guide the loss as well as the feature sampling and only for Ours w/ 3D-LHP add an additional small segmentation head. In contrast, the competing method Hidden Positives [35] relies on a computationally more expensive process to select features and introduces an additional segmentation head to fill their task-specific feature pool. To keep our computational overhead low, we make use of a pre-trained monocular depth estimation network with impressive zero-shot capabilities. We consider task specific training of the depth estimator not a necessity, since the model has zero-shot capabilities that generalize well to different scenes and domains. Therefore, in our experiments on a diverse array of scenes, we do not re-train the depth estimator, and consider the additional computational cost for generating the depth maps to be negligible. + +
MethodTrained withU. Acc.U. mIoU
ZoeDepth [3]Sensor Depth56.325.6
MiDaS [32]Sensor Depth53.025.0
KBR [36]Self-Supervision50.623.1
+ +Table 5. Different depth map sources. We experiment with different monocular depth estimators which were trained with either sensor depth or self-supervision. Overall, the model trained with depth maps from ZoeDepth performs best on COCO-Stuff 27. + +
LHPPropagation StrategyU. Acc.U. mIoU
X-56.325.6
✓3D-LHP (Depth)55.126.7
✓LHP (Attention)52.624.5
+ +Table 6. LHP Ablation. We compare the use of depth and attention for local hidden positives. We find that using depth improves unsupervised mIoU, while we find that using attention does not improve our method. + +# 6. Limitations + +While we have demonstrated our method's effectiveness for many real-world cases, our method's applicability is limited in settings unsuitable for depth estimation, such as slices of CT scans and other medical data domains. Furthermore, the experiments on Potsdam-3 have shown, our method can improve unsupervised semantic segmentation despite suboptimal viewing perspectives for the monocular depth estimator. We assume the scenario of aerial images represents a rare case where our method would profit from a domain-specific monocular depth estimator. We also present failure cases of our model in Figure 6. + +# 7. Conclusion & Future Work + +In this work, we have presented a novel method to induce spatial knowledge of the scene into our model for unsupervised semantic segmentation. We have proposed to correlate the feature space with the depth space and use the 3D information to more meaningfully sample features in a spatially informed way. Furthermore, we have demonstrated that these contributions produce state-of-the-art performance on many real-world datasets and thus foster the progress in unsupervised segmentation. The applicability of our approach for other tasks is further to be explored, since we hypothesize it can be useful beyond unsupervised segmentation as part of other contrastive processes. We consider this to be a promising direction for future work. Furthermore, it remains to be investigated what kind of information could be useful in domains where depth is not an obviously meaningful signal, like medical data. + +# References + +[1] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3 +[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 7, 8 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5 +[5] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1129-1139, 2022. 1, 2 +[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6 +[8] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2624-2632, 2019. 1 +[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1 +[10] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 2, 5, 6 +[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 5, 6 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 + +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1 +[14] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 4 +[15] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1 +[16] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6 +[17] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2 +[18] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11130-11140, 2021. 1, 2 +[19] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H Adelson. Learning visual groups from co-occurrences in space and time. arXiv preprint arXiv:1511.06811, 2015. 6 +[20] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9865-9874, 2019. 1, 2, 5, 6 +[21] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of stego for safe unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3788-3797, 2023. 6 +[22] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011, 2, 3 +[23] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[24] Harold W Kuhn. Variants of the hungarian method for assignment problems. Naval research logistics quarterly, 3(4):253-258, 1956. 5 +[25] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar + +guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 3 +[26] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 5 +[27] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1 +[29] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2):129-137, 1982. 2 +[30] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 +[32] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 3, 7, 8 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1 +[34] M Seitzer, M Horn, A Zadaianchuk, D Zietlow, T Xiao, C Simon-Gabriel, T He, Z Zhang, B Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In International Conference on Learning Representations (ICLR), 2023. 2, 5 +[35] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19540-19549, 2023. 1, 2, 5, 6, 8 +[36] Jaime Spencer, Chris Russell, Simon Hadfield, and Richard Bowden. Kick back & relax: Learning to reconstruct the world by watching slowtv. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 7, 8 +[37] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019. 1 +[38] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8515-8525, 2021. 1, 2 +[39] Risheng Wang, Tao Lei, Ruixia Cui, Bingtao Zhang, Hongying Meng, and Asoke K Nandi. Medical image segmentation using deep learning: A survey. IET Image Processing, 16(5):1243-1267, 2022. 1 +[40] Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. Multimodal token fusion for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12186-12195, 2022. 2 +[41] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1 +[42] Zhaoyuan Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, and Rong Jin. Transfgu: a top-down approach to fine-grained unsupervised semantic segmentation. In European conference on computer vision, pages 73-89. Springer, 2022. 5 +[43] Xiaowen Ying and Mooi Choo Chuah. Uctnet: Uncertainty-aware cross-modal transformer network for indoor rgb-d semantic segmentation. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2 +[44] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 1 +[45] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1 \ No newline at end of file diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/images.zip b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1c8df33cfe1075f9a3c6df63c17577be15e5594 --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:238925fe223e605cffbbc1599a69122f614d6caef5bf90fbef987d86954c1dce +size 509202 diff --git a/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/layout.json b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e5ab66de5d643bbda658e290d9c218d9e3a2e5ee --- /dev/null +++ b/2024/Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling/layout.json @@ -0,0 +1,7034 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 478, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 478, + 140 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 478, + 140 + ], + "type": "text", + "content": "Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 161, + 172, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 161, + 172, + 190 + ], + "spans": [ + { + "bbox": [ + 95, + 161, + 172, + 190 + ], + "type": "text", + "content": "Leon Sick Ulm University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 161, + 277, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 161, + 277, + 190 + ], + "spans": [ + { + "bbox": [ + 201, + 161, + 277, + 190 + ], + "type": "text", + "content": "Dominik Engel Ulm University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 161, + 392, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 392, + 188 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 392, + 188 + ], + "type": "text", + "content": "Pedro Hermosilla TU Vienna" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 421, + 161, + 497, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 161, + 497, + 190 + ], + "spans": [ + { + "bbox": [ + 421, + 161, + 497, + 190 + ], + "type": "text", + "content": "Timo Ropinski \nUlm University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 257, + 290, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 290, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 290, + 497 + ], + "type": "text", + "content": "Traditionally, training neural networks to perform semantic segmentation requires expensive human-made annotations. But more recently, advances in the field of unsupervised learning have made significant progress on this issue and towards closing the gap to supervised algorithms. To achieve this, semantic knowledge is distilled by learning to correlate randomly sampled features from images across an entire dataset. In this work, we build upon these advances by incorporating information about the structure of the scene into the training process through the use of depth information. We achieve this by (1) learning depth-feature correlation by spatially correlating the feature maps with the depth maps to induce knowledge about the structure of the scene and (2) exploiting farthest-point sampling to more effectively select relevant features by utilizing 3D sampling techniques on depth information of the scene. Finally, we demonstrate the effectiveness of our technical contributions through extensive experimentation and present significant improvements in performance across multiple benchmark datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "content": "Semantic segmentation plays a critical role in many of today's vision systems in a multitude of domains. These include, among others, autonomous driving [15], medical applications [33, 39], and many more [9, 27, 41, 45]. Until recently, the main body of research in this area was focused on supervised models that require a large amount pixel-level annotations for training. Not only is sourcing this image data often a labor intensive process, but also annotating the large datasets required for good performance comes at a high price. Several benchmark datasets report their annotation times. For example, the MS COCO dataset [28] required more than " + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "28\\mathrm{K}" + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "content": " hours of human annotation for around " + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "164\\mathrm{K}" + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "content": " images, and annotating a single image in the Cityscapes dataset [11] took 1.5 hours on average." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 614 + ], + "type": "text", + "content": "These costs have triggered the advent of unsupervised semantic segmentation [10, 16, 20, 35], which aims to remove the need for labeled training data in order to train segmentation models. Recently, work by Hamilton et al. [16] has accelerated the progress towards removing the need for labels to achieve good results on semantic segmentation tasks. Their model, STEGO, uses a DINO-pretrained [7] Vision Transformer (ViT) [13] to extract features that are then distilled across the entire dataset to learn semantically relevant features, using a contrastive learning approach. The to-be-distilled features are sampled randomly from feature maps produced from the same image, k-NN matched images as well as other negative images. Seong et al. [35] build on this process by trying to identify features that are most relevant to the model by discovering hidden positives. Their work exposes an inefficiency of random sampling in STEGO as hidden positives sampling leads to significant improvements. However, both approaches only operate on the pixel space and therefore fail to take into account the spatial layout of the scene. Not only do we humans perceive the world in 3D, but also previous work [5, 8, 18, 37, 38, 44] has shown that supervised semantic segmentation can benefit greatly from spatial information during training. Inspired by these observations, we propose to incorporate spatial information in the form of depth maps into the STEGO training process. Depth is considered a product of vision and does not provide a labeled training signal. To obtain depth information for the benchmark image datasets in our evaluations, we make use of an off-the-shelf zero-shot monocular depth estimator to obtain spatial information of the scene. This allows us to incorporate the depth information without the need for human annotations or sensor ground truth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "With our method, " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "DepthG" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", we propose to (1) guide the model to learn a rough spatial layout of the scene, since we hypothesize this will aid the network in differentiating objects much better. We achieve this by extending the contrastive process to the spatial dimension: We do not limit the model to learning only Feature-Feature Correlations, but also Depth-Feature Correlations. Through this process, the model is guided towards pulling apart the features with high" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3637" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 289, + 300 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 289, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 289, + 300 + ], + "type": "image", + "image_path": "40bcafca6e9638f4a9deefdaae77ca1abd41bf8d4c9de28e7095e01c1291f39b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 306, + 288, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 373 + ], + "type": "text", + "content": "Figure 1. Guiding the feature space for unsupervised segmentation with depth information. Our intuition behind the proposed approach is simple: For locations in the 3D space with a low distance, we guide the model to map their features closer together. Vice versa, the features are learned to be drawn apart in feature space if their distance in the metric space is large." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 439, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 477 + ], + "type": "text", + "content": "distances in 3D space, as well as mapping them closer together if their distance is low depth space. Figure 1 visualizes this process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "With the information about the spatial layout of the scene present, we furthermore propose to (2) spatially inform our feature sampling process by utilizing Farthest-Point Sampling (FPS) [14, 31] on the depth map, which equally samples scenes in 3D. We show these techniques in combination make our approach to unsupervised segmentation highly effective, since in our evaluations across multiple benchmark datasets, we demonstrate state-of-the-art performance. We include a short video presenting our method as part of the supplementary materials. To the best of our knowledge, we are the first to propose a mechanism to incorporate 3D knowledge of the scene into unsupervised learning for 2D images without encoding depth maps as part of the network input. This design aspect of our method alleviates the risk of the model developing an input dependency. Therefore, our approach does not rely on depth information during inference and is not affected by availability or quality of such depth information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 91, + 505, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 505, + 104 + ], + "type": "text", + "content": "2.1. Unsupervised Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 110, + 547, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 459 + ], + "type": "text", + "content": "Recent works [10, 16, 20, 35] have attempted to tackle semantic segmentation without the use of human annotations. Ji et al. [20] propose IIC, a method that aims to maximize the mutual information between different augmented versions of an image. PiCIE, published by Cho et al. [10], introduces an inductive bias based on the invariance to photometric transformations and equivariance to geometric manipulations. DINO [7] often serves as a critical component to unsupervised segmentation algorithms, since the self-supervised pre-trained ViT can produce semantically relevant features. Recent work by Seitzer et al. [34] builds upon this ability by training a model with slot attention [30] to reconstruct the feature maps produced by DINO from the different slots. The features of their object-centric model are clustered with k-means [29] where each slot is associated with a cluster. In their 2021 work, Hamilton et al. [16] have also built upon DINO features by introducing a feature distillation process with features from the same image, k-NN retrieved examples as well as random other images from the dataset. Their learned representations are finally clustered and refined with a CRF [22] for semantic segmentation. While STEGO's feature selection process is random, Seong et al. [35] introduce a more effective sampling strategy by discovering hidden positives. During training, they form task-agnostic and task-specific feature pools. For an anchor feature, they then compute the maximum similarity to any of the pool features and sample locations in the image with greater similarity than the determined value. A more detailed introduction to STEGO is provided in Section 3.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 467, + 491, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 491, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 491, + 480 + ], + "type": "text", + "content": "2.2. Depth For Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "Previous research [5, 17, 18, 38, 40, 43] has sought to incorporate depth for semantic segmentation in different settings. Wang et al. [38] propose to use depth for adapting segmentation models to new data domains. Their method adds depth estimation as an auxiliary task to strengthen the prediction of segmentation tasks. Furthermore, they approximate the pixel-wise adaption difficulty from source to target domain through the use of depth decoders. Work by Hoyer et al. [18] explores three further strategies of how depth can be useful for segmentation. First, they propose using a shared backbone to share learning features for segmentation and self-supervised depth estimation, similar to Wang et al. [38]. Second, they use depth maps to introduce a data augmentation that is informed by the structure of the scene. And lastly, they detail the integration of depth into an active learning loop as part of a student-teacher setup. Another work by Hou et al. [17] incorporates depth information into a pre-training algorithm with the aim to learn better representations for semantic segmentation. In their" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3638" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "work, Mask3D, they propose to incorporate depth in the form of a 3D prior by formulating a reconstruction task that operates on masked RGB and depth patches, enabling them to learn more useful features for semantic segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 102, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 102, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 102, + 144 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 153, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 287, + 190 + ], + "type": "text", + "content": "In the following, we detail our proposed method for guiding unsupervised segmentation with depth information. An overview of our approach is presented in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 198, + 126, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 126, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 126, + 210 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": "Our approach builds upon work by Hamilton et al. [16]. In their work, each image is 5-cropped and k-NN correspondences between these images are calculated using the DINO ViT [7]. Generally, STEGO uses a feature extractor " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:\\mathbb{R}^{3\\times H_{\\mathrm{in}}\\times W_{\\mathrm{in}}}\\to \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " with input image height " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": ", to calculate a feature map " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "f\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " and feature dimension " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " from the input image. These features are then further encoded by a segmentation head " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{S}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{Q\\times H\\times W}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " to calculate the code space " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "s\\in \\mathbb{R}^{Q\\times H\\times W}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " with code dimension " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": ". With the goal of forming compact clusters and amplifying the correlation of the learned features, let " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "f_{i}\\coloneqq \\mathcal{F}(x_{i})" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "f_{j}\\coloneqq \\mathcal{F}(x_{j})" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " be feature maps for a given input pair of " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": ", which are then used to calculate " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "s_i\\coloneqq S(f_i)" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "s_j\\coloneqq S(f_j)" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " from the segmentation head " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": ". In practice, STEGO samples " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "N^2" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " vectors from the feature map during training. Hamilton et al. [16] introduce the concept of constructing the feature correspondence tensor " + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\pmb {F}\\in \\mathbb{R}^{H\\times W\\times H\\times W}" + }, + { + "bbox": [ + 46, + 217, + 289, + 443 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 452, + 287, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 452, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 287, + 482 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {h w, u v} = \\frac {f _ {i} ^ {h w} \\cdot f _ {j} ^ {u v}}{\\| f _ {i} ^ {h w} \\| \\| f _ {j} ^ {u v} \\|} \\tag {1}", + "image_path": "38a50a56e3926cbb4351445519c8d794ea29fd6396278328a3bb130ec4583db0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "text", + "content": " denotes the dot product. Using the same formula we obtain " + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "inline_equation", + "content": "s_i, s_j" + }, + { + "bbox": [ + 46, + 491, + 287, + 526 + ], + "type": "text", + "content": ". Consequently, the feature correlation loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 535, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 535, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 71, + 535, + 287, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {C o r r}} := - \\sum_ {h w, u v} \\left(\\boldsymbol {F} _ {h w, u v} - b\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {2}", + "image_path": "c1f54eb0352d6f9d03939b0be84b886473c3c28ccbf9b3f43e94bb34232a9e37.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": " is a scalar bias hyperparameter. Empirical evaluations from STEGO have shown that applying spatial centering to the feature correlation loss along with zero-clamping further improves performance [16]. These correlations are calculated for two crops from the same image " + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{self}})" + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": " and one from a different but similar image, determined by the k-NN correspondence pre-processing " + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{knn}})" + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": ". Finally, negative images are sampled randomly " + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{random}})" + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": ". The final loss is a weighted sum of the different losses where each of them has their individual weight " + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 571, + 287, + 690 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 63, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S T E G O}} = \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k n n}} \\mathcal {L} _ {\\text {k n n}} + \\lambda_ {\\text {r a n d o m}} \\mathcal {L} _ {\\text {r a n d o m}} \\tag {3}", + "image_path": "a744d567f3d69eef3053be2f790709ca52f5f654eaf2d8424e9ad91a5df20b55.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "After training, the inferred feature maps for a test image are clustered using k-means and refined with a conditional random field (CRF) [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 114, + 437, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 114, + 437, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 437, + 126 + ], + "type": "text", + "content": "3.2. Depth Map Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": "Since in many cases, depth information about the scene is not readily available, we make use of recent progress in the field of monocular depth estimation [1-3, 25, 32] to obtain depth maps from RGB images. Recently, methods from this field have made significant advances in zero-shot depth estimation, i.e. predicting depth values for scenes from data domains not seen during training. This property makes them especially suitable for our method, since it enables us to obtain high-quality depth predictions for a wide variety of data domains without ever re-training the depth network. This property also limits the computational cost for our method. For our method, we experiment with different state-of-the-art monocular depth estimators, detailed in Section 5, and found ZoeDepth [3] to perform best in our evaluations. Given a cropped RGB image " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": ", we use the monocular depth estimator " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " together with average pooling to predict depth " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "d_{i}\\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " at feature resolution:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 342, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 342, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 386, + 342, + 545, + 354 + ], + "type": "interline_equation", + "content": "d _ {i} = \\operatorname {p o o l} \\left(\\mathcal {M} \\left(x _ {i}\\right)\\right) \\tag {4}", + "image_path": "1b1f83b1d324bbf0bc0157db492cad6ac6f138496c9d7c4f0ea4d6f1358164c0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 396 + ], + "type": "text", + "content": "The average pooling operation is used to match the dimensions of the feature map, which is required to sample nonoverlapping locations at the patch resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 402, + 479, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 479, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 479, + 414 + ], + "type": "text", + "content": "3.3. Depth-Feature Correlation Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "type": "text", + "content": "With our Depth-Feature Correlation loss, we aim to enforce spatial consistency in the feature map by transferring the distances from the depth map to the latent feature space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 456, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 696 + ], + "type": "text", + "content": "In contrastive learning, the network is incentivized to decrease the distance in feature space for similar instances, therefore learning to map their latent representations closer together. Likewise, different instances are drawn further apart in feature distance. We assume the same concept to be true in 3D space: The spatial distance between two points from the same depth plateau is smaller, while the distance between a point in the foreground and one in the background is larger. Since, in both spaces, the concept of measuring difference is represented by the distance between two points, we propose to align them through our concept of Depth-Feature Correlation: For large distances in the 3D space, we encourage the network to produce vectors that are further apart, and vice versa. With this, we induce the model with knowledge about the spatial structure of the scene, enabling it to better differentiate between objects. To achieve this we construct the depth correspondence tensor similar to the feature correspondence from Equation 1. The depth correspondence tensor " + }, + { + "bbox": [ + 304, + 456, + 545, + 696 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 456, + 545, + 696 + ], + "type": "text", + "content": " is computed from the depths of two different image crops as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {D} _ {h w, u v} = d _ {i} ^ {h w} d _ {j} ^ {u v}, \\tag {5}", + "image_path": "3d832ed91fa0ba3daab3e1d41c342393b71e6cefebcad552ffa10f9fa7665c4c.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3639" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 334, + 248 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 334, + 248 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 334, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 334, + 248 + ], + "type": "image", + "image_path": "4814db22c9dc63ad43748ca08fbc5988ddb67f17e0bed4334757c1e413bd4902.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 135, + 455, + 147 + ], + "lines": [ + { + "bbox": [ + 427, + 135, + 455, + 147 + ], + "spans": [ + { + "bbox": [ + 427, + 135, + 455, + 147 + ], + "type": "inline_equation", + "content": "D_{hw,uv}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "type": "text", + "content": "Figure 2. Overview of the DepthG training process. After 5-cropping the image, each crop is encoded by the DINO-pretrained ViT " + }, + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "type": "text", + "content": " to output a feature map. Using farthest point sampling (FPS), we sample the 3D space equally and convert the coordinates to select samples in the feature map. The sampled features are further transformed by the segmentation head " + }, + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 255, + 547, + 333 + ], + "type": "text", + "content": ". For both feature maps, the correlation tensor is computed. Following, we sample the depth map at the coordinates obtained by FPS and compute a correlation tensor in the same fashion. Finally, we compute our Depth-Feature Correlation loss and combine it with the feature distillation loss from STEGO. We guide the model to learn depth-feature correlation for crops of the same image, while the feature distillation loss is also applied to k-NN-selected and random images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "content": " represent the pixel positions in the depth maps " + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "inline_equation", + "content": "d_{j}" + }, + { + "bbox": [ + 46, + 346, + 288, + 381 + ], + "type": "text", + "content": " respectively. Together with the zero clamping, our Depth-Feature Correlation loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D e p t h G}} = - \\sum_ {h w, u v} \\left(\\boldsymbol {D} _ {h w, u v} - b _ {\\text {D e p t h G}}\\right) \\max \\left(\\boldsymbol {S} _ {h w, u v}, 0\\right) \\tag {6}", + "image_path": "5e3dd6e6524222716382de9204ecefd5293569aff09cb2084409122edb09b3d4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "inline_equation", + "content": "D_{hw,uv}" + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "content": " represents the depth correlation tensor, " + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{DepthG}}" + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "content": " is the bias for our loss term, and " + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "inline_equation", + "content": "S_{hw,uv}" + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "content": " represents the feature correlation tensor computed from the output features of the segmentation head " + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 428, + 287, + 523 + ], + "type": "text", + "content": ". By also using zero-clamping, we limit erroneous learning signals that aim to draw apart instances of the same class if they have large spatial differences. With this, we extend the STEGO loss so it can be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 533, + 287, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 533, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 81, + 533, + 287, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S T E G O + D e p t h G}} = \\mathcal {L} _ {\\text {S T E G O}} + \\lambda_ {\\text {D e p t h G}} \\mathcal {L} _ {\\text {D e p t h G}} \\tag {7}", + "image_path": "3ce75a7e9c5f1644cfdf12c195b7ffe3c2a69f13a169f5287f1cc2cb928c3d27.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 627 + ], + "type": "text", + "content": "with Depth-Feature Correlation weight " + }, + { + "bbox": [ + 46, + 555, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{DepthG}}" + }, + { + "bbox": [ + 46, + 555, + 287, + 627 + ], + "type": "text", + "content": ". By inducing depth knowledge during training without encoding the depth maps as part of the model input, our model can predict spatially informed segmentations on RGB images with its distilled knowledge and does not rely on depth to be available at test time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 224, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 224, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 224, + 648 + ], + "type": "text", + "content": "3.4. Depth-Guided Feature Sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "We also aim to make the feature sampling process informed by the spatial layout of the scene. To perform sampling in the 3D space, we transform the downsampled depth map " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "d(x_{i})" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": " into a point cloud with points " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{p_1,p_2,\\dots ,p_n\\}" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": ". On this point cloud, we apply farthest point sampling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": "(FPS) [14], in an iterative fashion by always selecting the next point " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": " as the point with the maximum distance in 3D space with respect to the already sampled points " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "\\{p_1, p_2, \\dots, p_{k-1}\\}" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": ". After having sampled " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "N^2" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": " points, we end up with a set of samples " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "\\{p_1, p_2, \\dots, p_{N^2}\\}" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": " which are consequently converted to 2D sampling indices for the feature maps " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 346, + 547, + 526 + ], + "type": "text", + "content": ". In contrast to the data-agnostic random sampling applied in STEGO, our feature selection process takes into account the geometry of the input scene and covers the spatial structure more equally. In our ablations in Section 5, we show this scene coverage from FPS of the depth space further increases the effectiveness of our Depth Feature Correlation loss, due to the increased diversity in selected 3D locations. We show a visual comparison between random and farthest point sampling in Figure 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 537, + 429, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 429, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 429, + 551 + ], + "type": "text", + "content": "3.5. Guidance Scheduling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "While our Depth-Feature Correlation loss is effective at enriching the model's learning process with spatial information of the scene, we aim to alleviate the danger of it interfering with the learning of feature correlations during model training. We hypothesize that our model most greatly benefits from depth information in the beginning of training when its only knowledge is encoded in the features maps by the frozen ViT backbone. To give it a head start, we increase the weight of our Depth-Feature Correlation loss in the beginning and gradually decrease its influence during training. Vice versa, the distillation process in the feature space will increasingly emphasised as the training progresses. In this way, the network builds upon the already learned rough spa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 287, + 175 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 175 + ], + "type": "image", + "image_path": "b2489f9a943cdf638050e6fdad377dc99fbf9e6cff4f93949ca8f4551da53e33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 251 + ], + "type": "text", + "content": "Figure 3. Local Hidden Positives. We visualize the use of depth and attention maps for local hidden positives. For this visualization, we sample the respective propagation maps at the yellow patch in the center of the crops. We observe the depth map to have sharper borders and more consistent propagation values. We experiment with both propagation strategies in Section 5." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "text", + "content": "tial structure of the scene achieved through our depth guidance. Therefore, we implement an exponential decay of the weight " + }, + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{DepthG}}" + }, + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{DepthG}}" + }, + { + "bbox": [ + 46, + 271, + 288, + 321 + ], + "type": "text", + "content": " of our loss component. We ablate on the use of guidance scheduling in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 328, + 236, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 236, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 236, + 342 + ], + "type": "text", + "content": "3.6. Local Hidden Positives In 3D Space" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "text", + "content": "We further explore the combination of our method with Hidden Positives [35], which also builds upon STEGO. As we demonstrate as part of our ablation on the individual influence of our contributions in Section 5, our feature sampling method, implemented through farthest point sampling, is integral to the effectiveness of DepthG. Therefore, we decide not to replace it with the global hidden positives sampling from Seong et al. [35]. Instead, we implement a depth-informed variant of local hidden positives, 3D-LHP, where the loss for an individual patch is propagated to eight neighboring patches proportionally to their attention values obtained from the feature extractor. We modify this strategy by instead propagating the learning signal to the closest patches in 3D space proportional to their relative distances, using the point cloud described in Section 3.4. As visualized in Figure 3, we observe that our depth-based propagation map has sharper and more consistent surfaces. To propagate the learning signal to the selected patches, we follow Hidden Positives and mix the features coming from the segmentation head " + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "text", + "content": " in proportion to their propagation values (point distances). The mixed representations are then fed through an additional projection head " + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "text", + "content": ". We then calculate " + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{STEGO+DepthG}}" + }, + { + "bbox": [ + 46, + 347, + 289, + 636 + ], + "type": "text", + "content": " again for the produced output features and combine it with the loss from the segmentation head output." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 128, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 128, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 128, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Datasets and Models. We conduct experiments on the COCO-Stuff [4], Cityscapes [11] and Potsdam-3 datasets. COCO-Stuff contains a wide variety of real-world scenes. In our evaluation, we follow [16, 20, 35] and provide re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "sults on the coarse class split, COCO-Stuff 27. In contrast, Cityscapes contains traffic scenes from 50 cities from a driver-like viewpoint. Lastly, the Potsdam-3 dataset is composed of aerial, top-down images of the city of Potsdam. We use the DINO [7] backbones ViT-Small (ViT-S) and ViT-Base (ViT-B), which were pre-trained in a self-supervised manner on ImageNet-1k [12]. We choose the models with a patch size of " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ", since they have been shown to perform best for semantic segmentation due to the higher resolution of the resulting feature maps [16, 35]. In the result tables, we refer to DepthG models trained with our DepthFeature Correlation loss and FPS as Ours. Models that utilize 3D-LHP propagation with depth maps in addition are displayed as Ours w/ 3D-LHP. We compare our approach against competing methods which were never trained with human annotations, i.e. human labels or language supervision, neither in the feature extractor nor the segmentation training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 449 + ], + "type": "text", + "content": "Evaluation Protocols. Similar to STEGO and related work [16, 35], we evaluate our models in the unsupervised, clustering-based setting, as well as linear probing. Since the output of our model is a pixel-level map of features and not class labels, these features are consequently clustered. Following, the pseudo-labeled clusters are aligned with the ground truth labels through Hungarian matching [23, 24] across the entire validation dataset. To perform linear probing, an additional linear layer is added on top of the model and trained with cross-entropy loss to learn classification of the features." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 460, + 545, + 641 + ], + "blocks": [ + { + "bbox": [ + 308, + 460, + 545, + 641 + ], + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 641 + ], + "type": "table", + "html": "
SettingUnsupervisedLinear
MethodModelAcc.mIoUAcc.mIoU
IIC [20]R18+FPN21.86.744.58.4
PiCIE [10]R18+FPN48.113.854.213.9
PiCIE+H [10]R18+FPN50.014.454.814.8
DINO [7]ViT-S/828.711.368.633.9
ACSeg [26]ViT-S/816.4---
TransFGU [42]ViT-S/817.552.7--
STEGO + HP [35]ViT-S/857.224.675.642.7
STEGO [16]ViT-S/848.324.574.438.3
+ OursViT-S/856.325.673.738.9
+ Ours w/ 3D-LHPViT-S/855.126.773.937.8
DINO [7,16]ViT-B/830.59.666.829.4
DINOSAUR [34]*ViT-B/844.924.0--
STEGO [16]ViT-B/856.928.276.141.0
+ OursViT-B/858.629.075.541.6
", + "image_path": "86303acc50b22f95b1a7afb0cb943bc2bc074cf870c395bdf9ab47569c3b3f59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 648, + 547, + 706 + ], + "lines": [ + { + "bbox": [ + 304, + 648, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 648, + 547, + 706 + ], + "type": "text", + "content": "Table 1. Evaluation on COCO-Stuff 27. We report results on COCO-Stuff with 27 high-level classes. Overall, our method outperforms STEGO and HP on unsupervised segmentation with the ViT-B/8, while showing competitive performance for the ViT-S/8. *Results obtained without post-processing optimization." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "3641" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "text", + "content": "4.1. COCO-Stuff" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": "We present our evaluation on COCO-Stuff 27 in Table 1. For the ViT-S/8, our experiments show that Ours is able to improve upon STEGO in most metrics, with improved unsupervised accuracy by " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+8.0\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": " and unsupervised mIoU increased by " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": ". Ours w/ 3D-LHP further increases this mIoU delta to " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+1.8\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": ", highlighting the effectiveness of our 3D-information propagation strategy in combination with DepthG. When comparing our approach to Hidden Positives, a method with more computational overhead, for the ViT-S/8, we show competitive performance for unsupervised accuracy and outperform their approach by " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": " on unsupervised mIoU with Ours and " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": " with Ours w/ 3D-LHP. When using the DINO ViT-B/8 encoder, our approach again outperforms STEGO, as well as all other presented methods on unsupervised metrics. Most notably, we are able to increase the unsupervised mIoU by " + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "inline_equation", + "content": "+0.8\\%" + }, + { + "bbox": [ + 46, + 90, + 289, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 291, + 121, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 121, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 121, + 304 + ], + "type": "text", + "content": "4.2. Cityscapes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "text", + "content": "We further evaluate our approach on the Cityscapes dataset [11], consisting of various scenes from 50 different cities. We follow the training setting from STEGO and, contrary to all other datasets, do not sample point-wise but use the full feature map for our learning process along with the full depth map. As can be seen in Table 2, our method significantly outperforms STEGO as well as Hidden Positives. For unsupervised mIoU, while Hidden Positives decreased performance compared to STEGO, we observe that our approach to achieves a " + }, + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "text", + "content": " increase. Similarly, we report state-of-the-art performance in accuracy, building upon Hidden Positives' already impressive improvements over STEGO and outperforming it by " + }, + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 46, + 310, + 289, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 111, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 111, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 111, + 487 + ], + "type": "text", + "content": "4.3. Potsdam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 494, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 289, + 662 + ], + "type": "text", + "content": "Our model is further evaluated on the Potsdam-3 dataset, containing aerial images of the German city of Potsdam. Contrary to the other benchmarks, which contain images in a first-person perspective, Potsdam-3 contains only birds-eye-view images, a perspective that is considered out-of-distribution for the monocular depth estimator. Despite this inherent limitation of our approach for aerial data, we are able to demonstrate relatively commendable performance in Table 3 by improving STEGO's performance and reporting state-of-the-art performance for the ViT-S backbone. In contrast, Hidden Positives [35] use a ViT-B/8, and with roughly twice the parameters as ours, reach an accuracy of " + }, + { + "bbox": [ + 46, + 494, + 289, + 662 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 46, + 494, + 289, + 662 + ], + "type": "text", + "content": ". We present a visual overview of the predicted Potsdam depth maps in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 160, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 160, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 160, + 683 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We present qualitative results of our method in Figure 4 and compare with segmentation maps from STEGO. On" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 173 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 173 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 173 + ], + "type": "table", + "html": "
MethodModelU. AccU. mIoU
IIC [20]R18+FPN47.96.4
PiCIE [10]R18+FPN65.612.3
DINO [7]ViT-B/843.611.8
STEGO + HP [35]ViT-B/879.518.4
STEGO [16]ViT-B/873.221.0
+ OursViT-B/881.623.1
", + "image_path": "6b3fe1628976c2526f1cd33b4f58949724bd637e4c2190c3cc6de27d16e5b242.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 233, + 520, + 348 + ], + "blocks": [ + { + "bbox": [ + 305, + 180, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 305, + 180, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 547, + 225 + ], + "type": "text", + "content": "Table 2. Results on Cityscapes. We report unsupervised accuracy and mIoU on Cityscapes. Our method outperforms both STEGO variants by substantial margins. Notably, our method is the first to improve upon unsupervised mIoU." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 233, + 520, + 348 + ], + "lines": [ + { + "bbox": [ + 332, + 233, + 520, + 348 + ], + "spans": [ + { + "bbox": [ + 332, + 233, + 520, + 348 + ], + "type": "table", + "html": "
MethodModelU. Acc.
CC [19]VGG1163.9
DeepCluster [6]VGG1141.7
IIC [20]VGG1165.1
DINO [7,21]ViT-S/871.3
STEGO [16,21]ViT-S/877.0
+ OursViT-S/880.4
+ Ours w/ 3D-LHPViT-S/880.4
", + "image_path": "b742b0cb1400e48a1963120a8e8d9702289c8042ff57bd2d89fd8f6ccda5cc6c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 418, + 547, + 520 + ], + "blocks": [ + { + "bbox": [ + 305, + 356, + 547, + 412 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 547, + 412 + ], + "type": "text", + "content": "Table 3. Results on Potsdam. We report unsupervised accuracy on the Potsdam dataset. Our method is able to improve upon STEGO. We hypothesize that with a zero-shot depth estimator more suitable for aerial images, the results for our method could further improve." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 418, + 547, + 520 + ], + "lines": [ + { + "bbox": [ + 307, + 418, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 547, + 520 + ], + "type": "table", + "html": "
MethodU. mIoU.U. Acc
STEGO [16]24.548.3
+ Depth-Feature Correlation (1)24.751.2
+ FPS (2)24.649.1
+ 3D-LHP (3)25.248.5
+ Ours (1, 2)25.656.3
+ Ours w/ 3D-LHP (1, 2, 3)26.755.1
", + "image_path": "b44384b9bf35062d5580ca68aeae15e70f520dd7e875d6fd932fe99ee6b79657.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 529, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 551 + ], + "type": "text", + "content": "Table 4. Effect of our contributions. We compare our individual contributions and the combination of all contributions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "multiple occasions, our depth guidance reduces erroneous predictions from the model caused by visual irritations in the pixel space. In the example of the boy with the baseball bat in Figure 4a, false classifications from STEGO are caused by shadows on the ground. Our model is able to correct this. Furthermore, it goes beyond the noisy label and also correctly classifies the glimpse of a plant that can be seen through a hole in the background. This is an indication that our model does not overfit to the depth map, since this visual cue is only observable in the pixel space, but not the depth map." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3642" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 293, + 278 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 293, + 278 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 293, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 293, + 278 + ], + "type": "image", + "image_path": "59e641edfed74e60fae016a3268151762a0cad18f01b708fda77b3a1348329b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 281, + 196, + 289 + ], + "lines": [ + { + "bbox": [ + 143, + 281, + 196, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 196, + 289 + ], + "type": "text", + "content": "(a) COCO-Stuff" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 68, + 547, + 278 + ], + "blocks": [ + { + "bbox": [ + 302, + 68, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 302, + 68, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 302, + 68, + 547, + 278 + ], + "type": "image", + "image_path": "652e6f12fc2a495b4f2261e12088cab65129cb38427224d8bcb0e4c49ee744fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 281, + 448, + 290 + ], + "lines": [ + { + "bbox": [ + 401, + 281, + 448, + 290 + ], + "spans": [ + { + "bbox": [ + 401, + 281, + 448, + 290 + ], + "type": "text", + "content": "(b) Cityscapes" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "type": "text", + "content": "Figure 4. Qualitative results. We show qualitative differences for plain STEGO compared to STEGO with our depth guidance, using ViT-S models for COCO and ViT-B for Cityscapes. Where STEGO struggles to differentiate instances, our model is able to correct this and successfully separates them for segmentation. In the case of the building in (a), our method alleviates visual irritations from the pixel space and corrects the segmentation of the building. In (b), our model is able to better handle visual inconsistencies from shadows." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 355, + 112, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 112, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 112, + 368 + ], + "type": "text", + "content": "5. Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 289, + 604 + ], + "type": "text", + "content": "Individual Influence. We investigate the effect of our technical contributions on training our model with a ViT-S/8 backbone on COCO-Stuff 27. Our observations in Table 4 show that our Depth-Feature Correlation loss itself already improves the performance of STEGO. This improvement is further increased through the use of FPS, which enables us to sample the depth space more meaningfully and therefore encourages more diversity in the depth correlation tensor " + }, + { + "bbox": [ + 46, + 375, + 289, + 604 + ], + "type": "inline_equation", + "content": "D_{hw,uv}" + }, + { + "bbox": [ + 46, + 375, + 289, + 604 + ], + "type": "text", + "content": ". Intuitively, this sampling diversity significantly amplifies our Depth-Feature Correlation for aligning the feature space with the depth space. We provide a visual comparison to random sampling in Figure 5 and additional illustrations in the appendix. FPS retrieves more diverse locations and specifically selects locations with depth discontinuities. This naturally benefits the Depth-Feature Correlation to learn sharper edges in this area for the output segmentation. Adding local hidden positives with depth maps further increases the unsupervised mIoU, while slightly lowering the accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Source Of Depth Maps. We investigate the effect of different monocular depth estimators to generate the depth maps used to train our model. In our experiments, we consider three options: The previously mentioned ZoeDepth [3], Kick Back & Relax (KBR) [36] which uses self-supervision to learn depth from Slow-TV videos, as well as MiDaS [32], the base model to ZoeDepth. Empirically, ZoeDepth produces the most accurate zero-shot" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 477 + ], + "type": "text", + "content": "monodepth results across indoor and outdoor datasets, followed by MiDaS and KBR [3,36]. We provide a qualitative comparison in the appendix. To evaluate the influence of the depth map quality on the performance of our model, we first generate depth maps for COCO-Stuff 27 using each of the introduced models. We then train our model with a ViT-S/8 backbone and report the results in Table 5. We observe that depth maps from ZoeDepth work best with our method, while the model trained with MiDaS depth has an edge over the KBR counterpart." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 489, + 545, + 630 + ], + "blocks": [ + { + "bbox": [ + 306, + 489, + 545, + 630 + ], + "lines": [ + { + "bbox": [ + 306, + 489, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 545, + 630 + ], + "type": "image", + "image_path": "c4f3c4efd1e84904ac46fe56ea15f26a851f17e279e0d2b0c10efd522e629a07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 638, + 547, + 704 + ], + "lines": [ + { + "bbox": [ + 304, + 638, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 547, + 704 + ], + "type": "text", + "content": "Figure 5. Random vs. Farthest Point Sampling. We observe that random sampling can miss entire structures like trees in the first top and the plane in the bottom row. In contrast, our method meaningfully samples the depth space and selects locations across the different structures and at depth edges. We show further illustrations of FPS in the appendix." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3643" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 224 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "type": "image", + "image_path": "d330e6a3c0ef13c3e80d3f32b1e61f8fa42269c0f7e0760153deb52c1c5d1c8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 287, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 278 + ], + "type": "text", + "content": "Figure 6. Failure cases. We show cases where our model fails to correctly segment and classify the scene. The top row is a prime example where the difference in depth is correctly distilled, though the model fails to correctly classify the snow region." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 292, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 495 + ], + "type": "text", + "content": "Signal Propagation With Local Hidden Positives. We ablate the implementation of LHP along with different propagation strategies. As described in Section 3.6, our method takes advantage of the depth information of the scene to propagate the learning signal to patches which are nearby in 3D space. We further implement utilizing the attention map from the DINO backbone and propagate proportionally to their values i.e., the approach utilized in Hidden Positives [35]. While, in Table 6, we show that our approach benefits from signal propagation with 3D-LHP, applying LHP with Attention leads to lower performance. We speculate 3D-LHP is more suitable for our approach since, for a given location, the signal from our Depth-Feature Correlation loss is calculated w.r.t. the depth at this sample. Propagating to locations with the same depth does not corrupt this signal, though this can happen with LHP (Attention), since it does not consider depth for propagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "Computational Cost. Our method only leads to an insignificant increase in runtime versus the baseline STEGO model, since we solely guide the loss as well as the feature sampling and only for Ours w/ 3D-LHP add an additional small segmentation head. In contrast, the competing method Hidden Positives [35] relies on a computationally more expensive process to select features and introduces an additional segmentation head to fill their task-specific feature pool. To keep our computational overhead low, we make use of a pre-trained monocular depth estimation network with impressive zero-shot capabilities. We consider task specific training of the depth estimator not a necessity, since the model has zero-shot capabilities that generalize well to different scenes and domains. Therefore, in our experiments on a diverse array of scenes, we do not re-train the depth estimator, and consider the additional computational cost for generating the depth maps to be negligible." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 131 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 131 + ], + "type": "table", + "html": "
MethodTrained withU. Acc.U. mIoU
ZoeDepth [3]Sensor Depth56.325.6
MiDaS [32]Sensor Depth53.025.0
KBR [36]Self-Supervision50.623.1
", + "image_path": "6b7920799ce5b47643586229418c9f94f593c3c240f8aac2a7aea2df72c67db7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 198, + 535, + 259 + ], + "blocks": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 183 + ], + "type": "text", + "content": "Table 5. Different depth map sources. We experiment with different monocular depth estimators which were trained with either sensor depth or self-supervision. Overall, the model trained with depth maps from ZoeDepth performs best on COCO-Stuff 27." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 198, + 535, + 259 + ], + "lines": [ + { + "bbox": [ + 316, + 198, + 535, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 535, + 259 + ], + "type": "table", + "html": "
LHPPropagation StrategyU. Acc.U. mIoU
X-56.325.6
✓3D-LHP (Depth)55.126.7
✓LHP (Attention)52.624.5
", + "image_path": "fed6098ada020a7f52e33fefa9261dadda94ad1990863270907beecd9da6ce9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 267, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 267, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 545, + 312 + ], + "type": "text", + "content": "Table 6. LHP Ablation. We compare the use of depth and attention for local hidden positives. We find that using depth improves unsupervised mIoU, while we find that using attention does not improve our method." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 330, + 381, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 381, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 381, + 342 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "content": "While we have demonstrated our method's effectiveness for many real-world cases, our method's applicability is limited in settings unsuitable for depth estimation, such as slices of CT scans and other medical data domains. Furthermore, the experiments on Potsdam-3 have shown, our method can improve unsupervised semantic segmentation despite suboptimal viewing perspectives for the monocular depth estimator. We assume the scenario of aerial images represents a rare case where our method would profit from a domain-specific monocular depth estimator. We also present failure cases of our model in Figure 6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 500, + 462, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 462, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 462, + 512 + ], + "type": "text", + "content": "7. Conclusion & Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "In this work, we have presented a novel method to induce spatial knowledge of the scene into our model for unsupervised semantic segmentation. We have proposed to correlate the feature space with the depth space and use the 3D information to more meaningfully sample features in a spatially informed way. Furthermore, we have demonstrated that these contributions produce state-of-the-art performance on many real-world datasets and thus foster the progress in unsupervised segmentation. The applicability of our approach for other tasks is further to be explored, since we hypothesize it can be useful beyond unsupervised segmentation as part of other contrastive processes. We consider this to be a promising direction for future work. Furthermore, it remains to be investigated what kind of information could be useful in domains where depth is not an obviously meaningful signal, like medical data." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3644" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 181 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 339 + ], + "type": "text", + "content": "[5] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1129-1139, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "type": "text", + "content": "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 441 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 444, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 508 + ], + "type": "text", + "content": "[8] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2624-2632, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 544 + ], + "type": "text", + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[10] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "text", + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "text", + "content": "[14] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 258 + ], + "type": "text", + "content": "[15] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 260, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 314 + ], + "type": "text", + "content": "[16] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[17] Ji Hou, Xiaoliang Dai, Zijian He, Angela Dai, and Matthias Nießner. Mask3d: Pre-training 2d vision transformers by learning masked 3d priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13510-13519, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "type": "text", + "content": "[18] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11130-11140, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "text", + "content": "[19] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H Adelson. Learning visual groups from co-occurrences in space and time. arXiv preprint arXiv:1511.06811, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "text", + "content": "[20] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9865-9874, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "text", + "content": "[21] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of stego for safe unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3788-3797, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 581, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 623 + ], + "type": "text", + "content": "[22] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "text", + "content": "[23] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[24] Harold W Kuhn. Variants of the hungarian method for assignment problems. Naval research logistics quarterly, 3(4):253-258, 1956. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3645" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "text", + "content": "[26] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[27] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 275 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "type": "text", + "content": "[29] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2):129-137, 1982. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "type": "text", + "content": "[30] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 288, + 399 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 288, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 288, + 455 + ], + "type": "text", + "content": "[32] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[34] M Seitzer, M Horn, A Zadaianchuk, D Zietlow, T Xiao, C Simon-Gabriel, T He, Z Zhang, B Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In International Conference on Learning Representations (ICLR), 2023. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[35] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19540-19549, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[36] Jaime Spencer, Chris Russell, Simon Hadfield, and Richard Bowden. Kick back & relax: Learning to reconstruct the world by watching slowtv. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[37] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Dada: Depth-aware domain adaptation in semantic segmentation. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 540 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 7364-7373, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[38] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8515-8525, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[39] Risheng Wang, Tao Lei, Ruixia Cui, Bingtao Zhang, Hongying Meng, and Asoke K Nandi. Medical image segmentation using deep learning: A survey. IET Image Processing, 16(5):1243-1267, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[40] Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. Multimodal token fusion for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12186-12195, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[41] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[42] Zhaoyuan Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, and Rong Jin. Transfgu: a top-down approach to fine-grained unsupervised semantic segmentation. In European conference on computer vision, pages 73-89. Springer, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[43] Xiaowen Ying and Mooi Choo Chuah. Uctnet: Uncertainty-aware cross-modal transformer network for indoor rgb-d semantic segmentation. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "type": "text", + "content": "[44] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 540 + ], + "type": "text", + "content": "[45] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3646" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_content_list.json b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4053d9df8733c81f648a320652b901ba80aa7612 --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_content_list.json @@ -0,0 +1,1473 @@ +[ + { + "type": "text", + "text": "Unsupervised Template-assisted Point Cloud Shape Correspondence Network", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacheng Deng $^{1}$ , Jiahao Lu $^{1}$ , Tianzhu Zhang $^{1,2,\\dagger}$", + "bbox": [ + 295, + 179, + 674, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2Deep Space Exploration Lab {dengjc, lujiahao} @mail.ustc.edu.cn, tzzhang@ustc.edu.cn", + "bbox": [ + 233, + 200, + 738, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods.", + "bbox": [ + 75, + 301, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 633, + 209, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud shape correspondence is a challenging task that identifies and densely matches the source and target point clouds with deformable 3D shapes. The task has significant implications for various industries, including augmented reality [1], gaming [15], and robotics [21, 32]. However, the unrestricted mobility of humans and animals and their unusual postures have made direct correspondence between unconventional shapes a longstanding challenge in point cloud shape correspondence.", + "bbox": [ + 75, + 661, + 468, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The shape correspondence problem has been thoroughly investigated for 3D mesh data. However, such spectral-based methods [2, 14, 19, 27, 35] are mainly limited by the time-consuming pre-processing step and the reliance on mesh vertical connectivity information, frequently absent", + "bbox": [ + 75, + 797, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0772c8e58c58b0dd860e9f0a0afe8d05508e32648dd23247f674e7574bcaf5ca.jpg", + "image_caption": [ + "Figure 1. Visualization of template-assisted shape correspondence results. Correspondences are depicted by transferring colors from the source to the target based on matching results. The baseline incorrectly aligns the hands with the head and knees between unconventional shapes. In contrast, our method leverages the template to establish accurate correspondences for the hands." + ], + "image_footnote": [], + "bbox": [ + 504, + 267, + 890, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in point cloud scenarios. Point-based methods [7, 8, 11, 13, 16, 24, 42] are directly applied to the raw point cloud data, relying solely on point coordinates without connectivity information. To mitigate the resource-intensive labeling in fully supervised methods [8, 11, 24], the unsupervised point cloud shape correspondence task garners significant attention and yields competitive results [7, 13, 16, 42]. DPC [16] designs construction losses to reduce outlier matches. HSTR [13] incorporates whitening losses to mitigate the impact of shape discrepancies. SE-ORNet [7] introduces an orientation estimation module to alleviate symmetry issues. The shapes of humans and animals are highly diverse and complex. For instance, the SURREAL dataset [11] comprises $230\\mathrm{K}$ human body shapes, while the SMAL dataset [45] contains $10\\mathrm{K}$ animal shapes, including numerous unconventional forms. Establishing accurate correspondences between two unconventional point cloud shapes remains a challenging problem in shape correspondence. Current methods [7, 13, 16, 42] rely on computing", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5250", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the similarity between point cloud features to establish correspondences. However, in the case of unconventional point cloud pairs, the complete structure is distorted or destroyed due to limb adhesion, bending, and other postures. Without explicit structural guidance, directly calculating similarity is often insufficient to guarantee accurate correspondences. Therefore, it is essential to consider how to construct template shapes as intermediaries to assist in establishing correct correspondences.", + "bbox": [ + 75, + 90, + 470, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By analyzing prior shape correspondence methods, we have identified two pivotal issues that must be considered in shape correspondence. 1) How to generate suitable templates for point cloud pairs? Traditional point cloud template construction [11, 37, 41] mainly relies on manual selection or autoencoders [41]. While effective for shape-stable rigid objects in the same category, balancing template complexity becomes challenging for diverse non-rigid objects of different categories. Simple templates are clear but lack structural information, risking information loss during correspondence propagation. Complex templates offer richer information but are sensitive to noise, compromising accuracy. Despite attempts [26, 43] with learnable approaches, incomplete template structure remains a challenge. A more effective modeling approach is needed to generate templates suitable for intricate point cloud pairs. 2) How to effectively leverage templates for better correspondences? Establishing direct correspondences between unconventional shapes is challenging, but building correspondences between an unconventional shape and a suitable template is relatively straightforward. A correct mapping can be formed by finding corresponding template points for the source points and identifying corresponding points on the target point cloud based on the template. Direct similarity calculations between point clouds may introduce noise and yield ambiguous similarity weights. The template allows each point in the point cloud to compute a correlation vector, providing more stable information. Seeking consensus among these correlation vectors enhances appropriate point correspondences while mitigating erroneous ones.", + "bbox": [ + 75, + 231, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve the above goal, we propose an Unsupervised Template-Assisted Point Cloud Shape Correspondence Network (TANet), including a template generation module and a template assistance module. In the template generation module, we incorporate a template bank with several learnable templates, effectively balancing template complexity in a data-driven manner. The space aligner enriches templates with comprehensive shape structures by establishing a mapping between the template feature space and the coordinate space. The space aligner is supervised by the coordinates and encoded features of the point cloud pairs. In the template assistance module, the adaptive selector chooses the most suitable template from the bank based on geometric and semantic attributes. The correlation fusion process", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhances point features with the template correlation vector via attention, suppressing noise and ambiguity. Moreover, we devise a transitive consistency loss to ensure coherence between the similarity computed through the template and directly computed between the source and target point clouds.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, the contributions of this work can be summarized as follows: (i) We propose an unsupervised template-assisted point cloud shape correspondence network, achieved by jointly designing a template generation module and a template assistance module. (ii) We introduce a template generation module comprising a template bank with learnable templates and the space aligner for building explicit structures. The template assistance module adaptively selects a suitable template for robust point representations and maintains transitive consistency in shape correspondences. (iii) Extensive experiments on TOSCA [3] and SHREC'19 [25] benchmarks show the proposed TANet outperforming state-of-the-art methods. Cross-dataset experiments on SMAL [45] and SURREAL [11] demonstrate our method's desirable generalization capabilities.", + "bbox": [ + 496, + 181, + 892, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 421, + 640, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this Section, we give a brief overview of related works on point cloud shape correspondence, including deep learning for point clouds, shape correspondence, and point cloud template.", + "bbox": [ + 496, + 446, + 890, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning for Point Clouds. PointNet [29] employs a symmetrical function to aggregate all point features, resulting in a permutation-invariant global feature representation. Building upon this, PointNet++ [30] takes the concept further by incorporating local information, enhancing the representation capabilities and efficiency. DGCNN [38] introduces EdgeConv [38] blocks, dynamically updating neighborhood information based on dynamic graphs, ultimately delivering improved performance in point cloud analysis. Similarly, KPCnv [36] introduces point-wise convolution operators for point cloud feature learning. In the era of transformers, Point Cloud Transformer [12] stands out as a purely global Transformer network, pioneering self-attention layers instead of encoder layers within the PointNet framework. Point Transformer [44] operates within the local neighborhood of the target point cloud, enabling hierarchical extraction of geometric and semantic features.", + "bbox": [ + 496, + 508, + 892, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shape Correspondence. The goal of shape correspondence is to accurately find the point-to-point correspondence between two point clouds. Traditional spectral-based methods [2, 14, 27, 35] utilize mesh data to project features onto Laplace-Beltrami Operator (LBO) eigenbasis and then learn a transformation between the eigendecomposition of source and target shapes, which will be translated to a dense shape correspondence. Cao et al. [4] transfer rich structure and connectivity information in mesh data", + "bbox": [ + 496, + 765, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5251", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94d278b8f0b473e22f8c68b2901823f9660c726abd02bbcb6f5f0c695b1decf2.jpg", + "image_caption": [ + "Figure 2. Illustration of the TANet. TANet comprises an encoder, a template generation module, and a template assistance module. The template generation module produces several learnable shape templates in the template bank with a space aligner. The template assistance module selects suitable templates for point cloud pairs and improve accuracy via correlation fusion and transitive consistency." + ], + "image_footnote": [], + "bbox": [ + 114, + 84, + 861, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to point cloud shape for robust matching. However, these methods are constrained by complex preprocessing and a reliance on vertex connectivity information. In contrast, point-based methods [5, 7, 9, 13, 16, 18, 34, 42] process point clouds directly without depending on connectivity information. CorrNet3D [42] and DPC [16] harness the established DGCNN [38] network to extract point representations through shape reconstruction. HSTR [13] has introduced a multi-receptive-field transformer point cloud encoder that considers local point cloud structures and long-range contextual information. SE-ORNet [7] incorporates an orientation estimation module to align the orientations of point cloud pairs, achieving precise matching results, particularly for symmetrical parts. These methods directly calculate point-wise similarities, making them susceptible to matching errors caused by noise and incomplete representations. In contrast, our approach constructs and utilizes templates to establish stable correspondences.", + "bbox": [ + 75, + 439, + 472, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Template. Point cloud templates, such as shape templates, scene priors, and semantic priors, are pivotal in various point cloud tasks, including 6D pose estimation [17, 37], reconstruction [10, 20], scene synthesis [26], and segmentation [23, 33, 39]. Tian et al. [37] develop an autoencoder trained on a collection of object models, which decodes the mean latent embedding to create shape templates. For similar rigid objects, generative methods can yield robust and stable templates. ScenePrior [26] encodes scene priors into a latent space and utilizes sampling to synthesize plausible 3D scenes. SemAffiNet [39] and Mask3D [33] integrate a learnable implicit semantic prior", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "via transformer blocks with learnable queries. However, the complex shape and category variations pose a challenge for non-rigid objects to obtain well-balanced explicit templates through generative methods. The learnable implicit modeling struggles to achieve explicit stuctures. Based on the above discussion, we propose TANet, which integrates the template generation and assistance modules within a unified framework. Through technical designs, the learnable templates not only balance shape complexity but also possess an explicitly structured shape.", + "bbox": [ + 496, + 439, + 893, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 604, + 591, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 630, + 609, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unsupervised point cloud correspondence task strives to establish accurate one-to-one correspondences between source and target point clouds without annotations. As illustrated in Figure 2, TANet primarily comprises three key modules: a general point cloud encoder, a Template Generation Module, and a Template Assistance Module. The choice of point cloud encoder follows the previous study [13]. The detailed design and optimization process for the Template Generation Module will be expounded upon in Section 3.2. Moreover, Section 3.3 will provide further insights into the Template Assistance Module. Finally, the training and inference process are discussed in Section 3.4.", + "bbox": [ + 496, + 654, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Template Generation Module", + "text_level": 1, + "bbox": [ + 500, + 845, + 764, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The template generation module consists primarily of two components: a template bank and a space aligner.", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5252", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Template bank. In the template bank, we integrate $K$ learnable explicit shape templates, each comprising learnable positions $P_T^i \\in \\mathbb{R}^{N \\times 3}$ and learnable embeddings $F_T^i \\in \\mathbb{R}^{N \\times d}$ . During the training phase, the positions and embeddings of templates are updated concurrently with the network. In the inference phase, the positions and embeddings of template shapes remain unchanged, eliminating the need for additional updates based on input point clouds. The learnable embeddings $F_T^i$ parameters are initialized using Gaussian random initialization. To expedite template training, we utilize random shapes in datasets to initialize the learnable positions $P_T^i$ .", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6850f72c935869667dbeeb3e14b4448529cf987b3fa7fca30c613c1629cae051.jpg", + "image_caption": [ + "Figure 3. Space aligner structure. The predicted point positions in the output are constrained to match the true point positions." + ], + "image_footnote": [], + "bbox": [ + 94, + 287, + 450, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Space aligner. Space aligner $\\phi$ is crafted to predict point positions through point embeddings, aiding in template learning. The specific structure is illustrated in Figure 3. The predicted point positions undergo adjustment to align with true positions:", + "bbox": [ + 75, + 407, + 470, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {S m o o t h} \\mathrm {L} _ {1} (\\mathrm {P}, \\phi (\\mathrm {F})),\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 491, + 377, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P$ are true positions, and $F$ are point embeddings. The space aligner is trained with the encoded feature and the coordinates from point cloud pairs. During the inference phase, the space aligner is not involved in the computation.", + "bbox": [ + 75, + 513, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, the training of templates also includes the template construction loss $\\mathcal{L}_{tc}$ , which enforces consistency between the template shape and the cross-constructed template shape by point cloud pairs. Specifically, the cross-construction process is computed as follows:", + "bbox": [ + 75, + 575, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {t} _ {x _ {i}} = \\sum_ {j \\in \\mathcal {N} _ {T} (x _ {i})} \\frac {e ^ {s _ {i j}}}{\\sum_ {l \\in \\mathcal {N} _ {T} (x _ {i})} e ^ {s _ {i l}}} t _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 657, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{i} \\in X$ , $t_{j} \\in P_{T}$ and $s_{ij} \\in S_{XT}$ . $S_{XT}$ is the similarity matrix between $F_{X}$ and $F_{T}$ . $\\mathcal{N}_T(x_i)$ represents the latent $k$ -nearest neighbors of $x_{i}$ in the target $F_{T}$ . The cross-construction of $P_{T}$ by the source point cloud $X$ is denoted $\\hat{T}_{X} \\in \\mathbb{R}^{N \\times 3}$ , where $\\hat{T}_{X}^{i} = \\hat{t}_{x_{i}}$ . After the introduction of the cross-construction process, the template construction loss $\\mathcal{L}_{tc}$ can be computed as follows:", + "bbox": [ + 75, + 702, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t c} = \\mathbf {C D} (P _ {T}, \\hat {T} _ {X}) + \\mathbf {C D} (P _ {T}, \\hat {T} _ {Y}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 814, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CD denotes chamfer distance. It is essential to highlight that the point encoder remains unaltered during the backpropagation of $\\mathcal{L}_{\\text {align }}$ and $\\mathcal{L}_{\\text {tc }}$ gradients, preventing interference with the feature extraction of point cloud pairs.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Template Assistance Module", + "text_level": 1, + "bbox": [ + 500, + 90, + 756, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The template assistance module aims to enhance the correspondence accuracy in point cloud pairs through templates. It comprises three primary components: Template Selector, Correlation Fusion, and Transitive Consistency.", + "bbox": [ + 498, + 114, + 890, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a36ecdcbaace7a50176579398b88fc58035bc5865d574c19b32e43347ab92ea3.jpg", + "image_caption": [ + "Figure 4. Template selector structure. The best suitable template is determined through geometric and semantic measures of similarity between templates and point cloud pairs." + ], + "image_footnote": [], + "bbox": [ + 513, + 190, + 874, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Template selector. The input to the Template Selector comprises point cloud coordinates $X \\& Y \\in \\mathbb{R}^{N \\times 3}$ and features $F_{X} \\& F_{Y} \\in \\mathbb{R}^{N \\times d}$ , along with template point cloud coordinates $\\{P_T^i | i = 1,2,\\dots,K\\}$ and embeddings $\\{F_T^i | i = 1,2,\\dots,K\\}$ from the template bank. Geometric and semantic similarities between the point cloud pair and templates guide the adaptive template selection. As illustrated in Figure 4, it involves calculating the chamfer distance as the geometric similarity between the point cloud pair and template point cloud. To get semantic similarity, the point cloud features and template embeddings undergo mix pooling and cosine similarity computation. The fusion of geometric and semantic similarities is employed in Gumbel-softmax to ascertain the most fitting template $T$ for the given point cloud pair. The specific computation formula is as follows:", + "bbox": [ + 498, + 321, + 892, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T = \\mathbf {G S} \\left\\{\\left(\\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {X}\\right) + \\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {Y}\\right)\\right) \\right. \\tag {3} \\\\ + \\left(\\mathbf {C D} \\left(P _ {T} ^ {i}, X\\right) + \\mathbf {C D} \\left(P _ {T} ^ {i}, Y\\right)\\right) | i = 1, 2, \\dots , K \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 553, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where GS represents Gumbel-Softmax, MP denotes Mix Pooing, and CD stands for chamfer distance.", + "bbox": [ + 500, + 598, + 888, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correlation fusion. As depicted in Figure 5, template embeddings $F_{T}$ and point cloud features $F_{X} / F_{Y}$ are input into the correlation fusion process to enhance point cloud features. Initially, a correlation between the point cloud and the template is established. Subsequently, an attention mechanism incorporates the correlation vectors into point cloud features, yielding $F_{X}' / F_{Y}' \\in \\mathbb{R}^{N \\times d}$ . The similarity in the attention mechanism is computed based on point cloud features to underscore the spatial distribution of shapes, enhancing perception to spatial relationships. The introduction of template correlation aids in refining point cloud features and mitigating ambiguous representations in certain noisy point clouds.", + "bbox": [ + 498, + 628, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transitive consistency. Utilizing the fused source features $F_{X}^{\\prime}$ and template embeddings $F_{T}$ , we compute the similarity matrix $S_{XT}$ , and $S_{TY}$ is obtained by calculating the similarity between $F_{T}$ and $F_{Y}^{\\prime}$ . The multiplication of $S_{XT}$ and $S_{TY}$ matrices yields the similarity matrix $S_{X TY}$ ,", + "bbox": [ + 498, + 825, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5253", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba78a9763622f2cefb092af425fcf3c9be703dda78d6f723475d6dc063c8b69b.jpg", + "image_caption": [ + "Figure 5. The correlation fusion process. Features of point cloud pairs and embeddings of templates compute correlation vectors, which are then fused using an attention mechanism." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 433, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reflecting the similarity between the source and target point clouds transmitted through the template. As shown in Figure 6, (a) Direct similarity illustrates the process of directly calculating the similarity between the source point cloud and the target point cloud:", + "bbox": [ + 75, + 291, + 470, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {X Y} = F _ {X} ^ {\\prime} F _ {Y} ^ {\\prime} ^ {\\top}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 378, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, (b) Template-assisted similarity demonstrates the computation of $S_{X TY}$ . The multiplication of matrices $S_{XT}$ and $S_{TY}$ essentially results in an indirect computation of the similarity relationship between the source and target point clouds. This process involves a weighted sum of the similarity calculated between the template point cloud and the target point cloud, providing a more stable and accurate similarity result for source and target points. The Template-assisted similarity method helps mitigate the impact of outliers and noise on the similarity calculation, leading to more reliable results. In order to guide the establishment of more stable and accurate similarity relationships between point clouds without compromising the inference speed, we introduce the Transitive Consistency loss $(\\mathcal{L}_{trans})$ , promoting the consistency between $S_{X TY}$ and the similarity matrix $S_{XY}$ . The Transitive Consistency loss is calculated as follows:", + "bbox": [ + 75, + 412, + 470, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s}} = \\mathbf {C E} \\left(S _ {X T Y}, S _ {X Y}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 672, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where CE means Cross-Entropy loss.", + "bbox": [ + 76, + 700, + 331, + 715 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eeb0efd0caf43b60588bdbda8d510ddfb437bb5546e3ae5b85d17aeccd60a094.jpg", + "image_caption": [ + "Figure 6. Transitive Consistency. (a) the direct similarity computation between the source and target point clouds. (b) a stable template is employed to transmit the similarity relation. $L_{trans}$ ensures the consistency between these two similarity results." + ], + "image_footnote": [], + "bbox": [ + 119, + 732, + 426, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training & Inference", + "text_level": 1, + "bbox": [ + 500, + 90, + 754, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the previous work [7, 13, 16], we also incorporate the construction loss $(\\mathcal{L}_{\\text{const}})$ , encompassing both cross-construction and self-construction between the source shape $X$ and the target shape $Y$ . The construction loss promotes feature smoothness and unique correspondences. The computation process and formula for cross-construction and self-construction are similar to Equation 1. Specific formulas are detailed in the supplementary material. To sum up, the total loss of TANet is formulated as:", + "bbox": [ + 496, + 113, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {t r a n s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {t c}} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t r}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 262, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{i}$ are hyperparameters, balancing the contribution of different loss terms. The training of both the template and the network is jointly achieved, enabling end-to-end training without requiring additional prolonged iterations.", + "bbox": [ + 496, + 292, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we set the closest point $y_{j^*}$ in the feature space for each point $x_{i}$ as its corresponding point. This selection rule can be formulated as:", + "bbox": [ + 496, + 354, + 890, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(x _ {i}\\right) = y _ {j ^ {*}}, j ^ {*} = \\underset {j} {\\operatorname {a r g m a x}} \\left(s _ {i j}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 411, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 436, + 633, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 460, + 689, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on the most popular benchmarks, including the TOSCA [3] dataset and the SHREC'19 [25] dataset. TOSCA comprises 41 different shapes of various animal species. Following previous methods [13, 16], we pair these shapes to create both the training and testing sets. SHREC'19 consists of 44 real human models paired into 430 shape correspondence samples. SURREAL [11] and SMAL [45] are two large-scale datasets that we employ to evaluate the generalization capabilities of our method. The SURREAL dataset comprises 230,000 training shapes, from which we randomly sample and pair 2,000 shape pairs for training. The SMAL dataset includes parameterized animal models for generating shape pairs to train shape correspondence models. Furthermore, robustness experiments are conducted on the real-scanned Owlii dataset [40] and the partial shape dataset SHREC'16 [6]. Besides, we maintain a consistent point count of $n = 1024$ in accordance with prior studies [13, 16, 42].", + "bbox": [ + 496, + 483, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. The primary evaluation metrics for the point cloud shape correspondence task include two key measures: average correspondence error and correspondence accuracy. The average correspondence error measures a source and target shape pair $(X,Y)$ as follows:", + "bbox": [ + 496, + 756, + 890, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne r r = \\frac {1}{n} \\sum_ {x _ {i} \\in X} | f (x _ {i}) - y _ {g t} | _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 842, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{gt} \\in Y$ represents the ground-truth matching point", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5254", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to $x_{i}$ . Correspondence accuracy is formulated as follows:", + "bbox": [ + 76, + 90, + 455, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\na c c (\\epsilon) = \\frac {1}{n} \\sum_ {x _ {i} \\in X} \\mathbb {1} \\left(\\left| f \\left(x _ {i}\\right) - y _ {g t} \\right| _ {2} < \\epsilon d\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 114, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ denotes the indicator function, $d$ is the maximum Euclidean distance between points in $Y$ , and $\\epsilon \\in [0,1]$ signifies the error tolerance.", + "bbox": [ + 76, + 160, + 468, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The number of learnable templates $(K)$ in the Template Bank is set to 4, with a template embedding dimension of 512. The Space Aligner incorporates two attention blocks, each with a dimensionality of 512. We employ a transformer encoder, akin to existing method [13], featuring four layers of attention blocks as our point cloud encoder. For the Correlation Fusion process, the dimensions of self-attention and MLP are set to 512. In Equation 6, $\\lambda_1$ , $\\lambda_2$ , $\\lambda_3$ , and $\\lambda_4$ are set to 0.5, 0.5, 1, and 1. All experiments with our method on various datasets are conducted on a single GeForce RTX 3090 device using the PyTorch 1.10.1 framework [28]. Model training utilizes the AdamW [22] optimizer with a learning rate of 5e-4, a weight decay of 5e-4, and a batch size of 4.", + "bbox": [ + 75, + 205, + 470, + 417 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3342c0126e771bae1809e56e6f2ddc354ca9f4434dd7cb3a8302760eb0e198f4.jpg", + "table_caption": [ + "Table 1. Comparison on TOSCA and SHREC'19 benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. Higher accuracy and lower error reflect a better result, with the best and second-best outcomes highlighted in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodInputTOSCASHREC'19
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh//5.9%0.2
CorrNet3D[42]Point0.3%32.70.4%33.8
DPC[16]Point34.7%2.815.3%5.6
SE-ORNet[7]Point38.3%2.717.5%5.1
HSTR[13]Point52.3%1.219.3%4.9
OursPoint65.1%0.721.5%4.5
", + "bbox": [ + 93, + 502, + 450, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 635, + 441, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted comprehensive experimental comparisons with existing methods on the TOSCA animal dataset and the SHREC'19 human dataset. Furthermore, to ensure a fair comparison with existing methods, our approach is implemented without the use of post-processing or additional connectivity information.", + "bbox": [ + 75, + 659, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on TOSCA dataset. As shown in Table 1, our approach demonstrates significant improvement on the TOSCA dataset, achieving a new state-of-the-art performance with a $12.8\\%$ increase in accuracy and a minimal average error of $0.7\\mathrm{cm}$ . For a comprehensive depiction of accuracy across different error tolerances, we present correspondence accuracy in Figure 7(a). Similarly, our method outperforms others at different error tolerances.", + "bbox": [ + 75, + 750, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on SHREC'19 dataset. As indicated in Table 1, our method demonstrates significant improvement on", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/973e5a12c8854770ddef2485edbbd74a74c4ba1f3ddb597d653e2880535b005e.jpg", + "image_caption": [ + "Figure 7. Correspondence accuracy at various error tolerances. (a) Corresponding accuracy of different methods on the TOSCA dataset under various error tolerances. (b) Experiments similar to (a) conducted on the SHREC'19 dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 696, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14bc645f377329aa3c2e07927c9ffcbe223ece013ae02a6a8e0f219c5b8b7e6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 90, + 893, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the SHREC'19 dataset, achieving an impressive $2.2\\%$ increase in accuracy and a reduction of $0.4\\mathrm{cm}$ in error, consequently establishing a new state-of-the-art performance. Figure 7(b) illustrates the correspondence accuracy of various competitive methods at different error tolerances, showcasing our method's clear superiority across different levels of error tolerance.", + "bbox": [ + 496, + 263, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e7ccaa5516b4efa80258a9d0b978f29f1285e6246e2e29afd0faa6e24ae368b.jpg", + "table_caption": [ + "Table 2. Evaluation of the model with different designs on TOSCA dataset. TGM denotes the template generation module, $L_{tc}$ stands for template construction loss, TS represents template selector, CF denotes correlation fusion process, and $L_{trans}$ is the transitive consistency loss." + ], + "table_footnote": [], + "table_body": "
TGMLtcTSCFLtransTOSCA
acc ↑err ↓
[A]XXXXX52.3%1.2
[B]✓XX✓X54.6%1.2
[C]✓XXX✓55.8%1.1
[D]✓✓X✓X59.3%1.0
[E]✓✓XX✓58.8%1.0
[F]✓✓✓✓X61.9%0.9
[G]✓✓✓X✓61.3%0.9
[H]✓✓✓✓✓65.1%0.7
", + "bbox": [ + 524, + 454, + 859, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 604, + 653, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation of the model with different designs. In Table 2, we conduct extensive ablation studies on the TOSCA dataset to validate the effectiveness of our model designs. Specifically, [A] presents a baseline method without any template-related design. [B] and [C] demonstrate the employment of the template generation module, assisted respectively through correlation fusion and transitive consistency loss, resulting in accuracy improvements of $2.3\\%$ and $3.5\\%$ . The limited improvement is attributed to the difficulty in structural learning of templates without appropriate loss constraints. [D] and [E] illustrate the introduction of template construction loss to aid in the learning and optimization of template structures, leading to advancements in accuracy and error compared to settings [B] and [C]. Building upon this, [F] and [G] further employ a template selector to identify the most suitable template for aiding the establishment of correspondence for specific shapes. The further improvements in accuracy and error indicate that adaptive", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5255", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "template selection significantly contributes to accurate correspondence establishment. Finally, in [H], we present the performance of the complete model, highlighting the essential roles played by different modules in collectively achieving accurate point cloud shape correspondences.", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce5a77a6d3d58d13dcafe8c3a9abde9b8cf437b05f656718547315dd4e84a1bc.jpg", + "image_caption": [ + "Figure 8. Visual comparison of template generation methods. The templates generated by the template generation module exhibit clearer structures in both animal and human templates." + ], + "image_footnote": [], + "bbox": [ + 89, + 180, + 454, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the template generation module. Figure 8 compares templates generated by our template generation module and the common autoencoder-based approach [37]. Our designed learnable templates, constrained by the space aligner and construction loss, provide more structural templates to aid point cloud shape correspondence. However, due to the significant shape variations of non-rigid bodies, it remains challenging to generate perfect templates without direct shape supervision. The existing templates are merely a sub-optimal form. Additionally, we carefully consider the number of templates in Table 3. A single or dual template falls short of capturing the shape variations, resulting in a significant decrease in accuracy. While eight templates can slightly enhance accuracy, they lead to a substantial increase in model parameters. Hence, we offer a trade-off to set up four learnable templates.", + "bbox": [ + 75, + 367, + 472, + 609 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06d9177fd4e876ff33eaf1c25b336413c3c914a7649462a301e38b63ce11c160.jpg", + "table_caption": [ + "Table 3. Ablation studies on Table 4. Comparison of simi-the number of templates. larity computation methods." + ], + "table_footnote": [], + "table_body": "
#TemplateParam.acc ↑
17.3M56.8%
27.8M61.4%
48.8M65.1%
810.9M66.2%
", + "bbox": [ + 80, + 654, + 261, + 729 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/490daf094c2b3a62852fd534d529d93aa2f210d310999882689aa476a6d7374e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsFLOPs(G)TOSCA
acc ↑err ↓
Direct1.0765.1%0.73
Transitive4.2966.3%0.67
", + "bbox": [ + 267, + 667, + 457, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the template assistance module. The effectiveness of correlation fusion has been validated in Table 2. Here, we further delve into the discussion regarding the computation of similarities between point cloud pairs. The purpose of transitive consistency loss is to constrain the direct similarity computation between the source and target point clouds to align with the similarity computed through template transitivity. As demonstrated in Table 4, similarity computation through template propagation has been proven to be a more stable and accurate approach. However, considering the increased computational complexity associated", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with the large number of point clouds, we adopt the direct similarity computation method during inference to enhance model efficiency. The transitive consistency loss only constrains the learning of point cloud representations through similarity-guided learning during training.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06942cb53460cd574e3f0ba9c2d08731448d6e9eb48414587a610717eb41260f.jpg", + "table_caption": [ + "Table 5. Cross-dataset generalization evaluation on SMAL and SURREAL benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. The best and second-best outcomes highlighted in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodInputSMALSURREAL
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh5.9%0.24.3%0.3
GeoFMNet[9]Mesh//8.2%0.2
Diff-FMaps[24]Point//4.0%7.1
3D-CODED[11]Point0.5%19.22.1%8.1
Elementary[8]Point0.5%13.72.3%7.6
CorrNet3D[42]Point5.3%9.86.0%6.9
DPC[16]Point33.2%5.817.7%6.1
SE-ORNet[7]Point36.4%3.921.5%4.6
HSTR[13]Point33.9%5.619.4%5.6
OursPoint37.1%3.720.6%4.8
", + "bbox": [ + 514, + 261, + 874, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Cross-dataset Generalization Performance", + "text_level": 1, + "bbox": [ + 498, + 460, + 861, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The cross-dataset generalization ability is a crucial evaluation metric for the point cloud shape correspondence task. Following the conventional setup [7, 13, 16], generalization evaluation involves training on one dataset and testing on another. Specifically, training on the SMAL dataset and testing on the TOSCA dataset, as well as training on the SURREAL dataset and testing on the SHREC'19 dataset. As shown in Table 5, our method surpasses existing state-of-the-art approaches, establishing a new state-of-the-art performance by overcoming domain differences on the SMAL dataset. Regarding generalization to the SURREAL dataset, our results are also competitive. Confronting challenges posed by human body symmetry and directional rotations in the SURREAL and SHREC'19 datasets, our approach slightly lags behind the SE-ORNet [7], specifically designed for this issue. However, TANet significantly outperforms other point cloud shape correspondence methods.", + "bbox": [ + 496, + 488, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visual Comparison", + "text_level": 1, + "bbox": [ + 500, + 767, + 684, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 9, we present visual comparisons of our method with competitive approaches. When the source point cloud forms an unconventional shape, such as a distorted standing posture, other methods struggle to distinguish between the hand and facial structures. Our method more accurately establishes the correspondence for hands and heads by leveraging learned templates as an intermediary.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5256", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef4d4d5c61d29f5ec77470168d79de831b9f7069df0450dd0d794389b3c595c6.jpg", + "image_caption": [ + "Figure 9. Visual comparison on TOSCA and SHREC'19. We visually compared the correspondence results of different methods on unconventional shape pairs, providing intuitive validation of the proposed approach's superiority in establishing correspondences." + ], + "image_footnote": [], + "bbox": [ + 127, + 58, + 848, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8bd9378d45f79188f11a607e1576ae864e841ccd56c9d9102108903fe4983cca.jpg", + "image_caption": [ + "Figure 10. Visualization of the correspondence results on partial-shape SHREC'16 dataset. The partial shapes contain four species, including horses, cats, dogs, and humans. The generation of partial shapes in the SHREC'16 dataset primarily involves two methods: cutting and holing to wipe off body parts." + ], + "image_footnote": [], + "bbox": [ + 86, + 323, + 455, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Robustness Analysis", + "text_level": 1, + "bbox": [ + 76, + 559, + 269, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the robustness and generalization capabilities of our method, we directly evaluate our trained model on the SHREC'16 [6] and Owlii [40] datasets, presenting visual results. As shown in Figure 10, the SHREC'16 dataset comprises point cloud pairs of partial shapes, simulating challenges encountered in real-world applications such as occlusion, missing parts, and noise. Although only the upper body of the cat is present, and the dog is represented by a sparse point cloud with minimal parts of its body, our method discriminatively models local structures and accurately establishes shape correspondences. Moreover, the correspondence results of the partial human shape are predictions of the model trained on the animal dataset, providing further validation of the generalization capability. Illustrated in Figure 11, the Owlii dataset includes real-scanned human motion sequences, introducing challenges with scan noise and variations in attire and accessories. Nevertheless, TANet accurately establishes the shape correspondences for these sequence point clouds, including details such as dresses worn by the individuals. The visual results on SHREC'16 and Owlii further substantiate", + "bbox": [ + 75, + 583, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/406c42f169da4b3cc39bd9291aaf62463dbc91a0bdef455648ca9d2fe8400c11.jpg", + "image_caption": [ + "Figure 11. Visualization of the correspondence results on real scanned Owlii dataset. The results contain two consecutive action sequences involving individuals wearing different attire." + ], + "image_footnote": [], + "bbox": [ + 576, + 323, + 813, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TANet's superior robustness and generalization.", + "bbox": [ + 500, + 530, + 816, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 560, + 617, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an unsupervised template-assisted point cloud shape correspondence network, including a template generation module and a template assistance module. Specifically, the template generation module is introduced to learn a set of learnable templates that balance shape complexity and possess an explicitly structured shape. The template assistance module is designed to leverage the generated templates adaptively to achieve more accurate shape correspondences from template selection, feature learning, and transitive consistency perspectives. Extensive experiments on the SHREC'19 and TOSCA benchmarks demonstrate the superiority of TANet. Cross-dataset experiments on SURREAL and SMAL showcase our method's desirable generalization capabilities.", + "bbox": [ + 496, + 585, + 890, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 814, + 687, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by the National Defense Basic Scientific Research program (JCKY2022911B002) and the National Nature Science Foundation of China (NSFC 12150007, NSFC 62121002).", + "bbox": [ + 498, + 839, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5257", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fabio Arena, Mario Collotta, Giovanni Pau, and Francesco Terminate. An overview of augmented reality. Computers, 11 (2):28, 2022. 1", + "[2] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Generalized multidimensional scaling: a framework for isometry-invariant partial surface matching. Proceedings of the National Academy of Sciences, 103(5):1168-1172, 2006. 1, 2", + "[3] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 2, 5", + "[4] Dongliang Cao and Florian Bernard. Self-supervised learning for multimodal non-rigid 3d shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17735-17744, 2023. 2", + "[5] Etienne Corman, Maks Ovsjanikov, and Antonin Chambolle. Supervised descriptor learning for non-rigid shape matching. In European conference on computer vision, pages 283-298. Springer, 2014. 3", + "[6] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, Y Sahillioglu, et al. Shrec'16: Partial matching of deformable shapes. Proc. 3DOR, 2(9): 12, 2016. 5, 8", + "[7] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 1, 3, 5, 6, 7", + "[8] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. Advances in Neural Information Processing Systems, 32, 2019. 1, 7", + "[9] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8592-8601, 2020. 3, 7", + "[10] Georgia Gkioxari, Nikhila Ravi, and Justin Johnson. Learning 3d object shape and layout without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1695-1704, 2022. 3", + "[11] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 230-246, 2018. 1, 2, 5, 7", + "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 2", + "[13] Jianfeng He, Jiacheng Deng, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Hierarchical shape-consistent transformer for unsupervised point cloud shape correspondence. IEEE Transactions on Image Processing, 2023. 1, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, pages 1449-1457. Wiley Online Library, 2008. 1, 2", + "[15] Dany Laksono and Trias Aditya. Utilizing a game engine for interactive 3d topographic data visualization. ISPRS International Journal of Geo-Information, 8(8):361, 2019. 1", + "[16] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. Dpc: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 1, 3, 5, 6, 7", + "[17] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In European Conference on Computer Vision, pages 19-34. Springer, 2022. 3", + "[18] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In Proceedings of the IEEE international conference on computer vision, pages 5659-5667, 2017. 3", + "[19] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. IEEE transactions on pattern analysis and machine intelligence, 36(1):171-180, 2013. 1", + "[20] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 3", + "[21] Ming Liu. Robotic online path planning on point cloud. IEEE transactions on cybernetics, 46(5):1217-1228, 2015. 1", + "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[23] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 3", + "[24] Riccardo Marin, Marie-Julie Rakotosaona, Simone Melzi, and Maks Ovsjanikov. Correspondence learning via linearly-invariant embedding. Advances in Neural Information Processing Systems, 33:1608-1620, 2020. 1, 7", + "[25] Simone Melzi, Riccardo Marin, Emanuele Rodola, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, page 3, 2019. 2, 5", + "[26] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Learning 3d scene priors with 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792-802, 2023. 2, 3", + "[27] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5258", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6", + "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2", + "[31] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1617-1627, 2019. 6, 7", + "[32] Radu Bogdan Rusu, Zoltan Csaba Marton, Nico Blodow, Mihai Dolha, and Michael Beetz. Towards 3d point cloud based object maps for household environments. Robotics and Autonomous Systems, 56(11):927-941, 2008. 1", + "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3", + "[34] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 344-362. Springer, 2022. 3", + "[35] Art Tevs, Alexander Berner, Michael Wand, Ivo Ihrke, and H-P Seidel. Intrinsic shape matching by planned landmark sampling. In Computer graphics forum, pages 543-552. Wiley Online Library, 2011. 1, 2", + "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2", + "[37] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 2, 3, 7", + "[38] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 3", + "[39] Ziyi Wang, Yongming Rao, Xumin Yu, Jie Zhou, and Jiwen Lu. Semaffinet: Semantic-affine transformation for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11819-11829, 2022. 3", + "[40] Yi Xu, Yao Lu, and Ziyu Wen. Owli dynamic human mesh sequence dataset. In ISO/IEC JTC1/SC29/WG11 m41658, 120th MPEG Meeting, 2017. 5, 8" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 2", + "[42] Yiming Zeng, Yue Qian, Zhiyu Zhu, Junhui Hou, Hui Yuan, and Ying He. Corrnet3d: Unsupervised end-to-end learning of dense correspondence for 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6052-6061, 2021. 1, 3, 5, 6, 7", + "[43] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 2", + "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2", + "[45] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 1, 2, 5" + ], + "bbox": [ + 501, + 92, + 893, + 414 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5259", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_model.json b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_model.json new file mode 100644 index 0000000000000000000000000000000000000000..827b0b4887bf858e12f589afc8759f32d6712226 --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Unsupervised Template-assisted Point Cloud Shape Correspondence Network" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.18, + 0.676, + 0.2 + ], + "angle": 0, + "content": "Jiacheng Deng\\(^{1}\\), Jiahao Lu\\(^{1}\\), Tianzhu Zhang\\(^{1,2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.201, + 0.739, + 0.234 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2Deep Space Exploration Lab {dengjc, lujiahao} @mail.ustc.edu.cn, tzzhang@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.21, + 0.65 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.47, + 0.798 + ], + "angle": 0, + "content": "Point cloud shape correspondence is a challenging task that identifies and densely matches the source and target point clouds with deformable 3D shapes. The task has significant implications for various industries, including augmented reality [1], gaming [15], and robotics [21, 32]. However, the unrestricted mobility of humans and animals and their unusual postures have made direct correspondence between unconventional shapes a longstanding challenge in point cloud shape correspondence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.47, + 0.875 + ], + "angle": 0, + "content": "The shape correspondence problem has been thoroughly investigated for 3D mesh data. However, such spectral-based methods [2, 14, 19, 27, 35] are mainly limited by the time-consuming pre-processing step and the reliance on mesh vertical connectivity information, frequently absent" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.268, + 0.891, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.593 + ], + "angle": 0, + "content": "Figure 1. Visualization of template-assisted shape correspondence results. Correspondences are depicted by transferring colors from the source to the target based on matching results. The baseline incorrectly aligns the hands with the head and knees between unconventional shapes. In contrast, our method leverages the template to establish accurate correspondences for the hands." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "in point cloud scenarios. Point-based methods [7, 8, 11, 13, 16, 24, 42] are directly applied to the raw point cloud data, relying solely on point coordinates without connectivity information. To mitigate the resource-intensive labeling in fully supervised methods [8, 11, 24], the unsupervised point cloud shape correspondence task garners significant attention and yields competitive results [7, 13, 16, 42]. DPC [16] designs construction losses to reduce outlier matches. HSTR [13] incorporates whitening losses to mitigate the impact of shape discrepancies. SE-ORNet [7] introduces an orientation estimation module to alleviate symmetry issues. The shapes of humans and animals are highly diverse and complex. For instance, the SURREAL dataset [11] comprises \\(230\\mathrm{K}\\) human body shapes, while the SMAL dataset [45] contains \\(10\\mathrm{K}\\) animal shapes, including numerous unconventional forms. Establishing accurate correspondences between two unconventional point cloud shapes remains a challenging problem in shape correspondence. Current methods [7, 13, 16, 42] rely on computing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "†Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5250" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.228 + ], + "angle": 0, + "content": "the similarity between point cloud features to establish correspondences. However, in the case of unconventional point cloud pairs, the complete structure is distorted or destroyed due to limb adhesion, bending, and other postures. Without explicit structural guidance, directly calculating similarity is often insufficient to guarantee accurate correspondences. Therefore, it is essential to consider how to construct template shapes as intermediaries to assist in establishing correct correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.686 + ], + "angle": 0, + "content": "By analyzing prior shape correspondence methods, we have identified two pivotal issues that must be considered in shape correspondence. 1) How to generate suitable templates for point cloud pairs? Traditional point cloud template construction [11, 37, 41] mainly relies on manual selection or autoencoders [41]. While effective for shape-stable rigid objects in the same category, balancing template complexity becomes challenging for diverse non-rigid objects of different categories. Simple templates are clear but lack structural information, risking information loss during correspondence propagation. Complex templates offer richer information but are sensitive to noise, compromising accuracy. Despite attempts [26, 43] with learnable approaches, incomplete template structure remains a challenge. A more effective modeling approach is needed to generate templates suitable for intricate point cloud pairs. 2) How to effectively leverage templates for better correspondences? Establishing direct correspondences between unconventional shapes is challenging, but building correspondences between an unconventional shape and a suitable template is relatively straightforward. A correct mapping can be formed by finding corresponding template points for the source points and identifying corresponding points on the target point cloud based on the template. Direct similarity calculations between point clouds may introduce noise and yield ambiguous similarity weights. The template allows each point in the point cloud to compute a correlation vector, providing more stable information. Seeking consensus among these correlation vectors enhances appropriate point correspondences while mitigating erroneous ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.902 + ], + "angle": 0, + "content": "To achieve the above goal, we propose an Unsupervised Template-Assisted Point Cloud Shape Correspondence Network (TANet), including a template generation module and a template assistance module. In the template generation module, we incorporate a template bank with several learnable templates, effectively balancing template complexity in a data-driven manner. The space aligner enriches templates with comprehensive shape structures by establishing a mapping between the template feature space and the coordinate space. The space aligner is supervised by the coordinates and encoded features of the point cloud pairs. In the template assistance module, the adaptive selector chooses the most suitable template from the bank based on geometric and semantic attributes. The correlation fusion process" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "enhances point features with the template correlation vector via attention, suppressing noise and ambiguity. Moreover, we devise a transitive consistency loss to ensure coherence between the similarity computed through the template and directly computed between the source and target point clouds." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.183, + 0.893, + 0.41 + ], + "angle": 0, + "content": "To sum up, the contributions of this work can be summarized as follows: (i) We propose an unsupervised template-assisted point cloud shape correspondence network, achieved by jointly designing a template generation module and a template assistance module. (ii) We introduce a template generation module comprising a template bank with learnable templates and the space aligner for building explicit structures. The template assistance module adaptively selects a suitable template for robust point representations and maintains transitive consistency in shape correspondences. (iii) Extensive experiments on TOSCA [3] and SHREC'19 [25] benchmarks show the proposed TANet outperforming state-of-the-art methods. Cross-dataset experiments on SMAL [45] and SURREAL [11] demonstrate our method's desirable generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.642, + 0.438 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.448, + 0.892, + 0.508 + ], + "angle": 0, + "content": "In this Section, we give a brief overview of related works on point cloud shape correspondence, including deep learning for point clouds, shape correspondence, and point cloud template." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.509, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Deep Learning for Point Clouds. PointNet [29] employs a symmetrical function to aggregate all point features, resulting in a permutation-invariant global feature representation. Building upon this, PointNet++ [30] takes the concept further by incorporating local information, enhancing the representation capabilities and efficiency. DGCNN [38] introduces EdgeConv [38] blocks, dynamically updating neighborhood information based on dynamic graphs, ultimately delivering improved performance in point cloud analysis. Similarly, KPCnv [36] introduces point-wise convolution operators for point cloud feature learning. In the era of transformers, Point Cloud Transformer [12] stands out as a purely global Transformer network, pioneering self-attention layers instead of encoder layers within the PointNet framework. Point Transformer [44] operates within the local neighborhood of the target point cloud, enabling hierarchical extraction of geometric and semantic features." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.766, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Shape Correspondence. The goal of shape correspondence is to accurately find the point-to-point correspondence between two point clouds. Traditional spectral-based methods [2, 14, 27, 35] utilize mesh data to project features onto Laplace-Beltrami Operator (LBO) eigenbasis and then learn a transformation between the eigendecomposition of source and target shapes, which will be translated to a dense shape correspondence. Cao et al. [4] transfer rich structure and connectivity information in mesh data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.085, + 0.862, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.895, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Illustration of the TANet. TANet comprises an encoder, a template generation module, and a template assistance module. The template generation module produces several learnable shape templates in the template bank with a space aligner. The template assistance module selects suitable templates for point cloud pairs and improve accuracy via correlation fusion and transitive consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.473, + 0.712 + ], + "angle": 0, + "content": "to point cloud shape for robust matching. However, these methods are constrained by complex preprocessing and a reliance on vertex connectivity information. In contrast, point-based methods [5, 7, 9, 13, 16, 18, 34, 42] process point clouds directly without depending on connectivity information. CorrNet3D [42] and DPC [16] harness the established DGCNN [38] network to extract point representations through shape reconstruction. HSTR [13] has introduced a multi-receptive-field transformer point cloud encoder that considers local point cloud structures and long-range contextual information. SE-ORNet [7] incorporates an orientation estimation module to align the orientations of point cloud pairs, achieving precise matching results, particularly for symmetrical parts. These methods directly calculate point-wise similarities, making them susceptible to matching errors caused by noise and incomplete representations. In contrast, our approach constructs and utilizes templates to establish stable correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Point Cloud Template. Point cloud templates, such as shape templates, scene priors, and semantic priors, are pivotal in various point cloud tasks, including 6D pose estimation [17, 37], reconstruction [10, 20], scene synthesis [26], and segmentation [23, 33, 39]. Tian et al. [37] develop an autoencoder trained on a collection of object models, which decodes the mean latent embedding to create shape templates. For similar rigid objects, generative methods can yield robust and stable templates. ScenePrior [26] encodes scene priors into a latent space and utilizes sampling to synthesize plausible 3D scenes. SemAffiNet [39] and Mask3D [33] integrate a learnable implicit semantic prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.593 + ], + "angle": 0, + "content": "via transformer blocks with learnable queries. However, the complex shape and category variations pose a challenge for non-rigid objects to obtain well-balanced explicit templates through generative methods. The learnable implicit modeling struggles to achieve explicit stuctures. Based on the above discussion, we propose TANet, which integrates the template generation and assistance modules within a unified framework. Through technical designs, the learnable templates not only balance shape complexity but also possess an explicitly structured shape." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.593, + 0.621 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.631, + 0.611, + 0.645 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.837 + ], + "angle": 0, + "content": "The unsupervised point cloud correspondence task strives to establish accurate one-to-one correspondences between source and target point clouds without annotations. As illustrated in Figure 2, TANet primarily comprises three key modules: a general point cloud encoder, a Template Generation Module, and a Template Assistance Module. The choice of point cloud encoder follows the previous study [13]. The detailed design and optimization process for the Template Generation Module will be expounded upon in Section 3.2. Moreover, Section 3.3 will provide further insights into the Template Assistance Module. Finally, the training and inference process are discussed in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.765, + 0.864 + ], + "angle": 0, + "content": "3.2. Template Generation Module" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The template generation module consists primarily of two components: a template bank and a space aligner." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5252" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "Template bank. In the template bank, we integrate \\(K\\) learnable explicit shape templates, each comprising learnable positions \\(P_T^i \\in \\mathbb{R}^{N \\times 3}\\) and learnable embeddings \\(F_T^i \\in \\mathbb{R}^{N \\times d}\\). During the training phase, the positions and embeddings of templates are updated concurrently with the network. In the inference phase, the positions and embeddings of template shapes remain unchanged, eliminating the need for additional updates based on input point clouds. The learnable embeddings \\(F_T^i\\) parameters are initialized using Gaussian random initialization. To expedite template training, we utilize random shapes in datasets to initialize the learnable positions \\(P_T^i\\)." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.288, + 0.451, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Space aligner structure. The predicted point positions in the output are constrained to match the true point positions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.484 + ], + "angle": 0, + "content": "Space aligner. Space aligner \\(\\phi\\) is crafted to predict point positions through point embeddings, aiding in template learning. The specific structure is illustrated in Figure 3. The predicted point positions undergo adjustment to align with true positions:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.492, + 0.378, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {S m o o t h} \\mathrm {L} _ {1} (\\mathrm {P}, \\phi (\\mathrm {F})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where \\( P \\) are true positions, and \\( F \\) are point embeddings. The space aligner is trained with the encoded feature and the coordinates from point cloud pairs. During the inference phase, the space aligner is not involved in the computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.469, + 0.651 + ], + "angle": 0, + "content": "Additionally, the training of templates also includes the template construction loss \\(\\mathcal{L}_{tc}\\), which enforces consistency between the template shape and the cross-constructed template shape by point cloud pairs. Specifically, the cross-construction process is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.658, + 0.47, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\hat {t} _ {x _ {i}} = \\sum_ {j \\in \\mathcal {N} _ {T} (x _ {i})} \\frac {e ^ {s _ {i j}}}{\\sum_ {l \\in \\mathcal {N} _ {T} (x _ {i})} e ^ {s _ {i l}}} t _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.809 + ], + "angle": 0, + "content": "where \\( x_{i} \\in X \\), \\( t_{j} \\in P_{T} \\) and \\( s_{ij} \\in S_{XT} \\). \\( S_{XT} \\) is the similarity matrix between \\( F_{X} \\) and \\( F_{T} \\). \\( \\mathcal{N}_T(x_i) \\) represents the latent \\( k \\)-nearest neighbors of \\( x_{i} \\) in the target \\( F_{T} \\). The cross-construction of \\( P_{T} \\) by the source point cloud \\( X \\) is denoted \\( \\hat{T}_{X} \\in \\mathbb{R}^{N \\times 3} \\), where \\( \\hat{T}_{X}^{i} = \\hat{t}_{x_{i}} \\). After the introduction of the cross-construction process, the template construction loss \\( \\mathcal{L}_{tc} \\) can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.815, + 0.469, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t c} = \\mathbf {C D} (P _ {T}, \\hat {T} _ {X}) + \\mathbf {C D} (P _ {T}, \\hat {T} _ {Y}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where CD denotes chamfer distance. It is essential to highlight that the point encoder remains unaltered during the backpropagation of \\(\\mathcal{L}_{\\text {align }}\\) and \\(\\mathcal{L}_{\\text {tc }}\\) gradients, preventing interference with the feature extraction of point cloud pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.757, + 0.107 + ], + "angle": 0, + "content": "3.3. Template Assistance Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.175 + ], + "angle": 0, + "content": "The template assistance module aims to enhance the correspondence accuracy in point cloud pairs through templates. It comprises three primary components: Template Selector, Correlation Fusion, and Transitive Consistency." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.191, + 0.875, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Figure 4. Template selector structure. The best suitable template is determined through geometric and semantic measures of similarity between templates and point cloud pairs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.323, + 0.893, + 0.546 + ], + "angle": 0, + "content": "Template selector. The input to the Template Selector comprises point cloud coordinates \\( X \\& Y \\in \\mathbb{R}^{N \\times 3} \\) and features \\( F_{X} \\& F_{Y} \\in \\mathbb{R}^{N \\times d} \\), along with template point cloud coordinates \\( \\{P_T^i | i = 1,2,\\dots,K\\} \\) and embeddings \\( \\{F_T^i | i = 1,2,\\dots,K\\} \\) from the template bank. Geometric and semantic similarities between the point cloud pair and templates guide the adaptive template selection. As illustrated in Figure 4, it involves calculating the chamfer distance as the geometric similarity between the point cloud pair and template point cloud. To get semantic similarity, the point cloud features and template embeddings undergo mix pooling and cosine similarity computation. The fusion of geometric and semantic similarities is employed in Gumbel-softmax to ascertain the most fitting template \\( T \\) for the given point cloud pair. The specific computation formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.554, + 0.892, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T = \\mathbf {G S} \\left\\{\\left(\\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {X}\\right) + \\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {Y}\\right)\\right) \\right. \\tag {3} \\\\ + \\left(\\mathbf {C D} \\left(P _ {T} ^ {i}, X\\right) + \\mathbf {C D} \\left(P _ {T} ^ {i}, Y\\right)\\right) | i = 1, 2, \\dots , K \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.89, + 0.628 + ], + "angle": 0, + "content": "where GS represents Gumbel-Softmax, MP denotes Mix Pooing, and CD stands for chamfer distance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Correlation fusion. As depicted in Figure 5, template embeddings \\( F_{T} \\) and point cloud features \\( F_{X} / F_{Y} \\) are input into the correlation fusion process to enhance point cloud features. Initially, a correlation between the point cloud and the template is established. Subsequently, an attention mechanism incorporates the correlation vectors into point cloud features, yielding \\( F_{X}' / F_{Y}' \\in \\mathbb{R}^{N \\times d} \\). The similarity in the attention mechanism is computed based on point cloud features to underscore the spatial distribution of shapes, enhancing perception to spatial relationships. The introduction of template correlation aids in refining point cloud features and mitigating ambiguous representations in certain noisy point clouds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Transitive consistency. Utilizing the fused source features \\(F_{X}^{\\prime}\\) and template embeddings \\(F_{T}\\), we compute the similarity matrix \\(S_{XT}\\), and \\(S_{TY}\\) is obtained by calculating the similarity between \\(F_{T}\\) and \\(F_{Y}^{\\prime}\\). The multiplication of \\(S_{XT}\\) and \\(S_{TY}\\) matrices yields the similarity matrix \\(S_{X TY}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.091, + 0.434, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.276 + ], + "angle": 0, + "content": "Figure 5. The correlation fusion process. Features of point cloud pairs and embeddings of templates compute correlation vectors, which are then fused using an attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.368 + ], + "angle": 0, + "content": "reflecting the similarity between the source and target point clouds transmitted through the template. As shown in Figure 6, (a) Direct similarity illustrates the process of directly calculating the similarity between the source point cloud and the target point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.38, + 0.47, + 0.401 + ], + "angle": 0, + "content": "\\[\nS _ {X Y} = F _ {X} ^ {\\prime} F _ {Y} ^ {\\prime} ^ {\\top}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.471, + 0.669 + ], + "angle": 0, + "content": "In contrast, (b) Template-assisted similarity demonstrates the computation of \\( S_{X TY} \\). The multiplication of matrices \\( S_{XT} \\) and \\( S_{TY} \\) essentially results in an indirect computation of the similarity relationship between the source and target point clouds. This process involves a weighted sum of the similarity calculated between the template point cloud and the target point cloud, providing a more stable and accurate similarity result for source and target points. The Template-assisted similarity method helps mitigate the impact of outliers and noise on the similarity calculation, leading to more reliable results. In order to guide the establishment of more stable and accurate similarity relationships between point clouds without compromising the inference speed, we introduce the Transitive Consistency loss \\( (\\mathcal{L}_{trans}) \\), promoting the consistency between \\( S_{X TY} \\) and the similarity matrix \\( S_{XY} \\). The Transitive Consistency loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.674, + 0.469, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s}} = \\mathbf {C E} \\left(S _ {X T Y}, S _ {X Y}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.332, + 0.716 + ], + "angle": 0, + "content": "where CE means Cross-Entropy loss." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.733, + 0.428, + 0.837 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.838, + 0.468, + 0.893 + ], + "angle": 0, + "content": "Figure 6. Transitive Consistency. (a) the direct similarity computation between the source and target point clouds. (b) a stable template is employed to transmit the similarity relation. \\( L_{trans} \\) ensures the consistency between these two similarity results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.755, + 0.108 + ], + "angle": 0, + "content": "3.4. Model Training & Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Following the previous work [7, 13, 16], we also incorporate the construction loss \\((\\mathcal{L}_{\\text{const}})\\), encompassing both cross-construction and self-construction between the source shape \\(X\\) and the target shape \\(Y\\). The construction loss promotes feature smoothness and unique correspondences. The computation process and formula for cross-construction and self-construction are similar to Equation 1. Specific formulas are detailed in the supplementary material. To sum up, the total loss of TANet is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.263, + 0.892, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {t r a n s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {t c}} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t r}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.354 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}\\) are hyperparameters, balancing the contribution of different loss terms. The training of both the template and the network is jointly achieved, enabling end-to-end training without requiring additional prolonged iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.398 + ], + "angle": 0, + "content": "During inference, we set the closest point \\( y_{j^*} \\) in the feature space for each point \\( x_{i} \\) as its corresponding point. This selection rule can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.412, + 0.892, + 0.437 + ], + "angle": 0, + "content": "\\[\nf \\left(x _ {i}\\right) = y _ {j ^ {*}}, j ^ {*} = \\underset {j} {\\operatorname {a r g m a x}} \\left(s _ {i j}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.437, + 0.634, + 0.453 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.461, + 0.691, + 0.478 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on the most popular benchmarks, including the TOSCA [3] dataset and the SHREC'19 [25] dataset. TOSCA comprises 41 different shapes of various animal species. Following previous methods [13, 16], we pair these shapes to create both the training and testing sets. SHREC'19 consists of 44 real human models paired into 430 shape correspondence samples. SURREAL [11] and SMAL [45] are two large-scale datasets that we employ to evaluate the generalization capabilities of our method. The SURREAL dataset comprises 230,000 training shapes, from which we randomly sample and pair 2,000 shape pairs for training. The SMAL dataset includes parameterized animal models for generating shape pairs to train shape correspondence models. Furthermore, robustness experiments are conducted on the real-scanned Owlii dataset [40] and the partial shape dataset SHREC'16 [6]. Besides, we maintain a consistent point count of \\( n = 1024 \\) in accordance with prior studies [13, 16, 42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Evaluation Metrics. The primary evaluation metrics for the point cloud shape correspondence task include two key measures: average correspondence error and correspondence accuracy. The average correspondence error measures a source and target shape pair \\((X,Y)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.843, + 0.892, + 0.88 + ], + "angle": 0, + "content": "\\[\ne r r = \\frac {1}{n} \\sum_ {x _ {i} \\in X} | f (x _ {i}) - y _ {g t} | _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( y_{gt} \\in Y \\) represents the ground-truth matching point" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5254" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.457, + 0.107 + ], + "angle": 0, + "content": "to \\( x_{i} \\). Correspondence accuracy is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.115, + 0.47, + 0.152 + ], + "angle": 0, + "content": "\\[\na c c (\\epsilon) = \\frac {1}{n} \\sum_ {x _ {i} \\in X} \\mathbb {1} \\left(\\left| f \\left(x _ {i}\\right) - y _ {g t} \\right| _ {2} < \\epsilon d\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.469, + 0.205 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) denotes the indicator function, \\(d\\) is the maximum Euclidean distance between points in \\(Y\\), and \\(\\epsilon \\in [0,1]\\) signifies the error tolerance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.471, + 0.419 + ], + "angle": 0, + "content": "Implementation details. The number of learnable templates \\((K)\\) in the Template Bank is set to 4, with a template embedding dimension of 512. The Space Aligner incorporates two attention blocks, each with a dimensionality of 512. We employ a transformer encoder, akin to existing method [13], featuring four layers of attention blocks as our point cloud encoder. For the Correlation Fusion process, the dimensions of self-attention and MLP are set to 512. In Equation 6, \\(\\lambda_1\\), \\(\\lambda_2\\), \\(\\lambda_3\\), and \\(\\lambda_4\\) are set to 0.5, 0.5, 1, and 1. All experiments with our method on various datasets are conducted on a single GeForce RTX 3090 device using the PyTorch 1.10.1 framework [28]. Model training utilizes the AdamW [22] optimizer with a learning rate of 5e-4, a weight decay of 5e-4, and a batch size of 4." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Table 1. Comparison on TOSCA and SHREC'19 benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. Higher accuracy and lower error reflect a better result, with the best and second-best outcomes highlighted in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.503, + 0.451, + 0.622 + ], + "angle": 0, + "content": "
MethodInputTOSCASHREC'19
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh//5.9%0.2
CorrNet3D[42]Point0.3%32.70.4%33.8
DPC[16]Point34.7%2.815.3%5.6
SE-ORNet[7]Point38.3%2.717.5%5.1
HSTR[13]Point52.3%1.219.3%4.9
OursPoint65.1%0.721.5%4.5
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.442, + 0.652 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.749 + ], + "angle": 0, + "content": "We conducted comprehensive experimental comparisons with existing methods on the TOSCA animal dataset and the SHREC'19 human dataset. Furthermore, to ensure a fair comparison with existing methods, our approach is implemented without the use of post-processing or additional connectivity information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Evaluation on TOSCA dataset. As shown in Table 1, our approach demonstrates significant improvement on the TOSCA dataset, achieving a new state-of-the-art performance with a \\(12.8\\%\\) increase in accuracy and a minimal average error of \\(0.7\\mathrm{cm}\\). For a comprehensive depiction of accuracy across different error tolerances, we present correspondence accuracy in Figure 7(a). Similarly, our method outperforms others at different error tolerances." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Evaluation on SHREC'19 dataset. As indicated in Table 1, our method demonstrates significant improvement on" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.697, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.091, + 0.894, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.19, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Figure 7. Correspondence accuracy at various error tolerances. (a) Corresponding accuracy of different methods on the TOSCA dataset under various error tolerances. (b) Experiments similar to (a) conducted on the SHREC'19 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.37 + ], + "angle": 0, + "content": "the SHREC'19 dataset, achieving an impressive \\(2.2\\%\\) increase in accuracy and a reduction of \\(0.4\\mathrm{cm}\\) in error, consequently establishing a new state-of-the-art performance. Figure 7(b) illustrates the correspondence accuracy of various competitive methods at different error tolerances, showcasing our method's clear superiority across different levels of error tolerance." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.382, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Table 2. Evaluation of the model with different designs on TOSCA dataset. TGM denotes the template generation module, \\( L_{tc} \\) stands for template construction loss, TS represents template selector, CF denotes correlation fusion process, and \\( L_{trans} \\) is the transitive consistency loss." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.455, + 0.861, + 0.591 + ], + "angle": 0, + "content": "
TGMLtcTSCFLtransTOSCA
acc ↑err ↓
[A]XXXXX52.3%1.2
[B]✓XX✓X54.6%1.2
[C]✓XXX✓55.8%1.1
[D]✓✓X✓X59.3%1.0
[E]✓✓XX✓58.8%1.0
[F]✓✓✓✓X61.9%0.9
[G]✓✓✓X✓61.3%0.9
[H]✓✓✓✓✓65.1%0.7
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.654, + 0.621 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation of the model with different designs. In Table 2, we conduct extensive ablation studies on the TOSCA dataset to validate the effectiveness of our model designs. Specifically, [A] presents a baseline method without any template-related design. [B] and [C] demonstrate the employment of the template generation module, assisted respectively through correlation fusion and transitive consistency loss, resulting in accuracy improvements of \\(2.3\\%\\) and \\(3.5\\%\\). The limited improvement is attributed to the difficulty in structural learning of templates without appropriate loss constraints. [D] and [E] illustrate the introduction of template construction loss to aid in the learning and optimization of template structures, leading to advancements in accuracy and error compared to settings [B] and [C]. Building upon this, [F] and [G] further employ a template selector to identify the most suitable template for aiding the establishment of correspondence for specific shapes. The further improvements in accuracy and error indicate that adaptive" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "template selection significantly contributes to accurate correspondence establishment. Finally, in [H], we present the performance of the complete model, highlighting the essential roles played by different modules in collectively achieving accurate point cloud shape correspondences." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.181, + 0.455, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of template generation methods. The templates generated by the template generation module exhibit clearer structures in both animal and human templates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.611 + ], + "angle": 0, + "content": "Effectiveness of the template generation module. Figure 8 compares templates generated by our template generation module and the common autoencoder-based approach [37]. Our designed learnable templates, constrained by the space aligner and construction loss, provide more structural templates to aid point cloud shape correspondence. However, due to the significant shape variations of non-rigid bodies, it remains challenging to generate perfect templates without direct shape supervision. The existing templates are merely a sub-optimal form. Additionally, we carefully consider the number of templates in Table 3. A single or dual template falls short of capturing the shape variations, resulting in a significant decrease in accuracy. While eight templates can slightly enhance accuracy, they lead to a substantial increase in model parameters. Hence, we offer a trade-off to set up four learnable templates." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.623, + 0.461, + 0.651 + ], + "angle": 0, + "content": "Table 3. Ablation studies on Table 4. Comparison of simi-the number of templates. larity computation methods." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.655, + 0.262, + 0.73 + ], + "angle": 0, + "content": "
#TemplateParam.acc ↑
17.3M56.8%
27.8M61.4%
48.8M65.1%
810.9M66.2%
" + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.669, + 0.458, + 0.721 + ], + "angle": 0, + "content": "
SettingsFLOPs(G)TOSCA
acc ↑err ↓
Direct1.0765.1%0.73
Transitive4.2966.3%0.67
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of the template assistance module. The effectiveness of correlation fusion has been validated in Table 2. Here, we further delve into the discussion regarding the computation of similarities between point cloud pairs. The purpose of transitive consistency loss is to constrain the direct similarity computation between the source and target point clouds to align with the similarity computed through template transitivity. As demonstrated in Table 4, similarity computation through template propagation has been proven to be a more stable and accurate approach. However, considering the increased computational complexity associated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "with the large number of point clouds, we adopt the direct similarity computation method during inference to enhance model efficiency. The transitive consistency loss only constrains the learning of point cloud representations through similarity-guided learning during training." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Table 5. Cross-dataset generalization evaluation on SMAL and SURREAL benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. The best and second-best outcomes highlighted in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.262, + 0.875, + 0.428 + ], + "angle": 0, + "content": "
MethodInputSMALSURREAL
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh5.9%0.24.3%0.3
GeoFMNet[9]Mesh//8.2%0.2
Diff-FMaps[24]Point//4.0%7.1
3D-CODED[11]Point0.5%19.22.1%8.1
Elementary[8]Point0.5%13.72.3%7.6
CorrNet3D[42]Point5.3%9.86.0%6.9
DPC[16]Point33.2%5.817.7%6.1
SE-ORNet[7]Point36.4%3.921.5%4.6
HSTR[13]Point33.9%5.619.4%5.6
OursPoint37.1%3.720.6%4.8
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.462, + 0.862, + 0.477 + ], + "angle": 0, + "content": "4.4. Cross-dataset Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.747 + ], + "angle": 0, + "content": "The cross-dataset generalization ability is a crucial evaluation metric for the point cloud shape correspondence task. Following the conventional setup [7, 13, 16], generalization evaluation involves training on one dataset and testing on another. Specifically, training on the SMAL dataset and testing on the TOSCA dataset, as well as training on the SURREAL dataset and testing on the SHREC'19 dataset. As shown in Table 5, our method surpasses existing state-of-the-art approaches, establishing a new state-of-the-art performance by overcoming domain differences on the SMAL dataset. Regarding generalization to the SURREAL dataset, our results are also competitive. Confronting challenges posed by human body symmetry and directional rotations in the SURREAL and SHREC'19 datasets, our approach slightly lags behind the SE-ORNet [7], specifically designed for this issue. However, TANet significantly outperforms other point cloud shape correspondence methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.685, + 0.784 + ], + "angle": 0, + "content": "4.5. Visual Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Figure 9, we present visual comparisons of our method with competitive approaches. When the source point cloud forms an unconventional shape, such as a distorted standing posture, other methods struggle to distinguish between the hand and facial structures. Our method more accurately establishes the correspondence for hands and heads by leveraging learned templates as an intermediary." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.059, + 0.849, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 9. Visual comparison on TOSCA and SHREC'19. We visually compared the correspondence results of different methods on unconventional shape pairs, providing intuitive validation of the proposed approach's superiority in establishing correspondences." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.324, + 0.456, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.547 + ], + "angle": 0, + "content": "Figure 10. Visualization of the correspondence results on partial-shape SHREC'16 dataset. The partial shapes contain four species, including horses, cats, dogs, and humans. The generation of partial shapes in the SHREC'16 dataset primarily involves two methods: cutting and holing to wipe off body parts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.271, + 0.577 + ], + "angle": 0, + "content": "4.6. Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To further validate the robustness and generalization capabilities of our method, we directly evaluate our trained model on the SHREC'16 [6] and Owlii [40] datasets, presenting visual results. As shown in Figure 10, the SHREC'16 dataset comprises point cloud pairs of partial shapes, simulating challenges encountered in real-world applications such as occlusion, missing parts, and noise. Although only the upper body of the cat is present, and the dog is represented by a sparse point cloud with minimal parts of its body, our method discriminatively models local structures and accurately establishes shape correspondences. Moreover, the correspondence results of the partial human shape are predictions of the model trained on the animal dataset, providing further validation of the generalization capability. Illustrated in Figure 11, the Owlii dataset includes real-scanned human motion sequences, introducing challenges with scan noise and variations in attire and accessories. Nevertheless, TANet accurately establishes the shape correspondences for these sequence point clouds, including details such as dresses worn by the individuals. The visual results on SHREC'16 and Owlii further substantiate" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.324, + 0.814, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.479, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Figure 11. Visualization of the correspondence results on real scanned Owlii dataset. The results contain two consecutive action sequences involving individuals wearing different attire." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.531, + 0.818, + 0.546 + ], + "angle": 0, + "content": "TANet's superior robustness and generalization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.619, + 0.577 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.8 + ], + "angle": 0, + "content": "In this paper, we propose an unsupervised template-assisted point cloud shape correspondence network, including a template generation module and a template assistance module. Specifically, the template generation module is introduced to learn a set of learnable templates that balance shape complexity and possess an explicitly structured shape. The template assistance module is designed to leverage the generated templates adaptively to achieve more accurate shape correspondences from template selection, feature learning, and transitive consistency perspectives. Extensive experiments on the SHREC'19 and TOSCA benchmarks demonstrate the superiority of TANet. Cross-dataset experiments on SURREAL and SMAL showcase our method's desirable generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.688, + 0.832 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This work was partially supported by the National Defense Basic Scientific Research program (JCKY2022911B002) and the National Nature Science Foundation of China (NSFC 12150007, NSFC 62121002)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5257" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Fabio Arena, Mario Collotta, Giovanni Pau, and Francesco Terminate. An overview of augmented reality. Computers, 11 (2):28, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.469, + 0.226 + ], + "angle": 0, + "content": "[2] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Generalized multidimensional scaling: a framework for isometry-invariant partial surface matching. Proceedings of the National Academy of Sciences, 103(5):1168-1172, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[3] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] Dongliang Cao and Florian Bernard. Self-supervised learning for multimodal non-rigid 3d shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17735-17744, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[5] Etienne Corman, Maks Ovsjanikov, and Antonin Chambolle. Supervised descriptor learning for non-rigid shape matching. In European conference on computer vision, pages 283-298. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.384, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[6] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, Y Sahillioglu, et al. Shrec'16: Partial matching of deformable shapes. Proc. 3DOR, 2(9): 12, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[7] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[8] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. Advances in Neural Information Processing Systems, 32, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[9] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8592-8601, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[10] Georgia Gkioxari, Nikhila Ravi, and Justin Johnson. Learning 3d object shape and layout without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1695-1704, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[11] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 230-246, 2018. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Jianfeng He, Jiacheng Deng, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Hierarchical shape-consistent transformer for unsupervised point cloud shape correspondence. IEEE Transactions on Image Processing, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, pages 1449-1457. Wiley Online Library, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[15] Dany Laksono and Trias Aditya. Utilizing a game engine for interactive 3d topographic data visualization. ISPRS International Journal of Geo-Information, 8(8):361, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. Dpc: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[17] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In European Conference on Computer Vision, pages 19-34. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In Proceedings of the IEEE international conference on computer vision, pages 5659-5667, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[19] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. IEEE transactions on pattern analysis and machine intelligence, 36(1):171-180, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[20] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[21] Ming Liu. Robotic online path planning on point cloud. IEEE transactions on cybernetics, 46(5):1217-1228, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[23] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Riccardo Marin, Marie-Julie Rakotosaona, Simone Melzi, and Maks Ovsjanikov. Correspondence learning via linearly-invariant embedding. Advances in Neural Information Processing Systems, 33:1608-1620, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[25] Simone Melzi, Riccardo Marin, Emanuele Rodola, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, page 3, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Learning 3d scene priors with 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792-802, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5258" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.216 + ], + "angle": 0, + "content": "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[31] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1617-1627, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[32] Radu Bogdan Rusu, Zoltan Csaba Marton, Nico Blodow, Mihai Dolha, and Michael Beetz. Towards 3d point cloud based object maps for household environments. Robotics and Autonomous Systems, 56(11):927-941, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[34] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 344-362. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[35] Art Tevs, Alexander Berner, Michael Wand, Ivo Ihrke, and H-P Seidel. Intrinsic shape matching by planned landmark sampling. In Computer graphics forum, pages 543-552. Wiley Online Library, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[37] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[38] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[39] Ziyi Wang, Yongming Rao, Xumin Yu, Jie Zhou, and Jiwen Lu. Semaffinet: Semantic-affine transformation for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11819-11829, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[40] Yi Xu, Yao Lu, and Ziyu Wen. Owli dynamic human mesh sequence dataset. In ISO/IEC JTC1/SC29/WG11 m41658, 120th MPEG Meeting, 2017. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[41] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[42] Yiming Zeng, Yue Qian, Zhiyu Zhu, Junhui Hou, Hui Yuan, and Ying He. Corrnet3d: Unsupervised end-to-end learning of dense correspondence for 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6052-6061, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[43] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.894, + 0.415 + ], + "angle": 0, + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5259" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_origin.pdf b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..221cf1f19d79764d4b82c20e80a8432b8e211e02 --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/812a1d24-1f0f-400f-8c84-da6720a2a353_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fe83ea8b33d9397bed4ac91cd86f66f65127b1736f309fddaa6a7768b5df86d +size 8683506 diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/full.md b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd3d2562d1a56baadeec4f18fd89c2b4e136f1aa --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/full.md @@ -0,0 +1,293 @@ +# Unsupervised Template-assisted Point Cloud Shape Correspondence Network + +Jiacheng Deng $^{1}$ , Jiahao Lu $^{1}$ , Tianzhu Zhang $^{1,2,\dagger}$ + +1University of Science and Technology of China, 2Deep Space Exploration Lab {dengjc, lujiahao} @mail.ustc.edu.cn, tzzhang@ustc.edu.cn + +# Abstract + +Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods. + +# 1. Introduction + +Point cloud shape correspondence is a challenging task that identifies and densely matches the source and target point clouds with deformable 3D shapes. The task has significant implications for various industries, including augmented reality [1], gaming [15], and robotics [21, 32]. However, the unrestricted mobility of humans and animals and their unusual postures have made direct correspondence between unconventional shapes a longstanding challenge in point cloud shape correspondence. + +The shape correspondence problem has been thoroughly investigated for 3D mesh data. However, such spectral-based methods [2, 14, 19, 27, 35] are mainly limited by the time-consuming pre-processing step and the reliance on mesh vertical connectivity information, frequently absent + +![](images/0772c8e58c58b0dd860e9f0a0afe8d05508e32648dd23247f674e7574bcaf5ca.jpg) +Figure 1. Visualization of template-assisted shape correspondence results. Correspondences are depicted by transferring colors from the source to the target based on matching results. The baseline incorrectly aligns the hands with the head and knees between unconventional shapes. In contrast, our method leverages the template to establish accurate correspondences for the hands. + +in point cloud scenarios. Point-based methods [7, 8, 11, 13, 16, 24, 42] are directly applied to the raw point cloud data, relying solely on point coordinates without connectivity information. To mitigate the resource-intensive labeling in fully supervised methods [8, 11, 24], the unsupervised point cloud shape correspondence task garners significant attention and yields competitive results [7, 13, 16, 42]. DPC [16] designs construction losses to reduce outlier matches. HSTR [13] incorporates whitening losses to mitigate the impact of shape discrepancies. SE-ORNet [7] introduces an orientation estimation module to alleviate symmetry issues. The shapes of humans and animals are highly diverse and complex. For instance, the SURREAL dataset [11] comprises $230\mathrm{K}$ human body shapes, while the SMAL dataset [45] contains $10\mathrm{K}$ animal shapes, including numerous unconventional forms. Establishing accurate correspondences between two unconventional point cloud shapes remains a challenging problem in shape correspondence. Current methods [7, 13, 16, 42] rely on computing + +the similarity between point cloud features to establish correspondences. However, in the case of unconventional point cloud pairs, the complete structure is distorted or destroyed due to limb adhesion, bending, and other postures. Without explicit structural guidance, directly calculating similarity is often insufficient to guarantee accurate correspondences. Therefore, it is essential to consider how to construct template shapes as intermediaries to assist in establishing correct correspondences. + +By analyzing prior shape correspondence methods, we have identified two pivotal issues that must be considered in shape correspondence. 1) How to generate suitable templates for point cloud pairs? Traditional point cloud template construction [11, 37, 41] mainly relies on manual selection or autoencoders [41]. While effective for shape-stable rigid objects in the same category, balancing template complexity becomes challenging for diverse non-rigid objects of different categories. Simple templates are clear but lack structural information, risking information loss during correspondence propagation. Complex templates offer richer information but are sensitive to noise, compromising accuracy. Despite attempts [26, 43] with learnable approaches, incomplete template structure remains a challenge. A more effective modeling approach is needed to generate templates suitable for intricate point cloud pairs. 2) How to effectively leverage templates for better correspondences? Establishing direct correspondences between unconventional shapes is challenging, but building correspondences between an unconventional shape and a suitable template is relatively straightforward. A correct mapping can be formed by finding corresponding template points for the source points and identifying corresponding points on the target point cloud based on the template. Direct similarity calculations between point clouds may introduce noise and yield ambiguous similarity weights. The template allows each point in the point cloud to compute a correlation vector, providing more stable information. Seeking consensus among these correlation vectors enhances appropriate point correspondences while mitigating erroneous ones. + +To achieve the above goal, we propose an Unsupervised Template-Assisted Point Cloud Shape Correspondence Network (TANet), including a template generation module and a template assistance module. In the template generation module, we incorporate a template bank with several learnable templates, effectively balancing template complexity in a data-driven manner. The space aligner enriches templates with comprehensive shape structures by establishing a mapping between the template feature space and the coordinate space. The space aligner is supervised by the coordinates and encoded features of the point cloud pairs. In the template assistance module, the adaptive selector chooses the most suitable template from the bank based on geometric and semantic attributes. The correlation fusion process + +enhances point features with the template correlation vector via attention, suppressing noise and ambiguity. Moreover, we devise a transitive consistency loss to ensure coherence between the similarity computed through the template and directly computed between the source and target point clouds. + +To sum up, the contributions of this work can be summarized as follows: (i) We propose an unsupervised template-assisted point cloud shape correspondence network, achieved by jointly designing a template generation module and a template assistance module. (ii) We introduce a template generation module comprising a template bank with learnable templates and the space aligner for building explicit structures. The template assistance module adaptively selects a suitable template for robust point representations and maintains transitive consistency in shape correspondences. (iii) Extensive experiments on TOSCA [3] and SHREC'19 [25] benchmarks show the proposed TANet outperforming state-of-the-art methods. Cross-dataset experiments on SMAL [45] and SURREAL [11] demonstrate our method's desirable generalization capabilities. + +# 2. Related Work + +In this Section, we give a brief overview of related works on point cloud shape correspondence, including deep learning for point clouds, shape correspondence, and point cloud template. + +Deep Learning for Point Clouds. PointNet [29] employs a symmetrical function to aggregate all point features, resulting in a permutation-invariant global feature representation. Building upon this, PointNet++ [30] takes the concept further by incorporating local information, enhancing the representation capabilities and efficiency. DGCNN [38] introduces EdgeConv [38] blocks, dynamically updating neighborhood information based on dynamic graphs, ultimately delivering improved performance in point cloud analysis. Similarly, KPCnv [36] introduces point-wise convolution operators for point cloud feature learning. In the era of transformers, Point Cloud Transformer [12] stands out as a purely global Transformer network, pioneering self-attention layers instead of encoder layers within the PointNet framework. Point Transformer [44] operates within the local neighborhood of the target point cloud, enabling hierarchical extraction of geometric and semantic features. + +Shape Correspondence. The goal of shape correspondence is to accurately find the point-to-point correspondence between two point clouds. Traditional spectral-based methods [2, 14, 27, 35] utilize mesh data to project features onto Laplace-Beltrami Operator (LBO) eigenbasis and then learn a transformation between the eigendecomposition of source and target shapes, which will be translated to a dense shape correspondence. Cao et al. [4] transfer rich structure and connectivity information in mesh data + +![](images/94d278b8f0b473e22f8c68b2901823f9660c726abd02bbcb6f5f0c695b1decf2.jpg) +Figure 2. Illustration of the TANet. TANet comprises an encoder, a template generation module, and a template assistance module. The template generation module produces several learnable shape templates in the template bank with a space aligner. The template assistance module selects suitable templates for point cloud pairs and improve accuracy via correlation fusion and transitive consistency. + +to point cloud shape for robust matching. However, these methods are constrained by complex preprocessing and a reliance on vertex connectivity information. In contrast, point-based methods [5, 7, 9, 13, 16, 18, 34, 42] process point clouds directly without depending on connectivity information. CorrNet3D [42] and DPC [16] harness the established DGCNN [38] network to extract point representations through shape reconstruction. HSTR [13] has introduced a multi-receptive-field transformer point cloud encoder that considers local point cloud structures and long-range contextual information. SE-ORNet [7] incorporates an orientation estimation module to align the orientations of point cloud pairs, achieving precise matching results, particularly for symmetrical parts. These methods directly calculate point-wise similarities, making them susceptible to matching errors caused by noise and incomplete representations. In contrast, our approach constructs and utilizes templates to establish stable correspondences. + +Point Cloud Template. Point cloud templates, such as shape templates, scene priors, and semantic priors, are pivotal in various point cloud tasks, including 6D pose estimation [17, 37], reconstruction [10, 20], scene synthesis [26], and segmentation [23, 33, 39]. Tian et al. [37] develop an autoencoder trained on a collection of object models, which decodes the mean latent embedding to create shape templates. For similar rigid objects, generative methods can yield robust and stable templates. ScenePrior [26] encodes scene priors into a latent space and utilizes sampling to synthesize plausible 3D scenes. SemAffiNet [39] and Mask3D [33] integrate a learnable implicit semantic prior + +via transformer blocks with learnable queries. However, the complex shape and category variations pose a challenge for non-rigid objects to obtain well-balanced explicit templates through generative methods. The learnable implicit modeling struggles to achieve explicit stuctures. Based on the above discussion, we propose TANet, which integrates the template generation and assistance modules within a unified framework. Through technical designs, the learnable templates not only balance shape complexity but also possess an explicitly structured shape. + +# 3. Method + +# 3.1. Overview + +The unsupervised point cloud correspondence task strives to establish accurate one-to-one correspondences between source and target point clouds without annotations. As illustrated in Figure 2, TANet primarily comprises three key modules: a general point cloud encoder, a Template Generation Module, and a Template Assistance Module. The choice of point cloud encoder follows the previous study [13]. The detailed design and optimization process for the Template Generation Module will be expounded upon in Section 3.2. Moreover, Section 3.3 will provide further insights into the Template Assistance Module. Finally, the training and inference process are discussed in Section 3.4. + +# 3.2. Template Generation Module + +The template generation module consists primarily of two components: a template bank and a space aligner. + +Template bank. In the template bank, we integrate $K$ learnable explicit shape templates, each comprising learnable positions $P_T^i \in \mathbb{R}^{N \times 3}$ and learnable embeddings $F_T^i \in \mathbb{R}^{N \times d}$ . During the training phase, the positions and embeddings of templates are updated concurrently with the network. In the inference phase, the positions and embeddings of template shapes remain unchanged, eliminating the need for additional updates based on input point clouds. The learnable embeddings $F_T^i$ parameters are initialized using Gaussian random initialization. To expedite template training, we utilize random shapes in datasets to initialize the learnable positions $P_T^i$ . + +![](images/6850f72c935869667dbeeb3e14b4448529cf987b3fa7fca30c613c1629cae051.jpg) +Figure 3. Space aligner structure. The predicted point positions in the output are constrained to match the true point positions. + +Space aligner. Space aligner $\phi$ is crafted to predict point positions through point embeddings, aiding in template learning. The specific structure is illustrated in Figure 3. The predicted point positions undergo adjustment to align with true positions: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \operatorname {S m o o t h} \mathrm {L} _ {1} (\mathrm {P}, \phi (\mathrm {F})), +$$ + +where $P$ are true positions, and $F$ are point embeddings. The space aligner is trained with the encoded feature and the coordinates from point cloud pairs. During the inference phase, the space aligner is not involved in the computation. + +Additionally, the training of templates also includes the template construction loss $\mathcal{L}_{tc}$ , which enforces consistency between the template shape and the cross-constructed template shape by point cloud pairs. Specifically, the cross-construction process is computed as follows: + +$$ +\hat {t} _ {x _ {i}} = \sum_ {j \in \mathcal {N} _ {T} (x _ {i})} \frac {e ^ {s _ {i j}}}{\sum_ {l \in \mathcal {N} _ {T} (x _ {i})} e ^ {s _ {i l}}} t _ {j}, \tag {1} +$$ + +where $x_{i} \in X$ , $t_{j} \in P_{T}$ and $s_{ij} \in S_{XT}$ . $S_{XT}$ is the similarity matrix between $F_{X}$ and $F_{T}$ . $\mathcal{N}_T(x_i)$ represents the latent $k$ -nearest neighbors of $x_{i}$ in the target $F_{T}$ . The cross-construction of $P_{T}$ by the source point cloud $X$ is denoted $\hat{T}_{X} \in \mathbb{R}^{N \times 3}$ , where $\hat{T}_{X}^{i} = \hat{t}_{x_{i}}$ . After the introduction of the cross-construction process, the template construction loss $\mathcal{L}_{tc}$ can be computed as follows: + +$$ +\mathcal {L} _ {t c} = \mathbf {C D} (P _ {T}, \hat {T} _ {X}) + \mathbf {C D} (P _ {T}, \hat {T} _ {Y}), \tag {2} +$$ + +where CD denotes chamfer distance. It is essential to highlight that the point encoder remains unaltered during the backpropagation of $\mathcal{L}_{\text {align }}$ and $\mathcal{L}_{\text {tc }}$ gradients, preventing interference with the feature extraction of point cloud pairs. + +# 3.3. Template Assistance Module + +The template assistance module aims to enhance the correspondence accuracy in point cloud pairs through templates. It comprises three primary components: Template Selector, Correlation Fusion, and Transitive Consistency. + +![](images/a36ecdcbaace7a50176579398b88fc58035bc5865d574c19b32e43347ab92ea3.jpg) +Figure 4. Template selector structure. The best suitable template is determined through geometric and semantic measures of similarity between templates and point cloud pairs. + +Template selector. The input to the Template Selector comprises point cloud coordinates $X \& Y \in \mathbb{R}^{N \times 3}$ and features $F_{X} \& F_{Y} \in \mathbb{R}^{N \times d}$ , along with template point cloud coordinates $\{P_T^i | i = 1,2,\dots,K\}$ and embeddings $\{F_T^i | i = 1,2,\dots,K\}$ from the template bank. Geometric and semantic similarities between the point cloud pair and templates guide the adaptive template selection. As illustrated in Figure 4, it involves calculating the chamfer distance as the geometric similarity between the point cloud pair and template point cloud. To get semantic similarity, the point cloud features and template embeddings undergo mix pooling and cosine similarity computation. The fusion of geometric and semantic similarities is employed in Gumbel-softmax to ascertain the most fitting template $T$ for the given point cloud pair. The specific computation formula is as follows: + +$$ +\begin{array}{l} T = \mathbf {G S} \left\{\left(\mathbf {M P} \left(F _ {T} ^ {i}\right) \cdot \mathbf {M P} \left(F _ {X}\right) + \mathbf {M P} \left(F _ {T} ^ {i}\right) \cdot \mathbf {M P} \left(F _ {Y}\right)\right) \right. \tag {3} \\ + \left(\mathbf {C D} \left(P _ {T} ^ {i}, X\right) + \mathbf {C D} \left(P _ {T} ^ {i}, Y\right)\right) | i = 1, 2, \dots , K \}, \\ \end{array} +$$ + +where GS represents Gumbel-Softmax, MP denotes Mix Pooing, and CD stands for chamfer distance. + +Correlation fusion. As depicted in Figure 5, template embeddings $F_{T}$ and point cloud features $F_{X} / F_{Y}$ are input into the correlation fusion process to enhance point cloud features. Initially, a correlation between the point cloud and the template is established. Subsequently, an attention mechanism incorporates the correlation vectors into point cloud features, yielding $F_{X}' / F_{Y}' \in \mathbb{R}^{N \times d}$ . The similarity in the attention mechanism is computed based on point cloud features to underscore the spatial distribution of shapes, enhancing perception to spatial relationships. The introduction of template correlation aids in refining point cloud features and mitigating ambiguous representations in certain noisy point clouds. + +Transitive consistency. Utilizing the fused source features $F_{X}^{\prime}$ and template embeddings $F_{T}$ , we compute the similarity matrix $S_{XT}$ , and $S_{TY}$ is obtained by calculating the similarity between $F_{T}$ and $F_{Y}^{\prime}$ . The multiplication of $S_{XT}$ and $S_{TY}$ matrices yields the similarity matrix $S_{X TY}$ , + +![](images/ba78a9763622f2cefb092af425fcf3c9be703dda78d6f723475d6dc063c8b69b.jpg) +Figure 5. The correlation fusion process. Features of point cloud pairs and embeddings of templates compute correlation vectors, which are then fused using an attention mechanism. + +reflecting the similarity between the source and target point clouds transmitted through the template. As shown in Figure 6, (a) Direct similarity illustrates the process of directly calculating the similarity between the source point cloud and the target point cloud: + +$$ +S _ {X Y} = F _ {X} ^ {\prime} F _ {Y} ^ {\prime} ^ {\top}. \tag {4} +$$ + +In contrast, (b) Template-assisted similarity demonstrates the computation of $S_{X TY}$ . The multiplication of matrices $S_{XT}$ and $S_{TY}$ essentially results in an indirect computation of the similarity relationship between the source and target point clouds. This process involves a weighted sum of the similarity calculated between the template point cloud and the target point cloud, providing a more stable and accurate similarity result for source and target points. The Template-assisted similarity method helps mitigate the impact of outliers and noise on the similarity calculation, leading to more reliable results. In order to guide the establishment of more stable and accurate similarity relationships between point clouds without compromising the inference speed, we introduce the Transitive Consistency loss $(\mathcal{L}_{trans})$ , promoting the consistency between $S_{X TY}$ and the similarity matrix $S_{XY}$ . The Transitive Consistency loss is calculated as follows: + +$$ +\mathcal {L} _ {\text {t r a n s}} = \mathbf {C E} \left(S _ {X T Y}, S _ {X Y}\right), \tag {5} +$$ + +where CE means Cross-Entropy loss. + +![](images/eeb0efd0caf43b60588bdbda8d510ddfb437bb5546e3ae5b85d17aeccd60a094.jpg) +Figure 6. Transitive Consistency. (a) the direct similarity computation between the source and target point clouds. (b) a stable template is employed to transmit the similarity relation. $L_{trans}$ ensures the consistency between these two similarity results. + +# 3.4. Model Training & Inference + +Following the previous work [7, 13, 16], we also incorporate the construction loss $(\mathcal{L}_{\text{const}})$ , encompassing both cross-construction and self-construction between the source shape $X$ and the target shape $Y$ . The construction loss promotes feature smoothness and unique correspondences. The computation process and formula for cross-construction and self-construction are similar to Equation 1. Specific formulas are detailed in the supplementary material. To sum up, the total loss of TANet is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {t r a n s}} + \lambda_ {2} \mathcal {L} _ {\text {a l i g n}} + \lambda_ {3} \mathcal {L} _ {\text {t c}} + \lambda_ {4} \mathcal {L} _ {\text {c o n s t r}}, \tag {6} +$$ + +where $\lambda_{i}$ are hyperparameters, balancing the contribution of different loss terms. The training of both the template and the network is jointly achieved, enabling end-to-end training without requiring additional prolonged iterations. + +During inference, we set the closest point $y_{j^*}$ in the feature space for each point $x_{i}$ as its corresponding point. This selection rule can be formulated as: + +$$ +f \left(x _ {i}\right) = y _ {j ^ {*}}, j ^ {*} = \underset {j} {\operatorname {a r g m a x}} \left(s _ {i j}\right). \tag {7} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on the most popular benchmarks, including the TOSCA [3] dataset and the SHREC'19 [25] dataset. TOSCA comprises 41 different shapes of various animal species. Following previous methods [13, 16], we pair these shapes to create both the training and testing sets. SHREC'19 consists of 44 real human models paired into 430 shape correspondence samples. SURREAL [11] and SMAL [45] are two large-scale datasets that we employ to evaluate the generalization capabilities of our method. The SURREAL dataset comprises 230,000 training shapes, from which we randomly sample and pair 2,000 shape pairs for training. The SMAL dataset includes parameterized animal models for generating shape pairs to train shape correspondence models. Furthermore, robustness experiments are conducted on the real-scanned Owlii dataset [40] and the partial shape dataset SHREC'16 [6]. Besides, we maintain a consistent point count of $n = 1024$ in accordance with prior studies [13, 16, 42]. + +Evaluation Metrics. The primary evaluation metrics for the point cloud shape correspondence task include two key measures: average correspondence error and correspondence accuracy. The average correspondence error measures a source and target shape pair $(X,Y)$ as follows: + +$$ +e r r = \frac {1}{n} \sum_ {x _ {i} \in X} | f (x _ {i}) - y _ {g t} | _ {2}, \tag {8} +$$ + +where $y_{gt} \in Y$ represents the ground-truth matching point + +to $x_{i}$ . Correspondence accuracy is formulated as follows: + +$$ +a c c (\epsilon) = \frac {1}{n} \sum_ {x _ {i} \in X} \mathbb {1} \left(\left| f \left(x _ {i}\right) - y _ {g t} \right| _ {2} < \epsilon d\right), \tag {9} +$$ + +where $\mathbb{1}(\cdot)$ denotes the indicator function, $d$ is the maximum Euclidean distance between points in $Y$ , and $\epsilon \in [0,1]$ signifies the error tolerance. + +Implementation details. The number of learnable templates $(K)$ in the Template Bank is set to 4, with a template embedding dimension of 512. The Space Aligner incorporates two attention blocks, each with a dimensionality of 512. We employ a transformer encoder, akin to existing method [13], featuring four layers of attention blocks as our point cloud encoder. For the Correlation Fusion process, the dimensions of self-attention and MLP are set to 512. In Equation 6, $\lambda_1$ , $\lambda_2$ , $\lambda_3$ , and $\lambda_4$ are set to 0.5, 0.5, 1, and 1. All experiments with our method on various datasets are conducted on a single GeForce RTX 3090 device using the PyTorch 1.10.1 framework [28]. Model training utilizes the AdamW [22] optimizer with a learning rate of 5e-4, a weight decay of 5e-4, and a batch size of 4. + +Table 1. Comparison on TOSCA and SHREC'19 benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. Higher accuracy and lower error reflect a better result, with the best and second-best outcomes highlighted in bold and underlined, respectively. + +
MethodInputTOSCASHREC'19
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh//5.9%0.2
CorrNet3D[42]Point0.3%32.70.4%33.8
DPC[16]Point34.7%2.815.3%5.6
SE-ORNet[7]Point38.3%2.717.5%5.1
HSTR[13]Point52.3%1.219.3%4.9
OursPoint65.1%0.721.5%4.5
+ +# 4.2. Comparison with State-of-the-art Methods + +We conducted comprehensive experimental comparisons with existing methods on the TOSCA animal dataset and the SHREC'19 human dataset. Furthermore, to ensure a fair comparison with existing methods, our approach is implemented without the use of post-processing or additional connectivity information. + +Evaluation on TOSCA dataset. As shown in Table 1, our approach demonstrates significant improvement on the TOSCA dataset, achieving a new state-of-the-art performance with a $12.8\%$ increase in accuracy and a minimal average error of $0.7\mathrm{cm}$ . For a comprehensive depiction of accuracy across different error tolerances, we present correspondence accuracy in Figure 7(a). Similarly, our method outperforms others at different error tolerances. + +Evaluation on SHREC'19 dataset. As indicated in Table 1, our method demonstrates significant improvement on + +![](images/973e5a12c8854770ddef2485edbbd74a74c4ba1f3ddb597d653e2880535b005e.jpg) +Figure 7. Correspondence accuracy at various error tolerances. (a) Corresponding accuracy of different methods on the TOSCA dataset under various error tolerances. (b) Experiments similar to (a) conducted on the SHREC'19 dataset. + +![](images/14bc645f377329aa3c2e07927c9ffcbe223ece013ae02a6a8e0f219c5b8b7e6c.jpg) + +the SHREC'19 dataset, achieving an impressive $2.2\%$ increase in accuracy and a reduction of $0.4\mathrm{cm}$ in error, consequently establishing a new state-of-the-art performance. Figure 7(b) illustrates the correspondence accuracy of various competitive methods at different error tolerances, showcasing our method's clear superiority across different levels of error tolerance. + +Table 2. Evaluation of the model with different designs on TOSCA dataset. TGM denotes the template generation module, $L_{tc}$ stands for template construction loss, TS represents template selector, CF denotes correlation fusion process, and $L_{trans}$ is the transitive consistency loss. + +
TGMLtcTSCFLtransTOSCA
acc ↑err ↓
[A]XXXXX52.3%1.2
[B]✓XX✓X54.6%1.2
[C]✓XXX✓55.8%1.1
[D]✓✓X✓X59.3%1.0
[E]✓✓XX✓58.8%1.0
[F]✓✓✓✓X61.9%0.9
[G]✓✓✓X✓61.3%0.9
[H]✓✓✓✓✓65.1%0.7
+ +# 4.3. Ablation Study + +Evaluation of the model with different designs. In Table 2, we conduct extensive ablation studies on the TOSCA dataset to validate the effectiveness of our model designs. Specifically, [A] presents a baseline method without any template-related design. [B] and [C] demonstrate the employment of the template generation module, assisted respectively through correlation fusion and transitive consistency loss, resulting in accuracy improvements of $2.3\%$ and $3.5\%$ . The limited improvement is attributed to the difficulty in structural learning of templates without appropriate loss constraints. [D] and [E] illustrate the introduction of template construction loss to aid in the learning and optimization of template structures, leading to advancements in accuracy and error compared to settings [B] and [C]. Building upon this, [F] and [G] further employ a template selector to identify the most suitable template for aiding the establishment of correspondence for specific shapes. The further improvements in accuracy and error indicate that adaptive + +template selection significantly contributes to accurate correspondence establishment. Finally, in [H], we present the performance of the complete model, highlighting the essential roles played by different modules in collectively achieving accurate point cloud shape correspondences. + +![](images/ce5a77a6d3d58d13dcafe8c3a9abde9b8cf437b05f656718547315dd4e84a1bc.jpg) +Figure 8. Visual comparison of template generation methods. The templates generated by the template generation module exhibit clearer structures in both animal and human templates. + +Effectiveness of the template generation module. Figure 8 compares templates generated by our template generation module and the common autoencoder-based approach [37]. Our designed learnable templates, constrained by the space aligner and construction loss, provide more structural templates to aid point cloud shape correspondence. However, due to the significant shape variations of non-rigid bodies, it remains challenging to generate perfect templates without direct shape supervision. The existing templates are merely a sub-optimal form. Additionally, we carefully consider the number of templates in Table 3. A single or dual template falls short of capturing the shape variations, resulting in a significant decrease in accuracy. While eight templates can slightly enhance accuracy, they lead to a substantial increase in model parameters. Hence, we offer a trade-off to set up four learnable templates. + +Table 3. Ablation studies on Table 4. Comparison of simi-the number of templates. larity computation methods. + +
#TemplateParam.acc ↑
17.3M56.8%
27.8M61.4%
48.8M65.1%
810.9M66.2%
+ +
SettingsFLOPs(G)TOSCA
acc ↑err ↓
Direct1.0765.1%0.73
Transitive4.2966.3%0.67
+ +Effectiveness of the template assistance module. The effectiveness of correlation fusion has been validated in Table 2. Here, we further delve into the discussion regarding the computation of similarities between point cloud pairs. The purpose of transitive consistency loss is to constrain the direct similarity computation between the source and target point clouds to align with the similarity computed through template transitivity. As demonstrated in Table 4, similarity computation through template propagation has been proven to be a more stable and accurate approach. However, considering the increased computational complexity associated + +with the large number of point clouds, we adopt the direct similarity computation method during inference to enhance model efficiency. The transitive consistency loss only constrains the learning of point cloud representations through similarity-guided learning during training. + +Table 5. Cross-dataset generalization evaluation on SMAL and SURREAL benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. The best and second-best outcomes highlighted in bold and underlined, respectively. + +
MethodInputSMALSURREAL
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh5.9%0.24.3%0.3
GeoFMNet[9]Mesh//8.2%0.2
Diff-FMaps[24]Point//4.0%7.1
3D-CODED[11]Point0.5%19.22.1%8.1
Elementary[8]Point0.5%13.72.3%7.6
CorrNet3D[42]Point5.3%9.86.0%6.9
DPC[16]Point33.2%5.817.7%6.1
SE-ORNet[7]Point36.4%3.921.5%4.6
HSTR[13]Point33.9%5.619.4%5.6
OursPoint37.1%3.720.6%4.8
+ +# 4.4. Cross-dataset Generalization Performance + +The cross-dataset generalization ability is a crucial evaluation metric for the point cloud shape correspondence task. Following the conventional setup [7, 13, 16], generalization evaluation involves training on one dataset and testing on another. Specifically, training on the SMAL dataset and testing on the TOSCA dataset, as well as training on the SURREAL dataset and testing on the SHREC'19 dataset. As shown in Table 5, our method surpasses existing state-of-the-art approaches, establishing a new state-of-the-art performance by overcoming domain differences on the SMAL dataset. Regarding generalization to the SURREAL dataset, our results are also competitive. Confronting challenges posed by human body symmetry and directional rotations in the SURREAL and SHREC'19 datasets, our approach slightly lags behind the SE-ORNet [7], specifically designed for this issue. However, TANet significantly outperforms other point cloud shape correspondence methods. + +# 4.5. Visual Comparison + +In Figure 9, we present visual comparisons of our method with competitive approaches. When the source point cloud forms an unconventional shape, such as a distorted standing posture, other methods struggle to distinguish between the hand and facial structures. Our method more accurately establishes the correspondence for hands and heads by leveraging learned templates as an intermediary. + +![](images/ef4d4d5c61d29f5ec77470168d79de831b9f7069df0450dd0d794389b3c595c6.jpg) +Figure 9. Visual comparison on TOSCA and SHREC'19. We visually compared the correspondence results of different methods on unconventional shape pairs, providing intuitive validation of the proposed approach's superiority in establishing correspondences. + +![](images/8bd9378d45f79188f11a607e1576ae864e841ccd56c9d9102108903fe4983cca.jpg) +Figure 10. Visualization of the correspondence results on partial-shape SHREC'16 dataset. The partial shapes contain four species, including horses, cats, dogs, and humans. The generation of partial shapes in the SHREC'16 dataset primarily involves two methods: cutting and holing to wipe off body parts. + +# 4.6. Robustness Analysis + +To further validate the robustness and generalization capabilities of our method, we directly evaluate our trained model on the SHREC'16 [6] and Owlii [40] datasets, presenting visual results. As shown in Figure 10, the SHREC'16 dataset comprises point cloud pairs of partial shapes, simulating challenges encountered in real-world applications such as occlusion, missing parts, and noise. Although only the upper body of the cat is present, and the dog is represented by a sparse point cloud with minimal parts of its body, our method discriminatively models local structures and accurately establishes shape correspondences. Moreover, the correspondence results of the partial human shape are predictions of the model trained on the animal dataset, providing further validation of the generalization capability. Illustrated in Figure 11, the Owlii dataset includes real-scanned human motion sequences, introducing challenges with scan noise and variations in attire and accessories. Nevertheless, TANet accurately establishes the shape correspondences for these sequence point clouds, including details such as dresses worn by the individuals. The visual results on SHREC'16 and Owlii further substantiate + +![](images/406c42f169da4b3cc39bd9291aaf62463dbc91a0bdef455648ca9d2fe8400c11.jpg) +Figure 11. Visualization of the correspondence results on real scanned Owlii dataset. The results contain two consecutive action sequences involving individuals wearing different attire. + +TANet's superior robustness and generalization. + +# 5. Conclusion + +In this paper, we propose an unsupervised template-assisted point cloud shape correspondence network, including a template generation module and a template assistance module. Specifically, the template generation module is introduced to learn a set of learnable templates that balance shape complexity and possess an explicitly structured shape. The template assistance module is designed to leverage the generated templates adaptively to achieve more accurate shape correspondences from template selection, feature learning, and transitive consistency perspectives. Extensive experiments on the SHREC'19 and TOSCA benchmarks demonstrate the superiority of TANet. Cross-dataset experiments on SURREAL and SMAL showcase our method's desirable generalization capabilities. + +# 6. Acknowledgements + +This work was partially supported by the National Defense Basic Scientific Research program (JCKY2022911B002) and the National Nature Science Foundation of China (NSFC 12150007, NSFC 62121002). + +# References + +[1] Fabio Arena, Mario Collotta, Giovanni Pau, and Francesco Terminate. An overview of augmented reality. Computers, 11 (2):28, 2022. 1 +[2] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Generalized multidimensional scaling: a framework for isometry-invariant partial surface matching. Proceedings of the National Academy of Sciences, 103(5):1168-1172, 2006. 1, 2 +[3] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 2, 5 +[4] Dongliang Cao and Florian Bernard. Self-supervised learning for multimodal non-rigid 3d shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17735-17744, 2023. 2 +[5] Etienne Corman, Maks Ovsjanikov, and Antonin Chambolle. Supervised descriptor learning for non-rigid shape matching. In European conference on computer vision, pages 283-298. Springer, 2014. 3 +[6] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, Y Sahillioglu, et al. Shrec'16: Partial matching of deformable shapes. Proc. 3DOR, 2(9): 12, 2016. 5, 8 +[7] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 1, 3, 5, 6, 7 +[8] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. Advances in Neural Information Processing Systems, 32, 2019. 1, 7 +[9] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8592-8601, 2020. 3, 7 +[10] Georgia Gkioxari, Nikhila Ravi, and Justin Johnson. Learning 3d object shape and layout without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1695-1704, 2022. 3 +[11] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 230-246, 2018. 1, 2, 5, 7 +[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 2 +[13] Jianfeng He, Jiacheng Deng, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Hierarchical shape-consistent transformer for unsupervised point cloud shape correspondence. IEEE Transactions on Image Processing, 2023. 1, 3, 5, 6, 7 + +[14] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, pages 1449-1457. Wiley Online Library, 2008. 1, 2 +[15] Dany Laksono and Trias Aditya. Utilizing a game engine for interactive 3d topographic data visualization. ISPRS International Journal of Geo-Information, 8(8):361, 2019. 1 +[16] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. Dpc: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 1, 3, 5, 6, 7 +[17] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In European Conference on Computer Vision, pages 19-34. Springer, 2022. 3 +[18] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In Proceedings of the IEEE international conference on computer vision, pages 5659-5667, 2017. 3 +[19] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. IEEE transactions on pattern analysis and machine intelligence, 36(1):171-180, 2013. 1 +[20] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 3 +[21] Ming Liu. Robotic online path planning on point cloud. IEEE transactions on cybernetics, 46(5):1217-1228, 2015. 1 +[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[23] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 3 +[24] Riccardo Marin, Marie-Julie Rakotosaona, Simone Melzi, and Maks Ovsjanikov. Correspondence learning via linearly-invariant embedding. Advances in Neural Information Processing Systems, 33:1608-1620, 2020. 1, 7 +[25] Simone Melzi, Riccardo Marin, Emanuele Rodola, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, page 3, 2019. 2, 5 +[26] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Learning 3d scene priors with 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792-802, 2023. 2, 3 +[27] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 1, 2 + +[28] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6 +[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 +[31] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1617-1627, 2019. 6, 7 +[32] Radu Bogdan Rusu, Zoltan Csaba Marton, Nico Blodow, Mihai Dolha, and Michael Beetz. Towards 3d point cloud based object maps for household environments. Robotics and Autonomous Systems, 56(11):927-941, 2008. 1 +[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3 +[34] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 344-362. Springer, 2022. 3 +[35] Art Tevs, Alexander Berner, Michael Wand, Ivo Ihrke, and H-P Seidel. Intrinsic shape matching by planned landmark sampling. In Computer graphics forum, pages 543-552. Wiley Online Library, 2011. 1, 2 +[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2 +[37] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 2, 3, 7 +[38] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 3 +[39] Ziyi Wang, Yongming Rao, Xumin Yu, Jie Zhou, and Jiwen Lu. Semaffinet: Semantic-affine transformation for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11819-11829, 2022. 3 +[40] Yi Xu, Yao Lu, and Ziyu Wen. Owli dynamic human mesh sequence dataset. In ISO/IEC JTC1/SC29/WG11 m41658, 120th MPEG Meeting, 2017. 5, 8 + +[41] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 2 +[42] Yiming Zeng, Yue Qian, Zhiyu Zhu, Junhui Hou, Hui Yuan, and Ying He. Corrnet3d: Unsupervised end-to-end learning of dense correspondence for 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6052-6061, 2021. 1, 3, 5, 6, 7 +[43] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 2 +[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2 +[45] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 1, 2, 5 \ No newline at end of file diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/images.zip b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ad3d4ef7175fc8dd8099c7b11789ad851bb202b8 --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b1a0ad76df17f7527496a3369a9d98f9f43b76e51fb0adcf2c5760062d1009a +size 536519 diff --git a/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/layout.json b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09492961a39466ca30325002309c8ac579db2425 --- /dev/null +++ b/2024/Unsupervised Template-assisted Point Cloud Shape Correspondence Network/layout.json @@ -0,0 +1,7829 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "Unsupervised Template-assisted Point Cloud Shape Correspondence Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "spans": [ + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "text", + "content": "Jiacheng Deng" + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "text", + "content": ", Jiahao Lu" + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "text", + "content": ", Tianzhu Zhang" + }, + { + "bbox": [ + 181, + 142, + 413, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 159, + 452, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 159, + 452, + 185 + ], + "spans": [ + { + "bbox": [ + 143, + 159, + 452, + 185 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2Deep Space Exploration Lab {dengjc, lujiahao} @mail.ustc.edu.cn, tzzhang@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "type": "text", + "content": "Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 524, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 287, + 632 + ], + "type": "text", + "content": "Point cloud shape correspondence is a challenging task that identifies and densely matches the source and target point clouds with deformable 3D shapes. The task has significant implications for various industries, including augmented reality [1], gaming [15], and robotics [21, 32]. However, the unrestricted mobility of humans and animals and their unusual postures have made direct correspondence between unconventional shapes a longstanding challenge in point cloud shape correspondence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "content": "The shape correspondence problem has been thoroughly investigated for 3D mesh data. However, such spectral-based methods [2, 14, 19, 27, 35] are mainly limited by the time-consuming pre-processing step and the reliance on mesh vertical connectivity information, frequently absent" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 212, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 309, + 212, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 309, + 212, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 309, + 212, + 545, + 397 + ], + "type": "image", + "image_path": "0772c8e58c58b0dd860e9f0a0afe8d05508e32648dd23247f674e7574bcaf5ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 402, + 547, + 469 + ], + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 469 + ], + "type": "text", + "content": "Figure 1. Visualization of template-assisted shape correspondence results. Correspondences are depicted by transferring colors from the source to the target based on matching results. The baseline incorrectly aligns the hands with the head and knees between unconventional shapes. In contrast, our method leverages the template to establish accurate correspondences for the hands." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "in point cloud scenarios. Point-based methods [7, 8, 11, 13, 16, 24, 42] are directly applied to the raw point cloud data, relying solely on point coordinates without connectivity information. To mitigate the resource-intensive labeling in fully supervised methods [8, 11, 24], the unsupervised point cloud shape correspondence task garners significant attention and yields competitive results [7, 13, 16, 42]. DPC [16] designs construction losses to reduce outlier matches. HSTR [13] incorporates whitening losses to mitigate the impact of shape discrepancies. SE-ORNet [7] introduces an orientation estimation module to alleviate symmetry issues. The shapes of humans and animals are highly diverse and complex. For instance, the SURREAL dataset [11] comprises " + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "inline_equation", + "content": "230\\mathrm{K}" + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": " human body shapes, while the SMAL dataset [45] contains " + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": " animal shapes, including numerous unconventional forms. Establishing accurate correspondences between two unconventional point cloud shapes remains a challenging problem in shape correspondence. Current methods [7, 13, 16, 42] rely on computing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "†Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5250" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "content": "the similarity between point cloud features to establish correspondences. However, in the case of unconventional point cloud pairs, the complete structure is distorted or destroyed due to limb adhesion, bending, and other postures. Without explicit structural guidance, directly calculating similarity is often insufficient to guarantee accurate correspondences. Therefore, it is essential to consider how to construct template shapes as intermediaries to assist in establishing correct correspondences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 543 + ], + "type": "text", + "content": "By analyzing prior shape correspondence methods, we have identified two pivotal issues that must be considered in shape correspondence. 1) How to generate suitable templates for point cloud pairs? Traditional point cloud template construction [11, 37, 41] mainly relies on manual selection or autoencoders [41]. While effective for shape-stable rigid objects in the same category, balancing template complexity becomes challenging for diverse non-rigid objects of different categories. Simple templates are clear but lack structural information, risking information loss during correspondence propagation. Complex templates offer richer information but are sensitive to noise, compromising accuracy. Despite attempts [26, 43] with learnable approaches, incomplete template structure remains a challenge. A more effective modeling approach is needed to generate templates suitable for intricate point cloud pairs. 2) How to effectively leverage templates for better correspondences? Establishing direct correspondences between unconventional shapes is challenging, but building correspondences between an unconventional shape and a suitable template is relatively straightforward. A correct mapping can be formed by finding corresponding template points for the source points and identifying corresponding points on the target point cloud based on the template. Direct similarity calculations between point clouds may introduce noise and yield ambiguous similarity weights. The template allows each point in the point cloud to compute a correlation vector, providing more stable information. Seeking consensus among these correlation vectors enhances appropriate point correspondences while mitigating erroneous ones." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "To achieve the above goal, we propose an Unsupervised Template-Assisted Point Cloud Shape Correspondence Network (TANet), including a template generation module and a template assistance module. In the template generation module, we incorporate a template bank with several learnable templates, effectively balancing template complexity in a data-driven manner. The space aligner enriches templates with comprehensive shape structures by establishing a mapping between the template feature space and the coordinate space. The space aligner is supervised by the coordinates and encoded features of the point cloud pairs. In the template assistance module, the adaptive selector chooses the most suitable template from the bank based on geometric and semantic attributes. The correlation fusion process" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "enhances point features with the template correlation vector via attention, suppressing noise and ambiguity. Moreover, we devise a transitive consistency loss to ensure coherence between the similarity computed through the template and directly computed between the source and target point clouds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 144, + 546, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 324 + ], + "type": "text", + "content": "To sum up, the contributions of this work can be summarized as follows: (i) We propose an unsupervised template-assisted point cloud shape correspondence network, achieved by jointly designing a template generation module and a template assistance module. (ii) We introduce a template generation module comprising a template bank with learnable templates and the space aligner for building explicit structures. The template assistance module adaptively selects a suitable template for robust point representations and maintains transitive consistency in shape correspondences. (iii) Extensive experiments on TOSCA [3] and SHREC'19 [25] benchmarks show the proposed TANet outperforming state-of-the-art methods. Cross-dataset experiments on SMAL [45] and SURREAL [11] demonstrate our method's desirable generalization capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 334, + 392, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 392, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 392, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 354, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 402 + ], + "type": "text", + "content": "In this Section, we give a brief overview of related works on point cloud shape correspondence, including deep learning for point clouds, shape correspondence, and point cloud template." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 605 + ], + "type": "text", + "content": "Deep Learning for Point Clouds. PointNet [29] employs a symmetrical function to aggregate all point features, resulting in a permutation-invariant global feature representation. Building upon this, PointNet++ [30] takes the concept further by incorporating local information, enhancing the representation capabilities and efficiency. DGCNN [38] introduces EdgeConv [38] blocks, dynamically updating neighborhood information based on dynamic graphs, ultimately delivering improved performance in point cloud analysis. Similarly, KPCnv [36] introduces point-wise convolution operators for point cloud feature learning. In the era of transformers, Point Cloud Transformer [12] stands out as a purely global Transformer network, pioneering self-attention layers instead of encoder layers within the PointNet framework. Point Transformer [44] operates within the local neighborhood of the target point cloud, enabling hierarchical extraction of geometric and semantic features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 713 + ], + "type": "text", + "content": "Shape Correspondence. The goal of shape correspondence is to accurately find the point-to-point correspondence between two point clouds. Traditional spectral-based methods [2, 14, 27, 35] utilize mesh data to project features onto Laplace-Beltrami Operator (LBO) eigenbasis and then learn a transformation between the eigendecomposition of source and target shapes, which will be translated to a dense shape correspondence. Cao et al. [4] transfer rich structure and connectivity information in mesh data" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5251" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 67, + 527, + 302 + ], + "blocks": [ + { + "bbox": [ + 70, + 67, + 527, + 302 + ], + "lines": [ + { + "bbox": [ + 70, + 67, + 527, + 302 + ], + "spans": [ + { + "bbox": [ + 70, + 67, + 527, + 302 + ], + "type": "image", + "image_path": "94d278b8f0b473e22f8c68b2901823f9660c726abd02bbcb6f5f0c695b1decf2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 547, + 338 + ], + "type": "text", + "content": "Figure 2. Illustration of the TANet. TANet comprises an encoder, a template generation module, and a template assistance module. The template generation module produces several learnable shape templates in the template bank with a space aligner. The template assistance module selects suitable templates for point cloud pairs and improve accuracy via correlation fusion and transitive consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 289, + 563 + ], + "type": "text", + "content": "to point cloud shape for robust matching. However, these methods are constrained by complex preprocessing and a reliance on vertex connectivity information. In contrast, point-based methods [5, 7, 9, 13, 16, 18, 34, 42] process point clouds directly without depending on connectivity information. CorrNet3D [42] and DPC [16] harness the established DGCNN [38] network to extract point representations through shape reconstruction. HSTR [13] has introduced a multi-receptive-field transformer point cloud encoder that considers local point cloud structures and long-range contextual information. SE-ORNet [7] incorporates an orientation estimation module to align the orientations of point cloud pairs, achieving precise matching results, particularly for symmetrical parts. These methods directly calculate point-wise similarities, making them susceptible to matching errors caused by noise and incomplete representations. In contrast, our approach constructs and utilizes templates to establish stable correspondences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Point Cloud Template. Point cloud templates, such as shape templates, scene priors, and semantic priors, are pivotal in various point cloud tasks, including 6D pose estimation [17, 37], reconstruction [10, 20], scene synthesis [26], and segmentation [23, 33, 39]. Tian et al. [37] develop an autoencoder trained on a collection of object models, which decodes the mean latent embedding to create shape templates. For similar rigid objects, generative methods can yield robust and stable templates. ScenePrior [26] encodes scene priors into a latent space and utilizes sampling to synthesize plausible 3D scenes. SemAffiNet [39] and Mask3D [33] integrate a learnable implicit semantic prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 348, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 469 + ], + "type": "text", + "content": "via transformer blocks with learnable queries. However, the complex shape and category variations pose a challenge for non-rigid objects to obtain well-balanced explicit templates through generative methods. The learnable implicit modeling struggles to achieve explicit stuctures. Based on the above discussion, we propose TANet, which integrates the template generation and assistance modules within a unified framework. Through technical designs, the learnable templates not only balance shape complexity but also possess an explicitly structured shape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 499, + 373, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 373, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 373, + 510 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 662 + ], + "type": "text", + "content": "The unsupervised point cloud correspondence task strives to establish accurate one-to-one correspondences between source and target point clouds without annotations. As illustrated in Figure 2, TANet primarily comprises three key modules: a general point cloud encoder, a Template Generation Module, and a Template Assistance Module. The choice of point cloud encoder follows the previous study [13]. The detailed design and optimization process for the Template Generation Module will be expounded upon in Section 3.2. Moreover, Section 3.3 will provide further insights into the Template Assistance Module. Finally, the training and inference process are discussed in Section 3.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 670, + 468, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 468, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 468, + 684 + ], + "type": "text", + "content": "3.2. Template Generation Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "The template generation module consists primarily of two components: a template bank and a space aligner." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5252" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "Template bank. In the template bank, we integrate " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " learnable explicit shape templates, each comprising learnable positions " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "P_T^i \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and learnable embeddings " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_T^i \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". During the training phase, the positions and embeddings of templates are updated concurrently with the network. In the inference phase, the positions and embeddings of template shapes remain unchanged, eliminating the need for additional updates based on input point clouds. The learnable embeddings " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_T^i" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " parameters are initialized using Gaussian random initialization. To expedite template training, we utilize random shapes in datasets to initialize the learnable positions " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "P_T^i" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 276, + 297 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 276, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 276, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 276, + 297 + ], + "type": "image", + "image_path": "6850f72c935869667dbeeb3e14b4448529cf987b3fa7fca30c613c1629cae051.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 287, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 322 + ], + "type": "text", + "content": "Figure 3. Space aligner structure. The predicted point positions in the output are constrained to match the true point positions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 288, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 383 + ], + "type": "text", + "content": "Space aligner. Space aligner " + }, + { + "bbox": [ + 46, + 323, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 323, + 288, + 383 + ], + "type": "text", + "content": " is crafted to predict point positions through point embeddings, aiding in template learning. The specific structure is illustrated in Figure 3. The predicted point positions undergo adjustment to align with true positions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 389, + 231, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 389, + 231, + 403 + ], + "spans": [ + { + "bbox": [ + 102, + 389, + 231, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {S m o o t h} \\mathrm {L} _ {1} (\\mathrm {P}, \\phi (\\mathrm {F})),", + "image_path": "0b16c525f574c01188c113343a131f2b63c4f91dd03af29e97e52a42dec40960.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "content": " are true positions, and " + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "content": " are point embeddings. The space aligner is trained with the encoded feature and the coordinates from point cloud pairs. During the inference phase, the space aligner is not involved in the computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": "Additionally, the training of templates also includes the template construction loss " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": ", which enforces consistency between the template shape and the cross-constructed template shape by point cloud pairs. Specifically, the cross-construction process is computed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 521, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 521, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 96, + 521, + 287, + 552 + ], + "type": "interline_equation", + "content": "\\hat {t} _ {x _ {i}} = \\sum_ {j \\in \\mathcal {N} _ {T} (x _ {i})} \\frac {e ^ {s _ {i j}}}{\\sum_ {l \\in \\mathcal {N} _ {T} (x _ {i})} e ^ {s _ {i l}}} t _ {j}, \\tag {1}", + "image_path": "6e0658b46db4f74009e92fa9611fd87cff9a9e6cce1ad452fa5e4d4168a110e2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "x_{i} \\in X" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "t_{j} \\in P_{T}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "s_{ij} \\in S_{XT}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "S_{XT}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " is the similarity matrix between " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "F_{X}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_T(x_i)" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " represents the latent " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " in the target " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ". The cross-construction of " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "P_{T}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " by the source point cloud " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " is denoted " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\hat{T}_{X} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\hat{T}_{X}^{i} = \\hat{t}_{x_{i}}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": ". After the introduction of the cross-construction process, the template construction loss " + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": " can be computed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 645, + 287, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 645, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 89, + 645, + 287, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t c} = \\mathbf {C D} (P _ {T}, \\hat {T} _ {X}) + \\mathbf {C D} (P _ {T}, \\hat {T} _ {Y}), \\tag {2}", + "image_path": "9f99e13631988a12242766f7970694a4ceb033ff9faa1a3b8b2d919f7d7390e7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where CD denotes chamfer distance. It is essential to highlight that the point encoder remains unaltered during the backpropagation of " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {align }}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {tc }}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " gradients, preventing interference with the feature extraction of point cloud pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 463, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 463, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 463, + 84 + ], + "type": "text", + "content": "3.3. Template Assistance Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "type": "text", + "content": "The template assistance module aims to enhance the correspondence accuracy in point cloud pairs through templates. It comprises three primary components: Template Selector, Correlation Fusion, and Transitive Consistency." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 151, + 535, + 217 + ], + "blocks": [ + { + "bbox": [ + 314, + 151, + 535, + 217 + ], + "lines": [ + { + "bbox": [ + 314, + 151, + 535, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 151, + 535, + 217 + ], + "type": "image", + "image_path": "a36ecdcbaace7a50176579398b88fc58035bc5865d574c19b32e43347ab92ea3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "type": "text", + "content": "Figure 4. Template selector structure. The best suitable template is determined through geometric and semantic measures of similarity between templates and point cloud pairs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": "Template selector. The input to the Template Selector comprises point cloud coordinates " + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "inline_equation", + "content": "X \\& Y \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "inline_equation", + "content": "F_{X} \\& F_{Y} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": ", along with template point cloud coordinates " + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "inline_equation", + "content": "\\{P_T^i | i = 1,2,\\dots,K\\}" + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": " and embeddings " + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "inline_equation", + "content": "\\{F_T^i | i = 1,2,\\dots,K\\}" + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": " from the template bank. Geometric and semantic similarities between the point cloud pair and templates guide the adaptive template selection. As illustrated in Figure 4, it involves calculating the chamfer distance as the geometric similarity between the point cloud pair and template point cloud. To get semantic similarity, the point cloud features and template embeddings undergo mix pooling and cosine similarity computation. The fusion of geometric and semantic similarities is employed in Gumbel-softmax to ascertain the most fitting template " + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 255, + 546, + 432 + ], + "type": "text", + "content": " for the given point cloud pair. The specific computation formula is as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 438, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 438, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 310, + 438, + 545, + 467 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T = \\mathbf {G S} \\left\\{\\left(\\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {X}\\right) + \\mathbf {M P} \\left(F _ {T} ^ {i}\\right) \\cdot \\mathbf {M P} \\left(F _ {Y}\\right)\\right) \\right. \\tag {3} \\\\ + \\left(\\mathbf {C D} \\left(P _ {T} ^ {i}, X\\right) + \\mathbf {C D} \\left(P _ {T} ^ {i}, Y\\right)\\right) | i = 1, 2, \\dots , K \\}, \\\\ \\end{array}", + "image_path": "94cef48c633bb716665c825e7137df4b661bde1904d0d514921e35abe5e6bb55.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 474, + 544, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 544, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 544, + 497 + ], + "type": "text", + "content": "where GS represents Gumbel-Softmax, MP denotes Mix Pooing, and CD stands for chamfer distance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "text", + "content": "Correlation fusion. As depicted in Figure 5, template embeddings " + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "text", + "content": " and point cloud features " + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "F_{X} / F_{Y}" + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "text", + "content": " are input into the correlation fusion process to enhance point cloud features. Initially, a correlation between the point cloud and the template is established. Subsequently, an attention mechanism incorporates the correlation vectors into point cloud features, yielding " + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "F_{X}' / F_{Y}' \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 305, + 498, + 545, + 653 + ], + "type": "text", + "content": ". The similarity in the attention mechanism is computed based on point cloud features to underscore the spatial distribution of shapes, enhancing perception to spatial relationships. The introduction of template correlation aids in refining point cloud features and mitigating ambiguous representations in certain noisy point clouds." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": "Transitive consistency. Utilizing the fused source features " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{X}^{\\prime}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": " and template embeddings " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": ", we compute the similarity matrix " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{XT}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{TY}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": " is obtained by calculating the similarity between " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{Y}^{\\prime}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": ". The multiplication of " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{XT}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{TY}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": " matrices yields the similarity matrix " + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{X TY}" + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5253" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 265, + 183 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 265, + 183 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 265, + 183 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 265, + 183 + ], + "type": "image", + "image_path": "ba78a9763622f2cefb092af425fcf3c9be703dda78d6f723475d6dc063c8b69b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 218 + ], + "type": "text", + "content": "Figure 5. The correlation fusion process. Features of point cloud pairs and embeddings of templates compute correlation vectors, which are then fused using an attention mechanism." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 288, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 291 + ], + "type": "text", + "content": "reflecting the similarity between the source and target point clouds transmitted through the template. As shown in Figure 6, (a) Direct similarity illustrates the process of directly calculating the similarity between the source point cloud and the target point cloud:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 287, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 287, + 317 + ], + "type": "interline_equation", + "content": "S _ {X Y} = F _ {X} ^ {\\prime} F _ {Y} ^ {\\prime} ^ {\\top}. \\tag {4}", + "image_path": "6c952f9a94b474bef4e5b5a5eef0a4ac02bbb13bab423fc389d58348d80c3e73.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": "In contrast, (b) Template-assisted similarity demonstrates the computation of " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "S_{X TY}" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": ". The multiplication of matrices " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "S_{XT}" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "S_{TY}" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": " essentially results in an indirect computation of the similarity relationship between the source and target point clouds. This process involves a weighted sum of the similarity calculated between the template point cloud and the target point cloud, providing a more stable and accurate similarity result for source and target points. The Template-assisted similarity method helps mitigate the impact of outliers and noise on the similarity calculation, leading to more reliable results. In order to guide the establishment of more stable and accurate similarity relationships between point clouds without compromising the inference speed, we introduce the Transitive Consistency loss " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{trans})" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": ", promoting the consistency between " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "S_{X TY}" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": " and the similarity matrix " + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "inline_equation", + "content": "S_{XY}" + }, + { + "bbox": [ + 46, + 327, + 288, + 529 + ], + "type": "text", + "content": ". The Transitive Consistency loss is calculated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 533, + 287, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 533, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 287, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s}} = \\mathbf {C E} \\left(S _ {X T Y}, S _ {X Y}\\right), \\tag {5}", + "image_path": "da59adb853449494207886956c5e7ace16ff63e77aeb57bb9c2689e2f03b0b57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 203, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 203, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 203, + 567 + ], + "type": "text", + "content": "where CE means Cross-Entropy loss." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 73, + 580, + 261, + 662 + ], + "blocks": [ + { + "bbox": [ + 73, + 580, + 261, + 662 + ], + "lines": [ + { + "bbox": [ + 73, + 580, + 261, + 662 + ], + "spans": [ + { + "bbox": [ + 73, + 580, + 261, + 662 + ], + "type": "image", + "image_path": "eeb0efd0caf43b60588bdbda8d510ddfb437bb5546e3ae5b85d17aeccd60a094.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 663, + 286, + 707 + ], + "lines": [ + { + "bbox": [ + 46, + 663, + 286, + 707 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 286, + 707 + ], + "type": "text", + "content": "Figure 6. Transitive Consistency. (a) the direct similarity computation between the source and target point clouds. (b) a stable template is employed to transmit the similarity relation. " + }, + { + "bbox": [ + 46, + 663, + 286, + 707 + ], + "type": "inline_equation", + "content": "L_{trans}" + }, + { + "bbox": [ + 46, + 663, + 286, + 707 + ], + "type": "text", + "content": " ensures the consistency between these two similarity results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 462, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 462, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 462, + 85 + ], + "type": "text", + "content": "3.4. Model Training & Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "content": "Following the previous work [7, 13, 16], we also incorporate the construction loss " + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\text{const}})" + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "content": ", encompassing both cross-construction and self-construction between the source shape " + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "content": " and the target shape " + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "content": ". The construction loss promotes feature smoothness and unique correspondences. The computation process and formula for cross-construction and self-construction are similar to Equation 1. Specific formulas are detailed in the supplementary material. To sum up, the total loss of TANet is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 208, + 545, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 208, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 208, + 545, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {t r a n s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {3} \\mathcal {L} _ {\\text {t c}} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t r}}, \\tag {6}", + "image_path": "7b7938e27053679b807cbac9094de12b2b1fdf6ae7369c54509f418cb7352606.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "text", + "content": " are hyperparameters, balancing the contribution of different loss terms. The training of both the template and the network is jointly achieved, enabling end-to-end training without requiring additional prolonged iterations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "text", + "content": "During inference, we set the closest point " + }, + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "inline_equation", + "content": "y_{j^*}" + }, + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "text", + "content": " in the feature space for each point " + }, + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 281, + 545, + 315 + ], + "type": "text", + "content": " as its corresponding point. This selection rule can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 326, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 326, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 358, + 326, + 545, + 346 + ], + "type": "interline_equation", + "content": "f \\left(x _ {i}\\right) = y _ {j ^ {*}}, j ^ {*} = \\underset {j} {\\operatorname {a r g m a x}} \\left(s _ {i j}\\right). \\tag {7}", + "image_path": "f375bc4aaafe95397c32c76040e74108f058ec4b3fc91810208fd3e7b4fc478b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 346, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 388, + 358 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 365, + 422, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 422, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 422, + 378 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 383, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 598 + ], + "type": "text", + "content": "Datasets. We evaluate our method on the most popular benchmarks, including the TOSCA [3] dataset and the SHREC'19 [25] dataset. TOSCA comprises 41 different shapes of various animal species. Following previous methods [13, 16], we pair these shapes to create both the training and testing sets. SHREC'19 consists of 44 real human models paired into 430 shape correspondence samples. SURREAL [11] and SMAL [45] are two large-scale datasets that we employ to evaluate the generalization capabilities of our method. The SURREAL dataset comprises 230,000 training shapes, from which we randomly sample and pair 2,000 shape pairs for training. The SMAL dataset includes parameterized animal models for generating shape pairs to train shape correspondence models. Furthermore, robustness experiments are conducted on the real-scanned Owlii dataset [40] and the partial shape dataset SHREC'16 [6]. Besides, we maintain a consistent point count of " + }, + { + "bbox": [ + 304, + 383, + 545, + 598 + ], + "type": "inline_equation", + "content": "n = 1024" + }, + { + "bbox": [ + 304, + 383, + 545, + 598 + ], + "type": "text", + "content": " in accordance with prior studies [13, 16, 42]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 599, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 545, + 659 + ], + "type": "text", + "content": "Evaluation Metrics. The primary evaluation metrics for the point cloud shape correspondence task include two key measures: average correspondence error and correspondence accuracy. The average correspondence error measures a source and target shape pair " + }, + { + "bbox": [ + 304, + 599, + 545, + 659 + ], + "type": "inline_equation", + "content": "(X,Y)" + }, + { + "bbox": [ + 304, + 599, + 545, + 659 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 667, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 667, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 362, + 667, + 545, + 696 + ], + "type": "interline_equation", + "content": "e r r = \\frac {1}{n} \\sum_ {x _ {i} \\in X} | f (x _ {i}) - y _ {g t} | _ {2}, \\tag {8}", + "image_path": "a9c6120bcc40eff1b89edc989845a92472cf053594ada330f226112a600b839c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "y_{gt} \\in Y" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " represents the ground-truth matching point" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5254" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 279, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 279, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 279, + 84 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 47, + 72, + 279, + 84 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 72, + 279, + 84 + ], + "type": "text", + "content": ". Correspondence accuracy is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 91, + 287, + 120 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 91, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 78, + 91, + 287, + 120 + ], + "type": "interline_equation", + "content": "a c c (\\epsilon) = \\frac {1}{n} \\sum_ {x _ {i} \\in X} \\mathbb {1} \\left(\\left| f \\left(x _ {i}\\right) - y _ {g t} \\right| _ {2} < \\epsilon d\\right), \\tag {9}", + "image_path": "6b1a818f2f40b8571590e0922187e88708dceb1070afe8c2fb51f646a6c9f960.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "content": " denotes the indicator function, " + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "content": " is the maximum Euclidean distance between points in " + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "inline_equation", + "content": "\\epsilon \\in [0,1]" + }, + { + "bbox": [ + 47, + 127, + 287, + 162 + ], + "type": "text", + "content": " signifies the error tolerance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": "Implementation details. The number of learnable templates " + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": " in the Template Bank is set to 4, with a template embedding dimension of 512. The Space Aligner incorporates two attention blocks, each with a dimensionality of 512. We employ a transformer encoder, akin to existing method [13], featuring four layers of attention blocks as our point cloud encoder. For the Correlation Fusion process, the dimensions of self-attention and MLP are set to 512. In Equation 6, " + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 46, + 163, + 288, + 331 + ], + "type": "text", + "content": " are set to 0.5, 0.5, 1, and 1. All experiments with our method on various datasets are conducted on a single GeForce RTX 3090 device using the PyTorch 1.10.1 framework [28]. Model training utilizes the AdamW [22] optimizer with a learning rate of 5e-4, a weight decay of 5e-4, and a batch size of 4." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 398, + 276, + 492 + ], + "blocks": [ + { + "bbox": [ + 46, + 339, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 396 + ], + "type": "text", + "content": "Table 1. Comparison on TOSCA and SHREC'19 benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. Higher accuracy and lower error reflect a better result, with the best and second-best outcomes highlighted in bold and underlined, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 398, + 276, + 492 + ], + "lines": [ + { + "bbox": [ + 57, + 398, + 276, + 492 + ], + "spans": [ + { + "bbox": [ + 57, + 398, + 276, + 492 + ], + "type": "table", + "html": "
MethodInputTOSCASHREC'19
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh//5.9%0.2
CorrNet3D[42]Point0.3%32.70.4%33.8
DPC[16]Point34.7%2.815.3%5.6
SE-ORNet[7]Point38.3%2.717.5%5.1
HSTR[13]Point52.3%1.219.3%4.9
OursPoint65.1%0.721.5%4.5
", + "image_path": "3342c0126e771bae1809e56e6f2ddc354ca9f4434dd7cb3a8302760eb0e198f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 270, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 270, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 270, + 516 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 593 + ], + "type": "text", + "content": "We conducted comprehensive experimental comparisons with existing methods on the TOSCA animal dataset and the SHREC'19 human dataset. Furthermore, to ensure a fair comparison with existing methods, our approach is implemented without the use of post-processing or additional connectivity information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": "Evaluation on TOSCA dataset. As shown in Table 1, our approach demonstrates significant improvement on the TOSCA dataset, achieving a new state-of-the-art performance with a " + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": " increase in accuracy and a minimal average error of " + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "inline_equation", + "content": "0.7\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": ". For a comprehensive depiction of accuracy across different error tolerances, we present correspondence accuracy in Figure 7(a). Similarly, our method outperforms others at different error tolerances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on SHREC'19 dataset. As indicated in Table 1, our method demonstrates significant improvement on" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 426, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 426, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 426, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 426, + 148 + ], + "type": "image", + "image_path": "973e5a12c8854770ddef2485edbbd74a74c4ba1f3ddb597d653e2880535b005e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 150, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 305, + 150, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 545, + 193 + ], + "type": "text", + "content": "Figure 7. Correspondence accuracy at various error tolerances. (a) Corresponding accuracy of different methods on the TOSCA dataset under various error tolerances. (b) Experiments similar to (a) conducted on the SHREC'19 dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 547, + 148 + ], + "type": "image", + "image_path": "14bc645f377329aa3c2e07927c9ffcbe223ece013ae02a6a8e0f219c5b8b7e6c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": "the SHREC'19 dataset, achieving an impressive " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " increase in accuracy and a reduction of " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " in error, consequently establishing a new state-of-the-art performance. Figure 7(b) illustrates the correspondence accuracy of various competitive methods at different error tolerances, showcasing our method's clear superiority across different levels of error tolerance." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 360, + 526, + 468 + ], + "blocks": [ + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "type": "text", + "content": "Table 2. Evaluation of the model with different designs on TOSCA dataset. TGM denotes the template generation module, " + }, + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "type": "inline_equation", + "content": "L_{tc}" + }, + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "type": "text", + "content": " stands for template construction loss, TS represents template selector, CF denotes correlation fusion process, and " + }, + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "type": "inline_equation", + "content": "L_{trans}" + }, + { + "bbox": [ + 305, + 302, + 546, + 357 + ], + "type": "text", + "content": " is the transitive consistency loss." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 360, + 526, + 468 + ], + "lines": [ + { + "bbox": [ + 321, + 360, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 321, + 360, + 526, + 468 + ], + "type": "table", + "html": "
TGMLtcTSCFLtransTOSCA
acc ↑err ↓
[A]XXXXX52.3%1.2
[B]✓XX✓X54.6%1.2
[C]✓XXX✓55.8%1.1
[D]✓✓X✓X59.3%1.0
[E]✓✓XX✓58.8%1.0
[F]✓✓✓✓X61.9%0.9
[G]✓✓✓X✓61.3%0.9
[H]✓✓✓✓✓65.1%0.7
", + "image_path": "0e7ccaa5516b4efa80258a9d0b978f29f1285e6246e2e29afd0faa6e24ae368b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Evaluation of the model with different designs. In Table 2, we conduct extensive ablation studies on the TOSCA dataset to validate the effectiveness of our model designs. Specifically, [A] presents a baseline method without any template-related design. [B] and [C] demonstrate the employment of the template generation module, assisted respectively through correlation fusion and transitive consistency loss, resulting in accuracy improvements of " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". The limited improvement is attributed to the difficulty in structural learning of templates without appropriate loss constraints. [D] and [E] illustrate the introduction of template construction loss to aid in the learning and optimization of template structures, leading to advancements in accuracy and error compared to settings [B] and [C]. Building upon this, [F] and [G] further employ a template selector to identify the most suitable template for aiding the establishment of correspondence for specific shapes. The further improvements in accuracy and error indicate that adaptive" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5255" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "template selection significantly contributes to accurate correspondence establishment. Finally, in [H], we present the performance of the complete model, highlighting the essential roles played by different modules in collectively achieving accurate point cloud shape correspondences." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 143, + 278, + 247 + ], + "blocks": [ + { + "bbox": [ + 55, + 143, + 278, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 278, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 278, + 247 + ], + "type": "image", + "image_path": "ce5a77a6d3d58d13dcafe8c3a9abde9b8cf437b05f656718547315dd4e84a1bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 283 + ], + "type": "text", + "content": "Figure 8. Visual comparison of template generation methods. The templates generated by the template generation module exhibit clearer structures in both animal and human templates." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 483 + ], + "type": "text", + "content": "Effectiveness of the template generation module. Figure 8 compares templates generated by our template generation module and the common autoencoder-based approach [37]. Our designed learnable templates, constrained by the space aligner and construction loss, provide more structural templates to aid point cloud shape correspondence. However, due to the significant shape variations of non-rigid bodies, it remains challenging to generate perfect templates without direct shape supervision. The existing templates are merely a sub-optimal form. Additionally, we carefully consider the number of templates in Table 3. A single or dual template falls short of capturing the shape variations, resulting in a significant decrease in accuracy. While eight templates can slightly enhance accuracy, they lead to a substantial increase in model parameters. Hence, we offer a trade-off to set up four learnable templates." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 518, + 160, + 578 + ], + "blocks": [ + { + "bbox": [ + 47, + 493, + 282, + 515 + ], + "lines": [ + { + "bbox": [ + 47, + 493, + 282, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 282, + 515 + ], + "type": "text", + "content": "Table 3. Ablation studies on Table 4. Comparison of simi-the number of templates. larity computation methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 518, + 160, + 578 + ], + "lines": [ + { + "bbox": [ + 49, + 518, + 160, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 518, + 160, + 578 + ], + "type": "table", + "html": "
#TemplateParam.acc ↑
17.3M56.8%
27.8M61.4%
48.8M65.1%
810.9M66.2%
", + "image_path": "06d9177fd4e876ff33eaf1c25b336413c3c914a7649462a301e38b63ce11c160.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 164, + 529, + 280, + 571 + ], + "blocks": [ + { + "bbox": [ + 164, + 529, + 280, + 571 + ], + "lines": [ + { + "bbox": [ + 164, + 529, + 280, + 571 + ], + "spans": [ + { + "bbox": [ + 164, + 529, + 280, + 571 + ], + "type": "table", + "html": "
SettingsFLOPs(G)TOSCA
acc ↑err ↓
Direct1.0765.1%0.73
Transitive4.2966.3%0.67
", + "image_path": "490daf094c2b3a62852fd534d529d93aa2f210d310999882689aa476a6d7374e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Effectiveness of the template assistance module. The effectiveness of correlation fusion has been validated in Table 2. Here, we further delve into the discussion regarding the computation of similarities between point cloud pairs. The purpose of transitive consistency loss is to constrain the direct similarity computation between the source and target point clouds to align with the similarity computed through template transitivity. As demonstrated in Table 4, similarity computation through template propagation has been proven to be a more stable and accurate approach. However, considering the increased computational complexity associated" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "with the large number of point clouds, we adopt the direct similarity computation method during inference to enhance model efficiency. The transitive consistency loss only constrains the learning of point cloud representations through similarity-guided learning during training." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 207, + 535, + 338 + ], + "blocks": [ + { + "bbox": [ + 304, + 148, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 204 + ], + "type": "text", + "content": "Table 5. Cross-dataset generalization evaluation on SMAL and SURREAL benchmarks. Acc signifies correspondence accuracy at 0.01 error tolerance, and err denotes average correspondence error. The best and second-best outcomes highlighted in bold and underlined, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 207, + 535, + 338 + ], + "lines": [ + { + "bbox": [ + 315, + 207, + 535, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 207, + 535, + 338 + ], + "type": "table", + "html": "
MethodInputSMALSURREAL
acc ↑err ↓acc ↑err ↓
SURFMNet[31]Mesh5.9%0.24.3%0.3
GeoFMNet[9]Mesh//8.2%0.2
Diff-FMaps[24]Point//4.0%7.1
3D-CODED[11]Point0.5%19.22.1%8.1
Elementary[8]Point0.5%13.72.3%7.6
CorrNet3D[42]Point5.3%9.86.0%6.9
DPC[16]Point33.2%5.817.7%6.1
SE-ORNet[7]Point36.4%3.921.5%4.6
HSTR[13]Point33.9%5.619.4%5.6
OursPoint37.1%3.720.6%4.8
", + "image_path": "06942cb53460cd574e3f0ba9c2d08731448d6e9eb48414587a610717eb41260f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 365, + 527, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 527, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 527, + 377 + ], + "type": "text", + "content": "4.4. Cross-dataset Generalization Performance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 591 + ], + "type": "text", + "content": "The cross-dataset generalization ability is a crucial evaluation metric for the point cloud shape correspondence task. Following the conventional setup [7, 13, 16], generalization evaluation involves training on one dataset and testing on another. Specifically, training on the SMAL dataset and testing on the TOSCA dataset, as well as training on the SURREAL dataset and testing on the SHREC'19 dataset. As shown in Table 5, our method surpasses existing state-of-the-art approaches, establishing a new state-of-the-art performance by overcoming domain differences on the SMAL dataset. Regarding generalization to the SURREAL dataset, our results are also competitive. Confronting challenges posed by human body symmetry and directional rotations in the SURREAL and SHREC'19 datasets, our approach slightly lags behind the SE-ORNet [7], specifically designed for this issue. However, TANet significantly outperforms other point cloud shape correspondence methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 608, + 419, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 419, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 419, + 620 + ], + "type": "text", + "content": "4.5. Visual Comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "In Figure 9, we present visual comparisons of our method with competitive approaches. When the source point cloud forms an unconventional shape, such as a distorted standing posture, other methods struggle to distinguish between the hand and facial structures. Our method more accurately establishes the correspondence for hands and heads by leveraging learned templates as an intermediary." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5256" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 46, + 519, + 213 + ], + "blocks": [ + { + "bbox": [ + 78, + 46, + 519, + 213 + ], + "lines": [ + { + "bbox": [ + 78, + 46, + 519, + 213 + ], + "spans": [ + { + "bbox": [ + 78, + 46, + 519, + 213 + ], + "type": "image", + "image_path": "ef4d4d5c61d29f5ec77470168d79de831b9f7069df0450dd0d794389b3c595c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "type": "text", + "content": "Figure 9. Visual comparison on TOSCA and SHREC'19. We visually compared the correspondence results of different methods on unconventional shape pairs, providing intuitive validation of the proposed approach's superiority in establishing correspondences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 256, + 279, + 373 + ], + "blocks": [ + { + "bbox": [ + 53, + 256, + 279, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 256, + 279, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 279, + 373 + ], + "type": "image", + "image_path": "8bd9378d45f79188f11a607e1576ae864e841ccd56c9d9102108903fe4983cca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "text", + "content": "Figure 10. Visualization of the correspondence results on partial-shape SHREC'16 dataset. The partial shapes contain four species, including horses, cats, dogs, and humans. The generation of partial shapes in the SHREC'16 dataset primarily involves two methods: cutting and holing to wipe off body parts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "type": "text", + "content": "4.6. Robustness Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": "To further validate the robustness and generalization capabilities of our method, we directly evaluate our trained model on the SHREC'16 [6] and Owlii [40] datasets, presenting visual results. As shown in Figure 10, the SHREC'16 dataset comprises point cloud pairs of partial shapes, simulating challenges encountered in real-world applications such as occlusion, missing parts, and noise. Although only the upper body of the cat is present, and the dog is represented by a sparse point cloud with minimal parts of its body, our method discriminatively models local structures and accurately establishes shape correspondences. Moreover, the correspondence results of the partial human shape are predictions of the model trained on the animal dataset, providing further validation of the generalization capability. Illustrated in Figure 11, the Owlii dataset includes real-scanned human motion sequences, introducing challenges with scan noise and variations in attire and accessories. Nevertheless, TANet accurately establishes the shape correspondences for these sequence point clouds, including details such as dresses worn by the individuals. The visual results on SHREC'16 and Owlii further substantiate" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 353, + 256, + 498, + 376 + ], + "blocks": [ + { + "bbox": [ + 353, + 256, + 498, + 376 + ], + "lines": [ + { + "bbox": [ + 353, + 256, + 498, + 376 + ], + "spans": [ + { + "bbox": [ + 353, + 256, + 498, + 376 + ], + "type": "image", + "image_path": "406c42f169da4b3cc39bd9291aaf62463dbc91a0bdef455648ca9d2fe8400c11.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 379, + 546, + 413 + ], + "lines": [ + { + "bbox": [ + 305, + 379, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 546, + 413 + ], + "type": "text", + "content": "Figure 11. Visualization of the correspondence results on real scanned Owlii dataset. The results contain two consecutive action sequences involving individuals wearing different attire." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 420, + 500, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 500, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 500, + 432 + ], + "type": "text", + "content": "TANet's superior robustness and generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": "In this paper, we propose an unsupervised template-assisted point cloud shape correspondence network, including a template generation module and a template assistance module. Specifically, the template generation module is introduced to learn a set of learnable templates that balance shape complexity and possess an explicitly structured shape. The template assistance module is designed to leverage the generated templates adaptively to achieve more accurate shape correspondences from template selection, feature learning, and transitive consistency perspectives. Extensive experiments on the SHREC'19 and TOSCA benchmarks demonstrate the superiority of TANet. Cross-dataset experiments on SURREAL and SMAL showcase our method's desirable generalization capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 645, + 421, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 421, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 421, + 658 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "content": "This work was partially supported by the National Defense Basic Scientific Research program (JCKY2022911B002) and the National Nature Science Foundation of China (NSFC 12150007, NSFC 62121002)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5257" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Fabio Arena, Mario Collotta, Giovanni Pau, and Francesco Terminate. An overview of augmented reality. Computers, 11 (2):28, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 178 + ], + "type": "text", + "content": "[2] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Generalized multidimensional scaling: a framework for isometry-invariant partial surface matching. Proceedings of the National Academy of Sciences, 103(5):1168-1172, 2006. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "text", + "content": "[3] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "text", + "content": "[4] Dongliang Cao and Florian Bernard. Self-supervised learning for multimodal non-rigid 3d shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17735-17744, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 302 + ], + "type": "text", + "content": "[5] Etienne Corman, Maks Ovsjanikov, and Antonin Chambolle. Supervised descriptor learning for non-rigid shape matching. In European conference on computer vision, pages 283-298. Springer, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 304, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 286, + 345 + ], + "type": "text", + "content": "[6] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, Y Sahillioglu, et al. Shrec'16: Partial matching of deformable shapes. Proc. 3DOR, 2(9): 12, 2016. 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 413 + ], + "type": "text", + "content": "[7] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "type": "text", + "content": "[8] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. Advances in Neural Information Processing Systems, 32, 2019. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "text", + "content": "[9] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8592-8601, 2020. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Georgia Gkioxari, Nikhila Ravi, and Justin Johnson. Learning 3d object shape and layout without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1695-1704, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[11] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 230-246, 2018. 1, 2, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[12] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Jianfeng He, Jiacheng Deng, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Hierarchical shape-consistent transformer for unsupervised point cloud shape correspondence. IEEE Transactions on Image Processing, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, pages 1449-1457. Wiley Online Library, 2008. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 151 + ], + "type": "text", + "content": "[15] Dany Laksono and Trias Aditya. Utilizing a game engine for interactive 3d topographic data visualization. ISPRS International Journal of Geo-Information, 8(8):361, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[16] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. Dpc: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "text", + "content": "[17] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In European Conference on Computer Vision, pages 19-34. Springer, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In Proceedings of the IEEE international conference on computer vision, pages 5659-5667, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[19] Roee Litman and Alexander M Bronstein. Learning spectral descriptors for deformable shape correspondence. IEEE transactions on pattern analysis and machine intelligence, 36(1):171-180, 2013. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "type": "text", + "content": "[20] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 441 + ], + "type": "text", + "content": "[21] Ming Liu. Robotic online path planning on point cloud. IEEE transactions on cybernetics, 46(5):1217-1228, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 445, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 466 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "type": "text", + "content": "[23] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[24] Riccardo Marin, Marie-Julie Rakotosaona, Simone Melzi, and Maks Ovsjanikov. Correspondence learning via linearly-invariant embedding. Advances in Neural Information Processing Systems, 33:1608-1620, 2020. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "text", + "content": "[25] Simone Melzi, Riccardo Marin, Emanuele Rodola, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, page 3, 2019. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Learning 3d scene priors with 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792-802, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5258" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[30] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[31] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1617-1627, 2019. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 316 + ], + "type": "text", + "content": "[32] Radu Bogdan Rusu, Zoltan Csaba Marton, Nico Blodow, Mihai Dolha, and Michael Beetz. Towards 3d point cloud based object maps for household environments. Robotics and Autonomous Systems, 56(11):927-941, 2008. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[33] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[34] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 344-362. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[35] Art Tevs, Alexander Berner, Michael Wand, Ivo Ihrke, and H-P Seidel. Intrinsic shape matching by planned landmark sampling. In Computer graphics forum, pages 543-552. Wiley Online Library, 2011. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "text", + "content": "[37] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "text", + "content": "[38] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[39] Ziyi Wang, Yongming Rao, Xumin Yu, Jie Zhou, and Jiwen Lu. Semaffinet: Semantic-affine transformation for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11819-11829, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[40] Yi Xu, Yao Lu, and Ziyu Wen. Owli dynamic human mesh sequence dataset. In ISO/IEC JTC1/SC29/WG11 m41658, 120th MPEG Meeting, 2017. 5, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 328 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[41] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[42] Yiming Zeng, Yue Qian, Zhiyu Zhu, Junhui Hou, Hui Yuan, and Ying He. Corrnet3d: Unsupervised end-to-end learning of dense correspondence for 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6052-6061, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 228 + ], + "type": "text", + "content": "[43] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "type": "text", + "content": "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 547, + 328 + ], + "type": "text", + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 1, 2, 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5259" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_content_list.json b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70a8f3282b7b62b207f0e094d91d18fc5421f1b0 --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_content_list.json @@ -0,0 +1,1762 @@ +[ + { + "type": "text", + "text": "Unsupervised Universal Image Segmentation", + "text_level": 1, + "bbox": [ + 258, + 130, + 712, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dantong Niu*† Xudong Wang*† Xinyang Han* Long Lian Roei Herzig Trevor Darrell Berkeley AI Research, UC Berkeley", + "bbox": [ + 111, + 179, + 854, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code: https://github.com/u2seg/U2Seg", + "bbox": [ + 364, + 219, + 599, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f6c80cf4672c46b0a61ea77a23d3c8c725ee3dabd3b4a1b5d3bb276bc9e5dd2.jpg", + "image_caption": [ + "Figure 1. We present U2Seg, a unified framework for Unsupervised Universal image Segmentation that consistently outperforms previous state-of-the-art methods designed for individual tasks: CutLER [60] for unsupervised instance segmentation, STEGO [24] for unsupervised semantic segmentation, and the naive combination of CutLER and STEGO for unsupervised panoptic segmentation. We visualize instance segmentation results with \"semantic label\" + confidence score and semantic predictions with \"semantic label\". Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 83, + 247, + 879, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 614, + 313, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized", + "bbox": [ + 73, + 647, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods tailored to each task: $a + 2.6AP^{box}$ boost (vs. CutLER) in unsupervised instance segmentation on COCO and $a + 7.0$ PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOSTuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by $+5.0AP^{mask}$ when trained on a low-data regime, e.g., only $1\\%$ COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation.", + "bbox": [ + 500, + 599, + 893, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 797, + 632, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of image segmentation has witnessed significant advancements in the recent years [4, 5, 12, 20, 26, 35, 38, 40]. Nonetheless, the effectiveness of these segmentation methods heavily depends on the availability of extensive densely human-labeled data for training these models,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal Contribution; †: Project Lead.", + "bbox": [ + 99, + 886, + 308, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22744", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which is both labor-intensive and costly and thus less scalable. In this paper, our objective is to explore the extent to which unsupervised image segmentation can be achieved without relying on any human-generated labels.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several recent works such as CutLER [60] and STEGO [24] have emerged as promising approaches for unsupervised image segmentation. CutLER leverages the property of the self-supervised model DINO [8] to 'discover' objects without supervision, and learns a state-of-the-art localization model on pseudo instance segmentation masks produced by MaskCut [60] (based on Normalize Cuts [45]). Similarly leveraging DINO [8], STEGO [24] introduces a novel framework that distills unsupervised features into discrete semantic labels. This is achieved using a contrastive loss that encourages pixel features to form compact clusters while preserving their relationships across the corpora [24]. However, these methods have limitations:", + "bbox": [ + 75, + 152, + 470, + 349 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The output of unsupervised instance segmentation methods such as CutLER [60] comprises class-agnostic segments for \"things\", ignoring the \"stuff\" categories that represent pixel semantics. Moreover, CutLER often treats several overlapping instances as one instance, especially when these instances belong to the same semantic class.", + "- On the other hand, unsupervised semantic segmentation methods such as STEGO [24] focus on the segmentation of semantically coherent regions, lacking the capability to distinguish between individual instances.", + "- Unsupervised panoptic segmentation has not been addressed. Supervised panoptic segmentation methods [12, 29, 31] predict both \"stuff\" and \"things\" classes simultaneously; to the best of our knowledge there has not been work on unsupervised panoptic segmentation heretofore." + ], + "bbox": [ + 76, + 351, + 468, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose U2Seg, a novel Unsupervised Universal image Segmentation model. U2Seg offers comprehensive scene understanding—instance, semantic and panoptic—without relying on human annotations, segmenting semantically meaningful regions in the image as well as identifying and differentiating between individual instances within those regions.", + "bbox": [ + 75, + 580, + 468, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "U2Seg is comprised of three steps. First, we create high-quality, discrete semantic labels for instance masks obtained from MaskCut and DINO, by clustering semantically similar instance masks into distinct fine-grained clusters, as in Sec. 3.2. Next, we amalgamate the semantically pseudolabeled \"things\" pixels (from the first step) with \"stuff\" pixels (from STEGO) to produce pseudo semantic labels for each pixel in the image. Lastly, a universal image segmentation model is trained using these pseudo-labels, resulting in a model capable of simultaneously predicting pixel-level (i.e., semantic and class-agnostic instance segmentation) and instance-level semantic labels, detailed in Sec. 3.3.", + "bbox": [ + 75, + 686, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the inherent noise in these pseudo-labels, self-training the model with them yields substantial performance", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gains over specialized methods tailored to each task: U2Seg achieves a $+2.6$ $\\mathbf{AP}^{\\mathrm{box}}$ boost (vs. CutLER) in unsupervised instance segmentation on COCO and a $+7.0$ PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation. We also find that the multi-task learning framework and learning unsupervised segmentor with semantic labels enable our model to generate a more discriminative feature space, which makes it a superior representation for downstream supervised detection and segmentation tasks. When trained on a low-data regime, such as $1\\%$ COCO labels, U2Seg surpasses CutLER by $+5.0$ $\\mathbf{AP}^{\\mathrm{mask}}$ .", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our main contribution is the first universal unsupervised image segmentation model that can tackle unsupervised semantic-aware instance, semantic and panoptic segmentation tasks using a unified framework. We establish a suite of benchmarks on unsupervised semantic-aware instance segmentation and panoptic segmentation, areas previously unexplored. Despite using a single framework, we demonstrate that U2Seg surpasses previous methods specialized for each task across all experimented benchmarks (instance, semantic, panoptic, etc.) and datasets (COCO, Cityscapes, UVO, VOC, etc.).", + "bbox": [ + 496, + 287, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 467, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Representation Learning focuses on feature learning from a large amount of unlabeled data without using human-made labels. Contrastive Learning-Based Methods [9, 27, 43, 62] learn representation by comparing similar instances or different versions of a single instance while separating dissimilar ones. Similarity-Based Self-Supervised Learning [10, 23] mainly reduces differences between different augmented versions of the same instance. Clustering-Based Feature Learning [2, 7, 55, 63, 65] finds natural data groups in the hidden space. Masked Autoencoders [3, 18, 28] learn by masking and then reconstructing masked parts of the image.", + "bbox": [ + 496, + 492, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Object Detection and Instance Segmentation. DINO [8] shows that self-supervised learning (SSL) Vision Transformers (ViT) [19] can reveal hidden semantic segmentation in images, which is not obvious in supervised counterparts [8, 66]. Extending this, LOST [46], TokenCut [61] and MaskDistill [50] use DINO's patch features to identify main objects in images. FreeSOLO [58] performs unsupervised class-agnostic instance segmentation by creating coarse masks first, which are later improved through self-training. Meanwhile, CutLER [59] introduces the MaskCut method, which aims to identify multiple instances in a single image. Yet, MaskCut frequently consolidates overlapping instances into a single segment and lacks the capability to assign semantic labels to each instance.", + "bbox": [ + 496, + 674, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Semantic Segmentation. IIC [30] maxi", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22745", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mizes mutual information for clustering, while PiCIE [14] uses invariance to photometric effects and equivariance to geometric transformations for segmentation. MaskContrast [49] learns unsupervised semantic segmentation by contrasting features within saliency masks. STEGO [24] refines pretrained SSL visual features to distill correspondence information embedded within these features, thereby fostering discrete semantic clusters.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Universal Segmentation has been introduced to deliver instance, semantic and panoptic segmentation tasks using a unified architecture [6, 11-13, 29, 33, 34, 37, 52, 64]. In this work, we propose U2Seg to tackle this challenging task without relying on human-annotated data.", + "bbox": [ + 75, + 212, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised Image Classification methods mainly focus on providing a semantic label for each query image that can be mapped to ground truth classes by hungarian matching. SCAN [48] proposes a three-stage pipeline that includes representation learning, deep clustering, and self-labeling. NNM [16] enhances SCAN by incorporating local and global nearest neighbor matching. RUC [44] further improves SCAN using a robust loss as training objective. However, these approaches only provide one classification prediction per image, whereas our method provides classification per-instance for instance segmentation and per-pixel for semantic segmentation.", + "bbox": [ + 75, + 287, + 467, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Unsupervised Universal Segmentation", + "text_level": 1, + "bbox": [ + 76, + 484, + 419, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 510, + 217, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first explain the previous Unsupervised Instance Segmentation method CutLER [60], and Unsupervised Semantic Segmentation method STEGO [24].", + "bbox": [ + 75, + 534, + 467, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CutLER [60] exploits self-supervised learning models like DINO [8] to 'discover' objects and train a state-of-the-art detection and segmentation model using a cut-and-learn pipeline. It first uses MaskCut to extract multiple initial masks from DINO [8] features. MaskCut first generates a patch-wise affinity matrix $W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|K_j\\|_2}$ using the \"key\" features $K_i$ for patch $i$ from DINO's last attention layer. Subsequently, the cut-based clustering method Normalized Cut [45] is employed on the affinity matrix by finding the eigenvector $x$ that corresponds to the second smallest eigenvalue. A foreground instance mask $M^s$ is derived through bi-partitioning of the vector $x$ , enabling segmentation of individual objects in the image. For multi-instance segmentation, MaskCut iteratively refines the affinity matrix by masking out already segmented objects, allowing for subsequent extractions", + "bbox": [ + 75, + 580, + 467, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 837, + 467, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and repeating above steps by $N$ times. CutLER then refines", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection and segmentation through a loss-dropping strategy and iterative self-training.", + "bbox": [ + 496, + 90, + 888, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "STEGO [24] harnesses the semantically rich feature correlations produced by unsupervised methods like DINO [8] for segmentation. It trains a segmentation head to refine these correlations within an image, with its K-Nearest Neighbors (KNNs), and across randomly chosen images. Specifically, STEGO distills DINO's unsupervised features into distinct semantic labels by optimizing a correspondence loss. This loss function measures the feature correspondences $F^{SC}$ between image feature pairs generated by DINO and the feature correspondence $S_{hwij}$ derived from a trainable, lightweight segmentation head [24]:", + "bbox": [ + 496, + 119, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o r r}} (x, y, b) = - \\sum_ {h w i j} \\left(F _ {h w i j} ^ {S C} - b\\right) \\max \\left(S _ {h w i j}, 0\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 297, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Unsupervised Instance Segmentation", + "text_level": 1, + "bbox": [ + 500, + 343, + 818, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although CutLER [60] provides high-quality instance segmentation masks without human annotations, the predicted masks are class-agnostic, and thus do not include semantic labels for each instance. Our method addresses this issue by grouping the detected instances with a clustering method. In this way, instances assigned to the same cluster are associated with identical or closely related semantic information, while instances residing in separate clusters exhibit semantic dissimilarity.", + "bbox": [ + 496, + 366, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo Semantic Labels. To train a detection and instance segmentation model, we vector quantize the model targets (pseudo semantic labels) by clustering the instance-level features of the entire dataset, under constraints derived from self-supervision. Specifically, our approach starts with the generation of instance segmentation masks using MaskCut [60]. Subsequently, we utilize the efficient $K$ -Means clustering method as implemented in USL [57] to cluster all segmentation masks into semantically meaningful clusters.", + "bbox": [ + 496, + 503, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ $K$ -Means clustering to partition $n$ instances into $C(\\leq n)$ clusters, where each cluster is represented by its centroid $c$ [22, 41]. Each instance is assigned to the cluster with the nearest centroid. Formally, we conduct a $C$ -way node partitioning, denoted as $\\mathcal{S} = S_1, S_2, \\ldots, S_C$ , that minimizes the within-cluster sum of squares [36]:", + "bbox": [ + 496, + 638, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} \\sum_ {V \\in S _ {i}} | V - c _ {i} | ^ {2} = \\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} | S _ {i} | \\operatorname {V a r} \\left(S _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 741, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This optimization process is carried out iteratively using the EM algorithm [42], starting from selecting random samples as initial centroids. As a result, this process assigns pseudo semantic labels, denoted as $y_{i}$ , to each instance $i$ with $y_{i}$ falling within the range of $[1,C]$ .", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The resulting semantic labels serve multiple purposes: 1) Semantic-aware copy-paste augmentation, which signifi-", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22746", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0cdd369e78f9e985922848723320a3606944d07490bdc0ae7ce4b47a1bee4c38.jpg", + "image_caption": [ + "Figure 2. Overview of the training and inference pipeline for the proposed Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b67dbb46ef3a9411dfb950452309367d0c4ad6b01b5b7b1920e958b6f2d903f7.jpg", + "image_caption": [ + "Figure 3. Pipeline overview for generating masks and their semantically meaningful pseudo labels in semantic-aware instance segmentation. We first use MaskCut to generate class-agnostic instance masks, which are then grouped into semantically meaningful clusters. These pseudo semantic labels are used for training a semantic-aware instance segmentor." + ], + "image_footnote": [], + "bbox": [ + 143, + 227, + 831, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cantly improves CutLER's capability to differentiate overlapping instances, especially when they share similar semantic information. 2) Training instance segmentation models: They serve as pseudo ground-truth labels for training a non-agnostic instance segmentor.", + "bbox": [ + 75, + 402, + 468, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic-aware Copy-Paste Augmentation. In cluttered natural scenes, previous unsupervised instance segmentation model often fail to distinguish instances from the same semantic class. This results in multiple instances being captured in the same mask. To distinguish multiple overlapping objects and small objects in existing unsupervised detectors, we employ semantic-aware copy-paste augmentation, which includes several steps:", + "bbox": [ + 75, + 478, + 468, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) We begin by randomly selecting two instances, denoted as $I_{1}$ and $I_{2}$ , both belonging to the same pseudocategory (or group/cluster). 2) One of these instances undergoes a transformation function $\\mathcal{T}$ , which randomly reizes and shifts the associated pseudo-masks. 3) The resized instance is then pasted onto another image, creating synthetic overlapping cases using the following equation:", + "bbox": [ + 75, + 598, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {3} = I _ {1} \\cdot (1 - \\mathcal {T} (M _ {c})) + I _ {2} \\cdot \\mathcal {T} (M _ {c}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 710, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cdot$ denotes element-wise multiplication.", + "bbox": [ + 75, + 734, + 372, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Unsupervised Instance Segmentor. Traditionally, unsupervised segmentation community focused primarily on class-agnostic instance segmentation [58, 60, 61], whose outputs lack class labels. However, by incorporating clustering information obtained from pseudo-labels on ImageNet, as discussed above, our method allows the model to predict not only the location and segmentation of an object but also its pseudo semantic labels.", + "bbox": [ + 75, + 750, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As observed by [60], \"ground-truth\" masks may miss instances. However, a standard detection loss penalizes", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predicted regions $r_i$ that do not overlap with the \"ground-truth\". Therefore, following [60], we drop the loss for each predicted region $r_i$ that has a maximum overlap of $\\tau^{\\mathrm{IoU}}$ with any of the 'ground-truth' instances: $\\mathcal{L}_{\\mathrm{drop}}(r_i) = \\mathbb{1}(\\mathrm{IoU}_i^{\\max} > \\tau^{\\mathrm{IoU}})\\mathcal{L}_{\\mathrm{vanilla}}(r_i)$ , where $\\mathrm{IoU}_i^{\\max}$ denotes the maximum IoU with all 'ground-truth' for $r_i$ and $\\mathcal{L}_{\\mathrm{vanilla}}$ is the vanilla loss function of detectors. $\\mathcal{L}_{\\mathrm{drop}}$ encourages the exploration of image regions missed in the \"ground-truth\".", + "bbox": [ + 496, + 402, + 892, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Unsupervised Universal Image Segmentation", + "text_level": 1, + "bbox": [ + 498, + 530, + 880, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo Labels for Panoptic Segmentation. For each pixel $(i,j)$ in the image, we vector quantize pixels with different semantics or instance membership, generating pseudo semantic labels for panoptic segmentation. We assign each pixel a semantic label based on \"stuff\" or \"things\" identity. This results in an instance label $(I(i,j))$ for \"things\" or a semantic label $(S(i,j))$ for \"stuff\". The critical challenge in this process is distinguishing between pixels associated with \"things\" (countable, often foreground) and \"stuff\" (uncountable, usually background) [1].", + "bbox": [ + 496, + 551, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To resolve this problem, our method unfolds in three steps: 1) Semantic Labeling for \"Things\": Utilizing the class-agnostic instance segmentation capabilities of CutLER [60], we first identify \"things\" within an image, generating class-agnostic instance masks. These masks then undergo deep clustering to attribute a semantic label $I_C(i,j)$ to each instance, detailed in Sec. 3.2. 2) Semantic Labeling for \"Stuff\": For \"stuff\" pixels, we deploy the unsupervised semantic segmentation model STEGO [24], which distills DINO's unsupervised features into discrete semantic labels, as outlined in 3.1. This step assigns a \"stuff\" semantic label to all pixels, including those of \"Things\" identified earlier. 3) Integrating Labels for \"Things\" and \"Stuff\". We deter", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/337bc35090a1995e040b92da3caf0ba23189d9156b13a6cb8895d1f2cef56a0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task → Datasets → Metric →Agn Instance Seg.Instance Seg.Semantic Seg.Panoptic Seg.
COCOCOCOVOCUVOCOCOCOCOCityscapes
APboxAPbox50APbox50ARbox100APbox50ARbox100APbox50ARbox100PixelAccmIoUPQSQRQPQSQRQ
FreeSOLO [58]9.64.2-------------
TokenCut [61]5.83.2-------------
DINO [8]--------30.59.6-----
PiCIE + H [14]--------48.113.8-----
STEGO [24]--------56.928.2-----
CutLER [59]21.912.3-------------
CutLER+--9.010.326.827.210.611.8-------
CutLER+STEGO----------12.464.915.512.436.115.2
U2Seg22.813.011.821.531.048.110.825.063.930.216.171.119.917.652.721.7
vs.prev.SOTA+0.9+0.7+2.8+11.2+4.2+20.9+0.2+13.2+7.0+2.0+3.7+6.2+4.4+5.2+16.6+6.5
", + "bbox": [ + 84, + 85, + 883, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. With a unified framework, U2Seg outperforms previous state-of-the-art methods tailored for individual tasks across various datasets, including CutLER for unsupervised instance segmentation, STEGO for unsupervised semantic segmentation, and CutLER+STEGO for unsupervised panoptic segmentation. \"Agn Instance Seg\" denotes class-agnostic instance segmentation.", + "bbox": [ + 75, + 258, + 892, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mine a pixel's classification as \"things\" or \"stuff\" using the following logic:", + "bbox": [ + 75, + 327, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (i, j) = \\left\\{ \\begin{array}{l l} I _ {C} (i, j), & \\text {i f} I _ {C} (i, j) \\neq 0 \\\\ S _ {S} (i, j), & \\text {i f} I _ {C} (i, j) = 0 \\& S _ {S} (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 371, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This process merges the semantic labels, assigning priority to \"things\" labels over \"stuff\" where applicable. We then train a universal segmentation model on these pseudo-labels for instance, semantic and panoptic segmentation tasks.", + "bbox": [ + 75, + 439, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning Unsupervised Universal Image Segmentor. After we obtain the pseudo labels for panoptic segmentation, following [33], we construct an unsupervised universal image segmentation model, that has two branches: instance segmentation branch and semantic segmentation branch, to address corresponding segmentation tasks. The model is trained jointly for both branches, employing the following loss function: $\\mathcal{L} = \\lambda_i(\\mathcal{L}_c + \\mathcal{L}_b + \\mathcal{L}_m) + \\lambda_s\\mathcal{L}_s$ , where $\\mathcal{L}_c$ represents the classification loss, $\\mathcal{L}_b$ is the detection loss, $\\mathcal{L}_m$ is the segmentation loss, and $\\mathcal{L}_s$ signifies the semantic loss. The $\\mathcal{L}_s$ is computed as a per-pixel cross-entropy loss between the predicted and ground-truth labels. The hyperparameters $\\lambda_{i}$ and $\\lambda_{s}$ balance these two parts.", + "bbox": [ + 75, + 501, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 713, + 310, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 739, + 267, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. Our model is trained on 1.3M unlabeled images from ImageNet [17] and is evaluated directly across various benchmarks, unless otherwise noted. For unsupervised semantic segmentation comparisons with STEGO [24], we additionally fine-tune our model using MSCOCO's unlabeled images, following STEGO [24].", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test Data. For unsupervised instance segmentation, we test our model on COCO val2017, PASCAL VOC val2012 [21] and UVO val [53]. For unsupervised", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "panoptic segmentation, we evaluate our model on COCO val2017 and Cityscapes val [15].", + "bbox": [ + 498, + 327, + 890, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use AP, $\\mathrm{AP}_{50}$ , $\\mathrm{AP}_{75}$ and $\\mathrm{AR}_{100}$ to evaluate the unsupervised instance segmentation; PixelAcc and mIoU for unsupervised semantic segmentation; PQ, RQ, SQ for unsupervised universal image segmentation. After predicting the instance with its semantic labels, we use Hungarian matching to map the semantic labels to class names in the real dataset (details in B). It evaluates the consistency of the predicted semantic segments with the ground truth labels, remaining unaffected by any permutations in the predicted class labels.", + "bbox": [ + 496, + 359, + 892, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following [32], we employ Panoptic Cascade Mask R-CNN [4, 32] with a ResNet50 backbone [25]. Following CutLER's training recipe [60], our model, initialized with DINO pre-trained weights, is trained on unlabeled ImageNet for two epochs. It starts with an initial learning rate of 0.01, which then decreases after the first epoch to $5 \\times 10^{-5}$ , with a batch size of 16 for all models. For unsupervised panoptic segmentation, we maintain the same training schedule as unsupervised instance segmentation for zero-shot evaluation. In non-zero-shot scenarios, the models undergo training on a combination of unlabeled COCO and ImageNet datasets, beginning with a learning rate of 0.01 over 90k steps.", + "bbox": [ + 496, + 512, + 893, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unsupervised Universal Image Segmentation", + "text_level": 1, + "bbox": [ + 498, + 722, + 880, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, U2Seg represents the first effort in addressing unsupervised semantic-aware instance, semantic and panoptic segmentation, all unified under a single framework. Due to the absence of benchmarks for unsupervised semantic-aware instance segmentation and panoptic segmentation, we establish comprehensive benchmarks and baselines for both tasks.", + "bbox": [ + 496, + 747, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we demonstrate that U2Seg, utilizing a unified framework, significantly outperforms all previous approaches across various benchmarks and datasets. For", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22748", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7c884a3a81406a16d8fabc1d4a8ff342f46d098f8830d874e27853656dcb03ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric\\( AP^{box} \\)\\( AP^{box}_{50} \\)\\( AP^{box}_{75} \\)\\( AR^{box}_{100} \\)\\( AP^{mask} \\)\\( AP^{mask}_{50} \\)\\( AP^{mask}_{75} \\)\\( AR^{mask}_{100} \\)
CutLER+5.99.06.110.35.38.65.59.3
U2Seg7.311.87.521.56.411.26.418.5
Δ+1.4+2.8+1.4+11.2+1.1+2.6+0.9+9.2
", + "bbox": [ + 78, + 88, + 467, + 145 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b4f89d70b82cad555d473754042dcf32903e35f2a312759c2006e3401874543.jpg", + "table_caption": [ + "Table 2. The results for zero-shot unsupervised object detection and instance segmentation on COCO val2017. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + ], + "table_footnote": [], + "table_body": "
MethodsAPboxAP50AP75AR100
CutLER+17.126.818.127.2
U2Seg19.031.019.548.1
Δ+1.9+4.2+1.4+20.9
", + "bbox": [ + 81, + 213, + 464, + 273 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/115ddc8823d04fe21a9a1a4223a37e966ea41f7430811b7a618ef3891d08ba4a.jpg", + "table_caption": [ + "Table 3. The results for zero-shot unsupervised object detection on PASCAL VOC va12012. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + ], + "table_footnote": [], + "table_body": "
Metric\\(AP^{box}\\)\\(AP^{box}\\)\\(50\\)\\(AR^{box}\\)\\(100\\)\\(AP^{mask}\\)\\(AP^{mask}\\)\\(50\\)\\(AR^{mask}\\)\\(100\\)
CutLER+6.310.611.86.09.010.4
U2Seg6.810.825.06.29.521.0
Δ+0.5+0.2+13.2+0.2+0.5+10.6
", + "bbox": [ + 81, + 342, + 464, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class-agnostic unsupervised instance segmentation, our method achieves an increase of $+0.9$ in $\\mathrm{AP^{box}}$ compared to CutLER [60]. This improvement is largely attributed to our novel semantic-aware copy-paste augmentation, as detailed in Sec. 3.2. For unsupervised semantic-aware instance segmentation, we benchmark against the advanced baseline CutLER+, derived from CutLER, and record a substantial gain of over $11.2\\%$ in AR. A more comprehensive analysis of these results is provided in Sec. 4.3. For unsupervised semantic segmentation, our approach surpasses the state-of-the-art STEGO with impressive margins of $+7.0$ in PixelAcc and $+2.0$ in mIoU. Lastly, for unsupervised panoptic segmentation, we compare against the strong baseline of CutLER+STEGO, a hybrid of CutLER+ and STEGO, and observe performance gains of over $6.2\\%$ in SQ on MSCOCO and a notable $16.6\\%$ improvement in SQ on Cityscapes. Further comparisons and discussions on this task are elaborated in Sec. 4.4.", + "bbox": [ + 75, + 489, + 467, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Unsupervised Instance Segmentation", + "text_level": 1, + "bbox": [ + 76, + 771, + 395, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We performed extensive experiments for zero-shot unsupervised instance segmentation. Given that prior methods [51, 58, 60, 61] are limited to class-agnostic instance segmentation, we developed CutLER+, a strong baseline for unsupervised semantic-aware instance segmentation, building upon the current state-of-the-art CutLER [60]. CutLER+ operates in two steps: it first uses the pre-trained", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc90c06b085655a8e9428566e73bdc0d232015a8fc9e14924b15d7ebedaccb1c.jpg", + "table_caption": [ + "Table 4. The results for zero-shot unsupervised object detection and instance segmentation on UVO va1. The model is trained on ImageNet with a cluster number of 800. We compare with CutLER+, a combination of CutLER and offline clustering." + ], + "table_footnote": [], + "table_body": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN15.746.619.8
non zero-shot methods
CutLER+STEGOCOCO12.436.115.2
U2SegCOCO15.451.519.0
U2SegCOCO+IN17.652.721.7
Δ+5.2+16.6+6.5
", + "bbox": [ + 501, + 89, + 890, + 203 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8fb8c50919c64fa25dc416fcdfd68847fbe08e52f0d5418a0662fcdc2547db7.jpg", + "table_caption": [ + "Table 5. Unsupervised Panoptic image segmentation on Cityscapes val. We show PQ, SQ and RQ on zero-shot and nonzero shot settings with the cluster number of 800. We compare with CutLER+STEGO, a combination of CutLER+ and STEGO." + ], + "table_footnote": [], + "table_body": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN11.160.113.7
non zero-shot methods
CutLER+STEGOCOCO12.464.915.5
U2SegCOCO15.366.519.1
U2SegCOCO+IN16.171.119.9
Δ+3.7+6.2+4.4
", + "bbox": [ + 504, + 272, + 887, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6. Unsupervised Panoptic image segmentation on COCOval2017. We show PQ, SQ and RQ on zero-shot and non-zero shot settings. We use CutLER+STEGO, a combination of CutLER+ and STEGO, as a strong baseline.", + "bbox": [ + 498, + 392, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CutLER to generate class-agnostic instance masks, and subsequently assigns semantic labels to all instance masks through offline clustering.", + "bbox": [ + 498, + 474, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 demonstrates that U2Seg markedly improves performance in both unsupervised object detection and instance segmentation on MSCOCO, delivering a +2.8 boost in $\\mathrm{AP}_{50}^{\\mathrm{box}}$ and a +2.6 rise in $\\mathrm{AP}_{50}^{\\mathrm{mask}}$ over CutLER+. Additionally, our method sees a substantial increase of approximately +10.0 in $\\mathrm{AR}_{100}$ . Results on PASCAL VOC val2012 and UVO val are detailed in Tab. 3 and Tab. 4, respectively. Notably, we achieve gains exceeding +20% in AR for PASCAL VOC and +10% for UVO.", + "bbox": [ + 496, + 518, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Unsupervised Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 666, + 821, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For unsupervised panoptic/universal image segmentation, our experiments span two scenarios. In the zero-shot setting, the model is trained exclusively on unlabeled ImageNet images. For non-zero-shot (in-domain) scenarios, we train on unlabeled COCO images or a mix of COCO and ImageNet. With no existing benchmarks for unsupervised panoptic segmentation, we establish a new baseline by integrating the state-of-the-art unsupervised semantic segmentation from STEGO [24] with semantic-aware instance segmentation from CutLER+ (discussed in Sec. 4.3), which are then merged to create panoptic/universal segmentation outcomes, referred to as CutLER+STEGO.", + "bbox": [ + 496, + 689, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 6 presents the PQ, SQ, and RQ scores of U2Seg on COCO val2017. U2Seg surpasses the strong baseline", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22749", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57d1555519e3d36a924bb9e55a21e78830f8f13c164617af05a82981a0b9c9bd.jpg", + "image_caption": [ + "Figure 4. Universal image segmentation visualization in COCO val2017. We present the results with cluster IDs predicted by U2Seg, categorizing athletes playing hockey (left columns) as “139”, those playing badminton (middle columns) as “52” and gentlemen (right columns) as “132”. After Hungarian matching, the IDs are automatically matched to the category “person” for quantitative evaluations." + ], + "image_footnote": [], + "bbox": [ + 83, + 95, + 176, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d424ff81ae77e5baee7a4d75b7198d0d2a0d2b6f50bcb4756b16bee6274d85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 95, + 251, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7a021221a3fee0363e41bda4dad055c90a7cf0b8d84137ba965ee586e87b31d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 95, + 354, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4200f1116b4bc2f76ba55e184002336dd93428cb00e3f6947e5c3b420ce0bec5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 95, + 449, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/045dcb50dd5752b8c82220b7b529c8f4daacb2e082d94a5eb0feeb4bc2824cfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 95, + 558, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01c2e184134b1cf9e405a63cd0a2b19210ccdaef546acfb4701813fd2c0d3ba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 95, + 637, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c8f717cb9bd25917d102ee3a56d6e441452ee896d0646d815e359b8e0126a9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 95, + 725, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6472ff337e015fb0d62bbc919fdf3a116f062518a8cdfb250e3d1fe92b1a1421.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 95, + 787, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/691a1b5dd5ef893a22d178fccefb94463f4258c51caeca36cb7f2ef96cf64018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 95, + 875, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f08e5400b55537038db8311b872ec1631027dbe0de811e03a6a86bf6ed6d656c.jpg", + "image_caption": [ + "Raw Image", + "U2Seg", + "Ground-Truth", + "Pseudo-Labels", + "Figure 5. Visualizations of U2Seg's unsupervised Panoptic segmentation results on COCO val2017 (after Hungarian matching). The pseudo label is the naive combination of previous state-of-the-art instance segmentation, i.e. CutLER [60], and semantic segmentation, i.e., STEGO [24], results." + ], + "image_footnote": [], + "bbox": [ + 89, + 250, + 455, + 454 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faa57eef681a40a5bb4b7deaf25aa8b595a7e95d2e1fd7054e8431c71a2a7e08.jpg", + "image_caption": [ + "Raw Image", + "U2Seg", + "Ground-Truth", + "Pseudo-Labels", + "Figure 6. Qualitative results of U2Seg's Panoptic image segmentation results on Cityscapes val (after Hungarian matching)." + ], + "image_footnote": [], + "bbox": [ + 89, + 565, + 455, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CutLER+STEGO with a +3.5 improvement in PQ and an increase of over +4.0 in RQ. Qualitative results of U2Seg's performance is provided in Fig. 4, with the predicted seman", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tic labels visualized. The qualitative results suggest that an over-clustering strategy in pseudo-label generation, e.g. setting the number of clusters to 300 or 800, leads to highly granular model predictions. For instance, as in Fig. 4, the model distinctly categorizes hockey players as “139”, badminton players as “52”, and gentlemen in suits as “132”, showcasing its potent discriminative capabilities.", + "bbox": [ + 496, + 255, + 890, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantitatively measure the quality of segmentation masks and their corresponding semantic labels, we use Hungarian matching (detailed in Appendix B) to align semantic labels with the category names from the test dataset; for instance, all three sub-clusters depicted in Fig. 4 are assigned to the \"person\" category. The qualitative outcomes post-Hungarian matching are shown in Fig. 5, where our model demonstrates superior panoptic segmentation mask quality. For instance, while the baseline tends to segment parts separately (as seen with the man's head and torso being treated as separate entities in the third row), our model correctly identifies them as parts of a single object. This level of recognition is also evident with the \"trunks of the motorcycle\" example in the second row. For additional results, please see Appendix D. We also present results of the more challenging dataset Cityscapes in Tab. 5 and Fig. 6.", + "bbox": [ + 496, + 362, + 892, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Efficient Learning", + "text_level": 1, + "bbox": [ + 500, + 617, + 676, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for object detection and instance segmentation, we employ our unsupervised instance segmentation model, with cluster count set to 300, to initialize the model weights. We adopt the recipe from [56, 59] for model fine-tuning across various annotation splits. For label-efficient panoptic segmentation, we fine-tune the model initialized with our zero-shot unsupervised framework on the same data splits.", + "bbox": [ + 496, + 642, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are depicted in Fig. 7, where our model's instance segmentation performance is benchmarked against MoCo-V2, DETReg, and CutLER. Our model consistently surpasses the state-of-the-art with consistent gains in both $\\mathrm{AP^{box}}$ and $\\mathrm{AP^{mask}}$ . In scenarios with panoptic image segmentation as the downstream task, we contrast our results with MoCo-V2 and CutLER in terms of PQ, SQ, and RQ metrics. The results illustrate a remarkable improvement, effectively doubling the performance boost from MoCo-V2 to CutLER, especially in few-shot contexts with limited an-", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22750", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf83c6c4ce582435bd099f332bf0d1ea509197b3dc0f0c5065a608a7d3e771a7.jpg", + "image_caption": [ + "Figure 7. We evaluate the label-efficient learning performance on 3 different tasks: object detection (the left), instance segmentation (the second left) and panoptic image segmentation (the last three)." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 251, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db12f59874e9387e28a98187214e676f8c0d4ab543a7c339d0d3c9bb4f9dc8e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 85, + 406, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00535cc70a101c8c1801a2aec1ff91ad2cffe36ef73ad5c84f9ddc2baae8040e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 85, + 562, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d4f66e960707b4a7140125fd020a3bb1605ae088a8126076e49089439809a2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 85, + 717, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae688957d09ee9443bf9d72e0c9b49b955c3bb8b596608560320bd616347f608.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 85, + 870, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f7e9dabd66cb0d7ceb2d5564a48a4496f35dad5d521b0bc3866592f8b0b68c0.jpg", + "image_caption": [ + "Figure 8. U2Seg learns features that are more discriminative than those learned by CutLER. The t-SNE [47] visualization of the features from the model's FC layer. We color-code each dot based on its ground-truth category." + ], + "image_footnote": [], + "bbox": [ + 106, + 280, + 267, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ec9a34b1971e2419382e4e33cee5a65505fe30ec262cb03d484d8fe4dbab502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 280, + 441, + 465 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2bdf9351a003bd72e3fe7d9275a5604ceb31e7611e3c328d5359959c772e945e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# ClusterCOCOUVOVOC
APbox50ARbox100APbox50ARbox100APbox50ARbox100
3009.320.19.822.629.645.7
80011.821.510.825.031.048.0
291113.322.115.125.831.648.3
", + "bbox": [ + 81, + 535, + 464, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "notations (1% or 2% labeled samples). This highlights the practical value of our approach in real-world unsupervised learning applications, where annotations are often scarce.", + "bbox": [ + 75, + 671, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We attribute the performance gains primarily to the discriminative features our model learns, as in Fig. 8, obtaining effective model initialization for few-shot learning.", + "bbox": [ + 75, + 717, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 771, + 240, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation study on U2Seg.", + "bbox": [ + 75, + 794, + 423, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Numbers of clusters. The choice of cluster quantity significantly affects the model's representation granularity. Our ablation study on various cluster counts, as detailed in Tab. 7, reveals their impact on model performance. Overclustering generally leads to a finer level of detail, prompting the model to learn more discriminative features.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2b45a1f35f4f8863eb1118c8cadc00f2fbc12741896cd7f8b78f3073434006e.jpg", + "table_caption": [ + "Table 7. Over-clustering can improve the model performance. We show results on different datasets for the unsupervised object detection using different cluster numbers." + ], + "table_footnote": [], + "table_body": "
conf# matchedAPbox50ARbox100
0.910910.913.1
0.722511.618.0
0.628211.819.7
0.438911.821.5
0.251311.321.8
0.07188.618.4
", + "bbox": [ + 501, + 277, + 684, + 367 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d429384adec8615b954f65a4ee9c07af6b954a65fb76b79cdfe783acaac19722.jpg", + "table_caption": [ + "(a) Conf's effect on accuracy." + ], + "table_footnote": [], + "table_body": "
IoU# matchedAPbox50ARbox100
0.929510.819.7
0.834811.420.7
0.441411.521.6
0.245011.521.1
0.04949.217.7
0.638911.821.5
", + "bbox": [ + 715, + 277, + 890, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) IoU's effect on accuracy.", + "bbox": [ + 715, + 369, + 890, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Impact of Confidence and IoU on Hungarian Matching Performance: The left table illustrates the outcomes at a fixed IoU of 0.6 while varying the confidence scores. Conversely, the right table displays the results with a constant confidence of 0.4, altering the IoU values. The cluster number is 800.", + "bbox": [ + 498, + 386, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hungarian matching. As our trained model could predict the instance with corresponding semantic labels, we are able to go further beyond unsupervised class-agnostic instance segmentation. To quantitatively evaluate the performance, Hungarian matching is employed to match the predicted semantic labels to the ground-truth dataset categories. See Appendix B for details of the adopted Hungarian matching used in our evaluation. As shown in Tab. 8, the two parameters conf threshold and IoU threshold also affect the precision and recall.", + "bbox": [ + 496, + 482, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Summary", + "text_level": 1, + "bbox": [ + 500, + 647, + 607, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present U2Seg, a novel Unsupervised Universal Image Segmentation model, adept at performing unsupervised instance, semantic, and panoptic segmentation tasks within a unified framework. Evaluated on extensive benchmarks, U2Seg consistently outperforms previous state-of-the-art methods designed for individual tasks. Additionally, U2Seg achieves the new state-of-the-art for label-efficient panoptic segmentation and instance segmentation. We anticipate that U2Seg, free from the constraints of human annotations, will demonstrate enhanced performance when scaled up with more training data, representing an important direction for our future research.", + "bbox": [ + 496, + 672, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: Trevor Darrell, Dantong Niu and Xudong Wang were funded by DoD including DARPA LwLL and the Berkeley AI Research (BAIR) Commons.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22751", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edward H Adelson. On seeing stuff: the perception of materials by humans and machines. In Human Vision and Electronic Imaging, 2001.", + "[2] YM Asano, C Rupprecht, and A Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2019.", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021.", + "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018.", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020.", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021.", + "[11] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-DeepLab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020.", + "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021.", + "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022.", + "[14] Jang Hyun Cho, U. Mall, K. Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. ArXiv, abs/2103.17070, 2021.", + "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. CoRR, abs/1604.01685, 2016." + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13693-13702, 2021.", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[18] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015.", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[20] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019.", + "[21] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.", + "[22] Edward W Forgy. Cluster analysis of multivariate data: efficiency versus interpretability of classifications. biometrics, 21:768-769, 1965.", + "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020.", + "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022.", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[26] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017.", + "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020.", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022.", + "[29] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. 2023." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22752", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019.", + "[31] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9396-9405, 2018.", + "[32] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019.", + "[33] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019.", + "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019.", + "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.", + "[36] Hans-Peter Kriegel, Erich Schubert, and Arthur Zimek. The (black) art of runtime evaluation: Are we comparing algorithms or implementations? Knowledge and Information Systems, 52(2):341-378, 2017.", + "[37] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. arXiv preprint arXiv:2108.07682, 2021.", + "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022.", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017.", + "[41] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982.", + "[42] Geoffrey J McLachlan and Thriyambakam Krishnan. The EM algorithm and extensions. John Wiley & Sons, 2007.", + "[43] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020.", + "[44] Sungwon Park, Sungwon Han, Sundong Kim, Danu Kim, Sungkyu Park, Seunghoon Hong, and Meeyoung Cha. Improving unsupervised image clustering with robust learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12278-12287, 2021.", + "[45] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021.", + "[47] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008.", + "[48] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020.", + "[49] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. arxiv preprint arxiv:2102.06191, 2021.", + "[50] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022.", + "[51] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021.", + "[52] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021.", + "[53] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, openworld segmentation. CoRR, abs/2104.04691, 2021.", + "[54] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10776-10785, 2021.", + "[55] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021.", + "[56] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2021.", + "[57] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In European Conference on Computer Vision, pages 427-445. Springer, 2022.", + "[58] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022.", + "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023.", + "[60] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22753", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023.", + "[61] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022.", + "[62] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3733-3742, 2018.", + "[63] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016.", + "[64] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019.", + "[65] Chengxu Zhuang, Alex Lin Zhai, Daniel Yamins, , et al. Local aggregation for unsupervised learning of visual embeddings. In ICCV, 2019.", + "[66] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 473 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22754", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_model.json b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f1a548becfb92680ff93ea9b4f32c0634c0e2dde --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_model.json @@ -0,0 +1,2587 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.131, + 0.713, + 0.154 + ], + "angle": 0, + "content": "Unsupervised Universal Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.18, + 0.856, + 0.217 + ], + "angle": 0, + "content": "Dantong Niu*† Xudong Wang*† Xinyang Han* Long Lian Roei Herzig Trevor Darrell Berkeley AI Research, UC Berkeley" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.22, + 0.601, + 0.234 + ], + "angle": 0, + "content": "Code: https://github.com/u2seg/U2Seg" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.248, + 0.88, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.526, + 0.893, + 0.582 + ], + "angle": 0, + "content": "Figure 1. We present U2Seg, a unified framework for Unsupervised Universal image Segmentation that consistently outperforms previous state-of-the-art methods designed for individual tasks: CutLER [60] for unsupervised instance segmentation, STEGO [24] for unsupervised semantic segmentation, and the naive combination of CutLER and STEGO for unsupervised panoptic segmentation. We visualize instance segmentation results with \"semantic label\" + confidence score and semantic predictions with \"semantic label\". Zoom in for the best view." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.615, + 0.314, + 0.631 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.648, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.601, + 0.895, + 0.768 + ], + "angle": 0, + "content": "methods tailored to each task: \\( a + 2.6AP^{box} \\) boost (vs. CutLER) in unsupervised instance segmentation on COCO and \\( a + 7.0 \\) PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOSTuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by \\( +5.0AP^{mask} \\) when trained on a low-data regime, e.g., only \\( 1\\% \\) COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.799, + 0.633, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The field of image segmentation has witnessed significant advancements in the recent years [4, 5, 12, 20, 26, 35, 38, 40]. Nonetheless, the effectiveness of these segmentation methods heavily depends on the availability of extensive densely human-labeled data for training these models," + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.31, + 0.901 + ], + "angle": 0, + "content": "*: Equal Contribution; †: Project Lead." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "which is both labor-intensive and costly and thus less scalable. In this paper, our objective is to explore the extent to which unsupervised image segmentation can be achieved without relying on any human-generated labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.471, + 0.35 + ], + "angle": 0, + "content": "Several recent works such as CutLER [60] and STEGO [24] have emerged as promising approaches for unsupervised image segmentation. CutLER leverages the property of the self-supervised model DINO [8] to 'discover' objects without supervision, and learns a state-of-the-art localization model on pseudo instance segmentation masks produced by MaskCut [60] (based on Normalize Cuts [45]). Similarly leveraging DINO [8], STEGO [24] introduces a novel framework that distills unsupervised features into discrete semantic labels. This is achieved using a contrastive loss that encourages pixel features to form compact clusters while preserving their relationships across the corpora [24]. However, these methods have limitations:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.352, + 0.47, + 0.442 + ], + "angle": 0, + "content": "- The output of unsupervised instance segmentation methods such as CutLER [60] comprises class-agnostic segments for \"things\", ignoring the \"stuff\" categories that represent pixel semantics. Moreover, CutLER often treats several overlapping instances as one instance, especially when these instances belong to the same semantic class." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.444, + 0.469, + 0.502 + ], + "angle": 0, + "content": "- On the other hand, unsupervised semantic segmentation methods such as STEGO [24] focus on the segmentation of semantically coherent regions, lacking the capability to distinguish between individual instances." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.504, + 0.469, + 0.579 + ], + "angle": 0, + "content": "- Unsupervised panoptic segmentation has not been addressed. Supervised panoptic segmentation methods [12, 29, 31] predict both \"stuff\" and \"things\" classes simultaneously; to the best of our knowledge there has not been work on unsupervised panoptic segmentation heretofore." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.352, + 0.47, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.686 + ], + "angle": 0, + "content": "To address these limitations, we propose U2Seg, a novel Unsupervised Universal image Segmentation model. U2Seg offers comprehensive scene understanding—instance, semantic and panoptic—without relying on human annotations, segmenting semantically meaningful regions in the image as well as identifying and differentiating between individual instances within those regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.868 + ], + "angle": 0, + "content": "U2Seg is comprised of three steps. First, we create high-quality, discrete semantic labels for instance masks obtained from MaskCut and DINO, by clustering semantically similar instance masks into distinct fine-grained clusters, as in Sec. 3.2. Next, we amalgamate the semantically pseudolabeled \"things\" pixels (from the first step) with \"stuff\" pixels (from STEGO) to produce pseudo semantic labels for each pixel in the image. Lastly, a universal image segmentation model is trained using these pseudo-labels, resulting in a model capable of simultaneously predicting pixel-level (i.e., semantic and class-agnostic instance segmentation) and instance-level semantic labels, detailed in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Despite the inherent noise in these pseudo-labels, self-training the model with them yields substantial performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "gains over specialized methods tailored to each task: U2Seg achieves a \\(+2.6\\) \\(\\mathbf{AP}^{\\mathrm{box}}\\) boost (vs. CutLER) in unsupervised instance segmentation on COCO and a \\(+7.0\\) PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation. We also find that the multi-task learning framework and learning unsupervised segmentor with semantic labels enable our model to generate a more discriminative feature space, which makes it a superior representation for downstream supervised detection and segmentation tasks. When trained on a low-data regime, such as \\(1\\%\\) COCO labels, U2Seg surpasses CutLER by \\(+5.0\\) \\(\\mathbf{AP}^{\\mathrm{mask}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Contributions. Our main contribution is the first universal unsupervised image segmentation model that can tackle unsupervised semantic-aware instance, semantic and panoptic segmentation tasks using a unified framework. We establish a suite of benchmarks on unsupervised semantic-aware instance segmentation and panoptic segmentation, areas previously unexplored. Despite using a single framework, we demonstrate that U2Seg surpasses previous methods specialized for each task across all experimented benchmarks (instance, semantic, panoptic, etc.) and datasets (COCO, Cityscapes, UVO, VOC, etc.)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Self-supervised Representation Learning focuses on feature learning from a large amount of unlabeled data without using human-made labels. Contrastive Learning-Based Methods [9, 27, 43, 62] learn representation by comparing similar instances or different versions of a single instance while separating dissimilar ones. Similarity-Based Self-Supervised Learning [10, 23] mainly reduces differences between different augmented versions of the same instance. Clustering-Based Feature Learning [2, 7, 55, 63, 65] finds natural data groups in the hidden space. Masked Autoencoders [3, 18, 28] learn by masking and then reconstructing masked parts of the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Unsupervised Object Detection and Instance Segmentation. DINO [8] shows that self-supervised learning (SSL) Vision Transformers (ViT) [19] can reveal hidden semantic segmentation in images, which is not obvious in supervised counterparts [8, 66]. Extending this, LOST [46], TokenCut [61] and MaskDistill [50] use DINO's patch features to identify main objects in images. FreeSOLO [58] performs unsupervised class-agnostic instance segmentation by creating coarse masks first, which are later improved through self-training. Meanwhile, CutLER [59] introduces the MaskCut method, which aims to identify multiple instances in a single image. Yet, MaskCut frequently consolidates overlapping instances into a single segment and lacks the capability to assign semantic labels to each instance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unsupervised Semantic Segmentation. IIC [30] maxi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22745" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "mizes mutual information for clustering, while PiCIE [14] uses invariance to photometric effects and equivariance to geometric transformations for segmentation. MaskContrast [49] learns unsupervised semantic segmentation by contrasting features within saliency masks. STEGO [24] refines pretrained SSL visual features to distill correspondence information embedded within these features, thereby fostering discrete semantic clusters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.468, + 0.288 + ], + "angle": 0, + "content": "Universal Segmentation has been introduced to deliver instance, semantic and panoptic segmentation tasks using a unified architecture [6, 11-13, 29, 33, 34, 37, 52, 64]. In this work, we propose U2Seg to tackle this challenging task without relying on human-annotated data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.468, + 0.471 + ], + "angle": 0, + "content": "Unsupervised Image Classification methods mainly focus on providing a semantic label for each query image that can be mapped to ground truth classes by hungarian matching. SCAN [48] proposes a three-stage pipeline that includes representation learning, deep clustering, and self-labeling. NNM [16] enhances SCAN by incorporating local and global nearest neighbor matching. RUC [44] further improves SCAN using a robust loss as training objective. However, these approaches only provide one classification prediction per image, whereas our method provides classification per-instance for instance segmentation and per-pixel for semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.486, + 0.42, + 0.503 + ], + "angle": 0, + "content": "3. Unsupervised Universal Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.218, + 0.525 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.468, + 0.58 + ], + "angle": 0, + "content": "We first explain the previous Unsupervised Instance Segmentation method CutLER [60], and Unsupervised Semantic Segmentation method STEGO [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.826 + ], + "angle": 0, + "content": "CutLER [60] exploits self-supervised learning models like DINO [8] to 'discover' objects and train a state-of-the-art detection and segmentation model using a cut-and-learn pipeline. It first uses MaskCut to extract multiple initial masks from DINO [8] features. MaskCut first generates a patch-wise affinity matrix \\( W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|K_j\\|_2} \\) using the \"key\" features \\( K_i \\) for patch \\( i \\) from DINO's last attention layer. Subsequently, the cut-based clustering method Normalized Cut [45] is employed on the affinity matrix by finding the eigenvector \\( x \\) that corresponds to the second smallest eigenvalue. A foreground instance mask \\( M^s \\) is derived through bi-partitioning of the vector \\( x \\), enabling segmentation of individual objects in the image. For multi-instance segmentation, MaskCut iteratively refines the affinity matrix by masking out already segmented objects, allowing for subsequent extractions" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.838, + 0.468, + 0.875 + ], + "angle": 0, + "content": "\\[\nW _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "and repeating above steps by \\(N\\) times. CutLER then refines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "detection and segmentation through a loss-dropping strategy and iterative self-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.289 + ], + "angle": 0, + "content": "STEGO [24] harnesses the semantically rich feature correlations produced by unsupervised methods like DINO [8] for segmentation. It trains a segmentation head to refine these correlations within an image, with its K-Nearest Neighbors (KNNs), and across randomly chosen images. Specifically, STEGO distills DINO's unsupervised features into distinct semantic labels by optimizing a correspondence loss. This loss function measures the feature correspondences \\( F^{SC} \\) between image feature pairs generated by DINO and the feature correspondence \\( S_{hwij} \\) derived from a trainable, lightweight segmentation head [24]:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.299, + 0.892, + 0.333 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o r r}} (x, y, b) = - \\sum_ {h w i j} \\left(F _ {h w i j} ^ {S C} - b\\right) \\max \\left(S _ {h w i j}, 0\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.82, + 0.36 + ], + "angle": 0, + "content": "3.2. Unsupervised Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Although CutLER [60] provides high-quality instance segmentation masks without human annotations, the predicted masks are class-agnostic, and thus do not include semantic labels for each instance. Our method addresses this issue by grouping the detected instances with a clustering method. In this way, instances assigned to the same cluster are associated with identical or closely related semantic information, while instances residing in separate clusters exhibit semantic dissimilarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Pseudo Semantic Labels. To train a detection and instance segmentation model, we vector quantize the model targets (pseudo semantic labels) by clustering the instance-level features of the entire dataset, under constraints derived from self-supervision. Specifically, our approach starts with the generation of instance segmentation masks using MaskCut [60]. Subsequently, we utilize the efficient \\( K \\)-Means clustering method as implemented in USL [57] to cluster all segmentation masks into semantically meaningful clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.731 + ], + "angle": 0, + "content": "We employ \\(K\\)-Means clustering to partition \\(n\\) instances into \\(C(\\leq n)\\) clusters, where each cluster is represented by its centroid \\(c\\) [22, 41]. Each instance is assigned to the cluster with the nearest centroid. Formally, we conduct a \\(C\\)-way node partitioning, denoted as \\(\\mathcal{S} = S_1, S_2, \\ldots, S_C\\), that minimizes the within-cluster sum of squares [36]:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.742, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} \\sum_ {V \\in S _ {i}} | V - c _ {i} | ^ {2} = \\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} | S _ {i} | \\operatorname {V a r} \\left(S _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "This optimization process is carried out iteratively using the EM algorithm [42], starting from selecting random samples as initial centroids. As a result, this process assigns pseudo semantic labels, denoted as \\( y_{i} \\), to each instance \\( i \\) with \\( y_{i} \\) falling within the range of \\( [1,C] \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The resulting semantic labels serve multiple purposes: 1) Semantic-aware copy-paste augmentation, which signifi-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22746" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.852, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.192, + 0.894, + 0.222 + ], + "angle": 0, + "content": "Figure 2. Overview of the training and inference pipeline for the proposed Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework." + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.228, + 0.832, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.893, + 0.378 + ], + "angle": 0, + "content": "Figure 3. Pipeline overview for generating masks and their semantically meaningful pseudo labels in semantic-aware instance segmentation. We first use MaskCut to generate class-agnostic instance masks, which are then grouped into semantically meaningful clusters. These pseudo semantic labels are used for training a semantic-aware instance segmentor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.478 + ], + "angle": 0, + "content": "cantly improves CutLER's capability to differentiate overlapping instances, especially when they share similar semantic information. 2) Training instance segmentation models: They serve as pseudo ground-truth labels for training a non-agnostic instance segmentor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Semantic-aware Copy-Paste Augmentation. In cluttered natural scenes, previous unsupervised instance segmentation model often fail to distinguish instances from the same semantic class. This results in multiple instances being captured in the same mask. To distinguish multiple overlapping objects and small objects in existing unsupervised detectors, we employ semantic-aware copy-paste augmentation, which includes several steps:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.706 + ], + "angle": 0, + "content": "1) We begin by randomly selecting two instances, denoted as \\( I_{1} \\) and \\( I_{2} \\), both belonging to the same pseudocategory (or group/cluster). 2) One of these instances undergoes a transformation function \\( \\mathcal{T} \\), which randomly reizes and shifts the associated pseudo-masks. 3) The resized instance is then pasted onto another image, creating synthetic overlapping cases using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.712, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\nI _ {3} = I _ {1} \\cdot (1 - \\mathcal {T} (M _ {c})) + I _ {2} \\cdot \\mathcal {T} (M _ {c}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.374, + 0.749 + ], + "angle": 0, + "content": "where \\(\\cdot\\) denotes element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Learning Unsupervised Instance Segmentor. Traditionally, unsupervised segmentation community focused primarily on class-agnostic instance segmentation [58, 60, 61], whose outputs lack class labels. However, by incorporating clustering information obtained from pseudo-labels on ImageNet, as discussed above, our method allows the model to predict not only the location and segmentation of an object but also its pseudo semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As observed by [60], \"ground-truth\" masks may miss instances. However, a standard detection loss penalizes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.525 + ], + "angle": 0, + "content": "predicted regions \\( r_i \\) that do not overlap with the \"ground-truth\". Therefore, following [60], we drop the loss for each predicted region \\( r_i \\) that has a maximum overlap of \\( \\tau^{\\mathrm{IoU}} \\) with any of the 'ground-truth' instances: \\( \\mathcal{L}_{\\mathrm{drop}}(r_i) = \\mathbb{1}(\\mathrm{IoU}_i^{\\max} > \\tau^{\\mathrm{IoU}})\\mathcal{L}_{\\mathrm{vanilla}}(r_i) \\), where \\( \\mathrm{IoU}_i^{\\max} \\) denotes the maximum IoU with all 'ground-truth' for \\( r_i \\) and \\( \\mathcal{L}_{\\mathrm{vanilla}} \\) is the vanilla loss function of detectors. \\( \\mathcal{L}_{\\mathrm{drop}} \\) encourages the exploration of image regions missed in the \"ground-truth\"." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.881, + 0.547 + ], + "angle": 0, + "content": "3.3. Unsupervised Universal Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Pseudo Labels for Panoptic Segmentation. For each pixel \\((i,j)\\) in the image, we vector quantize pixels with different semantics or instance membership, generating pseudo semantic labels for panoptic segmentation. We assign each pixel a semantic label based on \"stuff\" or \"things\" identity. This results in an instance label \\((I(i,j))\\) for \"things\" or a semantic label \\((S(i,j))\\) for \"stuff\". The critical challenge in this process is distinguishing between pixels associated with \"things\" (countable, often foreground) and \"stuff\" (uncountable, usually background) [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To resolve this problem, our method unfolds in three steps: 1) Semantic Labeling for \"Things\": Utilizing the class-agnostic instance segmentation capabilities of CutLER [60], we first identify \"things\" within an image, generating class-agnostic instance masks. These masks then undergo deep clustering to attribute a semantic label \\( I_C(i,j) \\) to each instance, detailed in Sec. 3.2. 2) Semantic Labeling for \"Stuff\": For \"stuff\" pixels, we deploy the unsupervised semantic segmentation model STEGO [24], which distills DINO's unsupervised features into discrete semantic labels, as outlined in 3.1. This step assigns a \"stuff\" semantic label to all pixels, including those of \"Things\" identified earlier. 3) Integrating Labels for \"Things\" and \"Stuff\". We deter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22747" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.087, + 0.885, + 0.257 + ], + "angle": 0, + "content": "
Task → Datasets → Metric →Agn Instance Seg.Instance Seg.Semantic Seg.Panoptic Seg.
COCOCOCOVOCUVOCOCOCOCOCityscapes
APboxAPbox50APbox50ARbox100APbox50ARbox100APbox50ARbox100PixelAccmIoUPQSQRQPQSQRQ
FreeSOLO [58]9.64.2-------------
TokenCut [61]5.83.2-------------
DINO [8]--------30.59.6-----
PiCIE + H [14]--------48.113.8-----
STEGO [24]--------56.928.2-----
CutLER [59]21.912.3-------------
CutLER+--9.010.326.827.210.611.8-------
CutLER+STEGO----------12.464.915.512.436.115.2
U2Seg22.813.011.821.531.048.110.825.063.930.216.171.119.917.652.721.7
vs.prev.SOTA+0.9+0.7+2.8+11.2+4.2+20.9+0.2+13.2+7.0+2.0+3.7+6.2+4.4+5.2+16.6+6.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.26, + 0.893, + 0.303 + ], + "angle": 0, + "content": "Table 1. With a unified framework, U2Seg outperforms previous state-of-the-art methods tailored for individual tasks across various datasets, including CutLER for unsupervised instance segmentation, STEGO for unsupervised semantic segmentation, and CutLER+STEGO for unsupervised panoptic segmentation. \"Agn Instance Seg\" denotes class-agnostic instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.358 + ], + "angle": 0, + "content": "mine a pixel's classification as \"things\" or \"stuff\" using the following logic:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.372, + 0.47, + 0.428 + ], + "angle": 0, + "content": "\\[\nI (i, j) = \\left\\{ \\begin{array}{l l} I _ {C} (i, j), & \\text {i f} I _ {C} (i, j) \\neq 0 \\\\ S _ {S} (i, j), & \\text {i f} I _ {C} (i, j) = 0 \\& S _ {S} (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.501 + ], + "angle": 0, + "content": "This process merges the semantic labels, assigning priority to \"things\" labels over \"stuff\" where applicable. We then train a universal segmentation model on these pseudo-labels for instance, semantic and panoptic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Learning Unsupervised Universal Image Segmentor. After we obtain the pseudo labels for panoptic segmentation, following [33], we construct an unsupervised universal image segmentation model, that has two branches: instance segmentation branch and semantic segmentation branch, to address corresponding segmentation tasks. The model is trained jointly for both branches, employing the following loss function: \\(\\mathcal{L} = \\lambda_i(\\mathcal{L}_c + \\mathcal{L}_b + \\mathcal{L}_m) + \\lambda_s\\mathcal{L}_s\\), where \\(\\mathcal{L}_c\\) represents the classification loss, \\(\\mathcal{L}_b\\) is the detection loss, \\(\\mathcal{L}_m\\) is the segmentation loss, and \\(\\mathcal{L}_s\\) signifies the semantic loss. The \\(\\mathcal{L}_s\\) is computed as a per-pixel cross-entropy loss between the predicted and ground-truth labels. The hyperparameters \\(\\lambda_{i}\\) and \\(\\lambda_{s}\\) balance these two parts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.714, + 0.312, + 0.731 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.268, + 0.758 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Training Data. Our model is trained on 1.3M unlabeled images from ImageNet [17] and is evaluated directly across various benchmarks, unless otherwise noted. For unsupervised semantic segmentation comparisons with STEGO [24], we additionally fine-tune our model using MSCOCO's unlabeled images, following STEGO [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Test Data. For unsupervised instance segmentation, we test our model on COCO val2017, PASCAL VOC val2012 [21] and UVO val [53]. For unsupervised" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.328, + 0.892, + 0.358 + ], + "angle": 0, + "content": "panoptic segmentation, we evaluate our model on COCO val2017 and Cityscapes val [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.893, + 0.511 + ], + "angle": 0, + "content": "Evaluation Metrics. We use AP, \\(\\mathrm{AP}_{50}\\), \\(\\mathrm{AP}_{75}\\) and \\(\\mathrm{AR}_{100}\\) to evaluate the unsupervised instance segmentation; PixelAcc and mIoU for unsupervised semantic segmentation; PQ, RQ, SQ for unsupervised universal image segmentation. After predicting the instance with its semantic labels, we use Hungarian matching to map the semantic labels to class names in the real dataset (details in B). It evaluates the consistency of the predicted semantic segments with the ground truth labels, remaining unaffected by any permutations in the predicted class labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.894, + 0.71 + ], + "angle": 0, + "content": "Implementation Details. Following [32], we employ Panoptic Cascade Mask R-CNN [4, 32] with a ResNet50 backbone [25]. Following CutLER's training recipe [60], our model, initialized with DINO pre-trained weights, is trained on unlabeled ImageNet for two epochs. It starts with an initial learning rate of 0.01, which then decreases after the first epoch to \\(5 \\times 10^{-5}\\), with a batch size of 16 for all models. For unsupervised panoptic segmentation, we maintain the same training schedule as unsupervised instance segmentation for zero-shot evaluation. In non-zero-shot scenarios, the models undergo training on a combination of unlabeled COCO and ImageNet datasets, beginning with a learning rate of 0.01 over 90k steps." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.723, + 0.881, + 0.74 + ], + "angle": 0, + "content": "4.2. Unsupervised Universal Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.853 + ], + "angle": 0, + "content": "To the best of our knowledge, U2Seg represents the first effort in addressing unsupervised semantic-aware instance, semantic and panoptic segmentation, all unified under a single framework. Due to the absence of benchmarks for unsupervised semantic-aware instance segmentation and panoptic segmentation, we establish comprehensive benchmarks and baselines for both tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 1, we demonstrate that U2Seg, utilizing a unified framework, significantly outperforms all previous approaches across various benchmarks and datasets. For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22748" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.146 + ], + "angle": 0, + "content": "
Metric\\( AP^{box} \\)\\( AP^{box}_{50} \\)\\( AP^{box}_{75} \\)\\( AR^{box}_{100} \\)\\( AP^{mask} \\)\\( AP^{mask}_{50} \\)\\( AP^{mask}_{75} \\)\\( AR^{mask}_{100} \\)
CutLER+5.99.06.110.35.38.65.59.3
U2Seg7.311.87.521.56.411.26.418.5
Δ+1.4+2.8+1.4+11.2+1.1+2.6+0.9+9.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.15, + 0.468, + 0.204 + ], + "angle": 0, + "content": "Table 2. The results for zero-shot unsupervised object detection and instance segmentation on COCO val2017. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.214, + 0.465, + 0.274 + ], + "angle": 0, + "content": "
MethodsAPboxAP50AP75AR100
CutLER+17.126.818.127.2
U2Seg19.031.019.548.1
Δ+1.9+4.2+1.4+20.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.278, + 0.468, + 0.333 + ], + "angle": 0, + "content": "Table 3. The results for zero-shot unsupervised object detection on PASCAL VOC va12012. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.343, + 0.465, + 0.402 + ], + "angle": 0, + "content": "
Metric\\(AP^{box}\\)\\(AP^{box}\\)\\(50\\)\\(AR^{box}\\)\\(100\\)\\(AP^{mask}\\)\\(AP^{mask}\\)\\(50\\)\\(AR^{mask}\\)\\(100\\)
CutLER+6.310.611.86.09.010.4
U2Seg6.810.825.06.29.521.0
Δ+0.5+0.2+13.2+0.2+0.5+10.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.406, + 0.468, + 0.462 + ], + "angle": 0, + "content": "Table 4. The results for zero-shot unsupervised object detection and instance segmentation on UVO va1. The model is trained on ImageNet with a cluster number of 800. We compare with CutLER+, a combination of CutLER and offline clustering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.468, + 0.76 + ], + "angle": 0, + "content": "class-agnostic unsupervised instance segmentation, our method achieves an increase of \\(+0.9\\) in \\(\\mathrm{AP^{box}}\\) compared to CutLER [60]. This improvement is largely attributed to our novel semantic-aware copy-paste augmentation, as detailed in Sec. 3.2. For unsupervised semantic-aware instance segmentation, we benchmark against the advanced baseline CutLER+, derived from CutLER, and record a substantial gain of over \\(11.2\\%\\) in AR. A more comprehensive analysis of these results is provided in Sec. 4.3. For unsupervised semantic segmentation, our approach surpasses the state-of-the-art STEGO with impressive margins of \\(+7.0\\) in PixelAcc and \\(+2.0\\) in mIoU. Lastly, for unsupervised panoptic segmentation, we compare against the strong baseline of CutLER+STEGO, a hybrid of CutLER+ and STEGO, and observe performance gains of over \\(6.2\\%\\) in SQ on MSCOCO and a notable \\(16.6\\%\\) improvement in SQ on Cityscapes. Further comparisons and discussions on this task are elaborated in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.396, + 0.787 + ], + "angle": 0, + "content": "4.3. Unsupervised Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We performed extensive experiments for zero-shot unsupervised instance segmentation. Given that prior methods [51, 58, 60, 61] are limited to class-agnostic instance segmentation, we developed CutLER+, a strong baseline for unsupervised semantic-aware instance segmentation, building upon the current state-of-the-art CutLER [60]. CutLER+ operates in two steps: it first uses the pre-trained" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.891, + 0.204 + ], + "angle": 0, + "content": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN15.746.619.8
non zero-shot methods
CutLER+STEGOCOCO12.436.115.2
U2SegCOCO15.451.519.0
U2SegCOCO+IN17.652.721.7
Δ+5.2+16.6+6.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.209, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Table 5. Unsupervised Panoptic image segmentation on Cityscapes val. We show PQ, SQ and RQ on zero-shot and nonzero shot settings with the cluster number of 800. We compare with CutLER+STEGO, a combination of CutLER+ and STEGO." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.273, + 0.888, + 0.389 + ], + "angle": 0, + "content": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN11.160.113.7
non zero-shot methods
CutLER+STEGOCOCO12.464.915.5
U2SegCOCO15.366.519.1
U2SegCOCO+IN16.171.119.9
Δ+3.7+6.2+4.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.393, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Table 6. Unsupervised Panoptic image segmentation on COCOval2017. We show PQ, SQ and RQ on zero-shot and non-zero shot settings. We use CutLER+STEGO, a combination of CutLER+ and STEGO, as a strong baseline." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.891, + 0.52 + ], + "angle": 0, + "content": "CutLER to generate class-agnostic instance masks, and subsequently assigns semantic labels to all instance masks through offline clustering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Tab. 2 demonstrates that U2Seg markedly improves performance in both unsupervised object detection and instance segmentation on MSCOCO, delivering a +2.8 boost in \\(\\mathrm{AP}_{50}^{\\mathrm{box}}\\) and a +2.6 rise in \\(\\mathrm{AP}_{50}^{\\mathrm{mask}}\\) over CutLER+. Additionally, our method sees a substantial increase of approximately +10.0 in \\(\\mathrm{AR}_{100}\\). Results on PASCAL VOC val2012 and UVO val are detailed in Tab. 3 and Tab. 4, respectively. Notably, we achieve gains exceeding +20% in AR for PASCAL VOC and +10% for UVO." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.822, + 0.683 + ], + "angle": 0, + "content": "4.4. Unsupervised Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.869 + ], + "angle": 0, + "content": "For unsupervised panoptic/universal image segmentation, our experiments span two scenarios. In the zero-shot setting, the model is trained exclusively on unlabeled ImageNet images. For non-zero-shot (in-domain) scenarios, we train on unlabeled COCO images or a mix of COCO and ImageNet. With no existing benchmarks for unsupervised panoptic segmentation, we establish a new baseline by integrating the state-of-the-art unsupervised semantic segmentation from STEGO [24] with semantic-aware instance segmentation from CutLER+ (discussed in Sec. 4.3), which are then merged to create panoptic/universal segmentation outcomes, referred to as CutLER+STEGO." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Tab. 6 presents the PQ, SQ, and RQ scores of U2Seg on COCO val2017. U2Seg surpasses the strong baseline" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22749" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.096, + 0.177, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.097, + 0.252, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.097, + 0.355, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.097, + 0.45, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.097, + 0.56, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.097, + 0.638, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.097, + 0.726, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.097, + 0.789, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.097, + 0.876, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.187, + 0.892, + 0.229 + ], + "angle": 0, + "content": "Figure 4. Universal image segmentation visualization in COCO val2017. We present the results with cluster IDs predicted by U2Seg, categorizing athletes playing hockey (left columns) as “139”, those playing badminton (middle columns) as “52” and gentlemen (right columns) as “132”. After Hungarian matching, the IDs are automatically matched to the category “person” for quantitative evaluations." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.251, + 0.457, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.459, + 0.171, + 0.471 + ], + "angle": 0, + "content": "Raw Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.459, + 0.247, + 0.471 + ], + "angle": 0, + "content": "U2Seg" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.459, + 0.361, + 0.47 + ], + "angle": 0, + "content": "Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.459, + 0.456, + 0.47 + ], + "angle": 0, + "content": "Pseudo-Labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Figure 5. Visualizations of U2Seg's unsupervised Panoptic segmentation results on COCO val2017 (after Hungarian matching). The pseudo label is the naive combination of previous state-of-the-art instance segmentation, i.e. CutLER [60], and semantic segmentation, i.e., STEGO [24], results." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.566, + 0.457, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.78, + 0.171, + 0.792 + ], + "angle": 0, + "content": "Raw Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.781, + 0.247, + 0.792 + ], + "angle": 0, + "content": "U2Seg" + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.781, + 0.361, + 0.792 + ], + "angle": 0, + "content": "Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.781, + 0.456, + 0.792 + ], + "angle": 0, + "content": "Pseudo-Labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.795, + 0.468, + 0.822 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of U2Seg's Panoptic image segmentation results on Cityscapes val (after Hungarian matching)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "CutLER+STEGO with a +3.5 improvement in PQ and an increase of over +4.0 in RQ. Qualitative results of U2Seg's performance is provided in Fig. 4, with the predicted seman" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.361 + ], + "angle": 0, + "content": "tic labels visualized. The qualitative results suggest that an over-clustering strategy in pseudo-label generation, e.g. setting the number of clusters to 300 or 800, leads to highly granular model predictions. For instance, as in Fig. 4, the model distinctly categorizes hockey players as “139”, badminton players as “52”, and gentlemen in suits as “132”, showcasing its potent discriminative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.893, + 0.605 + ], + "angle": 0, + "content": "To quantitatively measure the quality of segmentation masks and their corresponding semantic labels, we use Hungarian matching (detailed in Appendix B) to align semantic labels with the category names from the test dataset; for instance, all three sub-clusters depicted in Fig. 4 are assigned to the \"person\" category. The qualitative outcomes post-Hungarian matching are shown in Fig. 5, where our model demonstrates superior panoptic segmentation mask quality. For instance, while the baseline tends to segment parts separately (as seen with the man's head and torso being treated as separate entities in the third row), our model correctly identifies them as parts of a single object. This level of recognition is also evident with the \"trunks of the motorcycle\" example in the second row. For additional results, please see Appendix D. We also present results of the more challenging dataset Cityscapes in Tab. 5 and Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.677, + 0.635 + ], + "angle": 0, + "content": "4.5. Efficient Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Specifically, for object detection and instance segmentation, we employ our unsupervised instance segmentation model, with cluster count set to 300, to initialize the model weights. We adopt the recipe from [56, 59] for model fine-tuning across various annotation splits. For label-efficient panoptic segmentation, we fine-tune the model initialized with our zero-shot unsupervised framework on the same data splits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The results are depicted in Fig. 7, where our model's instance segmentation performance is benchmarked against MoCo-V2, DETReg, and CutLER. Our model consistently surpasses the state-of-the-art with consistent gains in both \\(\\mathrm{AP^{box}}\\) and \\(\\mathrm{AP^{mask}}\\). In scenarios with panoptic image segmentation as the downstream task, we contrast our results with MoCo-V2 and CutLER in terms of PQ, SQ, and RQ metrics. The results illustrate a remarkable improvement, effectively doubling the performance boost from MoCo-V2 to CutLER, especially in few-shot contexts with limited an-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22750" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.087, + 0.253, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.087, + 0.408, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.087, + 0.563, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.087, + 0.718, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.087, + 0.871, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 7. We evaluate the label-efficient learning performance on 3 different tasks: object detection (the left), instance segmentation (the second left) and panoptic image segmentation (the last three)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.281, + 0.268, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.281, + 0.442, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.525 + ], + "angle": 0, + "content": "Figure 8. U2Seg learns features that are more discriminative than those learned by CutLER. The t-SNE [47] visualization of the features from the model's FC layer. We color-code each dot based on its ground-truth category." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.536, + 0.465, + 0.604 + ], + "angle": 0, + "content": "
# ClusterCOCOUVOVOC
APbox50ARbox100APbox50ARbox100APbox50ARbox100
3009.320.19.822.629.645.7
80011.821.510.825.031.048.0
291113.322.115.125.831.648.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Table 7. Over-clustering can improve the model performance. We show results on different datasets for the unsupervised object detection using different cluster numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.717 + ], + "angle": 0, + "content": "notations (1% or 2% labeled samples). This highlights the practical value of our approach in real-world unsupervised learning applications, where annotations are often scarce." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.763 + ], + "angle": 0, + "content": "We attribute the performance gains primarily to the discriminative features our model learns, as in Fig. 8, obtaining effective model initialization for few-shot learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.241, + 0.787 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.424, + 0.81 + ], + "angle": 0, + "content": "In this section, we conduct ablation study on U2Seg." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Numbers of clusters. The choice of cluster quantity significantly affects the model's representation granularity. Our ablation study on various cluster counts, as detailed in Tab. 7, reveals their impact on model performance. Overclustering generally leads to a finer level of detail, prompting the model to learn more discriminative features." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.279, + 0.685, + 0.368 + ], + "angle": 0, + "content": "
conf# matchedAPbox50ARbox100
0.910910.913.1
0.722511.618.0
0.628211.819.7
0.438911.821.5
0.251311.321.8
0.07188.618.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.371, + 0.685, + 0.385 + ], + "angle": 0, + "content": "(a) Conf's effect on accuracy." + }, + { + "type": "table", + "bbox": [ + 0.716, + 0.279, + 0.892, + 0.369 + ], + "angle": 0, + "content": "
IoU# matchedAPbox50ARbox100
0.929510.819.7
0.834811.420.7
0.441411.521.6
0.245011.521.1
0.04949.217.7
0.638911.821.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.716, + 0.371, + 0.892, + 0.385 + ], + "angle": 0, + "content": "(b) IoU's effect on accuracy." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Table 8. Impact of Confidence and IoU on Hungarian Matching Performance: The left table illustrates the outcomes at a fixed IoU of 0.6 while varying the confidence scores. Conversely, the right table displays the results with a constant confidence of 0.4, altering the IoU values. The cluster number is 800." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Hungarian matching. As our trained model could predict the instance with corresponding semantic labels, we are able to go further beyond unsupervised class-agnostic instance segmentation. To quantitatively evaluate the performance, Hungarian matching is employed to match the predicted semantic labels to the ground-truth dataset categories. See Appendix B for details of the adopted Hungarian matching used in our evaluation. As shown in Tab. 8, the two parameters conf threshold and IoU threshold also affect the precision and recall." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.608, + 0.666 + ], + "angle": 0, + "content": "5. Summary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.853 + ], + "angle": 0, + "content": "We present U2Seg, a novel Unsupervised Universal Image Segmentation model, adept at performing unsupervised instance, semantic, and panoptic segmentation tasks within a unified framework. Evaluated on extensive benchmarks, U2Seg consistently outperforms previous state-of-the-art methods designed for individual tasks. Additionally, U2Seg achieves the new state-of-the-art for label-efficient panoptic segmentation and instance segmentation. We anticipate that U2Seg, free from the constraints of human annotations, will demonstrate enhanced performance when scaled up with more training data, representing an important direction for our future research." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgement: Trevor Darrell, Dantong Niu and Xudong Wang were funded by DoD including DARPA LwLL and the Berkeley AI Research (BAIR) Commons." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22751" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Edward H Adelson. On seeing stuff: the perception of materials by humans and machines. In Human Vision and Electronic Imaging, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] YM Asano, C Rupprecht, and A Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.468, + 0.243 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.246, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.303, + 0.468, + 0.343 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.346, + 0.468, + 0.386 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.389, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.446, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-DeepLab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[14] Jang Hyun Cho, U. Mall, K. Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. ArXiv, abs/2103.17070, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. CoRR, abs/1604.01685, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13693-13702, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[18] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[20] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[21] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[22] Edward W Forgy. Cluster analysis of multivariate data: efficiency versus interpretability of classifications. biometrics, 21:768-769, 1965." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[26] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22752" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[30] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9396-9405, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[33] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[36] Hans-Peter Kriegel, Erich Schubert, and Arthur Zimek. The (black) art of runtime evaluation: Are we comparing algorithms or implementations? Knowledge and Information Systems, 52(2):341-378, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[37] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. arXiv preprint arXiv:2108.07682, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[41] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[42] Geoffrey J McLachlan and Thriyambakam Krishnan. The EM algorithm and extensions. John Wiley & Sons, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[43] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[44] Sungwon Park, Sungwon Han, Sundong Kim, Danu Kim, Sungkyu Park, Seunghoon Hong, and Meeyoung Cha. Improving unsupervised image clustering with robust learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12278-12287, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[47] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[48] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[49] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. arxiv preprint arxiv:2102.06191, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[50] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[51] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[52] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[53] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, openworld segmentation. CoRR, abs/2104.04691, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[54] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10776-10785, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[55] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[56] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[57] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In European Conference on Computer Vision, pages 427-445. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[58] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[60] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22753" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[61] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[62] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3733-3742, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[63] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[64] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[65] Chengxu Zhuang, Alex Lin Zhai, Daniel Yamins, , et al. Local aggregation for unsupervised learning of visual embeddings. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.474 + ], + "angle": 0, + "content": "[66] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22754" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_origin.pdf b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bd2421af4060dd261099004e559c9c68b7dda3d --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/f08d4016-acc4-41d7-91e7-a46018d06d6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec026f3711ad8531706c20d73e8ec9254b0064aa29c82d12a0f52625d2db3d7d +size 6875227 diff --git a/2024/Unsupervised Universal Image Segmentation/full.md b/2024/Unsupervised Universal Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4415c6a9b400825c6b59447aba92d9d8d5e645e5 --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/full.md @@ -0,0 +1,355 @@ +# Unsupervised Universal Image Segmentation + +Dantong Niu*† Xudong Wang*† Xinyang Han* Long Lian Roei Herzig Trevor Darrell Berkeley AI Research, UC Berkeley + +Code: https://github.com/u2seg/U2Seg + +![](images/9f6c80cf4672c46b0a61ea77a23d3c8c725ee3dabd3b4a1b5d3bb276bc9e5dd2.jpg) +Figure 1. We present U2Seg, a unified framework for Unsupervised Universal image Segmentation that consistently outperforms previous state-of-the-art methods designed for individual tasks: CutLER [60] for unsupervised instance segmentation, STEGO [24] for unsupervised semantic segmentation, and the naive combination of CutLER and STEGO for unsupervised panoptic segmentation. We visualize instance segmentation results with "semantic label" + confidence score and semantic predictions with "semantic label". Zoom in for the best view. + +# Abstract + +Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized + +methods tailored to each task: $a + 2.6AP^{box}$ boost (vs. CutLER) in unsupervised instance segmentation on COCO and $a + 7.0$ PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOSTuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by $+5.0AP^{mask}$ when trained on a low-data regime, e.g., only $1\%$ COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation. + +# 1. Introduction + +The field of image segmentation has witnessed significant advancements in the recent years [4, 5, 12, 20, 26, 35, 38, 40]. Nonetheless, the effectiveness of these segmentation methods heavily depends on the availability of extensive densely human-labeled data for training these models, + +which is both labor-intensive and costly and thus less scalable. In this paper, our objective is to explore the extent to which unsupervised image segmentation can be achieved without relying on any human-generated labels. + +Several recent works such as CutLER [60] and STEGO [24] have emerged as promising approaches for unsupervised image segmentation. CutLER leverages the property of the self-supervised model DINO [8] to 'discover' objects without supervision, and learns a state-of-the-art localization model on pseudo instance segmentation masks produced by MaskCut [60] (based on Normalize Cuts [45]). Similarly leveraging DINO [8], STEGO [24] introduces a novel framework that distills unsupervised features into discrete semantic labels. This is achieved using a contrastive loss that encourages pixel features to form compact clusters while preserving their relationships across the corpora [24]. However, these methods have limitations: + +- The output of unsupervised instance segmentation methods such as CutLER [60] comprises class-agnostic segments for "things", ignoring the "stuff" categories that represent pixel semantics. Moreover, CutLER often treats several overlapping instances as one instance, especially when these instances belong to the same semantic class. +- On the other hand, unsupervised semantic segmentation methods such as STEGO [24] focus on the segmentation of semantically coherent regions, lacking the capability to distinguish between individual instances. +- Unsupervised panoptic segmentation has not been addressed. Supervised panoptic segmentation methods [12, 29, 31] predict both "stuff" and "things" classes simultaneously; to the best of our knowledge there has not been work on unsupervised panoptic segmentation heretofore. + +To address these limitations, we propose U2Seg, a novel Unsupervised Universal image Segmentation model. U2Seg offers comprehensive scene understanding—instance, semantic and panoptic—without relying on human annotations, segmenting semantically meaningful regions in the image as well as identifying and differentiating between individual instances within those regions. + +U2Seg is comprised of three steps. First, we create high-quality, discrete semantic labels for instance masks obtained from MaskCut and DINO, by clustering semantically similar instance masks into distinct fine-grained clusters, as in Sec. 3.2. Next, we amalgamate the semantically pseudolabeled "things" pixels (from the first step) with "stuff" pixels (from STEGO) to produce pseudo semantic labels for each pixel in the image. Lastly, a universal image segmentation model is trained using these pseudo-labels, resulting in a model capable of simultaneously predicting pixel-level (i.e., semantic and class-agnostic instance segmentation) and instance-level semantic labels, detailed in Sec. 3.3. + +Despite the inherent noise in these pseudo-labels, self-training the model with them yields substantial performance + +gains over specialized methods tailored to each task: U2Seg achieves a $+2.6$ $\mathbf{AP}^{\mathrm{box}}$ boost (vs. CutLER) in unsupervised instance segmentation on COCO and a $+7.0$ PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation. We also find that the multi-task learning framework and learning unsupervised segmentor with semantic labels enable our model to generate a more discriminative feature space, which makes it a superior representation for downstream supervised detection and segmentation tasks. When trained on a low-data regime, such as $1\%$ COCO labels, U2Seg surpasses CutLER by $+5.0$ $\mathbf{AP}^{\mathrm{mask}}$ . + +Contributions. Our main contribution is the first universal unsupervised image segmentation model that can tackle unsupervised semantic-aware instance, semantic and panoptic segmentation tasks using a unified framework. We establish a suite of benchmarks on unsupervised semantic-aware instance segmentation and panoptic segmentation, areas previously unexplored. Despite using a single framework, we demonstrate that U2Seg surpasses previous methods specialized for each task across all experimented benchmarks (instance, semantic, panoptic, etc.) and datasets (COCO, Cityscapes, UVO, VOC, etc.). + +# 2. Related Work + +Self-supervised Representation Learning focuses on feature learning from a large amount of unlabeled data without using human-made labels. Contrastive Learning-Based Methods [9, 27, 43, 62] learn representation by comparing similar instances or different versions of a single instance while separating dissimilar ones. Similarity-Based Self-Supervised Learning [10, 23] mainly reduces differences between different augmented versions of the same instance. Clustering-Based Feature Learning [2, 7, 55, 63, 65] finds natural data groups in the hidden space. Masked Autoencoders [3, 18, 28] learn by masking and then reconstructing masked parts of the image. + +Unsupervised Object Detection and Instance Segmentation. DINO [8] shows that self-supervised learning (SSL) Vision Transformers (ViT) [19] can reveal hidden semantic segmentation in images, which is not obvious in supervised counterparts [8, 66]. Extending this, LOST [46], TokenCut [61] and MaskDistill [50] use DINO's patch features to identify main objects in images. FreeSOLO [58] performs unsupervised class-agnostic instance segmentation by creating coarse masks first, which are later improved through self-training. Meanwhile, CutLER [59] introduces the MaskCut method, which aims to identify multiple instances in a single image. Yet, MaskCut frequently consolidates overlapping instances into a single segment and lacks the capability to assign semantic labels to each instance. + +Unsupervised Semantic Segmentation. IIC [30] maxi + +mizes mutual information for clustering, while PiCIE [14] uses invariance to photometric effects and equivariance to geometric transformations for segmentation. MaskContrast [49] learns unsupervised semantic segmentation by contrasting features within saliency masks. STEGO [24] refines pretrained SSL visual features to distill correspondence information embedded within these features, thereby fostering discrete semantic clusters. + +Universal Segmentation has been introduced to deliver instance, semantic and panoptic segmentation tasks using a unified architecture [6, 11-13, 29, 33, 34, 37, 52, 64]. In this work, we propose U2Seg to tackle this challenging task without relying on human-annotated data. + +Unsupervised Image Classification methods mainly focus on providing a semantic label for each query image that can be mapped to ground truth classes by hungarian matching. SCAN [48] proposes a three-stage pipeline that includes representation learning, deep clustering, and self-labeling. NNM [16] enhances SCAN by incorporating local and global nearest neighbor matching. RUC [44] further improves SCAN using a robust loss as training objective. However, these approaches only provide one classification prediction per image, whereas our method provides classification per-instance for instance segmentation and per-pixel for semantic segmentation. + +# 3. Unsupervised Universal Segmentation + +# 3.1. Preliminaries + +We first explain the previous Unsupervised Instance Segmentation method CutLER [60], and Unsupervised Semantic Segmentation method STEGO [24]. + +CutLER [60] exploits self-supervised learning models like DINO [8] to 'discover' objects and train a state-of-the-art detection and segmentation model using a cut-and-learn pipeline. It first uses MaskCut to extract multiple initial masks from DINO [8] features. MaskCut first generates a patch-wise affinity matrix $W_{ij} = \frac{K_i K_j}{\|K_i\|_2\|K_j\|_2}$ using the "key" features $K_i$ for patch $i$ from DINO's last attention layer. Subsequently, the cut-based clustering method Normalized Cut [45] is employed on the affinity matrix by finding the eigenvector $x$ that corresponds to the second smallest eigenvalue. A foreground instance mask $M^s$ is derived through bi-partitioning of the vector $x$ , enabling segmentation of individual objects in the image. For multi-instance segmentation, MaskCut iteratively refines the affinity matrix by masking out already segmented objects, allowing for subsequent extractions + +$$ +W _ {i j} ^ {t} = \frac {\left(K _ {i} \prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\right) \left(K _ {j} \prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\right)}{\| K _ {i} \| _ {2} \| K _ {j} \| _ {2}} \tag {1} +$$ + +and repeating above steps by $N$ times. CutLER then refines + +detection and segmentation through a loss-dropping strategy and iterative self-training. + +STEGO [24] harnesses the semantically rich feature correlations produced by unsupervised methods like DINO [8] for segmentation. It trains a segmentation head to refine these correlations within an image, with its K-Nearest Neighbors (KNNs), and across randomly chosen images. Specifically, STEGO distills DINO's unsupervised features into distinct semantic labels by optimizing a correspondence loss. This loss function measures the feature correspondences $F^{SC}$ between image feature pairs generated by DINO and the feature correspondence $S_{hwij}$ derived from a trainable, lightweight segmentation head [24]: + +$$ +L _ {\text {c o r r}} (x, y, b) = - \sum_ {h w i j} \left(F _ {h w i j} ^ {S C} - b\right) \max \left(S _ {h w i j}, 0\right) \tag {2} +$$ + +# 3.2. Unsupervised Instance Segmentation + +Although CutLER [60] provides high-quality instance segmentation masks without human annotations, the predicted masks are class-agnostic, and thus do not include semantic labels for each instance. Our method addresses this issue by grouping the detected instances with a clustering method. In this way, instances assigned to the same cluster are associated with identical or closely related semantic information, while instances residing in separate clusters exhibit semantic dissimilarity. + +Pseudo Semantic Labels. To train a detection and instance segmentation model, we vector quantize the model targets (pseudo semantic labels) by clustering the instance-level features of the entire dataset, under constraints derived from self-supervision. Specifically, our approach starts with the generation of instance segmentation masks using MaskCut [60]. Subsequently, we utilize the efficient $K$ -Means clustering method as implemented in USL [57] to cluster all segmentation masks into semantically meaningful clusters. + +We employ $K$ -Means clustering to partition $n$ instances into $C(\leq n)$ clusters, where each cluster is represented by its centroid $c$ [22, 41]. Each instance is assigned to the cluster with the nearest centroid. Formally, we conduct a $C$ -way node partitioning, denoted as $\mathcal{S} = S_1, S_2, \ldots, S_C$ , that minimizes the within-cluster sum of squares [36]: + +$$ +\min _ {\mathcal {S}} \sum_ {i = 1} ^ {C} \sum_ {V \in S _ {i}} | V - c _ {i} | ^ {2} = \min _ {\mathcal {S}} \sum_ {i = 1} ^ {C} | S _ {i} | \operatorname {V a r} \left(S _ {i}\right) \tag {3} +$$ + +This optimization process is carried out iteratively using the EM algorithm [42], starting from selecting random samples as initial centroids. As a result, this process assigns pseudo semantic labels, denoted as $y_{i}$ , to each instance $i$ with $y_{i}$ falling within the range of $[1,C]$ . + +The resulting semantic labels serve multiple purposes: 1) Semantic-aware copy-paste augmentation, which signifi- + +![](images/0cdd369e78f9e985922848723320a3606944d07490bdc0ae7ce4b47a1bee4c38.jpg) +Figure 2. Overview of the training and inference pipeline for the proposed Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework. + +![](images/b67dbb46ef3a9411dfb950452309367d0c4ad6b01b5b7b1920e958b6f2d903f7.jpg) +Figure 3. Pipeline overview for generating masks and their semantically meaningful pseudo labels in semantic-aware instance segmentation. We first use MaskCut to generate class-agnostic instance masks, which are then grouped into semantically meaningful clusters. These pseudo semantic labels are used for training a semantic-aware instance segmentor. + +cantly improves CutLER's capability to differentiate overlapping instances, especially when they share similar semantic information. 2) Training instance segmentation models: They serve as pseudo ground-truth labels for training a non-agnostic instance segmentor. + +Semantic-aware Copy-Paste Augmentation. In cluttered natural scenes, previous unsupervised instance segmentation model often fail to distinguish instances from the same semantic class. This results in multiple instances being captured in the same mask. To distinguish multiple overlapping objects and small objects in existing unsupervised detectors, we employ semantic-aware copy-paste augmentation, which includes several steps: + +1) We begin by randomly selecting two instances, denoted as $I_{1}$ and $I_{2}$ , both belonging to the same pseudocategory (or group/cluster). 2) One of these instances undergoes a transformation function $\mathcal{T}$ , which randomly reizes and shifts the associated pseudo-masks. 3) The resized instance is then pasted onto another image, creating synthetic overlapping cases using the following equation: + +$$ +I _ {3} = I _ {1} \cdot (1 - \mathcal {T} (M _ {c})) + I _ {2} \cdot \mathcal {T} (M _ {c}) \tag {4} +$$ + +where $\cdot$ denotes element-wise multiplication. + +Learning Unsupervised Instance Segmentor. Traditionally, unsupervised segmentation community focused primarily on class-agnostic instance segmentation [58, 60, 61], whose outputs lack class labels. However, by incorporating clustering information obtained from pseudo-labels on ImageNet, as discussed above, our method allows the model to predict not only the location and segmentation of an object but also its pseudo semantic labels. + +As observed by [60], "ground-truth" masks may miss instances. However, a standard detection loss penalizes + +predicted regions $r_i$ that do not overlap with the "ground-truth". Therefore, following [60], we drop the loss for each predicted region $r_i$ that has a maximum overlap of $\tau^{\mathrm{IoU}}$ with any of the 'ground-truth' instances: $\mathcal{L}_{\mathrm{drop}}(r_i) = \mathbb{1}(\mathrm{IoU}_i^{\max} > \tau^{\mathrm{IoU}})\mathcal{L}_{\mathrm{vanilla}}(r_i)$ , where $\mathrm{IoU}_i^{\max}$ denotes the maximum IoU with all 'ground-truth' for $r_i$ and $\mathcal{L}_{\mathrm{vanilla}}$ is the vanilla loss function of detectors. $\mathcal{L}_{\mathrm{drop}}$ encourages the exploration of image regions missed in the "ground-truth". + +# 3.3. Unsupervised Universal Image Segmentation + +Pseudo Labels for Panoptic Segmentation. For each pixel $(i,j)$ in the image, we vector quantize pixels with different semantics or instance membership, generating pseudo semantic labels for panoptic segmentation. We assign each pixel a semantic label based on "stuff" or "things" identity. This results in an instance label $(I(i,j))$ for "things" or a semantic label $(S(i,j))$ for "stuff". The critical challenge in this process is distinguishing between pixels associated with "things" (countable, often foreground) and "stuff" (uncountable, usually background) [1]. + +To resolve this problem, our method unfolds in three steps: 1) Semantic Labeling for "Things": Utilizing the class-agnostic instance segmentation capabilities of CutLER [60], we first identify "things" within an image, generating class-agnostic instance masks. These masks then undergo deep clustering to attribute a semantic label $I_C(i,j)$ to each instance, detailed in Sec. 3.2. 2) Semantic Labeling for "Stuff": For "stuff" pixels, we deploy the unsupervised semantic segmentation model STEGO [24], which distills DINO's unsupervised features into discrete semantic labels, as outlined in 3.1. This step assigns a "stuff" semantic label to all pixels, including those of "Things" identified earlier. 3) Integrating Labels for "Things" and "Stuff". We deter + +
Task → Datasets → Metric →Agn Instance Seg.Instance Seg.Semantic Seg.Panoptic Seg.
COCOCOCOVOCUVOCOCOCOCOCityscapes
APboxAPbox50APbox50ARbox100APbox50ARbox100APbox50ARbox100PixelAccmIoUPQSQRQPQSQRQ
FreeSOLO [58]9.64.2-------------
TokenCut [61]5.83.2-------------
DINO [8]--------30.59.6-----
PiCIE + H [14]--------48.113.8-----
STEGO [24]--------56.928.2-----
CutLER [59]21.912.3-------------
CutLER+--9.010.326.827.210.611.8-------
CutLER+STEGO----------12.464.915.512.436.115.2
U2Seg22.813.011.821.531.048.110.825.063.930.216.171.119.917.652.721.7
vs.prev.SOTA+0.9+0.7+2.8+11.2+4.2+20.9+0.2+13.2+7.0+2.0+3.7+6.2+4.4+5.2+16.6+6.5
+ +Table 1. With a unified framework, U2Seg outperforms previous state-of-the-art methods tailored for individual tasks across various datasets, including CutLER for unsupervised instance segmentation, STEGO for unsupervised semantic segmentation, and CutLER+STEGO for unsupervised panoptic segmentation. "Agn Instance Seg" denotes class-agnostic instance segmentation. + +mine a pixel's classification as "things" or "stuff" using the following logic: + +$$ +I (i, j) = \left\{ \begin{array}{l l} I _ {C} (i, j), & \text {i f} I _ {C} (i, j) \neq 0 \\ S _ {S} (i, j), & \text {i f} I _ {C} (i, j) = 0 \& S _ {S} (i, j) \neq 0 \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {5} +$$ + +This process merges the semantic labels, assigning priority to "things" labels over "stuff" where applicable. We then train a universal segmentation model on these pseudo-labels for instance, semantic and panoptic segmentation tasks. + +Learning Unsupervised Universal Image Segmentor. After we obtain the pseudo labels for panoptic segmentation, following [33], we construct an unsupervised universal image segmentation model, that has two branches: instance segmentation branch and semantic segmentation branch, to address corresponding segmentation tasks. The model is trained jointly for both branches, employing the following loss function: $\mathcal{L} = \lambda_i(\mathcal{L}_c + \mathcal{L}_b + \mathcal{L}_m) + \lambda_s\mathcal{L}_s$ , where $\mathcal{L}_c$ represents the classification loss, $\mathcal{L}_b$ is the detection loss, $\mathcal{L}_m$ is the segmentation loss, and $\mathcal{L}_s$ signifies the semantic loss. The $\mathcal{L}_s$ is computed as a per-pixel cross-entropy loss between the predicted and ground-truth labels. The hyperparameters $\lambda_{i}$ and $\lambda_{s}$ balance these two parts. + +# 4. Experiments and Results + +# 4.1. Experimental Setup + +Training Data. Our model is trained on 1.3M unlabeled images from ImageNet [17] and is evaluated directly across various benchmarks, unless otherwise noted. For unsupervised semantic segmentation comparisons with STEGO [24], we additionally fine-tune our model using MSCOCO's unlabeled images, following STEGO [24]. + +Test Data. For unsupervised instance segmentation, we test our model on COCO val2017, PASCAL VOC val2012 [21] and UVO val [53]. For unsupervised + +panoptic segmentation, we evaluate our model on COCO val2017 and Cityscapes val [15]. + +Evaluation Metrics. We use AP, $\mathrm{AP}_{50}$ , $\mathrm{AP}_{75}$ and $\mathrm{AR}_{100}$ to evaluate the unsupervised instance segmentation; PixelAcc and mIoU for unsupervised semantic segmentation; PQ, RQ, SQ for unsupervised universal image segmentation. After predicting the instance with its semantic labels, we use Hungarian matching to map the semantic labels to class names in the real dataset (details in B). It evaluates the consistency of the predicted semantic segments with the ground truth labels, remaining unaffected by any permutations in the predicted class labels. + +Implementation Details. Following [32], we employ Panoptic Cascade Mask R-CNN [4, 32] with a ResNet50 backbone [25]. Following CutLER's training recipe [60], our model, initialized with DINO pre-trained weights, is trained on unlabeled ImageNet for two epochs. It starts with an initial learning rate of 0.01, which then decreases after the first epoch to $5 \times 10^{-5}$ , with a batch size of 16 for all models. For unsupervised panoptic segmentation, we maintain the same training schedule as unsupervised instance segmentation for zero-shot evaluation. In non-zero-shot scenarios, the models undergo training on a combination of unlabeled COCO and ImageNet datasets, beginning with a learning rate of 0.01 over 90k steps. + +# 4.2. Unsupervised Universal Image Segmentation + +To the best of our knowledge, U2Seg represents the first effort in addressing unsupervised semantic-aware instance, semantic and panoptic segmentation, all unified under a single framework. Due to the absence of benchmarks for unsupervised semantic-aware instance segmentation and panoptic segmentation, we establish comprehensive benchmarks and baselines for both tasks. + +In Tab. 1, we demonstrate that U2Seg, utilizing a unified framework, significantly outperforms all previous approaches across various benchmarks and datasets. For + +
Metric\( AP^{box} \)\( AP^{box}_{50} \)\( AP^{box}_{75} \)\( AR^{box}_{100} \)\( AP^{mask} \)\( AP^{mask}_{50} \)\( AP^{mask}_{75} \)\( AR^{mask}_{100} \)
CutLER+5.99.06.110.35.38.65.59.3
U2Seg7.311.87.521.56.411.26.418.5
Δ+1.4+2.8+1.4+11.2+1.1+2.6+0.9+9.2
+ +Table 2. The results for zero-shot unsupervised object detection and instance segmentation on COCO val2017. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering. + +
MethodsAPboxAP50AP75AR100
CutLER+17.126.818.127.2
U2Seg19.031.019.548.1
Δ+1.9+4.2+1.4+20.9
+ +Table 3. The results for zero-shot unsupervised object detection on PASCAL VOC va12012. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering. + +
Metric\(AP^{box}\)\(AP^{box}\)\(50\)\(AR^{box}\)\(100\)\(AP^{mask}\)\(AP^{mask}\)\(50\)\(AR^{mask}\)\(100\)
CutLER+6.310.611.86.09.010.4
U2Seg6.810.825.06.29.521.0
Δ+0.5+0.2+13.2+0.2+0.5+10.6
+ +class-agnostic unsupervised instance segmentation, our method achieves an increase of $+0.9$ in $\mathrm{AP^{box}}$ compared to CutLER [60]. This improvement is largely attributed to our novel semantic-aware copy-paste augmentation, as detailed in Sec. 3.2. For unsupervised semantic-aware instance segmentation, we benchmark against the advanced baseline CutLER+, derived from CutLER, and record a substantial gain of over $11.2\%$ in AR. A more comprehensive analysis of these results is provided in Sec. 4.3. For unsupervised semantic segmentation, our approach surpasses the state-of-the-art STEGO with impressive margins of $+7.0$ in PixelAcc and $+2.0$ in mIoU. Lastly, for unsupervised panoptic segmentation, we compare against the strong baseline of CutLER+STEGO, a hybrid of CutLER+ and STEGO, and observe performance gains of over $6.2\%$ in SQ on MSCOCO and a notable $16.6\%$ improvement in SQ on Cityscapes. Further comparisons and discussions on this task are elaborated in Sec. 4.4. + +# 4.3. Unsupervised Instance Segmentation + +We performed extensive experiments for zero-shot unsupervised instance segmentation. Given that prior methods [51, 58, 60, 61] are limited to class-agnostic instance segmentation, we developed CutLER+, a strong baseline for unsupervised semantic-aware instance segmentation, building upon the current state-of-the-art CutLER [60]. CutLER+ operates in two steps: it first uses the pre-trained + +Table 4. The results for zero-shot unsupervised object detection and instance segmentation on UVO va1. The model is trained on ImageNet with a cluster number of 800. We compare with CutLER+, a combination of CutLER and offline clustering. + +
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN15.746.619.8
non zero-shot methods
CutLER+STEGOCOCO12.436.115.2
U2SegCOCO15.451.519.0
U2SegCOCO+IN17.652.721.7
Δ+5.2+16.6+6.5
+ +Table 5. Unsupervised Panoptic image segmentation on Cityscapes val. We show PQ, SQ and RQ on zero-shot and nonzero shot settings with the cluster number of 800. We compare with CutLER+STEGO, a combination of CutLER+ and STEGO. + +
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN11.160.113.7
non zero-shot methods
CutLER+STEGOCOCO12.464.915.5
U2SegCOCO15.366.519.1
U2SegCOCO+IN16.171.119.9
Δ+3.7+6.2+4.4
+ +Table 6. Unsupervised Panoptic image segmentation on COCOval2017. We show PQ, SQ and RQ on zero-shot and non-zero shot settings. We use CutLER+STEGO, a combination of CutLER+ and STEGO, as a strong baseline. + +CutLER to generate class-agnostic instance masks, and subsequently assigns semantic labels to all instance masks through offline clustering. + +Tab. 2 demonstrates that U2Seg markedly improves performance in both unsupervised object detection and instance segmentation on MSCOCO, delivering a +2.8 boost in $\mathrm{AP}_{50}^{\mathrm{box}}$ and a +2.6 rise in $\mathrm{AP}_{50}^{\mathrm{mask}}$ over CutLER+. Additionally, our method sees a substantial increase of approximately +10.0 in $\mathrm{AR}_{100}$ . Results on PASCAL VOC val2012 and UVO val are detailed in Tab. 3 and Tab. 4, respectively. Notably, we achieve gains exceeding +20% in AR for PASCAL VOC and +10% for UVO. + +# 4.4. Unsupervised Panoptic Segmentation + +For unsupervised panoptic/universal image segmentation, our experiments span two scenarios. In the zero-shot setting, the model is trained exclusively on unlabeled ImageNet images. For non-zero-shot (in-domain) scenarios, we train on unlabeled COCO images or a mix of COCO and ImageNet. With no existing benchmarks for unsupervised panoptic segmentation, we establish a new baseline by integrating the state-of-the-art unsupervised semantic segmentation from STEGO [24] with semantic-aware instance segmentation from CutLER+ (discussed in Sec. 4.3), which are then merged to create panoptic/universal segmentation outcomes, referred to as CutLER+STEGO. + +Tab. 6 presents the PQ, SQ, and RQ scores of U2Seg on COCO val2017. U2Seg surpasses the strong baseline + +![](images/57d1555519e3d36a924bb9e55a21e78830f8f13c164617af05a82981a0b9c9bd.jpg) +Figure 4. Universal image segmentation visualization in COCO val2017. We present the results with cluster IDs predicted by U2Seg, categorizing athletes playing hockey (left columns) as “139”, those playing badminton (middle columns) as “52” and gentlemen (right columns) as “132”. After Hungarian matching, the IDs are automatically matched to the category “person” for quantitative evaluations. + +![](images/0d424ff81ae77e5baee7a4d75b7198d0d2a0d2b6f50bcb4756b16bee6274d85d.jpg) + +![](images/e7a021221a3fee0363e41bda4dad055c90a7cf0b8d84137ba965ee586e87b31d.jpg) + +![](images/4200f1116b4bc2f76ba55e184002336dd93428cb00e3f6947e5c3b420ce0bec5.jpg) + +![](images/045dcb50dd5752b8c82220b7b529c8f4daacb2e082d94a5eb0feeb4bc2824cfc.jpg) + +![](images/01c2e184134b1cf9e405a63cd0a2b19210ccdaef546acfb4701813fd2c0d3ba3.jpg) + +![](images/4c8f717cb9bd25917d102ee3a56d6e441452ee896d0646d815e359b8e0126a9f.jpg) + +![](images/6472ff337e015fb0d62bbc919fdf3a116f062518a8cdfb250e3d1fe92b1a1421.jpg) + +![](images/691a1b5dd5ef893a22d178fccefb94463f4258c51caeca36cb7f2ef96cf64018.jpg) + +![](images/f08e5400b55537038db8311b872ec1631027dbe0de811e03a6a86bf6ed6d656c.jpg) +Raw Image +U2Seg +Ground-Truth +Pseudo-Labels +Figure 5. Visualizations of U2Seg's unsupervised Panoptic segmentation results on COCO val2017 (after Hungarian matching). The pseudo label is the naive combination of previous state-of-the-art instance segmentation, i.e. CutLER [60], and semantic segmentation, i.e., STEGO [24], results. + +![](images/faa57eef681a40a5bb4b7deaf25aa8b595a7e95d2e1fd7054e8431c71a2a7e08.jpg) +Raw Image +U2Seg +Ground-Truth +Pseudo-Labels +Figure 6. Qualitative results of U2Seg's Panoptic image segmentation results on Cityscapes val (after Hungarian matching). + +CutLER+STEGO with a +3.5 improvement in PQ and an increase of over +4.0 in RQ. Qualitative results of U2Seg's performance is provided in Fig. 4, with the predicted seman + +tic labels visualized. The qualitative results suggest that an over-clustering strategy in pseudo-label generation, e.g. setting the number of clusters to 300 or 800, leads to highly granular model predictions. For instance, as in Fig. 4, the model distinctly categorizes hockey players as “139”, badminton players as “52”, and gentlemen in suits as “132”, showcasing its potent discriminative capabilities. + +To quantitatively measure the quality of segmentation masks and their corresponding semantic labels, we use Hungarian matching (detailed in Appendix B) to align semantic labels with the category names from the test dataset; for instance, all three sub-clusters depicted in Fig. 4 are assigned to the "person" category. The qualitative outcomes post-Hungarian matching are shown in Fig. 5, where our model demonstrates superior panoptic segmentation mask quality. For instance, while the baseline tends to segment parts separately (as seen with the man's head and torso being treated as separate entities in the third row), our model correctly identifies them as parts of a single object. This level of recognition is also evident with the "trunks of the motorcycle" example in the second row. For additional results, please see Appendix D. We also present results of the more challenging dataset Cityscapes in Tab. 5 and Fig. 6. + +# 4.5. Efficient Learning + +Specifically, for object detection and instance segmentation, we employ our unsupervised instance segmentation model, with cluster count set to 300, to initialize the model weights. We adopt the recipe from [56, 59] for model fine-tuning across various annotation splits. For label-efficient panoptic segmentation, we fine-tune the model initialized with our zero-shot unsupervised framework on the same data splits. + +The results are depicted in Fig. 7, where our model's instance segmentation performance is benchmarked against MoCo-V2, DETReg, and CutLER. Our model consistently surpasses the state-of-the-art with consistent gains in both $\mathrm{AP^{box}}$ and $\mathrm{AP^{mask}}$ . In scenarios with panoptic image segmentation as the downstream task, we contrast our results with MoCo-V2 and CutLER in terms of PQ, SQ, and RQ metrics. The results illustrate a remarkable improvement, effectively doubling the performance boost from MoCo-V2 to CutLER, especially in few-shot contexts with limited an- + +![](images/bf83c6c4ce582435bd099f332bf0d1ea509197b3dc0f0c5065a608a7d3e771a7.jpg) +Figure 7. We evaluate the label-efficient learning performance on 3 different tasks: object detection (the left), instance segmentation (the second left) and panoptic image segmentation (the last three). + +![](images/db12f59874e9387e28a98187214e676f8c0d4ab543a7c339d0d3c9bb4f9dc8e6.jpg) + +![](images/00535cc70a101c8c1801a2aec1ff91ad2cffe36ef73ad5c84f9ddc2baae8040e.jpg) + +![](images/1d4f66e960707b4a7140125fd020a3bb1605ae088a8126076e49089439809a2d.jpg) + +![](images/ae688957d09ee9443bf9d72e0c9b49b955c3bb8b596608560320bd616347f608.jpg) + +![](images/0f7e9dabd66cb0d7ceb2d5564a48a4496f35dad5d521b0bc3866592f8b0b68c0.jpg) +Figure 8. U2Seg learns features that are more discriminative than those learned by CutLER. The t-SNE [47] visualization of the features from the model's FC layer. We color-code each dot based on its ground-truth category. + +![](images/2ec9a34b1971e2419382e4e33cee5a65505fe30ec262cb03d484d8fe4dbab502.jpg) + +
# ClusterCOCOUVOVOC
APbox50ARbox100APbox50ARbox100APbox50ARbox100
3009.320.19.822.629.645.7
80011.821.510.825.031.048.0
291113.322.115.125.831.648.3
+ +notations (1% or 2% labeled samples). This highlights the practical value of our approach in real-world unsupervised learning applications, where annotations are often scarce. + +We attribute the performance gains primarily to the discriminative features our model learns, as in Fig. 8, obtaining effective model initialization for few-shot learning. + +# 4.6. Ablation Studies + +In this section, we conduct ablation study on U2Seg. + +Numbers of clusters. The choice of cluster quantity significantly affects the model's representation granularity. Our ablation study on various cluster counts, as detailed in Tab. 7, reveals their impact on model performance. Overclustering generally leads to a finer level of detail, prompting the model to learn more discriminative features. + +Table 7. Over-clustering can improve the model performance. We show results on different datasets for the unsupervised object detection using different cluster numbers. + +
conf# matchedAPbox50ARbox100
0.910910.913.1
0.722511.618.0
0.628211.819.7
0.438911.821.5
0.251311.321.8
0.07188.618.4
+ +(a) Conf's effect on accuracy. + +
IoU# matchedAPbox50ARbox100
0.929510.819.7
0.834811.420.7
0.441411.521.6
0.245011.521.1
0.04949.217.7
0.638911.821.5
+ +(b) IoU's effect on accuracy. + +Table 8. Impact of Confidence and IoU on Hungarian Matching Performance: The left table illustrates the outcomes at a fixed IoU of 0.6 while varying the confidence scores. Conversely, the right table displays the results with a constant confidence of 0.4, altering the IoU values. The cluster number is 800. + +Hungarian matching. As our trained model could predict the instance with corresponding semantic labels, we are able to go further beyond unsupervised class-agnostic instance segmentation. To quantitatively evaluate the performance, Hungarian matching is employed to match the predicted semantic labels to the ground-truth dataset categories. See Appendix B for details of the adopted Hungarian matching used in our evaluation. As shown in Tab. 8, the two parameters conf threshold and IoU threshold also affect the precision and recall. + +# 5. Summary + +We present U2Seg, a novel Unsupervised Universal Image Segmentation model, adept at performing unsupervised instance, semantic, and panoptic segmentation tasks within a unified framework. Evaluated on extensive benchmarks, U2Seg consistently outperforms previous state-of-the-art methods designed for individual tasks. Additionally, U2Seg achieves the new state-of-the-art for label-efficient panoptic segmentation and instance segmentation. We anticipate that U2Seg, free from the constraints of human annotations, will demonstrate enhanced performance when scaled up with more training data, representing an important direction for our future research. + +Acknowledgement: Trevor Darrell, Dantong Niu and Xudong Wang were funded by DoD including DARPA LwLL and the Berkeley AI Research (BAIR) Commons. + +# References + +[1] Edward H Adelson. On seeing stuff: the perception of materials by humans and machines. In Human Vision and Electronic Imaging, 2001. +[2] YM Asano, C Rupprecht, and A Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2019. +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020. +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. +[11] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-DeepLab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. +[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021. +[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. +[14] Jang Hyun Cho, U. Mall, K. Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. ArXiv, abs/2103.17070, 2021. +[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. CoRR, abs/1604.01685, 2016. + +[16] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13693-13702, 2021. +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[18] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[20] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. +[21] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. +[22] Edward W Forgy. Cluster analysis of multivariate data: efficiency versus interpretability of classifications. biometrics, 21:768-769, 1965. +[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. +[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[26] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. +[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. +[29] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. 2023. + +[30] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019. +[31] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9396-9405, 2018. +[32] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. +[33] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. +[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. +[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. +[36] Hans-Peter Kriegel, Erich Schubert, and Arthur Zimek. The (black) art of runtime evaluation: Are we comparing algorithms or implementations? Knowledge and Information Systems, 52(2):341-378, 2017. +[37] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. arXiv preprint arXiv:2108.07682, 2021. +[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. +[41] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. +[42] Geoffrey J McLachlan and Thriyambakam Krishnan. The EM algorithm and extensions. John Wiley & Sons, 2007. +[43] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020. +[44] Sungwon Park, Sungwon Han, Sundong Kim, Danu Kim, Sungkyu Park, Seunghoon Hong, and Meeyoung Cha. Improving unsupervised image clustering with robust learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12278-12287, 2021. +[45] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. + +[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. +[47] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. +[48] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. +[49] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. arxiv preprint arxiv:2102.06191, 2021. +[50] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022. +[51] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. +[52] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021. +[53] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, openworld segmentation. CoRR, abs/2104.04691, 2021. +[54] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10776-10785, 2021. +[55] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. +[56] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2021. +[57] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In European Conference on Computer Vision, pages 427-445. Springer, 2022. +[58] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. +[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023. +[60] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance + +segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. +[61] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022. +[62] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3733-3742, 2018. +[63] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. +[64] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019. +[65] Chengxu Zhuang, Alex Lin Zhai, Daniel Yamins, , et al. Local aggregation for unsupervised learning of visual embeddings. In ICCV, 2019. +[66] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. \ No newline at end of file diff --git a/2024/Unsupervised Universal Image Segmentation/images.zip b/2024/Unsupervised Universal Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a19919f3129ccedcfc32415142706805da46cbb --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f6ccbafd3f4d67e20dcf105f482ccbabaa55af0141786d7447ee105e54fe940 +size 787990 diff --git a/2024/Unsupervised Universal Image Segmentation/layout.json b/2024/Unsupervised Universal Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40a2cc2e74501ab84c0a4196ea07f151ebf95418 --- /dev/null +++ b/2024/Unsupervised Universal Image Segmentation/layout.json @@ -0,0 +1,9456 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "spans": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "type": "text", + "content": "Unsupervised Universal Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 142, + 523, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 142, + 523, + 171 + ], + "spans": [ + { + "bbox": [ + 68, + 142, + 523, + 171 + ], + "type": "text", + "content": "Dantong Niu*† Xudong Wang*† Xinyang Han* Long Lian Roei Herzig Trevor Darrell Berkeley AI Research, UC Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "spans": [ + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "type": "text", + "content": "Code: https://github.com/u2seg/U2Seg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 196, + 538, + 415 + ], + "blocks": [ + { + "bbox": [ + 51, + 196, + 538, + 415 + ], + "lines": [ + { + "bbox": [ + 51, + 196, + 538, + 415 + ], + "spans": [ + { + "bbox": [ + 51, + 196, + 538, + 415 + ], + "type": "image", + "image_path": "9f6c80cf4672c46b0a61ea77a23d3c8c725ee3dabd3b4a1b5d3bb276bc9e5dd2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 416, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 546, + 460 + ], + "type": "text", + "content": "Figure 1. We present U2Seg, a unified framework for Unsupervised Universal image Segmentation that consistently outperforms previous state-of-the-art methods designed for individual tasks: CutLER [60] for unsupervised instance segmentation, STEGO [24] for unsupervised semantic segmentation, and the naive combination of CutLER and STEGO for unsupervised panoptic segmentation. We visualize instance segmentation results with \"semantic label\" + confidence score and semantic predictions with \"semantic label\". Zoom in for the best view." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "spans": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 513, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 513, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 513, + 288, + 693 + ], + "type": "text", + "content": "Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "content": "methods tailored to each task: " + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "inline_equation", + "content": "a + 2.6AP^{box}" + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "content": " boost (vs. CutLER) in unsupervised instance segmentation on COCO and " + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "inline_equation", + "content": "a + 7.0" + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "content": " PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOSTuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by " + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "inline_equation", + "content": "+5.0AP^{mask}" + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "content": " when trained on a low-data regime, e.g., only " + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 306, + 475, + 547, + 608 + ], + "type": "text", + "content": " COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 632, + 387, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 387, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 387, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The field of image segmentation has witnessed significant advancements in the recent years [4, 5, 12, 20, 26, 35, 38, 40]. Nonetheless, the effectiveness of these segmentation methods heavily depends on the availability of extensive densely human-labeled data for training these models," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 189, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 189, + 713 + ], + "type": "text", + "content": "*: Equal Contribution; †: Project Lead." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22744" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "which is both labor-intensive and costly and thus less scalable. In this paper, our objective is to explore the extent to which unsupervised image segmentation can be achieved without relying on any human-generated labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "type": "text", + "content": "Several recent works such as CutLER [60] and STEGO [24] have emerged as promising approaches for unsupervised image segmentation. CutLER leverages the property of the self-supervised model DINO [8] to 'discover' objects without supervision, and learns a state-of-the-art localization model on pseudo instance segmentation masks produced by MaskCut [60] (based on Normalize Cuts [45]). Similarly leveraging DINO [8], STEGO [24] introduces a novel framework that distills unsupervised features into discrete semantic labels. This is achieved using a contrastive loss that encourages pixel features to form compact clusters while preserving their relationships across the corpora [24]. However, these methods have limitations:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 278, + 287, + 458 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 278, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 350 + ], + "type": "text", + "content": "- The output of unsupervised instance segmentation methods such as CutLER [60] comprises class-agnostic segments for \"things\", ignoring the \"stuff\" categories that represent pixel semantics. Moreover, CutLER often treats several overlapping instances as one instance, especially when these instances belong to the same semantic class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 351, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 397 + ], + "type": "text", + "content": "- On the other hand, unsupervised semantic segmentation methods such as STEGO [24] focus on the segmentation of semantically coherent regions, lacking the capability to distinguish between individual instances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 399, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 458 + ], + "type": "text", + "content": "- Unsupervised panoptic segmentation has not been addressed. Supervised panoptic segmentation methods [12, 29, 31] predict both \"stuff\" and \"things\" classes simultaneously; to the best of our knowledge there has not been work on unsupervised panoptic segmentation heretofore." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 460, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 543 + ], + "type": "text", + "content": "To address these limitations, we propose U2Seg, a novel Unsupervised Universal image Segmentation model. U2Seg offers comprehensive scene understanding—instance, semantic and panoptic—without relying on human annotations, segmenting semantically meaningful regions in the image as well as identifying and differentiating between individual instances within those regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "content": "U2Seg is comprised of three steps. First, we create high-quality, discrete semantic labels for instance masks obtained from MaskCut and DINO, by clustering semantically similar instance masks into distinct fine-grained clusters, as in Sec. 3.2. Next, we amalgamate the semantically pseudolabeled \"things\" pixels (from the first step) with \"stuff\" pixels (from STEGO) to produce pseudo semantic labels for each pixel in the image. Lastly, a universal image segmentation model is trained using these pseudo-labels, resulting in a model capable of simultaneously predicting pixel-level (i.e., semantic and class-agnostic instance segmentation) and instance-level semantic labels, detailed in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Despite the inherent noise in these pseudo-labels, self-training the model with them yields substantial performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "gains over specialized methods tailored to each task: U2Seg achieves a " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "+2.6" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " boost (vs. CutLER) in unsupervised instance segmentation on COCO and a " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "+7.0" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation. We also find that the multi-task learning framework and learning unsupervised segmentor with semantic labels enable our model to generate a more discriminative feature space, which makes it a superior representation for downstream supervised detection and segmentation tasks. When trained on a low-data regime, such as " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " COCO labels, U2Seg surpasses CutLER by " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "+5.0" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{AP}^{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "type": "text", + "content": "Contributions. Our main contribution is the first universal unsupervised image segmentation model that can tackle unsupervised semantic-aware instance, semantic and panoptic segmentation tasks using a unified framework. We establish a suite of benchmarks on unsupervised semantic-aware instance segmentation and panoptic segmentation, areas previously unexplored. Despite using a single framework, we demonstrate that U2Seg surpasses previous methods specialized for each task across all experimented benchmarks (instance, semantic, panoptic, etc.) and datasets (COCO, Cityscapes, UVO, VOC, etc.)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "type": "text", + "content": "Self-supervised Representation Learning focuses on feature learning from a large amount of unlabeled data without using human-made labels. Contrastive Learning-Based Methods [9, 27, 43, 62] learn representation by comparing similar instances or different versions of a single instance while separating dissimilar ones. Similarity-Based Self-Supervised Learning [10, 23] mainly reduces differences between different augmented versions of the same instance. Clustering-Based Feature Learning [2, 7, 55, 63, 65] finds natural data groups in the hidden space. Masked Autoencoders [3, 18, 28] learn by masking and then reconstructing masked parts of the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 700 + ], + "type": "text", + "content": "Unsupervised Object Detection and Instance Segmentation. DINO [8] shows that self-supervised learning (SSL) Vision Transformers (ViT) [19] can reveal hidden semantic segmentation in images, which is not obvious in supervised counterparts [8, 66]. Extending this, LOST [46], TokenCut [61] and MaskDistill [50] use DINO's patch features to identify main objects in images. FreeSOLO [58] performs unsupervised class-agnostic instance segmentation by creating coarse masks first, which are later improved through self-training. Meanwhile, CutLER [59] introduces the MaskCut method, which aims to identify multiple instances in a single image. Yet, MaskCut frequently consolidates overlapping instances into a single segment and lacks the capability to assign semantic labels to each instance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Unsupervised Semantic Segmentation. IIC [30] maxi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22745" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "mizes mutual information for clustering, while PiCIE [14] uses invariance to photometric effects and equivariance to geometric transformations for segmentation. MaskContrast [49] learns unsupervised semantic segmentation by contrasting features within saliency masks. STEGO [24] refines pretrained SSL visual features to distill correspondence information embedded within these features, thereby fostering discrete semantic clusters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "content": "Universal Segmentation has been introduced to deliver instance, semantic and panoptic segmentation tasks using a unified architecture [6, 11-13, 29, 33, 34, 37, 52, 64]. In this work, we propose U2Seg to tackle this challenging task without relying on human-annotated data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 286, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 373 + ], + "type": "text", + "content": "Unsupervised Image Classification methods mainly focus on providing a semantic label for each query image that can be mapped to ground truth classes by hungarian matching. SCAN [48] proposes a three-stage pipeline that includes representation learning, deep clustering, and self-labeling. NNM [16] enhances SCAN by incorporating local and global nearest neighbor matching. RUC [44] further improves SCAN using a robust loss as training objective. However, these approaches only provide one classification prediction per image, whereas our method provides classification per-instance for instance segmentation and per-pixel for semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 384, + 257, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 257, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 257, + 398 + ], + "type": "text", + "content": "3. Unsupervised Universal Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 404, + 133, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 133, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 133, + 415 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 423, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 286, + 459 + ], + "type": "text", + "content": "We first explain the previous Unsupervised Instance Segmentation method CutLER [60], and Unsupervised Semantic Segmentation method STEGO [24]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": "CutLER [60] exploits self-supervised learning models like DINO [8] to 'discover' objects and train a state-of-the-art detection and segmentation model using a cut-and-learn pipeline. It first uses MaskCut to extract multiple initial masks from DINO [8] features. MaskCut first generates a patch-wise affinity matrix " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "W_{ij} = \\frac{K_i K_j}{\\|K_i\\|_2\\|K_j\\|_2}" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": " using the \"key\" features " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "K_i" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": " for patch " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": " from DINO's last attention layer. Subsequently, the cut-based clustering method Normalized Cut [45] is employed on the affinity matrix by finding the eigenvector " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": " that corresponds to the second smallest eigenvalue. A foreground instance mask " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "M^s" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": " is derived through bi-partitioning of the vector " + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 460, + 286, + 654 + ], + "type": "text", + "content": ", enabling segmentation of individual objects in the image. For multi-instance segmentation, MaskCut iteratively refines the affinity matrix by masking out already segmented objects, allowing for subsequent extractions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 663, + 286, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 663, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 87, + 663, + 286, + 693 + ], + "type": "interline_equation", + "content": "W _ {i j} ^ {t} = \\frac {\\left(K _ {i} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right) \\left(K _ {j} \\prod_ {s = 1} ^ {t} M _ {i j} ^ {s}\\right)}{\\| K _ {i} \\| _ {2} \\| K _ {j} \\| _ {2}} \\tag {1}", + "image_path": "505073bd445b36bf1ed2c28956f836505134aec9daf8b58f77b15b3904b48c2e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "and repeating above steps by " + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": " times. CutLER then refines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 544, + 95 + ], + "type": "text", + "content": "detection and segmentation through a loss-dropping strategy and iterative self-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "text", + "content": "STEGO [24] harnesses the semantically rich feature correlations produced by unsupervised methods like DINO [8] for segmentation. It trains a segmentation head to refine these correlations within an image, with its K-Nearest Neighbors (KNNs), and across randomly chosen images. Specifically, STEGO distills DINO's unsupervised features into distinct semantic labels by optimizing a correspondence loss. This loss function measures the feature correspondences " + }, + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "inline_equation", + "content": "F^{SC}" + }, + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "text", + "content": " between image feature pairs generated by DINO and the feature correspondence " + }, + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "inline_equation", + "content": "S_{hwij}" + }, + { + "bbox": [ + 304, + 95, + 545, + 228 + ], + "type": "text", + "content": " derived from a trainable, lightweight segmentation head [24]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 236, + 545, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 236, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 545, + 263 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o r r}} (x, y, b) = - \\sum_ {h w i j} \\left(F _ {h w i j} ^ {S C} - b\\right) \\max \\left(S _ {h w i j}, 0\\right) \\tag {2}", + "image_path": "df3c0b4cd6c62775c5c6bc3e3fcb6d87d1cd9650b94d8669d25ca1305fcf7d9a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 272, + 501, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 501, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 501, + 285 + ], + "type": "text", + "content": "3.2. Unsupervised Instance Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "type": "text", + "content": "Although CutLER [60] provides high-quality instance segmentation masks without human annotations, the predicted masks are class-agnostic, and thus do not include semantic labels for each instance. Our method addresses this issue by grouping the detected instances with a clustering method. In this way, instances assigned to the same cluster are associated with identical or closely related semantic information, while instances residing in separate clusters exhibit semantic dissimilarity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "Pseudo Semantic Labels. To train a detection and instance segmentation model, we vector quantize the model targets (pseudo semantic labels) by clustering the instance-level features of the entire dataset, under constraints derived from self-supervision. Specifically, our approach starts with the generation of instance segmentation masks using MaskCut [60]. Subsequently, we utilize the efficient " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "-Means clustering method as implemented in USL [57] to cluster all segmentation masks into semantically meaningful clusters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": "We employ " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": "-Means clustering to partition " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": " instances into " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "C(\\leq n)" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": " clusters, where each cluster is represented by its centroid " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": " [22, 41]. Each instance is assigned to the cluster with the nearest centroid. Formally, we conduct a " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": "-way node partitioning, denoted as " + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = S_1, S_2, \\ldots, S_C" + }, + { + "bbox": [ + 304, + 506, + 545, + 578 + ], + "type": "text", + "content": ", that minimizes the within-cluster sum of squares [36]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 587, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 330, + 587, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} \\sum_ {V \\in S _ {i}} | V - c _ {i} | ^ {2} = \\min _ {\\mathcal {S}} \\sum_ {i = 1} ^ {C} | S _ {i} | \\operatorname {V a r} \\left(S _ {i}\\right) \\tag {3}", + "image_path": "d89d031a3e80d9c6f0605df702563f0bcb8132e687fdbd4899388473140ea464.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "This optimization process is carried out iteratively using the EM algorithm [42], starting from selecting random samples as initial centroids. As a result, this process assigns pseudo semantic labels, denoted as " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": ", to each instance " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": " falling within the range of " + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "inline_equation", + "content": "[1,C]" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "The resulting semantic labels serve multiple purposes: 1) Semantic-aware copy-paste augmentation, which signifi-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22746" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 149 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 149 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 149 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 149 + ], + "type": "image", + "image_path": "0cdd369e78f9e985922848723320a3606944d07490bdc0ae7ce4b47a1bee4c38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "type": "text", + "content": "Figure 2. Overview of the training and inference pipeline for the proposed Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks—instance, semantic and panoptic—using a novel unified framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 180, + 509, + 262 + ], + "blocks": [ + { + "bbox": [ + 88, + 180, + 509, + 262 + ], + "lines": [ + { + "bbox": [ + 88, + 180, + 509, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 180, + 509, + 262 + ], + "type": "image", + "image_path": "b67dbb46ef3a9411dfb950452309367d0c4ad6b01b5b7b1920e958b6f2d903f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 546, + 299 + ], + "type": "text", + "content": "Figure 3. Pipeline overview for generating masks and their semantically meaningful pseudo labels in semantic-aware instance segmentation. We first use MaskCut to generate class-agnostic instance masks, which are then grouped into semantically meaningful clusters. These pseudo semantic labels are used for training a semantic-aware instance segmentor." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "content": "cantly improves CutLER's capability to differentiate overlapping instances, especially when they share similar semantic information. 2) Training instance segmentation models: They serve as pseudo ground-truth labels for training a non-agnostic instance segmentor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 474 + ], + "type": "text", + "content": "Semantic-aware Copy-Paste Augmentation. In cluttered natural scenes, previous unsupervised instance segmentation model often fail to distinguish instances from the same semantic class. This results in multiple instances being captured in the same mask. To distinguish multiple overlapping objects and small objects in existing unsupervised detectors, we employ semantic-aware copy-paste augmentation, which includes several steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "text", + "content": "1) We begin by randomly selecting two instances, denoted as " + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "text", + "content": ", both belonging to the same pseudocategory (or group/cluster). 2) One of these instances undergoes a transformation function " + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 474, + 287, + 559 + ], + "type": "text", + "content": ", which randomly reizes and shifts the associated pseudo-masks. 3) The resized instance is then pasted onto another image, creating synthetic overlapping cases using the following equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 563, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 563, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 287, + 576 + ], + "type": "interline_equation", + "content": "I _ {3} = I _ {1} \\cdot (1 - \\mathcal {T} (M _ {c})) + I _ {2} \\cdot \\mathcal {T} (M _ {c}) \\tag {4}", + "image_path": "dce6b9223f9ce889d0aacf9d00ef859b21d7e1ee6fce6e7d4bf3d4d1eb15ef49.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 228, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 228, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 228, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 228, + 593 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 582, + 228, + 593 + ], + "type": "text", + "content": " denotes element-wise multiplication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": "Learning Unsupervised Instance Segmentor. Traditionally, unsupervised segmentation community focused primarily on class-agnostic instance segmentation [58, 60, 61], whose outputs lack class labels. However, by incorporating clustering information obtained from pseudo-labels on ImageNet, as discussed above, our method allows the model to predict not only the location and segmentation of an object but also its pseudo semantic labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "As observed by [60], \"ground-truth\" masks may miss instances. However, a standard detection loss penalizes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": "predicted regions " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " that do not overlap with the \"ground-truth\". Therefore, following [60], we drop the loss for each predicted region " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " that has a maximum overlap of " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\tau^{\\mathrm{IoU}}" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " with any of the 'ground-truth' instances: " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{drop}}(r_i) = \\mathbb{1}(\\mathrm{IoU}_i^{\\max} > \\tau^{\\mathrm{IoU}})\\mathcal{L}_{\\mathrm{vanilla}}(r_i)" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_i^{\\max}" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " denotes the maximum IoU with all 'ground-truth' for " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vanilla}}" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " is the vanilla loss function of detectors. " + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{drop}}" + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": " encourages the exploration of image regions missed in the \"ground-truth\"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 420, + 539, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 539, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 539, + 433 + ], + "type": "text", + "content": "3.3. Unsupervised Universal Image Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "content": "Pseudo Labels for Panoptic Segmentation. For each pixel " + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "content": " in the image, we vector quantize pixels with different semantics or instance membership, generating pseudo semantic labels for panoptic segmentation. We assign each pixel a semantic label based on \"stuff\" or \"things\" identity. This results in an instance label " + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "(I(i,j))" + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "content": " for \"things\" or a semantic label " + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "inline_equation", + "content": "(S(i,j))" + }, + { + "bbox": [ + 304, + 437, + 545, + 557 + ], + "type": "text", + "content": " for \"stuff\". The critical challenge in this process is distinguishing between pixels associated with \"things\" (countable, often foreground) and \"stuff\" (uncountable, usually background) [1]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "To resolve this problem, our method unfolds in three steps: 1) Semantic Labeling for \"Things\": Utilizing the class-agnostic instance segmentation capabilities of CutLER [60], we first identify \"things\" within an image, generating class-agnostic instance masks. These masks then undergo deep clustering to attribute a semantic label " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "I_C(i,j)" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " to each instance, detailed in Sec. 3.2. 2) Semantic Labeling for \"Stuff\": For \"stuff\" pixels, we deploy the unsupervised semantic segmentation model STEGO [24], which distills DINO's unsupervised features into discrete semantic labels, as outlined in 3.1. This step assigns a \"stuff\" semantic label to all pixels, including those of \"Things\" identified earlier. 3) Integrating Labels for \"Things\" and \"Stuff\". We deter" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22747" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 68, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 541, + 203 + ], + "type": "table", + "html": "
Task → Datasets → Metric →Agn Instance Seg.Instance Seg.Semantic Seg.Panoptic Seg.
COCOCOCOVOCUVOCOCOCOCOCityscapes
APboxAPbox50APbox50ARbox100APbox50ARbox100APbox50ARbox100PixelAccmIoUPQSQRQPQSQRQ
FreeSOLO [58]9.64.2-------------
TokenCut [61]5.83.2-------------
DINO [8]--------30.59.6-----
PiCIE + H [14]--------48.113.8-----
STEGO [24]--------56.928.2-----
CutLER [59]21.912.3-------------
CutLER+--9.010.326.827.210.611.8-------
CutLER+STEGO----------12.464.915.512.436.115.2
U2Seg22.813.011.821.531.048.110.825.063.930.216.171.119.917.652.721.7
vs.prev.SOTA+0.9+0.7+2.8+11.2+4.2+20.9+0.2+13.2+7.0+2.0+3.7+6.2+4.4+5.2+16.6+6.5
", + "image_path": "337bc35090a1995e040b92da3caf0ba23189d9156b13a6cb8895d1f2cef56a0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 546, + 239 + ], + "type": "text", + "content": "Table 1. With a unified framework, U2Seg outperforms previous state-of-the-art methods tailored for individual tasks across various datasets, including CutLER for unsupervised instance segmentation, STEGO for unsupervised semantic segmentation, and CutLER+STEGO for unsupervised panoptic segmentation. \"Agn Instance Seg\" denotes class-agnostic instance segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "type": "text", + "content": "mine a pixel's classification as \"things\" or \"stuff\" using the following logic:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 287, + 338 + ], + "type": "interline_equation", + "content": "I (i, j) = \\left\\{ \\begin{array}{l l} I _ {C} (i, j), & \\text {i f} I _ {C} (i, j) \\neq 0 \\\\ S _ {S} (i, j), & \\text {i f} I _ {C} (i, j) = 0 \\& S _ {S} (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}", + "image_path": "265ea7740e669688baf8f8eb731fff011bee7d8c9e1ee5dcc3a91070e13012bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "content": "This process merges the semantic labels, assigning priority to \"things\" labels over \"stuff\" where applicable. We then train a universal segmentation model on these pseudo-labels for instance, semantic and panoptic segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": "Learning Unsupervised Universal Image Segmentor. After we obtain the pseudo labels for panoptic segmentation, following [33], we construct an unsupervised universal image segmentation model, that has two branches: instance segmentation branch and semantic segmentation branch, to address corresponding segmentation tasks. The model is trained jointly for both branches, employing the following loss function: " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\lambda_i(\\mathcal{L}_c + \\mathcal{L}_b + \\mathcal{L}_m) + \\lambda_s\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " represents the classification loss, " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " is the detection loss, " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " is the segmentation loss, and " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " signifies the semantic loss. The " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " is computed as a per-pixel cross-entropy loss between the predicted and ground-truth labels. The hyperparameters " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 397, + 287, + 553 + ], + "type": "text", + "content": " balance these two parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 565, + 190, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 190, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 190, + 578 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 164, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 164, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 164, + 600 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "Training Data. Our model is trained on 1.3M unlabeled images from ImageNet [17] and is evaluated directly across various benchmarks, unless otherwise noted. For unsupervised semantic segmentation comparisons with STEGO [24], we additionally fine-tune our model using MSCOCO's unlabeled images, following STEGO [24]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Test Data. For unsupervised instance segmentation, we test our model on COCO val2017, PASCAL VOC val2012 [21] and UVO val [53]. For unsupervised" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 259, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 259, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 545, + 283 + ], + "type": "text", + "content": "panoptic segmentation, we evaluate our model on COCO val2017 and Cityscapes val [15]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "text", + "content": "Evaluation Metrics. We use AP, " + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 304, + 285, + 546, + 404 + ], + "type": "text", + "content": " to evaluate the unsupervised instance segmentation; PixelAcc and mIoU for unsupervised semantic segmentation; PQ, RQ, SQ for unsupervised universal image segmentation. After predicting the instance with its semantic labels, we use Hungarian matching to map the semantic labels to class names in the real dataset (details in B). It evaluates the consistency of the predicted semantic segments with the ground truth labels, remaining unaffected by any permutations in the predicted class labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 406, + 547, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 562 + ], + "type": "text", + "content": "Implementation Details. Following [32], we employ Panoptic Cascade Mask R-CNN [4, 32] with a ResNet50 backbone [25]. Following CutLER's training recipe [60], our model, initialized with DINO pre-trained weights, is trained on unlabeled ImageNet for two epochs. It starts with an initial learning rate of 0.01, which then decreases after the first epoch to " + }, + { + "bbox": [ + 304, + 406, + 547, + 562 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 406, + 547, + 562 + ], + "type": "text", + "content": ", with a batch size of 16 for all models. For unsupervised panoptic segmentation, we maintain the same training schedule as unsupervised instance segmentation for zero-shot evaluation. In non-zero-shot scenarios, the models undergo training on a combination of unlabeled COCO and ImageNet datasets, beginning with a learning rate of 0.01 over 90k steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 572, + 539, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 572, + 539, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 572, + 539, + 586 + ], + "type": "text", + "content": "4.2. Unsupervised Universal Image Segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 592, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 675 + ], + "type": "text", + "content": "To the best of our knowledge, U2Seg represents the first effort in addressing unsupervised semantic-aware instance, semantic and panoptic segmentation, all unified under a single framework. Due to the absence of benchmarks for unsupervised semantic-aware instance segmentation and panoptic segmentation, we establish comprehensive benchmarks and baselines for both tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 1, we demonstrate that U2Seg, utilizing a unified framework, significantly outperforms all previous approaches across various benchmarks and datasets. For" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22748" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 115 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 115 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 115 + ], + "type": "table", + "html": "
Metric\\( AP^{box} \\)\\( AP^{box}_{50} \\)\\( AP^{box}_{75} \\)\\( AR^{box}_{100} \\)\\( AP^{mask} \\)\\( AP^{mask}_{50} \\)\\( AP^{mask}_{75} \\)\\( AR^{mask}_{100} \\)
CutLER+5.99.06.110.35.38.65.59.3
U2Seg7.311.87.521.56.411.26.418.5
Δ+1.4+2.8+1.4+11.2+1.1+2.6+0.9+9.2
", + "image_path": "7c884a3a81406a16d8fabc1d4a8ff342f46d098f8830d874e27853656dcb03ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 169, + 284, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 118, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 286, + 161 + ], + "type": "text", + "content": "Table 2. The results for zero-shot unsupervised object detection and instance segmentation on COCO val2017. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 169, + 284, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 169, + 284, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 169, + 284, + 217 + ], + "type": "table", + "html": "
MethodsAPboxAP50AP75AR100
CutLER+17.126.818.127.2
U2Seg19.031.019.548.1
Δ+1.9+4.2+1.4+20.9
", + "image_path": "3b4f89d70b82cad555d473754042dcf32903e35f2a312759c2006e3401874543.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 271, + 284, + 318 + ], + "blocks": [ + { + "bbox": [ + 47, + 220, + 286, + 263 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 286, + 263 + ], + "type": "text", + "content": "Table 3. The results for zero-shot unsupervised object detection on PASCAL VOC va12012. The model is trained on ImageNet with a cluster number of 800. We compare it with CutLER+, a combination of CutLER and offline clustering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 271, + 284, + 318 + ], + "lines": [ + { + "bbox": [ + 50, + 271, + 284, + 318 + ], + "spans": [ + { + "bbox": [ + 50, + 271, + 284, + 318 + ], + "type": "table", + "html": "
Metric\\(AP^{box}\\)\\(AP^{box}\\)\\(50\\)\\(AR^{box}\\)\\(100\\)\\(AP^{mask}\\)\\(AP^{mask}\\)\\(50\\)\\(AR^{mask}\\)\\(100\\)
CutLER+6.310.611.86.09.010.4
U2Seg6.810.825.06.29.521.0
Δ+0.5+0.2+13.2+0.2+0.5+10.6
", + "image_path": "115ddc8823d04fe21a9a1a4223a37e966ea41f7430811b7a618ef3891d08ba4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": "class-agnostic unsupervised instance segmentation, our method achieves an increase of " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{box}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " compared to CutLER [60]. This improvement is largely attributed to our novel semantic-aware copy-paste augmentation, as detailed in Sec. 3.2. For unsupervised semantic-aware instance segmentation, we benchmark against the advanced baseline CutLER+, derived from CutLER, and record a substantial gain of over " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " in AR. A more comprehensive analysis of these results is provided in Sec. 4.3. For unsupervised semantic segmentation, our approach surpasses the state-of-the-art STEGO with impressive margins of " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "+7.0" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " in PixelAcc and " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "+2.0" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " in mIoU. Lastly, for unsupervised panoptic segmentation, we compare against the strong baseline of CutLER+STEGO, a hybrid of CutLER+ and STEGO, and observe performance gains of over " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " in SQ on MSCOCO and a notable " + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "inline_equation", + "content": "16.6\\%" + }, + { + "bbox": [ + 46, + 388, + 286, + 601 + ], + "type": "text", + "content": " improvement in SQ on Cityscapes. Further comparisons and discussions on this task are elaborated in Sec. 4.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 242, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 242, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 242, + 623 + ], + "type": "text", + "content": "4.3. Unsupervised Instance Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "We performed extensive experiments for zero-shot unsupervised instance segmentation. Given that prior methods [51, 58, 60, 61] are limited to class-agnostic instance segmentation, we developed CutLER+, a strong baseline for unsupervised semantic-aware instance segmentation, building upon the current state-of-the-art CutLER [60]. CutLER+ operates in two steps: it first uses the pre-trained" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 46, + 321, + 286, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 286, + 365 + ], + "type": "text", + "content": "Table 4. The results for zero-shot unsupervised object detection and instance segmentation on UVO va1. The model is trained on ImageNet with a cluster number of 800. We compare with CutLER+, a combination of CutLER and offline clustering." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 71, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 161 + ], + "type": "table", + "html": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN15.746.619.8
non zero-shot methods
CutLER+STEGOCOCO12.436.115.2
U2SegCOCO15.451.519.0
U2SegCOCO+IN17.652.721.7
Δ+5.2+16.6+6.5
", + "image_path": "fc90c06b085655a8e9428566e73bdc0d232015a8fc9e14924b15d7ebedaccb1c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 216, + 543, + 308 + ], + "blocks": [ + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "type": "text", + "content": "Table 5. Unsupervised Panoptic image segmentation on Cityscapes val. We show PQ, SQ and RQ on zero-shot and nonzero shot settings with the cluster number of 800. We compare with CutLER+STEGO, a combination of CutLER+ and STEGO." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 216, + 543, + 308 + ], + "lines": [ + { + "bbox": [ + 309, + 216, + 543, + 308 + ], + "spans": [ + { + "bbox": [ + 309, + 216, + 543, + 308 + ], + "type": "table", + "html": "
MethodsPretrainPQSQRQ
zero-shot methods
U2SegIN11.160.113.7
non zero-shot methods
CutLER+STEGOCOCO12.464.915.5
U2SegCOCO15.366.519.1
U2SegCOCO+IN16.171.119.9
Δ+3.7+6.2+4.4
", + "image_path": "e8fb8c50919c64fa25dc416fcdfd68847fbe08e52f0d5418a0662fcdc2547db7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 311, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 305, + 311, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 545, + 354 + ], + "type": "text", + "content": "Table 6. Unsupervised Panoptic image segmentation on COCOval2017. We show PQ, SQ and RQ on zero-shot and non-zero shot settings. We use CutLER+STEGO, a combination of CutLER+ and STEGO, as a strong baseline." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 376, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 411 + ], + "type": "text", + "content": "CutLER to generate class-agnostic instance masks, and subsequently assigns semantic labels to all instance masks through offline clustering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": "Tab. 2 demonstrates that U2Seg markedly improves performance in both unsupervised object detection and instance segmentation on MSCOCO, delivering a +2.8 boost in " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": " and a +2.6 rise in " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": " over CutLER+. Additionally, our method sees a substantial increase of approximately +10.0 in " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{100}" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": ". Results on PASCAL VOC val2012 and UVO val are detailed in Tab. 3 and Tab. 4, respectively. Notably, we achieve gains exceeding +20% in AR for PASCAL VOC and +10% for UVO." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 528, + 503, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 503, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 503, + 540 + ], + "type": "text", + "content": "4.4. Unsupervised Panoptic Segmentation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 546, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 688 + ], + "type": "text", + "content": "For unsupervised panoptic/universal image segmentation, our experiments span two scenarios. In the zero-shot setting, the model is trained exclusively on unlabeled ImageNet images. For non-zero-shot (in-domain) scenarios, we train on unlabeled COCO images or a mix of COCO and ImageNet. With no existing benchmarks for unsupervised panoptic segmentation, we establish a new baseline by integrating the state-of-the-art unsupervised semantic segmentation from STEGO [24] with semantic-aware instance segmentation from CutLER+ (discussed in Sec. 4.3), which are then merged to create panoptic/universal segmentation outcomes, referred to as CutLER+STEGO." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Tab. 6 presents the PQ, SQ, and RQ scores of U2Seg on COCO val2017. U2Seg surpasses the strong baseline" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22749" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 76, + 108, + 144 + ], + "blocks": [ + { + "bbox": [ + 51, + 76, + 108, + 144 + ], + "lines": [ + { + "bbox": [ + 51, + 76, + 108, + 144 + ], + "spans": [ + { + "bbox": [ + 51, + 76, + 108, + 144 + ], + "type": "image", + "image_path": "57d1555519e3d36a924bb9e55a21e78830f8f13c164617af05a82981a0b9c9bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 148, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 545, + 181 + ], + "type": "text", + "content": "Figure 4. Universal image segmentation visualization in COCO val2017. We present the results with cluster IDs predicted by U2Seg, categorizing athletes playing hockey (left columns) as “139”, those playing badminton (middle columns) as “52” and gentlemen (right columns) as “132”. After Hungarian matching, the IDs are automatically matched to the category “person” for quantitative evaluations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 76, + 154, + 144 + ], + "blocks": [ + { + "bbox": [ + 108, + 76, + 154, + 144 + ], + "lines": [ + { + "bbox": [ + 108, + 76, + 154, + 144 + ], + "spans": [ + { + "bbox": [ + 108, + 76, + 154, + 144 + ], + "type": "image", + "image_path": "0d424ff81ae77e5baee7a4d75b7198d0d2a0d2b6f50bcb4756b16bee6274d85d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 76, + 217, + 144 + ], + "blocks": [ + { + "bbox": [ + 154, + 76, + 217, + 144 + ], + "lines": [ + { + "bbox": [ + 154, + 76, + 217, + 144 + ], + "spans": [ + { + "bbox": [ + 154, + 76, + 217, + 144 + ], + "type": "image", + "image_path": "e7a021221a3fee0363e41bda4dad055c90a7cf0b8d84137ba965ee586e87b31d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 76, + 275, + 144 + ], + "blocks": [ + { + "bbox": [ + 217, + 76, + 275, + 144 + ], + "lines": [ + { + "bbox": [ + 217, + 76, + 275, + 144 + ], + "spans": [ + { + "bbox": [ + 217, + 76, + 275, + 144 + ], + "type": "image", + "image_path": "4200f1116b4bc2f76ba55e184002336dd93428cb00e3f6947e5c3b420ce0bec5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 276, + 76, + 342, + 144 + ], + "blocks": [ + { + "bbox": [ + 276, + 76, + 342, + 144 + ], + "lines": [ + { + "bbox": [ + 276, + 76, + 342, + 144 + ], + "spans": [ + { + "bbox": [ + 276, + 76, + 342, + 144 + ], + "type": "image", + "image_path": "045dcb50dd5752b8c82220b7b529c8f4daacb2e082d94a5eb0feeb4bc2824cfc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 76, + 390, + 144 + ], + "blocks": [ + { + "bbox": [ + 343, + 76, + 390, + 144 + ], + "lines": [ + { + "bbox": [ + 343, + 76, + 390, + 144 + ], + "spans": [ + { + "bbox": [ + 343, + 76, + 390, + 144 + ], + "type": "image", + "image_path": "01c2e184134b1cf9e405a63cd0a2b19210ccdaef546acfb4701813fd2c0d3ba3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 391, + 76, + 444, + 144 + ], + "blocks": [ + { + "bbox": [ + 391, + 76, + 444, + 144 + ], + "lines": [ + { + "bbox": [ + 391, + 76, + 444, + 144 + ], + "spans": [ + { + "bbox": [ + 391, + 76, + 444, + 144 + ], + "type": "image", + "image_path": "4c8f717cb9bd25917d102ee3a56d6e441452ee896d0646d815e359b8e0126a9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 444, + 76, + 482, + 144 + ], + "blocks": [ + { + "bbox": [ + 444, + 76, + 482, + 144 + ], + "lines": [ + { + "bbox": [ + 444, + 76, + 482, + 144 + ], + "spans": [ + { + "bbox": [ + 444, + 76, + 482, + 144 + ], + "type": "image", + "image_path": "6472ff337e015fb0d62bbc919fdf3a116f062518a8cdfb250e3d1fe92b1a1421.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 483, + 76, + 536, + 144 + ], + "blocks": [ + { + "bbox": [ + 483, + 76, + 536, + 144 + ], + "lines": [ + { + "bbox": [ + 483, + 76, + 536, + 144 + ], + "spans": [ + { + "bbox": [ + 483, + 76, + 536, + 144 + ], + "type": "image", + "image_path": "691a1b5dd5ef893a22d178fccefb94463f4258c51caeca36cb7f2ef96cf64018.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 55, + 198, + 279, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 279, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 279, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 279, + 360 + ], + "type": "image", + "image_path": "f08e5400b55537038db8311b872ec1631027dbe0de811e03a6a86bf6ed6d656c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 363, + 104, + 373 + ], + "lines": [ + { + "bbox": [ + 62, + 363, + 104, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 104, + 373 + ], + "type": "text", + "content": "Raw Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 363, + 151, + 373 + ], + "lines": [ + { + "bbox": [ + 125, + 363, + 151, + 373 + ], + "spans": [ + { + "bbox": [ + 125, + 363, + 151, + 373 + ], + "type": "text", + "content": "U2Seg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 363, + 220, + 372 + ], + "lines": [ + { + "bbox": [ + 169, + 363, + 220, + 372 + ], + "spans": [ + { + "bbox": [ + 169, + 363, + 220, + 372 + ], + "type": "text", + "content": "Ground-Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 363, + 279, + 372 + ], + "lines": [ + { + "bbox": [ + 223, + 363, + 279, + 372 + ], + "spans": [ + { + "bbox": [ + 223, + 363, + 279, + 372 + ], + "type": "text", + "content": "Pseudo-Labels" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 375, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 430 + ], + "type": "text", + "content": "Figure 5. Visualizations of U2Seg's unsupervised Panoptic segmentation results on COCO val2017 (after Hungarian matching). The pseudo label is the naive combination of previous state-of-the-art instance segmentation, i.e. CutLER [60], and semantic segmentation, i.e., STEGO [24], results." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 55, + 448, + 279, + 616 + ], + "blocks": [ + { + "bbox": [ + 55, + 448, + 279, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 279, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 279, + 616 + ], + "type": "image", + "image_path": "faa57eef681a40a5bb4b7deaf25aa8b595a7e95d2e1fd7054e8431c71a2a7e08.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 617, + 104, + 627 + ], + "lines": [ + { + "bbox": [ + 62, + 617, + 104, + 627 + ], + "spans": [ + { + "bbox": [ + 62, + 617, + 104, + 627 + ], + "type": "text", + "content": "Raw Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 618, + 151, + 627 + ], + "lines": [ + { + "bbox": [ + 125, + 618, + 151, + 627 + ], + "spans": [ + { + "bbox": [ + 125, + 618, + 151, + 627 + ], + "type": "text", + "content": "U2Seg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 618, + 220, + 627 + ], + "lines": [ + { + "bbox": [ + 168, + 618, + 220, + 627 + ], + "spans": [ + { + "bbox": [ + 168, + 618, + 220, + 627 + ], + "type": "text", + "content": "Ground-Truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 618, + 279, + 627 + ], + "lines": [ + { + "bbox": [ + 223, + 618, + 279, + 627 + ], + "spans": [ + { + "bbox": [ + 223, + 618, + 279, + 627 + ], + "type": "text", + "content": "Pseudo-Labels" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 629, + 286, + 651 + ], + "lines": [ + { + "bbox": [ + 47, + 629, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 286, + 651 + ], + "type": "text", + "content": "Figure 6. Qualitative results of U2Seg's Panoptic image segmentation results on Cityscapes val (after Hungarian matching)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "CutLER+STEGO with a +3.5 improvement in PQ and an increase of over +4.0 in RQ. Qualitative results of U2Seg's performance is provided in Fig. 4, with the predicted seman" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 202, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 285 + ], + "type": "text", + "content": "tic labels visualized. The qualitative results suggest that an over-clustering strategy in pseudo-label generation, e.g. setting the number of clusters to 300 or 800, leads to highly granular model predictions. For instance, as in Fig. 4, the model distinctly categorizes hockey players as “139”, badminton players as “52”, and gentlemen in suits as “132”, showcasing its potent discriminative capabilities." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 287, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 546, + 479 + ], + "type": "text", + "content": "To quantitatively measure the quality of segmentation masks and their corresponding semantic labels, we use Hungarian matching (detailed in Appendix B) to align semantic labels with the category names from the test dataset; for instance, all three sub-clusters depicted in Fig. 4 are assigned to the \"person\" category. The qualitative outcomes post-Hungarian matching are shown in Fig. 5, where our model demonstrates superior panoptic segmentation mask quality. For instance, while the baseline tends to segment parts separately (as seen with the man's head and torso being treated as separate entities in the third row), our model correctly identifies them as parts of a single object. This level of recognition is also evident with the \"trunks of the motorcycle\" example in the second row. For additional results, please see Appendix D. We also present results of the more challenging dataset Cityscapes in Tab. 5 and Fig. 6." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 489, + 414, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 414, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 414, + 502 + ], + "type": "text", + "content": "4.5. Efficient Learning" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 509, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 592 + ], + "type": "text", + "content": "Specifically, for object detection and instance segmentation, we employ our unsupervised instance segmentation model, with cluster count set to 300, to initialize the model weights. We adopt the recipe from [56, 59] for model fine-tuning across various annotation splits. For label-efficient panoptic segmentation, we fine-tune the model initialized with our zero-shot unsupervised framework on the same data splits." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "The results are depicted in Fig. 7, where our model's instance segmentation performance is benchmarked against MoCo-V2, DETReg, and CutLER. Our model consistently surpasses the state-of-the-art with consistent gains in both " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{box}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". In scenarios with panoptic image segmentation as the downstream task, we contrast our results with MoCo-V2 and CutLER in terms of PQ, SQ, and RQ metrics. The results illustrate a remarkable improvement, effectively doubling the performance boost from MoCo-V2 to CutLER, especially in few-shot contexts with limited an-" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22750" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 154, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 154, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 154, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 154, + 185 + ], + "type": "image", + "image_path": "bf83c6c4ce582435bd099f332bf0d1ea509197b3dc0f0c5065a608a7d3e771a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "type": "text", + "content": "Figure 7. We evaluate the label-efficient learning performance on 3 different tasks: object detection (the left), instance segmentation (the second left) and panoptic image segmentation (the last three)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 68, + 249, + 185 + ], + "blocks": [ + { + "bbox": [ + 155, + 68, + 249, + 185 + ], + "lines": [ + { + "bbox": [ + 155, + 68, + 249, + 185 + ], + "spans": [ + { + "bbox": [ + 155, + 68, + 249, + 185 + ], + "type": "image", + "image_path": "db12f59874e9387e28a98187214e676f8c0d4ab543a7c339d0d3c9bb4f9dc8e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 249, + 68, + 344, + 186 + ], + "blocks": [ + { + "bbox": [ + 249, + 68, + 344, + 186 + ], + "lines": [ + { + "bbox": [ + 249, + 68, + 344, + 186 + ], + "spans": [ + { + "bbox": [ + 249, + 68, + 344, + 186 + ], + "type": "image", + "image_path": "00535cc70a101c8c1801a2aec1ff91ad2cffe36ef73ad5c84f9ddc2baae8040e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 344, + 68, + 439, + 186 + ], + "blocks": [ + { + "bbox": [ + 344, + 68, + 439, + 186 + ], + "lines": [ + { + "bbox": [ + 344, + 68, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 344, + 68, + 439, + 186 + ], + "type": "image", + "image_path": "1d4f66e960707b4a7140125fd020a3bb1605ae088a8126076e49089439809a2d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 439, + 68, + 533, + 186 + ], + "blocks": [ + { + "bbox": [ + 439, + 68, + 533, + 186 + ], + "lines": [ + { + "bbox": [ + 439, + 68, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 439, + 68, + 533, + 186 + ], + "type": "image", + "image_path": "ae688957d09ee9443bf9d72e0c9b49b955c3bb8b596608560320bd616347f608.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 222, + 164, + 370 + ], + "blocks": [ + { + "bbox": [ + 65, + 222, + 164, + 370 + ], + "lines": [ + { + "bbox": [ + 65, + 222, + 164, + 370 + ], + "spans": [ + { + "bbox": [ + 65, + 222, + 164, + 370 + ], + "type": "image", + "image_path": "0f7e9dabd66cb0d7ceb2d5564a48a4496f35dad5d521b0bc3866592f8b0b68c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 370, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 415 + ], + "type": "text", + "content": "Figure 8. U2Seg learns features that are more discriminative than those learned by CutLER. The t-SNE [47] visualization of the features from the model's FC layer. We color-code each dot based on its ground-truth category." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 222, + 270, + 369 + ], + "blocks": [ + { + "bbox": [ + 170, + 222, + 270, + 369 + ], + "lines": [ + { + "bbox": [ + 170, + 222, + 270, + 369 + ], + "spans": [ + { + "bbox": [ + 170, + 222, + 270, + 369 + ], + "type": "image", + "image_path": "2ec9a34b1971e2419382e4e33cee5a65505fe30ec262cb03d484d8fe4dbab502.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 424, + 284, + 478 + ], + "blocks": [ + { + "bbox": [ + 50, + 424, + 284, + 478 + ], + "lines": [ + { + "bbox": [ + 50, + 424, + 284, + 478 + ], + "spans": [ + { + "bbox": [ + 50, + 424, + 284, + 478 + ], + "type": "table", + "html": "
# ClusterCOCOUVOVOC
APbox50ARbox100APbox50ARbox100APbox50ARbox100
3009.320.19.822.629.645.7
80011.821.510.825.031.048.0
291113.322.115.125.831.648.3
", + "image_path": "2bdf9351a003bd72e3fe7d9275a5604ceb31e7611e3c328d5359959c772e945e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 532, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 567 + ], + "type": "text", + "content": "notations (1% or 2% labeled samples). This highlights the practical value of our approach in real-world unsupervised learning applications, where annotations are often scarce." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "content": "We attribute the performance gains primarily to the discriminative features our model learns, as in Fig. 8, obtaining effective model initialization for few-shot learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 147, + 623 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 259, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 259, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 259, + 641 + ], + "type": "text", + "content": "In this section, we conduct ablation study on U2Seg." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "Numbers of clusters. The choice of cluster quantity significantly affects the model's representation granularity. Our ablation study on various cluster counts, as detailed in Tab. 7, reveals their impact on model performance. Overclustering generally leads to a finer level of detail, prompting the model to learn more discriminative features." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 220, + 419, + 291 + ], + "blocks": [ + { + "bbox": [ + 46, + 479, + 287, + 511 + ], + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 511 + ], + "type": "text", + "content": "Table 7. Over-clustering can improve the model performance. We show results on different datasets for the unsupervised object detection using different cluster numbers." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 220, + 419, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 419, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 419, + 291 + ], + "type": "table", + "html": "
conf# matchedAPbox50ARbox100
0.910910.913.1
0.722511.618.0
0.628211.819.7
0.438911.821.5
0.251311.321.8
0.07188.618.4
", + "image_path": "e2b45a1f35f4f8863eb1118c8cadc00f2fbc12741896cd7f8b78f3073434006e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 438, + 220, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 293, + 419, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 419, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 419, + 304 + ], + "type": "text", + "content": "(a) Conf's effect on accuracy." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 220, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 438, + 220, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 438, + 220, + 545, + 292 + ], + "type": "table", + "html": "
IoU# matchedAPbox50ARbox100
0.929510.819.7
0.834811.420.7
0.441411.521.6
0.245011.521.1
0.04949.217.7
0.638911.821.5
", + "image_path": "d429384adec8615b954f65a4ee9c07af6b954a65fb76b79cdfe783acaac19722.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 438, + 293, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 438, + 293, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 438, + 293, + 545, + 304 + ], + "type": "text", + "content": "(b) IoU's effect on accuracy." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 306, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 361 + ], + "type": "text", + "content": "Table 8. Impact of Confidence and IoU on Hungarian Matching Performance: The left table illustrates the outcomes at a fixed IoU of 0.6 while varying the confidence scores. Conversely, the right table displays the results with a constant confidence of 0.4, altering the IoU values. The cluster number is 800." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "type": "text", + "content": "Hungarian matching. As our trained model could predict the instance with corresponding semantic labels, we are able to go further beyond unsupervised class-agnostic instance segmentation. To quantitatively evaluate the performance, Hungarian matching is employed to match the predicted semantic labels to the ground-truth dataset categories. See Appendix B for details of the adopted Hungarian matching used in our evaluation. As shown in Tab. 8, the two parameters conf threshold and IoU threshold also affect the precision and recall." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 513, + 372, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 372, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 372, + 527 + ], + "type": "text", + "content": "5. Summary" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 533, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 675 + ], + "type": "text", + "content": "We present U2Seg, a novel Unsupervised Universal Image Segmentation model, adept at performing unsupervised instance, semantic, and panoptic segmentation tasks within a unified framework. Evaluated on extensive benchmarks, U2Seg consistently outperforms previous state-of-the-art methods designed for individual tasks. Additionally, U2Seg achieves the new state-of-the-art for label-efficient panoptic segmentation and instance segmentation. We anticipate that U2Seg, free from the constraints of human annotations, will demonstrate enhanced performance when scaled up with more training data, representing an important direction for our future research." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement: Trevor Darrell, Dantong Niu and Xudong Wang were funded by DoD including DARPA LwLL and the Berkeley AI Research (BAIR) Commons." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22751" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Edward H Adelson. On seeing stuff: the perception of materials by humans and machines. In Human Vision and Electronic Imaging, 2001." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] YM Asano, C Rupprecht, and A Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 192 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "text", + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 286, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 286, + 305 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 308, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 286, + 351 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 353, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 286, + 407 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 497 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 286, + 543 + ], + "type": "text", + "content": "[11] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-DeepLab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 577 + ], + "type": "text", + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "type": "text", + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 655 + ], + "type": "text", + "content": "[14] Jang Hyun Cho, U. Mall, K. Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. ArXiv, abs/2103.17070, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. CoRR, abs/1604.01685, 2016." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13693-13702, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[18] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "text", + "content": "[20] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[21] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[22] Edward W Forgy. Cluster analysis of multivariate data: efficiency versus interpretability of classifications. biometrics, 21:768-769, 1965." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 457 + ], + "type": "text", + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 568 + ], + "type": "text", + "content": "[26] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[29] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. 2023." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22752" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9396-9405, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 218 + ], + "type": "text", + "content": "[33] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "text", + "content": "[36] Hans-Peter Kriegel, Erich Schubert, and Arthur Zimek. The (black) art of runtime evaluation: Are we comparing algorithms or implementations? Knowledge and Information Systems, 52(2):341-378, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "text", + "content": "[37] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. arXiv preprint arXiv:2108.07682, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 555 + ], + "type": "text", + "content": "[41] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 578 + ], + "type": "text", + "content": "[42] Geoffrey J McLachlan and Thriyambakam Krishnan. The EM algorithm and extensions. John Wiley & Sons, 2007." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[43] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "text", + "content": "[44] Sungwon Park, Sungwon Han, Sundong Kim, Danu Kim, Sungkyu Park, Seunghoon Hong, and Meeyoung Cha. Improving unsupervised image clustering with robust learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12278-12287, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[45] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "text", + "content": "[47] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[48] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[49] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. arxiv preprint arxiv:2102.06191, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[50] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "text", + "content": "[51] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[52] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "text", + "content": "[53] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, openworld segmentation. CoRR, abs/2104.04691, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "text", + "content": "[54] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10776-10785, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "text", + "content": "[55] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[56] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[57] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In European Conference on Computer Vision, pages 427-445. Springer, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[58] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 659, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 690 + ], + "type": "text", + "content": "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[60] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22753" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 375 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[61] Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufrey-daz. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. arXiv preprint arXiv:2209.00383, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[62] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3733-3742, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "text", + "content": "[63] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 296 + ], + "type": "text", + "content": "[64] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "text", + "content": "[65] Chengxu Zhuang, Alex Lin Zhai, Daniel Yamins, , et al. Local aggregation for unsupervised learning of visual embeddings. In ICCV, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 375 + ], + "type": "text", + "content": "[66] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22754" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_content_list.json b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2bb5b5b2d2a1243805abc9adaf60f39a63ef3f7d --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_content_list.json @@ -0,0 +1,1473 @@ +[ + { + "type": "text", + "text": "Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training", + "text_level": 1, + "bbox": [ + 124, + 128, + 846, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arun Reddy $^{1,2}$ , William Paul $^{1}$ , Corban Rivera $^{1}$ , Ketul Shah $^{2}$ , Celso M. de Melo $^{3}$ , Rama Chellappa $^{2}$", + "bbox": [ + 94, + 202, + 875, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Johns Hopkins University Applied Physics Laboratory", + "bbox": [ + 320, + 223, + 648, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Johns Hopkins University, Department of Electrical & Computer Engineering", + "bbox": [ + 250, + 239, + 717, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3DEVCOM U.S. Army Research Laboratory", + "bbox": [ + 349, + 257, + 617, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results.", + "bbox": [ + 75, + 340, + 473, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 603, + 209, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the field of video action recognition has undergone significant advancement, largely driven by deep learning-based techniques. These include approaches based on convolutional neural networks (CNNs) [6, 14, 15, 47, 51, 53] and vision transformers (ViTs) [1, 5, 13, 33, 39, 42]. Further enriching this landscape, models integrating video and language data have emerged [36, 55, 57], which effectively leverage large collections of captioned videos to achieve impressive video understanding capabilities.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite increasing performance on benchmark datasets, deep networks for video action recognition suffer from the same fundamental challenges as other machine learning models in the presence of distribution shift. This phenomenon, where a model's performance degrades when applied to data distributions different from those it was trained on, remains a critical obstacle in deploying the models in varied real-world scenarios.", + "bbox": [ + 75, + 765, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain adaptation (DA) seeks to mitigate the effects of", + "bbox": [ + 96, + 885, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/983754bd8c0575bab4a022d6803979f33aa1e57743d7a9d4b106a10e23bf56e0.jpg", + "image_caption": [ + "Figure 1. Overview of the UNITE pipeline for video UDA. A teacher model is used to guide the self-supervised learning process in the first stage, and is again used to improve pseudolabeling of target videos during self-training in the third stage." + ], + "image_footnote": [], + "bbox": [ + 521, + 306, + 874, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distribution shift and has been a major research area within the computer vision community for years [41]. A subproblem within DA is unsupervised domain adaptation (UDA), which tries to make use of unlabeled data from the target domain to improve transfer performance. UDA approaches specific to video have also been developed, with many focusing on explicit alignment of the source and target domains using adversarial techniques or domain discrepancy measures. In contrast, we present here a method for video UDA that relies on self-supervised masked pre-training and self-training, both aided by a powerful image-based teacher model (CLIP [44]). The authors of DALL-V [67] demonstrate the effectiveness of CLIP in the source-free video domain adaptation setting by directly adapting it to action recognition tasks despite fundamentally being an image-based technique. Instead, we leverage CLIP as a spatial teacher for training a spatiotemporal student model that can better operate on target domain videos.", + "bbox": [ + 496, + 595, + 892, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents, to the best of our knowledge, the first exploration of self-supervised masked distillation", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18919", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "techniques in the context of unsupervised video domain adaptation. Our approach, Unsupervised Adaptation with Teacher-Enhanced Learning (UNITE), successfully harnesses the capabilities of an image-based teacher in two ways: (1) to form the target representations in a masked self-supervised pre-training stage on target domain videos, and (2) to improve the quality of pseudolabels used during self-training in collaboration with the student video network (see Figure 1).", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of our work as follows:", + "bbox": [ + 76, + 242, + 447, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the UNITE pipeline for video domain adaptation, which consists of a novel combination of masked video modeling and self-training techniques.", + "- We conduct extensive experimentation by evaluating UNITE on three video domain adaptation benchmarks (Daily-DA, Sports-DA and UCF-HMDB) and observe significant performance gains compared to previously reported results.", + "- We present a series of ablation experiments that study the effectiveness of various aspects of the UNITE pipeline and assess alternative design choices. In particular, we show that masked distillation pre-training and masked self-training are best applied in concert to achieve the strongest domain transfer performance." + ], + "bbox": [ + 76, + 258, + 467, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 482, + 218, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Unsupervised Domain Adaptation (VUDA). Techniques for video unsupervised domain adaptation typically fall within a few common categories [59]. Adversarial methods attempt to align representations between the source and target domains by enforcing a domain confusion loss using a separate domain discriminator network, where the main network must learn to fool the discriminator. Some techniques which were developed for images, such as DANN [18] or MK-MMD [34], can be applied directly to video architectures. $\\mathrm{TA}^3\\mathrm{N}$ [8] is a video-specific approach that uses separate adversaries for the spatial and temporal dimensions. These approaches can result in unstable training due to competing objectives.", + "bbox": [ + 75, + 507, + 468, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another class of methods use contrastive learning and exploit the intrinsic structure of video. CoMix [45] uses contrastive learning with pseudolabeling and video-based augmentations such as sampling at varying frame rates or mixing the backgrounds of source videos and target videos. $\\mathrm{CO}^2\\mathrm{A}$ [52] optimizes 6 different losses simultaneously, including contrastive losses at the clip- and video-level, along with supervised contrastive learning across domains. UDAVT [10] adapts vision transformers using the information bottleneck principle by having separate projectors for source and target features and estimating a cross correlation matrix between features whose labels and pseudolabels match.", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some video domain adaptation techniques have been developed to address the related problem of source-free video video domain adaptation (SFVUDA), which focuses on scenarios where source data is unavailable during the adaptation phase. ATCoN [62] tries to ensure the trained model is as consistent with its features as the original model. EX-TERN [63] additionally uses masking augmentations as a factor for features to be consistent over. DALL-V [67] takes a different approach by adapting a CLIP [44] image encoder to source and target datasets using an adapter.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image & Video Modeling. Masked modeling first achieved success in natural language processing [11], and has since been applied to image and video data as a way to learn powerful feature encoding from unlabeled data. The basic idea behind masked modeling is to partially occlude patches of the input data and train a model to reconstruct the missing patches. This can be done at the pixel-level [16, 19, 31, 50, 54, 58], or the targets of the reconstruction can be formed by a neural network—either an exponential moving average (EMA) of the network being trained [2-4, 24, 32, 70], or even a separate teacher model. MVD [56] uses the latter concept to train a video transformer network using both image and video teachers, which can be viewed as a form of knowledge distillation [21].", + "bbox": [ + 496, + 244, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unmasked Teacher (UMT) [30] uses a different form of masked distillation to train powerful video encoders, where instead of reconstructing representations of masked regions, the student attempts to match the teacher representations of visible patches [66]. UMT is able to use a spatial encoder (CLIP [44]) to train a powerful spatiotemporal model. Because the video encoder only needs to process visible patches, and there is no need for an expensive decoder for reconstruction, UMT offers a highly efficient approach for self-supervision on videos. As such, we choose to leverage the UMT masked distillation objective to perform unsupervised video learning in UNITE.", + "bbox": [ + 496, + 458, + 890, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Modeling for Domain Adaptation. Masked modeling is fundamentally about ensuring a given feature is invariant to whether its corresponding input is masked or not. Inducing such invariances on novel domains can be a useful property to enforce for the purposes of domain adaptation. MAE-TTT [17] finetunes a masked autoencoder [19] on single images at test time, adapting the feature extractor to the statistics of the current image while leaving the prediction head unchanged. MIC [22] uses an EMA teacher model on unmasked images to produce pseudolabels for training on masked target images. PACMAC [43] selects target domain samples for pseudolabeling based on prediction consistency over masked views, and also applies the classification loss to masked target inputs. Since these approaches have been developed for image-based tasks, there is still an open question of how similar concepts could be applied to video recognition.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18920", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 89, + 217, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 114, + 281, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formalize the problem of unsupervised domain adaptation for action recognition as follows. Given a set of labeled videos $\\mathcal{D}_S\\coloneqq \\{(\\mathbf{x}_i^S,y_i^S)\\}_{i = 1}^{N_S}$ drawn from a source data distribution $\\mathcal{P}_S$ , and a set of unlabeled videos $\\mathcal{D}_T\\coloneqq \\{\\mathbf{x}_i^T\\}_{i = 1}^{N_T}$ drawn from a target distribution $\\mathcal{P}_T$ , our objective is to learn a function that can successfully classify videos drawn from $\\mathcal{P}_T$ . More precisely, we seek to learn the parameters $\\theta$ of a function $f_{\\theta}$ , in our case a transformer neural network, that minimize the empirical risk on a sampling of videos from $\\mathcal{P}_T$ . The challenge in domain adaptation lies in the fact that $\\mathcal{P}_S$ and $\\mathcal{P}_T$ differ from one another, leading to a gap in performance when directly applying a source trained model on target domain data. Thus, UDA techniques need to exploit information in unlabeled target domain samples to reduce this gap.", + "bbox": [ + 75, + 138, + 472, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Self-Supervised Initialization", + "text_level": 1, + "bbox": [ + 76, + 375, + 336, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to some other video domain adaptation works [10, 65] that use network weights from Kinetics-400 supervised pre-training, we initialize our network from self-supervised pre-training. As contended in [43] for UDA with images, we argue that supervised pre-training can potentially complicate the study of VUDA techniques. When the pre-trained network has been trained to classify categories present in the DA dataset, some DA techniques could perform well simply by preserving the capabilities of the pretrained model despite not generalizing well to DA tasks that do not have category overlap with the pre-training dataset. Moreover, the UDA problem formulation implies that labeled instances of relevant classes are only available from the source domain. This condition does not hold for the Daily-DA and Sports-DA benchmarks that we use for evaluation, as 6 out of their 8 classes also appear in the Kinetics-400 dataset. Further exacerbating the issue with supervised Kinetics pre-training is the fact that Kinetics serves as a target domain in both benchmarks. Thus, we initialize the network weights from UMT [30] single-modality (i.e., video only) self-supervised pre-training on Kinetics-710 [29]. Please refer to Sec. 10 in the Appendix for select results using a UMT pre-trained network with additional supervised fine-tuning on Kinetics-400, which unsurprisingly achieves better baseline performance.", + "bbox": [ + 75, + 398, + 472, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 791, + 168, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. UNITE", + "text_level": 1, + "bbox": [ + 76, + 816, + 171, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach to VUDA consists of three stages: (1) unsupervised pre-training on target domain data, (2) supervised fine-tuning on source domain data and (3) collaborative self-training using videos from both domains. We now", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/25629477d8c1a7622f021e15ce30078ede1bf0cfa12500d298519ff3a12d09c8.jpg", + "image_caption": [ + "Figure 2. Unmasked Teacher (UMT) training used in Stage 1 of UNITE to perform unsupervised representation learning on target domain videos." + ], + "image_footnote": [], + "bbox": [ + 509, + 90, + 883, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "describe each stage in detail.", + "bbox": [ + 500, + 388, + 692, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 1: Unsupervised Target Domain Pre-Training.", + "text_level": 1, + "bbox": [ + 498, + 417, + 867, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the extraction of discriminative features from target domain videos, the UNITE pipeline begins by performing additional unsupervised pre-training on the target dataset $\\mathcal{D}_T$ . For this, we adopt the UMT [30] self-supervised training objective. UMT training uses a pretrained and frozen spatial teacher model $g_{*}$ to train a spatiotemporal student model $g_{a}$ by enforcing alignment between feature representations of the two networks at multiple layers (depicted in Figure 2). A key aspect of the training process is that the student model takes as input a masked video view $m(\\mathbf{x}^T)$ , while the teacher model processes un-masked video frames from $\\mathbf{x}^T$ . Through UMT training, $g_{a}$ must learn to match the semantically rich (albeit only spatially informed) features of $g_{*}$ from heavily masked target domain videos. To ensure sampling of meaningful video patches, UMT performs an attention-guided masking operation $m(\\cdot)$ drawn from a multinomial distribution defined by the attention map of the final layer CLS token in $g_{*}$ . Attention-guided masking allows the use of a high masking ratio, which makes UMT training extremely efficient. We formalize the Stage 1 objective as follows. Let $\\mathbf{z}_a^l$ denote the L2 normalized $l$ -th layer representation of $g_{a}$ for a masked input $m(\\mathbf{x})$ , and let $\\mathbf{z}_{*}^{l}$ denote the normalized $l$ -th layer representation of $g_{*}$ for the full input $\\mathbf{x}$ , but only at the locations corresponding to the visible patches in $m(\\mathbf{x})$ . The loss function in Stage 1 can be represented as:", + "bbox": [ + 496, + 434, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U M T}} = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {P} _ {T}} \\left[ \\frac {1}{| \\mathcal {A} |} \\sum_ {l \\in \\mathcal {A}} \\operatorname {M S E} \\left(d ^ {l} (\\mathbf {z} _ {a} ^ {l}), \\mathbf {z} _ {*} ^ {l}\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 833, + 846, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ is a set of aligned layers, $d^{l}(\\cdot)$ is a linear projection", + "bbox": [ + 500, + 883, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a022215bbdb0dc2965c75f00e905c5a975574e999a33989630ad220214d22dd.jpg", + "image_caption": [ + "Figure 3. Overview of the collaborative self-training stage in UNITE. The student and teacher models work together to produce more accurate pseudolabels for target domain videos. The target domain classification loss is enforced on masked target videos to encourage stronger context learning. Source domain classification is included to stabilize training, which is especially important at the beginning of training when pseudolabels may have low accuracy." + ], + "image_footnote": [], + "bbox": [ + 199, + 87, + 772, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the $l$ -th layer and MSE indicates mean squared error. Following the original UMT method, we leverage a CLIP [44] ViT-B image encoder as $g_{*}$ . We keep the standard UMT masking ratio of $r = 0.8$ and perform alignment between the last 6 layers of the networks. $g_{a}$ is initialized as described in Sec. 3.2, along with the projections $d^{l}(\\cdot)$ , which are kept frozen throughout Stage 1.", + "bbox": [ + 75, + 439, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: Source Domain Fine-Tuning. The second stage of UNITE fits a classifier on source domain data using $g_{a}$ from Stage 1. We introduce label space information in a separate stage from target domain self-supervision because, as the authors of [17] observe for MAE training, we find that imposing a classification loss and UMT loss simultaneously can hinder convergence. To perform supervised fine-tuning, we introduce a linear classification head $h_{a}$ that operates on the mean pooled outputs of the final transformer layer in $g_{a}$ . The full student classifier network can now be represented as $f_{a} = h_{a}(g_{a}(\\cdot))$ . As is the case with other masked pretraining strategies, we find that representations learned via UMT pre-training are not linearly separable. Thus, we perform full fine-tuning of $f_{a}$ by minimizing a cross entropy (CE) loss $\\mathcal{L}_{CE}$ on unmasked source domain videos. The Stage 2 loss function is represented as:", + "bbox": [ + 75, + 560, + 470, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S F T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 815, + 468, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To help preserve target domain feature extraction learned in Stage 1, we apply a layer-wise learning rate decay when fine-tuning on source domain data.", + "bbox": [ + 75, + 840, + 468, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 3: Collaborative Self-Training (CST). After enhancing the ability of $g_{a}$ to extract meaningful features in the target domain (Stage 1) and fitting $f_{a}$ for the classification task using source videos (Stage 2), we further adapt $f_{a}$ to target data using a self-training process (depicted in Figure 3). In a conventional self-training framework (denoted as $\\mathcal{L}_{\\mathrm{ST}}$ in Eq. (2)), the overall loss is a sum of losses on labeled source domain samples and select target domain samples, chosen using mask $s(\\cdot)$ . The target domain loss leverages the model's own predictions $\\hat{y}$ (where $\\hat{y} = \\operatorname{argmax}_c \\sigma(f(\\mathbf{x}^T))[c]$ and $\\sigma(\\cdot)$ denotes the softmax function) in lieu of labels in the CE loss computation. Various strategies have been used to form the selection mask $s(\\cdot)$ , with the goal of choosing target samples whose predictions are more likely to be correct [43, 48, 68].", + "bbox": [ + 496, + 439, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] + \\tag {2} \\\\ \\lambda \\mathbb {E} _ {\\mathbf {x} ^ {T} \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} (f _ {a} (\\mathbf {x} ^ {T}), \\hat {y} _ {a} (\\mathbf {x} ^ {T})) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 686, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a modified form of self-training, which we refer to as collaborative self-training (CST), that makes use of the student-teacher setup from Stage 1 to improve the process of pseudolabel estimation. Specifically, we use the zero-shot classification capabilities of the CLIP image teacher model $f_{*}$ (see Sec. 5.2 for details), in combination with predictions from the video student model $f_{a}$ , to create refined target pseudolabels for self-training. We employ the MatchOrConf scheme proposed in [69] to combine the outputs of the two models as shown in Eq. (3), where $\\mathrm{Conf}(\\hat{y})$ denotes the maximum softmax probability", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18922", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[20], $\\max_c\\sigma (f(\\mathbf{x}^T))[c]$ , to create pseudolabel $\\tilde{y}$ :", + "bbox": [ + 76, + 90, + 403, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = \\left\\{ \\begin{array}{l l} \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} = \\hat {y} _ {*} \\\\ \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) > \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) \\leq \\gamma , \\\\ \\hat {y} _ {*} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) \\leq \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) > \\gamma , \\\\ - 1 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 130, + 485, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{x}^T$ is an argument of the above, but is omitted for clarity. The MatchOrConf scheme selects only those target samples where there is agreement between the predictions of the two models, or where one model's confidence exceeds threshold $\\gamma$ while the other's does not, resulting in the following selection mask:", + "bbox": [ + 75, + 215, + 468, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(\\mathbf {x} ^ {T}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {y} \\left(\\mathbf {x} ^ {T}\\right) \\neq - 1, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 318, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking inspiration from [22] and [43], we compute the target domain loss on masked videos, where we employ the same teacher-guided attention masking strategy (with $r = 0.8$ ) used during UMT pre-training in Stage 1. We find that training on masked videos in this stage results in substantial performance boosts compared to training on un-masked videos (see Table 6). Like [22], we include a loss weighting $q$ for each target video, based on the student network confidence:", + "bbox": [ + 75, + 362, + 472, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} ^ {T}\\right) = \\operatorname {C o n f} \\left(\\hat {y} _ {a} \\left(\\mathbf {x} ^ {T}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 512, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete loss function for the collaborative self-training stage can be expressed as:", + "bbox": [ + 75, + 536, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} (\\mathbf {x} ^ {S}), y ^ {S}\\right) \\right] + \\\\ \\lambda \\mathbb {E} _ {(\\mathbf {x} ^ {T}) \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) q (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} \\left(f _ {a} (m (\\mathbf {x} ^ {T})), \\tilde {y} (\\mathbf {x} ^ {T}) \\right] \\right. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 587, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 679, + 179, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approach on three video domain adaptation benchmarks: Daily-DA, Sports-DA, UCF $\\leftrightarrow$ HMDBfull.", + "bbox": [ + 75, + 700, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Daily-DA [64] is a recently introduced benchmark composed of videos from four domains: ARID (A) [60], HMDB51 (H) [28], Moments-in-Time (M) [38] and Kinetics (K) [7, 26]. It consists of 8 overlapping classes representing everyday actions, with a total volume of approximately 19K videos. Videos from the ARID domain present a unique challenge in this benchmark, as they were explicitly filmed under low-illumination conditions.", + "- Sports-DA [64] contains videos from three domains: Sports-1M (S) [25], UCF101 (U) [49] and Kinetics (K) [7, 26]. It includes 23 overlapping categories depicting" + ], + "bbox": [ + 76, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "various sporting activities, with a total volume of approximately 41K videos.", + "- $\\mathbf{UCF} \\leftrightarrow \\mathbf{HMDB}_{full}$ [8] contains approximately 3.2K videos across 12 action categories taken from the UCF101 (U) [49] and HMDB51 (K) [28] datasets." + ], + "bbox": [ + 500, + 90, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 203, + 718, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frame Sampling. During training, we perform random uniform frame sampling [53] on each video. We divide the video into $T = 8$ uniform segments and randomly sample one frame from each. All frames are resized to $224 \\times 224$ pixels, resulting in a $3 \\times 8 \\times 224 \\times 224$ tensor for each video. During testing, we average the network predictions across 3 spatial crops and 4 temporal clips, where clip $i$ is formed by splitting each of the $T$ segments into 4 sub-segments and sampling a frame from the $i$ -th sub-segment of each.", + "bbox": [ + 496, + 232, + 893, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture & Training. All of our experiments use a standard ViT-B/16 [12] architecture with a fixed sine and cosine positional encoding scheme. Our patch embedding projection operates on each frame independently, and self-attention is performed across both the space and time dimensions. All classification is performed using mean pooling of the final layer patch representations. We conduct our experiments using the PyTorch deep learning library [40] on 4 NVIDIA A5000 GPUs.", + "bbox": [ + 496, + 388, + 893, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During Stage 1, we perform UMT training for 50 epochs on the target dataset using a batch size of 256 and a learning rate of $1.5\\mathrm{e - }5$ . During Stage 2, we use a batch size of 28, a learning rate of $2.5\\mathrm{e - }5$ , and apply a layer-wise learning rate decay of 0.65. During stage 3, we use a batch size of 40 (20 videos from each domain), a learning rate of 1e-5, a MatchOrConf confidence threshold of $\\gamma = 0.1$ and a target domain loss weight $\\lambda = 1$ . Training in all stages uses the AdamW optimizer [35] and a weight decay of 0.05. Please refer to Sec. 8 in the Appendix for more training details for each stage.", + "bbox": [ + 496, + 532, + 893, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-Shot Classification with CLIP. In Stage 3 of UNITE, we use CLIP (ViT-B/16) in a zero-shot manner to refine the pseudolabels of target domain videos. We do so by passing each of the $T$ video frames through the CLIP image encoder and computing the cosine similarity with a set of text representations produced by the CLIP text encoder. The inputs to the text encoder are the class names associated with the particular action recognition task, embedded in the following template: \"A video of a person {class}.\" To form a video-level prediction, we simply average the softmax of the cosine similarity scores across the video frames. Refer to the Appendix (Sec. 11) for a listing of class names used for each task.", + "bbox": [ + 496, + 703, + 893, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18923", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2218beb92ffdae5226b40da60ccf651199043112fcd56b3462b2fb92e757c5ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTarget Domain Accuracy (Top-1%)
H→AM→AK→AA→HM→HK→HH→MA→MK→MM→KH→KA→KAvg.
DALL-V [67] LB17.534.715.614.644.647.925.515.535.761.645.117.831.3
Source Only40.452.136.549.668.357.941.536.343.379.348.041.749.6
ZSCLIP (ViT-B/16) [44]36.536.536.560.060.060.048.548.548.568.168.168.153.3
SFVUDAATCoN [62]17.927.217.226.747.348.230.717.232.557.748.531.033.5
EXTERNAL [63]26.218.123.926.253.755.840.718.235.268.157.651.439.6
DALL-V [67]24.024.024.057.965.452.547.045.747.078.176.775.051.4
UDADANN [18]14.222.821.220.143.337.529.519.721.758.838.227.029.5
MK-MMD [34]20.321.021.718.750.436.225.718.024.058.533.826.129.5
TA3N [8]14.421.619.914.943.037.725.715.631.555.538.423.428.5
UNITE w/o CST43.846.734.751.770.854.644.339.043.378.151.750.850.8
UNITE (Ours)48.044.137.567.974.265.851.850.048.089.969.963.659.2
DALL-V [67] UB26.926.926.970.470.470.461.561.561.588.988.988.961.9
Target Only68.568.568.584.684.684.673.073.073.098.398.398.381.1
", + "bbox": [ + 81, + 90, + 893, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00542bc2118f001c75a7746517a7c4e4c30ca134334a4385375f8093086439ee.jpg", + "table_caption": [ + "Table 1. UDA results on Daily-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]." + ], + "table_footnote": [], + "table_body": "
MethodTarget Domain Accuracy (Top-1 %)
U→SK→SS→UK→UU→KS→KAvg.
DALL-V [67] LB64.379.584.485.467.278.276.5
Source Only67.686.897.998.979.989.186.7
ZSCLIP (ViT-B/16) [44]85.085.093.393.391.391.389.9
SFVUDAATCoN [62]47.969.790.693.665.276.073.8
EXTERN [63]72.773.895.493.781.282.283.2
DALL-V [67]75.977.788.888.081.282.382.3
UDADANN [18]55.175.085.788.065.973.473.8
MK-MMD [34]55.667.990.990.266.173.674.0
TA3N [8]54.168.693.090.363.672.673.7
UNITE w/o CST73.987.398.299.284.690.288.9
UNITE (Ours)86.090.298.799.894.295.394.0
DALL-V [67] UB88.388.393.493.485.685.689.1
Target Only97.997.998.898.899.999.998.9
", + "bbox": [ + 243, + 347, + 733, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. UDA results on Sports-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61].", + "bbox": [ + 86, + 577, + 880, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Baselines", + "text_level": 1, + "bbox": [ + 76, + 618, + 184, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare results using UNITE with previously developed techniques in UDA, as well as in the closely related setting of SFUDA as reported by [67]. For Daily-DA and Sports-DA we compare against DANN [18], MK-MMD [34] and $\\mathrm{TA}^3\\mathrm{N}$ [8] in the UDA setting, and ATCoN [62], EXTERN [63] and DALL-V [67] for SFUDA. For $UCF\\leftrightarrow HMDB_{full}$ , we additionally report accuracies from $\\mathrm{CO}^2\\mathrm{A}$ [52] and UDAVT [10].", + "bbox": [ + 75, + 641, + 470, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Benchmark Results", + "text_level": 1, + "bbox": [ + 76, + 771, + 264, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 1, 2 and 3 show the results of our evaluation of UNITE on Daily-DA, Sports-DA and $UCF \\leftrightarrow HMDB_{full}$ . We include accuracies after Stage 2 of the pipeline (denoted as UNITE w/o CST) as well as after applying the full UNITE process. For reference, we include accuracies using our UMT pre-trained network with source domain supervised fine-tuning (Source Only) and with target domain", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "supervised fine-tuning (Target Only). We also include the source only and target only accuracies (denoted in the tables as LB and UB) of DALL-V [67] (a CLIP-based ResNet-50 model) for added context. We additionally display the zero-shot classification accuracy of the CLIP teacher model that we utilize during Stage 1 and Stage 3 of UNITE.", + "bbox": [ + 496, + 618, + 890, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An observation we make across all benchmarks is the enhanced baseline accuracy of the UMT pre-trained model, which speaks to the strength of the unsupervised masked distillation process and the ViT architecture. On the Daily-DA benchmark, we find that UNITE exceeds previously reported results on most domain shifts. On only one shift $(\\mathrm{M}\\rightarrow \\mathrm{A})$ we observe degradation in target accuracy after employing UNITE compared to source only, which is consistent with previous work.", + "bbox": [ + 496, + 714, + 890, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UNITE exhibits even stronger performance on the larger-scale Sports-DA benchmark, where it exceeds previous results by a large margin. We also see more consistent and", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18924", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ed99a58f6d2ac638770e2aa248076689cb3bb78b7cccc3432afc480c9cbe51ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
H→UU→HAvg.
SDALL-V [67] LB71.676.173.8
Source Only88.880.084.4
CLIP (ViT-B/16) [44]88.891.790.3
SEVUDAATCoN [62]85.379.782.5
EXTERN [63]91.988.990.4
DALL-V [67]93.188.991.0
UDADANN [18]74.475.174.8
MK-MMD [34]74.779.777.2
TA3N [8]78.184.881.5
CO2A [52]95.887.891.8
UDAVT [10]96.892.394.6
UNITE w/o CST92.183.687.9
UNITE (Ours)92.595.093.8
DALL-V [67] UB93.791.492.6
Target Only99.796.798.2
", + "bbox": [ + 135, + 88, + 415, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f5d2d4393376e31c3a9748c11b00f54c5a083a9fe8cd79fbfd4c40496b10d06.jpg", + "image_caption": [ + "Figure 4. Class-wise accuracy on ARID $\\rightarrow$ HMDB. Performances are shown for the CLIP teacher model and the student model before and after the collaborative self-training stage (Stage 3). For most classes, the Stage 3 student exceeds the accuracy of both the Stage 2 student and the CLIP teacher." + ], + "image_footnote": [], + "bbox": [ + 99, + 356, + 446, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "substantial improvement after the UMT target pre-training stage alone (UNITE w/o CST), suggesting that overtraining may be occurring on the smaller datasets in Daily-DA.", + "bbox": [ + 75, + 664, + 468, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On $UCF \\leftrightarrow HMDB_{full}$ we see state-of-the-art performance on $\\mathrm{U} \\rightarrow \\mathrm{H}$ , while UDAVT outperforms on $\\mathrm{H} \\rightarrow \\mathrm{U}$ . Nevertheless, UNITE still achieves strong overall performance on the benchmark.", + "bbox": [ + 75, + 710, + 468, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Additional Analysis & Discussion", + "text_level": 1, + "bbox": [ + 76, + 784, + 380, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Student exceeds the teacher. We see in the benchmark results that the video student model exceeds the zero-shot accuracy of the CLIP teacher in virtually all domain shifts. This suggests that UNITE successfully leverages the spatial modeling capabilities of CLIP to train an even more powerful spatiotemporal model for the action recognition task.", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87e0c8b77b64aaa10e585cd44f8da41bac04956b0248322fb8c148132c844ec1.jpg", + "table_caption": [ + "Table 3. UDA results on $UCF \\leftrightarrow HMDB_{full}$ . Rows in color use our UMT backbone, and rows without color are reported from [61]." + ], + "table_footnote": [], + "table_body": "
MethodPTCSTAccuracy (%)
H→AA→H
Source OnlyXX40.449.6
+ Target UMT Pre-Training✓X43.851.7
+ Collaborative Self-TrainingX✓42.060.8
UNITE✓✓48.067.9
", + "bbox": [ + 514, + 88, + 883, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of stages in UNITE on ARID ↔ HMDB from Daily-DA. PT and CST denote UMT pre-training and collaborative self-training, respectively.", + "bbox": [ + 498, + 189, + 890, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can see from Figure 4 that the teacher model and the student model prior to collaborative self-training have differing capabilities. For example, the image teacher is stronger at recognizing actions from the jump class than the Stage 2 student. On the other hand, the video student is significantly better than the teacher at recognizing wave at the start of CST. In both cases, we see that the student model, through collaborative self-training, comes to exceeds both the Stage 2 student and the CLIP image teacher. With the MatchOr-Conf pseudolabeling scheme, the two models are able to work together to achieve higher accuracy in the target domain. We see accuracy on the pick and run classes declines after CST, as self-training tends to reinforce the accuracy of well-performing classes, sometimes at the expense of more difficult ones that are infrequently assigned pseudolabels. Modifying CST to account for the learning status of each class, like in [68], may help remedy this issue.", + "bbox": [ + 496, + 250, + 892, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Stage 1 and 3. In Table 4 we ablate the stages in UNITE. While UMT pre-training and collaborative self-training each offer a benefit in the absence of the other, the largest gains in domain transfer performance are obtained when combining the two stages. For example, on the ARID $\\rightarrow$ HMDB domain shift in Daily-DA, Stage 1 and Stage 3 alone improve upon source only accuracy by $+1.8\\%$ and $+11.2\\%$ respectively. Meanwhile, the full UNITE pipeline results in $+18.3\\%$ accuracy improvement on the target domain, suggesting that the masked pre-training and masked self-training processes complement one another.", + "bbox": [ + 496, + 508, + 892, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of data domains during pre-training. In Table 5, we assess the impact of utilizing each of the two data domains during UMT pre-training on UDA performance. While the results vary across domain shifts, we observe the most consistent improvements when pre-training only on target domain videos. Although incorporating both domains during unsupervised pre-training has been successfully applied in image-based domain adaptation [27, 37, 43, 46], we find it to be suboptimal with UMT training on the VUDA tasks studied in this work.", + "bbox": [ + 496, + 674, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of masking during self-training. In Table 6, we see that enforcing the target domain CE loss on masked videos results in increased performance compared to using unmasked target videos, with smaller benefits for $\\mathrm{H}\\rightarrow \\mathrm{A}$ and much larger benefits for $\\mathrm{A}\\rightarrow \\mathrm{H}$ . We believe the benefits of", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f79697ed0e83c55de463a669435ff42f6f608a1e2cd304e36a18639fd13dbd33.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UMT Pre-Training DataAccuracy (%)
H→AA→H
None42.060.8
Source Only48.360.0
Target Only48.067.9
Source + Target45.267.1
", + "bbox": [ + 148, + 88, + 403, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/abe7b1463b7a026169250ceca59ef8480260846adb7207ca9d90f50dd6a30c9d.jpg", + "table_caption": [ + "Table 5. Comparison of data domains used during UMT pretraining. Accuracies are reported after the self-training stage. Bold indicates best accuracy and underline indicates second best." + ], + "table_footnote": [], + "table_body": "
Data for Target CE LossAccuracy (%)
H→AA→H
Unmasked46.654.6
Masked48.067.9
", + "bbox": [ + 145, + 227, + 406, + 291 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f142c2fb6f07dbfc6dcfc7ba37ef6a8e28920d6ebb23f6742e4684dc17323565.jpg", + "table_caption": [ + "Table 6. Effect of applying the target domain cross entropy loss on masked vs. unmasked target videos during the self-training stage." + ], + "table_footnote": [], + "table_body": "
Pseudolabeling StrategyAccuracy (%)
H→AA→H
Cons (k=2)39.857.5
ConsOrConf (k=2,T=0.5) [43]48.656.3
ConsAndConf (k=2,T=0.5)40.157.1
Zero-Shot CLIP Only (T=0.5)34.751.3
MatchOrConf (γ=0.1) [69]48.067.9
MatchAndConsOrConf (k=2,γ=0.1)47.460.8
", + "bbox": [ + 109, + 329, + 441, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "masking arise from forcing the model to use different cues from the same video to predict the pseudolabel, resulting in more robust target domain recognition. As we saw in Table 4, the masked pre-training stage plays a key role in unlocking the benefits of masked self-training.", + "bbox": [ + 75, + 477, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of pseudolabeling strategies. In Table 7, we explore the use of alternative pseudolabeling strategies to the MatchOrConf [69] scheme employed in UNITE. The first 3 rows of the table assess variations of the masked consistency pseudolabeling scheme proposed in PACMAC [43], where target samples are selected if there is agreement between predictions of $k$ masked views. This criterion can also be combined with a constraint on the confidence exceeding threshold $T$ . While we find the combination of these two constraints (ConsOrConf) to be effective for self-training, we find that MatchOrConf results in more consistent target accuracy improvements across domain shifts. We also consider combining MatchOrConf with a masked consistency constraint as a way to reduce cases where the student and teacher models agree on an incorrect prediction. This scheme, denoted as MatchAndConsOrConf, adds an additional constraint to the matching criterion by also requiring agreement between the student model predictions of $k$ masked views. We find that this approach underperforms compared to the vanilla MatchOrConf scheme due to low utilization of target domain samples. Our results also indicate that using both models together for pseudolableing outperforms using only zero-shot CLIP predictions.", + "bbox": [ + 75, + 553, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6356fb991535b06aa4414bb6b72af70e349fa90e3834627557bb1319413b68f.jpg", + "table_caption": [ + "Table 7. Comparison of pseudolabeling strategies." + ], + "table_footnote": [], + "table_body": "
Loss During Self-TrainingAccuracy (%)
H→AA→H
LCE(m(xT), y)33.742.1
LCE(m(xT), y) + LCE(xS, yS)48.067.9
", + "bbox": [ + 558, + 88, + 836, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ee0a8ff0c68464e69536666660c4215a0b7f2e1c39b640e9004adeb18fa446e3.jpg", + "table_caption": [ + "Table 8. Effect of including a source domain classification loss in addition to the target domain loss during the self-training stage." + ], + "table_footnote": [], + "table_body": "
Confidence ThresholdAccuracy (%)
H→AA→H
0.148.067.9
0.344.155.4
0.551.964.6
0.742.665.8
0.940.762.5
", + "bbox": [ + 581, + 186, + 815, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Effect of confidence threshold $\\gamma$ in MatchOrConf pseudolabeling scheme from Equation 3.", + "bbox": [ + 498, + 285, + 890, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of source classification during self-training. In Table 8, we study the importance of including a source domain classification loss while performing collaborative self-training on masked target videos. Indeed, we find that source classification is a crucial ingredient in this stage, as excluding it destabilizes training and results in performance degradation rather than improvement. We suspect that the ground truth labels on source data help to combat the effects of inaccurate target pseudolabels, which may be prevalent especially at the start of training.", + "bbox": [ + 496, + 329, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of pseudolabeling confidence threshold. In Table 9, we compare adaptation performance for various values of confidence threshold in the MatchOrConf pseudolabeling scheme. We observe significant variation across domain shifts, but find that $\\gamma = 0.1$ leads to relatively consistent improvement. We leave for future work the development of a more principled approach to determining an appropriate confidence threshold for a given domain shift.", + "bbox": [ + 496, + 479, + 890, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 614, + 625, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented UNITE, a three step process for unsupervised video domain adaptation. Our approach leverages a powerful spatial encoder to aid in the adaptation of a spatiotemporal network using unlabeled target domain videos—making significant progress towards bridging the domain gaps in various VUDA datasets. The advancements demonstrated by UNITE underscore the untapped potential of masked modeling techniques for video domain adaptation, which we hope will spur further research in this area.", + "bbox": [ + 496, + 640, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was sponsored by the Army Research Laboratory under Cooperative Agreement W911NF-21-2-0211. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.", + "bbox": [ + 496, + 787, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18926", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 1", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked Siamese Networks for Label-Efficient Learning. In ECCV, 2022. 2", + "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. In CVPR, 2023.", + "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In ICML, 2022. 2", + "[5] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In ICML, pages 813-824, 2021. 1", + "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In CVPR, 2017. 1", + "[7] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A Short Note about Kinetics-600, 2018. 5", + "[8] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal Attentive Alignment for Large-Scale Video Domain Adaptation. In ICCV, 2019. 2, 5, 6, 7", + "[9] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 iee. In CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 1", + "[10] Victor G. Turrisi Da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Unsupervised Domain Adaptation for Video Transformers in Action Recognition. In ICPR, pages 1258-1265, Montreal, QC, Canada, 2022. 2, 3, 6, 7", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5", + "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale Vision Transformers. In ICCV, 2021. 1", + "[14] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In CVPR, 2020. 1" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In ICCV, 2019. 1", + "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In NeurIPS, 2022. 2", + "[17] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 4", + "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 6, 7", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2021. 2", + "[20] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 5", + "[21] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 2", + "[22] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation. In CVPR, pages 11721–11732, Vancouver, BC, Canada, 2023. 2, 5", + "[23] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661. Springer, 2016. 1", + "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to Hide from Your Students: Attention-Guided Masked Image Modeling. In ECCV, pages 300-318, 2022. 2", + "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-Scale Video Classification with Convolutional Neural Networks. In CVPR, pages 1725–1732, 2014. 5", + "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5", + "[27] Donghyun Kim, Kuniaki Saito, Tae-Hyun Oh, Bryan A Plummer, Stan Sclaroff, and Kate Saenko. Cds: Cross-domain self-supervised pre-training. In ICCV, pages 9123-9132, 2021. 7", + "[28] H Kuehne, H Jhuang, E Garrote, T Poggio, and T Serre. HMDB: A Large Video Database for Human Motion Recognition. In ICCV, 2011. 5", + "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In ICCV, pages 1632-1643, 2023. 3", + "[30] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked Teacher: Towards Training-Efficient Video Foundation Models. In ICCV, 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. MST: Masked Self-Supervised Transformer for Visual Representation. In NeurIPS, pages 13165-13176, 2021. 2", + "[32] Han Lin, Guangxing Han, Jiawei Ma, Shiyuan Huang, Xudong Lin, and Shih-Fu Chang. Supervised masked knowledge distillation for few-shot transformers. In CVPR, pages 19649-19659, 2023. 2", + "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3192-3201, New Orleans, LA, USA, 2022. 1", + "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 2, 6, 7", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5, 1", + "[36] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-CLIP: End-to-End Multi-grained Contrastive Learning for Video-Text Retrieval. In ACM MM, 2022. 1", + "[37] Samarth Mishra, Kate Saenko, and Venkatesh Saligrama. Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency, 2021. 7", + "[38] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, and Dan Gutfreund. Moments in Time Dataset: One Million Videos for Event Understanding. PAMI, 42(2):502-508, 2020. 5", + "[39] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video Transformer Network. In ICCV, 2021. 1", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5", + "[41] Vishal M Patel, Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. Visual Domain Adaptation: A survey of recent advances. IEEE Signal Processing Magazine, 32(3):53-69, 2015. 1", + "[42] Aj Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning. In CVPR, pages 2214-2224, Vancouver, BC, Canada, 2023. 1", + "[43] Viraj Prabhu, Sriram Yenamandra, Aaditya Singh, and Judy Hoffman. Adapting self-supervised vision transformers by probing attention-conditioned masking consistency. NeurIPS, 35:23271-23283, 2022. 2, 3, 4, 5, 7, 8, 1", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 4, 6, 7", + "[45] Aadarsh Sahoo, Rutav Shah, R. Panda, Kate Saenko, and Abir Das. Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing. In NeurIPS, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Kendrick Shen, Robbie M. Jones, Ananya Kumar, Sang Michael Xie, Jeff Z. Haochen, Tengyu Ma, and Percy Liang. Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation. In ICML, pages 19847-19878, 2022. 7", + "[47] Karen Simonyan and Andrew Zisserman. Two-Stream Convolutional Networks for Action Recognition in Videos. In NeurIPS, 2014. 1", + "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In NeurIPS, pages 596-608, 2020. 4", + "[49] K. Soomro, A. Zamir, and M. Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. ArXiv, 2012. 5", + "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In NeurIPS, 2022. 2", + "[51] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning Spatiotemporal Features with 3D Convolutional Networks. In ICCV, pages 4489-4497, Santiago, Chile, 2015. 1", + "[52] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-Head Contrastive Domain Adaptation for Video Action Recognition. In WACV, pages 2234–2243, 2022. 2, 6, 7", + "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal Segment Networks for Action Recognition in Videos. PAMI, 2017. 1, 5", + "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. In CVPR, pages 14549-14560, Vancouver, BC, Canada, 2023. 2", + "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition, 2021. 1", + "[56] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning. In CVPR, pages 6312-6322, Vancouver, BC, Canada, 2023. 2", + "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. InternVideo: General Video Foundation Models via Generative and Discriminative Learning, 2022. 1", + "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18928", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Yuecong Xu, Haozhi Cao, Zhenghua Chen, Xiaoli Li, Lihua Xie, and Jianfei Yang. Video Unsupervised Domain Adaptation with Deep Learning: A Comprehensive Survey, 2022. 2", + "[60] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. ARID: A New Dataset for Recognizing Action in the Dark. In Communications in Computer and Information Science, 2022. 5", + "[61] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Aligning Correlation Information for Domain Adaptation in Action Recognition. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. 6, 7", + "[62] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-Free Video Domain Adaptation by Learning Temporal Consistency for Action Recognition. In ECCV, pages 147-164, Cham, 2022. 2, 6, 7", + "[63] Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Extern: Leveraging endo-temporal regularization for black-box video domain adaptation. arXiv preprint arXiv:2208.05187, 2022. 2, 6, 7", + "[64] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Zhengguo Li, and Zhenghua Chen. Multi-Source Video Domain Adaptation With Temporal Attentive Moment Alignment Network. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3860-3871, 2023. 5", + "[65] Yuecong Xu, Jianfei Yang, Yunjiao Zhou, Zhenghua Chen, Min Wu, and Xiaoli Li. Augmenting and Aligning Snippets for Few-Shot Video Domain Adaptation. In ICCV, 2023. 3", + "[66] Hongwei Xue, Peng Gao, Hongyang Li, Yu Qiao, Hao Sun, Houqiang Li, and Jiebo Luo. Stare at What You See: Masked Image Modeling without Reconstruction. In CVPR, pages 22732-22741, Vancouver, BC, Canada, 2023. 2", + "[67] Giacomo Zara, Alessandro Conti, Subhankar Roy, Stephane Lathuiliere, Paolo Rota, and Elisa Ricci. The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation. In ICCV, 2023. 1, 2, 6, 7", + "[68] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 34:18408-18419, 2021. 4, 7", + "[69] Wenyu Zhang, Li Shen, and Chuan-Sheng Foo. Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation. In CVPR, 2023. 4, 8", + "[70] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT Pre-training with Online Tokenizer. In ICLR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18929", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_model.json b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4acb54430fd1878ef55615a14ae46c49a9a3d2e --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.848, + 0.177 + ], + "angle": 0, + "content": "Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.203, + 0.877, + 0.223 + ], + "angle": 0, + "content": "Arun Reddy\\(^{1,2}\\), William Paul\\(^{1}\\), Corban Rivera\\(^{1}\\), Ketul Shah\\(^{2}\\), Celso M. de Melo\\(^{3}\\), Rama Chellappa\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.224, + 0.65, + 0.239 + ], + "angle": 0, + "content": "1 Johns Hopkins University Applied Physics Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.241, + 0.718, + 0.256 + ], + "angle": 0, + "content": "\\(^{2}\\)Johns Hopkins University, Department of Electrical & Computer Engineering" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.258, + 0.619, + 0.274 + ], + "angle": 0, + "content": "3DEVCOM U.S. Army Research Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.569 + ], + "angle": 0, + "content": "In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.604, + 0.21, + 0.62 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.765 + ], + "angle": 0, + "content": "In recent years, the field of video action recognition has undergone significant advancement, largely driven by deep learning-based techniques. These include approaches based on convolutional neural networks (CNNs) [6, 14, 15, 47, 51, 53] and vision transformers (ViTs) [1, 5, 13, 33, 39, 42]. Further enriching this landscape, models integrating video and language data have emerged [36, 55, 57], which effectively leverage large collections of captioned videos to achieve impressive video understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Despite increasing performance on benchmark datasets, deep networks for video action recognition suffer from the same fundamental challenges as other machine learning models in the presence of distribution shift. This phenomenon, where a model's performance degrades when applied to data distributions different from those it was trained on, remains a critical obstacle in deploying the models in varied real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Domain adaptation (DA) seeks to mitigate the effects of" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.307, + 0.875, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.51, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Overview of the UNITE pipeline for video UDA. A teacher model is used to guide the self-supervised learning process in the first stage, and is again used to improve pseudolabeling of target videos during self-training in the third stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.869 + ], + "angle": 0, + "content": "distribution shift and has been a major research area within the computer vision community for years [41]. A subproblem within DA is unsupervised domain adaptation (UDA), which tries to make use of unlabeled data from the target domain to improve transfer performance. UDA approaches specific to video have also been developed, with many focusing on explicit alignment of the source and target domains using adversarial techniques or domain discrepancy measures. In contrast, we present here a method for video UDA that relies on self-supervised masked pre-training and self-training, both aided by a powerful image-based teacher model (CLIP [44]). The authors of DALL-V [67] demonstrate the effectiveness of CLIP in the source-free video domain adaptation setting by directly adapting it to action recognition tasks despite fundamentally being an image-based technique. Instead, we leverage CLIP as a spatial teacher for training a spatiotemporal student model that can better operate on target domain videos." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This paper presents, to the best of our knowledge, the first exploration of self-supervised masked distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "techniques in the context of unsupervised video domain adaptation. Our approach, Unsupervised Adaptation with Teacher-Enhanced Learning (UNITE), successfully harnesses the capabilities of an image-based teacher in two ways: (1) to form the target representations in a masked self-supervised pre-training stage on target domain videos, and (2) to improve the quality of pseudolabels used during self-training in collaboration with the student video network (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.449, + 0.258 + ], + "angle": 0, + "content": "We summarize the contributions of our work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.259, + 0.468, + 0.304 + ], + "angle": 0, + "content": "- We introduce the UNITE pipeline for video domain adaptation, which consists of a novel combination of masked video modeling and self-training techniques." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.305, + 0.468, + 0.379 + ], + "angle": 0, + "content": "- We conduct extensive experimentation by evaluating UNITE on three video domain adaptation benchmarks (Daily-DA, Sports-DA and UCF-HMDB) and observe significant performance gains compared to previously reported results." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.381, + 0.468, + 0.47 + ], + "angle": 0, + "content": "- We present a series of ablation experiments that study the effectiveness of various aspects of the UNITE pipeline and assess alternative design choices. In particular, we show that masked distillation pre-training and masked self-training are best applied in concert to achieve the strongest domain transfer performance." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.259, + 0.468, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.483, + 0.22, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Video Unsupervised Domain Adaptation (VUDA). Techniques for video unsupervised domain adaptation typically fall within a few common categories [59]. Adversarial methods attempt to align representations between the source and target domains by enforcing a domain confusion loss using a separate domain discriminator network, where the main network must learn to fool the discriminator. Some techniques which were developed for images, such as DANN [18] or MK-MMD [34], can be applied directly to video architectures. \\(\\mathrm{TA}^3\\mathrm{N}\\) [8] is a video-specific approach that uses separate adversaries for the spatial and temporal dimensions. These approaches can result in unstable training due to competing objectives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Another class of methods use contrastive learning and exploit the intrinsic structure of video. CoMix [45] uses contrastive learning with pseudolabeling and video-based augmentations such as sampling at varying frame rates or mixing the backgrounds of source videos and target videos. \\(\\mathrm{CO}^2\\mathrm{A}\\) [52] optimizes 6 different losses simultaneously, including contrastive losses at the clip- and video-level, along with supervised contrastive learning across domains. UDAVT [10] adapts vision transformers using the information bottleneck principle by having separate projectors for source and target features and estimating a cross correlation matrix between features whose labels and pseudolabels match." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Some video domain adaptation techniques have been developed to address the related problem of source-free video video domain adaptation (SFVUDA), which focuses on scenarios where source data is unavailable during the adaptation phase. ATCoN [62] tries to ensure the trained model is as consistent with its features as the original model. EX-TERN [63] additionally uses masking augmentations as a factor for features to be consistent over. DALL-V [67] takes a different approach by adapting a CLIP [44] image encoder to source and target datasets using an adapter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Masked Image & Video Modeling. Masked modeling first achieved success in natural language processing [11], and has since been applied to image and video data as a way to learn powerful feature encoding from unlabeled data. The basic idea behind masked modeling is to partially occlude patches of the input data and train a model to reconstruct the missing patches. This can be done at the pixel-level [16, 19, 31, 50, 54, 58], or the targets of the reconstruction can be formed by a neural network—either an exponential moving average (EMA) of the network being trained [2-4, 24, 32, 70], or even a separate teacher model. MVD [56] uses the latter concept to train a video transformer network using both image and video teachers, which can be viewed as a form of knowledge distillation [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Unmasked Teacher (UMT) [30] uses a different form of masked distillation to train powerful video encoders, where instead of reconstructing representations of masked regions, the student attempts to match the teacher representations of visible patches [66]. UMT is able to use a spatial encoder (CLIP [44]) to train a powerful spatiotemporal model. Because the video encoder only needs to process visible patches, and there is no need for an expensive decoder for reconstruction, UMT offers a highly efficient approach for self-supervision on videos. As such, we choose to leverage the UMT masked distillation objective to perform unsupervised video learning in UNITE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Masked Modeling for Domain Adaptation. Masked modeling is fundamentally about ensuring a given feature is invariant to whether its corresponding input is masked or not. Inducing such invariances on novel domains can be a useful property to enforce for the purposes of domain adaptation. MAE-TTT [17] finetunes a masked autoencoder [19] on single images at test time, adapting the feature extractor to the statistics of the current image while leaving the prediction head unchanged. MIC [22] uses an EMA teacher model on unmasked images to produce pseudolabels for training on masked target images. PACMAC [43] selects target domain samples for pseudolabeling based on prediction consistency over masked views, and also applies the classification loss to masked target inputs. Since these approaches have been developed for image-based tasks, there is still an open question of how similar concepts could be applied to video recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18920" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.218, + 0.106 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.282, + 0.13 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.473, + 0.368 + ], + "angle": 0, + "content": "We formalize the problem of unsupervised domain adaptation for action recognition as follows. Given a set of labeled videos \\(\\mathcal{D}_S\\coloneqq \\{(\\mathbf{x}_i^S,y_i^S)\\}_{i = 1}^{N_S}\\) drawn from a source data distribution \\(\\mathcal{P}_S\\), and a set of unlabeled videos \\(\\mathcal{D}_T\\coloneqq \\{\\mathbf{x}_i^T\\}_{i = 1}^{N_T}\\) drawn from a target distribution \\(\\mathcal{P}_T\\), our objective is to learn a function that can successfully classify videos drawn from \\(\\mathcal{P}_T\\). More precisely, we seek to learn the parameters \\(\\theta\\) of a function \\(f_{\\theta}\\), in our case a transformer neural network, that minimize the empirical risk on a sampling of videos from \\(\\mathcal{P}_T\\). The challenge in domain adaptation lies in the fact that \\(\\mathcal{P}_S\\) and \\(\\mathcal{P}_T\\) differ from one another, leading to a gap in performance when directly applying a source trained model on target domain data. Thus, UDA techniques need to exploit information in unlabeled target domain samples to reduce this gap." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.338, + 0.392 + ], + "angle": 0, + "content": "3.2. Self-Supervised Initialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.78 + ], + "angle": 0, + "content": "In contrast to some other video domain adaptation works [10, 65] that use network weights from Kinetics-400 supervised pre-training, we initialize our network from self-supervised pre-training. As contended in [43] for UDA with images, we argue that supervised pre-training can potentially complicate the study of VUDA techniques. When the pre-trained network has been trained to classify categories present in the DA dataset, some DA techniques could perform well simply by preserving the capabilities of the pretrained model despite not generalizing well to DA tasks that do not have category overlap with the pre-training dataset. Moreover, the UDA problem formulation implies that labeled instances of relevant classes are only available from the source domain. This condition does not hold for the Daily-DA and Sports-DA benchmarks that we use for evaluation, as 6 out of their 8 classes also appear in the Kinetics-400 dataset. Further exacerbating the issue with supervised Kinetics pre-training is the fact that Kinetics serves as a target domain in both benchmarks. Thus, we initialize the network weights from UMT [30] single-modality (i.e., video only) self-supervised pre-training on Kinetics-710 [29]. Please refer to Sec. 10 in the Appendix for select results using a UMT pre-trained network with additional supervised fine-tuning on Kinetics-400, which unsurprisingly achieves better baseline performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.169, + 0.807 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.172, + 0.831 + ], + "angle": 0, + "content": "4.1. UNITE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Our approach to VUDA consists of three stages: (1) unsupervised pre-training on target domain data, (2) supervised fine-tuning on source domain data and (3) collaborative self-training using videos from both domains. We now" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.092, + 0.885, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.322, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Unmasked Teacher (UMT) training used in Stage 1 of UNITE to perform unsupervised representation learning on target domain videos." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.389, + 0.694, + 0.404 + ], + "angle": 0, + "content": "describe each stage in detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.419, + 0.868, + 0.434 + ], + "angle": 0, + "content": "Stage 1: Unsupervised Target Domain Pre-Training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.827 + ], + "angle": 0, + "content": "To enable the extraction of discriminative features from target domain videos, the UNITE pipeline begins by performing additional unsupervised pre-training on the target dataset \\(\\mathcal{D}_T\\). For this, we adopt the UMT [30] self-supervised training objective. UMT training uses a pretrained and frozen spatial teacher model \\(g_{*}\\) to train a spatiotemporal student model \\(g_{a}\\) by enforcing alignment between feature representations of the two networks at multiple layers (depicted in Figure 2). A key aspect of the training process is that the student model takes as input a masked video view \\(m(\\mathbf{x}^T)\\), while the teacher model processes un-masked video frames from \\(\\mathbf{x}^T\\). Through UMT training, \\(g_{a}\\) must learn to match the semantically rich (albeit only spatially informed) features of \\(g_{*}\\) from heavily masked target domain videos. To ensure sampling of meaningful video patches, UMT performs an attention-guided masking operation \\(m(\\cdot)\\) drawn from a multinomial distribution defined by the attention map of the final layer CLS token in \\(g_{*}\\). Attention-guided masking allows the use of a high masking ratio, which makes UMT training extremely efficient. We formalize the Stage 1 objective as follows. Let \\(\\mathbf{z}_a^l\\) denote the L2 normalized \\(l\\)-th layer representation of \\(g_{a}\\) for a masked input \\(m(\\mathbf{x})\\), and let \\(\\mathbf{z}_{*}^{l}\\) denote the normalized \\(l\\)-th layer representation of \\(g_{*}\\) for the full input \\(\\mathbf{x}\\), but only at the locations corresponding to the visible patches in \\(m(\\mathbf{x})\\). The loss function in Stage 1 can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.834, + 0.848, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U M T}} = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {P} _ {T}} \\left[ \\frac {1}{| \\mathcal {A} |} \\sum_ {l \\in \\mathcal {A}} \\operatorname {M S E} \\left(d ^ {l} (\\mathbf {z} _ {a} ^ {l}), \\mathbf {z} _ {*} ^ {l}\\right) \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) is a set of aligned layers, \\(d^{l}(\\cdot)\\) is a linear projection" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.2, + 0.088, + 0.773, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 3. Overview of the collaborative self-training stage in UNITE. The student and teacher models work together to produce more accurate pseudolabels for target domain videos. The target domain classification loss is enforced on masked target videos to encourage stronger context learning. Source domain classification is included to stabilize training, which is especially important at the beginning of training when pseudolabels may have low accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.546 + ], + "angle": 0, + "content": "for the \\(l\\)-th layer and MSE indicates mean squared error. Following the original UMT method, we leverage a CLIP [44] ViT-B image encoder as \\(g_{*}\\). We keep the standard UMT masking ratio of \\(r = 0.8\\) and perform alignment between the last 6 layers of the networks. \\(g_{a}\\) is initialized as described in Sec. 3.2, along with the projections \\(d^{l}(\\cdot)\\), which are kept frozen throughout Stage 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.471, + 0.803 + ], + "angle": 0, + "content": "Stage 2: Source Domain Fine-Tuning. The second stage of UNITE fits a classifier on source domain data using \\( g_{a} \\) from Stage 1. We introduce label space information in a separate stage from target domain self-supervision because, as the authors of [17] observe for MAE training, we find that imposing a classification loss and UMT loss simultaneously can hinder convergence. To perform supervised fine-tuning, we introduce a linear classification head \\( h_{a} \\) that operates on the mean pooled outputs of the final transformer layer in \\( g_{a} \\). The full student classifier network can now be represented as \\( f_{a} = h_{a}(g_{a}(\\cdot)) \\). As is the case with other masked pretraining strategies, we find that representations learned via UMT pre-training are not linearly separable. Thus, we perform full fine-tuning of \\( f_{a} \\) by minimizing a cross entropy (CE) loss \\( \\mathcal{L}_{CE} \\) on unmasked source domain videos. The Stage 2 loss function is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.816, + 0.469, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S F T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.886 + ], + "angle": 0, + "content": "To help preserve target domain feature extraction learned in Stage 1, we apply a layer-wise learning rate decay when fine-tuning on source domain data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Stage 3: Collaborative Self-Training (CST). After enhancing the ability of \\( g_{a} \\) to extract meaningful features in the target domain (Stage 1) and fitting \\( f_{a} \\) for the classification task using source videos (Stage 2), we further adapt \\( f_{a} \\) to target data using a self-training process (depicted in Figure 3). In a conventional self-training framework (denoted as \\( \\mathcal{L}_{\\mathrm{ST}} \\) in Eq. (2)), the overall loss is a sum of losses on labeled source domain samples and select target domain samples, chosen using mask \\( s(\\cdot) \\). The target domain loss leverages the model's own predictions \\( \\hat{y} \\) (where \\( \\hat{y} = \\operatorname{argmax}_c \\sigma(f(\\mathbf{x}^T))[c] \\) and \\( \\sigma(\\cdot) \\) denotes the softmax function) in lieu of labels in the CE loss computation. Various strategies have been used to form the selection mask \\( s(\\cdot) \\), with the goal of choosing target samples whose predictions are more likely to be correct [43, 48, 68]." + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.687, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] + \\tag {2} \\\\ \\lambda \\mathbb {E} _ {\\mathbf {x} ^ {T} \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} (f _ {a} (\\mathbf {x} ^ {T}), \\hat {y} _ {a} (\\mathbf {x} ^ {T})) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a modified form of self-training, which we refer to as collaborative self-training (CST), that makes use of the student-teacher setup from Stage 1 to improve the process of pseudolabel estimation. Specifically, we use the zero-shot classification capabilities of the CLIP image teacher model \\( f_{*} \\) (see Sec. 5.2 for details), in combination with predictions from the video student model \\( f_{a} \\), to create refined target pseudolabels for self-training. We employ the MatchOrConf scheme proposed in [69] to combine the outputs of the two models as shown in Eq. (3), where \\( \\mathrm{Conf}(\\hat{y}) \\) denotes the maximum softmax probability" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18922" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.405, + 0.108 + ], + "angle": 0, + "content": "[20], \\(\\max_c\\sigma (f(\\mathbf{x}^T))[c]\\), to create pseudolabel \\(\\tilde{y}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.131, + 0.486, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = \\left\\{ \\begin{array}{l l} \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} = \\hat {y} _ {*} \\\\ \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) > \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) \\leq \\gamma , \\\\ \\hat {y} _ {*} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) \\leq \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) > \\gamma , \\\\ - 1 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.307 + ], + "angle": 0, + "content": "\\(\\mathbf{x}^T\\) is an argument of the above, but is omitted for clarity. The MatchOrConf scheme selects only those target samples where there is agreement between the predictions of the two models, or where one model's confidence exceeds threshold \\(\\gamma\\) while the other's does not, resulting in the following selection mask:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.319, + 0.47, + 0.358 + ], + "angle": 0, + "content": "\\[\ns \\left(\\mathbf {x} ^ {T}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {y} \\left(\\mathbf {x} ^ {T}\\right) \\neq - 1, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.499 + ], + "angle": 0, + "content": "Taking inspiration from [22] and [43], we compute the target domain loss on masked videos, where we employ the same teacher-guided attention masking strategy (with \\( r = 0.8 \\)) used during UMT pre-training in Stage 1. We find that training on masked videos in this stage results in substantial performance boosts compared to training on un-masked videos (see Table 6). Like [22], we include a loss weighting \\( q \\) for each target video, based on the student network confidence:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.513, + 0.47, + 0.531 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} ^ {T}\\right) = \\operatorname {C o n f} \\left(\\hat {y} _ {a} \\left(\\mathbf {x} ^ {T}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.567 + ], + "angle": 0, + "content": "The complete loss function for the collaborative self-training stage can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.588, + 0.47, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} (\\mathbf {x} ^ {S}), y ^ {S}\\right) \\right] + \\\\ \\lambda \\mathbb {E} _ {(\\mathbf {x} ^ {T}) \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) q (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} \\left(f _ {a} (m (\\mathbf {x} ^ {T})), \\tilde {y} (\\mathbf {x} ^ {T}) \\right] \\right. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.21, + 0.672 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.181, + 0.695 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We evaluate our approach on three video domain adaptation benchmarks: Daily-DA, Sports-DA, UCF \\(\\leftrightarrow\\) HMDBfull." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.47, + 0.856 + ], + "angle": 0, + "content": "- Daily-DA [64] is a recently introduced benchmark composed of videos from four domains: ARID (A) [60], HMDB51 (H) [28], Moments-in-Time (M) [38] and Kinetics (K) [7, 26]. It consists of 8 overlapping classes representing everyday actions, with a total volume of approximately 19K videos. Videos from the ARID domain present a unique challenge in this benchmark, as they were explicitly filmed under low-illumination conditions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- Sports-DA [64] contains videos from three domains: Sports-1M (S) [25], UCF101 (U) [49] and Kinetics (K) [7, 26]. It includes 23 overlapping categories depicting" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "various sporting activities, with a total volume of approximately 41K videos." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.168 + ], + "angle": 0, + "content": "- \\( \\mathbf{UCF} \\leftrightarrow \\mathbf{HMDB}_{full} \\) [8] contains approximately 3.2K videos across 12 action categories taken from the UCF101 (U) [49] and HMDB51 (K) [28] datasets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.204, + 0.719, + 0.22 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.895, + 0.383 + ], + "angle": 0, + "content": "Frame Sampling. During training, we perform random uniform frame sampling [53] on each video. We divide the video into \\( T = 8 \\) uniform segments and randomly sample one frame from each. All frames are resized to \\( 224 \\times 224 \\) pixels, resulting in a \\( 3 \\times 8 \\times 224 \\times 224 \\) tensor for each video. During testing, we average the network predictions across 3 spatial crops and 4 temporal clips, where clip \\( i \\) is formed by splitting each of the \\( T \\) segments into 4 sub-segments and sampling a frame from the \\( i \\)-th sub-segment of each." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Network Architecture & Training. All of our experiments use a standard ViT-B/16 [12] architecture with a fixed sine and cosine positional encoding scheme. Our patch embedding projection operates on each frame independently, and self-attention is performed across both the space and time dimensions. All classification is performed using mean pooling of the final layer patch representations. We conduct our experiments using the PyTorch deep learning library [40] on 4 NVIDIA A5000 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.894, + 0.699 + ], + "angle": 0, + "content": "During Stage 1, we perform UMT training for 50 epochs on the target dataset using a batch size of 256 and a learning rate of \\(1.5\\mathrm{e - }5\\). During Stage 2, we use a batch size of 28, a learning rate of \\(2.5\\mathrm{e - }5\\), and apply a layer-wise learning rate decay of 0.65. During stage 3, we use a batch size of 40 (20 videos from each domain), a learning rate of 1e-5, a MatchOrConf confidence threshold of \\(\\gamma = 0.1\\) and a target domain loss weight \\(\\lambda = 1\\). Training in all stages uses the AdamW optimizer [35] and a weight decay of 0.05. Please refer to Sec. 8 in the Appendix for more training details for each stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Zero-Shot Classification with CLIP. In Stage 3 of UNITE, we use CLIP (ViT-B/16) in a zero-shot manner to refine the pseudolabels of target domain videos. We do so by passing each of the \\( T \\) video frames through the CLIP image encoder and computing the cosine similarity with a set of text representations produced by the CLIP text encoder. The inputs to the text encoder are the class names associated with the particular action recognition task, embedded in the following template: \"A video of a person {class}.\" To form a video-level prediction, we simply average the softmax of the cosine similarity scores across the video frames. Refer to the Appendix (Sec. 11) for a listing of class names used for each task." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18923" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.091, + 0.894, + 0.312 + ], + "angle": 0, + "content": "
MethodTarget Domain Accuracy (Top-1%)
H→AM→AK→AA→HM→HK→HH→MA→MK→MM→KH→KA→KAvg.
DALL-V [67] LB17.534.715.614.644.647.925.515.535.761.645.117.831.3
Source Only40.452.136.549.668.357.941.536.343.379.348.041.749.6
ZSCLIP (ViT-B/16) [44]36.536.536.560.060.060.048.548.548.568.168.168.153.3
SFVUDAATCoN [62]17.927.217.226.747.348.230.717.232.557.748.531.033.5
EXTERNAL [63]26.218.123.926.253.755.840.718.235.268.157.651.439.6
DALL-V [67]24.024.024.057.965.452.547.045.747.078.176.775.051.4
UDADANN [18]14.222.821.220.143.337.529.519.721.758.838.227.029.5
MK-MMD [34]20.321.021.718.750.436.225.718.024.058.533.826.129.5
TA3N [8]14.421.619.914.943.037.725.715.631.555.538.423.428.5
UNITE w/o CST43.846.734.751.770.854.644.339.043.378.151.750.850.8
UNITE (Ours)48.044.137.567.974.265.851.850.048.089.969.963.659.2
DALL-V [67] UB26.926.926.970.470.470.461.561.561.588.988.988.961.9
Target Only68.568.568.584.684.684.673.073.073.098.398.398.381.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.32, + 0.878, + 0.335 + ], + "angle": 0, + "content": "Table 1. UDA results on Daily-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.348, + 0.734, + 0.568 + ], + "angle": 0, + "content": "
MethodTarget Domain Accuracy (Top-1 %)
U→SK→SS→UK→UU→KS→KAvg.
DALL-V [67] LB64.379.584.485.467.278.276.5
Source Only67.686.897.998.979.989.186.7
ZSCLIP (ViT-B/16) [44]85.085.093.393.391.391.389.9
SFVUDAATCoN [62]47.969.790.693.665.276.073.8
EXTERN [63]72.773.895.493.781.282.283.2
DALL-V [67]75.977.788.888.081.282.382.3
UDADANN [18]55.175.085.788.065.973.473.8
MK-MMD [34]55.667.990.990.266.173.674.0
TA3N [8]54.168.693.090.363.672.673.7
UNITE w/o CST73.987.398.299.284.690.288.9
UNITE (Ours)86.090.298.799.894.295.394.0
DALL-V [67] UB88.388.393.493.485.685.689.1
Target Only97.997.998.898.899.999.998.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.578, + 0.882, + 0.593 + ], + "angle": 0, + "content": "Table 2. UDA results on Sports-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.186, + 0.633 + ], + "angle": 0, + "content": "5.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.763 + ], + "angle": 0, + "content": "We compare results using UNITE with previously developed techniques in UDA, as well as in the closely related setting of SFUDA as reported by [67]. For Daily-DA and Sports-DA we compare against DANN [18], MK-MMD [34] and \\(\\mathrm{TA}^3\\mathrm{N}\\) [8] in the UDA setting, and ATCoN [62], EXTERN [63] and DALL-V [67] for SFUDA. For \\(UCF\\leftrightarrow HMDB_{full}\\), we additionally report accuracies from \\(\\mathrm{CO}^2\\mathrm{A}\\) [52] and UDAVT [10]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.265, + 0.787 + ], + "angle": 0, + "content": "5.4. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Tables 1, 2 and 3 show the results of our evaluation of UNITE on Daily-DA, Sports-DA and \\( UCF \\leftrightarrow HMDB_{full} \\). We include accuracies after Stage 2 of the pipeline (denoted as UNITE w/o CST) as well as after applying the full UNITE process. For reference, we include accuracies using our UMT pre-trained network with source domain supervised fine-tuning (Source Only) and with target domain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.71 + ], + "angle": 0, + "content": "supervised fine-tuning (Target Only). We also include the source only and target only accuracies (denoted in the tables as LB and UB) of DALL-V [67] (a CLIP-based ResNet-50 model) for added context. We additionally display the zero-shot classification accuracy of the CLIP teacher model that we utilize during Stage 1 and Stage 3 of UNITE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.852 + ], + "angle": 0, + "content": "An observation we make across all benchmarks is the enhanced baseline accuracy of the UMT pre-trained model, which speaks to the strength of the unsupervised masked distillation process and the ViT architecture. On the Daily-DA benchmark, we find that UNITE exceeds previously reported results on most domain shifts. On only one shift \\((\\mathrm{M}\\rightarrow \\mathrm{A})\\) we observe degradation in target accuracy after employing UNITE compared to source only, which is consistent with previous work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "UNITE exhibits even stronger performance on the larger-scale Sports-DA benchmark, where it exceeds previous results by a large margin. We also see more consistent and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18924" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.416, + 0.317 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
H→UU→HAvg.
SDALL-V [67] LB71.676.173.8
Source Only88.880.084.4
CLIP (ViT-B/16) [44]88.891.790.3
SEVUDAATCoN [62]85.379.782.5
EXTERN [63]91.988.990.4
DALL-V [67]93.188.991.0
UDADANN [18]74.475.174.8
MK-MMD [34]74.779.777.2
TA3N [8]78.184.881.5
CO2A [52]95.887.891.8
UDAVT [10]96.892.394.6
UNITE w/o CST92.183.687.9
UNITE (Ours)92.595.093.8
DALL-V [67] UB93.791.492.6
Target Only99.796.798.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.319, + 0.472, + 0.348 + ], + "angle": 0, + "content": "Table 3. UDA results on \\(UCF \\leftrightarrow HMDB_{full}\\). Rows in color use our UMT backbone, and rows without color are reported from [61]." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.357, + 0.447, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.64 + ], + "angle": 0, + "content": "Figure 4. Class-wise accuracy on ARID \\(\\rightarrow\\) HMDB. Performances are shown for the CLIP teacher model and the student model before and after the collaborative self-training stage (Stage 3). For most classes, the Stage 3 student exceeds the accuracy of both the Stage 2 student and the CLIP teacher." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.711 + ], + "angle": 0, + "content": "substantial improvement after the UMT target pre-training stage alone (UNITE w/o CST), suggesting that overtraining may be occurring on the smaller datasets in Daily-DA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.771 + ], + "angle": 0, + "content": "On \\( UCF \\leftrightarrow HMDB_{full} \\) we see state-of-the-art performance on \\( \\mathrm{U} \\rightarrow \\mathrm{H} \\), while UDAVT outperforms on \\( \\mathrm{H} \\rightarrow \\mathrm{U} \\). Nevertheless, UNITE still achieves strong overall performance on the benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.382, + 0.803 + ], + "angle": 0, + "content": "6. Additional Analysis & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Student exceeds the teacher. We see in the benchmark results that the video student model exceeds the zero-shot accuracy of the CLIP teacher in virtually all domain shifts. This suggests that UNITE successfully leverages the spatial modeling capabilities of CLIP to train an even more powerful spatiotemporal model for the action recognition task." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.885, + 0.188 + ], + "angle": 0, + "content": "
MethodPTCSTAccuracy (%)
H→AA→H
Source OnlyXX40.449.6
+ Target UMT Pre-Training✓X43.851.7
+ Collaborative Self-TrainingX✓42.060.8
UNITE✓✓48.067.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.19, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 4. Ablation of stages in UNITE on ARID ↔ HMDB from Daily-DA. PT and CST denote UMT pre-training and collaborative self-training, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.893, + 0.508 + ], + "angle": 0, + "content": "We can see from Figure 4 that the teacher model and the student model prior to collaborative self-training have differing capabilities. For example, the image teacher is stronger at recognizing actions from the jump class than the Stage 2 student. On the other hand, the video student is significantly better than the teacher at recognizing wave at the start of CST. In both cases, we see that the student model, through collaborative self-training, comes to exceeds both the Stage 2 student and the CLIP image teacher. With the MatchOr-Conf pseudolabeling scheme, the two models are able to work together to achieve higher accuracy in the target domain. We see accuracy on the pick and run classes declines after CST, as self-training tends to reinforce the accuracy of well-performing classes, sometimes at the expense of more difficult ones that are infrequently assigned pseudolabels. Modifying CST to account for the learning status of each class, like in [68], may help remedy this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.674 + ], + "angle": 0, + "content": "Impact of Stage 1 and 3. In Table 4 we ablate the stages in UNITE. While UMT pre-training and collaborative self-training each offer a benefit in the absence of the other, the largest gains in domain transfer performance are obtained when combining the two stages. For example, on the ARID \\(\\rightarrow\\) HMDB domain shift in Daily-DA, Stage 1 and Stage 3 alone improve upon source only accuracy by \\(+1.8\\%\\) and \\(+11.2\\%\\) respectively. Meanwhile, the full UNITE pipeline results in \\(+18.3\\%\\) accuracy improvement on the target domain, suggesting that the masked pre-training and masked self-training processes complement one another." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Impact of data domains during pre-training. In Table 5, we assess the impact of utilizing each of the two data domains during UMT pre-training on UDA performance. While the results vary across domain shifts, we observe the most consistent improvements when pre-training only on target domain videos. Although incorporating both domains during unsupervised pre-training has been successfully applied in image-based domain adaptation [27, 37, 43, 46], we find it to be suboptimal with UMT training on the VUDA tasks studied in this work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Effect of masking during self-training. In Table 6, we see that enforcing the target domain CE loss on masked videos results in increased performance compared to using unmasked target videos, with smaller benefits for \\(\\mathrm{H}\\rightarrow \\mathrm{A}\\) and much larger benefits for \\(\\mathrm{A}\\rightarrow \\mathrm{H}\\). We believe the benefits of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18925" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.089, + 0.404, + 0.176 + ], + "angle": 0, + "content": "
UMT Pre-Training DataAccuracy (%)
H→AA→H
None42.060.8
Source Only48.360.0
Target Only48.067.9
Source + Target45.267.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.178, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Table 5. Comparison of data domains used during UMT pretraining. Accuracies are reported after the self-training stage. Bold indicates best accuracy and underline indicates second best." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.228, + 0.408, + 0.292 + ], + "angle": 0, + "content": "
Data for Target CE LossAccuracy (%)
H→AA→H
Unmasked46.654.6
Masked48.067.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.294, + 0.469, + 0.322 + ], + "angle": 0, + "content": "Table 6. Effect of applying the target domain cross entropy loss on masked vs. unmasked target videos during the self-training stage." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.33, + 0.442, + 0.442 + ], + "angle": 0, + "content": "
Pseudolabeling StrategyAccuracy (%)
H→AA→H
Cons (k=2)39.857.5
ConsOrConf (k=2,T=0.5) [43]48.656.3
ConsAndConf (k=2,T=0.5)40.157.1
Zero-Shot CLIP Only (T=0.5)34.751.3
MatchOrConf (γ=0.1) [69]48.067.9
MatchAndConsOrConf (k=2,γ=0.1)47.460.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.443, + 0.422, + 0.457 + ], + "angle": 0, + "content": "Table 7. Comparison of pseudolabeling strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.553 + ], + "angle": 0, + "content": "masking arise from forcing the model to use different cues from the same video to predict the pseudolabel, resulting in more robust target domain recognition. As we saw in Table 4, the masked pre-training stage plays a key role in unlocking the benefits of masked self-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comparison of pseudolabeling strategies. In Table 7, we explore the use of alternative pseudolabeling strategies to the MatchOrConf [69] scheme employed in UNITE. The first 3 rows of the table assess variations of the masked consistency pseudolabeling scheme proposed in PACMAC [43], where target samples are selected if there is agreement between predictions of \\( k \\) masked views. This criterion can also be combined with a constraint on the confidence exceeding threshold \\( T \\). While we find the combination of these two constraints (ConsOrConf) to be effective for self-training, we find that MatchOrConf results in more consistent target accuracy improvements across domain shifts. We also consider combining MatchOrConf with a masked consistency constraint as a way to reduce cases where the student and teacher models agree on an incorrect prediction. This scheme, denoted as MatchAndConsOrConf, adds an additional constraint to the matching criterion by also requiring agreement between the student model predictions of \\( k \\) masked views. We find that this approach underperforms compared to the vanilla MatchOrConf scheme due to low utilization of target domain samples. Our results also indicate that using both models together for pseudolableing outperforms using only zero-shot CLIP predictions." + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.089, + 0.838, + 0.15 + ], + "angle": 0, + "content": "
Loss During Self-TrainingAccuracy (%)
H→AA→H
LCE(m(xT), y)33.742.1
LCE(m(xT), y) + LCE(xS, yS)48.067.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Table 8. Effect of including a source domain classification loss in addition to the target domain loss during the self-training stage." + }, + { + "type": "table", + "bbox": [ + 0.582, + 0.188, + 0.816, + 0.284 + ], + "angle": 0, + "content": "
Confidence ThresholdAccuracy (%)
H→AA→H
0.148.067.9
0.344.155.4
0.551.964.6
0.742.665.8
0.940.762.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.286, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Table 9. Effect of confidence threshold \\(\\gamma\\) in MatchOrConf pseudolabeling scheme from Equation 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Effect of source classification during self-training. In Table 8, we study the importance of including a source domain classification loss while performing collaborative self-training on masked target videos. Indeed, we find that source classification is a crucial ingredient in this stage, as excluding it destabilizes training and results in performance degradation rather than improvement. We suspect that the ground truth labels on source data help to combat the effects of inaccurate target pseudolabels, which may be prevalent especially at the start of training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Choice of pseudolabeling confidence threshold. In Table 9, we compare adaptation performance for various values of confidence threshold in the MatchOrConf pseudolabeling scheme. We observe significant variation across domain shifts, but find that \\(\\gamma = 0.1\\) leads to relatively consistent improvement. We leave for future work the development of a more principled approach to determining an appropriate confidence threshold for a given domain shift." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.627, + 0.631 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this work, we presented UNITE, a three step process for unsupervised video domain adaptation. Our approach leverages a powerful spatial encoder to aid in the adaptation of a spatiotemporal network using unlabeled target domain videos—making significant progress towards bridging the domain gaps in various VUDA datasets. The advancements demonstrated by UNITE underscore the untapped potential of masked modeling techniques for video domain adaptation, which we hope will spur further research in this area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This research was sponsored by the Army Research Laboratory under Cooperative Agreement W911NF-21-2-0211. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18926" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked Siamese Networks for Label-Efficient Learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[5] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In ICML, pages 813-824, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.41 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[7] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A Short Note about Kinetics-600, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[8] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal Attentive Alignment for Large-Scale Video Domain Adaptation. In ICCV, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[9] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 iee. In CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.58, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[10] Victor G. Turrisi Da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Unsupervised Domain Adaptation for Video Transformers in Action Recognition. In ICPR, pages 1258-1265, Montreal, QC, Canada, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.649, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.746, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.83, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale Vision Transformers. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In CVPR, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[17] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[21] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[22] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation. In CVPR, pages 11721–11732, Vancouver, BC, Canada, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.498 + ], + "angle": 0, + "content": "[23] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to Hide from Your Students: Attention-Guided Masked Image Modeling. In ECCV, pages 300-318, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-Scale Video Classification with Convolutional Neural Networks. In CVPR, pages 1725–1732, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.691, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[27] Donghyun Kim, Kuniaki Saito, Tae-Hyun Oh, Bryan A Plummer, Stan Sclaroff, and Kate Saenko. Cds: Cross-domain self-supervised pre-training. In ICCV, pages 9123-9132, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[28] H Kuehne, H Jhuang, E Garrote, T Poggio, and T Serre. HMDB: A Large Video Database for Human Motion Recognition. In ICCV, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In ICCV, pages 1632-1643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked Teacher: Towards Training-Efficient Video Foundation Models. In ICCV, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18927" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[31] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. MST: Masked Self-Supervised Transformer for Visual Representation. In NeurIPS, pages 13165-13176, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[32] Han Lin, Guangxing Han, Jiawei Ma, Shiyuan Huang, Xudong Lin, and Shih-Fu Chang. Supervised masked knowledge distillation for few-shot transformers. In CVPR, pages 19649-19659, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3192-3201, New Orleans, LA, USA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[36] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-CLIP: End-to-End Multi-grained Contrastive Learning for Video-Text Retrieval. In ACM MM, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[37] Samarth Mishra, Kate Saenko, and Venkatesh Saligrama. Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[38] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, and Dan Gutfreund. Moments in Time Dataset: One Million Videos for Event Understanding. PAMI, 42(2):502-508, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[39] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video Transformer Network. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[41] Vishal M Patel, Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. Visual Domain Adaptation: A survey of recent advances. IEEE Signal Processing Magazine, 32(3):53-69, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[42] Aj Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning. In CVPR, pages 2214-2224, Vancouver, BC, Canada, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[43] Viraj Prabhu, Sriram Yenamandra, Aaditya Singh, and Judy Hoffman. Adapting self-supervised vision transformers by probing attention-conditioned masking consistency. NeurIPS, 35:23271-23283, 2022. 2, 3, 4, 5, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Aadarsh Sahoo, Rutav Shah, R. Panda, Kate Saenko, and Abir Das. Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing. In NeurIPS, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[46] Kendrick Shen, Robbie M. Jones, Ananya Kumar, Sang Michael Xie, Jeff Z. Haochen, Tengyu Ma, and Percy Liang. Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation. In ICML, pages 19847-19878, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[47] Karen Simonyan and Andrew Zisserman. Two-Stream Convolutional Networks for Action Recognition in Videos. In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In NeurIPS, pages 596-608, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[49] K. Soomro, A. Zamir, and M. Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. ArXiv, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[51] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning Spatiotemporal Features with 3D Convolutional Networks. In ICCV, pages 4489-4497, Santiago, Chile, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[52] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-Head Contrastive Domain Adaptation for Video Action Recognition. In WACV, pages 2234–2243, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal Segment Networks for Action Recognition in Videos. PAMI, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. In CVPR, pages 14549-14560, Vancouver, BC, Canada, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[56] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning. In CVPR, pages 6312-6322, Vancouver, BC, Canada, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. InternVideo: General Video Foundation Models via Generative and Discriminative Learning, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18928" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[59] Yuecong Xu, Haozhi Cao, Zhenghua Chen, Xiaoli Li, Lihua Xie, and Jianfei Yang. Video Unsupervised Domain Adaptation with Deep Learning: A Comprehensive Survey, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[60] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. ARID: A New Dataset for Recognizing Action in the Dark. In Communications in Computer and Information Science, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[61] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Aligning Correlation Information for Domain Adaptation in Action Recognition. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[62] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-Free Video Domain Adaptation by Learning Temporal Consistency for Action Recognition. In ECCV, pages 147-164, Cham, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[63] Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Extern: Leveraging endo-temporal regularization for black-box video domain adaptation. arXiv preprint arXiv:2208.05187, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[64] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Zhengguo Li, and Zhenghua Chen. Multi-Source Video Domain Adaptation With Temporal Attentive Moment Alignment Network. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3860-3871, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[65] Yuecong Xu, Jianfei Yang, Yunjiao Zhou, Zhenghua Chen, Min Wu, and Xiaoli Li. Augmenting and Aligning Snippets for Few-Shot Video Domain Adaptation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[66] Hongwei Xue, Peng Gao, Hongyang Li, Yu Qiao, Hao Sun, Houqiang Li, and Jiebo Luo. Stare at What You See: Masked Image Modeling without Reconstruction. In CVPR, pages 22732-22741, Vancouver, BC, Canada, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.56, + 0.471, + 0.627 + ], + "angle": 0, + "content": "[67] Giacomo Zara, Alessandro Conti, Subhankar Roy, Stephane Lathuiliere, Paolo Rota, and Elisa Ricci. The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.471, + 0.697 + ], + "angle": 0, + "content": "[68] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 34:18408-18419, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.701, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[69] Wenyu Zhang, Li Shen, and Chuan-Sheng Foo. Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation. In CVPR, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.743, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[70] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT Pre-training with Online Tokenizer. In ICLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18929" + } + ] +] \ No newline at end of file diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_origin.pdf b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b609a8ce469a485c59379de32ca2ab22af5ef911 --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/1edbe2de-8bfe-4c07-9630-27dfd62c6876_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d1c64785bfcfd5b71e7808c7550f47f712ade0662d6fef784a2aea0ab7c7501 +size 4759827 diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/full.md b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b63ea77ecf4028ad68eda8bd18b7dc9f53d2eee7 --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/full.md @@ -0,0 +1,308 @@ +# Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training + +Arun Reddy $^{1,2}$ , William Paul $^{1}$ , Corban Rivera $^{1}$ , Ketul Shah $^{2}$ , Celso M. de Melo $^{3}$ , Rama Chellappa $^{2}$ + +1 Johns Hopkins University Applied Physics Laboratory + +$^{2}$ Johns Hopkins University, Department of Electrical & Computer Engineering + +3DEVCOM U.S. Army Research Laboratory + +# Abstract + +In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results. + +# 1. Introduction + +In recent years, the field of video action recognition has undergone significant advancement, largely driven by deep learning-based techniques. These include approaches based on convolutional neural networks (CNNs) [6, 14, 15, 47, 51, 53] and vision transformers (ViTs) [1, 5, 13, 33, 39, 42]. Further enriching this landscape, models integrating video and language data have emerged [36, 55, 57], which effectively leverage large collections of captioned videos to achieve impressive video understanding capabilities. + +Despite increasing performance on benchmark datasets, deep networks for video action recognition suffer from the same fundamental challenges as other machine learning models in the presence of distribution shift. This phenomenon, where a model's performance degrades when applied to data distributions different from those it was trained on, remains a critical obstacle in deploying the models in varied real-world scenarios. + +Domain adaptation (DA) seeks to mitigate the effects of + +![](images/983754bd8c0575bab4a022d6803979f33aa1e57743d7a9d4b106a10e23bf56e0.jpg) +Figure 1. Overview of the UNITE pipeline for video UDA. A teacher model is used to guide the self-supervised learning process in the first stage, and is again used to improve pseudolabeling of target videos during self-training in the third stage. + +distribution shift and has been a major research area within the computer vision community for years [41]. A subproblem within DA is unsupervised domain adaptation (UDA), which tries to make use of unlabeled data from the target domain to improve transfer performance. UDA approaches specific to video have also been developed, with many focusing on explicit alignment of the source and target domains using adversarial techniques or domain discrepancy measures. In contrast, we present here a method for video UDA that relies on self-supervised masked pre-training and self-training, both aided by a powerful image-based teacher model (CLIP [44]). The authors of DALL-V [67] demonstrate the effectiveness of CLIP in the source-free video domain adaptation setting by directly adapting it to action recognition tasks despite fundamentally being an image-based technique. Instead, we leverage CLIP as a spatial teacher for training a spatiotemporal student model that can better operate on target domain videos. + +This paper presents, to the best of our knowledge, the first exploration of self-supervised masked distillation + +techniques in the context of unsupervised video domain adaptation. Our approach, Unsupervised Adaptation with Teacher-Enhanced Learning (UNITE), successfully harnesses the capabilities of an image-based teacher in two ways: (1) to form the target representations in a masked self-supervised pre-training stage on target domain videos, and (2) to improve the quality of pseudolabels used during self-training in collaboration with the student video network (see Figure 1). + +We summarize the contributions of our work as follows: + +- We introduce the UNITE pipeline for video domain adaptation, which consists of a novel combination of masked video modeling and self-training techniques. +- We conduct extensive experimentation by evaluating UNITE on three video domain adaptation benchmarks (Daily-DA, Sports-DA and UCF-HMDB) and observe significant performance gains compared to previously reported results. +- We present a series of ablation experiments that study the effectiveness of various aspects of the UNITE pipeline and assess alternative design choices. In particular, we show that masked distillation pre-training and masked self-training are best applied in concert to achieve the strongest domain transfer performance. + +# 2. Related Work + +Video Unsupervised Domain Adaptation (VUDA). Techniques for video unsupervised domain adaptation typically fall within a few common categories [59]. Adversarial methods attempt to align representations between the source and target domains by enforcing a domain confusion loss using a separate domain discriminator network, where the main network must learn to fool the discriminator. Some techniques which were developed for images, such as DANN [18] or MK-MMD [34], can be applied directly to video architectures. $\mathrm{TA}^3\mathrm{N}$ [8] is a video-specific approach that uses separate adversaries for the spatial and temporal dimensions. These approaches can result in unstable training due to competing objectives. + +Another class of methods use contrastive learning and exploit the intrinsic structure of video. CoMix [45] uses contrastive learning with pseudolabeling and video-based augmentations such as sampling at varying frame rates or mixing the backgrounds of source videos and target videos. $\mathrm{CO}^2\mathrm{A}$ [52] optimizes 6 different losses simultaneously, including contrastive losses at the clip- and video-level, along with supervised contrastive learning across domains. UDAVT [10] adapts vision transformers using the information bottleneck principle by having separate projectors for source and target features and estimating a cross correlation matrix between features whose labels and pseudolabels match. + +Some video domain adaptation techniques have been developed to address the related problem of source-free video video domain adaptation (SFVUDA), which focuses on scenarios where source data is unavailable during the adaptation phase. ATCoN [62] tries to ensure the trained model is as consistent with its features as the original model. EX-TERN [63] additionally uses masking augmentations as a factor for features to be consistent over. DALL-V [67] takes a different approach by adapting a CLIP [44] image encoder to source and target datasets using an adapter. + +Masked Image & Video Modeling. Masked modeling first achieved success in natural language processing [11], and has since been applied to image and video data as a way to learn powerful feature encoding from unlabeled data. The basic idea behind masked modeling is to partially occlude patches of the input data and train a model to reconstruct the missing patches. This can be done at the pixel-level [16, 19, 31, 50, 54, 58], or the targets of the reconstruction can be formed by a neural network—either an exponential moving average (EMA) of the network being trained [2-4, 24, 32, 70], or even a separate teacher model. MVD [56] uses the latter concept to train a video transformer network using both image and video teachers, which can be viewed as a form of knowledge distillation [21]. + +Unmasked Teacher (UMT) [30] uses a different form of masked distillation to train powerful video encoders, where instead of reconstructing representations of masked regions, the student attempts to match the teacher representations of visible patches [66]. UMT is able to use a spatial encoder (CLIP [44]) to train a powerful spatiotemporal model. Because the video encoder only needs to process visible patches, and there is no need for an expensive decoder for reconstruction, UMT offers a highly efficient approach for self-supervision on videos. As such, we choose to leverage the UMT masked distillation objective to perform unsupervised video learning in UNITE. + +Masked Modeling for Domain Adaptation. Masked modeling is fundamentally about ensuring a given feature is invariant to whether its corresponding input is masked or not. Inducing such invariances on novel domains can be a useful property to enforce for the purposes of domain adaptation. MAE-TTT [17] finetunes a masked autoencoder [19] on single images at test time, adapting the feature extractor to the statistics of the current image while leaving the prediction head unchanged. MIC [22] uses an EMA teacher model on unmasked images to produce pseudolabels for training on masked target images. PACMAC [43] selects target domain samples for pseudolabeling based on prediction consistency over masked views, and also applies the classification loss to masked target inputs. Since these approaches have been developed for image-based tasks, there is still an open question of how similar concepts could be applied to video recognition. + +# 3. Preliminaries + +# 3.1. Problem Formulation + +We formalize the problem of unsupervised domain adaptation for action recognition as follows. Given a set of labeled videos $\mathcal{D}_S\coloneqq \{(\mathbf{x}_i^S,y_i^S)\}_{i = 1}^{N_S}$ drawn from a source data distribution $\mathcal{P}_S$ , and a set of unlabeled videos $\mathcal{D}_T\coloneqq \{\mathbf{x}_i^T\}_{i = 1}^{N_T}$ drawn from a target distribution $\mathcal{P}_T$ , our objective is to learn a function that can successfully classify videos drawn from $\mathcal{P}_T$ . More precisely, we seek to learn the parameters $\theta$ of a function $f_{\theta}$ , in our case a transformer neural network, that minimize the empirical risk on a sampling of videos from $\mathcal{P}_T$ . The challenge in domain adaptation lies in the fact that $\mathcal{P}_S$ and $\mathcal{P}_T$ differ from one another, leading to a gap in performance when directly applying a source trained model on target domain data. Thus, UDA techniques need to exploit information in unlabeled target domain samples to reduce this gap. + +# 3.2. Self-Supervised Initialization + +In contrast to some other video domain adaptation works [10, 65] that use network weights from Kinetics-400 supervised pre-training, we initialize our network from self-supervised pre-training. As contended in [43] for UDA with images, we argue that supervised pre-training can potentially complicate the study of VUDA techniques. When the pre-trained network has been trained to classify categories present in the DA dataset, some DA techniques could perform well simply by preserving the capabilities of the pretrained model despite not generalizing well to DA tasks that do not have category overlap with the pre-training dataset. Moreover, the UDA problem formulation implies that labeled instances of relevant classes are only available from the source domain. This condition does not hold for the Daily-DA and Sports-DA benchmarks that we use for evaluation, as 6 out of their 8 classes also appear in the Kinetics-400 dataset. Further exacerbating the issue with supervised Kinetics pre-training is the fact that Kinetics serves as a target domain in both benchmarks. Thus, we initialize the network weights from UMT [30] single-modality (i.e., video only) self-supervised pre-training on Kinetics-710 [29]. Please refer to Sec. 10 in the Appendix for select results using a UMT pre-trained network with additional supervised fine-tuning on Kinetics-400, which unsurprisingly achieves better baseline performance. + +# 4. Method + +# 4.1. UNITE + +Our approach to VUDA consists of three stages: (1) unsupervised pre-training on target domain data, (2) supervised fine-tuning on source domain data and (3) collaborative self-training using videos from both domains. We now + +![](images/25629477d8c1a7622f021e15ce30078ede1bf0cfa12500d298519ff3a12d09c8.jpg) +Figure 2. Unmasked Teacher (UMT) training used in Stage 1 of UNITE to perform unsupervised representation learning on target domain videos. + +describe each stage in detail. + +# Stage 1: Unsupervised Target Domain Pre-Training. + +To enable the extraction of discriminative features from target domain videos, the UNITE pipeline begins by performing additional unsupervised pre-training on the target dataset $\mathcal{D}_T$ . For this, we adopt the UMT [30] self-supervised training objective. UMT training uses a pretrained and frozen spatial teacher model $g_{*}$ to train a spatiotemporal student model $g_{a}$ by enforcing alignment between feature representations of the two networks at multiple layers (depicted in Figure 2). A key aspect of the training process is that the student model takes as input a masked video view $m(\mathbf{x}^T)$ , while the teacher model processes un-masked video frames from $\mathbf{x}^T$ . Through UMT training, $g_{a}$ must learn to match the semantically rich (albeit only spatially informed) features of $g_{*}$ from heavily masked target domain videos. To ensure sampling of meaningful video patches, UMT performs an attention-guided masking operation $m(\cdot)$ drawn from a multinomial distribution defined by the attention map of the final layer CLS token in $g_{*}$ . Attention-guided masking allows the use of a high masking ratio, which makes UMT training extremely efficient. We formalize the Stage 1 objective as follows. Let $\mathbf{z}_a^l$ denote the L2 normalized $l$ -th layer representation of $g_{a}$ for a masked input $m(\mathbf{x})$ , and let $\mathbf{z}_{*}^{l}$ denote the normalized $l$ -th layer representation of $g_{*}$ for the full input $\mathbf{x}$ , but only at the locations corresponding to the visible patches in $m(\mathbf{x})$ . The loss function in Stage 1 can be represented as: + +$$ +\mathcal {L} _ {\mathrm {U M T}} = \mathbb {E} _ {\mathbf {x} \sim \mathcal {P} _ {T}} \left[ \frac {1}{| \mathcal {A} |} \sum_ {l \in \mathcal {A}} \operatorname {M S E} \left(d ^ {l} (\mathbf {z} _ {a} ^ {l}), \mathbf {z} _ {*} ^ {l}\right) \right] +$$ + +where $\mathcal{A}$ is a set of aligned layers, $d^{l}(\cdot)$ is a linear projection + +![](images/5a022215bbdb0dc2965c75f00e905c5a975574e999a33989630ad220214d22dd.jpg) +Figure 3. Overview of the collaborative self-training stage in UNITE. The student and teacher models work together to produce more accurate pseudolabels for target domain videos. The target domain classification loss is enforced on masked target videos to encourage stronger context learning. Source domain classification is included to stabilize training, which is especially important at the beginning of training when pseudolabels may have low accuracy. + +for the $l$ -th layer and MSE indicates mean squared error. Following the original UMT method, we leverage a CLIP [44] ViT-B image encoder as $g_{*}$ . We keep the standard UMT masking ratio of $r = 0.8$ and perform alignment between the last 6 layers of the networks. $g_{a}$ is initialized as described in Sec. 3.2, along with the projections $d^{l}(\cdot)$ , which are kept frozen throughout Stage 1. + +Stage 2: Source Domain Fine-Tuning. The second stage of UNITE fits a classifier on source domain data using $g_{a}$ from Stage 1. We introduce label space information in a separate stage from target domain self-supervision because, as the authors of [17] observe for MAE training, we find that imposing a classification loss and UMT loss simultaneously can hinder convergence. To perform supervised fine-tuning, we introduce a linear classification head $h_{a}$ that operates on the mean pooled outputs of the final transformer layer in $g_{a}$ . The full student classifier network can now be represented as $f_{a} = h_{a}(g_{a}(\cdot))$ . As is the case with other masked pretraining strategies, we find that representations learned via UMT pre-training are not linearly separable. Thus, we perform full fine-tuning of $f_{a}$ by minimizing a cross entropy (CE) loss $\mathcal{L}_{CE}$ on unmasked source domain videos. The Stage 2 loss function is represented as: + +$$ +\mathcal {L} _ {\mathrm {S F T}} = \mathbb {E} _ {(\mathbf {x} ^ {S}, y ^ {S}) \sim \mathcal {P} _ {S}} \left[ \mathcal {L} _ {C E} \left(f _ {a} \left(\mathbf {x} ^ {S}\right), y ^ {S}\right) \right] \tag {1} +$$ + +To help preserve target domain feature extraction learned in Stage 1, we apply a layer-wise learning rate decay when fine-tuning on source domain data. + +Stage 3: Collaborative Self-Training (CST). After enhancing the ability of $g_{a}$ to extract meaningful features in the target domain (Stage 1) and fitting $f_{a}$ for the classification task using source videos (Stage 2), we further adapt $f_{a}$ to target data using a self-training process (depicted in Figure 3). In a conventional self-training framework (denoted as $\mathcal{L}_{\mathrm{ST}}$ in Eq. (2)), the overall loss is a sum of losses on labeled source domain samples and select target domain samples, chosen using mask $s(\cdot)$ . The target domain loss leverages the model's own predictions $\hat{y}$ (where $\hat{y} = \operatorname{argmax}_c \sigma(f(\mathbf{x}^T))[c]$ and $\sigma(\cdot)$ denotes the softmax function) in lieu of labels in the CE loss computation. Various strategies have been used to form the selection mask $s(\cdot)$ , with the goal of choosing target samples whose predictions are more likely to be correct [43, 48, 68]. + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {S T}} = \mathbb {E} _ {(\mathbf {x} ^ {S}, y ^ {S}) \sim \mathcal {P} _ {S}} \left[ \mathcal {L} _ {C E} \left(f _ {a} \left(\mathbf {x} ^ {S}\right), y ^ {S}\right) \right] + \tag {2} \\ \lambda \mathbb {E} _ {\mathbf {x} ^ {T} \sim \mathcal {P} _ {T}} \left[ s (\mathbf {x} ^ {T}) \mathcal {L} _ {C E} (f _ {a} (\mathbf {x} ^ {T}), \hat {y} _ {a} (\mathbf {x} ^ {T})) \right] \\ \end{array} +$$ + +We introduce a modified form of self-training, which we refer to as collaborative self-training (CST), that makes use of the student-teacher setup from Stage 1 to improve the process of pseudolabel estimation. Specifically, we use the zero-shot classification capabilities of the CLIP image teacher model $f_{*}$ (see Sec. 5.2 for details), in combination with predictions from the video student model $f_{a}$ , to create refined target pseudolabels for self-training. We employ the MatchOrConf scheme proposed in [69] to combine the outputs of the two models as shown in Eq. (3), where $\mathrm{Conf}(\hat{y})$ denotes the maximum softmax probability + +[20], $\max_c\sigma (f(\mathbf{x}^T))[c]$ , to create pseudolabel $\tilde{y}$ : + +$$ +\tilde {y} = \left\{ \begin{array}{l l} \hat {y} _ {a} & \text {i f} \hat {y} _ {a} = \hat {y} _ {*} \\ \hat {y} _ {a} & \text {i f} \hat {y} _ {a} \neq \hat {y} _ {*} \text {a n d} \operatorname {C o n f} \left(\hat {y} _ {a}\right) > \gamma \text {a n d} \operatorname {C o n f} \left(\hat {y} _ {*}\right) \leq \gamma , \\ \hat {y} _ {*} & \text {i f} \hat {y} _ {a} \neq \hat {y} _ {*} \text {a n d} \operatorname {C o n f} \left(\hat {y} _ {a}\right) \leq \gamma \text {a n d} \operatorname {C o n f} \left(\hat {y} _ {*}\right) > \gamma , \\ - 1 & \text {o t h e r w i s e .} \end{array} \right. \tag {3} +$$ + +$\mathbf{x}^T$ is an argument of the above, but is omitted for clarity. The MatchOrConf scheme selects only those target samples where there is agreement between the predictions of the two models, or where one model's confidence exceeds threshold $\gamma$ while the other's does not, resulting in the following selection mask: + +$$ +s \left(\mathbf {x} ^ {T}\right) = \left\{ \begin{array}{l l} 1 & \text {i f} \tilde {y} \left(\mathbf {x} ^ {T}\right) \neq - 1, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Taking inspiration from [22] and [43], we compute the target domain loss on masked videos, where we employ the same teacher-guided attention masking strategy (with $r = 0.8$ ) used during UMT pre-training in Stage 1. We find that training on masked videos in this stage results in substantial performance boosts compared to training on un-masked videos (see Table 6). Like [22], we include a loss weighting $q$ for each target video, based on the student network confidence: + +$$ +q \left(\mathbf {x} ^ {T}\right) = \operatorname {C o n f} \left(\hat {y} _ {a} \left(\mathbf {x} ^ {T}\right)\right) \tag {5} +$$ + +The complete loss function for the collaborative self-training stage can be expressed as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C S T}} = \mathbb {E} _ {(\mathbf {x} ^ {S}, y ^ {S}) \sim \mathcal {P} _ {S}} \left[ \mathcal {L} _ {C E} \left(f _ {a} (\mathbf {x} ^ {S}), y ^ {S}\right) \right] + \\ \lambda \mathbb {E} _ {(\mathbf {x} ^ {T}) \sim \mathcal {P} _ {T}} \left[ s (\mathbf {x} ^ {T}) q (\mathbf {x} ^ {T}) \mathcal {L} _ {C E} \left(f _ {a} (m (\mathbf {x} ^ {T})), \tilde {y} (\mathbf {x} ^ {T}) \right] \right. \tag {6} \\ \end{array} +$$ + +# 5. Experiments + +# 5.1. Datasets + +We evaluate our approach on three video domain adaptation benchmarks: Daily-DA, Sports-DA, UCF $\leftrightarrow$ HMDBfull. + +- Daily-DA [64] is a recently introduced benchmark composed of videos from four domains: ARID (A) [60], HMDB51 (H) [28], Moments-in-Time (M) [38] and Kinetics (K) [7, 26]. It consists of 8 overlapping classes representing everyday actions, with a total volume of approximately 19K videos. Videos from the ARID domain present a unique challenge in this benchmark, as they were explicitly filmed under low-illumination conditions. +- Sports-DA [64] contains videos from three domains: Sports-1M (S) [25], UCF101 (U) [49] and Kinetics (K) [7, 26]. It includes 23 overlapping categories depicting + +various sporting activities, with a total volume of approximately 41K videos. +- $\mathbf{UCF} \leftrightarrow \mathbf{HMDB}_{full}$ [8] contains approximately 3.2K videos across 12 action categories taken from the UCF101 (U) [49] and HMDB51 (K) [28] datasets. + +# 5.2. Implementation Details + +Frame Sampling. During training, we perform random uniform frame sampling [53] on each video. We divide the video into $T = 8$ uniform segments and randomly sample one frame from each. All frames are resized to $224 \times 224$ pixels, resulting in a $3 \times 8 \times 224 \times 224$ tensor for each video. During testing, we average the network predictions across 3 spatial crops and 4 temporal clips, where clip $i$ is formed by splitting each of the $T$ segments into 4 sub-segments and sampling a frame from the $i$ -th sub-segment of each. + +Network Architecture & Training. All of our experiments use a standard ViT-B/16 [12] architecture with a fixed sine and cosine positional encoding scheme. Our patch embedding projection operates on each frame independently, and self-attention is performed across both the space and time dimensions. All classification is performed using mean pooling of the final layer patch representations. We conduct our experiments using the PyTorch deep learning library [40] on 4 NVIDIA A5000 GPUs. + +During Stage 1, we perform UMT training for 50 epochs on the target dataset using a batch size of 256 and a learning rate of $1.5\mathrm{e - }5$ . During Stage 2, we use a batch size of 28, a learning rate of $2.5\mathrm{e - }5$ , and apply a layer-wise learning rate decay of 0.65. During stage 3, we use a batch size of 40 (20 videos from each domain), a learning rate of 1e-5, a MatchOrConf confidence threshold of $\gamma = 0.1$ and a target domain loss weight $\lambda = 1$ . Training in all stages uses the AdamW optimizer [35] and a weight decay of 0.05. Please refer to Sec. 8 in the Appendix for more training details for each stage. + +Zero-Shot Classification with CLIP. In Stage 3 of UNITE, we use CLIP (ViT-B/16) in a zero-shot manner to refine the pseudolabels of target domain videos. We do so by passing each of the $T$ video frames through the CLIP image encoder and computing the cosine similarity with a set of text representations produced by the CLIP text encoder. The inputs to the text encoder are the class names associated with the particular action recognition task, embedded in the following template: "A video of a person {class}." To form a video-level prediction, we simply average the softmax of the cosine similarity scores across the video frames. Refer to the Appendix (Sec. 11) for a listing of class names used for each task. + +
MethodTarget Domain Accuracy (Top-1%)
H→AM→AK→AA→HM→HK→HH→MA→MK→MM→KH→KA→KAvg.
DALL-V [67] LB17.534.715.614.644.647.925.515.535.761.645.117.831.3
Source Only40.452.136.549.668.357.941.536.343.379.348.041.749.6
ZSCLIP (ViT-B/16) [44]36.536.536.560.060.060.048.548.548.568.168.168.153.3
SFVUDAATCoN [62]17.927.217.226.747.348.230.717.232.557.748.531.033.5
EXTERNAL [63]26.218.123.926.253.755.840.718.235.268.157.651.439.6
DALL-V [67]24.024.024.057.965.452.547.045.747.078.176.775.051.4
UDADANN [18]14.222.821.220.143.337.529.519.721.758.838.227.029.5
MK-MMD [34]20.321.021.718.750.436.225.718.024.058.533.826.129.5
TA3N [8]14.421.619.914.943.037.725.715.631.555.538.423.428.5
UNITE w/o CST43.846.734.751.770.854.644.339.043.378.151.750.850.8
UNITE (Ours)48.044.137.567.974.265.851.850.048.089.969.963.659.2
DALL-V [67] UB26.926.926.970.470.470.461.561.561.588.988.988.961.9
Target Only68.568.568.584.684.684.673.073.073.098.398.398.381.1
+ +Table 1. UDA results on Daily-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]. + +
MethodTarget Domain Accuracy (Top-1 %)
U→SK→SS→UK→UU→KS→KAvg.
DALL-V [67] LB64.379.584.485.467.278.276.5
Source Only67.686.897.998.979.989.186.7
ZSCLIP (ViT-B/16) [44]85.085.093.393.391.391.389.9
SFVUDAATCoN [62]47.969.790.693.665.276.073.8
EXTERN [63]72.773.895.493.781.282.283.2
DALL-V [67]75.977.788.888.081.282.382.3
UDADANN [18]55.175.085.788.065.973.473.8
MK-MMD [34]55.667.990.990.266.173.674.0
TA3N [8]54.168.693.090.363.672.673.7
UNITE w/o CST73.987.398.299.284.690.288.9
UNITE (Ours)86.090.298.799.894.295.394.0
DALL-V [67] UB88.388.393.493.485.685.689.1
Target Only97.997.998.898.899.999.998.9
+ +Table 2. UDA results on Sports-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]. + +# 5.3. Baselines + +We compare results using UNITE with previously developed techniques in UDA, as well as in the closely related setting of SFUDA as reported by [67]. For Daily-DA and Sports-DA we compare against DANN [18], MK-MMD [34] and $\mathrm{TA}^3\mathrm{N}$ [8] in the UDA setting, and ATCoN [62], EXTERN [63] and DALL-V [67] for SFUDA. For $UCF\leftrightarrow HMDB_{full}$ , we additionally report accuracies from $\mathrm{CO}^2\mathrm{A}$ [52] and UDAVT [10]. + +# 5.4. Benchmark Results + +Tables 1, 2 and 3 show the results of our evaluation of UNITE on Daily-DA, Sports-DA and $UCF \leftrightarrow HMDB_{full}$ . We include accuracies after Stage 2 of the pipeline (denoted as UNITE w/o CST) as well as after applying the full UNITE process. For reference, we include accuracies using our UMT pre-trained network with source domain supervised fine-tuning (Source Only) and with target domain + +supervised fine-tuning (Target Only). We also include the source only and target only accuracies (denoted in the tables as LB and UB) of DALL-V [67] (a CLIP-based ResNet-50 model) for added context. We additionally display the zero-shot classification accuracy of the CLIP teacher model that we utilize during Stage 1 and Stage 3 of UNITE. + +An observation we make across all benchmarks is the enhanced baseline accuracy of the UMT pre-trained model, which speaks to the strength of the unsupervised masked distillation process and the ViT architecture. On the Daily-DA benchmark, we find that UNITE exceeds previously reported results on most domain shifts. On only one shift $(\mathrm{M}\rightarrow \mathrm{A})$ we observe degradation in target accuracy after employing UNITE compared to source only, which is consistent with previous work. + +UNITE exhibits even stronger performance on the larger-scale Sports-DA benchmark, where it exceeds previous results by a large margin. We also see more consistent and + +
MethodAccuracy (%)
H→UU→HAvg.
SDALL-V [67] LB71.676.173.8
Source Only88.880.084.4
CLIP (ViT-B/16) [44]88.891.790.3
SEVUDAATCoN [62]85.379.782.5
EXTERN [63]91.988.990.4
DALL-V [67]93.188.991.0
UDADANN [18]74.475.174.8
MK-MMD [34]74.779.777.2
TA3N [8]78.184.881.5
CO2A [52]95.887.891.8
UDAVT [10]96.892.394.6
UNITE w/o CST92.183.687.9
UNITE (Ours)92.595.093.8
DALL-V [67] UB93.791.492.6
Target Only99.796.798.2
+ +![](images/3f5d2d4393376e31c3a9748c11b00f54c5a083a9fe8cd79fbfd4c40496b10d06.jpg) +Figure 4. Class-wise accuracy on ARID $\rightarrow$ HMDB. Performances are shown for the CLIP teacher model and the student model before and after the collaborative self-training stage (Stage 3). For most classes, the Stage 3 student exceeds the accuracy of both the Stage 2 student and the CLIP teacher. + +substantial improvement after the UMT target pre-training stage alone (UNITE w/o CST), suggesting that overtraining may be occurring on the smaller datasets in Daily-DA. + +On $UCF \leftrightarrow HMDB_{full}$ we see state-of-the-art performance on $\mathrm{U} \rightarrow \mathrm{H}$ , while UDAVT outperforms on $\mathrm{H} \rightarrow \mathrm{U}$ . Nevertheless, UNITE still achieves strong overall performance on the benchmark. + +# 6. Additional Analysis & Discussion + +Student exceeds the teacher. We see in the benchmark results that the video student model exceeds the zero-shot accuracy of the CLIP teacher in virtually all domain shifts. This suggests that UNITE successfully leverages the spatial modeling capabilities of CLIP to train an even more powerful spatiotemporal model for the action recognition task. + +Table 3. UDA results on $UCF \leftrightarrow HMDB_{full}$ . Rows in color use our UMT backbone, and rows without color are reported from [61]. + +
MethodPTCSTAccuracy (%)
H→AA→H
Source OnlyXX40.449.6
+ Target UMT Pre-Training✓X43.851.7
+ Collaborative Self-TrainingX✓42.060.8
UNITE✓✓48.067.9
+ +Table 4. Ablation of stages in UNITE on ARID ↔ HMDB from Daily-DA. PT and CST denote UMT pre-training and collaborative self-training, respectively. + +We can see from Figure 4 that the teacher model and the student model prior to collaborative self-training have differing capabilities. For example, the image teacher is stronger at recognizing actions from the jump class than the Stage 2 student. On the other hand, the video student is significantly better than the teacher at recognizing wave at the start of CST. In both cases, we see that the student model, through collaborative self-training, comes to exceeds both the Stage 2 student and the CLIP image teacher. With the MatchOr-Conf pseudolabeling scheme, the two models are able to work together to achieve higher accuracy in the target domain. We see accuracy on the pick and run classes declines after CST, as self-training tends to reinforce the accuracy of well-performing classes, sometimes at the expense of more difficult ones that are infrequently assigned pseudolabels. Modifying CST to account for the learning status of each class, like in [68], may help remedy this issue. + +Impact of Stage 1 and 3. In Table 4 we ablate the stages in UNITE. While UMT pre-training and collaborative self-training each offer a benefit in the absence of the other, the largest gains in domain transfer performance are obtained when combining the two stages. For example, on the ARID $\rightarrow$ HMDB domain shift in Daily-DA, Stage 1 and Stage 3 alone improve upon source only accuracy by $+1.8\%$ and $+11.2\%$ respectively. Meanwhile, the full UNITE pipeline results in $+18.3\%$ accuracy improvement on the target domain, suggesting that the masked pre-training and masked self-training processes complement one another. + +Impact of data domains during pre-training. In Table 5, we assess the impact of utilizing each of the two data domains during UMT pre-training on UDA performance. While the results vary across domain shifts, we observe the most consistent improvements when pre-training only on target domain videos. Although incorporating both domains during unsupervised pre-training has been successfully applied in image-based domain adaptation [27, 37, 43, 46], we find it to be suboptimal with UMT training on the VUDA tasks studied in this work. + +Effect of masking during self-training. In Table 6, we see that enforcing the target domain CE loss on masked videos results in increased performance compared to using unmasked target videos, with smaller benefits for $\mathrm{H}\rightarrow \mathrm{A}$ and much larger benefits for $\mathrm{A}\rightarrow \mathrm{H}$ . We believe the benefits of + +
UMT Pre-Training DataAccuracy (%)
H→AA→H
None42.060.8
Source Only48.360.0
Target Only48.067.9
Source + Target45.267.1
+ +Table 5. Comparison of data domains used during UMT pretraining. Accuracies are reported after the self-training stage. Bold indicates best accuracy and underline indicates second best. + +
Data for Target CE LossAccuracy (%)
H→AA→H
Unmasked46.654.6
Masked48.067.9
+ +Table 6. Effect of applying the target domain cross entropy loss on masked vs. unmasked target videos during the self-training stage. + +
Pseudolabeling StrategyAccuracy (%)
H→AA→H
Cons (k=2)39.857.5
ConsOrConf (k=2,T=0.5) [43]48.656.3
ConsAndConf (k=2,T=0.5)40.157.1
Zero-Shot CLIP Only (T=0.5)34.751.3
MatchOrConf (γ=0.1) [69]48.067.9
MatchAndConsOrConf (k=2,γ=0.1)47.460.8
+ +masking arise from forcing the model to use different cues from the same video to predict the pseudolabel, resulting in more robust target domain recognition. As we saw in Table 4, the masked pre-training stage plays a key role in unlocking the benefits of masked self-training. + +Comparison of pseudolabeling strategies. In Table 7, we explore the use of alternative pseudolabeling strategies to the MatchOrConf [69] scheme employed in UNITE. The first 3 rows of the table assess variations of the masked consistency pseudolabeling scheme proposed in PACMAC [43], where target samples are selected if there is agreement between predictions of $k$ masked views. This criterion can also be combined with a constraint on the confidence exceeding threshold $T$ . While we find the combination of these two constraints (ConsOrConf) to be effective for self-training, we find that MatchOrConf results in more consistent target accuracy improvements across domain shifts. We also consider combining MatchOrConf with a masked consistency constraint as a way to reduce cases where the student and teacher models agree on an incorrect prediction. This scheme, denoted as MatchAndConsOrConf, adds an additional constraint to the matching criterion by also requiring agreement between the student model predictions of $k$ masked views. We find that this approach underperforms compared to the vanilla MatchOrConf scheme due to low utilization of target domain samples. Our results also indicate that using both models together for pseudolableing outperforms using only zero-shot CLIP predictions. + +Table 7. Comparison of pseudolabeling strategies. + +
Loss During Self-TrainingAccuracy (%)
H→AA→H
LCE(m(xT), y)33.742.1
LCE(m(xT), y) + LCE(xS, yS)48.067.9
+ +Table 8. Effect of including a source domain classification loss in addition to the target domain loss during the self-training stage. + +
Confidence ThresholdAccuracy (%)
H→AA→H
0.148.067.9
0.344.155.4
0.551.964.6
0.742.665.8
0.940.762.5
+ +Table 9. Effect of confidence threshold $\gamma$ in MatchOrConf pseudolabeling scheme from Equation 3. + +Effect of source classification during self-training. In Table 8, we study the importance of including a source domain classification loss while performing collaborative self-training on masked target videos. Indeed, we find that source classification is a crucial ingredient in this stage, as excluding it destabilizes training and results in performance degradation rather than improvement. We suspect that the ground truth labels on source data help to combat the effects of inaccurate target pseudolabels, which may be prevalent especially at the start of training. + +Choice of pseudolabeling confidence threshold. In Table 9, we compare adaptation performance for various values of confidence threshold in the MatchOrConf pseudolabeling scheme. We observe significant variation across domain shifts, but find that $\gamma = 0.1$ leads to relatively consistent improvement. We leave for future work the development of a more principled approach to determining an appropriate confidence threshold for a given domain shift. + +# 7. Conclusions + +In this work, we presented UNITE, a three step process for unsupervised video domain adaptation. Our approach leverages a powerful spatial encoder to aid in the adaptation of a spatiotemporal network using unlabeled target domain videos—making significant progress towards bridging the domain gaps in various VUDA datasets. The advancements demonstrated by UNITE underscore the untapped potential of masked modeling techniques for video domain adaptation, which we hope will spur further research in this area. + +Acknowledgements. This research was sponsored by the Army Research Laboratory under Cooperative Agreement W911NF-21-2-0211. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 1 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked Siamese Networks for Label-Efficient Learning. In ECCV, 2022. 2 +[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. In CVPR, 2023. +[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In ICML, 2022. 2 +[5] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In ICML, pages 813-824, 2021. 1 +[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In CVPR, 2017. 1 +[7] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A Short Note about Kinetics-600, 2018. 5 +[8] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal Attentive Alignment for Large-Scale Video Domain Adaptation. In ICCV, 2019. 2, 5, 6, 7 +[9] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 iee. In CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 1 +[10] Victor G. Turrisi Da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Unsupervised Domain Adaptation for Video Transformers in Action Recognition. In ICPR, pages 1258-1265, Montreal, QC, Canada, 2022. 2, 3, 6, 7 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5 +[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale Vision Transformers. In ICCV, 2021. 1 +[14] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In CVPR, 2020. 1 + +[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In ICCV, 2019. 1 +[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In NeurIPS, 2022. 2 +[17] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 4 +[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 6, 7 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2021. 2 +[20] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 5 +[21] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 2 +[22] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation. In CVPR, pages 11721–11732, Vancouver, BC, Canada, 2023. 2, 5 +[23] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661. Springer, 2016. 1 +[24] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to Hide from Your Students: Attention-Guided Masked Image Modeling. In ECCV, pages 300-318, 2022. 2 +[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-Scale Video Classification with Convolutional Neural Networks. In CVPR, pages 1725–1732, 2014. 5 +[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5 +[27] Donghyun Kim, Kuniaki Saito, Tae-Hyun Oh, Bryan A Plummer, Stan Sclaroff, and Kate Saenko. Cds: Cross-domain self-supervised pre-training. In ICCV, pages 9123-9132, 2021. 7 +[28] H Kuehne, H Jhuang, E Garrote, T Poggio, and T Serre. HMDB: A Large Video Database for Human Motion Recognition. In ICCV, 2011. 5 +[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In ICCV, pages 1632-1643, 2023. 3 +[30] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked Teacher: Towards Training-Efficient Video Foundation Models. In ICCV, 2023. 2, 3 + +[31] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. MST: Masked Self-Supervised Transformer for Visual Representation. In NeurIPS, pages 13165-13176, 2021. 2 +[32] Han Lin, Guangxing Han, Jiawei Ma, Shiyuan Huang, Xudong Lin, and Shih-Fu Chang. Supervised masked knowledge distillation for few-shot transformers. In CVPR, pages 19649-19659, 2023. 2 +[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3192-3201, New Orleans, LA, USA, 2022. 1 +[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 2, 6, 7 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5, 1 +[36] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-CLIP: End-to-End Multi-grained Contrastive Learning for Video-Text Retrieval. In ACM MM, 2022. 1 +[37] Samarth Mishra, Kate Saenko, and Venkatesh Saligrama. Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency, 2021. 7 +[38] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, and Dan Gutfreund. Moments in Time Dataset: One Million Videos for Event Understanding. PAMI, 42(2):502-508, 2020. 5 +[39] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video Transformer Network. In ICCV, 2021. 1 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5 +[41] Vishal M Patel, Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. Visual Domain Adaptation: A survey of recent advances. IEEE Signal Processing Magazine, 32(3):53-69, 2015. 1 +[42] Aj Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning. In CVPR, pages 2214-2224, Vancouver, BC, Canada, 2023. 1 +[43] Viraj Prabhu, Sriram Yenamandra, Aaditya Singh, and Judy Hoffman. Adapting self-supervised vision transformers by probing attention-conditioned masking consistency. NeurIPS, 35:23271-23283, 2022. 2, 3, 4, 5, 7, 8, 1 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 4, 6, 7 +[45] Aadarsh Sahoo, Rutav Shah, R. Panda, Kate Saenko, and Abir Das. Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing. In NeurIPS, 2021. 2 + +[46] Kendrick Shen, Robbie M. Jones, Ananya Kumar, Sang Michael Xie, Jeff Z. Haochen, Tengyu Ma, and Percy Liang. Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation. In ICML, pages 19847-19878, 2022. 7 +[47] Karen Simonyan and Andrew Zisserman. Two-Stream Convolutional Networks for Action Recognition in Videos. In NeurIPS, 2014. 1 +[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In NeurIPS, pages 596-608, 2020. 4 +[49] K. Soomro, A. Zamir, and M. Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. ArXiv, 2012. 5 +[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In NeurIPS, 2022. 2 +[51] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning Spatiotemporal Features with 3D Convolutional Networks. In ICCV, pages 4489-4497, Santiago, Chile, 2015. 1 +[52] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-Head Contrastive Domain Adaptation for Video Action Recognition. In WACV, pages 2234–2243, 2022. 2, 6, 7 +[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal Segment Networks for Action Recognition in Videos. PAMI, 2017. 1, 5 +[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. In CVPR, pages 14549-14560, Vancouver, BC, Canada, 2023. 2 +[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition, 2021. 1 +[56] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning. In CVPR, pages 6312-6322, Vancouver, BC, Canada, 2023. 2 +[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. InternVideo: General Video Foundation Models via Generative and Discriminative Learning, 2022. 1 +[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2 + +[59] Yuecong Xu, Haozhi Cao, Zhenghua Chen, Xiaoli Li, Lihua Xie, and Jianfei Yang. Video Unsupervised Domain Adaptation with Deep Learning: A Comprehensive Survey, 2022. 2 +[60] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. ARID: A New Dataset for Recognizing Action in the Dark. In Communications in Computer and Information Science, 2022. 5 +[61] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Aligning Correlation Information for Domain Adaptation in Action Recognition. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. 6, 7 +[62] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-Free Video Domain Adaptation by Learning Temporal Consistency for Action Recognition. In ECCV, pages 147-164, Cham, 2022. 2, 6, 7 +[63] Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Extern: Leveraging endo-temporal regularization for black-box video domain adaptation. arXiv preprint arXiv:2208.05187, 2022. 2, 6, 7 +[64] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Zhengguo Li, and Zhenghua Chen. Multi-Source Video Domain Adaptation With Temporal Attentive Moment Alignment Network. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3860-3871, 2023. 5 +[65] Yuecong Xu, Jianfei Yang, Yunjiao Zhou, Zhenghua Chen, Min Wu, and Xiaoli Li. Augmenting and Aligning Snippets for Few-Shot Video Domain Adaptation. In ICCV, 2023. 3 +[66] Hongwei Xue, Peng Gao, Hongyang Li, Yu Qiao, Hao Sun, Houqiang Li, and Jiebo Luo. Stare at What You See: Masked Image Modeling without Reconstruction. In CVPR, pages 22732-22741, Vancouver, BC, Canada, 2023. 2 +[67] Giacomo Zara, Alessandro Conti, Subhankar Roy, Stephane Lathuiliere, Paolo Rota, and Elisa Ricci. The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation. In ICCV, 2023. 1, 2, 6, 7 +[68] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 34:18408-18419, 2021. 4, 7 +[69] Wenyu Zhang, Li Shen, and Chuan-Sheng Foo. Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation. In CVPR, 2023. 4, 8 +[70] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT Pre-training with Online Tokenizer. In ICLR, 2022. 2 \ No newline at end of file diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/images.zip b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d009851b6bb6e9e777120dde3ce7232f15b3db1e --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51df820df517ce1225d4fe018c2191669e1364c6b5dd841100b714fd13ec46b5 +size 502221 diff --git a/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/layout.json b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1df8f0eace1aec035786cabcaff533deb870aa0 --- /dev/null +++ b/2024/Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training/layout.json @@ -0,0 +1,8922 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 518, + 140 + ], + "type": "text", + "content": "Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": "Arun Reddy" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": ", William Paul" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Corban Rivera" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Ketul Shah" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Celso M. de Melo" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Rama Chellappa" + }, + { + "bbox": [ + 58, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 177, + 397, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 177, + 397, + 189 + ], + "spans": [ + { + "bbox": [ + 196, + 177, + 397, + 189 + ], + "type": "text", + "content": "1 Johns Hopkins University Applied Physics Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 190, + 439, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 190, + 439, + 202 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 439, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 190, + 439, + 202 + ], + "type": "text", + "content": "Johns Hopkins University, Department of Electrical & Computer Engineering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 204, + 378, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 204, + 378, + 217 + ], + "spans": [ + { + "bbox": [ + 214, + 204, + 378, + 217 + ], + "type": "text", + "content": "3DEVCOM U.S. Army Research Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 290, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 450 + ], + "type": "text", + "content": "In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 128, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "In recent years, the field of video action recognition has undergone significant advancement, largely driven by deep learning-based techniques. These include approaches based on convolutional neural networks (CNNs) [6, 14, 15, 47, 51, 53] and vision transformers (ViTs) [1, 5, 13, 33, 39, 42]. Further enriching this landscape, models integrating video and language data have emerged [36, 55, 57], which effectively leverage large collections of captioned videos to achieve impressive video understanding capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 606, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 700 + ], + "type": "text", + "content": "Despite increasing performance on benchmark datasets, deep networks for video action recognition suffer from the same fundamental challenges as other machine learning models in the presence of distribution shift. This phenomenon, where a model's performance degrades when applied to data distributions different from those it was trained on, remains a critical obstacle in deploying the models in varied real-world scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "content": "Domain adaptation (DA) seeks to mitigate the effects of" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 243, + 535, + 395 + ], + "blocks": [ + { + "bbox": [ + 319, + 243, + 535, + 395 + ], + "lines": [ + { + "bbox": [ + 319, + 243, + 535, + 395 + ], + "spans": [ + { + "bbox": [ + 319, + 243, + 535, + 395 + ], + "type": "image", + "image_path": "983754bd8c0575bab4a022d6803979f33aa1e57743d7a9d4b106a10e23bf56e0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 403, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 403, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 547, + 449 + ], + "type": "text", + "content": "Figure 1. Overview of the UNITE pipeline for video UDA. A teacher model is used to guide the self-supervised learning process in the first stage, and is again used to improve pseudolabeling of target videos during self-training in the third stage." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 688 + ], + "type": "text", + "content": "distribution shift and has been a major research area within the computer vision community for years [41]. A subproblem within DA is unsupervised domain adaptation (UDA), which tries to make use of unlabeled data from the target domain to improve transfer performance. UDA approaches specific to video have also been developed, with many focusing on explicit alignment of the source and target domains using adversarial techniques or domain discrepancy measures. In contrast, we present here a method for video UDA that relies on self-supervised masked pre-training and self-training, both aided by a powerful image-based teacher model (CLIP [44]). The authors of DALL-V [67] demonstrate the effectiveness of CLIP in the source-free video domain adaptation setting by directly adapting it to action recognition tasks despite fundamentally being an image-based technique. Instead, we leverage CLIP as a spatial teacher for training a spatiotemporal student model that can better operate on target domain videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "This paper presents, to the best of our knowledge, the first exploration of self-supervised masked distillation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18919" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "techniques in the context of unsupervised video domain adaptation. Our approach, Unsupervised Adaptation with Teacher-Enhanced Learning (UNITE), successfully harnesses the capabilities of an image-based teacher in two ways: (1) to form the target representations in a masked self-supervised pre-training stage on target domain videos, and (2) to improve the quality of pseudolabels used during self-training in collaboration with the student video network (see Figure 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 274, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 274, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 274, + 204 + ], + "type": "text", + "content": "We summarize the contributions of our work as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 286, + 372 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 205, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 286, + 240 + ], + "type": "text", + "content": "- We introduce the UNITE pipeline for video domain adaptation, which consists of a novel combination of masked video modeling and self-training techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 241, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 286, + 300 + ], + "type": "text", + "content": "- We conduct extensive experimentation by evaluating UNITE on three video domain adaptation benchmarks (Daily-DA, Sports-DA and UCF-HMDB) and observe significant performance gains compared to previously reported results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 301, + 286, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 286, + 372 + ], + "type": "text", + "content": "- We present a series of ablation experiments that study the effectiveness of various aspects of the UNITE pipeline and assess alternative design choices. In particular, we show that masked distillation pre-training and masked self-training are best applied in concert to achieve the strongest domain transfer performance." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": "Video Unsupervised Domain Adaptation (VUDA). Techniques for video unsupervised domain adaptation typically fall within a few common categories [59]. Adversarial methods attempt to align representations between the source and target domains by enforcing a domain confusion loss using a separate domain discriminator network, where the main network must learn to fool the discriminator. Some techniques which were developed for images, such as DANN [18] or MK-MMD [34], can be applied directly to video architectures. " + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^3\\mathrm{N}" + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": " [8] is a video-specific approach that uses separate adversaries for the spatial and temporal dimensions. These approaches can result in unstable training due to competing objectives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Another class of methods use contrastive learning and exploit the intrinsic structure of video. CoMix [45] uses contrastive learning with pseudolabeling and video-based augmentations such as sampling at varying frame rates or mixing the backgrounds of source videos and target videos. " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}^2\\mathrm{A}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " [52] optimizes 6 different losses simultaneously, including contrastive losses at the clip- and video-level, along with supervised contrastive learning across domains. UDAVT [10] adapts vision transformers using the information bottleneck principle by having separate projectors for source and target features and estimating a cross correlation matrix between features whose labels and pseudolabels match." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "Some video domain adaptation techniques have been developed to address the related problem of source-free video video domain adaptation (SFVUDA), which focuses on scenarios where source data is unavailable during the adaptation phase. ATCoN [62] tries to ensure the trained model is as consistent with its features as the original model. EX-TERN [63] additionally uses masking augmentations as a factor for features to be consistent over. DALL-V [67] takes a different approach by adapting a CLIP [44] image encoder to source and target datasets using an adapter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 194, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 361 + ], + "type": "text", + "content": "Masked Image & Video Modeling. Masked modeling first achieved success in natural language processing [11], and has since been applied to image and video data as a way to learn powerful feature encoding from unlabeled data. The basic idea behind masked modeling is to partially occlude patches of the input data and train a model to reconstruct the missing patches. This can be done at the pixel-level [16, 19, 31, 50, 54, 58], or the targets of the reconstruction can be formed by a neural network—either an exponential moving average (EMA) of the network being trained [2-4, 24, 32, 70], or even a separate teacher model. MVD [56] uses the latter concept to train a video transformer network using both image and video teachers, which can be viewed as a form of knowledge distillation [21]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 508 + ], + "type": "text", + "content": "Unmasked Teacher (UMT) [30] uses a different form of masked distillation to train powerful video encoders, where instead of reconstructing representations of masked regions, the student attempts to match the teacher representations of visible patches [66]. UMT is able to use a spatial encoder (CLIP [44]) to train a powerful spatiotemporal model. Because the video encoder only needs to process visible patches, and there is no need for an expensive decoder for reconstruction, UMT offers a highly efficient approach for self-supervision on videos. As such, we choose to leverage the UMT masked distillation objective to perform unsupervised video learning in UNITE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Masked Modeling for Domain Adaptation. Masked modeling is fundamentally about ensuring a given feature is invariant to whether its corresponding input is masked or not. Inducing such invariances on novel domains can be a useful property to enforce for the purposes of domain adaptation. MAE-TTT [17] finetunes a masked autoencoder [19] on single images at test time, adapting the feature extractor to the statistics of the current image while leaving the prediction head unchanged. MIC [22] uses an EMA teacher model on unmasked images to produce pseudolabels for training on masked target images. PACMAC [43] selects target domain samples for pseudolabeling based on prediction consistency over masked views, and also applies the classification loss to masked target inputs. Since these approaches have been developed for image-based tasks, there is still an open question of how similar concepts could be applied to video recognition." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18920" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 172, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 172, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 172, + 102 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": "We formalize the problem of unsupervised domain adaptation for action recognition as follows. Given a set of labeled videos " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S\\coloneqq \\{(\\mathbf{x}_i^S,y_i^S)\\}_{i = 1}^{N_S}" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": " drawn from a source data distribution " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_S" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": ", and a set of unlabeled videos " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T\\coloneqq \\{\\mathbf{x}_i^T\\}_{i = 1}^{N_T}" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": " drawn from a target distribution " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": ", our objective is to learn a function that can successfully classify videos drawn from " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": ". More precisely, we seek to learn the parameters " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": " of a function " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": ", in our case a transformer neural network, that minimize the empirical risk on a sampling of videos from " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": ". The challenge in domain adaptation lies in the fact that " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_S" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T" + }, + { + "bbox": [ + 46, + 110, + 289, + 291 + ], + "type": "text", + "content": " differ from one another, leading to a gap in performance when directly applying a source trained model on target domain data. Thus, UDA techniques need to exploit information in unlabeled target domain samples to reduce this gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 206, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 206, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 206, + 310 + ], + "type": "text", + "content": "3.2. Self-Supervised Initialization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 617 + ], + "type": "text", + "content": "In contrast to some other video domain adaptation works [10, 65] that use network weights from Kinetics-400 supervised pre-training, we initialize our network from self-supervised pre-training. As contended in [43] for UDA with images, we argue that supervised pre-training can potentially complicate the study of VUDA techniques. When the pre-trained network has been trained to classify categories present in the DA dataset, some DA techniques could perform well simply by preserving the capabilities of the pretrained model despite not generalizing well to DA tasks that do not have category overlap with the pre-training dataset. Moreover, the UDA problem formulation implies that labeled instances of relevant classes are only available from the source domain. This condition does not hold for the Daily-DA and Sports-DA benchmarks that we use for evaluation, as 6 out of their 8 classes also appear in the Kinetics-400 dataset. Further exacerbating the issue with supervised Kinetics pre-training is the fact that Kinetics serves as a target domain in both benchmarks. Thus, we initialize the network weights from UMT [30] single-modality (i.e., video only) self-supervised pre-training on Kinetics-710 [29]. Please refer to Sec. 10 in the Appendix for select results using a UMT pre-trained network with additional supervised fine-tuning on Kinetics-400, which unsurprisingly achieves better baseline performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 627, + 103, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 103, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 103, + 639 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "type": "text", + "content": "4.1. UNITE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Our approach to VUDA consists of three stages: (1) unsupervised pre-training on target domain data, (2) supervised fine-tuning on source domain data and (3) collaborative self-training using videos from both domains. We now" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 541, + 241 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 541, + 241 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 541, + 241 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 541, + 241 + ], + "type": "image", + "image_path": "25629477d8c1a7622f021e15ce30078ede1bf0cfa12500d298519ff3a12d09c8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 255, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 305, + 255, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 547, + 288 + ], + "type": "text", + "content": "Figure 2. Unmasked Teacher (UMT) training used in Stage 1 of UNITE to perform unsupervised representation learning on target domain videos." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 308, + 424, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 424, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 424, + 319 + ], + "type": "text", + "content": "describe each stage in detail." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 331, + 531, + 343 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 531, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 531, + 343 + ], + "type": "text", + "content": "Stage 1: Unsupervised Target Domain Pre-Training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": "To enable the extraction of discriminative features from target domain videos, the UNITE pipeline begins by performing additional unsupervised pre-training on the target dataset " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ". For this, we adopt the UMT [30] self-supervised training objective. UMT training uses a pretrained and frozen spatial teacher model " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{*}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " to train a spatiotemporal student model " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " by enforcing alignment between feature representations of the two networks at multiple layers (depicted in Figure 2). A key aspect of the training process is that the student model takes as input a masked video view " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "m(\\mathbf{x}^T)" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ", while the teacher model processes un-masked video frames from " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^T" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ". Through UMT training, " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " must learn to match the semantically rich (albeit only spatially informed) features of " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{*}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " from heavily masked target domain videos. To ensure sampling of meaningful video patches, UMT performs an attention-guided masking operation " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " drawn from a multinomial distribution defined by the attention map of the final layer CLS token in " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{*}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ". Attention-guided masking allows the use of a high masking ratio, which makes UMT training extremely efficient. We formalize the Stage 1 objective as follows. Let " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_a^l" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " denote the L2 normalized " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": "-th layer representation of " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " for a masked input " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "m(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{*}^{l}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " denote the normalized " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": "-th layer representation of " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "g_{*}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": " for the full input " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ", but only at the locations corresponding to the visible patches in " + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "inline_equation", + "content": "m(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 344, + 545, + 654 + ], + "type": "text", + "content": ". The loss function in Stage 1 can be represented as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 660, + 518, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 660, + 518, + 694 + ], + "spans": [ + { + "bbox": [ + 332, + 660, + 518, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U M T}} = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {P} _ {T}} \\left[ \\frac {1}{| \\mathcal {A} |} \\sum_ {l \\in \\mathcal {A}} \\operatorname {M S E} \\left(d ^ {l} (\\mathbf {z} _ {a} ^ {l}), \\mathbf {z} _ {*} ^ {l}\\right) \\right]", + "image_path": "e950684401d7c66612e98f2d5dac1402237175340cc6a6be03b484b9f92fa5b3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "content": " is a set of aligned layers, " + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "d^{l}(\\cdot)" + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "content": " is a linear projection" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18921" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 69, + 473, + 273 + ], + "blocks": [ + { + "bbox": [ + 122, + 69, + 473, + 273 + ], + "lines": [ + { + "bbox": [ + 122, + 69, + 473, + 273 + ], + "spans": [ + { + "bbox": [ + 122, + 69, + 473, + 273 + ], + "type": "image", + "image_path": "5a022215bbdb0dc2965c75f00e905c5a975574e999a33989630ad220214d22dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "type": "text", + "content": "Figure 3. Overview of the collaborative self-training stage in UNITE. The student and teacher models work together to produce more accurate pseudolabels for target domain videos. The target domain classification loss is enforced on masked target videos to encourage stronger context learning. Source domain classification is included to stabilize training, which is especially important at the beginning of training when pseudolabels may have low accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": "for the " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": "-th layer and MSE indicates mean squared error. Following the original UMT method, we leverage a CLIP [44] ViT-B image encoder as " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "g_{*}" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ". We keep the standard UMT masking ratio of " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "r = 0.8" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": " and perform alignment between the last 6 layers of the networks. " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": " is initialized as described in Sec. 3.2, along with the projections " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "d^{l}(\\cdot)" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ", which are kept frozen throughout Stage 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": "Stage 2: Source Domain Fine-Tuning. The second stage of UNITE fits a classifier on source domain data using " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": " from Stage 1. We introduce label space information in a separate stage from target domain self-supervision because, as the authors of [17] observe for MAE training, we find that imposing a classification loss and UMT loss simultaneously can hinder convergence. To perform supervised fine-tuning, we introduce a linear classification head " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "h_{a}" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": " that operates on the mean pooled outputs of the final transformer layer in " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": ". The full student classifier network can now be represented as " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "f_{a} = h_{a}(g_{a}(\\cdot))" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": ". As is the case with other masked pretraining strategies, we find that representations learned via UMT pre-training are not linearly separable. Thus, we perform full fine-tuning of " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": " by minimizing a cross entropy (CE) loss " + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 444, + 288, + 635 + ], + "type": "text", + "content": " on unmasked source domain videos. The Stage 2 loss function is represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 646, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 646, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 82, + 646, + 287, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S F T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] \\tag {1}", + "image_path": "b2a50e2e555a05622fcc772db3c37ce9126bc1d7bd9444bd7e5b1857cf740b01.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "type": "text", + "content": "To help preserve target domain feature extraction learned in Stage 1, we apply a layer-wise learning rate decay when fine-tuning on source domain data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": "Stage 3: Collaborative Self-Training (CST). After enhancing the ability of " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " to extract meaningful features in the target domain (Stage 1) and fitting " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " for the classification task using source videos (Stage 2), we further adapt " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " to target data using a self-training process (depicted in Figure 3). In a conventional self-training framework (denoted as " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " in Eq. (2)), the overall loss is a sum of losses on labeled source domain samples and select target domain samples, chosen using mask " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": ". The target domain loss leverages the model's own predictions " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\operatorname{argmax}_c \\sigma(f(\\mathbf{x}^T))[c]" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": " denotes the softmax function) in lieu of labels in the CE loss computation. Various strategies have been used to form the selection mask " + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 304, + 348, + 545, + 529 + ], + "type": "text", + "content": ", with the goal of choosing target samples whose predictions are more likely to be correct [43, 48, 68]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 330, + 544, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 544, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 330, + 544, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} \\left(\\mathbf {x} ^ {S}\\right), y ^ {S}\\right) \\right] + \\tag {2} \\\\ \\lambda \\mathbb {E} _ {\\mathbf {x} ^ {T} \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} (f _ {a} (\\mathbf {x} ^ {T}), \\hat {y} _ {a} (\\mathbf {x} ^ {T})) \\right] \\\\ \\end{array}", + "image_path": "5e2e31b13c6c1e002efb8ce365fbe5c388bab96cf4022483d0518f8aa3eba0aa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We introduce a modified form of self-training, which we refer to as collaborative self-training (CST), that makes use of the student-teacher setup from Stage 1 to improve the process of pseudolabel estimation. Specifically, we use the zero-shot classification capabilities of the CLIP image teacher model " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{*}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " (see Sec. 5.2 for details), in combination with predictions from the video student model " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", to create refined target pseudolabels for self-training. We employ the MatchOrConf scheme proposed in [69] to combine the outputs of the two models as shown in Eq. (3), where " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Conf}(\\hat{y})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " denotes the maximum softmax probability" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18922" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "text", + "content": "[20], " + }, + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "inline_equation", + "content": "\\max_c\\sigma (f(\\mathbf{x}^T))[c]" + }, + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "text", + "content": ", to create pseudolabel " + }, + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 103, + 297, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 297, + 171 + ], + "type": "interline_equation", + "content": "\\tilde {y} = \\left\\{ \\begin{array}{l l} \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} = \\hat {y} _ {*} \\\\ \\hat {y} _ {a} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) > \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) \\leq \\gamma , \\\\ \\hat {y} _ {*} & \\text {i f} \\hat {y} _ {a} \\neq \\hat {y} _ {*} \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {a}\\right) \\leq \\gamma \\text {a n d} \\operatorname {C o n f} \\left(\\hat {y} _ {*}\\right) > \\gamma , \\\\ - 1 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}", + "image_path": "ac9b5b98680334507c2d95135d1ec1ac36c3a40dac6593c105dd05be8a82c934.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^T" + }, + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "type": "text", + "content": " is an argument of the above, but is omitted for clarity. The MatchOrConf scheme selects only those target samples where there is agreement between the predictions of the two models, or where one model's confidence exceeds threshold " + }, + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 171, + 287, + 243 + ], + "type": "text", + "content": " while the other's does not, resulting in the following selection mask:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 252, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 252, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 102, + 252, + 287, + 283 + ], + "type": "interline_equation", + "content": "s \\left(\\mathbf {x} ^ {T}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {y} \\left(\\mathbf {x} ^ {T}\\right) \\neq - 1, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "3989b85a4a3491d2c109f947354abe36788a0d9d91ea41e1b51d3a9fa886f734.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "text", + "content": "Taking inspiration from [22] and [43], we compute the target domain loss on masked videos, where we employ the same teacher-guided attention masking strategy (with " + }, + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "inline_equation", + "content": "r = 0.8" + }, + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "text", + "content": ") used during UMT pre-training in Stage 1. We find that training on masked videos in this stage results in substantial performance boosts compared to training on un-masked videos (see Table 6). Like [22], we include a loss weighting " + }, + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 287, + 289, + 395 + ], + "type": "text", + "content": " for each target video, based on the student network confidence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 406, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 406, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 117, + 406, + 287, + 420 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} ^ {T}\\right) = \\operatorname {C o n f} \\left(\\hat {y} _ {a} \\left(\\mathbf {x} ^ {T}\\right)\\right) \\tag {5}", + "image_path": "a9a30627eaf84822102800abd13fe0b8f256f3a18c82bd6fc5d744744e6cc03c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "type": "text", + "content": "The complete loss function for the collaborative self-training stage can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 465, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 465, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 465, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S T}} = \\mathbb {E} _ {(\\mathbf {x} ^ {S}, y ^ {S}) \\sim \\mathcal {P} _ {S}} \\left[ \\mathcal {L} _ {C E} \\left(f _ {a} (\\mathbf {x} ^ {S}), y ^ {S}\\right) \\right] + \\\\ \\lambda \\mathbb {E} _ {(\\mathbf {x} ^ {T}) \\sim \\mathcal {P} _ {T}} \\left[ s (\\mathbf {x} ^ {T}) q (\\mathbf {x} ^ {T}) \\mathcal {L} _ {C E} \\left(f _ {a} (m (\\mathbf {x} ^ {T})), \\tilde {y} (\\mathbf {x} ^ {T}) \\right] \\right. \\tag {6} \\\\ \\end{array}", + "image_path": "dbb9dc0b815befdb391c338da14136d8e8b7e96c7879e59079e9dd589ee49ca0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 538, + 110, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 110, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 110, + 550 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 555, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 581 + ], + "type": "text", + "content": "We evaluate our approach on three video domain adaptation benchmarks: Daily-DA, Sports-DA, UCF " + }, + { + "bbox": [ + 46, + 555, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 46, + 555, + 287, + 581 + ], + "type": "text", + "content": " HMDBfull." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 582, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 677 + ], + "type": "text", + "content": "- Daily-DA [64] is a recently introduced benchmark composed of videos from four domains: ARID (A) [60], HMDB51 (H) [28], Moments-in-Time (M) [38] and Kinetics (K) [7, 26]. It consists of 8 overlapping classes representing everyday actions, with a total volume of approximately 19K videos. Videos from the ARID domain present a unique challenge in this benchmark, as they were explicitly filmed under low-illumination conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": "- Sports-DA [64] contains videos from three domains: Sports-1M (S) [25], UCF101 (U) [49] and Kinetics (K) [7, 26]. It includes 23 overlapping categories depicting" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 95 + ], + "type": "text", + "content": "various sporting activities, with a total volume of approximately 41K videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 96, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 133 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 96, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{UCF} \\leftrightarrow \\mathbf{HMDB}_{full}" + }, + { + "bbox": [ + 306, + 96, + 547, + 133 + ], + "type": "text", + "content": " [8] contains approximately 3.2K videos across 12 action categories taken from the UCF101 (U) [49] and HMDB51 (K) [28] datasets." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 161, + 440, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 440, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 440, + 174 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": "Frame Sampling. During training, we perform random uniform frame sampling [53] on each video. We divide the video into " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": " uniform segments and randomly sample one frame from each. All frames are resized to " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": " pixels, resulting in a " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 224 \\times 224" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": " tensor for each video. During testing, we average the network predictions across 3 spatial crops and 4 temporal clips, where clip " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": " is formed by splitting each of the " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": " segments into 4 sub-segments and sampling a frame from the " + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 184, + 547, + 303 + ], + "type": "text", + "content": "-th sub-segment of each." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 308, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 417 + ], + "type": "text", + "content": "Network Architecture & Training. All of our experiments use a standard ViT-B/16 [12] architecture with a fixed sine and cosine positional encoding scheme. Our patch embedding projection operates on each frame independently, and self-attention is performed across both the space and time dimensions. All classification is performed using mean pooling of the final layer patch representations. We conduct our experiments using the PyTorch deep learning library [40] on 4 NVIDIA A5000 GPUs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "content": "During Stage 1, we perform UMT training for 50 epochs on the target dataset using a batch size of 256 and a learning rate of " + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{e - }5" + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "content": ". During Stage 2, we use a batch size of 28, a learning rate of " + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{e - }5" + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "content": ", and apply a layer-wise learning rate decay of 0.65. During stage 3, we use a batch size of 40 (20 videos from each domain), a learning rate of 1e-5, a MatchOrConf confidence threshold of " + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "content": " and a target domain loss weight " + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 304, + 422, + 547, + 553 + ], + "type": "text", + "content": ". Training in all stages uses the AdamW optimizer [35] and a weight decay of 0.05. Please refer to Sec. 8 in the Appendix for more training details for each stage." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "text", + "content": "Zero-Shot Classification with CLIP. In Stage 3 of UNITE, we use CLIP (ViT-B/16) in a zero-shot manner to refine the pseudolabels of target domain videos. We do so by passing each of the " + }, + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "text", + "content": " video frames through the CLIP image encoder and computing the cosine similarity with a set of text representations produced by the CLIP text encoder. The inputs to the text encoder are the class names associated with the particular action recognition task, embedded in the following template: \"A video of a person {class}.\" To form a video-level prediction, we simply average the softmax of the cosine similarity scores across the video frames. Refer to the Appendix (Sec. 11) for a listing of class names used for each task." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18923" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 72, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 547, + 247 + ], + "type": "table", + "html": "
MethodTarget Domain Accuracy (Top-1%)
H→AM→AK→AA→HM→HK→HH→MA→MK→MM→KH→KA→KAvg.
DALL-V [67] LB17.534.715.614.644.647.925.515.535.761.645.117.831.3
Source Only40.452.136.549.668.357.941.536.343.379.348.041.749.6
ZSCLIP (ViT-B/16) [44]36.536.536.560.060.060.048.548.548.568.168.168.153.3
SFVUDAATCoN [62]17.927.217.226.747.348.230.717.232.557.748.531.033.5
EXTERNAL [63]26.218.123.926.253.755.840.718.235.268.157.651.439.6
DALL-V [67]24.024.024.057.965.452.547.045.747.078.176.775.051.4
UDADANN [18]14.222.821.220.143.337.529.519.721.758.838.227.029.5
MK-MMD [34]20.321.021.718.750.436.225.718.024.058.533.826.129.5
TA3N [8]14.421.619.914.943.037.725.715.631.555.538.423.428.5
UNITE w/o CST43.846.734.751.770.854.644.339.043.378.151.750.850.8
UNITE (Ours)48.044.137.567.974.265.851.850.048.089.969.963.659.2
DALL-V [67] UB26.926.926.970.470.470.461.561.561.588.988.988.961.9
Target Only68.568.568.584.684.684.673.073.073.098.398.398.381.1
", + "image_path": "2218beb92ffdae5226b40da60ccf651199043112fcd56b3462b2fb92e757c5ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 149, + 275, + 449, + 449 + ], + "blocks": [ + { + "bbox": [ + 55, + 253, + 537, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 537, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 537, + 265 + ], + "type": "text", + "content": "Table 1. UDA results on Daily-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 275, + 449, + 449 + ], + "lines": [ + { + "bbox": [ + 149, + 275, + 449, + 449 + ], + "spans": [ + { + "bbox": [ + 149, + 275, + 449, + 449 + ], + "type": "table", + "html": "
MethodTarget Domain Accuracy (Top-1 %)
U→SK→SS→UK→UU→KS→KAvg.
DALL-V [67] LB64.379.584.485.467.278.276.5
Source Only67.686.897.998.979.989.186.7
ZSCLIP (ViT-B/16) [44]85.085.093.393.391.391.389.9
SFVUDAATCoN [62]47.969.790.693.665.276.073.8
EXTERN [63]72.773.895.493.781.282.283.2
DALL-V [67]75.977.788.888.081.282.382.3
UDADANN [18]55.175.085.788.065.973.473.8
MK-MMD [34]55.667.990.990.266.173.674.0
TA3N [8]54.168.693.090.363.672.673.7
UNITE w/o CST73.987.398.299.284.690.288.9
UNITE (Ours)86.090.298.799.894.295.394.0
DALL-V [67] UB88.388.393.493.485.685.689.1
Target Only97.997.998.898.899.999.998.9
", + "image_path": "00542bc2118f001c75a7746517a7c4e4c30ca134334a4385375f8093086439ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 457, + 539, + 469 + ], + "lines": [ + { + "bbox": [ + 53, + 457, + 539, + 469 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 539, + 469 + ], + "type": "text", + "content": "Table 2. UDA results on Sports-DA. Rows in color use our UMT pre-trained backbone, and rows without color are reported from [61]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 490, + 113, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 113, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 113, + 501 + ], + "type": "text", + "content": "5.3. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "text", + "content": "We compare results using UNITE with previously developed techniques in UDA, as well as in the closely related setting of SFUDA as reported by [67]. For Daily-DA and Sports-DA we compare against DANN [18], MK-MMD [34] and " + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^3\\mathrm{N}" + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "text", + "content": " [8] in the UDA setting, and ATCoN [62], EXTERN [63] and DALL-V [67] for SFUDA. For " + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "inline_equation", + "content": "UCF\\leftrightarrow HMDB_{full}" + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "text", + "content": ", we additionally report accuracies from " + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}^2\\mathrm{A}" + }, + { + "bbox": [ + 46, + 508, + 288, + 604 + ], + "type": "text", + "content": " [52] and UDAVT [10]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "type": "text", + "content": "5.4. Benchmark Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Tables 1, 2 and 3 show the results of our evaluation of UNITE on Daily-DA, Sports-DA and " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "UCF \\leftrightarrow HMDB_{full}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". We include accuracies after Stage 2 of the pipeline (denoted as UNITE w/o CST) as well as after applying the full UNITE process. For reference, we include accuracies using our UMT pre-trained network with source domain supervised fine-tuning (Source Only) and with target domain" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 490, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 562 + ], + "type": "text", + "content": "supervised fine-tuning (Target Only). We also include the source only and target only accuracies (denoted in the tables as LB and UB) of DALL-V [67] (a CLIP-based ResNet-50 model) for added context. We additionally display the zero-shot classification accuracy of the CLIP teacher model that we utilize during Stage 1 and Stage 3 of UNITE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "text", + "content": "An observation we make across all benchmarks is the enhanced baseline accuracy of the UMT pre-trained model, which speaks to the strength of the unsupervised masked distillation process and the ViT architecture. On the Daily-DA benchmark, we find that UNITE exceeds previously reported results on most domain shifts. On only one shift " + }, + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "inline_equation", + "content": "(\\mathrm{M}\\rightarrow \\mathrm{A})" + }, + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "text", + "content": " we observe degradation in target accuracy after employing UNITE compared to source only, which is consistent with previous work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "UNITE exhibits even stronger performance on the larger-scale Sports-DA benchmark, where it exceeds previous results by a large margin. We also see more consistent and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18924" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 254, + 251 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 254, + 251 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 254, + 251 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 254, + 251 + ], + "type": "table", + "html": "
MethodAccuracy (%)
H→UU→HAvg.
SDALL-V [67] LB71.676.173.8
Source Only88.880.084.4
CLIP (ViT-B/16) [44]88.891.790.3
SEVUDAATCoN [62]85.379.782.5
EXTERN [63]91.988.990.4
DALL-V [67]93.188.991.0
UDADANN [18]74.475.174.8
MK-MMD [34]74.779.777.2
TA3N [8]78.184.881.5
CO2A [52]95.887.891.8
UDAVT [10]96.892.394.6
UNITE w/o CST92.183.687.9
UNITE (Ours)92.595.093.8
DALL-V [67] UB93.791.492.6
Target Only99.796.798.2
", + "image_path": "ed99a58f6d2ac638770e2aa248076689cb3bb78b7cccc3432afc480c9cbe51ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 282, + 273, + 443 + ], + "blocks": [ + { + "bbox": [ + 61, + 282, + 273, + 443 + ], + "lines": [ + { + "bbox": [ + 61, + 282, + 273, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 273, + 443 + ], + "type": "image", + "image_path": "3f5d2d4393376e31c3a9748c11b00f54c5a083a9fe8cd79fbfd4c40496b10d06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 451, + 287, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 506 + ], + "type": "text", + "content": "Figure 4. Class-wise accuracy on ARID " + }, + { + "bbox": [ + 46, + 451, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 451, + 287, + 506 + ], + "type": "text", + "content": " HMDB. Performances are shown for the CLIP teacher model and the student model before and after the collaborative self-training stage (Stage 3). For most classes, the Stage 3 student exceeds the accuracy of both the Stage 2 student and the CLIP teacher." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 526, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 563 + ], + "type": "text", + "content": "substantial improvement after the UMT target pre-training stage alone (UNITE w/o CST), suggesting that overtraining may be occurring on the smaller datasets in Daily-DA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "text", + "content": "On " + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "inline_equation", + "content": "UCF \\leftrightarrow HMDB_{full}" + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "text", + "content": " we see state-of-the-art performance on " + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{U} \\rightarrow \\mathrm{H}" + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "text", + "content": ", while UDAVT outperforms on " + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{H} \\rightarrow \\mathrm{U}" + }, + { + "bbox": [ + 46, + 563, + 287, + 610 + ], + "type": "text", + "content": ". Nevertheless, UNITE still achieves strong overall performance on the benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 621, + 233, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 233, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 233, + 635 + ], + "type": "text", + "content": "6. Additional Analysis & Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Student exceeds the teacher. We see in the benchmark results that the video student model exceeds the zero-shot accuracy of the CLIP teacher in virtually all domain shifts. This suggests that UNITE successfully leverages the spatial modeling capabilities of CLIP to train an even more powerful spatiotemporal model for the action recognition task." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 541, + 148 + ], + "blocks": [ + { + "bbox": [ + 47, + 252, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 252, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 288, + 275 + ], + "type": "text", + "content": "Table 3. UDA results on " + }, + { + "bbox": [ + 47, + 252, + 288, + 275 + ], + "type": "inline_equation", + "content": "UCF \\leftrightarrow HMDB_{full}" + }, + { + "bbox": [ + 47, + 252, + 288, + 275 + ], + "type": "text", + "content": ". Rows in color use our UMT backbone, and rows without color are reported from [61]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 541, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 541, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 541, + 148 + ], + "type": "table", + "html": "
MethodPTCSTAccuracy (%)
H→AA→H
Source OnlyXX40.449.6
+ Target UMT Pre-Training✓X43.851.7
+ Collaborative Self-TrainingX✓42.060.8
UNITE✓✓48.067.9
", + "image_path": "87e0c8b77b64aaa10e585cd44f8da41bac04956b0248322fb8c148132c844ec1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 150, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 305, + 150, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 545, + 183 + ], + "type": "text", + "content": "Table 4. Ablation of stages in UNITE on ARID ↔ HMDB from Daily-DA. PT and CST denote UMT pre-training and collaborative self-training, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 198, + 546, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 546, + 402 + ], + "type": "text", + "content": "We can see from Figure 4 that the teacher model and the student model prior to collaborative self-training have differing capabilities. For example, the image teacher is stronger at recognizing actions from the jump class than the Stage 2 student. On the other hand, the video student is significantly better than the teacher at recognizing wave at the start of CST. In both cases, we see that the student model, through collaborative self-training, comes to exceeds both the Stage 2 student and the CLIP image teacher. With the MatchOr-Conf pseudolabeling scheme, the two models are able to work together to achieve higher accuracy in the target domain. We see accuracy on the pick and run classes declines after CST, as self-training tends to reinforce the accuracy of well-performing classes, sometimes at the expense of more difficult ones that are infrequently assigned pseudolabels. Modifying CST to account for the learning status of each class, like in [68], may help remedy this issue." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "content": "Impact of Stage 1 and 3. In Table 4 we ablate the stages in UNITE. While UMT pre-training and collaborative self-training each offer a benefit in the absence of the other, the largest gains in domain transfer performance are obtained when combining the two stages. For example, on the ARID " + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "content": " HMDB domain shift in Daily-DA, Stage 1 and Stage 3 alone improve upon source only accuracy by " + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "inline_equation", + "content": "+1.8\\%" + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "inline_equation", + "content": "+11.2\\%" + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "content": " respectively. Meanwhile, the full UNITE pipeline results in " + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "inline_equation", + "content": "+18.3\\%" + }, + { + "bbox": [ + 304, + 403, + 546, + 533 + ], + "type": "text", + "content": " accuracy improvement on the target domain, suggesting that the masked pre-training and masked self-training processes complement one another." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 652 + ], + "type": "text", + "content": "Impact of data domains during pre-training. In Table 5, we assess the impact of utilizing each of the two data domains during UMT pre-training on UDA performance. While the results vary across domain shifts, we observe the most consistent improvements when pre-training only on target domain videos. Although incorporating both domains during unsupervised pre-training has been successfully applied in image-based domain adaptation [27, 37, 43, 46], we find it to be suboptimal with UMT training on the VUDA tasks studied in this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "Effect of masking during self-training. In Table 6, we see that enforcing the target domain CE loss on masked videos results in increased performance compared to using unmasked target videos, with smaller benefits for " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{H}\\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": " and much larger benefits for " + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\rightarrow \\mathrm{H}" + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": ". We believe the benefits of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18925" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 247, + 139 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 247, + 139 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 247, + 139 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 247, + 139 + ], + "type": "table", + "html": "
UMT Pre-Training DataAccuracy (%)
H→AA→H
None42.060.8
Source Only48.360.0
Target Only48.067.9
Source + Target45.267.1
", + "image_path": "f79697ed0e83c55de463a669435ff42f6f608a1e2cd304e36a18639fd13dbd33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 89, + 180, + 249, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "type": "text", + "content": "Table 5. Comparison of data domains used during UMT pretraining. Accuracies are reported after the self-training stage. Bold indicates best accuracy and underline indicates second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 180, + 249, + 231 + ], + "lines": [ + { + "bbox": [ + 89, + 180, + 249, + 231 + ], + "spans": [ + { + "bbox": [ + 89, + 180, + 249, + 231 + ], + "type": "table", + "html": "
Data for Target CE LossAccuracy (%)
H→AA→H
Unmasked46.654.6
Masked48.067.9
", + "image_path": "abe7b1463b7a026169250ceca59ef8480260846adb7207ca9d90f50dd6a30c9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 261, + 270, + 350 + ], + "blocks": [ + { + "bbox": [ + 47, + 232, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 287, + 255 + ], + "type": "text", + "content": "Table 6. Effect of applying the target domain cross entropy loss on masked vs. unmasked target videos during the self-training stage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 261, + 270, + 350 + ], + "lines": [ + { + "bbox": [ + 67, + 261, + 270, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 270, + 350 + ], + "type": "table", + "html": "
Pseudolabeling StrategyAccuracy (%)
H→AA→H
Cons (k=2)39.857.5
ConsOrConf (k=2,T=0.5) [43]48.656.3
ConsAndConf (k=2,T=0.5)40.157.1
Zero-Shot CLIP Only (T=0.5)34.751.3
MatchOrConf (γ=0.1) [69]48.067.9
MatchAndConsOrConf (k=2,γ=0.1)47.460.8
", + "image_path": "f142c2fb6f07dbfc6dcfc7ba37ef6a8e28920d6ebb23f6742e4684dc17323565.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "content": "masking arise from forcing the model to use different cues from the same video to predict the pseudolabel, resulting in more robust target domain recognition. As we saw in Table 4, the masked pre-training stage plays a key role in unlocking the benefits of masked self-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": "Comparison of pseudolabeling strategies. In Table 7, we explore the use of alternative pseudolabeling strategies to the MatchOrConf [69] scheme employed in UNITE. The first 3 rows of the table assess variations of the masked consistency pseudolabeling scheme proposed in PACMAC [43], where target samples are selected if there is agreement between predictions of " + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": " masked views. This criterion can also be combined with a constraint on the confidence exceeding threshold " + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": ". While we find the combination of these two constraints (ConsOrConf) to be effective for self-training, we find that MatchOrConf results in more consistent target accuracy improvements across domain shifts. We also consider combining MatchOrConf with a masked consistency constraint as a way to reduce cases where the student and teacher models agree on an incorrect prediction. This scheme, denoted as MatchAndConsOrConf, adds an additional constraint to the matching criterion by also requiring agreement between the student model predictions of " + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": " masked views. We find that this approach underperforms compared to the vanilla MatchOrConf scheme due to low utilization of target domain samples. Our results also indicate that using both models together for pseudolableing outperforms using only zero-shot CLIP predictions." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 342, + 70, + 512, + 118 + ], + "blocks": [ + { + "bbox": [ + 75, + 350, + 258, + 361 + ], + "lines": [ + { + "bbox": [ + 75, + 350, + 258, + 361 + ], + "spans": [ + { + "bbox": [ + 75, + 350, + 258, + 361 + ], + "type": "text", + "content": "Table 7. Comparison of pseudolabeling strategies." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 70, + 512, + 118 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 512, + 118 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 512, + 118 + ], + "type": "table", + "html": "
Loss During Self-TrainingAccuracy (%)
H→AA→H
LCE(m(xT), y)33.742.1
LCE(m(xT), y) + LCE(xS, yS)48.067.9
", + "image_path": "c6356fb991535b06aa4414bb6b72af70e349fa90e3834627557bb1319413b68f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 356, + 148, + 499, + 224 + ], + "blocks": [ + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "type": "text", + "content": "Table 8. Effect of including a source domain classification loss in addition to the target domain loss during the self-training stage." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 148, + 499, + 224 + ], + "lines": [ + { + "bbox": [ + 356, + 148, + 499, + 224 + ], + "spans": [ + { + "bbox": [ + 356, + 148, + 499, + 224 + ], + "type": "table", + "html": "
Confidence ThresholdAccuracy (%)
H→AA→H
0.148.067.9
0.344.155.4
0.551.964.6
0.742.665.8
0.940.762.5
", + "image_path": "ee0a8ff0c68464e69536666660c4215a0b7f2e1c39b640e9004adeb18fa446e3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 226, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 226, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 545, + 248 + ], + "type": "text", + "content": "Table 9. Effect of confidence threshold " + }, + { + "bbox": [ + 305, + 226, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 226, + 545, + 248 + ], + "type": "text", + "content": " in MatchOrConf pseudolabeling scheme from Equation 3." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "content": "Effect of source classification during self-training. In Table 8, we study the importance of including a source domain classification loss while performing collaborative self-training on masked target videos. Indeed, we find that source classification is a crucial ingredient in this stage, as excluding it destabilizes training and results in performance degradation rather than improvement. We suspect that the ground truth labels on source data help to combat the effects of inaccurate target pseudolabels, which may be prevalent especially at the start of training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 380, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 476 + ], + "type": "text", + "content": "Choice of pseudolabeling confidence threshold. In Table 9, we compare adaptation performance for various values of confidence threshold in the MatchOrConf pseudolabeling scheme. We observe significant variation across domain shifts, but find that " + }, + { + "bbox": [ + 304, + 380, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 304, + 380, + 545, + 476 + ], + "type": "text", + "content": " leads to relatively consistent improvement. We leave for future work the development of a more principled approach to determining an appropriate confidence threshold for a given domain shift." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 487, + 383, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 383, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 383, + 499 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "content": "In this work, we presented UNITE, a three step process for unsupervised video domain adaptation. Our approach leverages a powerful spatial encoder to aid in the adaptation of a spatiotemporal network using unlabeled target domain videos—making significant progress towards bridging the domain gaps in various VUDA datasets. The advancements demonstrated by UNITE underscore the untapped potential of masked modeling techniques for video domain adaptation, which we hope will spur further research in this area." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 624, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This research was sponsored by the Army Research Laboratory under Cooperative Agreement W911NF-21-2-0211. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18926" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video Vision Transformer. In ICCV, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked Siamese Networks for Label-Efficient Learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. In CVPR, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "text", + "content": "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In ICML, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "text", + "content": "[5] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In ICML, pages 813-824, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In CVPR, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "text", + "content": "[7] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A Short Note about Kinetics-600, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 402 + ], + "type": "text", + "content": "[8] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal Attentive Alignment for Large-Scale Video Domain Adaptation. In ICCV, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "text", + "content": "[9] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 iee. In CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 459, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 288, + 513 + ], + "type": "text", + "content": "[10] Victor G. Turrisi Da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Unsupervised Domain Adaptation for Video Transformers in Action Recognition. In ICPR, pages 1258-1265, Montreal, QC, Canada, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 288, + 590 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 590, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 288, + 656 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 657, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 288, + 690 + ], + "type": "text", + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale Vision Transformers. In ICCV, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "type": "text", + "content": "[14] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In CVPR, 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In ICCV, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[17] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "text", + "content": "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 293, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 316 + ], + "type": "text", + "content": "[21] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[22] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation. In CVPR, pages 11721–11732, Vancouver, BC, Canada, 2023. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 361, + 545, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 394 + ], + "type": "text", + "content": "[23] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, pages 646-661. Springer, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "text", + "content": "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to Hide from Your Students: Attention-Guided Masked Image Modeling. In ECCV, pages 300-318, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "text", + "content": "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-Scale Video Classification with Convolutional Neural Networks. In CVPR, pages 1725–1732, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 547 + ], + "type": "text", + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, 2017. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 547, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 547, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 547, + 545, + 592 + ], + "type": "text", + "content": "[27] Donghyun Kim, Kuniaki Saito, Tae-Hyun Oh, Bryan A Plummer, Stan Sclaroff, and Kate Saenko. Cds: Cross-domain self-supervised pre-training. In ICCV, pages 9123-9132, 2021. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 624 + ], + "type": "text", + "content": "[28] H Kuehne, H Jhuang, E Garrote, T Poggio, and T Serre. HMDB: A Large Video Database for Human Motion Recognition. In ICCV, 2011. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In ICCV, pages 1632-1643, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[30] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked Teacher: Towards Training-Efficient Video Foundation Models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18927" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[31] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. MST: Masked Self-Supervised Transformer for Visual Representation. In NeurIPS, pages 13165-13176, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "text", + "content": "[32] Han Lin, Guangxing Han, Jiawei Ma, Shiyuan Huang, Xudong Lin, and Shih-Fu Chang. Supervised masked knowledge distillation for few-shot transformers. In CVPR, pages 19649-19659, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "text", + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3192-3201, New Orleans, LA, USA, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "text", + "content": "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 260 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[36] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-CLIP: End-to-End Multi-grained Contrastive Learning for Video-Text Retrieval. In ACM MM, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "type": "text", + "content": "[37] Samarth Mishra, Kate Saenko, and Venkatesh Saligrama. Surprisingly Simple Semi-Supervised Domain Adaptation with Pretraining and Consistency, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[38] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, and Dan Gutfreund. Moments in Time Dataset: One Million Videos for Event Understanding. PAMI, 42(2):502-508, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 415 + ], + "type": "text", + "content": "[39] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video Transformer Network. In ICCV, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[41] Vishal M Patel, Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. Visual Domain Adaptation: A survey of recent advances. IEEE Signal Processing Magazine, 32(3):53-69, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "text", + "content": "[42] Aj Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning. In CVPR, pages 2214-2224, Vancouver, BC, Canada, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "text", + "content": "[43] Viraj Prabhu, Sriram Yenamandra, Aaditya Singh, and Judy Hoffman. Adapting self-supervised vision transformers by probing attention-conditioned masking consistency. NeurIPS, 35:23271-23283, 2022. 2, 3, 4, 5, 7, 8, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[45] Aadarsh Sahoo, Rutav Shah, R. Panda, Kate Saenko, and Abir Das. Contrast and Mix: Temporal Contrastive Video Domain Adaptation with Background Mixing. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "text", + "content": "[46] Kendrick Shen, Robbie M. Jones, Ananya Kumar, Sang Michael Xie, Jeff Z. Haochen, Tengyu Ma, and Percy Liang. Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation. In ICML, pages 19847-19878, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 163 + ], + "type": "text", + "content": "[47] Karen Simonyan and Andrew Zisserman. Two-Stream Convolutional Networks for Action Recognition in Videos. In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 220 + ], + "type": "text", + "content": "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In NeurIPS, pages 596-608, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "text", + "content": "[49] K. Soomro, A. Zamir, and M. Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. ArXiv, 2012. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "text", + "content": "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[51] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning Spatiotemporal Features with 3D Convolutional Networks. In ICCV, pages 4489-4497, Santiago, Chile, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "text", + "content": "[52] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-Head Contrastive Domain Adaptation for Video Action Recognition. In WACV, pages 2234–2243, 2022. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "type": "text", + "content": "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal Segment Networks for Action Recognition in Videos. PAMI, 2017. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 452, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 506 + ], + "type": "text", + "content": "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. In CVPR, pages 14549-14560, Vancouver, BC, Canada, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 508, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 541 + ], + "type": "text", + "content": "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "text", + "content": "[56] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked Video Distillation: Rethinking Masked Feature Modeling for Self-supervised Video Representation Learning. In CVPR, pages 6312-6322, Vancouver, BC, Canada, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 611, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 678 + ], + "type": "text", + "content": "[57] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. InternVideo: General Video Foundation Models via Generative and Discriminative Learning, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "text", + "content": "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A Simple Framework for Masked Image Modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18928" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[59] Yuecong Xu, Haozhi Cao, Zhenghua Chen, Xiaoli Li, Lihua Xie, and Jianfei Yang. Video Unsupervised Domain Adaptation with Deep Learning: A Comprehensive Survey, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[60] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. ARID: A New Dataset for Recognizing Action in the Dark. In Communications in Computer and Information Science, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "text", + "content": "[61] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Aligning Correlation Information for Domain Adaptation in Action Recognition. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 288, + 262 + ], + "type": "text", + "content": "[62] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-Free Video Domain Adaptation by Learning Temporal Consistency for Action Recognition. In ECCV, pages 147-164, Cham, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 288, + 308 + ], + "type": "text", + "content": "[63] Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Extern: Leveraging endo-temporal regularization for black-box video domain adaptation. arXiv preprint arXiv:2208.05187, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 288, + 363 + ], + "type": "text", + "content": "[64] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Zhengguo Li, and Zhenghua Chen. Multi-Source Video Domain Adaptation With Temporal Attentive Moment Alignment Network. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3860-3871, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 288, + 397 + ], + "type": "text", + "content": "[65] Yuecong Xu, Jianfei Yang, Yunjiao Zhou, Zhenghua Chen, Min Wu, and Xiaoli Li. Augmenting and Aligning Snippets for Few-Shot Video Domain Adaptation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 288, + 441 + ], + "type": "text", + "content": "[66] Hongwei Xue, Peng Gao, Hongyang Li, Yu Qiao, Hao Sun, Houqiang Li, and Jiebo Luo. Stare at What You See: Masked Image Modeling without Reconstruction. In CVPR, pages 22732-22741, Vancouver, BC, Canada, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 443, + 288, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 443, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 443, + 288, + 496 + ], + "type": "text", + "content": "[67] Giacomo Zara, Alessandro Conti, Subhankar Roy, Stephane Lathuiliere, Paolo Rota, and Elisa Ricci. The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 498, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 288, + 552 + ], + "type": "text", + "content": "[68] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 34:18408-18419, 2021. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 555, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 555, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 555, + 288, + 586 + ], + "type": "text", + "content": "[69] Wenyu Zhang, Li Shen, and Chuan-Sheng Foo. Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation. In CVPR, 2023. 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 588, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 588, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 588, + 288, + 620 + ], + "type": "text", + "content": "[70] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT Pre-training with Online Tokenizer. In ICLR, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18929" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_content_list.json b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97bab136aed635a5bf079543109f28cf47edc7e2 --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxuan Wang $^{1,2,3*}$ Tongtian Yue $^{1,2*}$ Yisi Zhang $^{4}$ Longteng Guo $^{1}$ Xingjian He $^{1}$ Xinlong Wang $^{3}$ Jing Liu $^{1,2\\dagger}$", + "bbox": [ + 138, + 200, + 830, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences (CASIA)", + "bbox": [ + 223, + 239, + 746, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 138, + 256, + 828, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Beijing Academy of Artificial Intelligence (BAAI)", + "bbox": [ + 277, + 273, + 691, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 University of Science and Technology Beijing (USTB)", + "bbox": [ + 261, + 292, + 707, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangwenxuan2023@ia.ac.cn, wangxinlong@baai.ac.cn, jliu@nlpr.ia.ac.cn}", + "bbox": [ + 171, + 311, + 787, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4215409bdd47d926dcce2e13dea41ea07440dc14afad1ff4d2d55d72c1b1509b.jpg", + "image_caption": [ + "Images" + ], + "image_footnote": [], + "bbox": [ + 158, + 371, + 246, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbc3328c82745a4b6a22d70666a91c4718bf26808b4d2704e2262755b5fa6a3a.jpg", + "image_caption": [ + "(a) \"Standing Lady on left\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 371, + 346, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb8c721ef56e1204bb01f690cca9d8ffb87371a4f4cd036ffc32d82ee7de2b8.jpg", + "image_caption": [ + "(a) \"Motorcycle on the left\"", + "Figure 1. Classic Referring Expression Segmentation (RES) only supports expressions that indicate a single target object, e.g., (a). Compared with classic RES, the proposed Multi-Granularity Referring Expression Segmentation (MRES) task supports expressions indicating the specific part-level regions of target objects, e.g., part-level expressions like (b)-(e) from our newly built RefCOCom benchmark." + ], + "image_footnote": [], + "bbox": [ + 259, + 422, + 346, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c71f2b66003c1c2aa26d45a5e887e33def7e3acc9f193c7d1e27746bae22999b.jpg", + "image_caption": [ + "Classic RES", + "(a) \"Standing Lady on left" + ], + "image_footnote": [], + "bbox": [ + 357, + 371, + 444, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72989edfd0c2cb4a7208cb18677905e0ff4027889f113e892fe23ef8428fe33a.jpg", + "image_caption": [ + "(a) \"Motorcycle on the left\"" + ], + "image_footnote": [], + "bbox": [ + 357, + 422, + 444, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8dd8a3911eaa263c703b098f78689f6a0c1eb7dfd56c5e1b24f5ab8e36ef82a.jpg", + "image_caption": [ + "MRES (Ours)", + "(b) \"Leg of lady walking on left" + ], + "image_footnote": [], + "bbox": [ + 446, + 371, + 532, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b536f2226f064f562556624889d0051583eff2cd58a1d850bf4788b64be66564.jpg", + "image_caption": [ + "(b) \"Silvery body of motorcycle on the left\"" + ], + "image_footnote": [], + "bbox": [ + 446, + 422, + 531, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f38fde52231ae2a6bc845c2b186b392753629f423c85c1a8c9a8c1bb1b9bef07.jpg", + "image_caption": [ + "(c) \"Middle girl arm\"" + ], + "image_footnote": [], + "bbox": [ + 534, + 371, + 620, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7c242ff774c85e64aae5243310e2657560ed6af24874b082c12481aa63f89aa.jpg", + "image_caption": [ + "(c) \"Center motorcycle", + "wheel\"" + ], + "image_footnote": [], + "bbox": [ + 534, + 422, + 620, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0e2bfe02d8fc8715b66e988b6464860be8adccc37fcf02f137a4c2d87cafda2.jpg", + "image_caption": [ + "(d) \"Head of center player\"" + ], + "image_footnote": [], + "bbox": [ + 622, + 371, + 707, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d545c6fdedf3b74e6c7cb0dcac3cc2192f62b05d43d55e91fcb7e86b48af7456.jpg", + "image_caption": [ + "(d) \"Front motor headlight\"" + ], + "image_footnote": [], + "bbox": [ + 622, + 422, + 707, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ac4e97f1a4cfe3f3f280283bbe4cb9a29004bb5553e71403029d026807f8179.jpg", + "image_caption": [ + "(e) \"Center player torso in white\"" + ], + "image_footnote": [], + "bbox": [ + 710, + 371, + 795, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0421597dd1692e22de9fe0803a7b084372ea284c70d72b9682643f48b9b52958.jpg", + "image_caption": [ + "(e) \"Transparent window of front motor\"" + ], + "image_footnote": [], + "bbox": [ + 710, + 422, + 795, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 560, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring expression segmentation (RES) aims at segmenting the foreground masks of the entities that match the descriptive natural language expression. Previous datasets and methods for classic RES task heavily rely on the prior assumption that one expression must refer to object-level targets. In this paper, we take a step further to finer-grained part-level RES task. To promote the object-level RES task towards finer-grained vision-language understanding, we put forward a new multi-granularity referring expression segmentation (MRES) task and construct an evaluation benchmark called RefCOCom by manual annotations. By employing our automatic model-assisted data engine, we build the largest visual grounding dataset namely MRES-32M, which comprises over 32.2M high-quality masks and captions on the provided 1M images. Besides, a simple yet strong model named UniRES is designed to accomplish the unified object-level and part-level grounding task. Extensive experiments on our RefCOCom for MRES and three datasets (i.e., RefCOCO(+/g)) for classic RES task demon", + "bbox": [ + 75, + 579, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strate the superiority of our method over previous state-of-the-art methods. To foster future research into fine-grained visual grounding, our benchmark RefCOCOM, the MRES-32M dataset and model UniRES will be publicly available at https://github.com/Rubics-Xuan/MRES.", + "bbox": [ + 500, + 561, + 890, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 664, + 630, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the most challenging tasks in vision-language understanding, referring expression segmentation (RES) aims to locate specific regions at the pixel level based on a descriptive language expression. Compared to traditional visual segmentation tasks that focus on images or videos alone, RES poses greater difficulties and challenges due to the necessity of strong comprehension across modalities, but it can simultaneously alleviate the problem of predefined categories in conventional object detection or segmentation. With the real-world scene often requiring diversity in target identification, RES task holds vast potential for applications, e.g., language-based human-object interaction and interactive image editing.", + "bbox": [ + 496, + 689, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the concept of RES task was initially proposed in", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 243, + 13, + 802, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12998", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/8871fc25628f00c85238b4ba6bef3bc5043fe642efabf6cb2214fba2003c7719.jpg", + "table_caption": [ + "Table 1. Comparison among different referring expression datasets, including ReferIt [18], RefCOCO(+/g) [27, 44], PhraseCut [41], and our proposed MRES-32M. Part-Level: expressions that specifies various parts of the target object in the given image." + ], + "table_footnote": [], + "table_body": "
ReferItRefCOCO(+/g)PhraseCutMRES-32M
Image SourceCLEF [11]COCO [24]VG [21]Object365 [34]
Object-Level✓✓✓✓
Part-Level×××✓
Expression Typefreefreetemplatedfree
", + "bbox": [ + 78, + 146, + 467, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[13], various multimodal frameworks such as [13, 19, 40, 42] towards precise RES have been proposed to deal with the challenging feature extraction and alignment problems between visual and linguistic modalities. However, current works are limited to the scope of object-level grounding. As shown in Table 1, the RefCOCO dataset [44] stands as one of the most widely used grounding benchmarks by far, basically containing only object-level visual and textual annotations. It does not take into account the part-level grounding task, which is crucial for future multimodal models to act as intelligent agents to realize the fine-grained perception of real world. The existing grounding datasets couldn't support the training and evaluation of such advanced capabilities. Thus, exploring how to transcend the current object-level constraints and delve into finer-grained part-level grounding, is highly meaningful and worthy of in-depth study, which is the primary focus of this work.", + "bbox": [ + 75, + 233, + 468, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In fact, prior to this work, a few works along another research line have made significant strides towards more fine-grained visual understanding. The demand for finer-grained visual understanding of objects arouses research community into constructing high-quality part-level dataset. Specifically, recent years have witnessed the introduction of various datasets that provide fine-grained part-level masks and bounding boxes annotations for either objects of specific categories [10, 33, 39] or general categories [4, 12, 26, 32]. However, these datasets essentially correspond to unimodal (i.e., visual) downstream tasks, lacking a deep connection between fine-grained part-level masks and rich textual descriptions. To our knowledge, few of previous studies have established this connection. Consequently, there appears to be a limited availability of fine-grained, large-scale vision-language datasets that facilitate part-level cross-modality understanding, which is necessary in terms of two aspects. Firstly, when describing an object, people always naturally gravitate towards detailing part-level local features, underscoring the indispensable need for multimodal agents to understand the part granularity. Secondly, a fine-grained understanding at the part level can positively promote the perception of object-level targets, especially under extreme conditions such as occlusion or deformation, which will consistently propel advancements in widely focused object-level tasks like classic visual grounding.", + "bbox": [ + 75, + 491, + 467, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we attempt to fill this impor", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tant blank space that has been neglected before and move towards finer-grained part-level RES task. Specifically, to push towards finer-grained vision-language understanding, we propose a new multi-granularity referring expression segmentation (MRES) task and an evaluation benchmark named RefCOCOm by manually annotating the part-level targets based on the previous benchmark RefCOCO with only object-level labels. We also construct the largest-scale visual grounding dataset which is also the first dataset to support part-level visual-textual annotations, and build a simple baseline model to accomplish the unified multi-granularity (i.e., object-level and part-level) RES task.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 517, + 273, + 880, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new MRES task (as shown in Fig. 1) with corresponding benchmark RefCOCOM for evaluation, pushing classic RES task towards finer-granularity understanding of real-world scenes.", + "- We build a multi-grained visual grounding dataset namely MRES-32M, which to the best of our knowledge is the first grounding dataset that supports part-level vision-language annotations and also the largest-scale visual grounding dataset.", + "- To effectively unify both the object-level and part-level RES tasks, we propose a simple yet strong model namely UniRES, which achieves the new state-of-the-art (SOTA) performance on three object-level benchmarks for classic RES tasks and our multi-granularity RefCOCOM benchmark for the proposed MRES task." + ], + "bbox": [ + 500, + 287, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 527, + 640, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Expression Segmentation. As a challenging visual grounding task the concept of RES is first proposed by [13]. Subsequent works such as [15, 16, 40, 43, 45] predominantly follow a two-step pipeline of encoding the linguistic and visual features separately and deriving the fused multimodal features from unimodal representations for mask prediction. In this way, the effectiveness of obtained multimodal representations essentially dominates the model performance, which has been continually studied in the following research [7, 22, 42, 48, 49]. Recently, there are some works [28, 36, 46] focusing on zero-shot RES task which has great application potential. In addition, a few recent works [14, 25] have been devoted to address the limitations of the existing benchmark datasets [18, 27, 41, 44] for visual grounding task. However, it's worth noting that previous works primarily focus on classic object-level RES methods and datasets, paying few attention to the vital fine-grained part-level grounding.", + "bbox": [ + 496, + 551, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Part-level Visual Perception. The increasing interest in the fine-grained understanding of objects has driven the creation of part-level annotated datasets across both specialized and general domains. Pioneering research in the former field has introduced datasets with part-level annotations, fo", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12999", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cusing on objects of particular categories like human body parts [10], animal body parts [33] and vehicle components [39]. In contrast, more general datasets that offer part annotations for a variety of common objects include Pascal-Part [4], PartNet [26], and PartImageNet [12]. Furthermore, recent work [6] introduces a finer-grained dataset with additional instance-specific part annotations for panoptic segmentation task. PACO [32] extends the available object-level datasets by including the annotations of part granularity and attribute for finer-grained segmentation and detection tasks. Moreover, VLPart [35] proposes a parsing pipeline for part segmentation and a detector that is capable of realizing the segmentation task of both open-vocabulary objects and their part regions.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-Granularity Grounding Benchmark", + "text_level": 1, + "bbox": [ + 76, + 316, + 457, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-Granularity RES Task", + "text_level": 1, + "bbox": [ + 76, + 342, + 336, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The interaction capability of a multimodal agent directly depends on the visual granularity it can perceive and understand. However, in the current research landscape, multi-granularity visual grounding remains under-explored. To break the limitations wherein visual-language alignment is constrained to the object level, we intend to take a step forward towards finer-grained vision-language understanding and propose the multi-granularity RES task, which requires models' proficiency in uniformly grounding entities across both part and object levels in response to various textual references. Within the single modality, it entails an enhancement of the capacity to apprehend complex, implicit knowledge structures, e.g., the granularity delineated by textual descriptions and the advanced level of pixel-wise semantic comprehension. Across the vision-language modalities, finer-grained references also pose new challenges for the accuracy and robustness of cross-modal alignment. Besides, since previous studies are mainly restricted to object-level referring segmentation and related benchmarks for the evaluation of part-level grounding performance are unavailable, we build a corresponding benchmark named RefCOCOM for performance evaluation, which will be elaborated below.", + "bbox": [ + 75, + 367, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. RefCOCOM Benchmark", + "text_level": 1, + "bbox": [ + 76, + 710, + 303, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manual Annotation Pipeline. To facilitate easy evaluation by community researchers without the need for any code modifications and build the multi-granularity benchmark with excellent quality, we split the existing object masks into part-level annotations through manual labeling on three subsets (i.e., validation, testA and testB) of the most prevalent RefCOCO [44] dataset. Initially, we establish a set of criteria for part masks delineation and identify the specific part categories requisite for each object. Subsequently, we employ SAM [20] to procure the preliminary segmentation results. We recruit 30 skilled annotators to refine and cap", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tioning these results through an online annotation tool developed by ourselves. To ensure the annotation quality, we conduct spot checks at three progress nodes of $20\\%$ , $50\\%$ , and $80\\%$ of the total workload with corrective feedback.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bfa889bcc52869eba53a52eb2296731a7efd7656f0d2aa8094b62d8634249620.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 509, + 162, + 723, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5cb87e73551e278622d387075ca584dd1f2185a8dcaebd70d6cda09692460bb5.jpg", + "image_caption": [ + "(b)", + "Figure 2. RefCOCom benchmark statistics. (a) the number of referring expressions per parts' category in the log scale. (b) the word cloud highlights the head categories." + ], + "image_footnote": [], + "bbox": [ + 725, + 161, + 880, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benchmark Details. Totally, we collect 70k part references and corresponding 26k masks. The average length of the references is 5.1, covering 80 object and 391 part categories, among which the number of referring expressions per parts' category and the word cloud that highlights the head categories are both presented in Fig. 2. Mixed with the original object-level annotations, our RefCOCOM has a total of 34k masks and 92k references. As an enhanced expansion of RefCOCO [44], our RefCOCOM puts forward higher requirements for referential understanding and visual perception capabilities. To better match the intention of multi-granularity unification, we leverage mean Intersection-over-Union (mIoU) as evaluation metrics.", + "bbox": [ + 496, + 339, + 893, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Multi-Granularity Grounding Dataset", + "text_level": 1, + "bbox": [ + 498, + 549, + 844, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Data Collection Engine", + "text_level": 1, + "bbox": [ + 498, + 574, + 715, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the intrinsic complexity of classic RES task, the associated training data necessitates extensive annotation costs across both textual and visual domains. When the granularity is promoted to part-level, the annotation intricacies are further exacerbated. We contend that the primary bottleneck hindering the emergence of open-world grounding is the limitation imposed by the current scalability of data. By leveraging robust foundation models for synergistic enhancement, we introduce an advanced data engine capable of automatically generating reliable visual grounding data.", + "bbox": [ + 496, + 598, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Grained Dense Captioner. As the large language models (LLMs) [5, 37] continue to empower the multimodal domain with profound capabilities, the current scope of large vision-language models (LVLMs) [1-3] has effectively expanded into the grounding domain. Formally, these models utilize numerical coordinates to represent bounding boxes, allowing positional information to be seamlessly integrated into the language model via word embeddings for contextual understanding. However, the research community's exploration into the part granularity remains nascent.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13000", + "bbox": [ + 480, + 931, + 519, + 945 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75deb53c97770984745486f7627f430217d633ff649018fea9e933d3281332a0.jpg", + "image_caption": [ + "(a) Multi-Grained Dense Captioner" + ], + "image_footnote": [], + "bbox": [ + 151, + 88, + 444, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af6fa3fe88e2cc4617bcf965ef9d6a55d037ffb2e9bf6b4ec8741412a14f48fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 89, + 813, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/558b5488a8c5d122f2d2f1455594f3594ed9ab18494079cc85e3d17fafebb97e.jpg", + "image_caption": [ + "(c) Part-Level Grounding Data Generation", + "Figure 3. The illustration of our data engine for building the MRES-32M dataset. (a) We start by fine-tuning an LVLM to create a capable dense captioner, which can effectively handle captioning tasks at three levels of granularity. (b) To generate object-level grounding data, we feed images and original bounding boxes into the dense captioner and a powerful segmenter to obtain the captions and masks of various objects. (c) We leverage the external knowledge from LLMs to decompose the existing object category annotations into a vocabulary set of part-level tags, which are sequentially fed into an open-vocabulary segmenter and our captioner to acquire the part-level annotations." + ], + "image_footnote": [], + "bbox": [ + 452, + 218, + 812, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The constrained training data have resulted in the fine-grained descriptive capabilities of LVLMs being confined merely to the object-level. To fully harness the open-world visual knowledge acquired from extensive pretraining on image-level and object-level data, we have devised a unified fine-tuning scheme that equips LVLMs to operate as dense captioners across all granularity levels, as depicted in Fig. 3 (a). Data for all three granularity levels are sourced from manual annotations to ensure reliability. The input comprises an image and the corresponding bounding box to be described, with bounding box coordinates normalized to integer values within the [0,999] range. For image granularity, we employ the COCO dataset [24], where all bboxes are uniformly represented as $(0,0)$ , $(999,999)$ . For object granularity, we utilize the Visual Genome dataset [21]. For part granularity, we draw upon unimodal semantic segmentation data [4, 12, 32] and employ a template in the form of PartNameX of ObjectNameY to construct dense captions. This unified multitask training approach can be synergistic across different granularity: it allows LVLMs to incorporate more comprehensive and detailed information to enhance part granularity descriptions. Concurrently, knowledge of part granularity assists LVLMs in generalizing knowledge within object interiors.", + "bbox": [ + 75, + 411, + 472, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model-Assisted Data Generation. For the data generation of object-level visual grounding, we capitalize on largescale object detection dataset Object365 [34] to furnish highly reliable bounding boxes. Moreover, the rich category labels inherent in it ensure a comprehensive knowledge coverage. As illustrated in Fig. 3 (b), the bounding boxes will serve as visual prompts, which are independently sent into both a promptable segmenter (i.e. segment anything model", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[20]) and our dense captioner to obtain the segmentation masks and detailed semantic descriptions, respectively.", + "bbox": [ + 498, + 411, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For part-level grounding data, we propose a hierarchical annotation scheme that builds upon existing object-level annotations, as shown in Fig. 3 (c). Specifically, employing GPT-4 [29] endowed with extensive external knowledge, we decompose the objects present in a given image to generate a customized part vocabulary set. This tailored vocabulary set is then fed into an open-vocabulary segmenter [35], which yields precise part masks and bounding boxes. These bounding boxes are subsequently sent into our dense captioner to acquire corresponding detailed captions.", + "bbox": [ + 496, + 441, + 892, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Filtering. After completing the multi-granularity annotation of all images, we further introduce CLIP [30] for filtering. The bounding box is cropped from the original image, and then sent to the encoder together with the dense caption to measure the similarity. To ensure the consistency between the visual and linguistic annotations to a great extent, we retain box-caption pairs with similarity greater than 0.5 as the final annotation results.", + "bbox": [ + 496, + 594, + 893, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. MRES-32M Dataset Details", + "text_level": 1, + "bbox": [ + 500, + 724, + 750, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As listed in Table 2, existing datasets, such as the most commonly used benchmark dataset RefCOCO [44], have limitations in terms of small data scale and lacking part-level dense annotations. As one of the pioneering works for finer-granularity unimodal visual comprehension, PACO [32] only contains semantic tags of objects or parts, without informative descriptions and visual context encompassed as referring expressions. Summarily, we compare the proposed MRES-32M with existing datasets and list some unique and significant properties of our dataset in Table", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13001", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/79d039ac35d2a83b4f5b70a4ca1df7815b6896f4987c7cea792820ce72068b2c.jpg", + "table_caption": [ + "Table 2. Comparisons with previous object-level visual grounding datasets and part-level segmentation datasets. # denotes the specific number, where Cats and Avg Len denote the object/part categories and the average length of referring expressions. “-” denotes the corresponding part masks or captions are unavailable." + ], + "table_footnote": [], + "table_body": "
Dataset#Images#Objs#Parts#Cats#Avg Len
Object-Level Visual Grouding
ReferIt [18]20K97K-238/-3.2
RefCOCO [44]20K50K-80/-3.6
RefCOCO+ [44]20K49K-80/-3.5
RefCOCOg [27]26K54K-80/-8.4
GRES [25]20K60K-80/-3.7
Part-Level Segmentation & Detection
PartsIN [12]24K24K112K158/609-
PascalPart [4]19K40K363K20/193-
PACO [32]20K260K641K75/456-
Multi-Grained Visual Grounding
MRES-32M (Ours)1M15.3M16.9M365/22994.6
", + "bbox": [ + 76, + 159, + 470, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Besides, we have also provided a few examples in our MRES-32M dataset in the appendix.", + "bbox": [ + 76, + 375, + 468, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified Multi-Granularity. In comparison to the grounding counterparts, our MRES-32M is the first visual grounding dataset covering both part and object granularity. In comparison to the part-level segmentation counterparts, our MRES-32M provides informative and unique fine-grained description for each part mask.", + "bbox": [ + 76, + 406, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More Diversified Categories. Our MRES-32M is composed of 365 object categories and an associated 2,299 part categories. Compared with existing datasets, it covers a wider range of multimodal knowledge and is an important step towards open-world understanding.", + "bbox": [ + 76, + 496, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Breakable Data Scales. To the best of our knowledge, MRES-32M is the largest dataset in the current grounding research community. In terms of the number of images and object instances, it surpasses the largest existing visual grounding dataset RefCOCOg [27] by factors of 38 and 283, respectively. Meanwhile, it encompasses part instance counts that exceed the largest existing part semantic segmentation dataset [32] by 58 times.", + "bbox": [ + 76, + 571, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More Complex References. Benefiting from our LVLM-based dense captioner, the reference of MRES-32M could be more fully combined with the visual context for entity (i.e., part and object) description. Without sticking to a specific template, the relationships and attributes of entities could be highlighted in free natural language expressions.", + "bbox": [ + 76, + 693, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Multi-Granularity RES Model", + "text_level": 1, + "bbox": [ + 76, + 799, + 357, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we describe the proposed multi-granularity RES model UniRES for the unified MRES task with the referring targets at both object and part granularity. Since our original intention is to establish a simple and easy-to-follow baseline model for the proposed multi-granularity RES task,", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the structure of our model UniRES is designed to be simple and clean. As shown in Fig. 4, UniRES has three major components, which will be illustrated below.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-Language Backbone. Taking account of the required attributes of both strong ability to capture the vision-language feature representations and promising scalability, we leverage the CLIP pre-trained weights of CLIP model [31], which learns transferable visual and linguistic concepts from vast image-text pairs, and adopt it as the backbone for our referring segmentation framework. The respective image and text encoders (i.e. Vision Transformer (ViT) [8] and Transformer [38] respectively) from CLIP are utilized to effectively extract visual and linguistic features.", + "bbox": [ + 496, + 136, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query-based Grouping Design. To effectively accomplish the proposed MRES task, it is essential to exploit both low-level local and high-level global visual features. In order to enhance the local-global visual representations of CLIP backbone without introducing much additional parameters or altering model's structure, we have incorporated 64 and 8 learnable tokens (set empirically) into the first and middle layers of CLIP visual backbone. These learnable tokens traverse the first and second halves of the visual backbone. We expect that the ViT's internal self-attention mechanism implicitly serves as a manner to perform visual grouping, obtaining representative group tokens that capture the low-level local and high-level global features simultaneously. Based on the fact that local features spread more fragmented, the number of appended low-level group tokens is greater than that of high-level counterparts. Then, group tokens from both levels are fed into a language-guided region filter (LRF) to select the language-related visual features by cross-attention mechanism, which is followed by concatenation to fuse these expression-related visual group tokens for subsequent vision-language decoding.", + "bbox": [ + 496, + 287, + 892, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-Stage V-L Decoder. Now that the visual and textual feature representations from backbone, and the expression-related group tokens with two different levels are obtained, the two-stage mask decoder which comprises stacked Transformer layers is utilized to generate the segmentation masks. Specifically, the first-stage V-L decoder takes the extracted visual and textual features as input and generates first-stage fused multimodal representations. Subsequently, these multimodal features are further integrated with the grouped expression-related region features of the two semantic levels (i.e., low-level & high-level) to realize further feature enhancement, which is followed by a linear projection layer to obtain the final segmentation masks.", + "bbox": [ + 496, + 604, + 890, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 815, + 705, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, comprehensive experiments are conducted on three classic RES datasets (i.e., RefCOCO [44], RefCOCO+ [44], RefCOCOg [27]) and our RefCOCom benchmark for multi-granularity RES.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13002", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b8862dd02c5505d6445148adf61783e93b874d375670eb69ca13edf10abdb924.jpg", + "image_caption": [ + "Figure 4. The architecture of our UniRES model as a simple baseline for the MRES task. UniRES mainly comprises three parts: the visual and textual backbone for feature extraction, the pixel grouping design for aggregating the low-level and high-level features, and the cascaded two-stage vision-language decoder for multimodal feature fusion and the generation of segmentation masks." + ], + "image_footnote": [], + "bbox": [ + 148, + 85, + 803, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 286, + 179, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The details about the three benchmark datasets for classic RES task can be found in the appendix.", + "bbox": [ + 75, + 309, + 468, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 349, + 267, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The implementation details are provided in the appendix.", + "bbox": [ + 75, + 373, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We adopt mean Intersection-over-Union (mIoU) and overall Intersection-over-Union (oIoU) as evaluation metrics. The mIoU measures the ratio between the intersection area and the union area of the predictions and labels among the test samples, while the oIoU calculates the total intersection area over total union area.", + "bbox": [ + 75, + 404, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Main Results", + "text_level": 1, + "bbox": [ + 76, + 505, + 215, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3.1 Multi-Granularity MRES Task", + "text_level": 1, + "bbox": [ + 76, + 529, + 346, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with SOTA Methods. To evaluate the multi-granularity grounding performance of our UniRES as the baseline for the proposed MRES task and previous RES methods, we further conduct experimental comparison on our newly built RefCOCOM benchmark dataset. As presented in Table 3, the classic RES methods including four specialist models (i.e., SeqTR [47], CRIS [40], LAVT [42]) and two generalist models (e.g., X-Decoder [48] and SEEM [49]) are incorporated. For fair comparisons, we re-implement these SOTA methods and report their performance on our RefCOCOM. It is clear that either the specialist models for classic RES task or the powerful generalist models perform poorly on RefCOCOM, which requires the crucial skills of both part-level and object-level referring segmentation. Benefiting from our MRES-32M dataset, our UniRES can better master the part-level RES skills and handle the multi-granularity RES task, achieving considerably higher segmentation accuracy. Due to the significantly higher difficulty of part-level RES (i.e. multi-granularity RES) compared to classic RES, the absolute value of segmentation accuracy is correspondingly lower. This further emphasizes the importance of researching finer-grained part grounding where previous SOTA methods have fallen short.", + "bbox": [ + 75, + 551, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7012d70d05bcb50ab8fd00baf18d25eb324d25d1e12c42ba409769a378098755.jpg", + "table_caption": [ + "Table 3. Comparison with previous SOTA methods on our Ref-COCOM benchmark in terms of mIoU. Part and Obj & Part denote part-only and multi-grained evaluation settings of our MRES task." + ], + "table_footnote": [], + "table_body": "
MethodsvaltestAtestB
PartObj & PartPartObj & PartPartObj & Part
Specialists
SeqTR [47]13.928.212.122.818.134.7
CRIS [40]10.625.410.121.212.930.0
LAVT [42]15.329.913.224.418.735.5
Generalists
X-Decoder [48]16.229.513.623.620.333.8
SEEM [49]16.129.413.623.420.433.9
UniRES (Ours)19.634.316.427.825.241.7
", + "bbox": [ + 503, + 327, + 887, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Analysis. We also conduct visual comparisons of CRIS [40], LAVT [42] and our UniRES on the MRES task, which is provided in appendix.", + "bbox": [ + 498, + 469, + 890, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3.2 Classic Object-Level RES Task", + "text_level": 1, + "bbox": [ + 500, + 529, + 769, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with SOTA Methods. To validate the superiority of our MRES-32M dataset and model UniRES, our framework is fairly evaluated against previous SOTA methods on RefCOCO [44], RefCOCO+ [44] and RefCOCOg [27] under both zero-shot and fine-tuning settings. As presented in Table 4, our method greatly outperforms previous methods in terms of segmentation accuracy across all the benchmark datasets. Via directly zero-shot transferring to the downstream classic RES task after pre-training on our MRES-32M dataset, Our UniRES model achieves a leading zero-shot segmentation accuracy (i.e., approximately $71\\%$ ) without any fine-tuning to adapt to downstream task's data, which is significantly higher than all the recently proposed zero-shot RES methods by a large margin (i.e., $\\uparrow 30 - 40\\%$ mIoU). At the same time, it is worth noting that our zero-shot segmentation performance is already better than many of previous RES methods under fine-tuning setting (e.g., CRIS [40] and ReSTR [19]), validating the potential of our MRES-32M dataset and UniRES model. Furthermore, after fine-tuning on the classic RES datasets, our UniRES greatly surpasses previous methods no matter the specialist models (e.g., LAVT [42] and LISA [22]) for classic RES or the generalist models (e.g., X-Decoder [48] and SEEM [49]).", + "bbox": [ + 496, + 551, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13003", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42c9d7e6f25cfdc01eda5a7f7491d1bc0b3ea38b398eb0e9b1fee2001539b730.jpg", + "table_caption": [ + "Table 4. Comparisons with the state-of-the-art approaches on previous three classic RES benchmark datasets under both the zero-shot and fine-tuning settings. “-” denotes that the result is not provided." + ], + "table_footnote": [], + "table_body": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Zero-Shot Methods
mIoURegion token [46] (CVPR-23)23.422.124.624.522.625.427.627.3
Cropping [46] (CVPR-23)24.822.625.726.324.126.531.930.9
Global-Local CLIP [46] (CVPR-23)26.224.926.627.825.627.833.533.7
SAM-CLIP [28] (arXiv-23)26.325.826.425.728.026.838.838.9
Ref-Diff [28] (arXiv-23)37.238.437.237.340.533.044.044.5
TAS [36] (arXiv-23)39.841.136.243.649.136.546.646.8
UniRES (Ours)71.274.866.059.966.751.462.363.2
Fine-Tune Methods
oIoUEFNet [9] (CVPR-21)62.865.759.751.555.243.0--
LTS [17] (CVPR-21)65.467.863.154.258.348.054.454.3
ReSTR [19] (CVPR-22)67.269.364.555.860.448.3--
ReLA [25] (CVPR-23)73.876.570.266.071.057.765.066.0
X-Decoder [48] (CVPR-23)------64.6-
SEEM [49] (arXiv-23)------65.7-
LISA [22] (arXiv-23)74.979.172.365.170.858.167.970.6
UniRES (Ours)77.480.974.769.476.161.469.071.7
mIoUVLT [7] (ICCV-21)65.768.362.755.559.249.453.056.7
RefTr [23] (NeurIPS-21)74.376.870.966.870.659.466.667.4
SeqTR [47] (ECCV-22)71.773.369.863.066.759.065.065.7
CRIS [40] (CVPR-22)70.573.266.162.368.153.759.960.4
LAVT [42] (CVPR-22)74.576.970.965.871.059.263.363.6
UniRES (Ours)79.281.676.673.078.165.871.773.2
", + "bbox": [ + 163, + 119, + 803, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 511, + 240, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on our RefCOCOM validation set. The tables below involves three evaluation settings on our RefCOCOM, using only objects as RES targets, using only parts as RES targets, and a mixed setting that combines both of the above two granularity levels.", + "bbox": [ + 75, + 535, + 470, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4.1 Ablation Study on MRES-32M Dataset", + "text_level": 1, + "bbox": [ + 76, + 619, + 403, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Granularity. We first probe into the effect of pretraining data's granularity with $50\\%$ of our MRES-32M dataset. As shown in Table 5, the baseline without pretraining on MRES-32M dataset obtains $75.2\\%$ , $15.8\\%$ and $32.0\\%$ mIoU separately on our RefCOCOM validation set under three granularity levels. Either introducing the object-level or part-level data from MRES-32M dataset for pretraining consistently results in an considerable accuracy increase across different granularity settings. In fact, pretraining on data at a certain granularity level to improve performance on corresponding benchmark aligns with common sense. However, it is noteworthy that incorporating part-level data into training also enhances model performance on the only object-level RefCOCOM. This underscores the significance of part-level understanding, as emphasized in the Sec. 1, where grasping the nuances of parts can yield benefits for object-level grounding. Besides, by", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "jointly incorporating the training samples of both granularity, our method attains $2.8\\%$ improvement against baseline under all the granularity settings on RefCOCOM, which fully demonstrate the benefit of exploiting our MRES-32M dataset for both object-level and part-level RES tasks.", + "bbox": [ + 496, + 512, + 890, + 588 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/86cb67d6e6e3a2a1e5232c2547a9b0c9369d1b61aad45ea907cad6d1fe83cae0.jpg", + "table_caption": [ + "Table 5. Ablation study on the granularity of pre-training data in our proposed MRES-32M. Object and Part denote the introduction of object-level and part-level data for pre-training." + ], + "table_footnote": [], + "table_body": "
ObjectPartRefCOCOM
Object-OnlyPart-OnlyObject&Part
75.215.830.5
✓77.515.931.1
✓75.918.432.6
✓✓78.018.633.3
", + "bbox": [ + 527, + 633, + 859, + 723 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07ead81a73f5514e52f0e4f2309ae78134f79afa5bdfeeb7b5e745b166426178.jpg", + "table_caption": [ + "Table 6. Ablation study on the data scale of MRES-32M dataset." + ], + "table_footnote": [], + "table_body": "
RatiosRefCOCOm
Object-OnlyPart-OnlyObject & Part
0%75.215.830.5
20%76.817.332.0
50%78.018.633.3
100%79.219.634.3
", + "bbox": [ + 547, + 752, + 841, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Scale. Next, we investigate the effect of different percentages of training samples in our MRES-32M dataset. The results are presented in Table 6. It is obvious in Table 6 that the model performance for our MRES tasks is", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13004", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "consistently improved with more and more employed training samples, which implicitly validates the high quality of our built dataset. As the ratios of introduced training samples continue to rise, there's no sign of diminishing gains in model performance, suggesting that our framework has significant potential with continually scaled up training data.", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/218d4d664e348d968d8e77aec2492523bd82d8c4436dac602ca63f76d24ea40e.jpg", + "table_caption": [ + "Table 7. Ablation study on the effectiveness of our MRES-32M dataset on RES SOTA methods." + ], + "table_footnote": [], + "table_body": "
MethodsMRES-32MRefCOCOM
Object-OnlyPart-OnlyObject & Part
CRIS70.510.625.4
CRIS✓73.115.529.7
LAVT74.515.329.9
LAVT✓75.719.333.2
", + "bbox": [ + 83, + 215, + 460, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Necessity. To prove the necessity and effectiveness of our MRES-32M dataset for the proposed MRES task, we take previous SOTA methods (i.e., CRIS [40] and LAVT [42]) on classic RES and compare the segmentation accuracy of the same models with or without pretraining on our MRES-32M. As shown in Table 7, the original CRIS and LAVT can already well handle the classic RES task with only object-level grounding skills, but it performs poorly on the multi-grained RefCOCOM benchmark, which requires part-level grounding capability. In contrast, pre-training on our MRES-32M consistently leads to great performance gains no matter on object-only, part-only or the multi-granularity RES tasks, because our high-quality MRES-32M dataset can effectively enable the model to handle the part-level visual grounding task and enhances the original object-level grounding capability.", + "bbox": [ + 75, + 310, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4.2 Ablation Study on UniRES Model", + "text_level": 1, + "bbox": [ + 76, + 561, + 369, + 575 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7709e178a5a64fe6db8b3cb18b2a253b6b89468ec929eeaa0e8241581d1eb1de.jpg", + "table_caption": [ + "Table 8. Ablation study on the query-based grouping design." + ], + "table_footnote": [], + "table_body": "
High-LevelLow-LevelRefCOCOM
Object-OnlyPart-OnlyObject&Part
✓74.414.929.6
74.915.230.0
✓74.915.430.1
✓✓75.215.830.5
", + "bbox": [ + 81, + 599, + 460, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, we additionally conduct ablation studies on the structural design of our model UniRES. As presented in Table 8, removing the simple but effective grouping design at two different levels (this simultaneously leads to discarding the second-stage decoder) will result in a considerable deterioration in model performance. Since the sequentially appended pixel grouping tokens are introduced to effectively capture the high-level and low-level visual features for further visual feature enhancement, losing any type of these group tokens will lead to a decrease in the model's segmentation accuracy on both object-level and part-level RES tasks. In order to verify that the high-level and low-level pixel grouping tokens can capture the clustering features of the corresponding level, we also visualize the group tokens", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the two levels (before being sent to LRF) for qualitative analysis, where the different colors represent the various areas of the same clustering group. The visualization results in Fig. 5 affirms that our introduced low-level and high-level group tokens respectively capture the part-level local features and aggregate the object-level global features with stronger semantics through the pixel grouping process, which aligns with our intentions for the grouping design.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a96cf496155f0279018f07a66691f1ed4e21b61ba6b1b3fa0acd0ebc831ea43.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 524, + 222, + 638, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dff899dcdbbb12475949cb8646a90136655ef9874e5a8e1ed11b84466837ec16.jpg", + "image_caption": [ + "(b) Low-Level" + ], + "image_footnote": [], + "bbox": [ + 640, + 222, + 751, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bac86dba758eb4ea3a68df8b2a6c2f0ccc213c34ff224fb50d9e5c3eb9d9050f.jpg", + "image_caption": [ + "(c) High-Level", + "Figure 5. Qualitative analysis for ablation study on the object-level and part-level grouping design in our model structure. (a) the input image. (b) low-level group tokens. (c) high-level group tokens." + ], + "image_footnote": [], + "bbox": [ + 754, + 222, + 867, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Broader Impact", + "text_level": 1, + "bbox": [ + 500, + 431, + 794, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we move beyond previous works that focused solely on object-level visual grounding tasks and take a step further to finer-grained part-level RES. We put forward a new multi-granularity referring expression segmentation task and establish an evaluation benchmark named RefCOCOM by manual annotation. To advance the visual grounding at both object and part levels towards finer-grained vision-language understanding, we build the largest visual grounding dataset MG-32M to date, which is also the first dataset to provide part-level vision-language annotations. Furthermore, we have developed a simple yet strong multi-grained referring segmentation model called UniRES. As a baseline for our newly proposed MRES task, UniRES achieves the new state-of-the-art performance on both our RefCOCOM for MRES task and three classic RES datasets. We plan to release our RefCOCOM benchmark, the MG32M dataset, and the UniRES model to the public, aspiring to foster future research in fine-grained visual grounding tasks and to inspire new research in this direction.", + "bbox": [ + 496, + 459, + 890, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 767, + 660, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Yepeng Tang for the helpful discussions on this work, all the Image & Video Analysis Group (IVA)'s members in CASIA for the technical support, and all the insightful reviewers for the helpful suggestions. This work was supported by the National Science and Technology Major Project (No.2022ZD0118801), National Natural Science Foundation of China (U21B2043, 62206279).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13005", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3", + "[2] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.", + "[3] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3", + "[4] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 2, 3, 4, 5", + "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3", + "[6] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 3", + "[7] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2, 7", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[9] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 7", + "[10] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 932-940, 2017. 2, 3", + "[11] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 2, 3, 4, 5", + "[13] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2", + "[14] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4067-4077, 2023. 2", + "[15] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 2", + "[16] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 2", + "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 7", + "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 5", + "[19] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 2, 6, 7", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3, 4", + "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 4", + "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 6, 7", + "[23] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13006", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in neural information processing systems, 34:19652-19664, 2021.7", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 4", + "[25] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 5, 7", + "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2, 3", + "[27] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2, 5, 6", + "[28] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 2, 7", + "[29] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 4", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[32] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 2, 3, 4, 5", + "[33] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 2, 3", + "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3, 4", + "[36] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial-aware zero-shot referring image segmentation. arXiv preprint arXiv:2310.18049, 2023. 2, 7", + "[37] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 3", + "[40] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2, 6, 7, 8", + "[41] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 2", + "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 2, 6, 7, 8", + "[43] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 2", + "[44] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 2,3,4,5,6", + "[45] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 2", + "[46] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 2, 7", + "[47] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13007", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 598-615. Springer, 2022. 6, 7", + "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2, 6, 7", + "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13008", + "bbox": [ + 480, + 931, + 517, + 944 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_model.json b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4355d55361e0b75ae35dcecf84ea03e70cf67292 --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.244, + 0.015, + 0.803, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.807, + 0.177 + ], + "angle": 0, + "content": "Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.202, + 0.831, + 0.24 + ], + "angle": 0, + "content": "Wenxuan Wang\\(^{1,2,3*}\\) Tongtian Yue\\(^{1,2*}\\) Yisi Zhang\\(^{4}\\) Longteng Guo\\(^{1}\\) Xingjian He\\(^{1}\\) Xinlong Wang\\(^{3}\\) Jing Liu\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.24, + 0.748, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Automation, Chinese Academy of Sciences (CASIA)" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.257, + 0.83, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.275, + 0.692, + 0.293 + ], + "angle": 0, + "content": "3 Beijing Academy of Artificial Intelligence (BAAI)" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.293, + 0.708, + 0.31 + ], + "angle": 0, + "content": "4 University of Science and Technology Beijing (USTB)" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.313, + 0.789, + 0.328 + ], + "angle": 0, + "content": "{wangwenxuan2023@ia.ac.cn, wangxinlong@baai.ac.cn, jliu@nlpr.ia.ac.cn}" + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.355, + 0.224, + 0.365 + ], + "angle": 0, + "content": "Images" + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.372, + 0.248, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.343, + 0.331, + 0.352 + ], + "angle": 0, + "content": "Classic RES" + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.353, + 0.331, + 0.371 + ], + "angle": 0, + "content": "(a) \"Standing Lady on left\"" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.372, + 0.347, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.423, + 0.347, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.476, + 0.339, + 0.494 + ], + "angle": 0, + "content": "(a) \"Motorcycle on the left\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.343, + 0.436, + 0.352 + ], + "angle": 0, + "content": "MRES (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.353, + 0.43, + 0.371 + ], + "angle": 0, + "content": "(a) \"Standing Lady on left" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.372, + 0.445, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.423, + 0.445, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.474, + 0.438, + 0.492 + ], + "angle": 0, + "content": "(a) \"Motorcycle on the left\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.353, + 0.526, + 0.371 + ], + "angle": 0, + "content": "(b) \"Leg of lady walking on left" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.372, + 0.533, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.423, + 0.532, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.474, + 0.526, + 0.502 + ], + "angle": 0, + "content": "(b) \"Silvery body of motorcycle on the left\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.353, + 0.61, + 0.37 + ], + "angle": 0, + "content": "(c) \"Middle girl arm\"" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.372, + 0.621, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.423, + 0.621, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.474, + 0.6, + 0.49 + ], + "angle": 0, + "content": "(c) \"Center motorcycle" + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.49, + 0.59, + 0.5 + ], + "angle": 0, + "content": "wheel\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.353, + 0.693, + 0.371 + ], + "angle": 0, + "content": "(d) \"Head of center player\"" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.372, + 0.708, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.423, + 0.708, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.475, + 0.701, + 0.493 + ], + "angle": 0, + "content": "(d) \"Front motor headlight\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.353, + 0.792, + 0.37 + ], + "angle": 0, + "content": "(e) \"Center player torso in white\"" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.372, + 0.797, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.423, + 0.797, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.475, + 0.787, + 0.503 + ], + "angle": 0, + "content": "(e) \"Transparent window of front motor\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.506, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Figure 1. Classic Referring Expression Segmentation (RES) only supports expressions that indicate a single target object, e.g., (a). Compared with classic RES, the proposed Multi-Granularity Referring Expression Segmentation (MRES) task supports expressions indicating the specific part-level regions of target objects, e.g., part-level expressions like (b)-(e) from our newly built RefCOCom benchmark." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.561, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Referring expression segmentation (RES) aims at segmenting the foreground masks of the entities that match the descriptive natural language expression. Previous datasets and methods for classic RES task heavily rely on the prior assumption that one expression must refer to object-level targets. In this paper, we take a step further to finer-grained part-level RES task. To promote the object-level RES task towards finer-grained vision-language understanding, we put forward a new multi-granularity referring expression segmentation (MRES) task and construct an evaluation benchmark called RefCOCom by manual annotations. By employing our automatic model-assisted data engine, we build the largest visual grounding dataset namely MRES-32M, which comprises over 32.2M high-quality masks and captions on the provided 1M images. Besides, a simple yet strong model named UniRES is designed to accomplish the unified object-level and part-level grounding task. Extensive experiments on our RefCOCom for MRES and three datasets (i.e., RefCOCO(+/g)) for classic RES task demon" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.562, + 0.892, + 0.637 + ], + "angle": 0, + "content": "strate the superiority of our method over previous state-of-the-art methods. To foster future research into fine-grained visual grounding, our benchmark RefCOCOM, the MRES-32M dataset and model UniRES will be publicly available at https://github.com/Rubics-Xuan/MRES." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.665, + 0.631, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.885 + ], + "angle": 0, + "content": "As one of the most challenging tasks in vision-language understanding, referring expression segmentation (RES) aims to locate specific regions at the pixel level based on a descriptive language expression. Compared to traditional visual segmentation tasks that focus on images or videos alone, RES poses greater difficulties and challenges due to the necessity of strong comprehension across modalities, but it can simultaneously alleviate the problem of predefined categories in conventional object detection or segmentation. With the real-world scene often requiring diversity in target identification, RES task holds vast potential for applications, e.g., language-based human-object interaction and interactive image editing." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since the concept of RES task was initially proposed in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.206, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "12998" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison among different referring expression datasets, including ReferIt [18], RefCOCO(+/g) [27, 44], PhraseCut [41], and our proposed MRES-32M. Part-Level: expressions that specifies various parts of the target object in the given image." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.147, + 0.468, + 0.224 + ], + "angle": 0, + "content": "
ReferItRefCOCO(+/g)PhraseCutMRES-32M
Image SourceCLEF [11]COCO [24]VG [21]Object365 [34]
Object-Level✓✓✓✓
Part-Level×××✓
Expression Typefreefreetemplatedfree
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.49 + ], + "angle": 0, + "content": "[13], various multimodal frameworks such as [13, 19, 40, 42] towards precise RES have been proposed to deal with the challenging feature extraction and alignment problems between visual and linguistic modalities. However, current works are limited to the scope of object-level grounding. As shown in Table 1, the RefCOCO dataset [44] stands as one of the most widely used grounding benchmarks by far, basically containing only object-level visual and textual annotations. It does not take into account the part-level grounding task, which is crucial for future multimodal models to act as intelligent agents to realize the fine-grained perception of real world. The existing grounding datasets couldn't support the training and evaluation of such advanced capabilities. Thus, exploring how to transcend the current object-level constraints and delve into finer-grained part-level grounding, is highly meaningful and worthy of in-depth study, which is the primary focus of this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.468, + 0.884 + ], + "angle": 0, + "content": "In fact, prior to this work, a few works along another research line have made significant strides towards more fine-grained visual understanding. The demand for finer-grained visual understanding of objects arouses research community into constructing high-quality part-level dataset. Specifically, recent years have witnessed the introduction of various datasets that provide fine-grained part-level masks and bounding boxes annotations for either objects of specific categories [10, 33, 39] or general categories [4, 12, 26, 32]. However, these datasets essentially correspond to unimodal (i.e., visual) downstream tasks, lacking a deep connection between fine-grained part-level masks and rich textual descriptions. To our knowledge, few of previous studies have established this connection. Consequently, there appears to be a limited availability of fine-grained, large-scale vision-language datasets that facilitate part-level cross-modality understanding, which is necessary in terms of two aspects. Firstly, when describing an object, people always naturally gravitate towards detailing part-level local features, underscoring the indispensable need for multimodal agents to understand the part granularity. Secondly, a fine-grained understanding at the part level can positively promote the perception of object-level targets, especially under extreme conditions such as occlusion or deformation, which will consistently propel advancements in widely focused object-level tasks like classic visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Therefore, in this work, we attempt to fill this impor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "tant blank space that has been neglected before and move towards finer-grained part-level RES task. Specifically, to push towards finer-grained vision-language understanding, we propose a new multi-granularity referring expression segmentation (MRES) task and an evaluation benchmark named RefCOCOm by manually annotating the part-level targets based on the previous benchmark RefCOCO with only object-level labels. We also construct the largest-scale visual grounding dataset which is also the first dataset to support part-level visual-textual annotations, and build a simple baseline model to accomplish the unified multi-granularity (i.e., object-level and part-level) RES task." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.274, + 0.882, + 0.287 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.891, + 0.348 + ], + "angle": 0, + "content": "- We propose a new MRES task (as shown in Fig. 1) with corresponding benchmark RefCOCOM for evaluation, pushing classic RES task towards finer-granularity understanding of real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.424 + ], + "angle": 0, + "content": "- We build a multi-grained visual grounding dataset namely MRES-32M, which to the best of our knowledge is the first grounding dataset that supports part-level vision-language annotations and also the largest-scale visual grounding dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.424, + 0.892, + 0.514 + ], + "angle": 0, + "content": "- To effectively unify both the object-level and part-level RES tasks, we propose a simple yet strong model namely UniRES, which achieves the new state-of-the-art (SOTA) performance on three object-level benchmarks for classic RES tasks and our multi-granularity RefCOCOM benchmark for the proposed MRES task." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.641, + 0.543 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Referring Expression Segmentation. As a challenging visual grounding task the concept of RES is first proposed by [13]. Subsequent works such as [15, 16, 40, 43, 45] predominantly follow a two-step pipeline of encoding the linguistic and visual features separately and deriving the fused multimodal features from unimodal representations for mask prediction. In this way, the effectiveness of obtained multimodal representations essentially dominates the model performance, which has been continually studied in the following research [7, 22, 42, 48, 49]. Recently, there are some works [28, 36, 46] focusing on zero-shot RES task which has great application potential. In addition, a few recent works [14, 25] have been devoted to address the limitations of the existing benchmark datasets [18, 27, 41, 44] for visual grounding task. However, it's worth noting that previous works primarily focus on classic object-level RES methods and datasets, paying few attention to the vital fine-grained part-level grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Part-level Visual Perception. The increasing interest in the fine-grained understanding of objects has driven the creation of part-level annotated datasets across both specialized and general domains. Pioneering research in the former field has introduced datasets with part-level annotations, fo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "12999" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "cusing on objects of particular categories like human body parts [10], animal body parts [33] and vehicle components [39]. In contrast, more general datasets that offer part annotations for a variety of common objects include Pascal-Part [4], PartNet [26], and PartImageNet [12]. Furthermore, recent work [6] introduces a finer-grained dataset with additional instance-specific part annotations for panoptic segmentation task. PACO [32] extends the available object-level datasets by including the annotations of part granularity and attribute for finer-grained segmentation and detection tasks. Moreover, VLPart [35] proposes a parsing pipeline for part segmentation and a detector that is capable of realizing the segmentation task of both open-vocabulary objects and their part regions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.318, + 0.458, + 0.337 + ], + "angle": 0, + "content": "3. Multi-Granularity Grounding Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.343, + 0.338, + 0.36 + ], + "angle": 0, + "content": "3.1. Multi-Granularity RES Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.701 + ], + "angle": 0, + "content": "The interaction capability of a multimodal agent directly depends on the visual granularity it can perceive and understand. However, in the current research landscape, multi-granularity visual grounding remains under-explored. To break the limitations wherein visual-language alignment is constrained to the object level, we intend to take a step forward towards finer-grained vision-language understanding and propose the multi-granularity RES task, which requires models' proficiency in uniformly grounding entities across both part and object levels in response to various textual references. Within the single modality, it entails an enhancement of the capacity to apprehend complex, implicit knowledge structures, e.g., the granularity delineated by textual descriptions and the advanced level of pixel-wise semantic comprehension. Across the vision-language modalities, finer-grained references also pose new challenges for the accuracy and robustness of cross-modal alignment. Besides, since previous studies are mainly restricted to object-level referring segmentation and related benchmarks for the evaluation of part-level grounding performance are unavailable, we build a corresponding benchmark named RefCOCOM for performance evaluation, which will be elaborated below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.305, + 0.727 + ], + "angle": 0, + "content": "3.2. RefCOCOM Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Manual Annotation Pipeline. To facilitate easy evaluation by community researchers without the need for any code modifications and build the multi-granularity benchmark with excellent quality, we split the existing object masks into part-level annotations through manual labeling on three subsets (i.e., validation, testA and testB) of the most prevalent RefCOCO [44] dataset. Initially, we establish a set of criteria for part masks delineation and identify the specific part categories requisite for each object. Subsequently, we employ SAM [20] to procure the preliminary segmentation results. We recruit 30 skilled annotators to refine and cap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tioning these results through an online annotation tool developed by ourselves. To ensure the annotation quality, we conduct spot checks at three progress nodes of \\(20\\%\\), \\(50\\%\\), and \\(80\\%\\) of the total workload with corrective feedback." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.163, + 0.725, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.283, + 0.634, + 0.294 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.162, + 0.882, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.283, + 0.81, + 0.294 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.338 + ], + "angle": 0, + "content": "Figure 2. RefCOCom benchmark statistics. (a) the number of referring expressions per parts' category in the log scale. (b) the word cloud highlights the head categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.895, + 0.536 + ], + "angle": 0, + "content": "Benchmark Details. Totally, we collect 70k part references and corresponding 26k masks. The average length of the references is 5.1, covering 80 object and 391 part categories, among which the number of referring expressions per parts' category and the word cloud that highlights the head categories are both presented in Fig. 2. Mixed with the original object-level annotations, our RefCOCOM has a total of 34k masks and 92k references. As an enhanced expansion of RefCOCO [44], our RefCOCOM puts forward higher requirements for referential understanding and visual perception capabilities. To better match the intention of multi-granularity unification, we leverage mean Intersection-over-Union (mIoU) as evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.55, + 0.846, + 0.568 + ], + "angle": 0, + "content": "4. Multi-Granularity Grounding Dataset" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.575, + 0.717, + 0.592 + ], + "angle": 0, + "content": "4.1. Data Collection Engine" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Given the intrinsic complexity of classic RES task, the associated training data necessitates extensive annotation costs across both textual and visual domains. When the granularity is promoted to part-level, the annotation intricacies are further exacerbated. We contend that the primary bottleneck hindering the emergence of open-world grounding is the limitation imposed by the current scalability of data. By leveraging robust foundation models for synergistic enhancement, we introduce an advanced data engine capable of automatically generating reliable visual grounding data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Multi-Grained Dense Captioner. As the large language models (LLMs) [5, 37] continue to empower the multimodal domain with profound capabilities, the current scope of large vision-language models (LVLMs) [1-3] has effectively expanded into the grounding domain. Formally, these models utilize numerical coordinates to represent bounding boxes, allowing positional information to be seamlessly integrated into the language model via word embeddings for contextual understanding. However, the research community's exploration into the part granularity remains nascent." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.52, + 0.946 + ], + "angle": 0, + "content": "13000" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.153, + 0.089, + 0.446, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.39, + 0.324 + ], + "angle": 0, + "content": "(a) Multi-Grained Dense Captioner" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.09, + 0.815, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.219, + 0.813, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.312, + 0.724, + 0.324 + ], + "angle": 0, + "content": "(c) Part-Level Grounding Data Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Figure 3. The illustration of our data engine for building the MRES-32M dataset. (a) We start by fine-tuning an LVLM to create a capable dense captioner, which can effectively handle captioning tasks at three levels of granularity. (b) To generate object-level grounding data, we feed images and original bounding boxes into the dense captioner and a powerful segmenter to obtain the captions and masks of various objects. (c) We leverage the external knowledge from LLMs to decompose the existing object category annotations into a vocabulary set of part-level tags, which are sequentially fed into an open-vocabulary segmenter and our captioner to acquire the part-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.775 + ], + "angle": 0, + "content": "The constrained training data have resulted in the fine-grained descriptive capabilities of LVLMs being confined merely to the object-level. To fully harness the open-world visual knowledge acquired from extensive pretraining on image-level and object-level data, we have devised a unified fine-tuning scheme that equips LVLMs to operate as dense captioners across all granularity levels, as depicted in Fig. 3 (a). Data for all three granularity levels are sourced from manual annotations to ensure reliability. The input comprises an image and the corresponding bounding box to be described, with bounding box coordinates normalized to integer values within the [0,999] range. For image granularity, we employ the COCO dataset [24], where all bboxes are uniformly represented as \\((0,0)\\), \\((999,999)\\). For object granularity, we utilize the Visual Genome dataset [21]. For part granularity, we draw upon unimodal semantic segmentation data [4, 12, 32] and employ a template in the form of PartNameX of ObjectNameY to construct dense captions. This unified multitask training approach can be synergistic across different granularity: it allows LVLMs to incorporate more comprehensive and detailed information to enhance part granularity descriptions. Concurrently, knowledge of part granularity assists LVLMs in generalizing knowledge within object interiors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Model-Assisted Data Generation. For the data generation of object-level visual grounding, we capitalize on largescale object detection dataset Object365 [34] to furnish highly reliable bounding boxes. Moreover, the rich category labels inherent in it ensure a comprehensive knowledge coverage. As illustrated in Fig. 3 (b), the bounding boxes will serve as visual prompts, which are independently sent into both a promptable segmenter (i.e. segment anything model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[20]) and our dense captioner to obtain the segmentation masks and detailed semantic descriptions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.893, + 0.593 + ], + "angle": 0, + "content": "For part-level grounding data, we propose a hierarchical annotation scheme that builds upon existing object-level annotations, as shown in Fig. 3 (c). Specifically, employing GPT-4 [29] endowed with extensive external knowledge, we decompose the objects present in a given image to generate a customized part vocabulary set. This tailored vocabulary set is then fed into an open-vocabulary segmenter [35], which yields precise part masks and bounding boxes. These bounding boxes are subsequently sent into our dense captioner to acquire corresponding detailed captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Data Filtering. After completing the multi-granularity annotation of all images, we further introduce CLIP [30] for filtering. The bounding box is cropped from the original image, and then sent to the encoder together with the dense caption to measure the similarity. To ensure the consistency between the visual and linguistic annotations to a great extent, we retain box-caption pairs with similarity greater than 0.5 as the final annotation results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.75, + 0.741 + ], + "angle": 0, + "content": "4.2. MRES-32M Dataset Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As listed in Table 2, existing datasets, such as the most commonly used benchmark dataset RefCOCO [44], have limitations in terms of small data scale and lacking part-level dense annotations. As one of the pioneering works for finer-granularity unimodal visual comprehension, PACO [32] only contains semantic tags of objects or parts, without informative descriptions and visual context encompassed as referring expressions. Summarily, we compare the proposed MRES-32M with existing datasets and list some unique and significant properties of our dataset in Table" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "13001" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.159 + ], + "angle": 0, + "content": "Table 2. Comparisons with previous object-level visual grounding datasets and part-level segmentation datasets. # denotes the specific number, where Cats and Avg Len denote the object/part categories and the average length of referring expressions. “-” denotes the corresponding part masks or captions are unavailable." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.16, + 0.472, + 0.366 + ], + "angle": 0, + "content": "
Dataset#Images#Objs#Parts#Cats#Avg Len
Object-Level Visual Grouding
ReferIt [18]20K97K-238/-3.2
RefCOCO [44]20K50K-80/-3.6
RefCOCO+ [44]20K49K-80/-3.5
RefCOCOg [27]26K54K-80/-8.4
GRES [25]20K60K-80/-3.7
Part-Level Segmentation & Detection
PartsIN [12]24K24K112K158/609-
PascalPart [4]19K40K363K20/193-
PACO [32]20K260K641K75/456-
Multi-Grained Visual Grounding
MRES-32M (Ours)1M15.3M16.9M365/22994.6
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.376, + 0.47, + 0.406 + ], + "angle": 0, + "content": "2. Besides, we have also provided a few examples in our MRES-32M dataset in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.47, + 0.497 + ], + "angle": 0, + "content": "Unified Multi-Granularity. In comparison to the grounding counterparts, our MRES-32M is the first visual grounding dataset covering both part and object granularity. In comparison to the part-level segmentation counterparts, our MRES-32M provides informative and unique fine-grained description for each part mask." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.47, + 0.572 + ], + "angle": 0, + "content": "More Diversified Categories. Our MRES-32M is composed of 365 object categories and an associated 2,299 part categories. Compared with existing datasets, it covers a wider range of multimodal knowledge and is an important step towards open-world understanding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Breakable Data Scales. To the best of our knowledge, MRES-32M is the largest dataset in the current grounding research community. In terms of the number of images and object instances, it surpasses the largest existing visual grounding dataset RefCOCOg [27] by factors of 38 and 283, respectively. Meanwhile, it encompasses part instance counts that exceed the largest existing part semantic segmentation dataset [32] by 58 times." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.694, + 0.47, + 0.785 + ], + "angle": 0, + "content": "More Complex References. Benefiting from our LVLM-based dense captioner, the reference of MRES-32M could be more fully combined with the visual context for entity (i.e., part and object) description. Without sticking to a specific template, the relationships and attributes of entities could be highlighted in free natural language expressions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.8, + 0.358, + 0.816 + ], + "angle": 0, + "content": "5. Multi-Granularity RES Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Next, we describe the proposed multi-granularity RES model UniRES for the unified MRES task with the referring targets at both object and part granularity. Since our original intention is to establish a simple and easy-to-follow baseline model for the proposed multi-granularity RES task," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "the structure of our model UniRES is designed to be simple and clean. As shown in Fig. 4, UniRES has three major components, which will be illustrated below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Vision-Language Backbone. Taking account of the required attributes of both strong ability to capture the vision-language feature representations and promising scalability, we leverage the CLIP pre-trained weights of CLIP model [31], which learns transferable visual and linguistic concepts from vast image-text pairs, and adopt it as the backbone for our referring segmentation framework. The respective image and text encoders (i.e. Vision Transformer (ViT) [8] and Transformer [38] respectively) from CLIP are utilized to effectively extract visual and linguistic features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.893, + 0.605 + ], + "angle": 0, + "content": "Query-based Grouping Design. To effectively accomplish the proposed MRES task, it is essential to exploit both low-level local and high-level global visual features. In order to enhance the local-global visual representations of CLIP backbone without introducing much additional parameters or altering model's structure, we have incorporated 64 and 8 learnable tokens (set empirically) into the first and middle layers of CLIP visual backbone. These learnable tokens traverse the first and second halves of the visual backbone. We expect that the ViT's internal self-attention mechanism implicitly serves as a manner to perform visual grouping, obtaining representative group tokens that capture the low-level local and high-level global features simultaneously. Based on the fact that local features spread more fragmented, the number of appended low-level group tokens is greater than that of high-level counterparts. Then, group tokens from both levels are fed into a language-guided region filter (LRF) to select the language-related visual features by cross-attention mechanism, which is followed by concatenation to fuse these expression-related visual group tokens for subsequent vision-language decoding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.802 + ], + "angle": 0, + "content": "Two-Stage V-L Decoder. Now that the visual and textual feature representations from backbone, and the expression-related group tokens with two different levels are obtained, the two-stage mask decoder which comprises stacked Transformer layers is utilized to generate the segmentation masks. Specifically, the first-stage V-L decoder takes the extracted visual and textual features as input and generates first-stage fused multimodal representations. Subsequently, these multimodal features are further integrated with the grouped expression-related region features of the two semantic levels (i.e., low-level & high-level) to realize further feature enhancement, which is followed by a linear projection layer to obtain the final segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.706, + 0.832 + ], + "angle": 0, + "content": "6. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, comprehensive experiments are conducted on three classic RES datasets (i.e., RefCOCO [44], RefCOCO+ [44], RefCOCOg [27]) and our RefCOCom benchmark for multi-granularity RES." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.519, + 0.945 + ], + "angle": 0, + "content": "13002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.087, + 0.805, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 4. The architecture of our UniRES model as a simple baseline for the MRES task. UniRES mainly comprises three parts: the visual and textual backbone for feature extraction, the pixel grouping design for aggregating the low-level and high-level features, and the cascaded two-stage vision-language decoder for multimodal feature fusion and the generation of segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.181, + 0.301 + ], + "angle": 0, + "content": "6.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.47, + 0.34 + ], + "angle": 0, + "content": "The details about the three benchmark datasets for classic RES task can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.351, + 0.269, + 0.367 + ], + "angle": 0, + "content": "6.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.469, + 0.403 + ], + "angle": 0, + "content": "Implementation Details. The implementation details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.469, + 0.493 + ], + "angle": 0, + "content": "Evaluation metrics. We adopt mean Intersection-over-Union (mIoU) and overall Intersection-over-Union (oIoU) as evaluation metrics. The mIoU measures the ratio between the intersection area and the union area of the predictions and labels among the test samples, while the oIoU calculates the total intersection area over total union area." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.506, + 0.216, + 0.52 + ], + "angle": 0, + "content": "6.3. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.348, + 0.544 + ], + "angle": 0, + "content": "6.3.1 Multi-Granularity MRES Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Comparison with SOTA Methods. To evaluate the multi-granularity grounding performance of our UniRES as the baseline for the proposed MRES task and previous RES methods, we further conduct experimental comparison on our newly built RefCOCOM benchmark dataset. As presented in Table 3, the classic RES methods including four specialist models (i.e., SeqTR [47], CRIS [40], LAVT [42]) and two generalist models (e.g., X-Decoder [48] and SEEM [49]) are incorporated. For fair comparisons, we re-implement these SOTA methods and report their performance on our RefCOCOM. It is clear that either the specialist models for classic RES task or the powerful generalist models perform poorly on RefCOCOM, which requires the crucial skills of both part-level and object-level referring segmentation. Benefiting from our MRES-32M dataset, our UniRES can better master the part-level RES skills and handle the multi-granularity RES task, achieving considerably higher segmentation accuracy. Due to the significantly higher difficulty of part-level RES (i.e. multi-granularity RES) compared to classic RES, the absolute value of segmentation accuracy is correspondingly lower. This further emphasizes the importance of researching finer-grained part grounding where previous SOTA methods have fallen short." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.285, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Table 3. Comparison with previous SOTA methods on our Ref-COCOM benchmark in terms of mIoU. Part and Obj & Part denote part-only and multi-grained evaluation settings of our MRES task." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.328, + 0.888, + 0.463 + ], + "angle": 0, + "content": "
MethodsvaltestAtestB
PartObj & PartPartObj & PartPartObj & Part
Specialists
SeqTR [47]13.928.212.122.818.134.7
CRIS [40]10.625.410.121.212.930.0
LAVT [42]15.329.913.224.418.735.5
Generalists
X-Decoder [48]16.229.513.623.620.333.8
SEEM [49]16.129.413.623.420.433.9
UniRES (Ours)19.634.316.427.825.241.7
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Qualitative Analysis. We also conduct visual comparisons of CRIS [40], LAVT [42] and our UniRES on the MRES task, which is provided in appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.771, + 0.545 + ], + "angle": 0, + "content": "6.3.2 Classic Object-Level RES Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Comparison with SOTA Methods. To validate the superiority of our MRES-32M dataset and model UniRES, our framework is fairly evaluated against previous SOTA methods on RefCOCO [44], RefCOCO+ [44] and RefCOCOg [27] under both zero-shot and fine-tuning settings. As presented in Table 4, our method greatly outperforms previous methods in terms of segmentation accuracy across all the benchmark datasets. Via directly zero-shot transferring to the downstream classic RES task after pre-training on our MRES-32M dataset, Our UniRES model achieves a leading zero-shot segmentation accuracy (i.e., approximately \\(71\\%\\)) without any fine-tuning to adapt to downstream task's data, which is significantly higher than all the recently proposed zero-shot RES methods by a large margin (i.e., \\(\\uparrow 30 - 40\\%\\) mIoU). At the same time, it is worth noting that our zero-shot segmentation performance is already better than many of previous RES methods under fine-tuning setting (e.g., CRIS [40] and ReSTR [19]), validating the potential of our MRES-32M dataset and UniRES model. Furthermore, after fine-tuning on the classic RES datasets, our UniRES greatly surpasses previous methods no matter the specialist models (e.g., LAVT [42] and LISA [22]) for classic RES or the generalist models (e.g., X-Decoder [48] and SEEM [49])." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "13003" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.091, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4. Comparisons with the state-of-the-art approaches on previous three classic RES benchmark datasets under both the zero-shot and fine-tuning settings. “-” denotes that the result is not provided." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.121, + 0.805, + 0.501 + ], + "angle": 0, + "content": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Zero-Shot Methods
mIoURegion token [46] (CVPR-23)23.422.124.624.522.625.427.627.3
Cropping [46] (CVPR-23)24.822.625.726.324.126.531.930.9
Global-Local CLIP [46] (CVPR-23)26.224.926.627.825.627.833.533.7
SAM-CLIP [28] (arXiv-23)26.325.826.425.728.026.838.838.9
Ref-Diff [28] (arXiv-23)37.238.437.237.340.533.044.044.5
TAS [36] (arXiv-23)39.841.136.243.649.136.546.646.8
UniRES (Ours)71.274.866.059.966.751.462.363.2
Fine-Tune Methods
oIoUEFNet [9] (CVPR-21)62.865.759.751.555.243.0--
LTS [17] (CVPR-21)65.467.863.154.258.348.054.454.3
ReSTR [19] (CVPR-22)67.269.364.555.860.448.3--
ReLA [25] (CVPR-23)73.876.570.266.071.057.765.066.0
X-Decoder [48] (CVPR-23)------64.6-
SEEM [49] (arXiv-23)------65.7-
LISA [22] (arXiv-23)74.979.172.365.170.858.167.970.6
UniRES (Ours)77.480.974.769.476.161.469.071.7
mIoUVLT [7] (ICCV-21)65.768.362.755.559.249.453.056.7
RefTr [23] (NeurIPS-21)74.376.870.966.870.659.466.667.4
SeqTR [47] (ECCV-22)71.773.369.863.066.759.065.065.7
CRIS [40] (CVPR-22)70.573.266.162.368.153.759.960.4
LAVT [42] (CVPR-22)74.576.970.965.871.059.263.363.6
UniRES (Ours)79.281.676.673.078.165.871.773.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.241, + 0.527 + ], + "angle": 0, + "content": "6.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.471, + 0.613 + ], + "angle": 0, + "content": "We conduct ablation experiments on our RefCOCOM validation set. The tables below involves three evaluation settings on our RefCOCOM, using only objects as RES targets, using only parts as RES targets, and a mixed setting that combines both of the above two granularity levels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.405, + 0.636 + ], + "angle": 0, + "content": "6.4.1 Ablation Study on MRES-32M Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Data Granularity. We first probe into the effect of pretraining data's granularity with \\(50\\%\\) of our MRES-32M dataset. As shown in Table 5, the baseline without pretraining on MRES-32M dataset obtains \\(75.2\\%\\), \\(15.8\\%\\) and \\(32.0\\%\\) mIoU separately on our RefCOCOM validation set under three granularity levels. Either introducing the object-level or part-level data from MRES-32M dataset for pretraining consistently results in an considerable accuracy increase across different granularity settings. In fact, pretraining on data at a certain granularity level to improve performance on corresponding benchmark aligns with common sense. However, it is noteworthy that incorporating part-level data into training also enhances model performance on the only object-level RefCOCOM. This underscores the significance of part-level understanding, as emphasized in the Sec. 1, where grasping the nuances of parts can yield benefits for object-level grounding. Besides, by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.589 + ], + "angle": 0, + "content": "jointly incorporating the training samples of both granularity, our method attains \\(2.8\\%\\) improvement against baseline under all the granularity settings on RefCOCOM, which fully demonstrate the benefit of exploiting our MRES-32M dataset for both object-level and part-level RES tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Table 5. Ablation study on the granularity of pre-training data in our proposed MRES-32M. Object and Part denote the introduction of object-level and part-level data for pre-training." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.635, + 0.86, + 0.724 + ], + "angle": 0, + "content": "
ObjectPartRefCOCOM
Object-OnlyPart-OnlyObject&Part
75.215.830.5
✓77.515.931.1
✓75.918.432.6
✓✓78.018.633.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.737, + 0.887, + 0.751 + ], + "angle": 0, + "content": "Table 6. Ablation study on the data scale of MRES-32M dataset." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.753, + 0.842, + 0.841 + ], + "angle": 0, + "content": "
RatiosRefCOCOm
Object-OnlyPart-OnlyObject & Part
0%75.215.830.5
20%76.817.332.0
50%78.018.633.3
100%79.219.634.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data Scale. Next, we investigate the effect of different percentages of training samples in our MRES-32M dataset. The results are presented in Table 6. It is obvious in Table 6 that the model performance for our MRES tasks is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.519, + 0.945 + ], + "angle": 0, + "content": "13004" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.184 + ], + "angle": 0, + "content": "consistently improved with more and more employed training samples, which implicitly validates the high quality of our built dataset. As the ratios of introduced training samples continue to rise, there's no sign of diminishing gains in model performance, suggesting that our framework has significant potential with continually scaled up training data." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.187, + 0.468, + 0.214 + ], + "angle": 0, + "content": "Table 7. Ablation study on the effectiveness of our MRES-32M dataset on RES SOTA methods." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.217, + 0.462, + 0.307 + ], + "angle": 0, + "content": "
MethodsMRES-32MRefCOCOM
Object-OnlyPart-OnlyObject & Part
CRIS70.510.625.4
CRIS✓73.115.529.7
LAVT74.515.329.9
LAVT✓75.719.333.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Dataset Necessity. To prove the necessity and effectiveness of our MRES-32M dataset for the proposed MRES task, we take previous SOTA methods (i.e., CRIS [40] and LAVT [42]) on classic RES and compare the segmentation accuracy of the same models with or without pretraining on our MRES-32M. As shown in Table 7, the original CRIS and LAVT can already well handle the classic RES task with only object-level grounding skills, but it performs poorly on the multi-grained RefCOCOM benchmark, which requires part-level grounding capability. In contrast, pre-training on our MRES-32M consistently leads to great performance gains no matter on object-only, part-only or the multi-granularity RES tasks, because our high-quality MRES-32M dataset can effectively enable the model to handle the part-level visual grounding task and enhances the original object-level grounding capability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.37, + 0.577 + ], + "angle": 0, + "content": "6.4.2 Ablation Study on UniRES Model" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.584, + 0.453, + 0.598 + ], + "angle": 0, + "content": "Table 8. Ablation study on the query-based grouping design." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.6, + 0.462, + 0.688 + ], + "angle": 0, + "content": "
High-LevelLow-LevelRefCOCOM
Object-OnlyPart-OnlyObject&Part
✓74.414.929.6
74.915.230.0
✓74.915.430.1
✓✓75.215.830.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Besides, we additionally conduct ablation studies on the structural design of our model UniRES. As presented in Table 8, removing the simple but effective grouping design at two different levels (this simultaneously leads to discarding the second-stage decoder) will result in a considerable deterioration in model performance. Since the sequentially appended pixel grouping tokens are introduced to effectively capture the high-level and low-level visual features for further visual feature enhancement, losing any type of these group tokens will lead to a decrease in the model's segmentation accuracy on both object-level and part-level RES tasks. In order to verify that the high-level and low-level pixel grouping tokens can capture the clustering features of the corresponding level, we also visualize the group tokens" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "of the two levels (before being sent to LRF) for qualitative analysis, where the different colors represent the various areas of the same clustering group. The visualization results in Fig. 5 affirms that our introduced low-level and high-level group tokens respectively capture the part-level local features and aggregate the object-level global features with stronger semantics through the pixel grouping process, which aligns with our intentions for the grouping design." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.223, + 0.639, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.361, + 0.605, + 0.371 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.223, + 0.753, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.361, + 0.728, + 0.371 + ], + "angle": 0, + "content": "(b) Low-Level" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.223, + 0.868, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.361, + 0.843, + 0.371 + ], + "angle": 0, + "content": "(c) High-Level" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.372, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Qualitative analysis for ablation study on the object-level and part-level grouping design in our model structure. (a) the input image. (b) low-level group tokens. (c) high-level group tokens." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.795, + 0.45 + ], + "angle": 0, + "content": "7. Conclusion and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.747 + ], + "angle": 0, + "content": "In this paper, we move beyond previous works that focused solely on object-level visual grounding tasks and take a step further to finer-grained part-level RES. We put forward a new multi-granularity referring expression segmentation task and establish an evaluation benchmark named RefCOCOM by manual annotation. To advance the visual grounding at both object and part levels towards finer-grained vision-language understanding, we build the largest visual grounding dataset MG-32M to date, which is also the first dataset to provide part-level vision-language annotations. Furthermore, we have developed a simple yet strong multi-grained referring segmentation model called UniRES. As a baseline for our newly proposed MRES task, UniRES achieves the new state-of-the-art performance on both our RefCOCOM for MRES task and three classic RES datasets. We plan to release our RefCOCOM benchmark, the MG32M dataset, and the UniRES model to the public, aspiring to foster future research in fine-grained visual grounding tasks and to inspire new research in this direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.661, + 0.785 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We thank Yepeng Tang for the helpful discussions on this work, all the Image & Video Analysis Group (IVA)'s members in CASIA for the technical support, and all the insightful reviewers for the helpful suggestions. This work was supported by the National Science and Technology Major Project (No.2022ZD0118801), National Natural Science Foundation of China (U21B2043, 62206279)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "13005" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[6] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[7] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[9] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[10] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 932-940, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[12] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[13] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[14] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4067-4077, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[15] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[16] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[19] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.519, + 0.945 + ], + "angle": 0, + "content": "13006" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "in neural information processing systems, 34:19652-19664, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[25] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[27] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[28] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[29] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[32] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[33] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[36] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial-aware zero-shot referring image segmentation. arXiv preprint arXiv:2310.18049, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[37] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[40] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[41] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[43] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[44] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 2,3,4,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[45] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[46] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.519, + 0.945 + ], + "angle": 0, + "content": "13007" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 598-615. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.933, + 0.518, + 0.945 + ], + "angle": 0, + "content": "13008" + } + ] +] \ No newline at end of file diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_origin.pdf b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01f2f89d20e0d48f9818cb4398a9c68c45b59c1f --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/0ef1dc22-21b7-474a-a254-d6f2d4943d35_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be08ecfe99724a20022972e05492bcc7479314d61aa8926b0b35e564af77cbbe +size 2227082 diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/full.md b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8e97ab502ff5d6904901ff170a916c12abc1d24 --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/full.md @@ -0,0 +1,323 @@ +# Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation + +Wenxuan Wang $^{1,2,3*}$ Tongtian Yue $^{1,2*}$ Yisi Zhang $^{4}$ Longteng Guo $^{1}$ Xingjian He $^{1}$ Xinlong Wang $^{3}$ Jing Liu $^{1,2\dagger}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences (CASIA) + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) + +3 Beijing Academy of Artificial Intelligence (BAAI) + +4 University of Science and Technology Beijing (USTB) + +{wangwenxuan2023@ia.ac.cn, wangxinlong@baai.ac.cn, jliu@nlpr.ia.ac.cn} + +![](images/4215409bdd47d926dcce2e13dea41ea07440dc14afad1ff4d2d55d72c1b1509b.jpg) +Images + +![](images/bbc3328c82745a4b6a22d70666a91c4718bf26808b4d2704e2262755b5fa6a3a.jpg) +(a) "Standing Lady on left" + +![](images/eeb8c721ef56e1204bb01f690cca9d8ffb87371a4f4cd036ffc32d82ee7de2b8.jpg) +(a) "Motorcycle on the left" +Figure 1. Classic Referring Expression Segmentation (RES) only supports expressions that indicate a single target object, e.g., (a). Compared with classic RES, the proposed Multi-Granularity Referring Expression Segmentation (MRES) task supports expressions indicating the specific part-level regions of target objects, e.g., part-level expressions like (b)-(e) from our newly built RefCOCom benchmark. + +![](images/c71f2b66003c1c2aa26d45a5e887e33def7e3acc9f193c7d1e27746bae22999b.jpg) +Classic RES +(a) "Standing Lady on left + +![](images/72989edfd0c2cb4a7208cb18677905e0ff4027889f113e892fe23ef8428fe33a.jpg) +(a) "Motorcycle on the left" + +![](images/e8dd8a3911eaa263c703b098f78689f6a0c1eb7dfd56c5e1b24f5ab8e36ef82a.jpg) +MRES (Ours) +(b) "Leg of lady walking on left + +![](images/b536f2226f064f562556624889d0051583eff2cd58a1d850bf4788b64be66564.jpg) +(b) "Silvery body of motorcycle on the left" + +![](images/f38fde52231ae2a6bc845c2b186b392753629f423c85c1a8c9a8c1bb1b9bef07.jpg) +(c) "Middle girl arm" + +![](images/b7c242ff774c85e64aae5243310e2657560ed6af24874b082c12481aa63f89aa.jpg) +(c) "Center motorcycle +wheel" + +![](images/e0e2bfe02d8fc8715b66e988b6464860be8adccc37fcf02f137a4c2d87cafda2.jpg) +(d) "Head of center player" + +![](images/d545c6fdedf3b74e6c7cb0dcac3cc2192f62b05d43d55e91fcb7e86b48af7456.jpg) +(d) "Front motor headlight" + +![](images/5ac4e97f1a4cfe3f3f280283bbe4cb9a29004bb5553e71403029d026807f8179.jpg) +(e) "Center player torso in white" + +![](images/0421597dd1692e22de9fe0803a7b084372ea284c70d72b9682643f48b9b52958.jpg) +(e) "Transparent window of front motor" + +# Abstract + +Referring expression segmentation (RES) aims at segmenting the foreground masks of the entities that match the descriptive natural language expression. Previous datasets and methods for classic RES task heavily rely on the prior assumption that one expression must refer to object-level targets. In this paper, we take a step further to finer-grained part-level RES task. To promote the object-level RES task towards finer-grained vision-language understanding, we put forward a new multi-granularity referring expression segmentation (MRES) task and construct an evaluation benchmark called RefCOCom by manual annotations. By employing our automatic model-assisted data engine, we build the largest visual grounding dataset namely MRES-32M, which comprises over 32.2M high-quality masks and captions on the provided 1M images. Besides, a simple yet strong model named UniRES is designed to accomplish the unified object-level and part-level grounding task. Extensive experiments on our RefCOCom for MRES and three datasets (i.e., RefCOCO(+/g)) for classic RES task demon + +strate the superiority of our method over previous state-of-the-art methods. To foster future research into fine-grained visual grounding, our benchmark RefCOCOM, the MRES-32M dataset and model UniRES will be publicly available at https://github.com/Rubics-Xuan/MRES. + +# 1. Introduction + +As one of the most challenging tasks in vision-language understanding, referring expression segmentation (RES) aims to locate specific regions at the pixel level based on a descriptive language expression. Compared to traditional visual segmentation tasks that focus on images or videos alone, RES poses greater difficulties and challenges due to the necessity of strong comprehension across modalities, but it can simultaneously alleviate the problem of predefined categories in conventional object detection or segmentation. With the real-world scene often requiring diversity in target identification, RES task holds vast potential for applications, e.g., language-based human-object interaction and interactive image editing. + +Since the concept of RES task was initially proposed in + +Table 1. Comparison among different referring expression datasets, including ReferIt [18], RefCOCO(+/g) [27, 44], PhraseCut [41], and our proposed MRES-32M. Part-Level: expressions that specifies various parts of the target object in the given image. + +
ReferItRefCOCO(+/g)PhraseCutMRES-32M
Image SourceCLEF [11]COCO [24]VG [21]Object365 [34]
Object-Level✓✓✓✓
Part-Level×××✓
Expression Typefreefreetemplatedfree
+ +[13], various multimodal frameworks such as [13, 19, 40, 42] towards precise RES have been proposed to deal with the challenging feature extraction and alignment problems between visual and linguistic modalities. However, current works are limited to the scope of object-level grounding. As shown in Table 1, the RefCOCO dataset [44] stands as one of the most widely used grounding benchmarks by far, basically containing only object-level visual and textual annotations. It does not take into account the part-level grounding task, which is crucial for future multimodal models to act as intelligent agents to realize the fine-grained perception of real world. The existing grounding datasets couldn't support the training and evaluation of such advanced capabilities. Thus, exploring how to transcend the current object-level constraints and delve into finer-grained part-level grounding, is highly meaningful and worthy of in-depth study, which is the primary focus of this work. + +In fact, prior to this work, a few works along another research line have made significant strides towards more fine-grained visual understanding. The demand for finer-grained visual understanding of objects arouses research community into constructing high-quality part-level dataset. Specifically, recent years have witnessed the introduction of various datasets that provide fine-grained part-level masks and bounding boxes annotations for either objects of specific categories [10, 33, 39] or general categories [4, 12, 26, 32]. However, these datasets essentially correspond to unimodal (i.e., visual) downstream tasks, lacking a deep connection between fine-grained part-level masks and rich textual descriptions. To our knowledge, few of previous studies have established this connection. Consequently, there appears to be a limited availability of fine-grained, large-scale vision-language datasets that facilitate part-level cross-modality understanding, which is necessary in terms of two aspects. Firstly, when describing an object, people always naturally gravitate towards detailing part-level local features, underscoring the indispensable need for multimodal agents to understand the part granularity. Secondly, a fine-grained understanding at the part level can positively promote the perception of object-level targets, especially under extreme conditions such as occlusion or deformation, which will consistently propel advancements in widely focused object-level tasks like classic visual grounding. + +Therefore, in this work, we attempt to fill this impor + +tant blank space that has been neglected before and move towards finer-grained part-level RES task. Specifically, to push towards finer-grained vision-language understanding, we propose a new multi-granularity referring expression segmentation (MRES) task and an evaluation benchmark named RefCOCOm by manually annotating the part-level targets based on the previous benchmark RefCOCO with only object-level labels. We also construct the largest-scale visual grounding dataset which is also the first dataset to support part-level visual-textual annotations, and build a simple baseline model to accomplish the unified multi-granularity (i.e., object-level and part-level) RES task. + +Our main contributions can be summarized as follows: + +- We propose a new MRES task (as shown in Fig. 1) with corresponding benchmark RefCOCOM for evaluation, pushing classic RES task towards finer-granularity understanding of real-world scenes. +- We build a multi-grained visual grounding dataset namely MRES-32M, which to the best of our knowledge is the first grounding dataset that supports part-level vision-language annotations and also the largest-scale visual grounding dataset. +- To effectively unify both the object-level and part-level RES tasks, we propose a simple yet strong model namely UniRES, which achieves the new state-of-the-art (SOTA) performance on three object-level benchmarks for classic RES tasks and our multi-granularity RefCOCOM benchmark for the proposed MRES task. + +# 2. Related Work + +Referring Expression Segmentation. As a challenging visual grounding task the concept of RES is first proposed by [13]. Subsequent works such as [15, 16, 40, 43, 45] predominantly follow a two-step pipeline of encoding the linguistic and visual features separately and deriving the fused multimodal features from unimodal representations for mask prediction. In this way, the effectiveness of obtained multimodal representations essentially dominates the model performance, which has been continually studied in the following research [7, 22, 42, 48, 49]. Recently, there are some works [28, 36, 46] focusing on zero-shot RES task which has great application potential. In addition, a few recent works [14, 25] have been devoted to address the limitations of the existing benchmark datasets [18, 27, 41, 44] for visual grounding task. However, it's worth noting that previous works primarily focus on classic object-level RES methods and datasets, paying few attention to the vital fine-grained part-level grounding. + +Part-level Visual Perception. The increasing interest in the fine-grained understanding of objects has driven the creation of part-level annotated datasets across both specialized and general domains. Pioneering research in the former field has introduced datasets with part-level annotations, fo + +cusing on objects of particular categories like human body parts [10], animal body parts [33] and vehicle components [39]. In contrast, more general datasets that offer part annotations for a variety of common objects include Pascal-Part [4], PartNet [26], and PartImageNet [12]. Furthermore, recent work [6] introduces a finer-grained dataset with additional instance-specific part annotations for panoptic segmentation task. PACO [32] extends the available object-level datasets by including the annotations of part granularity and attribute for finer-grained segmentation and detection tasks. Moreover, VLPart [35] proposes a parsing pipeline for part segmentation and a detector that is capable of realizing the segmentation task of both open-vocabulary objects and their part regions. + +# 3. Multi-Granularity Grounding Benchmark + +# 3.1. Multi-Granularity RES Task + +The interaction capability of a multimodal agent directly depends on the visual granularity it can perceive and understand. However, in the current research landscape, multi-granularity visual grounding remains under-explored. To break the limitations wherein visual-language alignment is constrained to the object level, we intend to take a step forward towards finer-grained vision-language understanding and propose the multi-granularity RES task, which requires models' proficiency in uniformly grounding entities across both part and object levels in response to various textual references. Within the single modality, it entails an enhancement of the capacity to apprehend complex, implicit knowledge structures, e.g., the granularity delineated by textual descriptions and the advanced level of pixel-wise semantic comprehension. Across the vision-language modalities, finer-grained references also pose new challenges for the accuracy and robustness of cross-modal alignment. Besides, since previous studies are mainly restricted to object-level referring segmentation and related benchmarks for the evaluation of part-level grounding performance are unavailable, we build a corresponding benchmark named RefCOCOM for performance evaluation, which will be elaborated below. + +# 3.2. RefCOCOM Benchmark + +Manual Annotation Pipeline. To facilitate easy evaluation by community researchers without the need for any code modifications and build the multi-granularity benchmark with excellent quality, we split the existing object masks into part-level annotations through manual labeling on three subsets (i.e., validation, testA and testB) of the most prevalent RefCOCO [44] dataset. Initially, we establish a set of criteria for part masks delineation and identify the specific part categories requisite for each object. Subsequently, we employ SAM [20] to procure the preliminary segmentation results. We recruit 30 skilled annotators to refine and cap + +tioning these results through an online annotation tool developed by ourselves. To ensure the annotation quality, we conduct spot checks at three progress nodes of $20\%$ , $50\%$ , and $80\%$ of the total workload with corrective feedback. + +![](images/bfa889bcc52869eba53a52eb2296731a7efd7656f0d2aa8094b62d8634249620.jpg) +(a) + +![](images/5cb87e73551e278622d387075ca584dd1f2185a8dcaebd70d6cda09692460bb5.jpg) +(b) +Figure 2. RefCOCom benchmark statistics. (a) the number of referring expressions per parts' category in the log scale. (b) the word cloud highlights the head categories. + +Benchmark Details. Totally, we collect 70k part references and corresponding 26k masks. The average length of the references is 5.1, covering 80 object and 391 part categories, among which the number of referring expressions per parts' category and the word cloud that highlights the head categories are both presented in Fig. 2. Mixed with the original object-level annotations, our RefCOCOM has a total of 34k masks and 92k references. As an enhanced expansion of RefCOCO [44], our RefCOCOM puts forward higher requirements for referential understanding and visual perception capabilities. To better match the intention of multi-granularity unification, we leverage mean Intersection-over-Union (mIoU) as evaluation metrics. + +# 4. Multi-Granularity Grounding Dataset + +# 4.1. Data Collection Engine + +Given the intrinsic complexity of classic RES task, the associated training data necessitates extensive annotation costs across both textual and visual domains. When the granularity is promoted to part-level, the annotation intricacies are further exacerbated. We contend that the primary bottleneck hindering the emergence of open-world grounding is the limitation imposed by the current scalability of data. By leveraging robust foundation models for synergistic enhancement, we introduce an advanced data engine capable of automatically generating reliable visual grounding data. + +Multi-Grained Dense Captioner. As the large language models (LLMs) [5, 37] continue to empower the multimodal domain with profound capabilities, the current scope of large vision-language models (LVLMs) [1-3] has effectively expanded into the grounding domain. Formally, these models utilize numerical coordinates to represent bounding boxes, allowing positional information to be seamlessly integrated into the language model via word embeddings for contextual understanding. However, the research community's exploration into the part granularity remains nascent. + +![](images/75deb53c97770984745486f7627f430217d633ff649018fea9e933d3281332a0.jpg) +(a) Multi-Grained Dense Captioner + +![](images/af6fa3fe88e2cc4617bcf965ef9d6a55d037ffb2e9bf6b4ec8741412a14f48fd.jpg) + +![](images/558b5488a8c5d122f2d2f1455594f3594ed9ab18494079cc85e3d17fafebb97e.jpg) +(c) Part-Level Grounding Data Generation +Figure 3. The illustration of our data engine for building the MRES-32M dataset. (a) We start by fine-tuning an LVLM to create a capable dense captioner, which can effectively handle captioning tasks at three levels of granularity. (b) To generate object-level grounding data, we feed images and original bounding boxes into the dense captioner and a powerful segmenter to obtain the captions and masks of various objects. (c) We leverage the external knowledge from LLMs to decompose the existing object category annotations into a vocabulary set of part-level tags, which are sequentially fed into an open-vocabulary segmenter and our captioner to acquire the part-level annotations. + +The constrained training data have resulted in the fine-grained descriptive capabilities of LVLMs being confined merely to the object-level. To fully harness the open-world visual knowledge acquired from extensive pretraining on image-level and object-level data, we have devised a unified fine-tuning scheme that equips LVLMs to operate as dense captioners across all granularity levels, as depicted in Fig. 3 (a). Data for all three granularity levels are sourced from manual annotations to ensure reliability. The input comprises an image and the corresponding bounding box to be described, with bounding box coordinates normalized to integer values within the [0,999] range. For image granularity, we employ the COCO dataset [24], where all bboxes are uniformly represented as $(0,0)$ , $(999,999)$ . For object granularity, we utilize the Visual Genome dataset [21]. For part granularity, we draw upon unimodal semantic segmentation data [4, 12, 32] and employ a template in the form of PartNameX of ObjectNameY to construct dense captions. This unified multitask training approach can be synergistic across different granularity: it allows LVLMs to incorporate more comprehensive and detailed information to enhance part granularity descriptions. Concurrently, knowledge of part granularity assists LVLMs in generalizing knowledge within object interiors. + +Model-Assisted Data Generation. For the data generation of object-level visual grounding, we capitalize on largescale object detection dataset Object365 [34] to furnish highly reliable bounding boxes. Moreover, the rich category labels inherent in it ensure a comprehensive knowledge coverage. As illustrated in Fig. 3 (b), the bounding boxes will serve as visual prompts, which are independently sent into both a promptable segmenter (i.e. segment anything model + +[20]) and our dense captioner to obtain the segmentation masks and detailed semantic descriptions, respectively. + +For part-level grounding data, we propose a hierarchical annotation scheme that builds upon existing object-level annotations, as shown in Fig. 3 (c). Specifically, employing GPT-4 [29] endowed with extensive external knowledge, we decompose the objects present in a given image to generate a customized part vocabulary set. This tailored vocabulary set is then fed into an open-vocabulary segmenter [35], which yields precise part masks and bounding boxes. These bounding boxes are subsequently sent into our dense captioner to acquire corresponding detailed captions. + +Data Filtering. After completing the multi-granularity annotation of all images, we further introduce CLIP [30] for filtering. The bounding box is cropped from the original image, and then sent to the encoder together with the dense caption to measure the similarity. To ensure the consistency between the visual and linguistic annotations to a great extent, we retain box-caption pairs with similarity greater than 0.5 as the final annotation results. + +# 4.2. MRES-32M Dataset Details + +As listed in Table 2, existing datasets, such as the most commonly used benchmark dataset RefCOCO [44], have limitations in terms of small data scale and lacking part-level dense annotations. As one of the pioneering works for finer-granularity unimodal visual comprehension, PACO [32] only contains semantic tags of objects or parts, without informative descriptions and visual context encompassed as referring expressions. Summarily, we compare the proposed MRES-32M with existing datasets and list some unique and significant properties of our dataset in Table + +Table 2. Comparisons with previous object-level visual grounding datasets and part-level segmentation datasets. # denotes the specific number, where Cats and Avg Len denote the object/part categories and the average length of referring expressions. “-” denotes the corresponding part masks or captions are unavailable. + +
Dataset#Images#Objs#Parts#Cats#Avg Len
Object-Level Visual Grouding
ReferIt [18]20K97K-238/-3.2
RefCOCO [44]20K50K-80/-3.6
RefCOCO+ [44]20K49K-80/-3.5
RefCOCOg [27]26K54K-80/-8.4
GRES [25]20K60K-80/-3.7
Part-Level Segmentation & Detection
PartsIN [12]24K24K112K158/609-
PascalPart [4]19K40K363K20/193-
PACO [32]20K260K641K75/456-
Multi-Grained Visual Grounding
MRES-32M (Ours)1M15.3M16.9M365/22994.6
+ +2. Besides, we have also provided a few examples in our MRES-32M dataset in the appendix. + +Unified Multi-Granularity. In comparison to the grounding counterparts, our MRES-32M is the first visual grounding dataset covering both part and object granularity. In comparison to the part-level segmentation counterparts, our MRES-32M provides informative and unique fine-grained description for each part mask. + +More Diversified Categories. Our MRES-32M is composed of 365 object categories and an associated 2,299 part categories. Compared with existing datasets, it covers a wider range of multimodal knowledge and is an important step towards open-world understanding. + +Breakable Data Scales. To the best of our knowledge, MRES-32M is the largest dataset in the current grounding research community. In terms of the number of images and object instances, it surpasses the largest existing visual grounding dataset RefCOCOg [27] by factors of 38 and 283, respectively. Meanwhile, it encompasses part instance counts that exceed the largest existing part semantic segmentation dataset [32] by 58 times. + +More Complex References. Benefiting from our LVLM-based dense captioner, the reference of MRES-32M could be more fully combined with the visual context for entity (i.e., part and object) description. Without sticking to a specific template, the relationships and attributes of entities could be highlighted in free natural language expressions. + +# 5. Multi-Granularity RES Model + +Next, we describe the proposed multi-granularity RES model UniRES for the unified MRES task with the referring targets at both object and part granularity. Since our original intention is to establish a simple and easy-to-follow baseline model for the proposed multi-granularity RES task, + +the structure of our model UniRES is designed to be simple and clean. As shown in Fig. 4, UniRES has three major components, which will be illustrated below. + +Vision-Language Backbone. Taking account of the required attributes of both strong ability to capture the vision-language feature representations and promising scalability, we leverage the CLIP pre-trained weights of CLIP model [31], which learns transferable visual and linguistic concepts from vast image-text pairs, and adopt it as the backbone for our referring segmentation framework. The respective image and text encoders (i.e. Vision Transformer (ViT) [8] and Transformer [38] respectively) from CLIP are utilized to effectively extract visual and linguistic features. + +Query-based Grouping Design. To effectively accomplish the proposed MRES task, it is essential to exploit both low-level local and high-level global visual features. In order to enhance the local-global visual representations of CLIP backbone without introducing much additional parameters or altering model's structure, we have incorporated 64 and 8 learnable tokens (set empirically) into the first and middle layers of CLIP visual backbone. These learnable tokens traverse the first and second halves of the visual backbone. We expect that the ViT's internal self-attention mechanism implicitly serves as a manner to perform visual grouping, obtaining representative group tokens that capture the low-level local and high-level global features simultaneously. Based on the fact that local features spread more fragmented, the number of appended low-level group tokens is greater than that of high-level counterparts. Then, group tokens from both levels are fed into a language-guided region filter (LRF) to select the language-related visual features by cross-attention mechanism, which is followed by concatenation to fuse these expression-related visual group tokens for subsequent vision-language decoding. + +Two-Stage V-L Decoder. Now that the visual and textual feature representations from backbone, and the expression-related group tokens with two different levels are obtained, the two-stage mask decoder which comprises stacked Transformer layers is utilized to generate the segmentation masks. Specifically, the first-stage V-L decoder takes the extracted visual and textual features as input and generates first-stage fused multimodal representations. Subsequently, these multimodal features are further integrated with the grouped expression-related region features of the two semantic levels (i.e., low-level & high-level) to realize further feature enhancement, which is followed by a linear projection layer to obtain the final segmentation masks. + +# 6. Experimental Results + +To evaluate the effectiveness of our method, comprehensive experiments are conducted on three classic RES datasets (i.e., RefCOCO [44], RefCOCO+ [44], RefCOCOg [27]) and our RefCOCom benchmark for multi-granularity RES. + +![](images/b8862dd02c5505d6445148adf61783e93b874d375670eb69ca13edf10abdb924.jpg) +Figure 4. The architecture of our UniRES model as a simple baseline for the MRES task. UniRES mainly comprises three parts: the visual and textual backbone for feature extraction, the pixel grouping design for aggregating the low-level and high-level features, and the cascaded two-stage vision-language decoder for multimodal feature fusion and the generation of segmentation masks. + +# 6.1. Datasets + +The details about the three benchmark datasets for classic RES task can be found in the appendix. + +# 6.2. Experimental Setup + +Implementation Details. The implementation details are provided in the appendix. + +Evaluation metrics. We adopt mean Intersection-over-Union (mIoU) and overall Intersection-over-Union (oIoU) as evaluation metrics. The mIoU measures the ratio between the intersection area and the union area of the predictions and labels among the test samples, while the oIoU calculates the total intersection area over total union area. + +# 6.3. Main Results + +# 6.3.1 Multi-Granularity MRES Task + +Comparison with SOTA Methods. To evaluate the multi-granularity grounding performance of our UniRES as the baseline for the proposed MRES task and previous RES methods, we further conduct experimental comparison on our newly built RefCOCOM benchmark dataset. As presented in Table 3, the classic RES methods including four specialist models (i.e., SeqTR [47], CRIS [40], LAVT [42]) and two generalist models (e.g., X-Decoder [48] and SEEM [49]) are incorporated. For fair comparisons, we re-implement these SOTA methods and report their performance on our RefCOCOM. It is clear that either the specialist models for classic RES task or the powerful generalist models perform poorly on RefCOCOM, which requires the crucial skills of both part-level and object-level referring segmentation. Benefiting from our MRES-32M dataset, our UniRES can better master the part-level RES skills and handle the multi-granularity RES task, achieving considerably higher segmentation accuracy. Due to the significantly higher difficulty of part-level RES (i.e. multi-granularity RES) compared to classic RES, the absolute value of segmentation accuracy is correspondingly lower. This further emphasizes the importance of researching finer-grained part grounding where previous SOTA methods have fallen short. + +Table 3. Comparison with previous SOTA methods on our Ref-COCOM benchmark in terms of mIoU. Part and Obj & Part denote part-only and multi-grained evaluation settings of our MRES task. + +
MethodsvaltestAtestB
PartObj & PartPartObj & PartPartObj & Part
Specialists
SeqTR [47]13.928.212.122.818.134.7
CRIS [40]10.625.410.121.212.930.0
LAVT [42]15.329.913.224.418.735.5
Generalists
X-Decoder [48]16.229.513.623.620.333.8
SEEM [49]16.129.413.623.420.433.9
UniRES (Ours)19.634.316.427.825.241.7
+ +Qualitative Analysis. We also conduct visual comparisons of CRIS [40], LAVT [42] and our UniRES on the MRES task, which is provided in appendix. + +# 6.3.2 Classic Object-Level RES Task + +Comparison with SOTA Methods. To validate the superiority of our MRES-32M dataset and model UniRES, our framework is fairly evaluated against previous SOTA methods on RefCOCO [44], RefCOCO+ [44] and RefCOCOg [27] under both zero-shot and fine-tuning settings. As presented in Table 4, our method greatly outperforms previous methods in terms of segmentation accuracy across all the benchmark datasets. Via directly zero-shot transferring to the downstream classic RES task after pre-training on our MRES-32M dataset, Our UniRES model achieves a leading zero-shot segmentation accuracy (i.e., approximately $71\%$ ) without any fine-tuning to adapt to downstream task's data, which is significantly higher than all the recently proposed zero-shot RES methods by a large margin (i.e., $\uparrow 30 - 40\%$ mIoU). At the same time, it is worth noting that our zero-shot segmentation performance is already better than many of previous RES methods under fine-tuning setting (e.g., CRIS [40] and ReSTR [19]), validating the potential of our MRES-32M dataset and UniRES model. Furthermore, after fine-tuning on the classic RES datasets, our UniRES greatly surpasses previous methods no matter the specialist models (e.g., LAVT [42] and LISA [22]) for classic RES or the generalist models (e.g., X-Decoder [48] and SEEM [49]). + +Table 4. Comparisons with the state-of-the-art approaches on previous three classic RES benchmark datasets under both the zero-shot and fine-tuning settings. “-” denotes that the result is not provided. + +
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Zero-Shot Methods
mIoURegion token [46] (CVPR-23)23.422.124.624.522.625.427.627.3
Cropping [46] (CVPR-23)24.822.625.726.324.126.531.930.9
Global-Local CLIP [46] (CVPR-23)26.224.926.627.825.627.833.533.7
SAM-CLIP [28] (arXiv-23)26.325.826.425.728.026.838.838.9
Ref-Diff [28] (arXiv-23)37.238.437.237.340.533.044.044.5
TAS [36] (arXiv-23)39.841.136.243.649.136.546.646.8
UniRES (Ours)71.274.866.059.966.751.462.363.2
Fine-Tune Methods
oIoUEFNet [9] (CVPR-21)62.865.759.751.555.243.0--
LTS [17] (CVPR-21)65.467.863.154.258.348.054.454.3
ReSTR [19] (CVPR-22)67.269.364.555.860.448.3--
ReLA [25] (CVPR-23)73.876.570.266.071.057.765.066.0
X-Decoder [48] (CVPR-23)------64.6-
SEEM [49] (arXiv-23)------65.7-
LISA [22] (arXiv-23)74.979.172.365.170.858.167.970.6
UniRES (Ours)77.480.974.769.476.161.469.071.7
mIoUVLT [7] (ICCV-21)65.768.362.755.559.249.453.056.7
RefTr [23] (NeurIPS-21)74.376.870.966.870.659.466.667.4
SeqTR [47] (ECCV-22)71.773.369.863.066.759.065.065.7
CRIS [40] (CVPR-22)70.573.266.162.368.153.759.960.4
LAVT [42] (CVPR-22)74.576.970.965.871.059.263.363.6
UniRES (Ours)79.281.676.673.078.165.871.773.2
+ +# 6.4. Ablation Studies + +We conduct ablation experiments on our RefCOCOM validation set. The tables below involves three evaluation settings on our RefCOCOM, using only objects as RES targets, using only parts as RES targets, and a mixed setting that combines both of the above two granularity levels. + +# 6.4.1 Ablation Study on MRES-32M Dataset + +Data Granularity. We first probe into the effect of pretraining data's granularity with $50\%$ of our MRES-32M dataset. As shown in Table 5, the baseline without pretraining on MRES-32M dataset obtains $75.2\%$ , $15.8\%$ and $32.0\%$ mIoU separately on our RefCOCOM validation set under three granularity levels. Either introducing the object-level or part-level data from MRES-32M dataset for pretraining consistently results in an considerable accuracy increase across different granularity settings. In fact, pretraining on data at a certain granularity level to improve performance on corresponding benchmark aligns with common sense. However, it is noteworthy that incorporating part-level data into training also enhances model performance on the only object-level RefCOCOM. This underscores the significance of part-level understanding, as emphasized in the Sec. 1, where grasping the nuances of parts can yield benefits for object-level grounding. Besides, by + +jointly incorporating the training samples of both granularity, our method attains $2.8\%$ improvement against baseline under all the granularity settings on RefCOCOM, which fully demonstrate the benefit of exploiting our MRES-32M dataset for both object-level and part-level RES tasks. + +Table 5. Ablation study on the granularity of pre-training data in our proposed MRES-32M. Object and Part denote the introduction of object-level and part-level data for pre-training. + +
ObjectPartRefCOCOM
Object-OnlyPart-OnlyObject&Part
75.215.830.5
✓77.515.931.1
✓75.918.432.6
✓✓78.018.633.3
+ +Table 6. Ablation study on the data scale of MRES-32M dataset. + +
RatiosRefCOCOm
Object-OnlyPart-OnlyObject & Part
0%75.215.830.5
20%76.817.332.0
50%78.018.633.3
100%79.219.634.3
+ +Data Scale. Next, we investigate the effect of different percentages of training samples in our MRES-32M dataset. The results are presented in Table 6. It is obvious in Table 6 that the model performance for our MRES tasks is + +consistently improved with more and more employed training samples, which implicitly validates the high quality of our built dataset. As the ratios of introduced training samples continue to rise, there's no sign of diminishing gains in model performance, suggesting that our framework has significant potential with continually scaled up training data. + +Table 7. Ablation study on the effectiveness of our MRES-32M dataset on RES SOTA methods. + +
MethodsMRES-32MRefCOCOM
Object-OnlyPart-OnlyObject & Part
CRIS70.510.625.4
CRIS✓73.115.529.7
LAVT74.515.329.9
LAVT✓75.719.333.2
+ +Dataset Necessity. To prove the necessity and effectiveness of our MRES-32M dataset for the proposed MRES task, we take previous SOTA methods (i.e., CRIS [40] and LAVT [42]) on classic RES and compare the segmentation accuracy of the same models with or without pretraining on our MRES-32M. As shown in Table 7, the original CRIS and LAVT can already well handle the classic RES task with only object-level grounding skills, but it performs poorly on the multi-grained RefCOCOM benchmark, which requires part-level grounding capability. In contrast, pre-training on our MRES-32M consistently leads to great performance gains no matter on object-only, part-only or the multi-granularity RES tasks, because our high-quality MRES-32M dataset can effectively enable the model to handle the part-level visual grounding task and enhances the original object-level grounding capability. + +# 6.4.2 Ablation Study on UniRES Model + +Table 8. Ablation study on the query-based grouping design. + +
High-LevelLow-LevelRefCOCOM
Object-OnlyPart-OnlyObject&Part
✓74.414.929.6
74.915.230.0
✓74.915.430.1
✓✓75.215.830.5
+ +Besides, we additionally conduct ablation studies on the structural design of our model UniRES. As presented in Table 8, removing the simple but effective grouping design at two different levels (this simultaneously leads to discarding the second-stage decoder) will result in a considerable deterioration in model performance. Since the sequentially appended pixel grouping tokens are introduced to effectively capture the high-level and low-level visual features for further visual feature enhancement, losing any type of these group tokens will lead to a decrease in the model's segmentation accuracy on both object-level and part-level RES tasks. In order to verify that the high-level and low-level pixel grouping tokens can capture the clustering features of the corresponding level, we also visualize the group tokens + +of the two levels (before being sent to LRF) for qualitative analysis, where the different colors represent the various areas of the same clustering group. The visualization results in Fig. 5 affirms that our introduced low-level and high-level group tokens respectively capture the part-level local features and aggregate the object-level global features with stronger semantics through the pixel grouping process, which aligns with our intentions for the grouping design. + +![](images/5a96cf496155f0279018f07a66691f1ed4e21b61ba6b1b3fa0acd0ebc831ea43.jpg) +(a) Image + +![](images/dff899dcdbbb12475949cb8646a90136655ef9874e5a8e1ed11b84466837ec16.jpg) +(b) Low-Level + +![](images/bac86dba758eb4ea3a68df8b2a6c2f0ccc213c34ff224fb50d9e5c3eb9d9050f.jpg) +(c) High-Level +Figure 5. Qualitative analysis for ablation study on the object-level and part-level grouping design in our model structure. (a) the input image. (b) low-level group tokens. (c) high-level group tokens. + +# 7. Conclusion and Broader Impact + +In this paper, we move beyond previous works that focused solely on object-level visual grounding tasks and take a step further to finer-grained part-level RES. We put forward a new multi-granularity referring expression segmentation task and establish an evaluation benchmark named RefCOCOM by manual annotation. To advance the visual grounding at both object and part levels towards finer-grained vision-language understanding, we build the largest visual grounding dataset MG-32M to date, which is also the first dataset to provide part-level vision-language annotations. Furthermore, we have developed a simple yet strong multi-grained referring segmentation model called UniRES. As a baseline for our newly proposed MRES task, UniRES achieves the new state-of-the-art performance on both our RefCOCOM for MRES task and three classic RES datasets. We plan to release our RefCOCOM benchmark, the MG32M dataset, and the UniRES model to the public, aspiring to foster future research in fine-grained visual grounding tasks and to inspire new research in this direction. + +# Acknowledgement + +We thank Yepeng Tang for the helpful discussions on this work, all the Image & Video Analysis Group (IVA)'s members in CASIA for the technical support, and all the insightful reviewers for the helpful suggestions. This work was supported by the National Science and Technology Major Project (No.2022ZD0118801), National Natural Science Foundation of China (U21B2043, 62206279). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3 +[2] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. +[3] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3 +[4] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 2, 3, 4, 5 +[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3 +[6] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 3 +[7] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2, 7 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[9] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 7 +[10] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 932-940, 2017. 2, 3 +[11] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 2 + +[12] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 2, 3, 4, 5 +[13] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2 +[14] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4067-4077, 2023. 2 +[15] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 2 +[16] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 2 +[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 7 +[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 5 +[19] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 2, 6, 7 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3, 4 +[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 4 +[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 6, 7 +[23] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances + +in neural information processing systems, 34:19652-19664, 2021.7 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 4 +[25] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 5, 7 +[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2, 3 +[27] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2, 5, 6 +[28] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 2, 7 +[29] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 4 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[32] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 2, 3, 4, 5 +[33] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 2, 3 +[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4 + +[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3, 4 +[36] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial-aware zero-shot referring image segmentation. arXiv preprint arXiv:2310.18049, 2023. 2, 7 +[37] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 3 +[40] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2, 6, 7, 8 +[41] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 2 +[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 2, 6, 7, 8 +[43] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 2 +[44] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 2,3,4,5,6 +[45] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 2 +[46] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 2, 7 +[47] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, + +and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 598-615. Springer, 2022. 6, 7 +[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2, 6, 7 +[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 6, 7 \ No newline at end of file diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/images.zip b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..203fe6d5a9f4cdf03829cebad9b4ecd5a5aa8375 --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:401dfdf27394a85a7b1962bcb9da7c23da678846922a1ba380d8de45dd378d44 +size 579698 diff --git a/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/layout.json b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..99f427781787270de3cc3c61c8b206a10f110800 --- /dev/null +++ b/2024/Unveiling Parts Beyond Objects_ Towards Finer-Granularity Referring Expression Segmentation/layout.json @@ -0,0 +1,8084 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 140 + ], + "type": "text", + "content": "Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": "Wenxuan Wang" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Tongtian Yue" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Yisi Zhang" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Longteng Guo" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Xingjian He" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Xinlong Wang" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "text", + "content": " Jing Liu" + }, + { + "bbox": [ + 85, + 159, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 190, + 457, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 190, + 457, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 190, + 457, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 190, + 457, + 203 + ], + "type": "text", + "content": " Institute of Automation, Chinese Academy of Sciences (CASIA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 203, + 507, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 203, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 85, + 203, + 507, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 203, + 507, + 217 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 217, + 423, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 217, + 423, + 232 + ], + "spans": [ + { + "bbox": [ + 170, + 217, + 423, + 232 + ], + "type": "text", + "content": "3 Beijing Academy of Artificial Intelligence (BAAI)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 232, + 433, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 232, + 433, + 245 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 433, + 245 + ], + "type": "text", + "content": "4 University of Science and Technology Beijing (USTB)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 247, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 247, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 482, + 259 + ], + "type": "text", + "content": "{wangwenxuan2023@ia.ac.cn, wangxinlong@baai.ac.cn, jliu@nlpr.ia.ac.cn}" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 97, + 294, + 151, + 376 + ], + "blocks": [ + { + "bbox": [ + 112, + 281, + 137, + 289 + ], + "lines": [ + { + "bbox": [ + 112, + 281, + 137, + 289 + ], + "spans": [ + { + "bbox": [ + 112, + 281, + 137, + 289 + ], + "type": "text", + "content": "Images" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 97, + 294, + 151, + 376 + ], + "lines": [ + { + "bbox": [ + 97, + 294, + 151, + 376 + ], + "spans": [ + { + "bbox": [ + 97, + 294, + 151, + 376 + ], + "type": "image", + "image_path": "4215409bdd47d926dcce2e13dea41ea07440dc14afad1ff4d2d55d72c1b1509b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 159, + 294, + 212, + 333 + ], + "blocks": [ + { + "bbox": [ + 166, + 279, + 202, + 293 + ], + "lines": [ + { + "bbox": [ + 166, + 279, + 202, + 293 + ], + "spans": [ + { + "bbox": [ + 166, + 279, + 202, + 293 + ], + "type": "text", + "content": "(a) \"Standing Lady on left\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 294, + 212, + 333 + ], + "lines": [ + { + "bbox": [ + 159, + 294, + 212, + 333 + ], + "spans": [ + { + "bbox": [ + 159, + 294, + 212, + 333 + ], + "type": "image", + "image_path": "bbc3328c82745a4b6a22d70666a91c4718bf26808b4d2704e2262755b5fa6a3a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 159, + 335, + 212, + 373 + ], + "blocks": [ + { + "bbox": [ + 159, + 335, + 212, + 373 + ], + "lines": [ + { + "bbox": [ + 159, + 335, + 212, + 373 + ], + "spans": [ + { + "bbox": [ + 159, + 335, + 212, + 373 + ], + "type": "image", + "image_path": "eeb8c721ef56e1204bb01f690cca9d8ffb87371a4f4cd036ffc32d82ee7de2b8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 376, + 207, + 391 + ], + "lines": [ + { + "bbox": [ + 164, + 376, + 207, + 391 + ], + "spans": [ + { + "bbox": [ + 164, + 376, + 207, + 391 + ], + "type": "text", + "content": "(a) \"Motorcycle on the left\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 400, + 546, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 546, + 434 + ], + "type": "text", + "content": "Figure 1. Classic Referring Expression Segmentation (RES) only supports expressions that indicate a single target object, e.g., (a). Compared with classic RES, the proposed Multi-Granularity Referring Expression Segmentation (MRES) task supports expressions indicating the specific part-level regions of target objects, e.g., part-level expressions like (b)-(e) from our newly built RefCOCom benchmark." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 219, + 294, + 272, + 332 + ], + "blocks": [ + { + "bbox": [ + 166, + 271, + 202, + 278 + ], + "lines": [ + { + "bbox": [ + 166, + 271, + 202, + 278 + ], + "spans": [ + { + "bbox": [ + 166, + 271, + 202, + 278 + ], + "type": "text", + "content": "Classic RES" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 279, + 263, + 293 + ], + "lines": [ + { + "bbox": [ + 227, + 279, + 263, + 293 + ], + "spans": [ + { + "bbox": [ + 227, + 279, + 263, + 293 + ], + "type": "text", + "content": "(a) \"Standing Lady on left" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 294, + 272, + 332 + ], + "lines": [ + { + "bbox": [ + 219, + 294, + 272, + 332 + ], + "spans": [ + { + "bbox": [ + 219, + 294, + 272, + 332 + ], + "type": "image", + "image_path": "c71f2b66003c1c2aa26d45a5e887e33def7e3acc9f193c7d1e27746bae22999b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 219, + 335, + 272, + 373 + ], + "blocks": [ + { + "bbox": [ + 219, + 335, + 272, + 373 + ], + "lines": [ + { + "bbox": [ + 219, + 335, + 272, + 373 + ], + "spans": [ + { + "bbox": [ + 219, + 335, + 272, + 373 + ], + "type": "image", + "image_path": "72989edfd0c2cb4a7208cb18677905e0ff4027889f113e892fe23ef8428fe33a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 375, + 268, + 389 + ], + "lines": [ + { + "bbox": [ + 225, + 375, + 268, + 389 + ], + "spans": [ + { + "bbox": [ + 225, + 375, + 268, + 389 + ], + "type": "text", + "content": "(a) \"Motorcycle on the left\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 273, + 294, + 326, + 332 + ], + "blocks": [ + { + "bbox": [ + 227, + 271, + 266, + 278 + ], + "lines": [ + { + "bbox": [ + 227, + 271, + 266, + 278 + ], + "spans": [ + { + "bbox": [ + 227, + 271, + 266, + 278 + ], + "type": "text", + "content": "MRES (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 277, + 279, + 321, + 293 + ], + "lines": [ + { + "bbox": [ + 277, + 279, + 321, + 293 + ], + "spans": [ + { + "bbox": [ + 277, + 279, + 321, + 293 + ], + "type": "text", + "content": "(b) \"Leg of lady walking on left" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 273, + 294, + 326, + 332 + ], + "lines": [ + { + "bbox": [ + 273, + 294, + 326, + 332 + ], + "spans": [ + { + "bbox": [ + 273, + 294, + 326, + 332 + ], + "type": "image", + "image_path": "e8dd8a3911eaa263c703b098f78689f6a0c1eb7dfd56c5e1b24f5ab8e36ef82a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 273, + 335, + 325, + 373 + ], + "blocks": [ + { + "bbox": [ + 273, + 335, + 325, + 373 + ], + "lines": [ + { + "bbox": [ + 273, + 335, + 325, + 373 + ], + "spans": [ + { + "bbox": [ + 273, + 335, + 325, + 373 + ], + "type": "image", + "image_path": "b536f2226f064f562556624889d0051583eff2cd58a1d850bf4788b64be66564.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 375, + 321, + 397 + ], + "lines": [ + { + "bbox": [ + 277, + 375, + 321, + 397 + ], + "spans": [ + { + "bbox": [ + 277, + 375, + 321, + 397 + ], + "type": "text", + "content": "(b) \"Silvery body of motorcycle on the left\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 327, + 294, + 380, + 333 + ], + "blocks": [ + { + "bbox": [ + 333, + 279, + 373, + 293 + ], + "lines": [ + { + "bbox": [ + 333, + 279, + 373, + 293 + ], + "spans": [ + { + "bbox": [ + 333, + 279, + 373, + 293 + ], + "type": "text", + "content": "(c) \"Middle girl arm\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 294, + 380, + 333 + ], + "lines": [ + { + "bbox": [ + 327, + 294, + 380, + 333 + ], + "spans": [ + { + "bbox": [ + 327, + 294, + 380, + 333 + ], + "type": "image", + "image_path": "f38fde52231ae2a6bc845c2b186b392753629f423c85c1a8c9a8c1bb1b9bef07.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 327, + 335, + 380, + 373 + ], + "blocks": [ + { + "bbox": [ + 327, + 335, + 380, + 373 + ], + "lines": [ + { + "bbox": [ + 327, + 335, + 380, + 373 + ], + "spans": [ + { + "bbox": [ + 327, + 335, + 380, + 373 + ], + "type": "image", + "image_path": "b7c242ff774c85e64aae5243310e2657560ed6af24874b082c12481aa63f89aa.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 375, + 367, + 388 + ], + "lines": [ + { + "bbox": [ + 337, + 375, + 367, + 388 + ], + "spans": [ + { + "bbox": [ + 337, + 375, + 367, + 388 + ], + "type": "text", + "content": "(c) \"Center motorcycle" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 388, + 361, + 396 + ], + "lines": [ + { + "bbox": [ + 337, + 388, + 361, + 396 + ], + "spans": [ + { + "bbox": [ + 337, + 388, + 361, + 396 + ], + "type": "text", + "content": "wheel\"" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 381, + 294, + 433, + 333 + ], + "blocks": [ + { + "bbox": [ + 386, + 279, + 424, + 293 + ], + "lines": [ + { + "bbox": [ + 386, + 279, + 424, + 293 + ], + "spans": [ + { + "bbox": [ + 386, + 279, + 424, + 293 + ], + "type": "text", + "content": "(d) \"Head of center player\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 294, + 433, + 333 + ], + "lines": [ + { + "bbox": [ + 381, + 294, + 433, + 333 + ], + "spans": [ + { + "bbox": [ + 381, + 294, + 433, + 333 + ], + "type": "image", + "image_path": "e0e2bfe02d8fc8715b66e988b6464860be8adccc37fcf02f137a4c2d87cafda2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 381, + 335, + 433, + 373 + ], + "blocks": [ + { + "bbox": [ + 381, + 335, + 433, + 373 + ], + "lines": [ + { + "bbox": [ + 381, + 335, + 433, + 373 + ], + "spans": [ + { + "bbox": [ + 381, + 335, + 433, + 373 + ], + "type": "image", + "image_path": "d545c6fdedf3b74e6c7cb0dcac3cc2192f62b05d43d55e91fcb7e86b48af7456.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 376, + 429, + 390 + ], + "lines": [ + { + "bbox": [ + 384, + 376, + 429, + 390 + ], + "spans": [ + { + "bbox": [ + 384, + 376, + 429, + 390 + ], + "type": "text", + "content": "(d) \"Front motor headlight\"" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 435, + 294, + 487, + 333 + ], + "blocks": [ + { + "bbox": [ + 436, + 279, + 484, + 293 + ], + "lines": [ + { + "bbox": [ + 436, + 279, + 484, + 293 + ], + "spans": [ + { + "bbox": [ + 436, + 279, + 484, + 293 + ], + "type": "text", + "content": "(e) \"Center player torso in white\"" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 294, + 487, + 333 + ], + "lines": [ + { + "bbox": [ + 435, + 294, + 487, + 333 + ], + "spans": [ + { + "bbox": [ + 435, + 294, + 487, + 333 + ], + "type": "image", + "image_path": "5ac4e97f1a4cfe3f3f280283bbe4cb9a29004bb5553e71403029d026807f8179.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 435, + 335, + 487, + 373 + ], + "blocks": [ + { + "bbox": [ + 435, + 335, + 487, + 373 + ], + "lines": [ + { + "bbox": [ + 435, + 335, + 487, + 373 + ], + "spans": [ + { + "bbox": [ + 435, + 335, + 487, + 373 + ], + "type": "image", + "image_path": "0421597dd1692e22de9fe0803a7b084372ea284c70d72b9682643f48b9b52958.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 376, + 481, + 398 + ], + "lines": [ + { + "bbox": [ + 438, + 376, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 438, + 376, + 481, + 398 + ], + "type": "text", + "content": "(e) \"Transparent window of front motor\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 143, + 444, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 444, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 444, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 459, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 288, + 685 + ], + "type": "text", + "content": "Referring expression segmentation (RES) aims at segmenting the foreground masks of the entities that match the descriptive natural language expression. Previous datasets and methods for classic RES task heavily rely on the prior assumption that one expression must refer to object-level targets. In this paper, we take a step further to finer-grained part-level RES task. To promote the object-level RES task towards finer-grained vision-language understanding, we put forward a new multi-granularity referring expression segmentation (MRES) task and construct an evaluation benchmark called RefCOCom by manual annotations. By employing our automatic model-assisted data engine, we build the largest visual grounding dataset namely MRES-32M, which comprises over 32.2M high-quality masks and captions on the provided 1M images. Besides, a simple yet strong model named UniRES is designed to accomplish the unified object-level and part-level grounding task. Extensive experiments on our RefCOCom for MRES and three datasets (i.e., RefCOCO(+/g)) for classic RES task demon" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 445, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 545, + 504 + ], + "type": "text", + "content": "strate the superiority of our method over previous state-of-the-art methods. To foster future research into fine-grained visual grounding, our benchmark RefCOCOM, the MRES-32M dataset and model UniRES will be publicly available at https://github.com/Rubics-Xuan/MRES." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 526, + 386, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 386, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 386, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "type": "text", + "content": "As one of the most challenging tasks in vision-language understanding, referring expression segmentation (RES) aims to locate specific regions at the pixel level based on a descriptive language expression. Compared to traditional visual segmentation tasks that focus on images or videos alone, RES poses greater difficulties and challenges due to the necessity of strong comprehension across modalities, but it can simultaneously alleviate the problem of predefined categories in conventional object detection or segmentation. With the real-world scene often requiring diversity in target identification, RES task holds vast potential for applications, e.g., language-based human-object interaction and interactive image editing." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "Since the concept of RES task was initially proposed in" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 11, + 491, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 11, + 491, + 45 + ], + "spans": [ + { + "bbox": [ + 149, + 11, + 491, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "12998" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 116, + 286, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "type": "text", + "content": "Table 1. Comparison among different referring expression datasets, including ReferIt [18], RefCOCO(+/g) [27, 44], PhraseCut [41], and our proposed MRES-32M. Part-Level: expressions that specifies various parts of the target object in the given image." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 116, + 286, + 177 + ], + "lines": [ + { + "bbox": [ + 48, + 116, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 286, + 177 + ], + "type": "table", + "html": "
ReferItRefCOCO(+/g)PhraseCutMRES-32M
Image SourceCLEF [11]COCO [24]VG [21]Object365 [34]
Object-Level✓✓✓✓
Part-Level×××✓
Expression Typefreefreetemplatedfree
", + "image_path": "8871fc25628f00c85238b4ba6bef3bc5043fe642efabf6cb2214fba2003c7719.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 388 + ], + "type": "text", + "content": "[13], various multimodal frameworks such as [13, 19, 40, 42] towards precise RES have been proposed to deal with the challenging feature extraction and alignment problems between visual and linguistic modalities. However, current works are limited to the scope of object-level grounding. As shown in Table 1, the RefCOCO dataset [44] stands as one of the most widely used grounding benchmarks by far, basically containing only object-level visual and textual annotations. It does not take into account the part-level grounding task, which is crucial for future multimodal models to act as intelligent agents to realize the fine-grained perception of real world. The existing grounding datasets couldn't support the training and evaluation of such advanced capabilities. Thus, exploring how to transcend the current object-level constraints and delve into finer-grained part-level grounding, is highly meaningful and worthy of in-depth study, which is the primary focus of this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 389, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 286, + 700 + ], + "type": "text", + "content": "In fact, prior to this work, a few works along another research line have made significant strides towards more fine-grained visual understanding. The demand for finer-grained visual understanding of objects arouses research community into constructing high-quality part-level dataset. Specifically, recent years have witnessed the introduction of various datasets that provide fine-grained part-level masks and bounding boxes annotations for either objects of specific categories [10, 33, 39] or general categories [4, 12, 26, 32]. However, these datasets essentially correspond to unimodal (i.e., visual) downstream tasks, lacking a deep connection between fine-grained part-level masks and rich textual descriptions. To our knowledge, few of previous studies have established this connection. Consequently, there appears to be a limited availability of fine-grained, large-scale vision-language datasets that facilitate part-level cross-modality understanding, which is necessary in terms of two aspects. Firstly, when describing an object, people always naturally gravitate towards detailing part-level local features, underscoring the indispensable need for multimodal agents to understand the part granularity. Secondly, a fine-grained understanding at the part level can positively promote the perception of object-level targets, especially under extreme conditions such as occlusion or deformation, which will consistently propel advancements in widely focused object-level tasks like classic visual grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Therefore, in this work, we attempt to fill this impor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "tant blank space that has been neglected before and move towards finer-grained part-level RES task. Specifically, to push towards finer-grained vision-language understanding, we propose a new multi-granularity referring expression segmentation (MRES) task and an evaluation benchmark named RefCOCOm by manually annotating the part-level targets based on the previous benchmark RefCOCO with only object-level labels. We also construct the largest-scale visual grounding dataset which is also the first dataset to support part-level visual-textual annotations, and build a simple baseline model to accomplish the unified multi-granularity (i.e., object-level and part-level) RES task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 217, + 539, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 539, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 539, + 227 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 228, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 228, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 275 + ], + "type": "text", + "content": "- We propose a new MRES task (as shown in Fig. 1) with corresponding benchmark RefCOCOM for evaluation, pushing classic RES task towards finer-granularity understanding of real-world scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 276, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 335 + ], + "type": "text", + "content": "- We build a multi-grained visual grounding dataset namely MRES-32M, which to the best of our knowledge is the first grounding dataset that supports part-level vision-language annotations and also the largest-scale visual grounding dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 335, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 407 + ], + "type": "text", + "content": "- To effectively unify both the object-level and part-level RES tasks, we propose a simple yet strong model namely UniRES, which achieves the new state-of-the-art (SOTA) performance on three object-level benchmarks for classic RES tasks and our multi-granularity RefCOCOM benchmark for the proposed MRES task." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 653 + ], + "type": "text", + "content": "Referring Expression Segmentation. As a challenging visual grounding task the concept of RES is first proposed by [13]. Subsequent works such as [15, 16, 40, 43, 45] predominantly follow a two-step pipeline of encoding the linguistic and visual features separately and deriving the fused multimodal features from unimodal representations for mask prediction. In this way, the effectiveness of obtained multimodal representations essentially dominates the model performance, which has been continually studied in the following research [7, 22, 42, 48, 49]. Recently, there are some works [28, 36, 46] focusing on zero-shot RES task which has great application potential. In addition, a few recent works [14, 25] have been devoted to address the limitations of the existing benchmark datasets [18, 27, 41, 44] for visual grounding task. However, it's worth noting that previous works primarily focus on classic object-level RES methods and datasets, paying few attention to the vital fine-grained part-level grounding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Part-level Visual Perception. The increasing interest in the fine-grained understanding of objects has driven the creation of part-level annotated datasets across both specialized and general domains. Pioneering research in the former field has introduced datasets with part-level annotations, fo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "12999" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "cusing on objects of particular categories like human body parts [10], animal body parts [33] and vehicle components [39]. In contrast, more general datasets that offer part annotations for a variety of common objects include Pascal-Part [4], PartNet [26], and PartImageNet [12]. Furthermore, recent work [6] introduces a finer-grained dataset with additional instance-specific part annotations for panoptic segmentation task. PACO [32] extends the available object-level datasets by including the annotations of part granularity and attribute for finer-grained segmentation and detection tasks. Moreover, VLPart [35] proposes a parsing pipeline for part segmentation and a detector that is capable of realizing the segmentation task of both open-vocabulary objects and their part regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 251, + 280, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 280, + 266 + ], + "type": "text", + "content": "3. Multi-Granularity Grounding Benchmark" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 271, + 206, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 206, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 206, + 285 + ], + "type": "text", + "content": "3.1. Multi-Granularity RES Task" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 555 + ], + "type": "text", + "content": "The interaction capability of a multimodal agent directly depends on the visual granularity it can perceive and understand. However, in the current research landscape, multi-granularity visual grounding remains under-explored. To break the limitations wherein visual-language alignment is constrained to the object level, we intend to take a step forward towards finer-grained vision-language understanding and propose the multi-granularity RES task, which requires models' proficiency in uniformly grounding entities across both part and object levels in response to various textual references. Within the single modality, it entails an enhancement of the capacity to apprehend complex, implicit knowledge structures, e.g., the granularity delineated by textual descriptions and the advanced level of pixel-wise semantic comprehension. Across the vision-language modalities, finer-grained references also pose new challenges for the accuracy and robustness of cross-modal alignment. Besides, since previous studies are mainly restricted to object-level referring segmentation and related benchmarks for the evaluation of part-level grounding performance are unavailable, we build a corresponding benchmark named RefCOCOM for performance evaluation, which will be elaborated below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 186, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 186, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 186, + 575 + ], + "type": "text", + "content": "3.2. RefCOCOM Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Manual Annotation Pipeline. To facilitate easy evaluation by community researchers without the need for any code modifications and build the multi-granularity benchmark with excellent quality, we split the existing object masks into part-level annotations through manual labeling on three subsets (i.e., validation, testA and testB) of the most prevalent RefCOCO [44] dataset. Initially, we establish a set of criteria for part masks delineation and identify the specific part categories requisite for each object. Subsequently, we employ SAM [20] to procure the preliminary segmentation results. We recruit 30 skilled annotators to refine and cap" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "tioning these results through an online annotation tool developed by ourselves. To ensure the annotation quality, we conduct spot checks at three progress nodes of " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " of the total workload with corrective feedback." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 129, + 443, + 222 + ], + "blocks": [ + { + "bbox": [ + 312, + 129, + 443, + 222 + ], + "lines": [ + { + "bbox": [ + 312, + 129, + 443, + 222 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 443, + 222 + ], + "type": "image", + "image_path": "bfa889bcc52869eba53a52eb2296731a7efd7656f0d2aa8094b62d8634249620.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 224, + 388, + 232 + ], + "lines": [ + { + "bbox": [ + 378, + 224, + 388, + 232 + ], + "spans": [ + { + "bbox": [ + 378, + 224, + 388, + 232 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 444, + 128, + 539, + 222 + ], + "blocks": [ + { + "bbox": [ + 444, + 128, + 539, + 222 + ], + "lines": [ + { + "bbox": [ + 444, + 128, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 444, + 128, + 539, + 222 + ], + "type": "image", + "image_path": "5cb87e73551e278622d387075ca584dd1f2185a8dcaebd70d6cda09692460bb5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 224, + 495, + 232 + ], + "lines": [ + { + "bbox": [ + 485, + 224, + 495, + 232 + ], + "spans": [ + { + "bbox": [ + 485, + 224, + 495, + 232 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 235, + 547, + 267 + ], + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 267 + ], + "type": "text", + "content": "Figure 2. RefCOCom benchmark statistics. (a) the number of referring expressions per parts' category in the log scale. (b) the word cloud highlights the head categories." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 269, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 547, + 424 + ], + "type": "text", + "content": "Benchmark Details. Totally, we collect 70k part references and corresponding 26k masks. The average length of the references is 5.1, covering 80 object and 391 part categories, among which the number of referring expressions per parts' category and the word cloud that highlights the head categories are both presented in Fig. 2. Mixed with the original object-level annotations, our RefCOCOM has a total of 34k masks and 92k references. As an enhanced expansion of RefCOCO [44], our RefCOCOM puts forward higher requirements for referential understanding and visual perception capabilities. To better match the intention of multi-granularity unification, we leverage mean Intersection-over-Union (mIoU) as evaluation metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 435, + 517, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 517, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 517, + 449 + ], + "type": "text", + "content": "4. Multi-Granularity Grounding Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 455, + 438, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 438, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 438, + 468 + ], + "type": "text", + "content": "4.1. Data Collection Engine" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": "Given the intrinsic complexity of classic RES task, the associated training data necessitates extensive annotation costs across both textual and visual domains. When the granularity is promoted to part-level, the annotation intricacies are further exacerbated. We contend that the primary bottleneck hindering the emergence of open-world grounding is the limitation imposed by the current scalability of data. By leveraging robust foundation models for synergistic enhancement, we introduce an advanced data engine capable of automatically generating reliable visual grounding data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Multi-Grained Dense Captioner. As the large language models (LLMs) [5, 37] continue to empower the multimodal domain with profound capabilities, the current scope of large vision-language models (LVLMs) [1-3] has effectively expanded into the grounding domain. Formally, these models utilize numerical coordinates to represent bounding boxes, allowing positional information to be seamlessly integrated into the language model via word embeddings for contextual understanding. However, the research community's exploration into the part granularity remains nascent." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 318, + 749 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 318, + 749 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 318, + 749 + ], + "type": "text", + "content": "13000" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 70, + 272, + 244 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 272, + 244 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 272, + 244 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 272, + 244 + ], + "type": "image", + "image_path": "75deb53c97770984745486f7627f430217d633ff649018fea9e933d3281332a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 238, + 256 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 238, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 238, + 256 + ], + "type": "text", + "content": "(a) Multi-Grained Dense Captioner" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 71, + 498, + 166 + ], + "blocks": [ + { + "bbox": [ + 276, + 71, + 498, + 166 + ], + "lines": [ + { + "bbox": [ + 276, + 71, + 498, + 166 + ], + "spans": [ + { + "bbox": [ + 276, + 71, + 498, + 166 + ], + "type": "image", + "image_path": "af6fa3fe88e2cc4617bcf965ef9d6a55d037ffb2e9bf6b4ec8741412a14f48fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 277, + 173, + 497, + 244 + ], + "blocks": [ + { + "bbox": [ + 277, + 173, + 497, + 244 + ], + "lines": [ + { + "bbox": [ + 277, + 173, + 497, + 244 + ], + "spans": [ + { + "bbox": [ + 277, + 173, + 497, + 244 + ], + "type": "image", + "image_path": "558b5488a8c5d122f2d2f1455594f3594ed9ab18494079cc85e3d17fafebb97e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 443, + 256 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 443, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 443, + 256 + ], + "type": "text", + "content": "(c) Part-Level Grounding Data Generation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 261, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 318 + ], + "type": "text", + "content": "Figure 3. The illustration of our data engine for building the MRES-32M dataset. (a) We start by fine-tuning an LVLM to create a capable dense captioner, which can effectively handle captioning tasks at three levels of granularity. (b) To generate object-level grounding data, we feed images and original bounding boxes into the dense captioner and a powerful segmenter to obtain the captions and masks of various objects. (c) We leverage the external knowledge from LLMs to decompose the existing object category annotations into a vocabulary set of part-level tags, which are sequentially fed into an open-vocabulary segmenter and our captioner to acquire the part-level annotations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "content": "The constrained training data have resulted in the fine-grained descriptive capabilities of LVLMs being confined merely to the object-level. To fully harness the open-world visual knowledge acquired from extensive pretraining on image-level and object-level data, we have devised a unified fine-tuning scheme that equips LVLMs to operate as dense captioners across all granularity levels, as depicted in Fig. 3 (a). Data for all three granularity levels are sourced from manual annotations to ensure reliability. The input comprises an image and the corresponding bounding box to be described, with bounding box coordinates normalized to integer values within the [0,999] range. For image granularity, we employ the COCO dataset [24], where all bboxes are uniformly represented as " + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "inline_equation", + "content": "(999,999)" + }, + { + "bbox": [ + 46, + 326, + 289, + 613 + ], + "type": "text", + "content": ". For object granularity, we utilize the Visual Genome dataset [21]. For part granularity, we draw upon unimodal semantic segmentation data [4, 12, 32] and employ a template in the form of PartNameX of ObjectNameY to construct dense captions. This unified multitask training approach can be synergistic across different granularity: it allows LVLMs to incorporate more comprehensive and detailed information to enhance part granularity descriptions. Concurrently, knowledge of part granularity assists LVLMs in generalizing knowledge within object interiors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Model-Assisted Data Generation. For the data generation of object-level visual grounding, we capitalize on largescale object detection dataset Object365 [34] to furnish highly reliable bounding boxes. Moreover, the rich category labels inherent in it ensure a comprehensive knowledge coverage. As illustrated in Fig. 3 (b), the bounding boxes will serve as visual prompts, which are independently sent into both a promptable segmenter (i.e. segment anything model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "type": "text", + "content": "[20]) and our dense captioner to obtain the segmentation masks and detailed semantic descriptions, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 546, + 469 + ], + "type": "text", + "content": "For part-level grounding data, we propose a hierarchical annotation scheme that builds upon existing object-level annotations, as shown in Fig. 3 (c). Specifically, employing GPT-4 [29] endowed with extensive external knowledge, we decompose the objects present in a given image to generate a customized part vocabulary set. This tailored vocabulary set is then fed into an open-vocabulary segmenter [35], which yields precise part masks and bounding boxes. These bounding boxes are subsequently sent into our dense captioner to acquire corresponding detailed captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 471, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 567 + ], + "type": "text", + "content": "Data Filtering. After completing the multi-granularity annotation of all images, we further introduce CLIP [30] for filtering. The bounding box is cropped from the original image, and then sent to the encoder together with the dense caption to measure the similarity. To ensure the consistency between the visual and linguistic annotations to a great extent, we retain box-caption pairs with similarity greater than 0.5 as the final annotation results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 574, + 459, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 459, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 459, + 586 + ], + "type": "text", + "content": "4.2. MRES-32M Dataset Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "As listed in Table 2, existing datasets, such as the most commonly used benchmark dataset RefCOCO [44], have limitations in terms of small data scale and lacking part-level dense annotations. As one of the pioneering works for finer-granularity unimodal visual comprehension, PACO [32] only contains semantic tags of objects or parts, without informative descriptions and visual context encompassed as referring expressions. Summarily, we compare the proposed MRES-32M with existing datasets and list some unique and significant properties of our dataset in Table" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13001" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 126, + 288, + 289 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 125 + ], + "type": "text", + "content": "Table 2. Comparisons with previous object-level visual grounding datasets and part-level segmentation datasets. # denotes the specific number, where Cats and Avg Len denote the object/part categories and the average length of referring expressions. “-” denotes the corresponding part masks or captions are unavailable." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 126, + 288, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 126, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 288, + 289 + ], + "type": "table", + "html": "
Dataset#Images#Objs#Parts#Cats#Avg Len
Object-Level Visual Grouding
ReferIt [18]20K97K-238/-3.2
RefCOCO [44]20K50K-80/-3.6
RefCOCO+ [44]20K49K-80/-3.5
RefCOCOg [27]26K54K-80/-8.4
GRES [25]20K60K-80/-3.7
Part-Level Segmentation & Detection
PartsIN [12]24K24K112K158/609-
PascalPart [4]19K40K363K20/193-
PACO [32]20K260K641K75/456-
Multi-Grained Visual Grounding
MRES-32M (Ours)1M15.3M16.9M365/22994.6
", + "image_path": "79d039ac35d2a83b4f5b70a4ca1df7815b6896f4987c7cea792820ce72068b2c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 297, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 321 + ], + "type": "text", + "content": "2. Besides, we have also provided a few examples in our MRES-32M dataset in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 287, + 393 + ], + "type": "text", + "content": "Unified Multi-Granularity. In comparison to the grounding counterparts, our MRES-32M is the first visual grounding dataset covering both part and object granularity. In comparison to the part-level segmentation counterparts, our MRES-32M provides informative and unique fine-grained description for each part mask." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 393, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 453 + ], + "type": "text", + "content": "More Diversified Categories. Our MRES-32M is composed of 365 object categories and an associated 2,299 part categories. Compared with existing datasets, it covers a wider range of multimodal knowledge and is an important step towards open-world understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 287, + 548 + ], + "type": "text", + "content": "Breakable Data Scales. To the best of our knowledge, MRES-32M is the largest dataset in the current grounding research community. In terms of the number of images and object instances, it surpasses the largest existing visual grounding dataset RefCOCOg [27] by factors of 38 and 283, respectively. Meanwhile, it encompasses part instance counts that exceed the largest existing part semantic segmentation dataset [32] by 58 times." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 287, + 621 + ], + "type": "text", + "content": "More Complex References. Benefiting from our LVLM-based dense captioner, the reference of MRES-32M could be more fully combined with the visual context for entity (i.e., part and object) description. Without sticking to a specific template, the relationships and attributes of entities could be highlighted in free natural language expressions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 219, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 219, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 219, + 646 + ], + "type": "text", + "content": "5. Multi-Granularity RES Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "Next, we describe the proposed multi-granularity RES model UniRES for the unified MRES task with the referring targets at both object and part granularity. Since our original intention is to establish a simple and easy-to-follow baseline model for the proposed multi-granularity RES task," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "the structure of our model UniRES is designed to be simple and clean. As shown in Fig. 4, UniRES has three major components, which will be illustrated below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 228 + ], + "type": "text", + "content": "Vision-Language Backbone. Taking account of the required attributes of both strong ability to capture the vision-language feature representations and promising scalability, we leverage the CLIP pre-trained weights of CLIP model [31], which learns transferable visual and linguistic concepts from vast image-text pairs, and adopt it as the backbone for our referring segmentation framework. The respective image and text encoders (i.e. Vision Transformer (ViT) [8] and Transformer [38] respectively) from CLIP are utilized to effectively extract visual and linguistic features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 228, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 479 + ], + "type": "text", + "content": "Query-based Grouping Design. To effectively accomplish the proposed MRES task, it is essential to exploit both low-level local and high-level global visual features. In order to enhance the local-global visual representations of CLIP backbone without introducing much additional parameters or altering model's structure, we have incorporated 64 and 8 learnable tokens (set empirically) into the first and middle layers of CLIP visual backbone. These learnable tokens traverse the first and second halves of the visual backbone. We expect that the ViT's internal self-attention mechanism implicitly serves as a manner to perform visual grouping, obtaining representative group tokens that capture the low-level local and high-level global features simultaneously. Based on the fact that local features spread more fragmented, the number of appended low-level group tokens is greater than that of high-level counterparts. Then, group tokens from both levels are fed into a language-guided region filter (LRF) to select the language-related visual features by cross-attention mechanism, which is followed by concatenation to fuse these expression-related visual group tokens for subsequent vision-language decoding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 479, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 635 + ], + "type": "text", + "content": "Two-Stage V-L Decoder. Now that the visual and textual feature representations from backbone, and the expression-related group tokens with two different levels are obtained, the two-stage mask decoder which comprises stacked Transformer layers is utilized to generate the segmentation masks. Specifically, the first-stage V-L decoder takes the extracted visual and textual features as input and generates first-stage fused multimodal representations. Subsequently, these multimodal features are further integrated with the grouped expression-related region features of the two semantic levels (i.e., low-level & high-level) to realize further feature enhancement, which is followed by a linear projection layer to obtain the final segmentation masks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 646, + 432, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 432, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 432, + 658 + ], + "type": "text", + "content": "6. Experimental Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, comprehensive experiments are conducted on three classic RES datasets (i.e., RefCOCO [44], RefCOCO+ [44], RefCOCOg [27]) and our RefCOCom benchmark for multi-granularity RES." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13002" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 68, + 492, + 186 + ], + "blocks": [ + { + "bbox": [ + 91, + 68, + 492, + 186 + ], + "lines": [ + { + "bbox": [ + 91, + 68, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 91, + 68, + 492, + 186 + ], + "type": "image", + "image_path": "b8862dd02c5505d6445148adf61783e93b874d375670eb69ca13edf10abdb924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 222 + ], + "type": "text", + "content": "Figure 4. The architecture of our UniRES model as a simple baseline for the MRES task. UniRES mainly comprises three parts: the visual and textual backbone for feature extraction, the pixel grouping design for aggregating the low-level and high-level features, and the cascaded two-stage vision-language decoder for multimodal feature fusion and the generation of segmentation masks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 227, + 110, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 110, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 110, + 238 + ], + "type": "text", + "content": "6.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 269 + ], + "type": "text", + "content": "The details about the three benchmark datasets for classic RES task can be found in the appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 277, + 164, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 164, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 164, + 290 + ], + "type": "text", + "content": "6.2. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 296, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 319 + ], + "type": "text", + "content": "Implementation Details. The implementation details are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 320, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 390 + ], + "type": "text", + "content": "Evaluation metrics. We adopt mean Intersection-over-Union (mIoU) and overall Intersection-over-Union (oIoU) as evaluation metrics. The mIoU measures the ratio between the intersection area and the union area of the predictions and labels among the test samples, while the oIoU calculates the total intersection area over total union area." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 400, + 132, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 132, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 132, + 411 + ], + "type": "text", + "content": "6.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 419, + 212, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 212, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 212, + 430 + ], + "type": "text", + "content": "6.3.1 Multi-Granularity MRES Task" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "type": "text", + "content": "Comparison with SOTA Methods. To evaluate the multi-granularity grounding performance of our UniRES as the baseline for the proposed MRES task and previous RES methods, we further conduct experimental comparison on our newly built RefCOCOM benchmark dataset. As presented in Table 3, the classic RES methods including four specialist models (i.e., SeqTR [47], CRIS [40], LAVT [42]) and two generalist models (e.g., X-Decoder [48] and SEEM [49]) are incorporated. For fair comparisons, we re-implement these SOTA methods and report their performance on our RefCOCOM. It is clear that either the specialist models for classic RES task or the powerful generalist models perform poorly on RefCOCOM, which requires the crucial skills of both part-level and object-level referring segmentation. Benefiting from our MRES-32M dataset, our UniRES can better master the part-level RES skills and handle the multi-granularity RES task, achieving considerably higher segmentation accuracy. Due to the significantly higher difficulty of part-level RES (i.e. multi-granularity RES) compared to classic RES, the absolute value of segmentation accuracy is correspondingly lower. This further emphasizes the importance of researching finer-grained part grounding where previous SOTA methods have fallen short." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 259, + 543, + 366 + ], + "blocks": [ + { + "bbox": [ + 305, + 225, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 225, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 545, + 258 + ], + "type": "text", + "content": "Table 3. Comparison with previous SOTA methods on our Ref-COCOM benchmark in terms of mIoU. Part and Obj & Part denote part-only and multi-grained evaluation settings of our MRES task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 259, + 543, + 366 + ], + "lines": [ + { + "bbox": [ + 308, + 259, + 543, + 366 + ], + "spans": [ + { + "bbox": [ + 308, + 259, + 543, + 366 + ], + "type": "table", + "html": "
MethodsvaltestAtestB
PartObj & PartPartObj & PartPartObj & Part
Specialists
SeqTR [47]13.928.212.122.818.134.7
CRIS [40]10.625.410.121.212.930.0
LAVT [42]15.329.913.224.418.735.5
Generalists
X-Decoder [48]16.229.513.623.620.333.8
SEEM [49]16.129.413.623.420.433.9
UniRES (Ours)19.634.316.427.825.241.7
", + "image_path": "7012d70d05bcb50ab8fd00baf18d25eb324d25d1e12c42ba409769a378098755.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 372, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 408 + ], + "type": "text", + "content": "Qualitative Analysis. We also conduct visual comparisons of CRIS [40], LAVT [42] and our UniRES on the MRES task, which is provided in appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 419, + 471, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 471, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 471, + 431 + ], + "type": "text", + "content": "6.3.2 Classic Object-Level RES Task" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": "Comparison with SOTA Methods. To validate the superiority of our MRES-32M dataset and model UniRES, our framework is fairly evaluated against previous SOTA methods on RefCOCO [44], RefCOCO+ [44] and RefCOCOg [27] under both zero-shot and fine-tuning settings. As presented in Table 4, our method greatly outperforms previous methods in terms of segmentation accuracy across all the benchmark datasets. Via directly zero-shot transferring to the downstream classic RES task after pre-training on our MRES-32M dataset, Our UniRES model achieves a leading zero-shot segmentation accuracy (i.e., approximately " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": ") without any fine-tuning to adapt to downstream task's data, which is significantly higher than all the recently proposed zero-shot RES methods by a large margin (i.e., " + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow 30 - 40\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 715 + ], + "type": "text", + "content": " mIoU). At the same time, it is worth noting that our zero-shot segmentation performance is already better than many of previous RES methods under fine-tuning setting (e.g., CRIS [40] and ReSTR [19]), validating the potential of our MRES-32M dataset and UniRES model. Furthermore, after fine-tuning on the classic RES datasets, our UniRES greatly surpasses previous methods no matter the specialist models (e.g., LAVT [42] and LISA [22]) for classic RES or the generalist models (e.g., X-Decoder [48] and SEEM [49])." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13003" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 95, + 492, + 396 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 545, + 94 + ], + "type": "text", + "content": "Table 4. Comparisons with the state-of-the-art approaches on previous three classic RES benchmark datasets under both the zero-shot and fine-tuning settings. “-” denotes that the result is not provided." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 95, + 492, + 396 + ], + "lines": [ + { + "bbox": [ + 100, + 95, + 492, + 396 + ], + "spans": [ + { + "bbox": [ + 100, + 95, + 492, + 396 + ], + "type": "table", + "html": "
MethodRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Zero-Shot Methods
mIoURegion token [46] (CVPR-23)23.422.124.624.522.625.427.627.3
Cropping [46] (CVPR-23)24.822.625.726.324.126.531.930.9
Global-Local CLIP [46] (CVPR-23)26.224.926.627.825.627.833.533.7
SAM-CLIP [28] (arXiv-23)26.325.826.425.728.026.838.838.9
Ref-Diff [28] (arXiv-23)37.238.437.237.340.533.044.044.5
TAS [36] (arXiv-23)39.841.136.243.649.136.546.646.8
UniRES (Ours)71.274.866.059.966.751.462.363.2
Fine-Tune Methods
oIoUEFNet [9] (CVPR-21)62.865.759.751.555.243.0--
LTS [17] (CVPR-21)65.467.863.154.258.348.054.454.3
ReSTR [19] (CVPR-22)67.269.364.555.860.448.3--
ReLA [25] (CVPR-23)73.876.570.266.071.057.765.066.0
X-Decoder [48] (CVPR-23)------64.6-
SEEM [49] (arXiv-23)------65.7-
LISA [22] (arXiv-23)74.979.172.365.170.858.167.970.6
UniRES (Ours)77.480.974.769.476.161.469.071.7
mIoUVLT [7] (ICCV-21)65.768.362.755.559.249.453.056.7
RefTr [23] (NeurIPS-21)74.376.870.966.870.659.466.667.4
SeqTR [47] (ECCV-22)71.773.369.863.066.759.065.065.7
CRIS [40] (CVPR-22)70.573.266.162.368.153.759.960.4
LAVT [42] (CVPR-22)74.576.970.965.871.059.263.363.6
UniRES (Ours)79.281.676.673.078.165.871.773.2
", + "image_path": "42c9d7e6f25cfdc01eda5a7f7491d1bc0b3ea38b398eb0e9b1fee2001539b730.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 405, + 147, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 147, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 147, + 417 + ], + "type": "text", + "content": "6.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 424, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 288, + 485 + ], + "type": "text", + "content": "We conduct ablation experiments on our RefCOCOM validation set. The tables below involves three evaluation settings on our RefCOCOM, using only objects as RES targets, using only parts as RES targets, and a mixed setting that combines both of the above two granularity levels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 247, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 247, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 247, + 503 + ], + "type": "text", + "content": "6.4.1 Ablation Study on MRES-32M Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "Data Granularity. We first probe into the effect of pretraining data's granularity with " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": " of our MRES-32M dataset. As shown in Table 5, the baseline without pretraining on MRES-32M dataset obtains " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "75.2\\%" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "32.0\\%" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": " mIoU separately on our RefCOCOM validation set under three granularity levels. Either introducing the object-level or part-level data from MRES-32M dataset for pretraining consistently results in an considerable accuracy increase across different granularity settings. In fact, pretraining on data at a certain granularity level to improve performance on corresponding benchmark aligns with common sense. However, it is noteworthy that incorporating part-level data into training also enhances model performance on the only object-level RefCOCOM. This underscores the significance of part-level understanding, as emphasized in the Sec. 1, where grasping the nuances of parts can yield benefits for object-level grounding. Besides, by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "text", + "content": "jointly incorporating the training samples of both granularity, our method attains " + }, + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 406, + 545, + 466 + ], + "type": "text", + "content": " improvement against baseline under all the granularity settings on RefCOCOM, which fully demonstrate the benefit of exploiting our MRES-32M dataset for both object-level and part-level RES tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 502, + 526, + 573 + ], + "blocks": [ + { + "bbox": [ + 304, + 468, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 502 + ], + "type": "text", + "content": "Table 5. Ablation study on the granularity of pre-training data in our proposed MRES-32M. Object and Part denote the introduction of object-level and part-level data for pre-training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 502, + 526, + 573 + ], + "lines": [ + { + "bbox": [ + 323, + 502, + 526, + 573 + ], + "spans": [ + { + "bbox": [ + 323, + 502, + 526, + 573 + ], + "type": "table", + "html": "
ObjectPartRefCOCOM
Object-OnlyPart-OnlyObject&Part
75.215.830.5
✓77.515.931.1
✓75.918.432.6
✓✓78.018.633.3
", + "image_path": "86cb67d6e6e3a2a1e5232c2547a9b0c9369d1b61aad45ea907cad6d1fe83cae0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 596, + 515, + 666 + ], + "blocks": [ + { + "bbox": [ + 307, + 583, + 542, + 594 + ], + "lines": [ + { + "bbox": [ + 307, + 583, + 542, + 594 + ], + "spans": [ + { + "bbox": [ + 307, + 583, + 542, + 594 + ], + "type": "text", + "content": "Table 6. Ablation study on the data scale of MRES-32M dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 596, + 515, + 666 + ], + "lines": [ + { + "bbox": [ + 335, + 596, + 515, + 666 + ], + "spans": [ + { + "bbox": [ + 335, + 596, + 515, + 666 + ], + "type": "table", + "html": "
RatiosRefCOCOm
Object-OnlyPart-OnlyObject & Part
0%75.215.830.5
20%76.817.332.0
50%78.018.633.3
100%79.219.634.3
", + "image_path": "07ead81a73f5514e52f0e4f2309ae78134f79afa5bdfeeb7b5e745b166426178.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Data Scale. Next, we investigate the effect of different percentages of training samples in our MRES-32M dataset. The results are presented in Table 6. It is obvious in Table 6 that the model performance for our MRES tasks is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13004" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "consistently improved with more and more employed training samples, which implicitly validates the high quality of our built dataset. As the ratios of introduced training samples continue to rise, there's no sign of diminishing gains in model performance, suggesting that our framework has significant potential with continually scaled up training data." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 171, + 282, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 148, + 286, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 148, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 286, + 169 + ], + "type": "text", + "content": "Table 7. Ablation study on the effectiveness of our MRES-32M dataset on RES SOTA methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 171, + 282, + 243 + ], + "lines": [ + { + "bbox": [ + 51, + 171, + 282, + 243 + ], + "spans": [ + { + "bbox": [ + 51, + 171, + 282, + 243 + ], + "type": "table", + "html": "
MethodsMRES-32MRefCOCOM
Object-OnlyPart-OnlyObject & Part
CRIS70.510.625.4
CRIS✓73.115.529.7
LAVT74.515.329.9
LAVT✓75.719.333.2
", + "image_path": "218d4d664e348d968d8e77aec2492523bd82d8c4436dac602ca63f76d24ea40e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 246, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 437 + ], + "type": "text", + "content": "Dataset Necessity. To prove the necessity and effectiveness of our MRES-32M dataset for the proposed MRES task, we take previous SOTA methods (i.e., CRIS [40] and LAVT [42]) on classic RES and compare the segmentation accuracy of the same models with or without pretraining on our MRES-32M. As shown in Table 7, the original CRIS and LAVT can already well handle the classic RES task with only object-level grounding skills, but it performs poorly on the multi-grained RefCOCOM benchmark, which requires part-level grounding capability. In contrast, pre-training on our MRES-32M consistently leads to great performance gains no matter on object-only, part-only or the multi-granularity RES tasks, because our high-quality MRES-32M dataset can effectively enable the model to handle the part-level visual grounding task and enhances the original object-level grounding capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 445, + 226, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 226, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 226, + 456 + ], + "type": "text", + "content": "6.4.2 Ablation Study on UniRES Model" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 475, + 282, + 544 + ], + "blocks": [ + { + "bbox": [ + 56, + 462, + 277, + 473 + ], + "lines": [ + { + "bbox": [ + 56, + 462, + 277, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 277, + 473 + ], + "type": "text", + "content": "Table 8. Ablation study on the query-based grouping design." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 475, + 282, + 544 + ], + "lines": [ + { + "bbox": [ + 50, + 475, + 282, + 544 + ], + "spans": [ + { + "bbox": [ + 50, + 475, + 282, + 544 + ], + "type": "table", + "html": "
High-LevelLow-LevelRefCOCOM
Object-OnlyPart-OnlyObject&Part
✓74.414.929.6
74.915.230.0
✓74.915.430.1
✓✓75.215.830.5
", + "image_path": "7709e178a5a64fe6db8b3cb18b2a253b6b89468ec929eeaa0e8241581d1eb1de.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "Besides, we additionally conduct ablation studies on the structural design of our model UniRES. As presented in Table 8, removing the simple but effective grouping design at two different levels (this simultaneously leads to discarding the second-stage decoder) will result in a considerable deterioration in model performance. Since the sequentially appended pixel grouping tokens are introduced to effectively capture the high-level and low-level visual features for further visual feature enhancement, losing any type of these group tokens will lead to a decrease in the model's segmentation accuracy on both object-level and part-level RES tasks. In order to verify that the high-level and low-level pixel grouping tokens can capture the clustering features of the corresponding level, we also visualize the group tokens" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "of the two levels (before being sent to LRF) for qualitative analysis, where the different colors represent the various areas of the same clustering group. The visualization results in Fig. 5 affirms that our introduced low-level and high-level group tokens respectively capture the part-level local features and aggregate the object-level global features with stronger semantics through the pixel grouping process, which aligns with our intentions for the grouping design." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 176, + 391, + 281 + ], + "blocks": [ + { + "bbox": [ + 321, + 176, + 391, + 281 + ], + "lines": [ + { + "bbox": [ + 321, + 176, + 391, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 176, + 391, + 281 + ], + "type": "image", + "image_path": "5a96cf496155f0279018f07a66691f1ed4e21b61ba6b1b3fa0acd0ebc831ea43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 285, + 370, + 293 + ], + "lines": [ + { + "bbox": [ + 342, + 285, + 370, + 293 + ], + "spans": [ + { + "bbox": [ + 342, + 285, + 370, + 293 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 176, + 460, + 281 + ], + "blocks": [ + { + "bbox": [ + 392, + 176, + 460, + 281 + ], + "lines": [ + { + "bbox": [ + 392, + 176, + 460, + 281 + ], + "spans": [ + { + "bbox": [ + 392, + 176, + 460, + 281 + ], + "type": "image", + "image_path": "dff899dcdbbb12475949cb8646a90136655ef9874e5a8e1ed11b84466837ec16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 285, + 445, + 293 + ], + "lines": [ + { + "bbox": [ + 404, + 285, + 445, + 293 + ], + "spans": [ + { + "bbox": [ + 404, + 285, + 445, + 293 + ], + "type": "text", + "content": "(b) Low-Level" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 462, + 176, + 531, + 280 + ], + "blocks": [ + { + "bbox": [ + 462, + 176, + 531, + 280 + ], + "lines": [ + { + "bbox": [ + 462, + 176, + 531, + 280 + ], + "spans": [ + { + "bbox": [ + 462, + 176, + 531, + 280 + ], + "type": "image", + "image_path": "bac86dba758eb4ea3a68df8b2a6c2f0ccc213c34ff224fb50d9e5c3eb9d9050f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 285, + 515, + 293 + ], + "lines": [ + { + "bbox": [ + 473, + 285, + 515, + 293 + ], + "spans": [ + { + "bbox": [ + 473, + 285, + 515, + 293 + ], + "type": "text", + "content": "(c) High-Level" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 294, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 328 + ], + "type": "text", + "content": "Figure 5. Qualitative analysis for ablation study on the object-level and part-level grouping design in our model structure. (a) the input image. (b) low-level group tokens. (c) high-level group tokens." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 342, + 486, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 486, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 486, + 356 + ], + "type": "text", + "content": "7. Conclusion and Broader Impact" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 364, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 591 + ], + "type": "text", + "content": "In this paper, we move beyond previous works that focused solely on object-level visual grounding tasks and take a step further to finer-grained part-level RES. We put forward a new multi-granularity referring expression segmentation task and establish an evaluation benchmark named RefCOCOM by manual annotation. To advance the visual grounding at both object and part levels towards finer-grained vision-language understanding, we build the largest visual grounding dataset MG-32M to date, which is also the first dataset to provide part-level vision-language annotations. Furthermore, we have developed a simple yet strong multi-grained referring segmentation model called UniRES. As a baseline for our newly proposed MRES task, UniRES achieves the new state-of-the-art performance on both our RefCOCOM for MRES task and three classic RES datasets. We plan to release our RefCOCOM benchmark, the MG32M dataset, and the UniRES model to the public, aspiring to foster future research in fine-grained visual grounding tasks and to inspire new research in this direction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 404, + 621 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "We thank Yepeng Tang for the helpful discussions on this work, all the Image & Video Analysis Group (IVA)'s members in CASIA for the technical support, and all the insightful reviewers for the helpful suggestions. This work was supported by the National Science and Technology Major Project (No.2022ZD0118801), National Natural Science Foundation of China (U21B2043, 62206279)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13005" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "text", + "content": "[2] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "text", + "content": "[3] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 312 + ], + "type": "text", + "content": "[4] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 378 + ], + "type": "text", + "content": "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 314, + 288, + 378 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 314, + 288, + 378 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "text", + "content": "[6] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5485-5494, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 490 + ], + "type": "text", + "content": "[7] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 557 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "text", + "content": "[9] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "type": "text", + "content": "[10] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 932-940, 2017. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[11] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "text", + "content": "[12] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[13] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "type": "text", + "content": "[14] Yutao Hu, Qixiong Wang, Wenqi Shao, Enze Xie, Zhenguo Li, Jungong Han, and Ping Luo. Beyond one-to-one: Rethinking the referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4067-4077, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[15] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 297, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 362 + ], + "type": "text", + "content": "[16] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "text", + "content": "[17] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tie-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "text", + "content": "[19] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 578, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 643 + ], + "type": "text", + "content": "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "text", + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[23] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13006" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "in neural information processing systems, 34:19652-19664, 2021.7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "text", + "content": "[25] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 206, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 271 + ], + "type": "text", + "content": "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 287, + 337 + ], + "type": "text", + "content": "[27] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "text", + "content": "[28] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 383, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 287, + 403 + ], + "type": "text", + "content": "[29] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 602 + ], + "type": "text", + "content": "[32] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 603, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 658 + ], + "type": "text", + "content": "[33] N Dinesh Reddy, Minh Vo, and Srinivasa G Narasimhan. Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1906-1915, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "type": "text", + "content": "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "text", + "content": "[36] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial-aware zero-shot referring image segmentation. arXiv preprint arXiv:2310.18049, 2023. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[37] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "text", + "content": "[40] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[41] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "text", + "content": "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[43] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "type": "text", + "content": "[44] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 2,3,4,5,6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[45] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[46] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[47] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13007" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 239 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "type": "text", + "content": "and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 598-615. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "text", + "content": "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "text", + "content": "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. 2, 6, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "spans": [ + { + "bbox": [ + 294, + 738, + 317, + 748 + ], + "type": "text", + "content": "13008" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_content_list.json b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8bf1c3aaab375d618ca4f2546b78bd94d93f398d --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shentong Mo", + "bbox": [ + 269, + 204, + 380, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carnegie Mellon University", + "bbox": [ + 214, + 223, + 437, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shentong.mo@sv.cmu.edu", + "bbox": [ + 196, + 243, + 455, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pedro Morgado", + "bbox": [ + 576, + 204, + 704, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Wisconsin-Madison", + "bbox": [ + 506, + 223, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pmorgado@wisc.edu", + "bbox": [ + 537, + 243, + 741, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/stoneMo/DeepAVFusion", + "bbox": [ + 250, + 289, + 712, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 334, + 313, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures.", + "bbox": [ + 75, + 369, + 473, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 811, + 209, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans naturally integrate audio-visual information to perceive and understand the environment. Several studies in cognitive psychology and neuroscience, as well as classic perceptual illusions such as the McGurk effect, demonstrate", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that such audio-visual fusion can occur early on in the perceptual stack, enabling deeper integration of the two modalities. Early fusion models aim to emulate this human-like perception. Specifically, it refers to the process of integrating auditory and visual cues at an early stage of a multi-modal perception model, in order to leverage the synergistic effects of both modalities. This can be especially powerful if the fusion process can attend to and establish connections between local components of the visual and audio signal (e.g., the frequencies of someone's voice with the pixels of their lips). In sum, early fusion of local audio-visual interactions holds the promise of a deeper and more sophisticated understanding of audio-visual content, critical for many real-world applications such as visually guided source separation, localization/segmentation, and multi-modal recognition.", + "bbox": [ + 496, + 335, + 895, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the potential, audio-visual early fusion architectures have remained under-explored, with the majority of the recent literature focusing on late fusion or no fusion at all. One recent prominent research direction [1, 47] seeks to learn independent uni-modal encoders through contrastive learning by requiring the representations of associated audio-visual pairs to be synchronized in latent space. However, contrastive learning is not compatible with early fusion, as the connections between uni-modal representations lead to trivial solutions to the contrastive learning problem. Another line of recent work [20, 27] learns encoders with late audio-visual fusion through a combination of contrastive learning and masked reconstruction, focusing on downstream tasks that require shallow integration of audio-visual features, such as audio-visual action recognition.", + "bbox": [ + 496, + 566, + 895, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One potential reason is that early fusion architectures pose a significant training challenge due to the substantial increase in model expressivity it offers. As a result, robust learning frameworks are required to ensure the creation of high-quality models and effectively harness their enhanced capabilities. In this work, we demonstrate the effectiveness of a masked reconstruction framework for training audio-", + "bbox": [ + 496, + 795, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27186", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c8228083643d84dcd15370528386369b224f674d9c48f231fdd14daaa2744e5.jpg", + "image_caption": [ + "Figure 1. Unlike prior works on Audio-Visual MAEs which learn either late fusion encoders or separate encoders altogether, we demonstrate that masked reconstruction is especially useful for learning deeply integrated audio-visual encoders with powerful early fusion modules." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 267, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f834e35a30fa9a2ae49b6f5ca1133773d70b001ec1f968ce3dc04d7364add01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 97, + 464, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual transformers with early-fusion of local interactions (Fig. 1). Our hypothesis is that the simultaneous reconstruction of audio and visual inputs from a limited audio-visual context promotes the learning of deeply integrated representations, as the model is required to understand the fine interactions between the two modalities. Furthermore, masked reconstruction has already been successfully deployed in training uni-modal representations for a variety of input signals, including text [11], images [3, 22], and audio [28], as well as for learning late fusion representations [20, 27]. However, unlike what was previously observed with uni-modal or late fusion models, our work shows that masked audio-visual reconstruction with early fusion encoders results in an interesting emergent property. While uni-modal representations encode the low-level details required for masked reconstruction, we observed that high-level semantics (surprisingly) emerged from the tokens used for audio-visual fusion.", + "bbox": [ + 75, + 339, + 472, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to demonstrating the effectiveness of masked reconstruction for early fusion transformers, we propose a novel attention-based fusion module, which can effectively attend to the interactions between local audio and visual representations, thereby enhancing its capacity to capture localized interactions between the two modalities. Take, for instance, the sound of a dog barking. The sound signals the presence of a dog and offers clues about its state, such as an open mouth. While the link between the dog and its bark can be made at a global level, our fusion module allows the model to delve deeper into local interactions, like the correlation between the dog's mouth and the barking sound's spectral frequencies. As a result, localized fusion boosts the quality of learned representations for numerous tasks. However, it necessitates the model to explicitly represent all pairwise interactions of local audio and visual tokens. Given the vast number of tokens that current transformers process, dense local interactions can quickly become unmanageable. To enhance efficiency, our fusion module factorizes unimodal representations, thereby circumventing the need to", + "bbox": [ + 75, + 598, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model all pairwise interactions.", + "bbox": [ + 500, + 90, + 709, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In sum, we propose an attention-based fusion module that can attend to local audio-visual interactions and deploy it to learn deeply integrated audio-visual representations by fusing uni-modal representations at early layers in the architecture. We also demonstrate that when paired with masked reconstruction, our framework, denoted DEEPAVFUSION, can learn strong audio-visual representations, yielding state-of-the-art performance on a variety of audio-visual tasks, including visually guided source separation, localization, and segmentation. We extensively evaluated DEEPAVFUSION on a variety of audio-visual datasets, and conducted thorough ablation studies, showing the importance of early fusion, dense interactions, and uni-modal pre-training when learning deeply integrated audio-visual representations.", + "bbox": [ + 496, + 108, + 895, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 347, + 640, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-Visual Representations Learning. Audio-visual representations learning has been addressed in many previous works [1, 2, 14, 16, 23, 24, 30, 40, 43, 45–47, 50, 51, 65, 66] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for many audio-visual tasks, such as audio-event localization [31, 32, 37, 41, 42, 53, 62], audio-visual spatialization [4, 17, 44, 45], audio-visual navigation [4–6] and audio-visual parsing [33, 39, 54, 61]. In this work, our main focus is to learn transferrable audio-visual representations from masked audio-visual reconstruction, which is more challenging than the above-mentioned tasks.", + "bbox": [ + 496, + 367, + 893, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Representations Learning. Masked representation learning aims to learn self-supervised representations by reconstructing desired features of masked data given unmasked parts as clues. In the recent years, masked representation learning has achieved promising results in natural language processing [9, 11, 35, 52, 59] and computer vision [3, 8, 12, 15, 22, 58, 60, 63] community. Typically, BERT [11] randomly masked $15\\%$ of word tokens in the sentence and recovered them with unmasked words to learn generalizable textual features via a self-attention transformer [56]. A block-wise masking strategy was proposed in BEiT [3] to reconstruct discrete tokens of masked image patches for pre-training transferrable visual representations. To simplify the masked image encoding framework, MAE [22] directly reconstructed missing pixels of $75\\%$ masked patches using vision transformers [13] for self-supervised pre-training.", + "bbox": [ + 496, + 550, + 895, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual Masked Autoencoders. More recently, researchers introduced diverse masking pipelines [19, 20, 27] to show the effectiveness of masked modeling in learning audio-visual representations. For example, CAV-MAE [20] combined contrastive learning and masked modeling to capture a joint and coordinated audio-visual representation.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27187", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "They tried to add a joint encoder on audio-visual features from the last attention block of single-modality encoders to fuse cross-modal information for audio-visual contrastive objectives. MAViL [27] extended masked audio-video reconstruction with masked intra- and inter-modal contrastive learning and self-training by recovering joint audio-video contextualized representations. Despite their promising performance, they ignored the importance of the early fusion of audio-visual features in masked audio-visual reconstruction. In contrast to AV-MAE [19] that simply concatenated the audio and visual tokens before passing them through the joint transformer for early fusion, we will design an early fusion module with interactions between local and visual representations for audio-visual masked auto-encoders. Our audio-visual interactions are different from fusion bottlenecks in MBT [48] that forces information between different modalities to pass through a small number of bottleneck latent. However, we develop a fully novel attention-based fusion module that can effectively attend to interactions between local audio and visual representations, thus enhancing its ability to capture fine-grained interactions between the two modalities.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-visual Early Fusion. Audio-visual fusion has been the topic of extensive research, with several works proposing a variety of architectures to aggregate multi-modal representations [19, 48, 49]. An example of an early fusion architecture was proposed by Owens et al. [49] to learn representations from audio-visual correspondences, by concatenating features from small unimodal encoders and feeding them into a fused audio-visual network for the early fusion of multisensory features. Beyond this early work, most recent papers on audio-visual fusion [19, 48] propose alternative fusion mechanisms, either through weight sharing [19] or token-based fusion [48]. However, since all these works focused on downstream tasks (mostly classification) that do not require fine multi-modal understanding, the benefits of early fusion were not evident, with mid-level fusion often achieving optimal performance. Different from them, we proposed an early fusion model that is designed for fine multi-modal understanding, by attending to local interactions between the audio and the visual while performing early fusion. The new architecture design paired with masked reconstruction objective is shown to outperform many of the prior works on a variety of downstream multi-modal applications beyond classification.", + "bbox": [ + 76, + 425, + 472, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 796, + 169, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to learn audio-visual representations that can be effectively transferred to a variety of downstream tasks that require a detailed understanding of audio-visual interactions. To accomplish this, we introduce a novel audio-visual transformer architecture, named DEEPAVFUSION, that enables the", + "bbox": [ + 76, + 824, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "early fusion of audio and visual tokens through dense local interactions. We also show how to effectively train such early fusion transformers using multi-modal masked reconstruction. In this section, we first describe the proposed early fusion transformer and then introduce the multi-modal masked reconstruction framework used for self-supervised pre-training.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. A Transformer Architecture for Joint Audio-Visual Encoding", + "text_level": 1, + "bbox": [ + 498, + 205, + 893, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Downstream tasks such as visually guided sound source separation, localization, and segmentation require a deep understanding of fine audio-visual interactions present in the data. However, late fusion often lacks the expressive power to represent such interactions. We introduce a transformer architecture that enables the early fusion of multi-modal representations through factorized local interactions between audio and visual tokens.", + "bbox": [ + 496, + 247, + 893, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early fusion architecture While we aim to learn deeply integrated audio-visual representations, foundational uni-modal transformers can still provide a strong starting point for multi-modal learning. As such, we introduce a modular architecture to repurpose existing pre-trained uni-modal transformers. Specifically, we propose a three-branch architecture, illustrated in Fig. 2. The first two branches are uni-modal transformers that process audio and visual tokens, respectively, while the third branch is a multi-modal transformer that updates a set of learnable fusion tokens to fuse audio-visual information. Formally, let $\\mathbf{X}_0^f \\in \\Re^{F \\times D}$ be $F$ learnable fusion tokens. We progressively update these tokens, at each layer $l$ , through a fusion block $\\Phi_l^f$ , which can interact with and aggregate information from modality-specific features $\\mathbf{X}_{l - 1}^v$ , $\\mathbf{X}_{l - 1}^a$ . Furthermore, fusion tokens $\\mathbf{X}_{l - 1}^f$ are fed into modality-specific blocks, $\\Phi_l^v$ and $\\Phi_l^a$ , to modulate uni-modal representations from early stages in the network. Overall, representations are updated as follows", + "bbox": [ + 496, + 369, + 895, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {l} ^ {f} = \\Phi_ {l} ^ {f} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {a}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 648, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {l} ^ {v} = \\Phi_ {l} ^ {v} \\left(\\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 671, + 890, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {l} ^ {a} = \\Phi_ {l} ^ {a} \\left(\\boldsymbol {X} _ {l - 1} ^ {a}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 693, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This token update strategy is repeated for a total of $L$ layers to compute the final representations $\\mathbf{X}_L^v, \\mathbf{X}_L^a$ and $\\mathbf{X}_L^f$ .", + "bbox": [ + 496, + 715, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modality-specific blocks We use standard transformer self-attention for the modality-specific blocks, $\\Phi^v$ and $\\Phi^a$ , while allowing them to further attend to fusion tokens. Unless otherwise specified, we initialize $\\Phi^v$ and $\\Phi^a$ from foundational uni-modal models. This is possible since the parameter space of the modality-specific blocks remains unchanged. In particular, we empirically validated the benefits of bootstrapping from [28] and [22], which are pre-trained on large-scale audio and visual datasets, respectively, through a uni-modal masked reconstruction objective.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27188", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5115c4a420005adddb207674d25a2c220139b865387eab1d8e5f06701148c7fb.jpg", + "image_caption": [ + "Figure 2. Audio-Visual Early Fusion Transformer with Dense Interactions. The architecture ① is composed of three interconnected branches - the visual, audio, and fusion branches - each with an equal number of transformer blocks. The visual and audio branches process visual and audio tokens, respectively, while simultaneously attending to fusion tokens. The fusion branch updates a set of learnable fusion tokens to fuse audio-visual information. To better attend to local audio-visual interactions, we introduce a new fusion block ②. At each layer $l$ , all patch-level audio (visual) tokens, $\\bar{X}_{l-1}^{a}(\\bar{X}_{l-1}^{v})$ , are first aggregated into a small set of aggregation tokens, $\\bar{X}_l^a (\\bar{X}_l^v)$ by cross-attention. Pairwise audio-visual interactions are then encoded through a linear layer into latents $\\bar{X}_l^{av}$ , which are used to update the fusion tokens $\\bar{X}_l^f$ by cross-attention." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 398, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd6002b3aba7b5ca98a5b9a39fdcbf97ad92f5bf452d5e485275ef3a59ce7e25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 98, + 870, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion blocks The design of the fusion blocks $\\Phi^f$ is critical to representing fine audio-visual interactions. In the context of transformers, a natural way to combine audio-visual representations is through a standard self-attention block which, in addition to the fusion tokens $X^f$ themselves, it can further attend to all image and audio tokens, $X_{l - 1}^v$ and $X_{l - 1}^{a}$ . This approach, which we refer to as token fusion, has been used in previous works such as MBT [48]. While intuitive, standard self-attention cannot directly exploit interactions between local audio-visual representations, which limits their expressivity. To address this limitation, we designed a fusion block that attends to local interactions.", + "bbox": [ + 75, + 318, + 472, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion with Dense Audio-Visual Interactions Interactions between audio and visual data occur between local regions in the image and the audio spectrogram. Given a set of $n_a$ audio tokens $X^a$ and $n_v$ visual tokens $X^v$ , aggregation of local multi-modal interactions requires attending to representations from all possible $n_a \\times n_v$ pairs, to identify the ones that are most useful for the task at hand. While we experimented with several ways to represent multi-modal interactions, from bilinear aggregation to kernel-based aggregation, we found that linear aggregation was sufficient to capture audio-visual interactions for our tasks. Formally,", + "bbox": [ + 75, + 502, + 472, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {i j} ^ {a v} = W ^ {a} \\boldsymbol {X} _ {i} ^ {a} + W ^ {v} \\boldsymbol {X} _ {j} ^ {v} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 680, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {a v} = \\left[ \\boldsymbol {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {a} \\forall j = 1, \\dots , n _ {v} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 700, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion tokens $X_{l}^{f}$ at layer $l$ are then updated using standard cross-attention [56] over the representation of all local interactions $X^{av}$", + "bbox": [ + 75, + 732, + 470, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}; \\boldsymbol {X} _ {l - 1} ^ {\\text {a v}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 787, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Factorized Audio-Visual Interactions While the procedure above can attend to local interactions, it requires the model to explicitly represent all possible $n_a \\times n_v$ pairs. Given that the number of audio and visual tokens is typically large, the number of interactions can quickly become intractable,", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resulting in high memory consumption and low throughput. To address this limitation, we introduce modality-specific aggregation tokens updated by aggregation blocks. The goal of the aggregation blocks is to summarize the large number of audio/visual tokens $X^a / X^v$ into a small set of $n_{\\bar{a}} / n_{\\bar{v}}$ aggregation tokens $\\bar{X}^a / \\bar{X}^v$ , respectively. Specifically, at layer $l$ , aggregation tokens are updated by cross-attention", + "bbox": [ + 496, + 319, + 893, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} _ {l} ^ {a} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {a}, X _ {l - 1} ^ {a}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 435, + 890, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} _ {l} ^ {v} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {v}, X _ {l - 1} ^ {v}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 455, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, instead of explicitly representing all $n_a \\times n_v$ pairwise interactions, we restrict audio-visual interactions to the sets of aggregated tokens. Formally, factorized interactions between aggregated tokens $\\bar{X}_i^a$ and $\\bar{X}_j^v$ are represented as", + "bbox": [ + 496, + 484, + 893, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} _ {i j} ^ {a v} = W ^ {\\bar {a}} \\bar {X} _ {i} ^ {a} + W ^ {\\bar {v}} \\bar {X} _ {j} ^ {v} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 556, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} ^ {a v} = \\left[ \\bar {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {\\bar {a}} \\forall j = 1, \\dots , n _ {\\bar {v}} \\right] \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 577, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the fusion tokens are updated at layer $l$ by", + "bbox": [ + 500, + 606, + 805, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\bar {\\boldsymbol {X}} _ {l} ^ {a v}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 632, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we only represent interactions between the aggregated uni-modal tokens, the number of possible interactions is greatly reduced. We observed that a $700 \\times$ reduction still maintains the benefits obtained with local dense interactions.", + "bbox": [ + 496, + 666, + 893, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Early-Fusion Transformers Through Audio-Visual Masked Auto-Encoding", + "text_level": 1, + "bbox": [ + 500, + 738, + 890, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked auto-encoders have been shown to learn effective representations for a wide variety of input signals, including images [3, 22] and audio [28]. Formally, the input signal $x$ is partitioned into a set of patch tokens $x = \\{x_{t}\\}_{t=1}^{T}$ , which is randomly split into the visible set $X_{\\mathrm{vis}} = \\{x_{t}\\}_{t \\in \\mathcal{V}}$ and the masked set $X_{\\mathrm{mask}} = \\{x_{t}\\}_{t \\in \\mathcal{M}}$ ( $\\mathcal{V}$ and $\\mathcal{M}$ are non-overlapping index sets). In the case of images, visible and masked tokens, $X_{\\mathrm{vis}}$ and $X_{\\mathrm{mask}}$ , are obtained from a grid", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27189", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of small image patches [3, 22], while in the case of audio signals, tokens are obtained either from small windows of a raw waveform or from small 2D patches of the audio spectrogram [28]. Then, to learn representations of the input data, a transformer encoder $f$ is first used to encode the visible tokens $X_{\\mathrm{vis}} = f(X_{\\mathrm{vis}}) \\in \\Re^{|\\mathcal{V}| \\times d}$ , where $d$ denotes the dimensionality of the token representation. These representations together with a series of (position marked) mask tokens $M_{\\mathrm{mask}}$ are then fed to a decoder $g(\\cdot)$ that seeks to reconstruct the masked input tokens $\\hat{X}_{\\mathrm{mask}} = g(X_{\\mathrm{vis}}, M_{\\mathrm{mask}})$ . To accomplish that, the model is trained to minimize the mean squared reconstruction error", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}}, \\hat {X} _ {\\text {m a s k}}\\right) = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\substack {x \\in X _ {\\text {m a s k}} \\\\ \\hat {x} \\in \\hat {X} _ {\\text {m a s k}}}} \\| x - \\hat {x} \\| _ {2} ^ {2}. \\tag{12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 277, + 470, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior work [20, 27], masked reconstruction can be easily extended to multi-modal signals, by simultaneously masking tokens from both audio and visual modalities. In the case of early-fusion transformers, the encoder $f$ is then used to encode the visible tokens, $X_{\\mathrm{vis}}^v$ and $X_{\\mathrm{vis}}^a$ ,", + "bbox": [ + 75, + 335, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}, \\boldsymbol {X} ^ {f} = f \\left(\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 417, + 468, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which, in addition to the uni-modal representations, $X_{\\mathrm{vis}}^v$ and $X_{\\mathrm{vis}}^a$ , also returns a set of fusion representations $\\pmb{X}^{f}$ . These representations are then fed to modality-specific decoders to reconstruct both audio and visual masked tokens, $\\hat{X}_{\\mathrm{mask}}^v = g_v(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^v)$ and $\\hat{X}_{\\mathrm{mask}}^a = g_a(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^a)$ , by minimizing the loss", + "bbox": [ + 76, + 441, + 472, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A V - M A E}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {v}, \\hat {X} _ {\\text {m a s k}} ^ {v}\\right) + \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {a}, \\hat {X} _ {\\text {m a s k}} ^ {a}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 539, + 468, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{MAE}}$ is the auto-encoder loss of Eq. 12.", + "bbox": [ + 76, + 578, + 388, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While multi-modal masked reconstruction has been previously explored in [20, 27], these approaches focus on learning separate audio and image encoders, $f_{a}$ and $f_{v}$ , with multi-modal representations obtained through a late fusion module $f_{f}$ . In contrast, we propose to learn a single multimodal early fusion encoder $f$ , and demonstrate that masked reconstruction is especially useful for learning multi-modal representations capable of representing fine audio-visual interactions through early fusion of local interactions between audio and visual tokens.", + "bbox": [ + 75, + 595, + 470, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 761, + 209, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now evaluate the proposed DEEPAVFUSION on a variety of audio-visual downstream tasks that require a deep understanding of audio-visual interactions.", + "bbox": [ + 75, + 789, + 470, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 844, + 290, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training. For pre-training, we experiment with two datasets, namely VGG-Sounds [7] (with 144k training sam", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ples) and AudioSet [18] (with 1.73M training samples), both of them containing diverse videos focused on audio events and sound sources. The audio is represented by log mel spectrograms extracted from $3s$ of audio at a sample rate of $16000\\mathrm{Hz}$ . To compute the log spectrograms, we apply an STFT using approximately 50ms windows with a hop size of $15\\mathrm{ms}$ , resulting in an input tensor of size $128\\times 196$ (128 mel frequency bands over 196 time steps). Since we use image models for the visual component, we randomly extract single frames from within the time window defined by the $3s$ audio snippets. Input frames are augmented by random crops with a minimum area of 0.5 and random horizontal flips, and resized into a $224\\times 224$ resolution.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unless otherwise specified, we initialize the uni-modal transformer blocks using modality-specific encoders pretrained with uni-modal masked reconstruction objectives [22, 28]. Specifically, we use the ViT-Base model from [22] pre-trained on ImageNet [10], and the Spec-MAE model (with the same ViT-Base architecture) pre-trained on AudioSet [18]. In our base configuration, we add multi-modal fusion blocks on all 12 layers of the model, together with 16 multi-modal fusion tokens, 8 visual aggregation tokens, and 8 audio aggregation tokens. For efficiency, we reduced the dimensionality of the MLP embeddings using an MLP ratio of 1 (instead of the standard 4), as well as the dimensionality of space in which the similarity for self and cross-attention is computed to 16 (as opposed to the standard 64).", + "bbox": [ + 496, + 289, + 893, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The models were trained using the Adam optimizer [29]. Ablation and parametric studies were conducted on VG-GSound, training for 200 epochs with a total batch size of 512 (split between 2 GPUs and 4 iterations of gradient accumulation). Large-scale training on AudioSet was conducted for 200 epochs with a total batch size of 2048 (split between 16 GPUs and 4 iterations of gradient accumulation). In all cases, we use a base learning rate of $1.5e - 4$ adjusted for the effective batch size by the linear scaling rule $lr = blr \\times bs / 256$ [21], a 40 epoch learning rate warm-up schedule followed by cosine learning rate decay, and a weight decay of 0.05.", + "bbox": [ + 496, + 503, + 893, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream tasks. After pre-training, we fine-tune our models on a variety of downstream tasks.", + "bbox": [ + 498, + 686, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visually-guided sound source separation. Following [65], we use a mix-and-separate strategy for training. The pretrained model is used to obtain audio-visual representations that are fed to a UNet decoder to generate the separated audio. We use a similar separation strategy by predicting a separation mask that is directly applied to the input mixture SFTF. To compute the separation masks, we deploy a 10-layer UNet decoder, which conditions its predictions on (1) the spatially pooled visual representations and (2) five intermediate audio representations obtained during the encoding stage at layers 1, 3, 6, 9, and 12. Following [65], the", + "bbox": [ + 496, + 719, + 893, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27190", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "predicted separation masks are trained to regress the STFT magnitude ratio between the separated source and the mixture at each time-frequency bin. We evaluate on three different datasets: the MUSIC [65] dataset and two subsets from VGG-Sounds [7]. MUSIC [65] consists of 448 untrimmed YouTube music videos of solos and duets from 11 instrument categories, where we use 358 solo videos for training and 90 solo videos for evaluation. This dataset was slightly smaller than the original MUSIC dataset since some videos are no longer publicly available to be downloaded. We also use two subsets of VGG-Sounds [7]: the VGGSound-Instruments subset [25] which includes 32k video clips of 10s lengths from 36 musical instrument classes, and a broader newly defined \"VGGSound-Music\" subset which includes 40,908 video clips from 49 music categories for training and 1201 clips for testing. As commonly done in the source separation literature, we measure performance using Signal-to-Distortion Ratio (SDR), Signal-to-Artifact Ratio (SAR), and Signal-to-Interference Ratio (SIR).", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio-visual segmentation. We fine-tune our models on AVSBench [67] for the tasks of Single Source Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS). Task S4 aims to predict pixel-level segmentation maps of the visible sound sources, while AVSS further requires the model to recognize the class of the sound source (in addition to the segmentation maps). We use the standard train/val/test split [67] and measure performance using mIoU and F1 scores. Similarly to the source separation task, audio-visual representations, obtained from the pre-trained model, are fed to a UNet decoder to generate the segmentation maps. This decoder conditions its predictions on a globally pooled audio representation and intermediate visual feature maps obtained at layers 1, 3, 6, 9, and 12.", + "bbox": [ + 75, + 380, + 472, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recognition through Linear Probing or Finetuning. We further assess the pre-trained representations on multi-modal recognition tasks by conducting linear probing and finetuning evaluations on the VGGSound-Music, VGGSound-All, and the balanced AudioSet dataset. In all cases, we simply attach three linear classifiers on the average pooled visual, audio, and fusion representations, respectively. The three classification logits are then averaged to obtain the final prediction. As usual, on VGGSound datasets, we measure performance by class accuracy, while on AudioSet (a multi-label dataset) we use the average precision (AP) and Area under the ROC curve (AUC) scores averaged across classes.", + "bbox": [ + 75, + 594, + 472, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to prior work", + "text_level": 1, + "bbox": [ + 76, + 799, + 316, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed method, we extensively compare it to prior work on audio-visual downstream tasks that require a detailed understanding of audio-visual data. Unless otherwise specified, we use our model trained on VGGSound for comparisons to prior work,", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/807c906ab07cd6ffdb81c47ce6cdd8635006b1713a25d5cf44423ef69083b314.jpg", + "table_caption": [ + "Table 1. Sound source separation performance on the MUSIC and VG-GSound datasets." + ], + "table_footnote": [], + "table_body": "
MethodMUSICVGGS-InstrumentsVGGS-Music
SDRSIRSARSDRSIRSARSDRSIRSAR
NMF [57]-0.892.386.28-3.520.786.95-5.060.157.03
RPCA [26]0.784.627.580.362.387.95-1.681.578.26
SoP [65]3.627.959.632.725.679.851.564.5910.15
MP-Net [64]3.988.369.863.056.1210.171.955.0210.59
CCoL [55]5.278.7510.524.076.9310.852.735.8611.03
OneAVM [38]6.2710.6812.355.897.8511.233.676.5311.85
DEEPÄVFUSION7.9512.4112.806.959.5213.235.798.2413.82
", + "bbox": [ + 501, + 114, + 893, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cefcd34699574bc96202a54e085dfac9916e57e29a175462ba9f76aeb35d3329.jpg", + "table_caption": [ + "Table 2. Audio-visual segmentation performance on the AVSBench dataset for the Sound Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS) tasks." + ], + "table_footnote": [], + "table_body": "
MethodAVSBench-S4AVSBench-AVSS
mIoUF1mIoUF1
AVS [67, 68]78.7487.9029.7735.2
LAVISH [34]80.10---
MMVAE [36]81.7490.10--
DEEPAVFUSION89.9492.3452.0558.29
", + "bbox": [ + 553, + 279, + 844, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as this enables us to make the fairest comparisons with the largest number of works that require pre-training.", + "bbox": [ + 496, + 380, + 890, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sound source separation. Table 1 shows the comparison between DEEPAVFUSION and several prior works, including traditional signal processing methods, NMF [57] and RPCA [26], deep learning methods specialized on the source separation task, Sound-of-Pixels [65] and MP-Net [64], and deep learning methods trained for simultaneous localization and separation, CCoL [55] and OneAVM [38]. While all listed deep learning approaches to source separation use (supervised) ImageNet pre-trained visual encoders, our model shows strong performance using completely unsupervised pre-training. Table 1 shows that DEEPAVFUSION outperforms prior works on all metrics. For example, on VGGSound-Music, we outperform the recently proposed OneAVM [38] by 2.12 SDR and 1.71 SIR, which uses supervised pre-training for the vision component and uses source separation as just one of multiple pre-training objectives. Although the different models use different architectures for their audio-visual encoders (which make direct comparison challenging to analyze), the improvements demonstrated on this task indicate that visual source separation can benefit from early-fusion to extract rich representations from fine audio-visual interactions.", + "bbox": [ + 496, + 411, + 893, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio-Visual Segmentation. We assess the effectiveness of DEEPAVFUSION for audio-visual segmentation on the AVS-Bench dataset [67, 68]. We compared the proposed approach to recent state-of-the-art methods (AVS [67], LAVISH [34] and MMVAE [36]) on the S4 and AVSS downstream tasks (sound source segmentation without and with semantics, respectively). We compare against the mIoU and F1 scores shown in the original papers in Table 2. As can be seen, DEEPAVFUSION outperforms all prior work on both tasks", + "bbox": [ + 496, + 731, + 893, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio-visual classification. The above experiments demonstrate the effectiveness of DEEPAVFUSION on audio-visual", + "bbox": [ + 498, + 869, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27191", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ede8d27f6484fe165ef589999017011d21fa79a783625ede05f5ecf6b00b79b1.jpg", + "table_caption": [ + "Table 3. Comparing DEEPAVFUSION trained on VGGSound and AudioSet dataset." + ], + "table_footnote": [], + "table_body": "
Pre-training DatasetPre-training EpochsVGGSound CLSAudioSet CLSVGGSS-Music SEPAVSBench S4
LP AccFt AccLP APFt APSDRSIRSARmIoUF1
VGGSound20053.0858.1930.2532.885.798.2413.8289.9492.34
AudioSet20053.0857.9132.6934.716.939.9313.4990.2792.49
", + "bbox": [ + 189, + 103, + 784, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c387c652b39b1492b8cf7aa5e79b8c3e4f79212c9e3489264c9cd9e12f2316c.jpg", + "table_caption": [ + "Table 4. Audio-visual classification performance on VGGSound-Music and VGGSound-All datasets, using linear probing (LP) and finetuning (FT) protocols. Performance is measured in terms of classification accuracy $[\\%]$ ." + ], + "table_footnote": [], + "table_body": "
MethodVGGS-MusicVGGS-All
LPFTLPFT
(Image) MAE [22]48.2553.1837.1243.86
AudioMAE [28]52.7358.2942.0549.53
CAV-MAE [20]59.1867.5849.8355.32
DEEPAVFUSION65.3272.2553.0858.19
", + "bbox": [ + 130, + 212, + 419, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tasks that require a detailed understanding of audio-visual interactions. We further evaluated the learned representations on recognition tasks. In particular, we compared our model to uni-modal masked reconstruction methods, MAE [22] and AudioMAE [28], as well as to CAV-MAE [20], a late-fusion model trained with both masked reconstruction and contrastive learning objectives. For a fair comparison, we use the same ViT-Base architecture for the uni-modal encoders and tune the model on image-audio pairs obtained from the VGG-Sound dataset.", + "bbox": [ + 75, + 313, + 470, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen in Table 4, we outperform both uni-modal pre-trained encoders on linear probing and fine-tuning by significant margins. DEEPAVFUSION outperforms the image MAE [22] and AudioMAE [28] by $19.62\\%$ and $13.26\\%$ on VGGSound-Music and $19.08\\%$ and $12.49\\%$ on the full VGSSound dataset. These direct comparisons to uni-modal encoders highlight the importance of learning joint representations for multi-modal recognition. Furthermore, our DEEPAVFUSION also significantly outperforms CAV-MAE which optimizes the same audio-visual reconstruction objective of Eq. 14 but does not leverage an early fusion encoder.", + "bbox": [ + 75, + 465, + 468, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It should be noted that multi-modal recognition could be further improved by learning from audio-video data (instead of audio-image pairs). However, video modeling is beyond the scope of this work. The improvements in Table 4 demonstrate nevertheless that even tasks like audio-visual classification of sound sources can benefit from more deeply integrated multi-modal representations.", + "bbox": [ + 75, + 635, + 468, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Pre-training data", + "text_level": 1, + "bbox": [ + 76, + 753, + 246, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparisons above were conducted using VGGSound for pre-training. To study the impact of pre-training data distribution and scale, we also trained our model on the larger AudioSet dataset [18] (with 1.73M training samples). We then tuned the resulting model on several downstream tasks, from classification on both VGGSounds and AudioSet to source separation on VGGSS-Musion and sound source segmentation on AVSBench (S4 task). As can be seen in", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7cdd9a180b213bc9d6fa536f5a9af2483ee59f39a9ad720c5b251e54aa7c0c54.jpg", + "table_caption": [ + "Table 5. Ablation study of major components of DEEPAVFUSION." + ], + "table_footnote": [], + "table_body": "
(#)Unimodal \nPre-TrainingFusionLocal AV \nInteractionsLinear \nAccSep \nSDRSegm \nmIoU
(1)✓Early✓53.086.5352.05
(2)×Early✓44.362.0329.55
(3)✓Early×51.196.1548.71
(4)✓Mid✓51.215.8548.04
(5)✓Late✓46.525.3244.32
(6)✓None×39.674.2338.32
", + "bbox": [ + 514, + 186, + 883, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3, pre-training on AudioSet yields better performance on most downstream tasks, demonstrating the scalability of the proposed method. The only exception is classification on VGGSound, where pre-training on VGGSound yields slightly better performance. This is likely due to a smaller domain gap between pre-training and fine-tuning data.", + "bbox": [ + 496, + 328, + 893, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Understanding DEEPAVFUSION", + "text_level": 1, + "bbox": [ + 500, + 434, + 748, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The previous section demonstrated the efficacy of early fusion of local audio-visual interactions, with the proposed method outperforming the state-of-the-art on a wide range of audio-visual tasks. In this section, we conduct a thorough analysis of our framework, revealing emergent properties in our model, and assessing the benefits of uni-modal pretraining, fusion depth, and other major design choices of the proposed methodology. To provide a full-picture comparison, we evaluate all models on five downstream tasks, namely linear probing and fine-tuning on VGGSound, sound source separation on VGGSound-Music, sound source localization on Flickr-SoundNet, and audio-visual segmentation with semantics on AVSBench.", + "bbox": [ + 496, + 460, + 893, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies. We begin our analysis with a series of ablation studies in Table 5 to measure the impact of three important components of the proposed framework, namely 1) uni-modal pre-training, 2) early fusion, and 3) audio-visual interactions. All models use the same backbone architecture for the uni-modal encoders and are trained to optimize the Audio-Visual MAE objective. Model #1 is the full DEEPAVFUSION framework, and the remaining models relax each of the three components in turn. Model #2 skips uni-modal pre-training, significantly hurting performance when compared to model #1 on all tasks, emphasizing the value of uni-modal pre-training, when learning deeply integrated audio-visual encoders. Model #3 performs early fusion but uses the \"token fusion\" strategy similar to that used in MBT [48]. Token fusion attends to uni-modal representations and cannot directly capture audio-visual inter", + "bbox": [ + 496, + 657, + 895, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27192", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/656e011f347636236edf63343d465e4a1f592414bbafb53498c9100d939d4eab.jpg", + "table_caption": [ + "Table 6. Exploration studies on the efficiency of factorized interactions." + ], + "table_footnote": [], + "table_body": "
Early FusionFactorized InteractionsSpeed (video/sec)Memory (GB)Linear AccSep SDRSegm mIoU
××14.510.346.525.3244.32
✓×4.331.652.826.3052.13
✓✓12.813.153.086.5352.05
", + "bbox": [ + 91, + 104, + 460, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4535a5279b37193eaf522f7851ff2594b05ded8ae7c184a6c07df064991a2def.jpg", + "image_caption": [ + "Figure 3. Nearest neighbor accuracy (NNAcc) vs training steps (Step) using Audio, Image, Fusion Tokens from an DEEPAVFUSION trained on VGGSounds." + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 450, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "actions, limiting the expressivity of the fusion block, and the overall performance of the model. Models #4, #5 and #6 use the same fusion modules with dense interactions but perform mid-fusion (at layers 9-12), late-fusion (at layer 12), and no fusion at all. The results indicate that mid-fusion outperforms, late-fusion which in turn is substantially better than no-fusion. Nevertheless, early fusion achieves the best results across all downstream tasks.", + "bbox": [ + 75, + 412, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency of factorized interactions. While audio-visual interactions are crucial for the performance of early fusion, they also introduce a significant computational overhead, if not handled properly. Table 6 lists the VRAM consumption and model throughput during inference for three models: a baseline model without fusion, an early fusion model with dense interactions, and our proposed early-fusion model with factorized interactions. Dense interactions enhance the model's effectiveness on subsequent tasks, but at a high computational cost, both in terms of throughput (3.3 times slower) and memory usage (3.1 times higher). Our proposed factorized interactions significantly cut down the computational overhead, while maintaining (and even slightly enhancing) the performance of early fusion with dense interactions.", + "bbox": [ + 75, + 535, + 470, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emergent semantics. In addition to robust performance across various tasks, the emergent semantics in fusion tokens are worth highlighting. Despite the masked auto-encoder objective typically favoring representations rich in low-level details, fusion tokens surprisingly yield higher-level semantic representations. Fig. 3 shows the nearest neighbor retrieval performance of our DEEPAVFUSION using either the learned visual, audio, or fusion tokens as queries. As can be seen, the retrieval using fusion tokens is substantially better than uni-modal representations (i.e. better aligned with the", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e9326675f2c00064cc207bf9d1f866e64b5e8c3789fddd5e3a4dea826b60e68.jpg", + "table_caption": [ + "Table 7. Impact of early fusion and fusion tokens." + ], + "table_footnote": [], + "table_body": "
#Linear AccSep SDRSegm mIoU
1253.086.5352.05
952.766.1549.93
652.296.1249.02
351.215.8548.04
146.525.3244.32
039.674.2338.32
", + "bbox": [ + 521, + 103, + 686, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(a) # Fusion layers", + "bbox": [ + 549, + 232, + 656, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c22cbb41f839dd342419d945e9ae2269b66b2d803db5239024034421d18dd1c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Linear AccSep SDRSegm mIoU
151.865.9450.67
851.975.8351.14
1653.086.5352.05
3252.186.4350.60
", + "bbox": [ + 707, + 116, + 870, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/481c11b56f0626fef493f4d525ff041c288293d94d78cb482b68b4082e0d8705.jpg", + "table_caption": [ + "(b) # Fusion tokens" + ], + "table_footnote": [], + "table_body": "
#Linear AccSep SDRSegm mIoU
451.785.6849.59
853.086.5352.05
1652.256.1650.89
", + "bbox": [ + 614, + 248, + 777, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) # Aggr tokens", + "bbox": [ + 647, + 334, + 745, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "semantics of the dataset). This suggests that fusion tokens aggregate information that is indicative of high-level semantics, while uni-modal representations encode the low-level details required for masked reconstruction.", + "bbox": [ + 496, + 359, + 893, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of miscellaneous design choices. Lastly, we examine the effects of the number of fusion layers, fusion tokens and aggregation tokens in Table 7. These results reveal that early fusion is crucial for optimal performance, with model performance improving as fusion depth increases. Furthermore, DEEPAVFUSION can benefit from a relatively large number of tokens, with the performance saturating after 16 fusion tokens and 8 aggregation tokens.", + "bbox": [ + 496, + 421, + 893, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 617, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present DEEPAVFUSION, a simple yet effective early fusion approach with dense interactions that achieves efficient audio-visual pre-training for audio-visual masked auto-encoders. We introduce learnable fusion tokens to aggregate modality-specific information with early fusion from each transformer block of audio and visual encoders, where two variants of parallel and sequential flows are proposed to achieve early fusion between fusion tokens and audio-visual patch tokens. Furthermore, we leverage multi-modal blocks with dense interactions to fuse fusion tokens and patches across audio-visual transformer blocks and achieve efficient downstream fine-tuning with factorized attention blocks. Empirical experiments on Flickr-SoundNet, VGG-Instruments, VGG-Music, VGGSound-All, and AVS-Bench datasets demonstrate the state-of-the-art performance of our DEEPAVFUSION in linear probing, fine-tuning classification, visual sound localization, sound separation, and audio-visual segmentation.", + "bbox": [ + 496, + 585, + 893, + 858 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27193", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 609-617, 2017. 1, 2", + "[2] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2016. 2", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 2, 4, 5", + "[4] Changan Chen, Unnat Jain, Carl Schissler, S. V. A. Garí, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Proceedings of European Conference on Computer Vision (ECCV), pages 17–36, 2020. 2", + "[5] Changan Chen, Sagnik Majumder, Al-Halah Ziad, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In Proceedings of International Conference on Learning Representations (ICLR), 2021.", + "[6] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. 2", + "[7] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 5, 6", + "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2", + "[9] Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2019. 2", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia. Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 5", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[12] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for BERT pre-training of vision transformers. In Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, 2023. 2", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 2", + "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 2", + "[15] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[16] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10478-10487, 2020. 2", + "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 324-333, 2019. 2", + "[18] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In Proceedings of 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780, 2017. 5, 7", + "[19] Mariana-Iuliana Georgescu, Eduardo Fonseca, Radu Tudor Ionescu, Mario Lucic, Cordelia Schmid, and Anurag Arnab. Audiovisual masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16144-16154, 2023. 2, 3", + "[20] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In Proceedings of The Eleventh International Conference on Learning Representations (ICLR), 2023. 1, 2, 5, 7", + "[21] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7", + "[23] John Hershey and Michael Casey. Audio-visual sound separation via hidden markov models. Advances in Neural Information Processing Systems, 14, 2001. 2", + "[24] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 2", + "[25] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 6" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27194", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Po-Sen Huang, Scott Deann Chen, Paris Smaragdis, and Mark Hasegawa-Johnson. Singing-voice separation from monaural recordings using robust principal component analysis. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 57–60, 2012. 6", + "[27] Po-Yao Huang, Vasu Sharma, Hu Xu, Chaitanya Ryali, Haoqi Fan, Yanghao Li, Shang-Wen Li, Gargi Ghosh, Jitendra Malik, and Christoph Feichtenhofer. Mavil: Masked audio-video learners. arXiv preprint arXiv:2212.08071, 2022. 1, 2, 3, 5", + "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 5, 7", + "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[30] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[31] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2", + "[32] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002–2006, 2019. 2", + "[33] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[34] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6", + "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2", + "[36] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 954–965, 2023. 6", + "[37] Shentong Mo and Pedro Morgado. Benchmarking weakly-supervised audio-visual sound localization. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2", + "[38] Shentong Mo and Pedro Morgado. A unified audio-visual learning framework for localization, separation, and recognition. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6", + "[39] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[40] Shentong Mo and Yapeng Tian. Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2", + "[41] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056, 2023. 2", + "[42] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2", + "[43] Shentong Mo, Jing Shi, and Yapeng Tian. DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903, 2023. 2", + "[44] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[45] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4733–4744, 2020. 2", + "[46] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12934–12945, 2021.", + "[47] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12475-12486, 2021. 1, 2", + "[48] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 7", + "[49] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3", + "[50] Andrew Owens, Jiajun Wu, Josh H. McDermott, William T. Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-816, 2016. 2", + "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 2", + "[52] Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223, 2019. 2", + "[53] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27195", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proceedings of European Conference on Computer Vision (ECCV), page 436–454, 2020. 2", + "[55] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2745-2754, 2021. 6", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), page 5998-6008, 2017. 2, 4", + "[57] Tuomas Virtanen. Monaural sound source separation by nonnegative matrix factorization with temporal continuity and sparseness criteria. IEEE Transactions on Audio, Speech, and Language Processing, 15(3):1066-1074, 2007. 6", + "[58] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan L. Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[59] Alexander Wettig, Tianyu Gao, Zexuan Zhong, and Danqi Chen. Should you mask $15\\%$ in masked language modeling? In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2985-3000, 2023. 2", + "[60] Jiantao Wu and Shentong Mo. Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338, 2022. 2", + "[61] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1326–1335, 2021. 2", + "[62] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 6291–6299, 2019. 2", + "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[64] Xudong Xu, Bo Dai, and Dahua Lin. Recursive visual sound separation using minus-plus net. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6", + "[65] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 2, 5, 6", + "[66] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1735-1744, 2019. 2", + "[67] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang, and Yiran Zhong. Audio-visual segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 6", + "[68] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 6" + ], + "bbox": [ + 503, + 92, + 893, + 202 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27196", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_model.json b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..753fcdbea333fc7d1b5611ec051acbc38d826fa0 --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.205, + 0.382, + 0.222 + ], + "angle": 0, + "content": "Shentong Mo" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.224, + 0.438, + 0.242 + ], + "angle": 0, + "content": "Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.244, + 0.456, + 0.26 + ], + "angle": 0, + "content": "shentong.mo@sv.cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.205, + 0.705, + 0.222 + ], + "angle": 0, + "content": "Pedro Morgado" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.224, + 0.776, + 0.241 + ], + "angle": 0, + "content": "University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.244, + 0.743, + 0.26 + ], + "angle": 0, + "content": "pmorgado@wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.29, + 0.714, + 0.307 + ], + "angle": 0, + "content": "https://github.com/stoneMo/DeepAVFusion" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.335, + 0.314, + 0.352 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.474, + 0.78 + ], + "angle": 0, + "content": "Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.813, + 0.21, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Humans naturally integrate audio-visual information to perceive and understand the environment. Several studies in cognitive psychology and neuroscience, as well as classic perceptual illusions such as the McGurk effect, demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.896, + 0.565 + ], + "angle": 0, + "content": "that such audio-visual fusion can occur early on in the perceptual stack, enabling deeper integration of the two modalities. Early fusion models aim to emulate this human-like perception. Specifically, it refers to the process of integrating auditory and visual cues at an early stage of a multi-modal perception model, in order to leverage the synergistic effects of both modalities. This can be especially powerful if the fusion process can attend to and establish connections between local components of the visual and audio signal (e.g., the frequencies of someone's voice with the pixels of their lips). In sum, early fusion of local audio-visual interactions holds the promise of a deeper and more sophisticated understanding of audio-visual content, critical for many real-world applications such as visually guided source separation, localization/segmentation, and multi-modal recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.897, + 0.793 + ], + "angle": 0, + "content": "Despite the potential, audio-visual early fusion architectures have remained under-explored, with the majority of the recent literature focusing on late fusion or no fusion at all. One recent prominent research direction [1, 47] seeks to learn independent uni-modal encoders through contrastive learning by requiring the representations of associated audio-visual pairs to be synchronized in latent space. However, contrastive learning is not compatible with early fusion, as the connections between uni-modal representations lead to trivial solutions to the contrastive learning problem. Another line of recent work [20, 27] learns encoders with late audio-visual fusion through a combination of contrastive learning and masked reconstruction, focusing on downstream tasks that require shallow integration of audio-visual features, such as audio-visual action recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.903 + ], + "angle": 0, + "content": "One potential reason is that early fusion architectures pose a significant training challenge due to the substantial increase in model expressivity it offers. As a result, robust learning frameworks are required to ensure the creation of high-quality models and effectively harness their enhanced capabilities. In this work, we demonstrate the effectiveness of a masked reconstruction framework for training audio-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.268, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.098, + 0.465, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 1. Unlike prior works on Audio-Visual MAEs which learn either late fusion encoders or separate encoders altogether, we demonstrate that masked reconstruction is especially useful for learning deeply integrated audio-visual encoders with powerful early fusion modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.597 + ], + "angle": 0, + "content": "visual transformers with early-fusion of local interactions (Fig. 1). Our hypothesis is that the simultaneous reconstruction of audio and visual inputs from a limited audio-visual context promotes the learning of deeply integrated representations, as the model is required to understand the fine interactions between the two modalities. Furthermore, masked reconstruction has already been successfully deployed in training uni-modal representations for a variety of input signals, including text [11], images [3, 22], and audio [28], as well as for learning late fusion representations [20, 27]. However, unlike what was previously observed with uni-modal or late fusion models, our work shows that masked audio-visual reconstruction with early fusion encoders results in an interesting emergent property. While uni-modal representations encode the low-level details required for masked reconstruction, we observed that high-level semantics (surprisingly) emerged from the tokens used for audio-visual fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In addition to demonstrating the effectiveness of masked reconstruction for early fusion transformers, we propose a novel attention-based fusion module, which can effectively attend to the interactions between local audio and visual representations, thereby enhancing its capacity to capture localized interactions between the two modalities. Take, for instance, the sound of a dog barking. The sound signals the presence of a dog and offers clues about its state, such as an open mouth. While the link between the dog and its bark can be made at a global level, our fusion module allows the model to delve deeper into local interactions, like the correlation between the dog's mouth and the barking sound's spectral frequencies. As a result, localized fusion boosts the quality of learned representations for numerous tasks. However, it necessitates the model to explicitly represent all pairwise interactions of local audio and visual tokens. Given the vast number of tokens that current transformers process, dense local interactions can quickly become unmanageable. To enhance efficiency, our fusion module factorizes unimodal representations, thereby circumventing the need to" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.71, + 0.106 + ], + "angle": 0, + "content": "model all pairwise interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.109, + 0.896, + 0.321 + ], + "angle": 0, + "content": "In sum, we propose an attention-based fusion module that can attend to local audio-visual interactions and deploy it to learn deeply integrated audio-visual representations by fusing uni-modal representations at early layers in the architecture. We also demonstrate that when paired with masked reconstruction, our framework, denoted DEEPAVFUSION, can learn strong audio-visual representations, yielding state-of-the-art performance on a variety of audio-visual tasks, including visually guided source separation, localization, and segmentation. We extensively evaluated DEEPAVFUSION on a variety of audio-visual datasets, and conducted thorough ablation studies, showing the importance of early fusion, dense interactions, and uni-modal pre-training when learning deeply integrated audio-visual representations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.348, + 0.642, + 0.364 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.549 + ], + "angle": 0, + "content": "Audio-Visual Representations Learning. Audio-visual representations learning has been addressed in many previous works [1, 2, 14, 16, 23, 24, 30, 40, 43, 45–47, 50, 51, 65, 66] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for many audio-visual tasks, such as audio-event localization [31, 32, 37, 41, 42, 53, 62], audio-visual spatialization [4, 17, 44, 45], audio-visual navigation [4–6] and audio-visual parsing [33, 39, 54, 61]. In this work, our main focus is to learn transferrable audio-visual representations from masked audio-visual reconstruction, which is more challenging than the above-mentioned tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.897, + 0.809 + ], + "angle": 0, + "content": "Masked Representations Learning. Masked representation learning aims to learn self-supervised representations by reconstructing desired features of masked data given unmasked parts as clues. In the recent years, masked representation learning has achieved promising results in natural language processing [9, 11, 35, 52, 59] and computer vision [3, 8, 12, 15, 22, 58, 60, 63] community. Typically, BERT [11] randomly masked \\(15\\%\\) of word tokens in the sentence and recovered them with unmasked words to learn generalizable textual features via a self-attention transformer [56]. A block-wise masking strategy was proposed in BEiT [3] to reconstruct discrete tokens of masked image patches for pre-training transferrable visual representations. To simplify the masked image encoding framework, MAE [22] directly reconstructed missing pixels of \\(75\\%\\) masked patches using vision transformers [13] for self-supervised pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Audio-visual Masked Autoencoders. More recently, researchers introduced diverse masking pipelines [19, 20, 27] to show the effectiveness of masked modeling in learning audio-visual representations. For example, CAV-MAE [20] combined contrastive learning and masked modeling to capture a joint and coordinated audio-visual representation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27187" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "They tried to add a joint encoder on audio-visual features from the last attention block of single-modality encoders to fuse cross-modal information for audio-visual contrastive objectives. MAViL [27] extended masked audio-video reconstruction with masked intra- and inter-modal contrastive learning and self-training by recovering joint audio-video contextualized representations. Despite their promising performance, they ignored the importance of the early fusion of audio-visual features in masked audio-visual reconstruction. In contrast to AV-MAE [19] that simply concatenated the audio and visual tokens before passing them through the joint transformer for early fusion, we will design an early fusion module with interactions between local and visual representations for audio-visual masked auto-encoders. Our audio-visual interactions are different from fusion bottlenecks in MBT [48] that forces information between different modalities to pass through a small number of bottleneck latent. However, we develop a fully novel attention-based fusion module that can effectively attend to interactions between local audio and visual representations, thus enhancing its ability to capture fine-grained interactions between the two modalities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Audio-visual Early Fusion. Audio-visual fusion has been the topic of extensive research, with several works proposing a variety of architectures to aggregate multi-modal representations [19, 48, 49]. An example of an early fusion architecture was proposed by Owens et al. [49] to learn representations from audio-visual correspondences, by concatenating features from small unimodal encoders and feeding them into a fused audio-visual network for the early fusion of multisensory features. Beyond this early work, most recent papers on audio-visual fusion [19, 48] propose alternative fusion mechanisms, either through weight sharing [19] or token-based fusion [48]. However, since all these works focused on downstream tasks (mostly classification) that do not require fine multi-modal understanding, the benefits of early fusion were not evident, with mid-level fusion often achieving optimal performance. Different from them, we proposed an early fusion model that is designed for fine multi-modal understanding, by attending to local interactions between the audio and the visual while performing early fusion. The new architecture design paired with masked reconstruction objective is shown to outperform many of the prior works on a variety of downstream multi-modal applications beyond classification." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.797, + 0.17, + 0.814 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.472, + 0.901 + ], + "angle": 0, + "content": "We aim to learn audio-visual representations that can be effectively transferred to a variety of downstream tasks that require a detailed understanding of audio-visual interactions. To accomplish this, we introduce a novel audio-visual transformer architecture, named DEEPAVFUSION, that enables the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "early fusion of audio and visual tokens through dense local interactions. We also show how to effectively train such early fusion transformers using multi-modal masked reconstruction. In this section, we first describe the proposed early fusion transformer and then introduce the multi-modal masked reconstruction framework used for self-supervised pre-training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.207, + 0.895, + 0.238 + ], + "angle": 0, + "content": "3.1. A Transformer Architecture for Joint Audio-Visual Encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Downstream tasks such as visually guided sound source separation, localization, and segmentation require a deep understanding of fine audio-visual interactions present in the data. However, late fusion often lacks the expressive power to represent such interactions. We introduce a transformer architecture that enables the early fusion of multi-modal representations through factorized local interactions between audio and visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.897, + 0.646 + ], + "angle": 0, + "content": "Early fusion architecture While we aim to learn deeply integrated audio-visual representations, foundational uni-modal transformers can still provide a strong starting point for multi-modal learning. As such, we introduce a modular architecture to repurpose existing pre-trained uni-modal transformers. Specifically, we propose a three-branch architecture, illustrated in Fig. 2. The first two branches are uni-modal transformers that process audio and visual tokens, respectively, while the third branch is a multi-modal transformer that updates a set of learnable fusion tokens to fuse audio-visual information. Formally, let \\( \\mathbf{X}_0^f \\in \\Re^{F \\times D} \\) be \\( F \\) learnable fusion tokens. We progressively update these tokens, at each layer \\( l \\), through a fusion block \\( \\Phi_l^f \\), which can interact with and aggregate information from modality-specific features \\( \\mathbf{X}_{l - 1}^v \\), \\( \\mathbf{X}_{l - 1}^a \\). Furthermore, fusion tokens \\( \\mathbf{X}_{l - 1}^f \\) are fed into modality-specific blocks, \\( \\Phi_l^v \\) and \\( \\Phi_l^a \\), to modulate uni-modal representations from early stages in the network. Overall, representations are updated as follows" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.649, + 0.892, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {l} ^ {f} = \\Phi_ {l} ^ {f} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {a}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.672, + 0.892, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {l} ^ {v} = \\Phi_ {l} ^ {v} \\left(\\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.694, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {l} ^ {a} = \\Phi_ {l} ^ {a} \\left(\\boldsymbol {X} _ {l - 1} ^ {a}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.749 + ], + "angle": 0, + "content": "This token update strategy is repeated for a total of \\(L\\) layers to compute the final representations \\(\\mathbf{X}_L^v, \\mathbf{X}_L^a\\) and \\(\\mathbf{X}_L^f\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Modality-specific blocks We use standard transformer self-attention for the modality-specific blocks, \\(\\Phi^v\\) and \\(\\Phi^a\\), while allowing them to further attend to fusion tokens. Unless otherwise specified, we initialize \\(\\Phi^v\\) and \\(\\Phi^a\\) from foundational uni-modal models. This is possible since the parameter space of the modality-specific blocks remains unchanged. In particular, we empirically validated the benefits of bootstrapping from [28] and [22], which are pre-trained on large-scale audio and visual datasets, respectively, through a uni-modal masked reconstruction objective." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.399, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.099, + 0.871, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.894, + 0.313 + ], + "angle": 0, + "content": "Figure 2. Audio-Visual Early Fusion Transformer with Dense Interactions. The architecture ① is composed of three interconnected branches - the visual, audio, and fusion branches - each with an equal number of transformer blocks. The visual and audio branches process visual and audio tokens, respectively, while simultaneously attending to fusion tokens. The fusion branch updates a set of learnable fusion tokens to fuse audio-visual information. To better attend to local audio-visual interactions, we introduce a new fusion block ②. At each layer \\(l\\), all patch-level audio (visual) tokens, \\(\\bar{X}_{l-1}^{a}(\\bar{X}_{l-1}^{v})\\), are first aggregated into a small set of aggregation tokens, \\(\\bar{X}_l^a (\\bar{X}_l^v)\\) by cross-attention. Pairwise audio-visual interactions are then encoded through a linear layer into latents \\(\\bar{X}_l^{av}\\), which are used to update the fusion tokens \\(\\bar{X}_l^f\\) by cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Fusion blocks The design of the fusion blocks \\(\\Phi^f\\) is critical to representing fine audio-visual interactions. In the context of transformers, a natural way to combine audio-visual representations is through a standard self-attention block which, in addition to the fusion tokens \\(X^f\\) themselves, it can further attend to all image and audio tokens, \\(X_{l - 1}^v\\) and \\(X_{l - 1}^{a}\\). This approach, which we refer to as token fusion, has been used in previous works such as MBT [48]. While intuitive, standard self-attention cannot directly exploit interactions between local audio-visual representations, which limits their expressivity. To address this limitation, we designed a fusion block that attends to local interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.473, + 0.671 + ], + "angle": 0, + "content": "Fusion with Dense Audio-Visual Interactions Interactions between audio and visual data occur between local regions in the image and the audio spectrogram. Given a set of \\( n_a \\) audio tokens \\( X^a \\) and \\( n_v \\) visual tokens \\( X^v \\), aggregation of local multi-modal interactions requires attending to representations from all possible \\( n_a \\times n_v \\) pairs, to identify the ones that are most useful for the task at hand. While we experimented with several ways to represent multi-modal interactions, from bilinear aggregation to kernel-based aggregation, we found that linear aggregation was sufficient to capture audio-visual interactions for our tasks. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.681, + 0.47, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {i j} ^ {a v} = W ^ {a} \\boldsymbol {X} _ {i} ^ {a} + W ^ {v} \\boldsymbol {X} _ {j} ^ {v} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.702, + 0.47, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {a v} = \\left[ \\boldsymbol {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {a} \\forall j = 1, \\dots , n _ {v} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.472, + 0.779 + ], + "angle": 0, + "content": "Fusion tokens \\( X_{l}^{f} \\) at layer \\( l \\) are then updated using standard cross-attention [56] over the representation of all local interactions \\( X^{av} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.788, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}; \\boldsymbol {X} _ {l - 1} ^ {\\text {a v}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Factorized Audio-Visual Interactions While the procedure above can attend to local interactions, it requires the model to explicitly represent all possible \\( n_a \\times n_v \\) pairs. Given that the number of audio and visual tokens is typically large, the number of interactions can quickly become intractable," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.895, + 0.427 + ], + "angle": 0, + "content": "resulting in high memory consumption and low throughput. To address this limitation, we introduce modality-specific aggregation tokens updated by aggregation blocks. The goal of the aggregation blocks is to summarize the large number of audio/visual tokens \\(X^a / X^v\\) into a small set of \\(n_{\\bar{a}} / n_{\\bar{v}}\\) aggregation tokens \\(\\bar{X}^a / \\bar{X}^v\\), respectively. Specifically, at layer \\(l\\), aggregation tokens are updated by cross-attention" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.436, + 0.892, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\bar {X} _ {l} ^ {a} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {a}, X _ {l - 1} ^ {a}\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.457, + 0.892, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\bar {X} _ {l} ^ {v} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {v}, X _ {l - 1} ^ {v}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.547 + ], + "angle": 0, + "content": "Then, instead of explicitly representing all \\(n_a \\times n_v\\) pairwise interactions, we restrict audio-visual interactions to the sets of aggregated tokens. Formally, factorized interactions between aggregated tokens \\(\\bar{X}_i^a\\) and \\(\\bar{X}_j^v\\) are represented as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.557, + 0.892, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\bar {X} _ {i j} ^ {a v} = W ^ {\\bar {a}} \\bar {X} _ {i} ^ {a} + W ^ {\\bar {v}} \\bar {X} _ {j} ^ {v} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.578, + 0.892, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\bar {X} ^ {a v} = \\left[ \\bar {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {\\bar {a}} \\forall j = 1, \\dots , n _ {\\bar {v}} \\right] \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.607, + 0.807, + 0.622 + ], + "angle": 0, + "content": "and the fusion tokens are updated at layer \\(l\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.633, + 0.892, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\bar {\\boldsymbol {X}} _ {l} ^ {a v}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.894, + 0.728 + ], + "angle": 0, + "content": "Since we only represent interactions between the aggregated uni-modal tokens, the number of possible interactions is greatly reduced. We observed that a \\(700 \\times\\) reduction still maintains the benefits obtained with local dense interactions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.892, + 0.771 + ], + "angle": 0, + "content": "3.2. Learning Early-Fusion Transformers Through Audio-Visual Masked Auto-Encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Masked auto-encoders have been shown to learn effective representations for a wide variety of input signals, including images [3, 22] and audio [28]. Formally, the input signal \\( x \\) is partitioned into a set of patch tokens \\( x = \\{x_{t}\\}_{t=1}^{T} \\), which is randomly split into the visible set \\( X_{\\mathrm{vis}} = \\{x_{t}\\}_{t \\in \\mathcal{V}} \\) and the masked set \\( X_{\\mathrm{mask}} = \\{x_{t}\\}_{t \\in \\mathcal{M}} \\) (\\( \\mathcal{V} \\) and \\( \\mathcal{M} \\) are non-overlapping index sets). In the case of images, visible and masked tokens, \\( X_{\\mathrm{vis}} \\) and \\( X_{\\mathrm{mask}} \\), are obtained from a grid" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27189" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "of small image patches [3, 22], while in the case of audio signals, tokens are obtained either from small windows of a raw waveform or from small 2D patches of the audio spectrogram [28]. Then, to learn representations of the input data, a transformer encoder \\( f \\) is first used to encode the visible tokens \\( X_{\\mathrm{vis}} = f(X_{\\mathrm{vis}}) \\in \\Re^{|\\mathcal{V}| \\times d} \\), where \\( d \\) denotes the dimensionality of the token representation. These representations together with a series of (position marked) mask tokens \\( M_{\\mathrm{mask}} \\) are then fed to a decoder \\( g(\\cdot) \\) that seeks to reconstruct the masked input tokens \\( \\hat{X}_{\\mathrm{mask}} = g(X_{\\mathrm{vis}}, M_{\\mathrm{mask}}) \\). To accomplish that, the model is trained to minimize the mean squared reconstruction error" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.279, + 0.471, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}}, \\hat {X} _ {\\text {m a s k}}\\right) = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\substack {x \\in X _ {\\text {m a s k}} \\\\ \\hat {x} \\in \\hat {X} _ {\\text {m a s k}}}} \\| x - \\hat {x} \\| _ {2} ^ {2}. \\tag{12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.413 + ], + "angle": 0, + "content": "Following prior work [20, 27], masked reconstruction can be easily extended to multi-modal signals, by simultaneously masking tokens from both audio and visual modalities. In the case of early-fusion transformers, the encoder \\( f \\) is then used to encode the visible tokens, \\( X_{\\mathrm{vis}}^v \\) and \\( X_{\\mathrm{vis}}^a \\)," + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.418, + 0.47, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}, \\boldsymbol {X} ^ {f} = f \\left(\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.443, + 0.473, + 0.534 + ], + "angle": 0, + "content": "which, in addition to the uni-modal representations, \\(X_{\\mathrm{vis}}^v\\) and \\(X_{\\mathrm{vis}}^a\\), also returns a set of fusion representations \\(\\pmb{X}^{f}\\). These representations are then fed to modality-specific decoders to reconstruct both audio and visual masked tokens, \\(\\hat{X}_{\\mathrm{mask}}^v = g_v(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^v)\\) and \\(\\hat{X}_{\\mathrm{mask}}^a = g_a(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^a)\\), by minimizing the loss" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.54, + 0.47, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A V - M A E}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {v}, \\hat {X} _ {\\text {m a s k}} ^ {v}\\right) + \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {a}, \\hat {X} _ {\\text {m a s k}} ^ {a}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.579, + 0.389, + 0.594 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{MAE}}\\) is the auto-encoder loss of Eq. 12." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.471, + 0.747 + ], + "angle": 0, + "content": "While multi-modal masked reconstruction has been previously explored in [20, 27], these approaches focus on learning separate audio and image encoders, \\( f_{a} \\) and \\( f_{v} \\), with multi-modal representations obtained through a late fusion module \\( f_{f} \\). In contrast, we propose to learn a single multimodal early fusion encoder \\( f \\), and demonstrate that masked reconstruction is especially useful for learning multi-modal representations capable of representing fine audio-visual interactions through early fusion of local interactions between audio and visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.21, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.471, + 0.835 + ], + "angle": 0, + "content": "We now evaluate the proposed DEEPAVFUSION on a variety of audio-visual downstream tasks that require a deep understanding of audio-visual interactions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.292, + 0.861 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Pre-training. For pre-training, we experiment with two datasets, namely VGG-Sounds [7] (with 144k training sam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "ples) and AudioSet [18] (with 1.73M training samples), both of them containing diverse videos focused on audio events and sound sources. The audio is represented by log mel spectrograms extracted from \\(3s\\) of audio at a sample rate of \\(16000\\mathrm{Hz}\\). To compute the log spectrograms, we apply an STFT using approximately 50ms windows with a hop size of \\(15\\mathrm{ms}\\), resulting in an input tensor of size \\(128\\times 196\\) (128 mel frequency bands over 196 time steps). Since we use image models for the visual component, we randomly extract single frames from within the time window defined by the \\(3s\\) audio snippets. Input frames are augmented by random crops with a minimum area of 0.5 and random horizontal flips, and resized into a \\(224\\times 224\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.895, + 0.502 + ], + "angle": 0, + "content": "Unless otherwise specified, we initialize the uni-modal transformer blocks using modality-specific encoders pretrained with uni-modal masked reconstruction objectives [22, 28]. Specifically, we use the ViT-Base model from [22] pre-trained on ImageNet [10], and the Spec-MAE model (with the same ViT-Base architecture) pre-trained on AudioSet [18]. In our base configuration, we add multi-modal fusion blocks on all 12 layers of the model, together with 16 multi-modal fusion tokens, 8 visual aggregation tokens, and 8 audio aggregation tokens. For efficiency, we reduced the dimensionality of the MLP embeddings using an MLP ratio of 1 (instead of the standard 4), as well as the dimensionality of space in which the similarity for self and cross-attention is computed to 16 (as opposed to the standard 64)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.686 + ], + "angle": 0, + "content": "The models were trained using the Adam optimizer [29]. Ablation and parametric studies were conducted on VG-GSound, training for 200 epochs with a total batch size of 512 (split between 2 GPUs and 4 iterations of gradient accumulation). Large-scale training on AudioSet was conducted for 200 epochs with a total batch size of 2048 (split between 16 GPUs and 4 iterations of gradient accumulation). In all cases, we use a base learning rate of \\(1.5e - 4\\) adjusted for the effective batch size by the linear scaling rule \\(lr = blr \\times bs / 256\\) [21], a 40 epoch learning rate warm-up schedule followed by cosine learning rate decay, and a weight decay of 0.05." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Downstream tasks. After pre-training, we fine-tune our models on a variety of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.887 + ], + "angle": 0, + "content": "Visually-guided sound source separation. Following [65], we use a mix-and-separate strategy for training. The pretrained model is used to obtain audio-visual representations that are fed to a UNet decoder to generate the separated audio. We use a similar separation strategy by predicting a separation mask that is directly applied to the input mixture SFTF. To compute the separation masks, we deploy a 10-layer UNet decoder, which conditions its predictions on (1) the spatially pooled visual representations and (2) five intermediate audio representations obtained during the encoding stage at layers 1, 3, 6, 9, and 12. Following [65], the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.379 + ], + "angle": 0, + "content": "predicted separation masks are trained to regress the STFT magnitude ratio between the separated source and the mixture at each time-frequency bin. We evaluate on three different datasets: the MUSIC [65] dataset and two subsets from VGG-Sounds [7]. MUSIC [65] consists of 448 untrimmed YouTube music videos of solos and duets from 11 instrument categories, where we use 358 solo videos for training and 90 solo videos for evaluation. This dataset was slightly smaller than the original MUSIC dataset since some videos are no longer publicly available to be downloaded. We also use two subsets of VGG-Sounds [7]: the VGGSound-Instruments subset [25] which includes 32k video clips of 10s lengths from 36 musical instrument classes, and a broader newly defined \"VGGSound-Music\" subset which includes 40,908 video clips from 49 music categories for training and 1201 clips for testing. As commonly done in the source separation literature, we measure performance using Signal-to-Distortion Ratio (SDR), Signal-to-Artifact Ratio (SAR), and Signal-to-Interference Ratio (SIR)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.593 + ], + "angle": 0, + "content": "Audio-visual segmentation. We fine-tune our models on AVSBench [67] for the tasks of Single Source Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS). Task S4 aims to predict pixel-level segmentation maps of the visible sound sources, while AVSS further requires the model to recognize the class of the sound source (in addition to the segmentation maps). We use the standard train/val/test split [67] and measure performance using mIoU and F1 scores. Similarly to the source separation task, audio-visual representations, obtained from the pre-trained model, are fed to a UNet decoder to generate the segmentation maps. This decoder conditions its predictions on a globally pooled audio representation and intermediate visual feature maps obtained at layers 1, 3, 6, 9, and 12." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Recognition through Linear Probing or Finetuning. We further assess the pre-trained representations on multi-modal recognition tasks by conducting linear probing and finetuning evaluations on the VGGSound-Music, VGGSound-All, and the balanced AudioSet dataset. In all cases, we simply attach three linear classifiers on the average pooled visual, audio, and fusion representations, respectively. The three classification logits are then averaged to obtain the final prediction. As usual, on VGGSound datasets, we measure performance by class accuracy, while on AudioSet (a multi-label dataset) we use the average precision (AP) and Area under the ROC curve (AUC) scores averaged across classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.317, + 0.816 + ], + "angle": 0, + "content": "4.2. Comparison to prior work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed method, we extensively compare it to prior work on audio-visual downstream tasks that require a detailed understanding of audio-visual data. Unless otherwise specified, we use our model trained on VGGSound for comparisons to prior work," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.113 + ], + "angle": 0, + "content": "Table 1. Sound source separation performance on the MUSIC and VG-GSound datasets." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.115, + 0.895, + 0.23 + ], + "angle": 0, + "content": "
MethodMUSICVGGS-InstrumentsVGGS-Music
SDRSIRSARSDRSIRSARSDRSIRSAR
NMF [57]-0.892.386.28-3.520.786.95-5.060.157.03
RPCA [26]0.784.627.580.362.387.95-1.681.578.26
SoP [65]3.627.959.632.725.679.851.564.5910.15
MP-Net [64]3.988.369.863.056.1210.171.955.0210.59
CCoL [55]5.278.7510.524.076.9310.852.735.8611.03
OneAVM [38]6.2710.6812.355.897.8511.233.676.5311.85
DEEPÄVFUSION7.9512.4112.806.959.5213.235.798.2413.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.242, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Table 2. Audio-visual segmentation performance on the AVSBench dataset for the Sound Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS) tasks." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.28, + 0.846, + 0.374 + ], + "angle": 0, + "content": "
MethodAVSBench-S4AVSBench-AVSS
mIoUF1mIoUF1
AVS [67, 68]78.7487.9029.7735.2
LAVISH [34]80.10---
MMVAE [36]81.7490.10--
DEEPAVFUSION89.9492.3452.0558.29
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.41 + ], + "angle": 0, + "content": "as this enables us to make the fairest comparisons with the largest number of works that require pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.73 + ], + "angle": 0, + "content": "Sound source separation. Table 1 shows the comparison between DEEPAVFUSION and several prior works, including traditional signal processing methods, NMF [57] and RPCA [26], deep learning methods specialized on the source separation task, Sound-of-Pixels [65] and MP-Net [64], and deep learning methods trained for simultaneous localization and separation, CCoL [55] and OneAVM [38]. While all listed deep learning approaches to source separation use (supervised) ImageNet pre-trained visual encoders, our model shows strong performance using completely unsupervised pre-training. Table 1 shows that DEEPAVFUSION outperforms prior works on all metrics. For example, on VGGSound-Music, we outperform the recently proposed OneAVM [38] by 2.12 SDR and 1.71 SIR, which uses supervised pre-training for the vision component and uses source separation as just one of multiple pre-training objectives. Although the different models use different architectures for their audio-visual encoders (which make direct comparison challenging to analyze), the improvements demonstrated on this task indicate that visual source separation can benefit from early-fusion to extract rich representations from fine audio-visual interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Audio-Visual Segmentation. We assess the effectiveness of DEEPAVFUSION for audio-visual segmentation on the AVS-Bench dataset [67, 68]. We compared the proposed approach to recent state-of-the-art methods (AVS [67], LAVISH [34] and MMVAE [36]) on the S4 and AVSS downstream tasks (sound source segmentation without and with semantics, respectively). We compare against the mIoU and F1 scores shown in the original papers in Table 2. As can be seen, DEEPAVFUSION outperforms all prior work on both tasks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Audio-visual classification. The above experiments demonstrate the effectiveness of DEEPAVFUSION on audio-visual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27191" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.263, + 0.089, + 0.706, + 0.102 + ], + "angle": 0, + "content": "Table 3. Comparing DEEPAVFUSION trained on VGGSound and AudioSet dataset." + }, + { + "type": "table", + "bbox": [ + 0.19, + 0.104, + 0.785, + 0.171 + ], + "angle": 0, + "content": "
Pre-training DatasetPre-training EpochsVGGSound CLSAudioSet CLSVGGSS-Music SEPAVSBench S4
LP AccFt AccLP APFt APSDRSIRSARmIoUF1
VGGSound20053.0858.1930.2532.885.798.2413.8289.9492.34
AudioSet20053.0857.9132.6934.716.939.9313.4990.2792.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.174, + 0.471, + 0.211 + ], + "angle": 0, + "content": "Table 4. Audio-visual classification performance on VGGSound-Music and VGGSound-All datasets, using linear probing (LP) and finetuning (FT) protocols. Performance is measured in terms of classification accuracy \\([\\%]\\)." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.213, + 0.421, + 0.306 + ], + "angle": 0, + "content": "
MethodVGGS-MusicVGGS-All
LPFTLPFT
(Image) MAE [22]48.2553.1837.1243.86
AudioMAE [28]52.7358.2942.0549.53
CAV-MAE [20]59.1867.5849.8355.32
DEEPAVFUSION65.3272.2553.0858.19
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.471, + 0.463 + ], + "angle": 0, + "content": "tasks that require a detailed understanding of audio-visual interactions. We further evaluated the learned representations on recognition tasks. In particular, we compared our model to uni-modal masked reconstruction methods, MAE [22] and AudioMAE [28], as well as to CAV-MAE [20], a late-fusion model trained with both masked reconstruction and contrastive learning objectives. For a fair comparison, we use the same ViT-Base architecture for the uni-modal encoders and tune the model on image-audio pairs obtained from the VGG-Sound dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.634 + ], + "angle": 0, + "content": "As can be seen in Table 4, we outperform both uni-modal pre-trained encoders on linear probing and fine-tuning by significant margins. DEEPAVFUSION outperforms the image MAE [22] and AudioMAE [28] by \\(19.62\\%\\) and \\(13.26\\%\\) on VGGSound-Music and \\(19.08\\%\\) and \\(12.49\\%\\) on the full VGSSound dataset. These direct comparisons to uni-modal encoders highlight the importance of learning joint representations for multi-modal recognition. Furthermore, our DEEPAVFUSION also significantly outperforms CAV-MAE which optimizes the same audio-visual reconstruction objective of Eq. 14 but does not leverage an early fusion encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.742 + ], + "angle": 0, + "content": "It should be noted that multi-modal recognition could be further improved by learning from audio-video data (instead of audio-image pairs). However, video modeling is beyond the scope of this work. The improvements in Table 4 demonstrate nevertheless that even tasks like audio-visual classification of sound sources can benefit from more deeply integrated multi-modal representations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.248, + 0.77 + ], + "angle": 0, + "content": "4.3. Pre-training data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The comparisons above were conducted using VGGSound for pre-training. To study the impact of pre-training data distribution and scale, we also trained our model on the larger AudioSet dataset [18] (with 1.73M training samples). We then tuned the resulting model on several downstream tasks, from classification on both VGGSounds and AudioSet to source separation on VGGSS-Musion and sound source segmentation on AVSBench (S4 task). As can be seen in" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.174, + 0.871, + 0.187 + ], + "angle": 0, + "content": "Table 5. Ablation study of major components of DEEPAVFUSION." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.188, + 0.884, + 0.32 + ], + "angle": 0, + "content": "
(#)Unimodal \nPre-TrainingFusionLocal AV \nInteractionsLinear \nAccSep \nSDRSegm \nmIoU
(1)✓Early✓53.086.5352.05
(2)×Early✓44.362.0329.55
(3)✓Early×51.196.1548.71
(4)✓Mid✓51.215.8548.04
(5)✓Late✓46.525.3244.32
(6)✓None×39.674.2338.32
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.894, + 0.421 + ], + "angle": 0, + "content": "Table 3, pre-training on AudioSet yields better performance on most downstream tasks, demonstrating the scalability of the proposed method. The only exception is classification on VGGSound, where pre-training on VGGSound yields slightly better performance. This is likely due to a smaller domain gap between pre-training and fine-tuning data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.749, + 0.451 + ], + "angle": 0, + "content": "4.4. Understanding DEEPAVFUSION" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.656 + ], + "angle": 0, + "content": "The previous section demonstrated the efficacy of early fusion of local audio-visual interactions, with the proposed method outperforming the state-of-the-art on a wide range of audio-visual tasks. In this section, we conduct a thorough analysis of our framework, revealing emergent properties in our model, and assessing the benefits of uni-modal pretraining, fusion depth, and other major design choices of the proposed methodology. To provide a full-picture comparison, we evaluate all models on five downstream tasks, namely linear probing and fine-tuning on VGGSound, sound source separation on VGGSound-Music, sound source localization on Flickr-SoundNet, and audio-visual segmentation with semantics on AVSBench." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Ablation studies. We begin our analysis with a series of ablation studies in Table 5 to measure the impact of three important components of the proposed framework, namely 1) uni-modal pre-training, 2) early fusion, and 3) audio-visual interactions. All models use the same backbone architecture for the uni-modal encoders and are trained to optimize the Audio-Visual MAE objective. Model #1 is the full DEEPAVFUSION framework, and the remaining models relax each of the three components in turn. Model #2 skips uni-modal pre-training, significantly hurting performance when compared to model #1 on all tasks, emphasizing the value of uni-modal pre-training, when learning deeply integrated audio-visual encoders. Model #3 performs early fusion but uses the \"token fusion\" strategy similar to that used in MBT [48]. Token fusion attends to uni-modal representations and cannot directly capture audio-visual inter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27192" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.102 + ], + "angle": 0, + "content": "Table 6. Exploration studies on the efficiency of factorized interactions." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.105, + 0.462, + 0.187 + ], + "angle": 0, + "content": "
Early FusionFactorized InteractionsSpeed (video/sec)Memory (GB)Linear AccSep SDRSegm mIoU
××14.510.346.525.3244.32
✓×4.331.652.826.3052.13
✓✓12.813.153.086.5352.05
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.206, + 0.451, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.471, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Nearest neighbor accuracy (NNAcc) vs training steps (Step) using Audio, Image, Fusion Tokens from an DEEPAVFUSION trained on VGGSounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.533 + ], + "angle": 0, + "content": "actions, limiting the expressivity of the fusion block, and the overall performance of the model. Models #4, #5 and #6 use the same fusion modules with dense interactions but perform mid-fusion (at layers 9-12), late-fusion (at layer 12), and no fusion at all. The results indicate that mid-fusion outperforms, late-fusion which in turn is substantially better than no-fusion. Nevertheless, early fusion achieves the best results across all downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.471, + 0.748 + ], + "angle": 0, + "content": "Efficiency of factorized interactions. While audio-visual interactions are crucial for the performance of early fusion, they also introduce a significant computational overhead, if not handled properly. Table 6 lists the VRAM consumption and model throughput during inference for three models: a baseline model without fusion, an early fusion model with dense interactions, and our proposed early-fusion model with factorized interactions. Dense interactions enhance the model's effectiveness on subsequent tasks, but at a high computational cost, both in terms of throughput (3.3 times slower) and memory usage (3.1 times higher). Our proposed factorized interactions significantly cut down the computational overhead, while maintaining (and even slightly enhancing) the performance of early fusion with dense interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Emergent semantics. In addition to robust performance across various tasks, the emergent semantics in fusion tokens are worth highlighting. Despite the masked auto-encoder objective typically favoring representations rich in low-level details, fusion tokens surprisingly yield higher-level semantic representations. Fig. 3 shows the nearest neighbor retrieval performance of our DEEPAVFUSION using either the learned visual, audio, or fusion tokens as queries. As can be seen, the retrieval using fusion tokens is substantially better than uni-modal representations (i.e. better aligned with the" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.089, + 0.826, + 0.102 + ], + "angle": 0, + "content": "Table 7. Impact of early fusion and fusion tokens." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.104, + 0.687, + 0.23 + ], + "angle": 0, + "content": "
#Linear AccSep SDRSegm mIoU
1253.086.5352.05
952.766.1549.93
652.296.1249.02
351.215.8548.04
146.525.3244.32
039.674.2338.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.233, + 0.657, + 0.245 + ], + "angle": 0, + "content": "(a) # Fusion layers" + }, + { + "type": "table", + "bbox": [ + 0.708, + 0.117, + 0.871, + 0.216 + ], + "angle": 0, + "content": "
#Linear AccSep SDRSegm mIoU
151.865.9450.67
851.975.8351.14
1653.086.5352.05
3252.186.4350.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.735, + 0.218, + 0.845, + 0.23 + ], + "angle": 0, + "content": "(b) # Fusion tokens" + }, + { + "type": "table", + "bbox": [ + 0.615, + 0.249, + 0.779, + 0.334 + ], + "angle": 0, + "content": "
#Linear AccSep SDRSegm mIoU
451.785.6849.59
853.086.5352.05
1652.256.1650.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.648, + 0.335, + 0.746, + 0.348 + ], + "angle": 0, + "content": "(c) # Aggr tokens" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.42 + ], + "angle": 0, + "content": "semantics of the dataset). This suggests that fusion tokens aggregate information that is indicative of high-level semantics, while uni-modal representations encode the low-level details required for masked reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.543 + ], + "angle": 0, + "content": "Impact of miscellaneous design choices. Lastly, we examine the effects of the number of fusion layers, fusion tokens and aggregation tokens in Table 7. These results reveal that early fusion is crucial for optimal performance, with model performance improving as fusion depth increases. Furthermore, DEEPAVFUSION can benefit from a relatively large number of tokens, with the performance saturating after 16 fusion tokens and 8 aggregation tokens." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.619, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.895, + 0.859 + ], + "angle": 0, + "content": "In this work, we present DEEPAVFUSION, a simple yet effective early fusion approach with dense interactions that achieves efficient audio-visual pre-training for audio-visual masked auto-encoders. We introduce learnable fusion tokens to aggregate modality-specific information with early fusion from each transformer block of audio and visual encoders, where two variants of parallel and sequential flows are proposed to achieve early fusion between fusion tokens and audio-visual patch tokens. Furthermore, we leverage multi-modal blocks with dense interactions to fuse fusion tokens and patches across audio-visual transformer blocks and achieve efficient downstream fine-tuning with factorized attention blocks. Empirical experiments on Flickr-SoundNet, VGG-Instruments, VGG-Music, VGGSound-All, and AVS-Bench datasets demonstrate the state-of-the-art performance of our DEEPAVFUSION in linear probing, fine-tuning classification, visual sound localization, sound separation, and audio-visual segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27193" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 609-617, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, S. V. A. Garí, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Proceedings of European Conference on Computer Vision (ECCV), pages 17–36, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[5] Changan Chen, Sagnik Majumder, Al-Halah Ziad, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In Proceedings of International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[7] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[9] Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia. Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[12] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for BERT pre-training of vision transformers. In Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.894, + 0.355 + ], + "angle": 0, + "content": "[16] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10478-10487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 324-333, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[18] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In Proceedings of 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[19] Mariana-Iuliana Georgescu, Eduardo Fonseca, Radu Tudor Ionescu, Mario Lucic, Cordelia Schmid, and Anurag Arnab. Audiovisual masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16144-16154, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.894, + 0.622 + ], + "angle": 0, + "content": "[20] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In Proceedings of The Eleventh International Conference on Learning Representations (ICLR), 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[21] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[23] John Hershey and Michael Casey. Audio-visual sound separation via hidden markov models. Advances in Neural Information Processing Systems, 14, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[24] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[25] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27194" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.162 + ], + "angle": 0, + "content": "[26] Po-Sen Huang, Scott Deann Chen, Paris Smaragdis, and Mark Hasegawa-Johnson. Singing-voice separation from monaural recordings using robust principal component analysis. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 57–60, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[27] Po-Yao Huang, Vasu Sharma, Hu Xu, Chaitanya Ryali, Haoqi Fan, Yanghao Li, Shang-Wen Li, Gargi Ghosh, Jitendra Malik, and Christoph Feichtenhofer. Mavil: Masked audio-video learners. arXiv preprint arXiv:2212.08071, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[30] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[31] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.472, + 0.504 + ], + "angle": 0, + "content": "[32] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002–2006, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[33] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[34] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.472, + 0.77 + ], + "angle": 0, + "content": "[36] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 954–965, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[37] Shentong Mo and Pedro Morgado. Benchmarking weakly-supervised audio-visual sound localization. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[38] Shentong Mo and Pedro Morgado. A unified audio-visual learning framework for localization, separation, and recognition. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[39] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[40] Shentong Mo and Yapeng Tian. Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[41] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[42] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[43] Shentong Mo, Jing Shi, and Yapeng Tian. DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[44] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.894, + 0.413 + ], + "angle": 0, + "content": "[45] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4733–4744, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[46] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12934–12945, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.894, + 0.537 + ], + "angle": 0, + "content": "[47] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12475-12486, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.894, + 0.595 + ], + "angle": 0, + "content": "[48] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.597, + 0.894, + 0.651 + ], + "angle": 0, + "content": "[49] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.652, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[50] Andrew Owens, Jiajun Wu, Josh H. McDermott, William T. Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-816, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[52] Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[53] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[54] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proceedings of European Conference on Computer Vision (ECCV), page 436–454, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.216 + ], + "angle": 0, + "content": "[55] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2745-2754, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), page 5998-6008, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.289, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[57] Tuomas Virtanen. Monaural sound source separation by nonnegative matrix factorization with temporal continuity and sparseness criteria. IEEE Transactions on Audio, Speech, and Language Processing, 15(3):1066-1074, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[58] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan L. Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[59] Alexander Wettig, Tianyu Gao, Zexuan Zhong, and Danqi Chen. Should you mask \\(15\\%\\) in masked language modeling? In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2985-3000, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.472, + 0.523 + ], + "angle": 0, + "content": "[60] Jiantao Wu and Shentong Mo. Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[61] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1326–1335, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.472, + 0.635 + ], + "angle": 0, + "content": "[62] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 6291–6299, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.706, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[64] Xudong Xu, Bo Dai, and Dahua Lin. Recursive visual sound separation using minus-plus net. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[65] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[66] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1735-1744, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[67] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Wang, and Yiran Zhong. Audio-visual segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.895, + 0.203 + ], + "angle": 0, + "content": "[68] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27196" + } + ] +] \ No newline at end of file diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_origin.pdf b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e8f0a5d3005e4229d6ea6f5932d6ac0ff734c5a --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/6a06bd12-56a6-4c87-8823-cfa2f03902c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ff0a4849f3f2652ae79b861cb5d3942877328d15c954f03a6b5b03082d409c1 +size 587497 diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/full.md b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..15a454ad535c952dfe08540af33f900856b5cb4c --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/full.md @@ -0,0 +1,354 @@ +# Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling + +Shentong Mo + +Carnegie Mellon University + +shentong.mo@sv.cmu.edu + +Pedro Morgado + +University of Wisconsin-Madison + +pmorgado@wisc.edu + +https://github.com/stoneMo/DeepAVFusion + +# Abstract + +Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures. + +# 1. Introduction + +Humans naturally integrate audio-visual information to perceive and understand the environment. Several studies in cognitive psychology and neuroscience, as well as classic perceptual illusions such as the McGurk effect, demonstrate + +that such audio-visual fusion can occur early on in the perceptual stack, enabling deeper integration of the two modalities. Early fusion models aim to emulate this human-like perception. Specifically, it refers to the process of integrating auditory and visual cues at an early stage of a multi-modal perception model, in order to leverage the synergistic effects of both modalities. This can be especially powerful if the fusion process can attend to and establish connections between local components of the visual and audio signal (e.g., the frequencies of someone's voice with the pixels of their lips). In sum, early fusion of local audio-visual interactions holds the promise of a deeper and more sophisticated understanding of audio-visual content, critical for many real-world applications such as visually guided source separation, localization/segmentation, and multi-modal recognition. + +Despite the potential, audio-visual early fusion architectures have remained under-explored, with the majority of the recent literature focusing on late fusion or no fusion at all. One recent prominent research direction [1, 47] seeks to learn independent uni-modal encoders through contrastive learning by requiring the representations of associated audio-visual pairs to be synchronized in latent space. However, contrastive learning is not compatible with early fusion, as the connections between uni-modal representations lead to trivial solutions to the contrastive learning problem. Another line of recent work [20, 27] learns encoders with late audio-visual fusion through a combination of contrastive learning and masked reconstruction, focusing on downstream tasks that require shallow integration of audio-visual features, such as audio-visual action recognition. + +One potential reason is that early fusion architectures pose a significant training challenge due to the substantial increase in model expressivity it offers. As a result, robust learning frameworks are required to ensure the creation of high-quality models and effectively harness their enhanced capabilities. In this work, we demonstrate the effectiveness of a masked reconstruction framework for training audio- + +![](images/5c8228083643d84dcd15370528386369b224f674d9c48f231fdd14daaa2744e5.jpg) +Figure 1. Unlike prior works on Audio-Visual MAEs which learn either late fusion encoders or separate encoders altogether, we demonstrate that masked reconstruction is especially useful for learning deeply integrated audio-visual encoders with powerful early fusion modules. + +![](images/0f834e35a30fa9a2ae49b6f5ca1133773d70b001ec1f968ce3dc04d7364add01.jpg) + +visual transformers with early-fusion of local interactions (Fig. 1). Our hypothesis is that the simultaneous reconstruction of audio and visual inputs from a limited audio-visual context promotes the learning of deeply integrated representations, as the model is required to understand the fine interactions between the two modalities. Furthermore, masked reconstruction has already been successfully deployed in training uni-modal representations for a variety of input signals, including text [11], images [3, 22], and audio [28], as well as for learning late fusion representations [20, 27]. However, unlike what was previously observed with uni-modal or late fusion models, our work shows that masked audio-visual reconstruction with early fusion encoders results in an interesting emergent property. While uni-modal representations encode the low-level details required for masked reconstruction, we observed that high-level semantics (surprisingly) emerged from the tokens used for audio-visual fusion. + +In addition to demonstrating the effectiveness of masked reconstruction for early fusion transformers, we propose a novel attention-based fusion module, which can effectively attend to the interactions between local audio and visual representations, thereby enhancing its capacity to capture localized interactions between the two modalities. Take, for instance, the sound of a dog barking. The sound signals the presence of a dog and offers clues about its state, such as an open mouth. While the link between the dog and its bark can be made at a global level, our fusion module allows the model to delve deeper into local interactions, like the correlation between the dog's mouth and the barking sound's spectral frequencies. As a result, localized fusion boosts the quality of learned representations for numerous tasks. However, it necessitates the model to explicitly represent all pairwise interactions of local audio and visual tokens. Given the vast number of tokens that current transformers process, dense local interactions can quickly become unmanageable. To enhance efficiency, our fusion module factorizes unimodal representations, thereby circumventing the need to + +model all pairwise interactions. + +In sum, we propose an attention-based fusion module that can attend to local audio-visual interactions and deploy it to learn deeply integrated audio-visual representations by fusing uni-modal representations at early layers in the architecture. We also demonstrate that when paired with masked reconstruction, our framework, denoted DEEPAVFUSION, can learn strong audio-visual representations, yielding state-of-the-art performance on a variety of audio-visual tasks, including visually guided source separation, localization, and segmentation. We extensively evaluated DEEPAVFUSION on a variety of audio-visual datasets, and conducted thorough ablation studies, showing the importance of early fusion, dense interactions, and uni-modal pre-training when learning deeply integrated audio-visual representations. + +# 2. Related Work + +Audio-Visual Representations Learning. Audio-visual representations learning has been addressed in many previous works [1, 2, 14, 16, 23, 24, 30, 40, 43, 45–47, 50, 51, 65, 66] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for many audio-visual tasks, such as audio-event localization [31, 32, 37, 41, 42, 53, 62], audio-visual spatialization [4, 17, 44, 45], audio-visual navigation [4–6] and audio-visual parsing [33, 39, 54, 61]. In this work, our main focus is to learn transferrable audio-visual representations from masked audio-visual reconstruction, which is more challenging than the above-mentioned tasks. + +Masked Representations Learning. Masked representation learning aims to learn self-supervised representations by reconstructing desired features of masked data given unmasked parts as clues. In the recent years, masked representation learning has achieved promising results in natural language processing [9, 11, 35, 52, 59] and computer vision [3, 8, 12, 15, 22, 58, 60, 63] community. Typically, BERT [11] randomly masked $15\%$ of word tokens in the sentence and recovered them with unmasked words to learn generalizable textual features via a self-attention transformer [56]. A block-wise masking strategy was proposed in BEiT [3] to reconstruct discrete tokens of masked image patches for pre-training transferrable visual representations. To simplify the masked image encoding framework, MAE [22] directly reconstructed missing pixels of $75\%$ masked patches using vision transformers [13] for self-supervised pre-training. + +Audio-visual Masked Autoencoders. More recently, researchers introduced diverse masking pipelines [19, 20, 27] to show the effectiveness of masked modeling in learning audio-visual representations. For example, CAV-MAE [20] combined contrastive learning and masked modeling to capture a joint and coordinated audio-visual representation. + +They tried to add a joint encoder on audio-visual features from the last attention block of single-modality encoders to fuse cross-modal information for audio-visual contrastive objectives. MAViL [27] extended masked audio-video reconstruction with masked intra- and inter-modal contrastive learning and self-training by recovering joint audio-video contextualized representations. Despite their promising performance, they ignored the importance of the early fusion of audio-visual features in masked audio-visual reconstruction. In contrast to AV-MAE [19] that simply concatenated the audio and visual tokens before passing them through the joint transformer for early fusion, we will design an early fusion module with interactions between local and visual representations for audio-visual masked auto-encoders. Our audio-visual interactions are different from fusion bottlenecks in MBT [48] that forces information between different modalities to pass through a small number of bottleneck latent. However, we develop a fully novel attention-based fusion module that can effectively attend to interactions between local audio and visual representations, thus enhancing its ability to capture fine-grained interactions between the two modalities. + +Audio-visual Early Fusion. Audio-visual fusion has been the topic of extensive research, with several works proposing a variety of architectures to aggregate multi-modal representations [19, 48, 49]. An example of an early fusion architecture was proposed by Owens et al. [49] to learn representations from audio-visual correspondences, by concatenating features from small unimodal encoders and feeding them into a fused audio-visual network for the early fusion of multisensory features. Beyond this early work, most recent papers on audio-visual fusion [19, 48] propose alternative fusion mechanisms, either through weight sharing [19] or token-based fusion [48]. However, since all these works focused on downstream tasks (mostly classification) that do not require fine multi-modal understanding, the benefits of early fusion were not evident, with mid-level fusion often achieving optimal performance. Different from them, we proposed an early fusion model that is designed for fine multi-modal understanding, by attending to local interactions between the audio and the visual while performing early fusion. The new architecture design paired with masked reconstruction objective is shown to outperform many of the prior works on a variety of downstream multi-modal applications beyond classification. + +# 3. Method + +We aim to learn audio-visual representations that can be effectively transferred to a variety of downstream tasks that require a detailed understanding of audio-visual interactions. To accomplish this, we introduce a novel audio-visual transformer architecture, named DEEPAVFUSION, that enables the + +early fusion of audio and visual tokens through dense local interactions. We also show how to effectively train such early fusion transformers using multi-modal masked reconstruction. In this section, we first describe the proposed early fusion transformer and then introduce the multi-modal masked reconstruction framework used for self-supervised pre-training. + +# 3.1. A Transformer Architecture for Joint Audio-Visual Encoding + +Downstream tasks such as visually guided sound source separation, localization, and segmentation require a deep understanding of fine audio-visual interactions present in the data. However, late fusion often lacks the expressive power to represent such interactions. We introduce a transformer architecture that enables the early fusion of multi-modal representations through factorized local interactions between audio and visual tokens. + +Early fusion architecture While we aim to learn deeply integrated audio-visual representations, foundational uni-modal transformers can still provide a strong starting point for multi-modal learning. As such, we introduce a modular architecture to repurpose existing pre-trained uni-modal transformers. Specifically, we propose a three-branch architecture, illustrated in Fig. 2. The first two branches are uni-modal transformers that process audio and visual tokens, respectively, while the third branch is a multi-modal transformer that updates a set of learnable fusion tokens to fuse audio-visual information. Formally, let $\mathbf{X}_0^f \in \Re^{F \times D}$ be $F$ learnable fusion tokens. We progressively update these tokens, at each layer $l$ , through a fusion block $\Phi_l^f$ , which can interact with and aggregate information from modality-specific features $\mathbf{X}_{l - 1}^v$ , $\mathbf{X}_{l - 1}^a$ . Furthermore, fusion tokens $\mathbf{X}_{l - 1}^f$ are fed into modality-specific blocks, $\Phi_l^v$ and $\Phi_l^a$ , to modulate uni-modal representations from early stages in the network. Overall, representations are updated as follows + +$$ +\boldsymbol {X} _ {l} ^ {f} = \Phi_ {l} ^ {f} \left(\boldsymbol {X} _ {l - 1} ^ {f}, \boldsymbol {X} _ {l - 1} ^ {v}, \boldsymbol {X} _ {l - 1} ^ {a}\right) \tag {1} +$$ + +$$ +\boldsymbol {X} _ {l} ^ {v} = \Phi_ {l} ^ {v} \left(\boldsymbol {X} _ {l - 1} ^ {v}, \boldsymbol {X} _ {l - 1} ^ {f}\right) \tag {2} +$$ + +$$ +\boldsymbol {X} _ {l} ^ {a} = \Phi_ {l} ^ {a} \left(\boldsymbol {X} _ {l - 1} ^ {a}, \boldsymbol {X} _ {l - 1} ^ {f}\right) \tag {3} +$$ + +This token update strategy is repeated for a total of $L$ layers to compute the final representations $\mathbf{X}_L^v, \mathbf{X}_L^a$ and $\mathbf{X}_L^f$ . + +Modality-specific blocks We use standard transformer self-attention for the modality-specific blocks, $\Phi^v$ and $\Phi^a$ , while allowing them to further attend to fusion tokens. Unless otherwise specified, we initialize $\Phi^v$ and $\Phi^a$ from foundational uni-modal models. This is possible since the parameter space of the modality-specific blocks remains unchanged. In particular, we empirically validated the benefits of bootstrapping from [28] and [22], which are pre-trained on large-scale audio and visual datasets, respectively, through a uni-modal masked reconstruction objective. + +![](images/5115c4a420005adddb207674d25a2c220139b865387eab1d8e5f06701148c7fb.jpg) +Figure 2. Audio-Visual Early Fusion Transformer with Dense Interactions. The architecture ① is composed of three interconnected branches - the visual, audio, and fusion branches - each with an equal number of transformer blocks. The visual and audio branches process visual and audio tokens, respectively, while simultaneously attending to fusion tokens. The fusion branch updates a set of learnable fusion tokens to fuse audio-visual information. To better attend to local audio-visual interactions, we introduce a new fusion block ②. At each layer $l$ , all patch-level audio (visual) tokens, $\bar{X}_{l-1}^{a}(\bar{X}_{l-1}^{v})$ , are first aggregated into a small set of aggregation tokens, $\bar{X}_l^a (\bar{X}_l^v)$ by cross-attention. Pairwise audio-visual interactions are then encoded through a linear layer into latents $\bar{X}_l^{av}$ , which are used to update the fusion tokens $\bar{X}_l^f$ by cross-attention. + +![](images/fd6002b3aba7b5ca98a5b9a39fdcbf97ad92f5bf452d5e485275ef3a59ce7e25.jpg) + +Fusion blocks The design of the fusion blocks $\Phi^f$ is critical to representing fine audio-visual interactions. In the context of transformers, a natural way to combine audio-visual representations is through a standard self-attention block which, in addition to the fusion tokens $X^f$ themselves, it can further attend to all image and audio tokens, $X_{l - 1}^v$ and $X_{l - 1}^{a}$ . This approach, which we refer to as token fusion, has been used in previous works such as MBT [48]. While intuitive, standard self-attention cannot directly exploit interactions between local audio-visual representations, which limits their expressivity. To address this limitation, we designed a fusion block that attends to local interactions. + +Fusion with Dense Audio-Visual Interactions Interactions between audio and visual data occur between local regions in the image and the audio spectrogram. Given a set of $n_a$ audio tokens $X^a$ and $n_v$ visual tokens $X^v$ , aggregation of local multi-modal interactions requires attending to representations from all possible $n_a \times n_v$ pairs, to identify the ones that are most useful for the task at hand. While we experimented with several ways to represent multi-modal interactions, from bilinear aggregation to kernel-based aggregation, we found that linear aggregation was sufficient to capture audio-visual interactions for our tasks. Formally, + +$$ +\boldsymbol {X} _ {i j} ^ {a v} = W ^ {a} \boldsymbol {X} _ {i} ^ {a} + W ^ {v} \boldsymbol {X} _ {j} ^ {v} \tag {4} +$$ + +$$ +\boldsymbol {X} ^ {a v} = \left[ \boldsymbol {X} _ {i j} ^ {a v}; \forall i = 1, \dots , n _ {a} \forall j = 1, \dots , n _ {v} \right]. \tag {5} +$$ + +Fusion tokens $X_{l}^{f}$ at layer $l$ are then updated using standard cross-attention [56] over the representation of all local interactions $X^{av}$ + +$$ +\boldsymbol {X} _ {l} ^ {f} = \text {C r o s s A t t e n t i o n} \left(\boldsymbol {X} _ {l - 1} ^ {f}; \boldsymbol {X} _ {l - 1} ^ {\text {a v}}\right). \tag {6} +$$ + +Factorized Audio-Visual Interactions While the procedure above can attend to local interactions, it requires the model to explicitly represent all possible $n_a \times n_v$ pairs. Given that the number of audio and visual tokens is typically large, the number of interactions can quickly become intractable, + +resulting in high memory consumption and low throughput. To address this limitation, we introduce modality-specific aggregation tokens updated by aggregation blocks. The goal of the aggregation blocks is to summarize the large number of audio/visual tokens $X^a / X^v$ into a small set of $n_{\bar{a}} / n_{\bar{v}}$ aggregation tokens $\bar{X}^a / \bar{X}^v$ , respectively. Specifically, at layer $l$ , aggregation tokens are updated by cross-attention + +$$ +\bar {X} _ {l} ^ {a} = \text {C r o s s A t t e n t i o n} \left(\bar {X} _ {l - 1} ^ {a}, X _ {l - 1} ^ {a}\right) \tag {7} +$$ + +$$ +\bar {X} _ {l} ^ {v} = \text {C r o s s A t t e n t i o n} \left(\bar {X} _ {l - 1} ^ {v}, X _ {l - 1} ^ {v}\right). \tag {8} +$$ + +Then, instead of explicitly representing all $n_a \times n_v$ pairwise interactions, we restrict audio-visual interactions to the sets of aggregated tokens. Formally, factorized interactions between aggregated tokens $\bar{X}_i^a$ and $\bar{X}_j^v$ are represented as + +$$ +\bar {X} _ {i j} ^ {a v} = W ^ {\bar {a}} \bar {X} _ {i} ^ {a} + W ^ {\bar {v}} \bar {X} _ {j} ^ {v} \tag {9} +$$ + +$$ +\bar {X} ^ {a v} = \left[ \bar {X} _ {i j} ^ {a v}; \forall i = 1, \dots , n _ {\bar {a}} \forall j = 1, \dots , n _ {\bar {v}} \right] \tag {10} +$$ + +and the fusion tokens are updated at layer $l$ by + +$$ +\boldsymbol {X} _ {l} ^ {f} = \text {C r o s s A t t e n t i o n} \left(\boldsymbol {X} _ {l - 1} ^ {f}, \bar {\boldsymbol {X}} _ {l} ^ {a v}\right). \tag {11} +$$ + +Since we only represent interactions between the aggregated uni-modal tokens, the number of possible interactions is greatly reduced. We observed that a $700 \times$ reduction still maintains the benefits obtained with local dense interactions. + +# 3.2. Learning Early-Fusion Transformers Through Audio-Visual Masked Auto-Encoding + +Masked auto-encoders have been shown to learn effective representations for a wide variety of input signals, including images [3, 22] and audio [28]. Formally, the input signal $x$ is partitioned into a set of patch tokens $x = \{x_{t}\}_{t=1}^{T}$ , which is randomly split into the visible set $X_{\mathrm{vis}} = \{x_{t}\}_{t \in \mathcal{V}}$ and the masked set $X_{\mathrm{mask}} = \{x_{t}\}_{t \in \mathcal{M}}$ ( $\mathcal{V}$ and $\mathcal{M}$ are non-overlapping index sets). In the case of images, visible and masked tokens, $X_{\mathrm{vis}}$ and $X_{\mathrm{mask}}$ , are obtained from a grid + +of small image patches [3, 22], while in the case of audio signals, tokens are obtained either from small windows of a raw waveform or from small 2D patches of the audio spectrogram [28]. Then, to learn representations of the input data, a transformer encoder $f$ is first used to encode the visible tokens $X_{\mathrm{vis}} = f(X_{\mathrm{vis}}) \in \Re^{|\mathcal{V}| \times d}$ , where $d$ denotes the dimensionality of the token representation. These representations together with a series of (position marked) mask tokens $M_{\mathrm{mask}}$ are then fed to a decoder $g(\cdot)$ that seeks to reconstruct the masked input tokens $\hat{X}_{\mathrm{mask}} = g(X_{\mathrm{vis}}, M_{\mathrm{mask}})$ . To accomplish that, the model is trained to minimize the mean squared reconstruction error + +$$ +\mathcal {L} _ {\mathrm {M A E}} \left(X _ {\text {m a s k}}, \hat {X} _ {\text {m a s k}}\right) = \frac {1}{| \mathcal {M} |} \sum_ {\substack {x \in X _ {\text {m a s k}} \\ \hat {x} \in \hat {X} _ {\text {m a s k}}}} \| x - \hat {x} \| _ {2} ^ {2}. \tag{12} +$$ + +Following prior work [20, 27], masked reconstruction can be easily extended to multi-modal signals, by simultaneously masking tokens from both audio and visual modalities. In the case of early-fusion transformers, the encoder $f$ is then used to encode the visible tokens, $X_{\mathrm{vis}}^v$ and $X_{\mathrm{vis}}^a$ , + +$$ +\boldsymbol {X} _ {\mathrm {v i s}} ^ {v}, \boldsymbol {X} _ {\mathrm {v i s}} ^ {a}, \boldsymbol {X} ^ {f} = f \left(\boldsymbol {X} _ {\mathrm {v i s}} ^ {v}, \boldsymbol {X} _ {\mathrm {v i s}} ^ {a}\right) \tag {13} +$$ + +which, in addition to the uni-modal representations, $X_{\mathrm{vis}}^v$ and $X_{\mathrm{vis}}^a$ , also returns a set of fusion representations $\pmb{X}^{f}$ . These representations are then fed to modality-specific decoders to reconstruct both audio and visual masked tokens, $\hat{X}_{\mathrm{mask}}^v = g_v(X_{\mathrm{vis}}^f,M_{\mathrm{mask}}^v)$ and $\hat{X}_{\mathrm{mask}}^a = g_a(X_{\mathrm{vis}}^f,M_{\mathrm{mask}}^a)$ , by minimizing the loss + +$$ +\mathcal {L} _ {\mathrm {A V - M A E}} = \mathcal {L} _ {\mathrm {M A E}} \left(X _ {\text {m a s k}} ^ {v}, \hat {X} _ {\text {m a s k}} ^ {v}\right) + \mathcal {L} _ {\mathrm {M A E}} \left(X _ {\text {m a s k}} ^ {a}, \hat {X} _ {\text {m a s k}} ^ {a}\right) \tag {14} +$$ + +where $\mathcal{L}_{\mathrm{MAE}}$ is the auto-encoder loss of Eq. 12. + +While multi-modal masked reconstruction has been previously explored in [20, 27], these approaches focus on learning separate audio and image encoders, $f_{a}$ and $f_{v}$ , with multi-modal representations obtained through a late fusion module $f_{f}$ . In contrast, we propose to learn a single multimodal early fusion encoder $f$ , and demonstrate that masked reconstruction is especially useful for learning multi-modal representations capable of representing fine audio-visual interactions through early fusion of local interactions between audio and visual tokens. + +# 4. Experiments + +We now evaluate the proposed DEEPAVFUSION on a variety of audio-visual downstream tasks that require a deep understanding of audio-visual interactions. + +# 4.1. Implementation details + +Pre-training. For pre-training, we experiment with two datasets, namely VGG-Sounds [7] (with 144k training sam + +ples) and AudioSet [18] (with 1.73M training samples), both of them containing diverse videos focused on audio events and sound sources. The audio is represented by log mel spectrograms extracted from $3s$ of audio at a sample rate of $16000\mathrm{Hz}$ . To compute the log spectrograms, we apply an STFT using approximately 50ms windows with a hop size of $15\mathrm{ms}$ , resulting in an input tensor of size $128\times 196$ (128 mel frequency bands over 196 time steps). Since we use image models for the visual component, we randomly extract single frames from within the time window defined by the $3s$ audio snippets. Input frames are augmented by random crops with a minimum area of 0.5 and random horizontal flips, and resized into a $224\times 224$ resolution. + +Unless otherwise specified, we initialize the uni-modal transformer blocks using modality-specific encoders pretrained with uni-modal masked reconstruction objectives [22, 28]. Specifically, we use the ViT-Base model from [22] pre-trained on ImageNet [10], and the Spec-MAE model (with the same ViT-Base architecture) pre-trained on AudioSet [18]. In our base configuration, we add multi-modal fusion blocks on all 12 layers of the model, together with 16 multi-modal fusion tokens, 8 visual aggregation tokens, and 8 audio aggregation tokens. For efficiency, we reduced the dimensionality of the MLP embeddings using an MLP ratio of 1 (instead of the standard 4), as well as the dimensionality of space in which the similarity for self and cross-attention is computed to 16 (as opposed to the standard 64). + +The models were trained using the Adam optimizer [29]. Ablation and parametric studies were conducted on VG-GSound, training for 200 epochs with a total batch size of 512 (split between 2 GPUs and 4 iterations of gradient accumulation). Large-scale training on AudioSet was conducted for 200 epochs with a total batch size of 2048 (split between 16 GPUs and 4 iterations of gradient accumulation). In all cases, we use a base learning rate of $1.5e - 4$ adjusted for the effective batch size by the linear scaling rule $lr = blr \times bs / 256$ [21], a 40 epoch learning rate warm-up schedule followed by cosine learning rate decay, and a weight decay of 0.05. + +Downstream tasks. After pre-training, we fine-tune our models on a variety of downstream tasks. + +Visually-guided sound source separation. Following [65], we use a mix-and-separate strategy for training. The pretrained model is used to obtain audio-visual representations that are fed to a UNet decoder to generate the separated audio. We use a similar separation strategy by predicting a separation mask that is directly applied to the input mixture SFTF. To compute the separation masks, we deploy a 10-layer UNet decoder, which conditions its predictions on (1) the spatially pooled visual representations and (2) five intermediate audio representations obtained during the encoding stage at layers 1, 3, 6, 9, and 12. Following [65], the + +predicted separation masks are trained to regress the STFT magnitude ratio between the separated source and the mixture at each time-frequency bin. We evaluate on three different datasets: the MUSIC [65] dataset and two subsets from VGG-Sounds [7]. MUSIC [65] consists of 448 untrimmed YouTube music videos of solos and duets from 11 instrument categories, where we use 358 solo videos for training and 90 solo videos for evaluation. This dataset was slightly smaller than the original MUSIC dataset since some videos are no longer publicly available to be downloaded. We also use two subsets of VGG-Sounds [7]: the VGGSound-Instruments subset [25] which includes 32k video clips of 10s lengths from 36 musical instrument classes, and a broader newly defined "VGGSound-Music" subset which includes 40,908 video clips from 49 music categories for training and 1201 clips for testing. As commonly done in the source separation literature, we measure performance using Signal-to-Distortion Ratio (SDR), Signal-to-Artifact Ratio (SAR), and Signal-to-Interference Ratio (SIR). + +Audio-visual segmentation. We fine-tune our models on AVSBench [67] for the tasks of Single Source Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS). Task S4 aims to predict pixel-level segmentation maps of the visible sound sources, while AVSS further requires the model to recognize the class of the sound source (in addition to the segmentation maps). We use the standard train/val/test split [67] and measure performance using mIoU and F1 scores. Similarly to the source separation task, audio-visual representations, obtained from the pre-trained model, are fed to a UNet decoder to generate the segmentation maps. This decoder conditions its predictions on a globally pooled audio representation and intermediate visual feature maps obtained at layers 1, 3, 6, 9, and 12. + +Recognition through Linear Probing or Finetuning. We further assess the pre-trained representations on multi-modal recognition tasks by conducting linear probing and finetuning evaluations on the VGGSound-Music, VGGSound-All, and the balanced AudioSet dataset. In all cases, we simply attach three linear classifiers on the average pooled visual, audio, and fusion representations, respectively. The three classification logits are then averaged to obtain the final prediction. As usual, on VGGSound datasets, we measure performance by class accuracy, while on AudioSet (a multi-label dataset) we use the average precision (AP) and Area under the ROC curve (AUC) scores averaged across classes. + +# 4.2. Comparison to prior work + +To demonstrate the effectiveness of the proposed method, we extensively compare it to prior work on audio-visual downstream tasks that require a detailed understanding of audio-visual data. Unless otherwise specified, we use our model trained on VGGSound for comparisons to prior work, + +Table 1. Sound source separation performance on the MUSIC and VG-GSound datasets. + +
MethodMUSICVGGS-InstrumentsVGGS-Music
SDRSIRSARSDRSIRSARSDRSIRSAR
NMF [57]-0.892.386.28-3.520.786.95-5.060.157.03
RPCA [26]0.784.627.580.362.387.95-1.681.578.26
SoP [65]3.627.959.632.725.679.851.564.5910.15
MP-Net [64]3.988.369.863.056.1210.171.955.0210.59
CCoL [55]5.278.7510.524.076.9310.852.735.8611.03
OneAVM [38]6.2710.6812.355.897.8511.233.676.5311.85
DEEPÄVFUSION7.9512.4112.806.959.5213.235.798.2413.82
+ +Table 2. Audio-visual segmentation performance on the AVSBench dataset for the Sound Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS) tasks. + +
MethodAVSBench-S4AVSBench-AVSS
mIoUF1mIoUF1
AVS [67, 68]78.7487.9029.7735.2
LAVISH [34]80.10---
MMVAE [36]81.7490.10--
DEEPAVFUSION89.9492.3452.0558.29
+ +as this enables us to make the fairest comparisons with the largest number of works that require pre-training. + +Sound source separation. Table 1 shows the comparison between DEEPAVFUSION and several prior works, including traditional signal processing methods, NMF [57] and RPCA [26], deep learning methods specialized on the source separation task, Sound-of-Pixels [65] and MP-Net [64], and deep learning methods trained for simultaneous localization and separation, CCoL [55] and OneAVM [38]. While all listed deep learning approaches to source separation use (supervised) ImageNet pre-trained visual encoders, our model shows strong performance using completely unsupervised pre-training. Table 1 shows that DEEPAVFUSION outperforms prior works on all metrics. For example, on VGGSound-Music, we outperform the recently proposed OneAVM [38] by 2.12 SDR and 1.71 SIR, which uses supervised pre-training for the vision component and uses source separation as just one of multiple pre-training objectives. Although the different models use different architectures for their audio-visual encoders (which make direct comparison challenging to analyze), the improvements demonstrated on this task indicate that visual source separation can benefit from early-fusion to extract rich representations from fine audio-visual interactions. + +Audio-Visual Segmentation. We assess the effectiveness of DEEPAVFUSION for audio-visual segmentation on the AVS-Bench dataset [67, 68]. We compared the proposed approach to recent state-of-the-art methods (AVS [67], LAVISH [34] and MMVAE [36]) on the S4 and AVSS downstream tasks (sound source segmentation without and with semantics, respectively). We compare against the mIoU and F1 scores shown in the original papers in Table 2. As can be seen, DEEPAVFUSION outperforms all prior work on both tasks + +Audio-visual classification. The above experiments demonstrate the effectiveness of DEEPAVFUSION on audio-visual + +Table 3. Comparing DEEPAVFUSION trained on VGGSound and AudioSet dataset. + +
Pre-training DatasetPre-training EpochsVGGSound CLSAudioSet CLSVGGSS-Music SEPAVSBench S4
LP AccFt AccLP APFt APSDRSIRSARmIoUF1
VGGSound20053.0858.1930.2532.885.798.2413.8289.9492.34
AudioSet20053.0857.9132.6934.716.939.9313.4990.2792.49
+ +Table 4. Audio-visual classification performance on VGGSound-Music and VGGSound-All datasets, using linear probing (LP) and finetuning (FT) protocols. Performance is measured in terms of classification accuracy $[\%]$ . + +
MethodVGGS-MusicVGGS-All
LPFTLPFT
(Image) MAE [22]48.2553.1837.1243.86
AudioMAE [28]52.7358.2942.0549.53
CAV-MAE [20]59.1867.5849.8355.32
DEEPAVFUSION65.3272.2553.0858.19
+ +tasks that require a detailed understanding of audio-visual interactions. We further evaluated the learned representations on recognition tasks. In particular, we compared our model to uni-modal masked reconstruction methods, MAE [22] and AudioMAE [28], as well as to CAV-MAE [20], a late-fusion model trained with both masked reconstruction and contrastive learning objectives. For a fair comparison, we use the same ViT-Base architecture for the uni-modal encoders and tune the model on image-audio pairs obtained from the VGG-Sound dataset. + +As can be seen in Table 4, we outperform both uni-modal pre-trained encoders on linear probing and fine-tuning by significant margins. DEEPAVFUSION outperforms the image MAE [22] and AudioMAE [28] by $19.62\%$ and $13.26\%$ on VGGSound-Music and $19.08\%$ and $12.49\%$ on the full VGSSound dataset. These direct comparisons to uni-modal encoders highlight the importance of learning joint representations for multi-modal recognition. Furthermore, our DEEPAVFUSION also significantly outperforms CAV-MAE which optimizes the same audio-visual reconstruction objective of Eq. 14 but does not leverage an early fusion encoder. + +It should be noted that multi-modal recognition could be further improved by learning from audio-video data (instead of audio-image pairs). However, video modeling is beyond the scope of this work. The improvements in Table 4 demonstrate nevertheless that even tasks like audio-visual classification of sound sources can benefit from more deeply integrated multi-modal representations. + +# 4.3. Pre-training data + +The comparisons above were conducted using VGGSound for pre-training. To study the impact of pre-training data distribution and scale, we also trained our model on the larger AudioSet dataset [18] (with 1.73M training samples). We then tuned the resulting model on several downstream tasks, from classification on both VGGSounds and AudioSet to source separation on VGGSS-Musion and sound source segmentation on AVSBench (S4 task). As can be seen in + +Table 5. Ablation study of major components of DEEPAVFUSION. + +
(#)Unimodal +Pre-TrainingFusionLocal AV +InteractionsLinear +AccSep +SDRSegm +mIoU
(1)✓Early✓53.086.5352.05
(2)×Early✓44.362.0329.55
(3)✓Early×51.196.1548.71
(4)✓Mid✓51.215.8548.04
(5)✓Late✓46.525.3244.32
(6)✓None×39.674.2338.32
+ +Table 3, pre-training on AudioSet yields better performance on most downstream tasks, demonstrating the scalability of the proposed method. The only exception is classification on VGGSound, where pre-training on VGGSound yields slightly better performance. This is likely due to a smaller domain gap between pre-training and fine-tuning data. + +# 4.4. Understanding DEEPAVFUSION + +The previous section demonstrated the efficacy of early fusion of local audio-visual interactions, with the proposed method outperforming the state-of-the-art on a wide range of audio-visual tasks. In this section, we conduct a thorough analysis of our framework, revealing emergent properties in our model, and assessing the benefits of uni-modal pretraining, fusion depth, and other major design choices of the proposed methodology. To provide a full-picture comparison, we evaluate all models on five downstream tasks, namely linear probing and fine-tuning on VGGSound, sound source separation on VGGSound-Music, sound source localization on Flickr-SoundNet, and audio-visual segmentation with semantics on AVSBench. + +Ablation studies. We begin our analysis with a series of ablation studies in Table 5 to measure the impact of three important components of the proposed framework, namely 1) uni-modal pre-training, 2) early fusion, and 3) audio-visual interactions. All models use the same backbone architecture for the uni-modal encoders and are trained to optimize the Audio-Visual MAE objective. Model #1 is the full DEEPAVFUSION framework, and the remaining models relax each of the three components in turn. Model #2 skips uni-modal pre-training, significantly hurting performance when compared to model #1 on all tasks, emphasizing the value of uni-modal pre-training, when learning deeply integrated audio-visual encoders. Model #3 performs early fusion but uses the "token fusion" strategy similar to that used in MBT [48]. Token fusion attends to uni-modal representations and cannot directly capture audio-visual inter + +Table 6. Exploration studies on the efficiency of factorized interactions. + +
Early FusionFactorized InteractionsSpeed (video/sec)Memory (GB)Linear AccSep SDRSegm mIoU
××14.510.346.525.3244.32
✓×4.331.652.826.3052.13
✓✓12.813.153.086.5352.05
+ +![](images/4535a5279b37193eaf522f7851ff2594b05ded8ae7c184a6c07df064991a2def.jpg) +Figure 3. Nearest neighbor accuracy (NNAcc) vs training steps (Step) using Audio, Image, Fusion Tokens from an DEEPAVFUSION trained on VGGSounds. + +actions, limiting the expressivity of the fusion block, and the overall performance of the model. Models #4, #5 and #6 use the same fusion modules with dense interactions but perform mid-fusion (at layers 9-12), late-fusion (at layer 12), and no fusion at all. The results indicate that mid-fusion outperforms, late-fusion which in turn is substantially better than no-fusion. Nevertheless, early fusion achieves the best results across all downstream tasks. + +Efficiency of factorized interactions. While audio-visual interactions are crucial for the performance of early fusion, they also introduce a significant computational overhead, if not handled properly. Table 6 lists the VRAM consumption and model throughput during inference for three models: a baseline model without fusion, an early fusion model with dense interactions, and our proposed early-fusion model with factorized interactions. Dense interactions enhance the model's effectiveness on subsequent tasks, but at a high computational cost, both in terms of throughput (3.3 times slower) and memory usage (3.1 times higher). Our proposed factorized interactions significantly cut down the computational overhead, while maintaining (and even slightly enhancing) the performance of early fusion with dense interactions. + +Emergent semantics. In addition to robust performance across various tasks, the emergent semantics in fusion tokens are worth highlighting. Despite the masked auto-encoder objective typically favoring representations rich in low-level details, fusion tokens surprisingly yield higher-level semantic representations. Fig. 3 shows the nearest neighbor retrieval performance of our DEEPAVFUSION using either the learned visual, audio, or fusion tokens as queries. As can be seen, the retrieval using fusion tokens is substantially better than uni-modal representations (i.e. better aligned with the + +Table 7. Impact of early fusion and fusion tokens. + +
#Linear AccSep SDRSegm mIoU
1253.086.5352.05
952.766.1549.93
652.296.1249.02
351.215.8548.04
146.525.3244.32
039.674.2338.32
+ +(a) # Fusion layers + +
#Linear AccSep SDRSegm mIoU
151.865.9450.67
851.975.8351.14
1653.086.5352.05
3252.186.4350.60
+ +(b) # Fusion tokens + +
#Linear AccSep SDRSegm mIoU
451.785.6849.59
853.086.5352.05
1652.256.1650.89
+ +(c) # Aggr tokens + +semantics of the dataset). This suggests that fusion tokens aggregate information that is indicative of high-level semantics, while uni-modal representations encode the low-level details required for masked reconstruction. + +Impact of miscellaneous design choices. Lastly, we examine the effects of the number of fusion layers, fusion tokens and aggregation tokens in Table 7. These results reveal that early fusion is crucial for optimal performance, with model performance improving as fusion depth increases. Furthermore, DEEPAVFUSION can benefit from a relatively large number of tokens, with the performance saturating after 16 fusion tokens and 8 aggregation tokens. + +# 5. Conclusion + +In this work, we present DEEPAVFUSION, a simple yet effective early fusion approach with dense interactions that achieves efficient audio-visual pre-training for audio-visual masked auto-encoders. We introduce learnable fusion tokens to aggregate modality-specific information with early fusion from each transformer block of audio and visual encoders, where two variants of parallel and sequential flows are proposed to achieve early fusion between fusion tokens and audio-visual patch tokens. Furthermore, we leverage multi-modal blocks with dense interactions to fuse fusion tokens and patches across audio-visual transformer blocks and achieve efficient downstream fine-tuning with factorized attention blocks. Empirical experiments on Flickr-SoundNet, VGG-Instruments, VGG-Music, VGGSound-All, and AVS-Bench datasets demonstrate the state-of-the-art performance of our DEEPAVFUSION in linear probing, fine-tuning classification, visual sound localization, sound separation, and audio-visual segmentation. + +# References + +[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 609-617, 2017. 1, 2 +[2] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2016. 2 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 2, 4, 5 +[4] Changan Chen, Unnat Jain, Carl Schissler, S. V. A. Garí, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Proceedings of European Conference on Computer Vision (ECCV), pages 17–36, 2020. 2 +[5] Changan Chen, Sagnik Majumder, Al-Halah Ziad, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In Proceedings of International Conference on Learning Representations (ICLR), 2021. +[6] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. 2 +[7] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 5, 6 +[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2 +[9] Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2019. 2 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia. Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 5 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[12] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for BERT pre-training of vision transformers. In Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, 2023. 2 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 2 +[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 2 +[15] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[16] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10478-10487, 2020. 2 +[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 324-333, 2019. 2 +[18] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In Proceedings of 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780, 2017. 5, 7 +[19] Mariana-Iuliana Georgescu, Eduardo Fonseca, Radu Tudor Ionescu, Mario Lucic, Cordelia Schmid, and Anurag Arnab. Audiovisual masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16144-16154, 2023. 2, 3 +[20] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In Proceedings of The Eleventh International Conference on Learning Representations (ICLR), 2023. 1, 2, 5, 7 +[21] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7 +[23] John Hershey and Michael Casey. Audio-visual sound separation via hidden markov models. Advances in Neural Information Processing Systems, 14, 2001. 2 +[24] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 2 +[25] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 6 + +[26] Po-Sen Huang, Scott Deann Chen, Paris Smaragdis, and Mark Hasegawa-Johnson. Singing-voice separation from monaural recordings using robust principal component analysis. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 57–60, 2012. 6 +[27] Po-Yao Huang, Vasu Sharma, Hu Xu, Chaitanya Ryali, Haoqi Fan, Yanghao Li, Shang-Wen Li, Gargi Ghosh, Jitendra Malik, and Christoph Feichtenhofer. Mavil: Masked audio-video learners. arXiv preprint arXiv:2212.08071, 2022. 1, 2, 3, 5 +[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 5, 7 +[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[30] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[31] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2 +[32] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002–2006, 2019. 2 +[33] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[34] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6 +[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2 +[36] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 954–965, 2023. 6 +[37] Shentong Mo and Pedro Morgado. Benchmarking weakly-supervised audio-visual sound localization. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2 +[38] Shentong Mo and Pedro Morgado. A unified audio-visual learning framework for localization, separation, and recognition. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6 +[39] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In + +Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[40] Shentong Mo and Yapeng Tian. Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2 +[41] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056, 2023. 2 +[42] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2 +[43] Shentong Mo, Jing Shi, and Yapeng Tian. DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903, 2023. 2 +[44] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[45] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4733–4744, 2020. 2 +[46] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12934–12945, 2021. +[47] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12475-12486, 2021. 1, 2 +[48] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 7 +[49] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3 +[50] Andrew Owens, Jiajun Wu, Josh H. McDermott, William T. Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-816, 2016. 2 +[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 2 +[52] Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223, 2019. 2 +[53] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 2 + +[54] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proceedings of European Conference on Computer Vision (ECCV), page 436–454, 2020. 2 +[55] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2745-2754, 2021. 6 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), page 5998-6008, 2017. 2, 4 +[57] Tuomas Virtanen. Monaural sound source separation by nonnegative matrix factorization with temporal continuity and sparseness criteria. IEEE Transactions on Audio, Speech, and Language Processing, 15(3):1066-1074, 2007. 6 +[58] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan L. Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[59] Alexander Wettig, Tianyu Gao, Zexuan Zhong, and Danqi Chen. Should you mask $15\%$ in masked language modeling? In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2985-3000, 2023. 2 +[60] Jiantao Wu and Shentong Mo. Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338, 2022. 2 +[61] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1326–1335, 2021. 2 +[62] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 6291–6299, 2019. 2 +[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[64] Xudong Xu, Bo Dai, and Dahua Lin. Recursive visual sound separation using minus-plus net. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6 +[65] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 2, 5, 6 +[66] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1735-1744, 2019. 2 +[67] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng + +Wang, and Yiran Zhong. Audio-visual segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 6 +[68] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 6 \ No newline at end of file diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/images.zip b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0cf8bd80100bcc2888acb84f752c06db10d22a24 --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a774f1fe0e70f75a825ae220ff6fb77ea6d507b0d07984f3902d624592b961b +size 393985 diff --git a/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/layout.json b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d87f1f8f9561a09e995c00699160fd0e35642f5 --- /dev/null +++ b/2024/Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling/layout.json @@ -0,0 +1,8976 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense Interactions through Masked Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 162, + 233, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 162, + 233, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 162, + 233, + 175 + ], + "type": "text", + "content": "Shentong Mo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 177, + 268, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 177, + 268, + 191 + ], + "spans": [ + { + "bbox": [ + 131, + 177, + 268, + 191 + ], + "type": "text", + "content": "Carnegie Mellon University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 193, + 279, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 193, + 279, + 205 + ], + "spans": [ + { + "bbox": [ + 120, + 193, + 279, + 205 + ], + "type": "text", + "content": "shentong.mo@sv.cmu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 353, + 162, + 431, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 162, + 431, + 175 + ], + "spans": [ + { + "bbox": [ + 353, + 162, + 431, + 175 + ], + "type": "text", + "content": "Pedro Morgado" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 177, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 177, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 177, + 474, + 190 + ], + "type": "text", + "content": "University of Wisconsin-Madison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 193, + 454, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 193, + 454, + 205 + ], + "spans": [ + { + "bbox": [ + 329, + 193, + 454, + 205 + ], + "type": "text", + "content": "pmorgado@wisc.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 153, + 229, + 436, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 229, + 436, + 243 + ], + "spans": [ + { + "bbox": [ + 153, + 229, + 436, + 243 + ], + "type": "text", + "content": "https://github.com/stoneMo/DeepAVFusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 265, + 192, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 265, + 192, + 278 + ], + "spans": [ + { + "bbox": [ + 143, + 265, + 192, + 278 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 293, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 290, + 617 + ], + "type": "text", + "content": "Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Humans naturally integrate audio-visual information to perceive and understand the environment. Several studies in cognitive psychology and neuroscience, as well as classic perceptual illusions such as the McGurk effect, demonstrate" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 266, + 548, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 548, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 548, + 447 + ], + "type": "text", + "content": "that such audio-visual fusion can occur early on in the perceptual stack, enabling deeper integration of the two modalities. Early fusion models aim to emulate this human-like perception. Specifically, it refers to the process of integrating auditory and visual cues at an early stage of a multi-modal perception model, in order to leverage the synergistic effects of both modalities. This can be especially powerful if the fusion process can attend to and establish connections between local components of the visual and audio signal (e.g., the frequencies of someone's voice with the pixels of their lips). In sum, early fusion of local audio-visual interactions holds the promise of a deeper and more sophisticated understanding of audio-visual content, critical for many real-world applications such as visually guided source separation, localization/segmentation, and multi-modal recognition." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 548, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 548, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 548, + 628 + ], + "type": "text", + "content": "Despite the potential, audio-visual early fusion architectures have remained under-explored, with the majority of the recent literature focusing on late fusion or no fusion at all. One recent prominent research direction [1, 47] seeks to learn independent uni-modal encoders through contrastive learning by requiring the representations of associated audio-visual pairs to be synchronized in latent space. However, contrastive learning is not compatible with early fusion, as the connections between uni-modal representations lead to trivial solutions to the contrastive learning problem. Another line of recent work [20, 27] learns encoders with late audio-visual fusion through a combination of contrastive learning and masked reconstruction, focusing on downstream tasks that require shallow integration of audio-visual features, such as audio-visual action recognition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "content": "One potential reason is that early fusion architectures pose a significant training challenge due to the substantial increase in model expressivity it offers. As a result, robust learning frameworks are required to ensure the creation of high-quality models and effectively harness their enhanced capabilities. In this work, we demonstrate the effectiveness of a masked reconstruction framework for training audio-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27186" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 164, + 212 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 164, + 212 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 164, + 212 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 164, + 212 + ], + "type": "image", + "image_path": "5c8228083643d84dcd15370528386369b224f674d9c48f231fdd14daaa2744e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 257 + ], + "type": "text", + "content": "Figure 1. Unlike prior works on Audio-Visual MAEs which learn either late fusion encoders or separate encoders altogether, we demonstrate that masked reconstruction is especially useful for learning deeply integrated audio-visual encoders with powerful early fusion modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 77, + 284, + 213 + ], + "blocks": [ + { + "bbox": [ + 169, + 77, + 284, + 213 + ], + "lines": [ + { + "bbox": [ + 169, + 77, + 284, + 213 + ], + "spans": [ + { + "bbox": [ + 169, + 77, + 284, + 213 + ], + "type": "image", + "image_path": "0f834e35a30fa9a2ae49b6f5ca1133773d70b001ec1f968ce3dc04d7364add01.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 472 + ], + "type": "text", + "content": "visual transformers with early-fusion of local interactions (Fig. 1). Our hypothesis is that the simultaneous reconstruction of audio and visual inputs from a limited audio-visual context promotes the learning of deeply integrated representations, as the model is required to understand the fine interactions between the two modalities. Furthermore, masked reconstruction has already been successfully deployed in training uni-modal representations for a variety of input signals, including text [11], images [3, 22], and audio [28], as well as for learning late fusion representations [20, 27]. However, unlike what was previously observed with uni-modal or late fusion models, our work shows that masked audio-visual reconstruction with early fusion encoders results in an interesting emergent property. While uni-modal representations encode the low-level details required for masked reconstruction, we observed that high-level semantics (surprisingly) emerged from the tokens used for audio-visual fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": "In addition to demonstrating the effectiveness of masked reconstruction for early fusion transformers, we propose a novel attention-based fusion module, which can effectively attend to the interactions between local audio and visual representations, thereby enhancing its capacity to capture localized interactions between the two modalities. Take, for instance, the sound of a dog barking. The sound signals the presence of a dog and offers clues about its state, such as an open mouth. While the link between the dog and its bark can be made at a global level, our fusion module allows the model to delve deeper into local interactions, like the correlation between the dog's mouth and the barking sound's spectral frequencies. As a result, localized fusion boosts the quality of learned representations for numerous tasks. However, it necessitates the model to explicitly represent all pairwise interactions of local audio and visual tokens. Given the vast number of tokens that current transformers process, dense local interactions can quickly become unmanageable. To enhance efficiency, our fusion module factorizes unimodal representations, thereby circumventing the need to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "content": "model all pairwise interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 86, + 548, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 86, + 548, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 86, + 548, + 254 + ], + "type": "text", + "content": "In sum, we propose an attention-based fusion module that can attend to local audio-visual interactions and deploy it to learn deeply integrated audio-visual representations by fusing uni-modal representations at early layers in the architecture. We also demonstrate that when paired with masked reconstruction, our framework, denoted DEEPAVFUSION, can learn strong audio-visual representations, yielding state-of-the-art performance on a variety of audio-visual tasks, including visually guided source separation, localization, and segmentation. We extensively evaluated DEEPAVFUSION on a variety of audio-visual datasets, and conducted thorough ablation studies, showing the importance of early fusion, dense interactions, and uni-modal pre-training when learning deeply integrated audio-visual representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 275, + 392, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 392, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 392, + 288 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 291, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 434 + ], + "type": "text", + "content": "Audio-Visual Representations Learning. Audio-visual representations learning has been addressed in many previous works [1, 2, 14, 16, 23, 24, 30, 40, 43, 45–47, 50, 51, 65, 66] to learn the audio-visual correlation between two distinct modalities from videos. Such cross-modal alignments are beneficial for many audio-visual tasks, such as audio-event localization [31, 32, 37, 41, 42, 53, 62], audio-visual spatialization [4, 17, 44, 45], audio-visual navigation [4–6] and audio-visual parsing [33, 39, 54, 61]. In this work, our main focus is to learn transferrable audio-visual representations from masked audio-visual reconstruction, which is more challenging than the above-mentioned tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "text", + "content": "Masked Representations Learning. Masked representation learning aims to learn self-supervised representations by reconstructing desired features of masked data given unmasked parts as clues. In the recent years, masked representation learning has achieved promising results in natural language processing [9, 11, 35, 52, 59] and computer vision [3, 8, 12, 15, 22, 58, 60, 63] community. Typically, BERT [11] randomly masked " + }, + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "text", + "content": " of word tokens in the sentence and recovered them with unmasked words to learn generalizable textual features via a self-attention transformer [56]. A block-wise masking strategy was proposed in BEiT [3] to reconstruct discrete tokens of masked image patches for pre-training transferrable visual representations. To simplify the masked image encoding framework, MAE [22] directly reconstructed missing pixels of " + }, + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 436, + 548, + 640 + ], + "type": "text", + "content": " masked patches using vision transformers [13] for self-supervised pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Audio-visual Masked Autoencoders. More recently, researchers introduced diverse masking pipelines [19, 20, 27] to show the effectiveness of masked modeling in learning audio-visual representations. For example, CAV-MAE [20] combined contrastive learning and masked modeling to capture a joint and coordinated audio-visual representation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "They tried to add a joint encoder on audio-visual features from the last attention block of single-modality encoders to fuse cross-modal information for audio-visual contrastive objectives. MAViL [27] extended masked audio-video reconstruction with masked intra- and inter-modal contrastive learning and self-training by recovering joint audio-video contextualized representations. Despite their promising performance, they ignored the importance of the early fusion of audio-visual features in masked audio-visual reconstruction. In contrast to AV-MAE [19] that simply concatenated the audio and visual tokens before passing them through the joint transformer for early fusion, we will design an early fusion module with interactions between local and visual representations for audio-visual masked auto-encoders. Our audio-visual interactions are different from fusion bottlenecks in MBT [48] that forces information between different modalities to pass through a small number of bottleneck latent. However, we develop a fully novel attention-based fusion module that can effectively attend to interactions between local audio and visual representations, thus enhancing its ability to capture fine-grained interactions between the two modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 337, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 289, + 613 + ], + "type": "text", + "content": "Audio-visual Early Fusion. Audio-visual fusion has been the topic of extensive research, with several works proposing a variety of architectures to aggregate multi-modal representations [19, 48, 49]. An example of an early fusion architecture was proposed by Owens et al. [49] to learn representations from audio-visual correspondences, by concatenating features from small unimodal encoders and feeding them into a fused audio-visual network for the early fusion of multisensory features. Beyond this early work, most recent papers on audio-visual fusion [19, 48] propose alternative fusion mechanisms, either through weight sharing [19] or token-based fusion [48]. However, since all these works focused on downstream tasks (mostly classification) that do not require fine multi-modal understanding, the benefits of early fusion were not evident, with mid-level fusion often achieving optimal performance. Different from them, we proposed an early fusion model that is designed for fine multi-modal understanding, by attending to local interactions between the audio and the visual while performing early fusion. The new architecture design paired with masked reconstruction objective is shown to outperform many of the prior works on a variety of downstream multi-modal applications beyond classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 631, + 104, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 104, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 104, + 644 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": "We aim to learn audio-visual representations that can be effectively transferred to a variety of downstream tasks that require a detailed understanding of audio-visual interactions. To accomplish this, we introduce a novel audio-visual transformer architecture, named DEEPAVFUSION, that enables the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "early fusion of audio and visual tokens through dense local interactions. We also show how to effectively train such early fusion transformers using multi-modal masked reconstruction. In this section, we first describe the proposed early fusion transformer and then introduce the multi-modal masked reconstruction framework used for self-supervised pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 163, + 547, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 547, + 188 + ], + "type": "text", + "content": "3.1. A Transformer Architecture for Joint Audio-Visual Encoding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 196, + 547, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 291 + ], + "type": "text", + "content": "Downstream tasks such as visually guided sound source separation, localization, and segmentation require a deep understanding of fine audio-visual interactions present in the data. However, late fusion often lacks the expressive power to represent such interactions. We introduce a transformer architecture that enables the early fusion of multi-modal representations through factorized local interactions between audio and visual tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": "Early fusion architecture While we aim to learn deeply integrated audio-visual representations, foundational uni-modal transformers can still provide a strong starting point for multi-modal learning. As such, we introduce a modular architecture to repurpose existing pre-trained uni-modal transformers. Specifically, we propose a three-branch architecture, illustrated in Fig. 2. The first two branches are uni-modal transformers that process audio and visual tokens, respectively, while the third branch is a multi-modal transformer that updates a set of learnable fusion tokens to fuse audio-visual information. Formally, let " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0^f \\in \\Re^{F \\times D}" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": " learnable fusion tokens. We progressively update these tokens, at each layer " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": ", through a fusion block " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\Phi_l^f" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": ", which can interact with and aggregate information from modality-specific features " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{l - 1}^v" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{l - 1}^a" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": ". Furthermore, fusion tokens " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{l - 1}^f" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": " are fed into modality-specific blocks, " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\Phi_l^v" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "inline_equation", + "content": "\\Phi_l^a" + }, + { + "bbox": [ + 304, + 293, + 548, + 511 + ], + "type": "text", + "content": ", to modulate uni-modal representations from early stages in the network. Overall, representations are updated as follows" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 353, + 514, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 514, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 353, + 514, + 545, + 529 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {l} ^ {f} = \\Phi_ {l} ^ {f} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {a}\\right) \\tag {1}", + "image_path": "9cd2290560cf32263db66e2b8ae821d652e7c23041a1d59a353a2678fc148415.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 532, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 532, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 354, + 532, + 545, + 547 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {l} ^ {v} = \\Phi_ {l} ^ {v} \\left(\\boldsymbol {X} _ {l - 1} ^ {v}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {2}", + "image_path": "6290845ae17541c91db0fc920590e9456c4936b5dde9a7cd85517ba25350ebb8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 549, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 549, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 354, + 549, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {l} ^ {a} = \\Phi_ {l} ^ {a} \\left(\\boldsymbol {X} _ {l - 1} ^ {a}, \\boldsymbol {X} _ {l - 1} ^ {f}\\right) \\tag {3}", + "image_path": "4dc327da9f6badc5713fc08cc66f79cdccf0befef45eef0f5d7fea0140dfe609.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "text", + "content": "This token update strategy is repeated for a total of " + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "text", + "content": " layers to compute the final representations " + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_L^v, \\mathbf{X}_L^a" + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_L^f" + }, + { + "bbox": [ + 304, + 567, + 545, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Modality-specific blocks We use standard transformer self-attention for the modality-specific blocks, " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^v" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^a" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": ", while allowing them to further attend to fusion tokens. Unless otherwise specified, we initialize " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^v" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^a" + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": " from foundational uni-modal models. This is possible since the parameter space of the modality-specific blocks remains unchanged. In particular, we empirically validated the benefits of bootstrapping from [28] and [22], which are pre-trained on large-scale audio and visual datasets, respectively, through a uni-modal masked reconstruction objective." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27188" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 244, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 244, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 244, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 244, + 183 + ], + "type": "image", + "image_path": "5115c4a420005adddb207674d25a2c220139b865387eab1d8e5f06701148c7fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": "Figure 2. Audio-Visual Early Fusion Transformer with Dense Interactions. The architecture ① is composed of three interconnected branches - the visual, audio, and fusion branches - each with an equal number of transformer blocks. The visual and audio branches process visual and audio tokens, respectively, while simultaneously attending to fusion tokens. The fusion branch updates a set of learnable fusion tokens to fuse audio-visual information. To better attend to local audio-visual interactions, we introduce a new fusion block ②. At each layer " + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": ", all patch-level audio (visual) tokens, " + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\bar{X}_{l-1}^{a}(\\bar{X}_{l-1}^{v})" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": ", are first aggregated into a small set of aggregation tokens, " + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\bar{X}_l^a (\\bar{X}_l^v)" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": " by cross-attention. Pairwise audio-visual interactions are then encoded through a linear layer into latents " + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\bar{X}_l^{av}" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": ", which are used to update the fusion tokens " + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\bar{X}_l^f" + }, + { + "bbox": [ + 46, + 185, + 547, + 247 + ], + "type": "text", + "content": " by cross-attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 246, + 78, + 533, + 184 + ], + "blocks": [ + { + "bbox": [ + 246, + 78, + 533, + 184 + ], + "lines": [ + { + "bbox": [ + 246, + 78, + 533, + 184 + ], + "spans": [ + { + "bbox": [ + 246, + 78, + 533, + 184 + ], + "type": "image", + "image_path": "fd6002b3aba7b5ca98a5b9a39fdcbf97ad92f5bf452d5e485275ef3a59ce7e25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": "Fusion blocks The design of the fusion blocks " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\Phi^f" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " is critical to representing fine audio-visual interactions. In the context of transformers, a natural way to combine audio-visual representations is through a standard self-attention block which, in addition to the fusion tokens " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "X^f" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " themselves, it can further attend to all image and audio tokens, " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "X_{l - 1}^v" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "inline_equation", + "content": "X_{l - 1}^{a}" + }, + { + "bbox": [ + 46, + 252, + 289, + 396 + ], + "type": "text", + "content": ". This approach, which we refer to as token fusion, has been used in previous works such as MBT [48]. While intuitive, standard self-attention cannot directly exploit interactions between local audio-visual representations, which limits their expressivity. To address this limitation, we designed a fusion block that attends to local interactions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": "Fusion with Dense Audio-Visual Interactions Interactions between audio and visual data occur between local regions in the image and the audio spectrogram. Given a set of " + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "inline_equation", + "content": "n_a" + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": " audio tokens " + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "inline_equation", + "content": "X^a" + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": " visual tokens " + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "inline_equation", + "content": "X^v" + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": ", aggregation of local multi-modal interactions requires attending to representations from all possible " + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "inline_equation", + "content": "n_a \\times n_v" + }, + { + "bbox": [ + 46, + 398, + 289, + 531 + ], + "type": "text", + "content": " pairs, to identify the ones that are most useful for the task at hand. While we experimented with several ways to represent multi-modal interactions, from bilinear aggregation to kernel-based aggregation, we found that linear aggregation was sufficient to capture audio-visual interactions for our tasks. Formally," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 539, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 539, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 63, + 539, + 287, + 554 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {i j} ^ {a v} = W ^ {a} \\boldsymbol {X} _ {i} ^ {a} + W ^ {v} \\boldsymbol {X} _ {j} ^ {v} \\tag {4}", + "image_path": "0345822f983e342c362b2c9adc6c68d4a3a69dc0f5e396c7816dbec6d858fd1d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 555, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 555, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 63, + 555, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {a v} = \\left[ \\boldsymbol {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {a} \\forall j = 1, \\dots , n _ {v} \\right]. \\tag {5}", + "image_path": "aa28386df01ba81d4831532ebb057875f28dd72e47b3027cfb57be55bb44c841.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "content": "Fusion tokens " + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "inline_equation", + "content": "X_{l}^{f}" + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "text", + "content": " are then updated using standard cross-attention [56] over the representation of all local interactions " + }, + { + "bbox": [ + 46, + 580, + 288, + 616 + ], + "type": "inline_equation", + "content": "X^{av}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 624, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 624, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 87, + 624, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}; \\boldsymbol {X} _ {l - 1} ^ {\\text {a v}}\\right). \\tag {6}", + "image_path": "f2479eef6ac70d43c05eb1fb7a193ff2f8d6a5349eac0aa5235cddbf2b0ab96f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Factorized Audio-Visual Interactions While the procedure above can attend to local interactions, it requires the model to explicitly represent all possible " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "n_a \\times n_v" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " pairs. Given that the number of audio and visual tokens is typically large, the number of interactions can quickly become intractable," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "content": "resulting in high memory consumption and low throughput. To address this limitation, we introduce modality-specific aggregation tokens updated by aggregation blocks. The goal of the aggregation blocks is to summarize the large number of audio/visual tokens " + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "X^a / X^v" + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "content": " into a small set of " + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "n_{\\bar{a}} / n_{\\bar{v}}" + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "content": " aggregation tokens " + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "\\bar{X}^a / \\bar{X}^v" + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "content": ", respectively. Specifically, at layer " + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 253, + 547, + 338 + ], + "type": "text", + "content": ", aggregation tokens are updated by cross-attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 345, + 545, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 345, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 340, + 345, + 545, + 359 + ], + "type": "interline_equation", + "content": "\\bar {X} _ {l} ^ {a} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {a}, X _ {l - 1} ^ {a}\\right) \\tag {7}", + "image_path": "04165125d6b1066f0b498f097dffaa4cc6e0da765dec162710571aead885e386.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 361, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 361, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 341, + 361, + 545, + 376 + ], + "type": "interline_equation", + "content": "\\bar {X} _ {l} ^ {v} = \\text {C r o s s A t t e n t i o n} \\left(\\bar {X} _ {l - 1} ^ {v}, X _ {l - 1} ^ {v}\\right). \\tag {8}", + "image_path": "7090ab579c6f6554ab50818a38c1127d2b2d5e5ea569469d94c70b7492e9a0f1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "content": "Then, instead of explicitly representing all " + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "inline_equation", + "content": "n_a \\times n_v" + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "content": " pairwise interactions, we restrict audio-visual interactions to the sets of aggregated tokens. Formally, factorized interactions between aggregated tokens " + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\bar{X}_i^a" + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\bar{X}_j^v" + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "content": " are represented as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 441, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 441, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 324, + 441, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\bar {X} _ {i j} ^ {a v} = W ^ {\\bar {a}} \\bar {X} _ {i} ^ {a} + W ^ {\\bar {v}} \\bar {X} _ {j} ^ {v} \\tag {9}", + "image_path": "b3559651cd811d0c60b95ef3f31a966b3d00739572083e49f695e51c5b62790c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 457, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 457, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 324, + 457, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\bar {X} ^ {a v} = \\left[ \\bar {X} _ {i j} ^ {a v}; \\forall i = 1, \\dots , n _ {\\bar {a}} \\forall j = 1, \\dots , n _ {\\bar {v}} \\right] \\tag {10}", + "image_path": "10a3cc0fc0f3382d747fa9697b1113005d8a1025857a245210cd0af9ba86bdc9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 480, + 493, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 493, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 493, + 492 + ], + "type": "text", + "content": "and the fusion tokens are updated at layer " + }, + { + "bbox": [ + 306, + 480, + 493, + 492 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 480, + 493, + 492 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 501, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 501, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 348, + 501, + 545, + 521 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {l} ^ {f} = \\text {C r o s s A t t e n t i o n} \\left(\\boldsymbol {X} _ {l - 1} ^ {f}, \\bar {\\boldsymbol {X}} _ {l} ^ {a v}\\right). \\tag {11}", + "image_path": "da7a6cde5a53db1491805f2361892c489ec95a60c6b657cbf01deee1711fe7f2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 528, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 547, + 576 + ], + "type": "text", + "content": "Since we only represent interactions between the aggregated uni-modal tokens, the number of possible interactions is greatly reduced. We observed that a " + }, + { + "bbox": [ + 304, + 528, + 547, + 576 + ], + "type": "inline_equation", + "content": "700 \\times" + }, + { + "bbox": [ + 304, + 528, + 547, + 576 + ], + "type": "text", + "content": " reduction still maintains the benefits obtained with local dense interactions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 585, + 545, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 545, + 610 + ], + "type": "text", + "content": "3.2. Learning Early-Fusion Transformers Through Audio-Visual Masked Auto-Encoding" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Masked auto-encoders have been shown to learn effective representations for a wide variety of input signals, including images [3, 22] and audio [28]. Formally, the input signal " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " is partitioned into a set of patch tokens " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "x = \\{x_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", which is randomly split into the visible set " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}} = \\{x_{t}\\}_{t \\in \\mathcal{V}}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and the masked set " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mask}} = \\{x_{t}\\}_{t \\in \\mathcal{M}}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " are non-overlapping index sets). In the case of images, visible and masked tokens, " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", are obtained from a grid" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27189" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "of small image patches [3, 22], while in the case of audio signals, tokens are obtained either from small windows of a raw waveform or from small 2D patches of the audio spectrogram [28]. Then, to learn representations of the input data, a transformer encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is first used to encode the visible tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}} = f(X_{\\mathrm{vis}}) \\in \\Re^{|\\mathcal{V}| \\times d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " denotes the dimensionality of the token representation. These representations together with a series of (position marked) mask tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are then fed to a decoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " that seeks to reconstruct the masked input tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{mask}} = g(X_{\\mathrm{vis}}, M_{\\mathrm{mask}})" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". To accomplish that, the model is trained to minimize the mean squared reconstruction error" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 220, + 288, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 288, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}}, \\hat {X} _ {\\text {m a s k}}\\right) = \\frac {1}{| \\mathcal {M} |} \\sum_ {\\substack {x \\in X _ {\\text {m a s k}} \\\\ \\hat {x} \\in \\hat {X} _ {\\text {m a s k}}}} \\| x - \\hat {x} \\| _ {2} ^ {2}. \\tag{12}", + "image_path": "e0bceb9f0a59e8002edb56c22aec25f63ed5b4a0bbbd707b99820f3c2f90dfe6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "content": "Following prior work [20, 27], masked reconstruction can be easily extended to multi-modal signals, by simultaneously masking tokens from both audio and visual modalities. In the case of early-fusion transformers, the encoder " + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "content": " is then used to encode the visible tokens, " + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}}^v" + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}}^a" + }, + { + "bbox": [ + 46, + 266, + 287, + 327 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 331, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 331, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 102, + 331, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}, \\boldsymbol {X} ^ {f} = f \\left(\\boldsymbol {X} _ {\\mathrm {v i s}} ^ {v}, \\boldsymbol {X} _ {\\mathrm {v i s}} ^ {a}\\right) \\tag {13}", + "image_path": "951b7070ef71b1a9bce3e4854f292111f56613156aab3326c6d53ffc826acff8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": "which, in addition to the uni-modal representations, " + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}}^v" + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{vis}}^a" + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": ", also returns a set of fusion representations " + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\pmb{X}^{f}" + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": ". These representations are then fed to modality-specific decoders to reconstruct both audio and visual masked tokens, " + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{mask}}^v = g_v(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^v)" + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{mask}}^a = g_a(X_{\\mathrm{vis}}^f,M_{\\mathrm{mask}}^a)" + }, + { + "bbox": [ + 47, + 350, + 289, + 422 + ], + "type": "text", + "content": ", by minimizing the loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 427, + 287, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 287, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A V - M A E}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {v}, \\hat {X} _ {\\text {m a s k}} ^ {v}\\right) + \\mathcal {L} _ {\\mathrm {M A E}} \\left(X _ {\\text {m a s k}} ^ {a}, \\hat {X} _ {\\text {m a s k}} ^ {a}\\right) \\tag {14}", + "image_path": "28792552c0f2cab2f84c50c931302f71036208b3a2ba5af60c019b5a3ab157f9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 458, + 238, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 238, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 238, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 458, + 238, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 47, + 458, + 238, + 470 + ], + "type": "text", + "content": " is the auto-encoder loss of Eq. 12." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "content": "While multi-modal masked reconstruction has been previously explored in [20, 27], these approaches focus on learning separate audio and image encoders, " + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "content": ", with multi-modal representations obtained through a late fusion module " + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "content": ". In contrast, we propose to learn a single multimodal early fusion encoder " + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 472, + 288, + 591 + ], + "type": "text", + "content": ", and demonstrate that masked reconstruction is especially useful for learning multi-modal representations capable of representing fine audio-visual interactions through early fusion of local interactions between audio and visual tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 603, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 128, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 661 + ], + "type": "text", + "content": "We now evaluate the proposed DEEPAVFUSION on a variety of audio-visual downstream tasks that require a deep understanding of audio-visual interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 669, + 178, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 178, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 178, + 681 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Pre-training. For pre-training, we experiment with two datasets, namely VGG-Sounds [7] (with 144k training sam" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "ples) and AudioSet [18] (with 1.73M training samples), both of them containing diverse videos focused on audio events and sound sources. The audio is represented by log mel spectrograms extracted from " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "3s" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " of audio at a sample rate of " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "16000\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". To compute the log spectrograms, we apply an STFT using approximately 50ms windows with a hop size of " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "15\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ", resulting in an input tensor of size " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "128\\times 196" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " (128 mel frequency bands over 196 time steps). Since we use image models for the visual component, we randomly extract single frames from within the time window defined by the " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "3s" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " audio snippets. Input frames are augmented by random crops with a minimum area of 0.5 and random horizontal flips, and resized into a " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "type": "text", + "content": "Unless otherwise specified, we initialize the uni-modal transformer blocks using modality-specific encoders pretrained with uni-modal masked reconstruction objectives [22, 28]. Specifically, we use the ViT-Base model from [22] pre-trained on ImageNet [10], and the Spec-MAE model (with the same ViT-Base architecture) pre-trained on AudioSet [18]. In our base configuration, we add multi-modal fusion blocks on all 12 layers of the model, together with 16 multi-modal fusion tokens, 8 visual aggregation tokens, and 8 audio aggregation tokens. For efficiency, we reduced the dimensionality of the MLP embeddings using an MLP ratio of 1 (instead of the standard 4), as well as the dimensionality of space in which the similarity for self and cross-attention is computed to 16 (as opposed to the standard 64)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "text", + "content": "The models were trained using the Adam optimizer [29]. Ablation and parametric studies were conducted on VG-GSound, training for 200 epochs with a total batch size of 512 (split between 2 GPUs and 4 iterations of gradient accumulation). Large-scale training on AudioSet was conducted for 200 epochs with a total batch size of 2048 (split between 16 GPUs and 4 iterations of gradient accumulation). In all cases, we use a base learning rate of " + }, + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "inline_equation", + "content": "1.5e - 4" + }, + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "text", + "content": " adjusted for the effective batch size by the linear scaling rule " + }, + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "inline_equation", + "content": "lr = blr \\times bs / 256" + }, + { + "bbox": [ + 304, + 399, + 547, + 543 + ], + "type": "text", + "content": " [21], a 40 epoch learning rate warm-up schedule followed by cosine learning rate decay, and a weight decay of 0.05." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 544, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 568 + ], + "type": "text", + "content": "Downstream tasks. After pre-training, we fine-tune our models on a variety of downstream tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "type": "text", + "content": "Visually-guided sound source separation. Following [65], we use a mix-and-separate strategy for training. The pretrained model is used to obtain audio-visual representations that are fed to a UNet decoder to generate the separated audio. We use a similar separation strategy by predicting a separation mask that is directly applied to the input mixture SFTF. To compute the separation masks, we deploy a 10-layer UNet decoder, which conditions its predictions on (1) the spatially pooled visual representations and (2) five intermediate audio representations obtained during the encoding stage at layers 1, 3, 6, 9, and 12. Following [65], the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27190" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "predicted separation masks are trained to regress the STFT magnitude ratio between the separated source and the mixture at each time-frequency bin. We evaluate on three different datasets: the MUSIC [65] dataset and two subsets from VGG-Sounds [7]. MUSIC [65] consists of 448 untrimmed YouTube music videos of solos and duets from 11 instrument categories, where we use 358 solo videos for training and 90 solo videos for evaluation. This dataset was slightly smaller than the original MUSIC dataset since some videos are no longer publicly available to be downloaded. We also use two subsets of VGG-Sounds [7]: the VGGSound-Instruments subset [25] which includes 32k video clips of 10s lengths from 36 musical instrument classes, and a broader newly defined \"VGGSound-Music\" subset which includes 40,908 video clips from 49 music categories for training and 1201 clips for testing. As commonly done in the source separation literature, we measure performance using Signal-to-Distortion Ratio (SDR), Signal-to-Artifact Ratio (SAR), and Signal-to-Interference Ratio (SIR)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "type": "text", + "content": "Audio-visual segmentation. We fine-tune our models on AVSBench [67] for the tasks of Single Source Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS). Task S4 aims to predict pixel-level segmentation maps of the visible sound sources, while AVSS further requires the model to recognize the class of the sound source (in addition to the segmentation maps). We use the standard train/val/test split [67] and measure performance using mIoU and F1 scores. Similarly to the source separation task, audio-visual representations, obtained from the pre-trained model, are fed to a UNet decoder to generate the segmentation maps. This decoder conditions its predictions on a globally pooled audio representation and intermediate visual feature maps obtained at layers 1, 3, 6, 9, and 12." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 471, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 289, + 615 + ], + "type": "text", + "content": "Recognition through Linear Probing or Finetuning. We further assess the pre-trained representations on multi-modal recognition tasks by conducting linear probing and finetuning evaluations on the VGGSound-Music, VGGSound-All, and the balanced AudioSet dataset. In all cases, we simply attach three linear classifiers on the average pooled visual, audio, and fusion representations, respectively. The three classification logits are then averaged to obtain the final prediction. As usual, on VGGSound datasets, we measure performance by class accuracy, while on AudioSet (a multi-label dataset) we use the average precision (AP) and Area under the ROC curve (AUC) scores averaged across classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 633, + 194, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 194, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 194, + 646 + ], + "type": "text", + "content": "4.2. Comparison to prior work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed method, we extensively compare it to prior work on audio-visual downstream tasks that require a detailed understanding of audio-visual data. Unless otherwise specified, we use our model trained on VGGSound for comparisons to prior work," + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 91, + 547, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 89 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 89 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 89 + ], + "type": "text", + "content": "Table 1. Sound source separation performance on the MUSIC and VG-GSound datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 91, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 547, + 182 + ], + "type": "table", + "html": "
MethodMUSICVGGS-InstrumentsVGGS-Music
SDRSIRSARSDRSIRSARSDRSIRSAR
NMF [57]-0.892.386.28-3.520.786.95-5.060.157.03
RPCA [26]0.784.627.580.362.387.95-1.681.578.26
SoP [65]3.627.959.632.725.679.851.564.5910.15
MP-Net [64]3.988.369.863.056.1210.171.955.0210.59
CCoL [55]5.278.7510.524.076.9310.852.735.8611.03
OneAVM [38]6.2710.6812.355.897.8511.233.676.5311.85
DEEPÄVFUSION7.9512.4112.806.959.5213.235.798.2413.82
", + "image_path": "807c906ab07cd6ffdb81c47ce6cdd8635006b1713a25d5cf44423ef69083b314.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 339, + 221, + 517, + 296 + ], + "blocks": [ + { + "bbox": [ + 305, + 191, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 547, + 220 + ], + "type": "text", + "content": "Table 2. Audio-visual segmentation performance on the AVSBench dataset for the Sound Source Segmentation (S4) and Audio-Visual Segmentation with Semantics (AVSS) tasks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 221, + 517, + 296 + ], + "lines": [ + { + "bbox": [ + 339, + 221, + 517, + 296 + ], + "spans": [ + { + "bbox": [ + 339, + 221, + 517, + 296 + ], + "type": "table", + "html": "
MethodAVSBench-S4AVSBench-AVSS
mIoUF1mIoUF1
AVS [67, 68]78.7487.9029.7735.2
LAVISH [34]80.10---
MMVAE [36]81.7490.10--
DEEPAVFUSION89.9492.3452.0558.29
", + "image_path": "cefcd34699574bc96202a54e085dfac9916e57e29a175462ba9f76aeb35d3329.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 301, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 324 + ], + "type": "text", + "content": "as this enables us to make the fairest comparisons with the largest number of works that require pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 578 + ], + "type": "text", + "content": "Sound source separation. Table 1 shows the comparison between DEEPAVFUSION and several prior works, including traditional signal processing methods, NMF [57] and RPCA [26], deep learning methods specialized on the source separation task, Sound-of-Pixels [65] and MP-Net [64], and deep learning methods trained for simultaneous localization and separation, CCoL [55] and OneAVM [38]. While all listed deep learning approaches to source separation use (supervised) ImageNet pre-trained visual encoders, our model shows strong performance using completely unsupervised pre-training. Table 1 shows that DEEPAVFUSION outperforms prior works on all metrics. For example, on VGGSound-Music, we outperform the recently proposed OneAVM [38] by 2.12 SDR and 1.71 SIR, which uses supervised pre-training for the vision component and uses source separation as just one of multiple pre-training objectives. Although the different models use different architectures for their audio-visual encoders (which make direct comparison challenging to analyze), the improvements demonstrated on this task indicate that visual source separation can benefit from early-fusion to extract rich representations from fine audio-visual interactions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "type": "text", + "content": "Audio-Visual Segmentation. We assess the effectiveness of DEEPAVFUSION for audio-visual segmentation on the AVS-Bench dataset [67, 68]. We compared the proposed approach to recent state-of-the-art methods (AVS [67], LAVISH [34] and MMVAE [36]) on the S4 and AVSS downstream tasks (sound source segmentation without and with semantics, respectively). We compare against the mIoU and F1 scores shown in the original papers in Table 2. As can be seen, DEEPAVFUSION outperforms all prior work on both tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "type": "text", + "content": "Audio-visual classification. The above experiments demonstrate the effectiveness of DEEPAVFUSION on audio-visual" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27191" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 82, + 480, + 135 + ], + "blocks": [ + { + "bbox": [ + 160, + 70, + 432, + 80 + ], + "lines": [ + { + "bbox": [ + 160, + 70, + 432, + 80 + ], + "spans": [ + { + "bbox": [ + 160, + 70, + 432, + 80 + ], + "type": "text", + "content": "Table 3. Comparing DEEPAVFUSION trained on VGGSound and AudioSet dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 82, + 480, + 135 + ], + "lines": [ + { + "bbox": [ + 116, + 82, + 480, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 82, + 480, + 135 + ], + "type": "table", + "html": "
Pre-training DatasetPre-training EpochsVGGSound CLSAudioSet CLSVGGSS-Music SEPAVSBench S4
LP AccFt AccLP APFt APSDRSIRSARmIoUF1
VGGSound20053.0858.1930.2532.885.798.2413.8289.9492.34
AudioSet20053.0857.9132.6934.716.939.9313.4990.2792.49
", + "image_path": "ede8d27f6484fe165ef589999017011d21fa79a783625ede05f5ecf6b00b79b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 80, + 168, + 257, + 242 + ], + "blocks": [ + { + "bbox": [ + 47, + 137, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 137, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 288, + 167 + ], + "type": "text", + "content": "Table 4. Audio-visual classification performance on VGGSound-Music and VGGSound-All datasets, using linear probing (LP) and finetuning (FT) protocols. Performance is measured in terms of classification accuracy " + }, + { + "bbox": [ + 47, + 137, + 288, + 167 + ], + "type": "inline_equation", + "content": "[\\%]" + }, + { + "bbox": [ + 47, + 137, + 288, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 168, + 257, + 242 + ], + "lines": [ + { + "bbox": [ + 80, + 168, + 257, + 242 + ], + "spans": [ + { + "bbox": [ + 80, + 168, + 257, + 242 + ], + "type": "table", + "html": "
MethodVGGS-MusicVGGS-All
LPFTLPFT
(Image) MAE [22]48.2553.1837.1243.86
AudioMAE [28]52.7358.2942.0549.53
CAV-MAE [20]59.1867.5849.8355.32
DEEPAVFUSION65.3272.2553.0858.19
", + "image_path": "1c387c652b39b1492b8cf7aa5e79b8c3e4f79212c9e3489264c9cd9e12f2316c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 248, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 288, + 366 + ], + "type": "text", + "content": "tasks that require a detailed understanding of audio-visual interactions. We further evaluated the learned representations on recognition tasks. In particular, we compared our model to uni-modal masked reconstruction methods, MAE [22] and AudioMAE [28], as well as to CAV-MAE [20], a late-fusion model trained with both masked reconstruction and contrastive learning objectives. For a fair comparison, we use the same ViT-Base architecture for the uni-modal encoders and tune the model on image-audio pairs obtained from the VGG-Sound dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "content": "As can be seen in Table 4, we outperform both uni-modal pre-trained encoders on linear probing and fine-tuning by significant margins. DEEPAVFUSION outperforms the image MAE [22] and AudioMAE [28] by " + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "inline_equation", + "content": "19.62\\%" + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "inline_equation", + "content": "13.26\\%" + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "content": " on VGGSound-Music and " + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "inline_equation", + "content": "19.08\\%" + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "inline_equation", + "content": "12.49\\%" + }, + { + "bbox": [ + 46, + 369, + 287, + 502 + ], + "type": "text", + "content": " on the full VGSSound dataset. These direct comparisons to uni-modal encoders highlight the importance of learning joint representations for multi-modal recognition. Furthermore, our DEEPAVFUSION also significantly outperforms CAV-MAE which optimizes the same audio-visual reconstruction objective of Eq. 14 but does not leverage an early fusion encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 503, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 587 + ], + "type": "text", + "content": "It should be noted that multi-modal recognition could be further improved by learning from audio-video data (instead of audio-image pairs). However, video modeling is beyond the scope of this work. The improvements in Table 4 demonstrate nevertheless that even tasks like audio-visual classification of sound sources can benefit from more deeply integrated multi-modal representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 151, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 151, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 151, + 609 + ], + "type": "text", + "content": "4.3. Pre-training data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "The comparisons above were conducted using VGGSound for pre-training. To study the impact of pre-training data distribution and scale, we also trained our model on the larger AudioSet dataset [18] (with 1.73M training samples). We then tuned the resulting model on several downstream tasks, from classification on both VGGSounds and AudioSet to source separation on VGGSS-Musion and sound source segmentation on AVSBench (S4 task). As can be seen in" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 148, + 541, + 253 + ], + "blocks": [ + { + "bbox": [ + 315, + 137, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 533, + 148 + ], + "type": "text", + "content": "Table 5. Ablation study of major components of DEEPAVFUSION." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 148, + 541, + 253 + ], + "lines": [ + { + "bbox": [ + 315, + 148, + 541, + 253 + ], + "spans": [ + { + "bbox": [ + 315, + 148, + 541, + 253 + ], + "type": "table", + "html": "
(#)Unimodal \nPre-TrainingFusionLocal AV \nInteractionsLinear \nAccSep \nSDRSegm \nmIoU
(1)✓Early✓53.086.5352.05
(2)×Early✓44.362.0329.55
(3)✓Early×51.196.1548.71
(4)✓Mid✓51.215.8548.04
(5)✓Late✓46.525.3244.32
(6)✓None×39.674.2338.32
", + "image_path": "7cdd9a180b213bc9d6fa536f5a9af2483ee59f39a9ad720c5b251e54aa7c0c54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 260, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 547, + 333 + ], + "type": "text", + "content": "Table 3, pre-training on AudioSet yields better performance on most downstream tasks, demonstrating the scalability of the proposed method. The only exception is classification on VGGSound, where pre-training on VGGSound yields slightly better performance. This is likely due to a smaller domain gap between pre-training and fine-tuning data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 344, + 458, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 458, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 458, + 357 + ], + "type": "text", + "content": "4.4. Understanding DEEPAVFUSION" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "type": "text", + "content": "The previous section demonstrated the efficacy of early fusion of local audio-visual interactions, with the proposed method outperforming the state-of-the-art on a wide range of audio-visual tasks. In this section, we conduct a thorough analysis of our framework, revealing emergent properties in our model, and assessing the benefits of uni-modal pretraining, fusion depth, and other major design choices of the proposed methodology. To provide a full-picture comparison, we evaluate all models on five downstream tasks, namely linear probing and fine-tuning on VGGSound, sound source separation on VGGSound-Music, sound source localization on Flickr-SoundNet, and audio-visual segmentation with semantics on AVSBench." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": "Ablation studies. We begin our analysis with a series of ablation studies in Table 5 to measure the impact of three important components of the proposed framework, namely 1) uni-modal pre-training, 2) early fusion, and 3) audio-visual interactions. All models use the same backbone architecture for the uni-modal encoders and are trained to optimize the Audio-Visual MAE objective. Model #1 is the full DEEPAVFUSION framework, and the remaining models relax each of the three components in turn. Model #2 skips uni-modal pre-training, significantly hurting performance when compared to model #1 on all tasks, emphasizing the value of uni-modal pre-training, when learning deeply integrated audio-visual encoders. Model #3 performs early fusion but uses the \"token fusion\" strategy similar to that used in MBT [48]. Token fusion attends to uni-modal representations and cannot directly capture audio-visual inter" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27192" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 83, + 282, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 80 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 80 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 80 + ], + "type": "text", + "content": "Table 6. Exploration studies on the efficiency of factorized interactions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 83, + 282, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 282, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 282, + 148 + ], + "type": "table", + "html": "
Early FusionFactorized InteractionsSpeed (video/sec)Memory (GB)Linear AccSep SDRSegm mIoU
××14.510.346.525.3244.32
✓×4.331.652.826.3052.13
✓✓12.813.153.086.5352.05
", + "image_path": "656e011f347636236edf63343d465e4a1f592414bbafb53498c9100d939d4eab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 276, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 163, + 276, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 276, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 276, + 285 + ], + "type": "image", + "image_path": "4535a5279b37193eaf522f7851ff2594b05ded8ae7c184a6c07df064991a2def.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 288, + 316 + ], + "type": "text", + "content": "Figure 3. Nearest neighbor accuracy (NNAcc) vs training steps (Step) using Audio, Image, Fusion Tokens from an DEEPAVFUSION trained on VGGSounds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "content": "actions, limiting the expressivity of the fusion block, and the overall performance of the model. Models #4, #5 and #6 use the same fusion modules with dense interactions but perform mid-fusion (at layers 9-12), late-fusion (at layer 12), and no fusion at all. The results indicate that mid-fusion outperforms, late-fusion which in turn is substantially better than no-fusion. Nevertheless, early fusion achieves the best results across all downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "type": "text", + "content": "Efficiency of factorized interactions. While audio-visual interactions are crucial for the performance of early fusion, they also introduce a significant computational overhead, if not handled properly. Table 6 lists the VRAM consumption and model throughput during inference for three models: a baseline model without fusion, an early fusion model with dense interactions, and our proposed early-fusion model with factorized interactions. Dense interactions enhance the model's effectiveness on subsequent tasks, but at a high computational cost, both in terms of throughput (3.3 times slower) and memory usage (3.1 times higher). Our proposed factorized interactions significantly cut down the computational overhead, while maintaining (and even slightly enhancing) the performance of early fusion with dense interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Emergent semantics. In addition to robust performance across various tasks, the emergent semantics in fusion tokens are worth highlighting. Despite the masked auto-encoder objective typically favoring representations rich in low-level details, fusion tokens surprisingly yield higher-level semantic representations. Fig. 3 shows the nearest neighbor retrieval performance of our DEEPAVFUSION using either the learned visual, audio, or fusion tokens as queries. As can be seen, the retrieval using fusion tokens is substantially better than uni-modal representations (i.e. better aligned with the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 82, + 420, + 182 + ], + "blocks": [ + { + "bbox": [ + 343, + 70, + 505, + 80 + ], + "lines": [ + { + "bbox": [ + 343, + 70, + 505, + 80 + ], + "spans": [ + { + "bbox": [ + 343, + 70, + 505, + 80 + ], + "type": "text", + "content": "Table 7. Impact of early fusion and fusion tokens." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 82, + 420, + 182 + ], + "lines": [ + { + "bbox": [ + 319, + 82, + 420, + 182 + ], + "spans": [ + { + "bbox": [ + 319, + 82, + 420, + 182 + ], + "type": "table", + "html": "
#Linear AccSep SDRSegm mIoU
1253.086.5352.05
952.766.1549.93
652.296.1249.02
351.215.8548.04
146.525.3244.32
039.674.2338.32
", + "image_path": "4e9326675f2c00064cc207bf9d1f866e64b5e8c3789fddd5e3a4dea826b60e68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 184, + 402, + 194 + ], + "lines": [ + { + "bbox": [ + 336, + 184, + 402, + 194 + ], + "spans": [ + { + "bbox": [ + 336, + 184, + 402, + 194 + ], + "type": "text", + "content": "(a) # Fusion layers" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 433, + 92, + 533, + 171 + ], + "blocks": [ + { + "bbox": [ + 433, + 92, + 533, + 171 + ], + "lines": [ + { + "bbox": [ + 433, + 92, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 433, + 92, + 533, + 171 + ], + "type": "table", + "html": "
#Linear AccSep SDRSegm mIoU
151.865.9450.67
851.975.8351.14
1653.086.5352.05
3252.186.4350.60
", + "image_path": "c22cbb41f839dd342419d945e9ae2269b66b2d803db5239024034421d18dd1c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 376, + 197, + 476, + 264 + ], + "blocks": [ + { + "bbox": [ + 449, + 172, + 517, + 182 + ], + "lines": [ + { + "bbox": [ + 449, + 172, + 517, + 182 + ], + "spans": [ + { + "bbox": [ + 449, + 172, + 517, + 182 + ], + "type": "text", + "content": "(b) # Fusion tokens" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 197, + 476, + 264 + ], + "lines": [ + { + "bbox": [ + 376, + 197, + 476, + 264 + ], + "spans": [ + { + "bbox": [ + 376, + 197, + 476, + 264 + ], + "type": "table", + "html": "
#Linear AccSep SDRSegm mIoU
451.785.6849.59
853.086.5352.05
1652.256.1650.89
", + "image_path": "481c11b56f0626fef493f4d525ff041c288293d94d78cb482b68b4082e0d8705.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 396, + 265, + 456, + 275 + ], + "lines": [ + { + "bbox": [ + 396, + 265, + 456, + 275 + ], + "spans": [ + { + "bbox": [ + 396, + 265, + 456, + 275 + ], + "type": "text", + "content": "(c) # Aggr tokens" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 285, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 332 + ], + "type": "text", + "content": "semantics of the dataset). This suggests that fusion tokens aggregate information that is indicative of high-level semantics, while uni-modal representations encode the low-level details required for masked reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 334, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 430 + ], + "type": "text", + "content": "Impact of miscellaneous design choices. Lastly, we examine the effects of the number of fusion layers, fusion tokens and aggregation tokens in Table 7. These results reveal that early fusion is crucial for optimal performance, with model performance improving as fusion depth increases. Furthermore, DEEPAVFUSION can benefit from a relatively large number of tokens, with the performance saturating after 16 fusion tokens and 8 aggregation tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 547, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 680 + ], + "type": "text", + "content": "In this work, we present DEEPAVFUSION, a simple yet effective early fusion approach with dense interactions that achieves efficient audio-visual pre-training for audio-visual masked auto-encoders. We introduce learnable fusion tokens to aggregate modality-specific information with early fusion from each transformer block of audio and visual encoders, where two variants of parallel and sequential flows are proposed to achieve early fusion between fusion tokens and audio-visual patch tokens. Furthermore, we leverage multi-modal blocks with dense interactions to fuse fusion tokens and patches across audio-visual transformer blocks and achieve efficient downstream fine-tuning with factorized attention blocks. Empirical experiments on Flickr-SoundNet, VGG-Instruments, VGG-Music, VGGSound-All, and AVS-Bench datasets demonstrate the state-of-the-art performance of our DEEPAVFUSION in linear probing, fine-tuning classification, visual sound localization, sound separation, and audio-visual segmentation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27193" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 609-617, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "text", + "content": "[2] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 213 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "text", + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, S. V. A. Garí, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Proceedings of European Conference on Computer Vision (ECCV), pages 17–36, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 323 + ], + "type": "text", + "content": "[5] Changan Chen, Sagnik Majumder, Al-Halah Ziad, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In Proceedings of International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 288, + 390 + ], + "type": "text", + "content": "[6] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "text", + "content": "[7] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 500 + ], + "type": "text", + "content": "[8] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 535 + ], + "type": "text", + "content": "[9] Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia. Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 634 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[12] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for BERT pre-training of vision transformers. In Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 128 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "text", + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 281 + ], + "type": "text", + "content": "[16] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10478-10487, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 282, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 547, + 316 + ], + "type": "text", + "content": "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 324-333, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 547, + 382 + ], + "type": "text", + "content": "[18] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In Proceedings of 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780, 2017. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 547, + 437 + ], + "type": "text", + "content": "[19] Mariana-Iuliana Georgescu, Eduardo Fonseca, Radu Tudor Ionescu, Mario Lucic, Cordelia Schmid, and Anurag Arnab. Audiovisual masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16144-16154, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 547, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 492 + ], + "type": "text", + "content": "[20] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In Proceedings of The Eleventh International Conference on Learning Representations (ICLR), 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "type": "text", + "content": "[21] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 547, + 624 + ], + "type": "text", + "content": "[23] John Hershey and Michael Casey. Audio-visual sound separation via hidden markov models. Advances in Neural Information Processing Systems, 14, 2001. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "type": "text", + "content": "[24] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[25] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27194" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "type": "text", + "content": "[26] Po-Sen Huang, Scott Deann Chen, Paris Smaragdis, and Mark Hasegawa-Johnson. Singing-voice separation from monaural recordings using robust principal component analysis. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 57–60, 2012. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[27] Po-Yao Huang, Vasu Sharma, Hu Xu, Chaitanya Ryali, Haoqi Fan, Yanghao Li, Shang-Wen Li, Gargi Ghosh, Jitendra Malik, and Christoph Feichtenhofer. Mavil: Masked audio-video learners. arXiv preprint arXiv:2212.08071, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "text", + "content": "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "text", + "content": "[30] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "text", + "content": "[31] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 399 + ], + "type": "text", + "content": "[32] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2002–2006, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[33] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video parsing. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "text", + "content": "[34] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 609 + ], + "type": "text", + "content": "[36] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 954–965, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 645 + ], + "type": "text", + "content": "[37] Shentong Mo and Pedro Morgado. Benchmarking weakly-supervised audio-visual sound localization. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[38] Shentong Mo and Pedro Morgado. A unified audio-visual learning framework for localization, separation, and recognition. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[39] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "type": "text", + "content": "[40] Shentong Mo and Yapeng Tian. Semantic-aware multi-modal grouping for weakly-supervised audio-visual video parsing. In European Conference on Computer Vision (ECCV) Workshop, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 171 + ], + "type": "text", + "content": "[41] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. arXiv preprint arXiv:2303.17056, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "text", + "content": "[42] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 239 + ], + "type": "text", + "content": "[43] Shentong Mo, Jing Shi, and Yapeng Tian. DiffAVA: Personalized text-to-audio generation with visual alignment. arXiv preprint arXiv:2305.12903, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 240, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 547, + 282 + ], + "type": "text", + "content": "[44] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 284, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 547, + 327 + ], + "type": "text", + "content": "[45] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4733–4744, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 328, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 547, + 371 + ], + "type": "text", + "content": "[46] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12934–12945, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 373, + 547, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 547, + 425 + ], + "type": "text", + "content": "[47] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12475-12486, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 427, + 547, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 547, + 471 + ], + "type": "text", + "content": "[48] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 472, + 547, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 547, + 515 + ], + "type": "text", + "content": "[49] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 516, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 516, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 516, + 547, + 568 + ], + "type": "text", + "content": "[50] Andrew Owens, Jiajun Wu, Josh H. McDermott, William T. Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-816, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 547, + 624 + ], + "type": "text", + "content": "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 669 + ], + "type": "text", + "content": "[52] Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[53] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27195" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[54] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proceedings of European Conference on Computer Vision (ECCV), page 436–454, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "type": "text", + "content": "[55] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2745-2754, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 172, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 288, + 227 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), page 5998-6008, 2017. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 228, + 288, + 271 + ], + "type": "text", + "content": "[57] Tuomas Virtanen. Monaural sound source separation by nonnegative matrix factorization with temporal continuity and sparseness criteria. IEEE Transactions on Audio, Speech, and Language Processing, 15(3):1066-1074, 2007. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 288, + 326 + ], + "type": "text", + "content": "[58] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan L. Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 288, + 381 + ], + "type": "text", + "content": "[59] Alexander Wettig, Tianyu Gao, Zexuan Zhong, and Danqi Chen. Should you mask " + }, + { + "bbox": [ + 49, + 327, + 288, + 381 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 49, + 327, + 288, + 381 + ], + "type": "text", + "content": " in masked language modeling? In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2985-3000, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 383, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 288, + 414 + ], + "type": "text", + "content": "[60] Jiantao Wu and Shentong Mo. Object-wise masked autoencoders for fast pre-training. arXiv preprint arXiv:2205.14338, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 415, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 288, + 459 + ], + "type": "text", + "content": "[61] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1326–1335, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 288, + 502 + ], + "type": "text", + "content": "[62] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 6291–6299, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 504, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 288, + 557 + ], + "type": "text", + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 288, + 602 + ], + "type": "text", + "content": "[64] Xudong Xu, Bo Dai, and Dahua Lin. Recursive visual sound separation using minus-plus net. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 603, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 288, + 647 + ], + "type": "text", + "content": "[65] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "text", + "content": "[66] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1735-1744, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 288, + 714 + ], + "type": "text", + "content": "[67] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 160 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Wang, and Yiran Zhong. Audio-visual segmentation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 160 + ], + "type": "text", + "content": "[68] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27196" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_content_list.json b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9c64980636159b3fc4cf98767eec5b9383a16406 --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_content_list.json @@ -0,0 +1,1391 @@ +[ + { + "type": "text", + "text": "Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation", + "text_level": 1, + "bbox": [ + 132, + 130, + 839, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fuli Wan*, Han Zhao*, Xu Yang†, and Cheng Deng† \nSchool of Electronic Engineering, Xidian University, Xi'an710071, China \n{wfl128.xidian, hzhao1698, xuyang.xd, chdeng.xd}@gmail.com", + "bbox": [ + 191, + 202, + 777, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-Set Source-Free Domain Adaptation aims to transfer knowledge in realistic scenarios where the target domain has additional unknown classes compared to the limited-access source domain. Due to the absence of information on unknown classes, existing methods mainly transfer knowledge of known classes while roughly grouping unknown classes as one, attenuating the knowledge transfer and generalization. In contrast, this paper advocates that exploring unknown classes can better identify known ones, and proposes a domain adaptation model to transfer knowledge on known and unknown classes jointly. Specifically, given a source pre-trained model, we first introduce an unknown diffuser that can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider space distribution, we enhance the reliability of known class knowledge in the source pre-trained model through contrastive constraint. Finally, various supervision information, including reliable known class knowledge and clustered pseudo-labels, optimize the model for impressive knowledge transfer and generalization. Extensive experiments show that our network can achieve superior exploration and knowledge generalization on unknown classes, while with excellent known class transfer. The code is available at https://github.com/xdwfl/UPUK.", + "bbox": [ + 75, + 323, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By transferring knowledge from label-rich source domains to unlabeled target domains, unsupervised domain adaptation (UDA) has exhibited huge potential to address the challenges of costly data labeling and domain distribution shifts in real-world scenarios. Existing methods mainly learn the domain invariant representation by distribution matching", + "bbox": [ + 75, + 772, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7258fd0de622d507a6a6acaa4714813d3feb5bf1a3e67c9e1a3389df1e1bb8ef.jpg", + "image_caption": [ + "Figure 1. Comparison of existing OS-SFDA methods and our proposed method. Existing methods ignore the unknown classes by grouping unknown classes as one, while our method can further explore the unknown class information." + ], + "image_footnote": [], + "bbox": [ + 506, + 292, + 893, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[19, 20, 34, 40, 46] and adversarial optimization [7, 21, 35], to mitigate domain distribution shift. They can achieve impressive transfer effects by following the close-set and real-time data-accessible assumption, which shares the same class space in source and target domains, and accesses the source data at any time. However, in the real world, the open-set phenomenon, where the target domain has additional unknown class space to the source domain, always appears along with limited access to source data [5, 16]. In such most realistic scenarios, these methods fail to transfer knowledge without accessing source data and shared class distribution, making the knowledge transfer in such scenarios challenging.", + "bbox": [ + 496, + 582, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, to tackle such a challenging and realistic task, the open-set source-free domain adaptation (OS-SFDA) has emerged to perform knowledge transfer in the open-set scenario by accessing only the source pre-trained model, not the source data. Considering the lack of information on target unknown classes, existing approaches [6, 22, 31, 37] focus on transferring knowledge of known classes through pseudo-supervision, while grouping unknown classes as", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†The corresponding authors.", + "bbox": [ + 96, + 887, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24015", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "one, as seen Fig. 1. However, such a grouping way restricts the thorough exploration of target class information, thereby diminishing models' transfer and generalization abilities on known and unknown classes, respectively.", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present a domain adaptation network for OS-SFDA, to explore target class space and transfer knowledge on known and unknown classes jointly. Specifically, given the source pre-trained model, an unknown diffuser is introduced to determine to-be-split/merged classes through similarity measures, generating a wider class space distribution, including known and unknown classes. Based on a wider distribution in the feature space, the reliability of the known class knowledge from the source pre-trained model is enhanced through contrastive constraint. Finally, the reliable known class knowledge, along with the clustered pseudo-labels spanning the extensive class space, can collaboratively serve as supervision during model optimization. This collaborative approach results in notable knowledge transfer and enhanced generalization capabilities. Our contributions are summarized as follows:", + "bbox": [ + 75, + 152, + 467, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a domain adaptation network that is the first to excavate the unknown class space for boosting knowledge transfer and generalization to the target domain in OS-SFDA.", + "- We design an unknown diffuser to explore wider and more precise target class space in the target domain, further benefiting known knowledge transfer and unknown generalization.", + "- To realize effective known knowledge transfer and unknown generalization, we utilize the reliable known knowledge and clustered pseudo-labels in the broader class space, to serve as supervision during optimization.", + "- Experiments can demonstrate that our method can explore target class space precisely, and yield state-of-the-art results among multiple OS-SFDA benchmarks." + ], + "bbox": [ + 76, + 397, + 467, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 638, + 217, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Source-Free Domain Adaptation", + "text_level": 1, + "bbox": [ + 76, + 665, + 361, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, with accessing only the source pre-trained model for transferring knowledge to the target domain, source-free domain adaptation (SFDA) methods [4, 16, 27, 41, 43-45, 47] have been widely used for overcoming source data privacy to achieve knowledge transfer in real-world scenarios. The existing methods are mainly divided into two categories, including methods based on pseudo-labeling [14, 16] and based on neighborhood clustering [43, 44]. For the former approaches, the classical method - SHOT [16] freezes the source classifier and optimizes the feature extractor, through mutual information maximization and source-target feature alignment based on self-supervised pseudo-labeling. To fully utilize the source and target domain knowledge, CoWA-JWDS [14] intro", + "bbox": [ + 75, + 688, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "duces a new sample confidence score – JMDS better to exert the importance of pseudo-label confidence in distinguishing samples. For the latter approaches, G-SFDA [43] proposes a method of clustering target features and semantically similar neighborhoods to make the model better adapt to the target domain. AaD [44] has introduced prediction consistency of local neighborhood features and discrete prediction of different potential features, which achieve efficient feature clustering and distribution to better transfer knowledge. In addition, there are also some approaches to tackling SFDA from other aspects. U-SFAN [27] uses predicted quantified uncertainty to guide target adaptation. CRS [47] embeds class relationship similarity to transfer domain-invariant class relationships.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Open-Set Domain Adaptation", + "text_level": 1, + "bbox": [ + 500, + 313, + 764, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, open-set domain adaptation (OSDA) tackles the realistic open-set challenge, where the target domain has additional unknown class space to the source domain. Currently, existing methods [2, 10, 15, 18, 29, 30, 33, 38] transfer known class knowledge while grouping unknown classes as one. For example, OSBP [30] uses pseudo labels to guide classifier learning and to construct constraints for the optimization of the model. Later, to better utilize pseudo-labels, STA [18] proposes an end-to-end method to gradually separate known and unknown classes while aligning feature distributions. OVANET [29] introduces the one-vs-all classifier to learn the inter- and intra-class distance, and minimizes class entropy to distinguish known and unknown classes. OSLPP [38] learns a common subspace from the source and target domains while gradually selecting and rejecting pseudo-labeled target data to promote the model's transfer on known classes. By introducing adversarial learning, UADAL [10] aligns the source- and the target-known distribution while segregating the target-unknown distribution, achieving better knowledge transfer.", + "bbox": [ + 496, + 335, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Open-Set Source-Free Domain Adaptation", + "text_level": 1, + "bbox": [ + 500, + 650, + 861, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When the open-set phenomenon appears, along with limited access to source data, the methods without accessing source data and shared class distribution cannot transfer favorable knowledge, causing them to fail in real-world scenarios. OS-SFDA [6, 11, 13, 17, 22, 26, 37, 42] has emerged to perform knowledge transfer in the open-set scenario with accessing only the source pre-trained model, not the source data. Specifically, the existing methods generally transfer known knowledge and group unknown classes as one. Among these methods, without source data, FS [13] designs the self-adaptive model to transfer the source task-specific knowledge to the target domain effectively, and the target unknown classes are grouped into one class under the open-set setting due to the lack of unknown class information. To tackle OS-SFDA, OSHT [6] adopts pseudo", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24016", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "labeling for adaptation and the entropy-based metric to reject unknown classes. Distill-SODA [37] introduces novel style-based adversarial data augmentation and the closed-set affinity score for better knowledge distillation from a self-supervised ViT during adaptation. SF-PGL [22] introduces a balanced pseudo-labeling strategy to progressive graph learning for adaptation and distinguishes known and unknown classes by applying the confidence threshold. In addition, the SFDA methods [14, 16, 44, 47] separate known and unknown classes and apply their strategies mentioned in Sec. 2.1 for known knowledge transfer under the open-set setting to tackle the OS-SFDA.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, under the OS-SFDA setting, where the target domain has a additional abundant class distribution than the source domain, mentioned OS-SFDA methods group unknown classes as one, which restricts the thorough exploration of target class information, hampering knowledge transfer and generalization. Our work advocates exploring unknown classes can better identify known ones. It should be noted that recently some methods have attempted to explore unknown classes in OSDA [11, 23, 49], which requires source data access to realize domain adaptation. On the contrary, our focused OS-SFDA needs no source data for exploring unknown class space that is more realistic and challenging, essentially differing from them. Concretely, their source data access results in the failure to deal with the most realistic OS-SFDA. Moreover, they utilize 'unknown class number prior' to predefine [11, 23] or enumerate [49] unknown class numbers to explore the unknown class space, whose reliance on unknown prior further limits their scalability in real world. Instead, during unknown class exploration, our method requires no source data and unknown class number prior, truly achieving spontaneous and applicable unknown class exploration in OS-SFDA. We present a domain adaptation network for OS-SFDA, which can fully explore the target class space to transfer knowledge on known classes and generalize on unknown classes.", + "bbox": [ + 75, + 272, + 472, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 662, + 169, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will introduce the details of our proposed method, and the framework is depicted in Fig. 2.", + "bbox": [ + 75, + 688, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setting. OS-SFDA aims to solve the knowledge-transferring problem in more realistic scenarios, where the source data have restricted access, and the target domain exhibits wider class distribution than the source domain. In such a task with restricted-accessible source data, we are given an unlabeled target domain $\\mathcal{D}^t = \\{(x_i^t)\\}_{i=1}^n$ with $n$ target samples, and a pre-trained source model. More importantly, the class space $\\mathcal{Y}^t$ of the target domain is broader than the one $\\mathcal{Y}^s$ of the source domain: $\\mathcal{Y}^t = \\mathcal{Y}^s \\cup \\mathcal{Y}^{ukn}$ . In other words, in addition to the source domain's class space $\\mathcal{Y}^s$ , $\\mathcal{Y}^t$ also contains an additional unknown class space $\\mathcal{Y}^{ukn}$ . Here we call $\\mathcal{Y}^s$ and $\\mathcal{Y}^{ukn}$ as the target domain's", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "known class and unknown class spaces, respectively, whose numbers of categories are denoted as $K_{n}$ and $K_{un}$ . The source domain pre-trained model is obtained by pre-training on the source domain in $\\mathcal{Y}^s$ , mainly consisting of feature extractor $f(\\cdot)$ and classifier $g(\\cdot)$ . Therefore, it is expected to enable the source pre-trained model to overcome the class distribution discrepancy between domains, while achieving further knowledge transfer on the known classes and exploration on the unknown classes.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Exploration of the Target Class Space", + "text_level": 1, + "bbox": [ + 498, + 234, + 826, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the effectiveness of no-parameter clustering methods [1, 25, 32, 48] in exploring the class space, we introduce an unknown diffuser based on the Multilayer Perceptron (MLP), which includes the main network $C$ and the sub-network $C_k^{sub}$ in the current class space, whose goal is to explore the target class space by inferring the target class number $K$ . Reasonably, we first initialize the target class space with the source known class space; that is, we assume that the initial value of the target class number $K$ is the number of known classes $K_n$ . The unknown diffuser progressively explores a wider class space in the target domain through continuous inferring and optimization of $K$ . Note that the value of $K$ changes dynamically with the inference process during training. The unknown diffuser completes the exploration of the target class space by performing two training processes: optimization in the latest class space and hierarchical class exploration.", + "bbox": [ + 496, + 257, + 893, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization in Latest Class Space. Since the target domain has no labels during exploration, we need to use the clusters obtained by clustering the target data to guide us in exploring the target class space. Simultaneously, we transform the exploration of the target class space into the exploration of the target cluster number. In this part, we learn the discriminative cluster assignments through the constraints of the feature-cluster assignment, which is beneficial for exploring the target cluster number. Firstly, the feature is the output of feature extractor $f(\\cdot)$ in the source pre-trained model, denoted as $z_{i} = f(\\pmb{x}_{i}^{t})\\in \\mathbb{R}^{d}$ , where $d$ is the dimension of the feature space, $i$ represents the $i$ -th sample. The features of $n$ samples are denoted as $Z$ . Then, the main network $C$ in the unknown diffuser maps the feature into the cluster assignment $\\widetilde{\\boldsymbol{Z}} = c(\\boldsymbol {Z})$ , where $c(\\cdot)$ is the mapping operation of $C$ , and the cluster assignment of feature $z_{i}$ is denoted as $\\widetilde{z}_i\\in \\mathbb{R}^K$ . The discriminative cluster assignments under the current class space $(K)$ are obtained by optimizing the cluster assignment under the guidance of the discriminative feature $Z$ . Specifically, to obtain the pseudo cluster-assignment $Y^{Z} = \\{y_{i}^{Z}\\}_{i = 1}^{n}, y_{i}^{Z}\\in \\mathbb{R}$ , we first use $k$ -means under current target class space $(K)$ on the feature set $Z$ . Then, according to $Y^{Z}$ , a teacher cluster soft-distribution $\\widetilde{Y}\\in \\mathbb{R}^{n\\times K}$ is generated based on the feature:", + "bbox": [ + 496, + 515, + 895, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {y} _ {i, k} = \\left\\| \\boldsymbol {z} _ {i} - \\boldsymbol {\\eta} _ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 882, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24017", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ce7c259dd2bd690febebc3e065781cf69691d9c52954a6b4ea10168bdab24d9a.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed method. Given the source pre-trained model, we first introduce an unknown diffuser on its initial known class space. The unknown diffuser can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider class space, we improve the reliability of the known class knowledge inside the source model by contrastive learning. In this way, the reliable known class knowledge and clustered pseudo-labels over the wider class space, can jointly serve as supervision during model optimization, leading to powerful knowledge transfer and generalization capabilities." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 872, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widetilde{y}_{i,k}$ is the $k$ -th element in $\\widetilde{\\mathbf{y}}_i$ , and $\\widetilde{\\mathbf{y}}_i$ is the teacher cluster soft-distribution of feature $\\mathbf{z}_i$ . $\\eta_k = \\sum_{i + 1}^{n_k}z_i^k /n_k$ is the center feature of the $k$ -th center and $n_k$ is the feature number of the $k$ -th master cluster. And $\\boldsymbol {z}_i^k$ is the $i$ -th element in $\\mathbf{Z}^k\\in \\mathbb{R}^{n_k\\times d}$ , which is the feature matrix of those features with $y_{i}^{Z} = k$ . Note that the teacher cluster soft-distribution of each feature is normalized. Then, we utilize the alignment loss to enforce the consistency of teacher cluster soft-distribution $\\widetilde{\\mathbf{Y}}$ with the cluster assignment $\\widetilde{\\mathbf{Z}}$ :", + "bbox": [ + 75, + 462, + 472, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l g} = \\sum_ {i = 1} ^ {n} L _ {a l g} \\left(x _ {i} ^ {t}\\right) = \\sum_ {i = 1} ^ {n} \\widetilde {\\boldsymbol {z}} _ {i} \\cdot \\widetilde {\\boldsymbol {y}} _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 609, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cluster assignment under the current target class space $(K)$ can be promoted through optimization on $c(\\cdot)$ in the training process, so as to promote the exploration of the target class space.", + "bbox": [ + 75, + 659, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Class Exploration. To fully explore the target class space by inferring the precise target class number $K$ , inspired by [3], the Metropolis-Hastings framework (M-H) is utilized as the underlying framework in the class diffusion process, which computes the Hasting ratio for changing $K$ through splitting and fusion. Besides that, inspired by the hierarchical clustering [39], we design a hierarchical class exploration method to explore the target class space based on the M-H framework, that performs splitting and merging on more dispersive and coupled clusters, respectively. Firstly, we obtain the discriminative cluster assignment after splitting/merging in all possible split-", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ting/merging cases. We then select the potential clusters to be split or merged based on the assignment distance between clusters. In the end, these selected clusters are put into the candidate set to be split or merged, while we use the M-H framework [8] on those clusters in the candidate set to determine the splitting and merging decisions. We assume that the current cluster distributions and those under all possible splitting/merging cases are called the distributions before and after splitting/merging.", + "bbox": [ + 496, + 462, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first obtain the discriminative cluster distribution after splitting/merging in all possible splitting/merging cases, in which the cluster before splitting/merging is already learned above. Then, the cluster distribution after merging can be acquired directly by merging the assignments of two clusters. However, obtaining the sub-cluster distribution after splitting is apparently unreasonable, by randomly dividing the target data in the master cluster. For subsequent convenience, we define the clusters before/after splitting as the master cluster and sub-cluster, respectively. To gain the discriminative sub-cluster distribution, we introduce the sub-network $C_k^{sub}$ into the unknown diffuser, and $C_k^{sub}$ contains $K$ sub-clustering layers with the output dimension equaling to 2. One of these sub-network $C_k^{sub}$ is applied on each master cluster to generate the sub-cluster assignment $\\widetilde{Z}^{sub,k} = c_k^{sub}(Z^k)$ , in which $c_k^{sub}(\\cdot)$ represents the operation of the $k$ -th sub-network applying on the $k$ -th master cluster. Then, to obtain the pseudo-sub-cluster assignment, the $k$ -means first is performed on the", + "bbox": [ + 496, + 608, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24018", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature matrix $\\mathbf{Z}^k$ of the $k$ -th master cluster, similar to cluster optimization with Eq. (1), Subsequently, a teacher subcluster assignment $\\widetilde{\\mathbf{Y}}^k \\in \\mathbb{R}^{n_k \\times 2}$ is also generated for the $k$ -th master cluster:", + "bbox": [ + 76, + 89, + 468, + 155 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {i, j} ^ {k} = \\left\\| \\boldsymbol {z} _ {i} ^ {k} - \\boldsymbol {\\eta} _ {j} ^ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 159, + 468, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{y}_{i,j}^{k}$ is the $i$ -th row and $j$ -th column element in $\\widetilde{\\pmb{Y}}^k$ . $\\pmb{\\eta}_j^k$ is the center feature of the $j$ -th sub-cluster in $k$ -th master cluster. And we utilize the alignment loss to optimize them:", + "bbox": [ + 76, + 191, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l g} ^ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\widetilde {\\boldsymbol {z}} _ {i} ^ {\\text {s u b}, k} \\cdot \\widetilde {\\boldsymbol {y}} _ {i} ^ {k}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 250, + 468, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{z}_i^{sub,k}$ is the $i$ -th element in $\\widetilde{\\mathbf{Z}}^{sub,k}$ . The training process on the sub-network $C_k^{sub}$ is synchronous with the one on the main network $C$ .", + "bbox": [ + 76, + 299, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, hierarchical clustering [39] gradually divides data by merging pairs of clusters with smaller distances or splitting the two most distant sub-clusters in one cluster. According to the assignment distance between clusters, the candidate sets to be split/merged are generated, inspired by the hierarchical clustering. In our method, to explore more different classes, those cluster pairs (sub-cluster pairs) with far assignment distances are considered as those likely to be split. While those with close assignment distances are considered as those likely to be merged due to high similarity between two main clusters. The candidate sets to be split/merged based on this can be written as:", + "bbox": [ + 75, + 348, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC S _ {m e r g} ^ {E D} = \\left\\{\\left(k _ {1}, k _ {2}\\right) \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} E D \\left(\\boldsymbol {\\eta} _ {k _ {1}}, \\boldsymbol {\\eta} _ {k _ {2}}\\right) \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 536, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC S _ {s p l} ^ {E D} = \\left\\{k \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} - E D \\left(\\boldsymbol {\\eta} _ {1} ^ {k}, \\boldsymbol {\\eta} _ {2} ^ {k}\\right) \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 561, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $ED(\\cdot, \\cdot)$ computes the Euclidean distance of the cluster embedding center. $n_{if}$ are the number of cluster(s) with potential to be merged and split.", + "bbox": [ + 76, + 583, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the candidate set, we can infer more accurate $K$ and explore a more accurate target class space, by calculating the Hasting ratio from the M-H framework [8] among (sub)cluster pairs in the candidate set and comparing it with 1. More details are provided in the supplementary material. In this way, we use the inferred latest target class number $K$ to update the unknown diffuser (the output dimension of $C$ and the number of sub-clustering layers) at the next optimization step.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the optimization in the latest more accurate class space, we learn the discriminative cluster distribution under the latest updated target class number $K$ . According to these cluster distributions, we decide the splitting and merging decisions to change the target class number $K$ in the process of hierarchical class exploration. The more precise $K$ and corresponding cluster assignments are generated, through operating these two processes iteratively, which is beneficial to fully explore the target class space.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 The training procedure of our proposed model \nInput: data, source pre-trained model, main network $C$ and sub-network $C_k^{sub}$ of unknown diffuser", + "bbox": [ + 500, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: // Exploration of the target class space", + "2: for $iter_1 = 1$ to $iter_1^{max}$ do" + ], + "bbox": [ + 509, + 140, + 781, + 170 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3: Obtain clusteredpesudo-distribution $\\widetilde{Y}$ and $\\widetilde{Y}^k$ by $k$ -means", + "4: Optimize $C$ and $C_k^{sub}$ guided by $\\widetilde{Y}$ and $\\widetilde{Y}^k$ in Eqs. (2) and (4)", + "5: Determine $CS_{slp} / CS_{merg}$ by similarity measurement in Eqs. (5) and (6)", + "6: Diffuse the class space by updating $K$ on $CS_{slp} / CS_{merg}$ , based on M-H framework", + "7: Update the unknown diffuser by the updated $K$", + "8: end for", + "9: // Knowledge transfer and generalization", + "10: for iter2 = 1 to iter2 max do", + "11: Promote reliability of known knowledge by contrastive constraint in Eq. (7)", + "12: Jointly optimize $C$ and the pre-trained model in the explored class space via $\\mathcal{L}_{kt}$ and $\\mathcal{L}_{alg}$ in Eq. (8)", + "13: end for" + ], + "bbox": [ + 504, + 171, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Knowledge Transfer and Generalization", + "text_level": 1, + "bbox": [ + 500, + 455, + 843, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After exploring the target class space in Sec. 3.1, we aim to transfer knowledge on known classes more effectively and improve the model's generalization on unknown classes. In this subsection, we promote the reliability of known knowledge and further optimize over wider class distribution.", + "bbox": [ + 496, + 479, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Promotion of Known Knowledge Reliability. To better utilize known knowledge, we hope to improve the reliability of known knowledge obtained by the classification layer of the pre-trained source model during training. The input of the model is all target samples, including known and unknown samples, from which we need to select the known ones. According to the logit output of known samples having smaller entropy than the unknown ones, we select the top $50\\%$ of samples with the lowest entropy as known samples, which are placed in set $\\mathcal{M}$ . After selection, the contrastive constraint $\\mathcal{L}_{con}$ is utilized on $\\mathcal{M}$ to enhance the discrimination of the known samples. Concretely, it brings the known features of similar predictions closer and those of dissimilar predictions farther away, achieving discrimination feature assignment to strengthen the reliability of known knowledge. The contrastive constraint is written as:", + "bbox": [ + 496, + 556, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = - \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {+} + \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {-}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 809, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_i = \\sigma(g(\\pmb{z}_i)) \\in \\mathbb{R}^d$ is the output of classifier $g$ and $\\sigma$ is the softmax function.", + "bbox": [ + 498, + 854, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization over Wider Class Distribution. After", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24019", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that, we optimize the unknown diffuser over the wider class space by taking advantage of the reliable known knowledge and alignment constraint on the explored space. In such an optimization based on the explored wider class space, we first select the high-confidence known samples as the reliable known knowledge. Then, we hope to use the hard pseudo-labels of the high-confidence known samples as supervision information, thereby leveraging them for better knowledge transfer on the explored class space. Specifically, for a known sample, we take the maximum logit output by the classifier in the source pre-trained model as its confidence. And the known sample is considered as the high-confidence known sample when it is greater than a manually-set confidence threshold, whose hard pseudolabel is represented by the one-hot paradigm $\\pmb{y}_i^h (x_i^h\\in \\mathcal{H})$ . Note that $\\mathcal{H}$ is the set of the high-confidence known samples. Therefore, we can perform better knowledge transfer on the explored class space by leveraging the reliable known knowledge $\\pmb{y}_i^h$ obtained on the source pre-trained model. In addition, we also perform alignment constraint on the explored class space, similar to Eq. (2), to complete the transfer on known samples further and achieve generalization on unknown samples. Finally, we leverage the supervision of the high-confidence known sample labels and the alignment constraint, to jointly achieve superior knowledge transfer on known classes and generalization on unknown classes over the wider class space:", + "bbox": [ + 76, + 90, + 472, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {k t} + \\mathcal {L} _ {a l g} = \\sum_ {i} ^ {\\mathcal {H}} L _ {k t} (\\widetilde {\\boldsymbol {z}} _ {i}, \\boldsymbol {y} _ {i} ^ {h}) + \\sum_ {i = 1} ^ {n} L _ {a l g} (x _ {i} ^ {t}) \\\\ = \\underbrace {\\sum_ {i} ^ {\\mathcal {H}} L _ {k t} \\left(\\widetilde {\\boldsymbol {z}} _ {i} , \\boldsymbol {y} _ {i} ^ {h}\\right) + \\sum_ {j _ {1} = 1} ^ {n _ {1}} L _ {a l g} \\left(x _ {j _ {1}} ^ {t}\\right)} _ {\\text {k n o w n}} + \\underbrace {\\sum_ {j _ {2} = 1} ^ {n _ {2}} L _ {a l g} \\left(x _ {j _ {2}} ^ {t}\\right)} _ {\\text {u n k n o w n}}, \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 508, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{kt}$ represents the guidance of high-confidence known samples, which is the cross-entropy loss between their pseudo-labels and soft distributions. $\\mathcal{L}_{alg}$ is the alignment constraint to achieve superior knowledge transfer on known classes and generalization on unknown classes. $\\widetilde{z}_i$ represents the soft-distribution output by the main network $C$ on the explored class space. $n_1$ and $n_2$ represent the number of known and unknown samples, respectively.", + "bbox": [ + 76, + 626, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 784, + 267, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three benchmark datasets for OS-SFDA through experiments, which are Office-31 [28], Office-Home [36], and VisDA-C 2017 [24]. Due to limited space, the results and analysis on VisDA-C 2017 are summarized in supplementary material.", + "bbox": [ + 76, + 806, + 467, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Office-31 [28] consists of three domains (Amazon, We", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bcam, DSLR), with 31 classes and 4652 images of common objects. Following the protocols [18] on Office-31 and Office-Home, we select the first 20 classes as the shared classes of the source and target domain while the remaining 11 classes are considered unknown.", + "bbox": [ + 509, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Office-Home [36] is a more challenging dataset. It contains four domains (Art, Product, Clipart, RealWorld) with 65 classes and 15500 images of common objects. Like Office-31, we select the first 25 classes as classes shared by the source and target domain, and the remaining 40 classes are considered unknown.", + "bbox": [ + 500, + 167, + 890, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. For OS-SFDA, we compare our method with the clustering on Resnet [9], the classic method SHOT [16], and the advanced method AaD [44], which do sufficient experiments under the OS-SFDA setting. In addition, our method uses the unknown diffuser based on clustering (class number-agnostic) to explore the target class space. For a fair comparison, we thoroughly compare our method with the no-parameter clustering methods (FINCH [32], COMIC [25], DenMune [1]).", + "bbox": [ + 496, + 260, + 890, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. As shown in Tab. 1 and Tab. 2, we report domain adaptation results for $OS^{*}$ and $UNK$ , which represent the average accuracy of the known and the unknown class, respectively. Simultaneously, $HOS = \\frac{2 \\times OS^{*} \\times UNK}{OS^{*} + UNK}$ is denoted as the harmonic mean between $OS^{*}$ and $UNK$ , which is the key indicator for evaluating performance. Moreover, the inferred target class number $K$ and accuracy ACC are used to evaluate the effectiveness of exploring the target class space in Tab. 3 and Tab. 4. The setting of ACC follows the previous protocols [12].", + "bbox": [ + 496, + 397, + 890, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For the source pre-trained model, we utilize the backbone of ResNet-50 [9] for Office-31 and Office-Home. For a fair comparison, we use the same network architecture with SHOT [16] and AaD [44], which includes a backbone and two additional fully connected layers. We adopt SGD with a momentum of 0.9 and a batch size of 64 for all datasets. The learning rate is set to 1e-3 on all datasets for all layers of the backbone, and the last two added fully-connected layers are applied 1e-2. We train 15 epochs for all datasets. The unknown diffuser includes one convolutional layer with 256 hidden dimensions. For a fair comparison, our network and the compared no-parameter clustering methods use the same input features, output by the source pre-trained model.", + "bbox": [ + 496, + 549, + 890, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 771, + 694, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1 and Tab. 2, the UNK and HOS for our method are much higher than SHOT on all datasets. Our method achieves $10.4\\%$ and $2.3\\%$ relative improvements over the state-of-art AaD for average OS* and HOS on Office-31, respectively. On the relatively small-scale Office-31 with fewer classes, our method demonstrates competitiveness across various indicators, standing in con", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24020", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ccdf777ca692021fc6e7c77ae47671ec8314e75128605bb4e3ab481f21219952.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodSFA→DA→WD→AD→WW→AW→DAvg
OS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOS
Cluster✗10.064.017.07.061.013.05.069.010.08.049.014.05.064.010.010.047.017.07.559.013.5
SHOT✓87.250.055.789.038.653.878.244.256.597.472.383.073.547.657.889.675.081.785.854.664.8
AaD✓69.986.777.462.083.971.360.881.869.883.297.489.858.481.568.082.395.288.369.487.877.4
Ours(w/o s)✓78.473.475.887.990.389.172.376.174.192.880.586.270.976.873.785.879.882.781.479.580.3
Ours(w/s)✓82.371.876.790.072.080.072.079.475.579.587.383.274.282.778.280.688.384.379.880.379.7
", + "bbox": [ + 96, + 88, + 874, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/13712ff936ebcb14398017864fa06991e683284495c11375b67b6a78bb999802.jpg", + "table_caption": [ + "Table 1. Accuracy (%) on Office-31 for OS-SFDA. The three indicators(OS*, UNK, HOS) have been explained in Sec. 4.1. ‘w/o s’ and ‘w/s’ indicate with or without supervision using reliable known knowledge. Note that ‘A→D’ represents that the model per-trained on the source domain Amazon transfers to the target domain DSLR. Bold and ‘_’ represent the best and second results, respectively." + ], + "table_footnote": [], + "table_body": "
methodSFAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwAvg (first 6 tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX6.063.011.05.055.010.04.062.08.08.059.014.05.055.09.05.056.08.05.056.08.05.558.310.0
SHOT✓67.028.039.581.826.339.887.532.147.066.846.254.677.527.240.280.025.939.176.831.043.4
AaD✓50.567.457.764.066.465.172.269.570.847.180.359.364.768.266.465.071.067.860.670.564.5
Ours(w/o s)✓47.669.256.470.081.875.470.685.677.460.374.766.765.682.773.272.184.277.664.479.771.1
Ours(w/s)✓49.064.855.868.188.076.771.586.678.461.372.666.466.581.673.171.684.877.664.779.771.3
methodSFPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrRw→PrAvg (all tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX8.053.015.06.050.010.05.061.09.08.051.013.05.064.010.05.050.010.05.856.610.6
SHOT✓66.351.157.759.331.040.885.831.646.273.550.659.965.328.940.184.428.242.374.633.945.6
AaD✓46.983.160.045.072.655.669.072.370.656.077.465.048.367.656.467.769.368.558.072.163.6
Ours(w/o s)✓60.076.767.352.666.258.670.785.177.256.583.367.449.170.757.966.182.173.261.878.569.0
Ours(w/s)✓55.985.667.645.470.255.173.983.978.656.784.167.849.374.659.469.580.074.461.679.769.2
", + "bbox": [ + 94, + 265, + 877, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Accuracy (%) on Office-Home for OS-SFDA. Note that ‘Ar→Cl’ represents that the model per-trained on the source domain Art transfers to the target domain Clipart. The representations of OS*, UNK, HOS, ‘w/o s’, and ‘w/s’ are the same with Tab. 1.", + "bbox": [ + 75, + 494, + 892, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trast to methods such as AaD and SHOT, where only one indicator exhibits competitive performance. SHOT tends to excel in OS*, while AaD typically performs well on UNK. In contrast, our method displays a well-balanced performance across different measurement dimensions, achieving competitive results overall. For the more challenging Office-Home, our method displays more apparent superiority, which achieves $61.6\\%$ , $79.7\\%$ , and $69.2\\%$ average accuracy for OS*, UNK, and HOS, with $3.6\\%$ , $7.6\\%$ , and $5.6\\%$ relative improvement compared with AaD. Furthermore, our method outperforms all no-parameter clustering methods (FINCH, DenMune, and COMIC) for ACC and $K$ on all tasks in Tab. 3 and Tab. 4. The above results demonstrate the superiority and effectiveness of our method.", + "bbox": [ + 75, + 537, + 472, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Exploration for the Target Class Space. Compared with OS-SFDA methods (SHOT and AaD) that group all target unknown classes as one, our method tends to explore class space and excavate different unknown classes, which has great application value in more realistic scenarios. The experimental results show that our method achieves excellent improvement over the state-of-the-art method AaD for the known accuracy $\\mathrm{OS^{*}}$ on the more challenging Office Home. In addition, the unknown accuracy UNK for our method is much higher than SHOT", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on Office-31 and Aad on Office-Home. This phenomenon shows that our method can improve knowledge transfer on known classes when exploring class space and simultaneously improve generalization on unknown classes. To more fully demonstrate the effectiveness and advancement of our exploration, we compare our method with classical no-parameter clustering methods for exploring the class space. Concretely, ours provides the more precise inferred target class number $K$ and higher accuracy ACC on all tasks, where the precision of $K$ represents the efficacy in exploring the target class space, as discussed in Sec. 3.1. The above analysis conclusively demonstrates that our method of exploring class space is exceptionally effective, improving knowledge transfer and generalization.", + "bbox": [ + 496, + 537, + 893, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Reliable Known Knowledge as Supervision. As introduced in Sec. 3.2, we hope to utilize the reliable known knowledge to optimize the unknown diffuser for better knowledge transfer and generalization. Concretely, our method that uses the reliable known knowledge(ours (w/ s)) outperforms all compared methods(including ours(w/o s)) on more challenging Office-Home for average HOS, ACC, and $K$ in Tabs. 1 to 4. Therefore, the more effectiveness brought by reliable known knowledge ('ours (w/ s)') confirms it guides the model for better known and unknown", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24021", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ab423e646ede09859f9823eb36263394f2465b816b76b89757dc15bfdc88c1d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 103, + 267, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af9dd5e01ac10a377c37e314bee6102371cc81e27656a37f85322e057d5c943a.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) t-SNE of SHOT on $\\mathbb{R}\\to \\mathbb{P}$ . (b) t-SNE of AaD on $\\mathbb{R}\\to \\mathbb{P}$ . (c) t-SNE of our method on $\\mathbb{R}\\to \\mathbb{P}$ . (d) Parameter analysis on $\\mathrm{A}\\rightarrow \\mathrm{W}$ ." + ], + "image_footnote": [], + "bbox": [ + 295, + 103, + 464, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ff444d523cf7f5fe0392fbc3fea6ad95c9e56df8187c93cff1f15e2d64751df.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 496, + 103, + 661, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be0297374d1be915c72b94367dec59dc52cfe1f74e5f50696723cba779430542.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 687, + 102, + 864, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f333b158594b31c74b2e6ac79c75629b30e86b40d9514760457e764cdb4394a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsA→DA→WD→AD→WW→AW→D
FINCH31.0 (8)8.0 (2)23.0 (7)12.0 (3)28.0 (9)23.0 (5)
DenMune20.0(5)52.0(15)58.0(29)46.0(14)65.0(35)22.0(5)
COMIC56.8 (22)47.3 (24)16.1 (40)46.4 (20)12.1 (32)56.6 (22)
Ours(w/o s)69.0(24)80.6(28)60.4(37)82.5(30)59.6(38)68.9(27)
Ours(w/ s)71.2(25)73.3(28)63.1(37)76.3(32)64.2(38)74.5(28)
", + "bbox": [ + 78, + 292, + 467, + 391 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c54e13aa563c0c8ee426c099dd2b6aa1ef23edd436cc2a9fe6282e582d790696.jpg", + "table_caption": [ + "Table 3. Accuracy ACC(%) and the inferred target class number $K$ on Office-31. Note that the values before $\\left( \\cdot \\right)$ ’ and inside $\\left( \\cdot \\right)$ ’ represent the performance ACC and $K$ ,respectively." + ], + "table_footnote": [], + "table_body": "
MethodsAr→ClAr→PrAr→RwCl→ArCl→PrCl→Rw
FINCH5.0(4)7.0(3)6.0(3)8.0(3)7.0(4)7.0(4)
DenMune15.0(12)58.0(45)54.0(39)16.0(7)55.0(40)40.0(32)
COMIC6.9(65)50.3(92)9.43(34)4.0(1)10.4(51)2.3(2)
Ours(w/o s)35.5(69)60.5(77)59.1(77)47.3(61)58.9(74)55.6(60)
Ours(w/s)35.9(69)61.4(80)61.3(77)46.9(62)58.9(78)61.0(60)
MethodsPr→ArPr→ClPr→RwRw→ArRw→PrRw→Cl
FINCH12.0(4)5.0(4)4.0(2)11.0(4)12.0(7)7.0(4)
DenMune16.0(7)7.0(6)52.0(37)26.0(13)9.0(8)62.0(43)
COMIC4.2(2)3.7(24)5.7(14)4.1(1)5.2(43)13.3(65)
Ours(w/o s)48.5(51)37.2(56)63.6(75)38.1(28)21.6(50)50.4(75)
Ours(w/s)47.8(51)35.5(56)64.0(68)44.2(35)21.9(50)51.8(75)
", + "bbox": [ + 88, + 458, + 460, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. ACC(%) and $K$ on Office-Home.", + "bbox": [ + 145, + 661, + 398, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "class discrimination, achieving better knowledge transfer on known classes and generalization on unknown classes.", + "bbox": [ + 75, + 688, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter Analysis. We study the impact of supervision information on unknown diffuser optimization during training by applying a variable parameter to the supervision loss. As shown in Fig. 3(d), the chart indicates that when there is no supervision information, the optimization cannot be completed well to obtain a favorable transfer effect on known or unknown classes. As the parameters increase within a certain range, UNK and HOS also improve; but parameters that are too large will lead to poor optimization results. In addition, it can be seen that OS* is not sensitive to this parameter, and its change trend is relatively stable. Therefore, we should select appropriate high", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "confidence known pseudo-labels as supervision information to prevent model overfitting for better guiding the optimization of unknown diffuser.", + "bbox": [ + 496, + 295, + 890, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Unknown Diffuser. Tabs. 1 to 4 show our superior performance in exploring unknown class space and domain adaptation in OS-SFDA. The superiority in unknown class exploration indicates that our unknown diffuser can excavate potential unknown classes under the guidance of current discriminative class distributions based on the Metropolis-Hastings (M-H) framework. The impressive domain adaptation effects based on the explored wider space also demonstrate the necessity of using the unknown diffuser for unknown class exploration.", + "bbox": [ + 496, + 342, + 892, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of features. As shown in Fig. 3(a)-(c), t-SNE visualizations indicate that compared with existing methods, our method explores a more accurate and extensive class space, exhibiting remarkable class discrimination.", + "bbox": [ + 496, + 494, + 890, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 570, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a domain adaptation network in OS-SFDA for exploring unknown classes, thereby jointly benefiting known knowledge transfer and unknown generalize. We introduce an unknown diffuser into the source pretrained model to excavate a wider target class space. And, the reliable known class knowledge and clustered pseudolabels over the wider class space are captured to serve as supervision during model optimization, realizing impressive knowledge transfer and generalization. Extensive experiments verify the superiority of our method.", + "bbox": [ + 496, + 597, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 763, + 687, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key Research and Development Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62132016, 62171343, and 62201436), and Fundamental Research Funds for the Central Universities (ZDRC2102).", + "bbox": [ + 496, + 787, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24022", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Abbas, Adel El-Zoghabi, and Amin Shoukry. Denmune: Density peak based clustering using mutual nearest neighbors. Pattern Recognit., 109:107589, 2021. 3, 6", + "[2] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 422-438. Springer, 2020. 2", + "[3] Jason Chang and John W Fisher III. Parallel sampling of dp mixture models using sub-cluster splits. Proc. Adv. Neural Inf. Process. Syst., 26, 2013. 4", + "[4] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7212-7222, 2022. 2", + "[5] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 1", + "[6] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE Trans. Image Process., 30:6473-6484, 2021. 1, 2", + "[7] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. J. Mach. Learn. Res., 17(1):2096-2030, 2016. 1", + "[8] W Keith Hastings. Monte carlo sampling methods using markov chains and their applications. 1970. 4, 5", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6", + "[10] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35:16755-16767, 2022. 2", + "[11] Taotao Jing, Hongfu Liu, and Zhengming Ding. Towards novel target discovery through open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9322-9331, 2021. 2, 3", + "[12] Harold W Kuhn. The hungarian method for the assignment problem. Nav. Res. Logist. Q., 2(1-2):83-97, 1955. 6", + "[13] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12376-12385, 2020. 2", + "[14] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 12365-12377. PMLR, 2022. 2, 3", + "[15] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 24110-24119, 2023. 2", + "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 6028-6039. PMLR, 2020. 1, 2, 3, 6", + "[17] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 2", + "[18] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2927-2936, 2019. 2, 6", + "[19] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Proc. Adv. Neural Inf. Process. Syst., 29, 2016. 1", + "[20] Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(12):3071-3085, 2018. 1", + "[21] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 31, 2018. 1", + "[22] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 2, 3", + "[23] Yingwei Pan, Ting Yao, Yehao Li, Chong-Wah Ngo, and Tao Mei. Exploring category-agnostic clusters for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 13867–13875, 2020. 3", + "[24] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017. 6", + "[25] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In Proc. Int. Conf. Mach. Learn., pages 5092-5101. PMLR, 2019. 3, 6", + "[26] Masoud Pourreza. Open-set source-free domain adaptation in fundus images analysis. 2023. 2", + "[27] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 537-555. Springer, 2022. 2", + "[28] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In Proc. Eur. Conf. Comput. Vis., pages 213-226. Springer, 2010. 6", + "[29] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9000-9009, 2021. 2", + "[30] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proc. Eur. Conf. Comput. Vis., pages 153-168, 2018. 2", + "[31] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. Proc. Adv. Neural Inf. Process. Syst., 33:16282-16292, 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24023", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Saquib Sarfraz, Vivek Sharma, and Rainer Stiefelhagen. Efficient parameter-free clustering using first neighbor relations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8934-8943, 2019. 3, 6", + "[33] Tasfia Shermin, Guojun Lu, Shyh Wei Teng, Manzur Murshed, and Ferdous Sohel. Adversarial network with multiple classifiers for open set domain adaptation. IEEE Trans. Multimed., 23:2732-2744, 2020. 2", + "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proc. AAAI Conf. Artif. Intell., 2016. 1", + "[35] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darryll. Adversarial discriminative domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7167-7176, 2017. 1", + "[36] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 5018-5027, 2017. 6", + "[37] Guillaume Vray, Devavrat Tomar, Behzad Bozorgtabar, and Jean-Philippe Thiran. Source-free open-set domain adaptation for histopathological images via distilling self-supervised vision transformer. arXiv preprint arXiv:2307.04596, 2023. 1, 2, 3", + "[38] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2", + "[39] Joe H Ward Jr. Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc., 58(301):236-244, 1963. 4, 5", + "[40] Kun Wei, Xu Yang, Zhe Xu, and Cheng Deng. Class incremental unsupervised domain adaptation via pseudolabel distillation. IEEE Trans. Image Process., 2024. 1", + "[41] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9010-9019, 2021. 2", + "[42] Jiexi Yan, Zhihui Yin, Erkun Yang, Yanhua Yang, and Heng Huang. Learning with diversity: Self-expanded equalization for better generalized deep metric learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 19365-19374, 2023. 2", + "[43] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8978-8987, 2021. 2", + "[44] Shiqi Yang, Shangling Jui, Joost van de Weijer, et al. Attracting and dispersing: A simple approach for source-free domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35: 5802-5815, 2022. 2, 3, 6", + "[45] Shiqi Yang, Yaxing Wang, Luis Herranz, Shangling Jui, and Joost van de Weijer. Casting a bait for offline and online source-free domain adaptation. Comput. Vis. Image Underst., page 103747, 2023. 2", + "[46] Xu Yang, Cheng Deng, Tongliang Liu, and Dacheng Tao. Heterogeneous graph attention network for unsupervised multiple-target domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1992-2003, 2020. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Yixin Zhang, Zilei Wang, and Weinan He. Class relationship embedded learning for source-free unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7619-7629, 2023. 2, 3", + "[48] Han Zhao, Xu Yang, and Cheng Deng. Parameter-agnostic deep graph clustering. ACM Trans. Knowl. Discov. Data, 18 (3):1–20, 2024. 3", + "[49] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Open set domain adaptation by novel class discovery. arXiv preprint arXiv:2203.03329, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 233 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24024", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_model.json b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7985b04142d2df4e7ab882d90765dcfc5ca0577a --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.84, + 0.177 + ], + "angle": 0, + "content": "Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.203, + 0.778, + 0.258 + ], + "angle": 0, + "content": "Fuli Wan*, Han Zhao*, Xu Yang†, and Cheng Deng† \nSchool of Electronic Engineering, Xidian University, Xi'an710071, China \n{wfl128.xidian, hzhao1698, xuyang.xd, chdeng.xd}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Open-Set Source-Free Domain Adaptation aims to transfer knowledge in realistic scenarios where the target domain has additional unknown classes compared to the limited-access source domain. Due to the absence of information on unknown classes, existing methods mainly transfer knowledge of known classes while roughly grouping unknown classes as one, attenuating the knowledge transfer and generalization. In contrast, this paper advocates that exploring unknown classes can better identify known ones, and proposes a domain adaptation model to transfer knowledge on known and unknown classes jointly. Specifically, given a source pre-trained model, we first introduce an unknown diffuser that can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider space distribution, we enhance the reliability of known class knowledge in the source pre-trained model through contrastive constraint. Finally, various supervision information, including reliable known class knowledge and clustered pseudo-labels, optimize the model for impressive knowledge transfer and generalization. Extensive experiments show that our network can achieve superior exploration and knowledge generalization on unknown classes, while with excellent known class transfer. The code is available at https://github.com/xdwfl/UPUK." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.866 + ], + "angle": 0, + "content": "By transferring knowledge from label-rich source domains to unlabeled target domains, unsupervised domain adaptation (UDA) has exhibited huge potential to address the challenges of costly data labeling and domain distribution shifts in real-world scenarios. Existing methods mainly learn the domain invariant representation by distribution matching" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.293, + 0.895, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.501, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Comparison of existing OS-SFDA methods and our proposed method. Existing methods ignore the unknown classes by grouping unknown classes as one, while our method can further explore the unknown class information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.78 + ], + "angle": 0, + "content": "[19, 20, 34, 40, 46] and adversarial optimization [7, 21, 35], to mitigate domain distribution shift. They can achieve impressive transfer effects by following the close-set and real-time data-accessible assumption, which shares the same class space in source and target domains, and accesses the source data at any time. However, in the real world, the open-set phenomenon, where the target domain has additional unknown class space to the source domain, always appears along with limited access to source data [5, 16]. In such most realistic scenarios, these methods fail to transfer knowledge without accessing source data and shared class distribution, making the knowledge transfer in such scenarios challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, to tackle such a challenging and realistic task, the open-set source-free domain adaptation (OS-SFDA) has emerged to perform knowledge transfer in the open-set scenario by accessing only the source pre-trained model, not the source data. Considering the lack of information on target unknown classes, existing approaches [6, 22, 31, 37] focus on transferring knowledge of known classes through pseudo-supervision, while grouping unknown classes as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.248, + 0.9 + ], + "angle": 0, + "content": "†The corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.248, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24015" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "one, as seen Fig. 1. However, such a grouping way restricts the thorough exploration of target class information, thereby diminishing models' transfer and generalization abilities on known and unknown classes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.392 + ], + "angle": 0, + "content": "To this end, we present a domain adaptation network for OS-SFDA, to explore target class space and transfer knowledge on known and unknown classes jointly. Specifically, given the source pre-trained model, an unknown diffuser is introduced to determine to-be-split/merged classes through similarity measures, generating a wider class space distribution, including known and unknown classes. Based on a wider distribution in the feature space, the reliability of the known class knowledge from the source pre-trained model is enhanced through contrastive constraint. Finally, the reliable known class knowledge, along with the clustered pseudo-labels spanning the extensive class space, can collaboratively serve as supervision during model optimization. This collaborative approach results in notable knowledge transfer and enhanced generalization capabilities. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.398, + 0.468, + 0.456 + ], + "angle": 0, + "content": "- We propose a domain adaptation network that is the first to excavate the unknown class space for boosting knowledge transfer and generalization to the target domain in OS-SFDA." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.459, + 0.468, + 0.517 + ], + "angle": 0, + "content": "- We design an unknown diffuser to explore wider and more precise target class space in the target domain, further benefiting known knowledge transfer and unknown generalization." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.52, + 0.468, + 0.578 + ], + "angle": 0, + "content": "- To realize effective known knowledge transfer and unknown generalization, we utilize the reliable known knowledge and clustered pseudo-labels in the broader class space, to serve as supervision during optimization." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.58, + 0.468, + 0.623 + ], + "angle": 0, + "content": "- Experiments can demonstrate that our method can explore target class space precisely, and yield state-of-the-art results among multiple OS-SFDA benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.398, + 0.468, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.218, + 0.655 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.362, + 0.681 + ], + "angle": 0, + "content": "2.1. Source-Free Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In recent years, with accessing only the source pre-trained model for transferring knowledge to the target domain, source-free domain adaptation (SFDA) methods [4, 16, 27, 41, 43-45, 47] have been widely used for overcoming source data privacy to achieve knowledge transfer in real-world scenarios. The existing methods are mainly divided into two categories, including methods based on pseudo-labeling [14, 16] and based on neighborhood clustering [43, 44]. For the former approaches, the classical method - SHOT [16] freezes the source classifier and optimizes the feature extractor, through mutual information maximization and source-target feature alignment based on self-supervised pseudo-labeling. To fully utilize the source and target domain knowledge, CoWA-JWDS [14] intro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "duces a new sample confidence score – JMDS better to exert the importance of pseudo-label confidence in distinguishing samples. For the latter approaches, G-SFDA [43] proposes a method of clustering target features and semantically similar neighborhoods to make the model better adapt to the target domain. AaD [44] has introduced prediction consistency of local neighborhood features and discrete prediction of different potential features, which achieve efficient feature clustering and distribution to better transfer knowledge. In addition, there are also some approaches to tackling SFDA from other aspects. U-SFAN [27] uses predicted quantified uncertainty to guide target adaptation. CRS [47] embeds class relationship similarity to transfer domain-invariant class relationships." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.765, + 0.33 + ], + "angle": 0, + "content": "2.2. Open-Set Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Recently, open-set domain adaptation (OSDA) tackles the realistic open-set challenge, where the target domain has additional unknown class space to the source domain. Currently, existing methods [2, 10, 15, 18, 29, 30, 33, 38] transfer known class knowledge while grouping unknown classes as one. For example, OSBP [30] uses pseudo labels to guide classifier learning and to construct constraints for the optimization of the model. Later, to better utilize pseudo-labels, STA [18] proposes an end-to-end method to gradually separate known and unknown classes while aligning feature distributions. OVANET [29] introduces the one-vs-all classifier to learn the inter- and intra-class distance, and minimizes class entropy to distinguish known and unknown classes. OSLPP [38] learns a common subspace from the source and target domains while gradually selecting and rejecting pseudo-labeled target data to promote the model's transfer on known classes. By introducing adversarial learning, UADAL [10] aligns the source- and the target-known distribution while segregating the target-unknown distribution, achieving better knowledge transfer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.862, + 0.667 + ], + "angle": 0, + "content": "2.3. Open-Set Source-Free Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "When the open-set phenomenon appears, along with limited access to source data, the methods without accessing source data and shared class distribution cannot transfer favorable knowledge, causing them to fail in real-world scenarios. OS-SFDA [6, 11, 13, 17, 22, 26, 37, 42] has emerged to perform knowledge transfer in the open-set scenario with accessing only the source pre-trained model, not the source data. Specifically, the existing methods generally transfer known knowledge and group unknown classes as one. Among these methods, without source data, FS [13] designs the self-adaptive model to transfer the source task-specific knowledge to the target domain effectively, and the target unknown classes are grouped into one class under the open-set setting due to the lack of unknown class information. To tackle OS-SFDA, OSHT [6] adopts pseudo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "labeling for adaptation and the entropy-based metric to reject unknown classes. Distill-SODA [37] introduces novel style-based adversarial data augmentation and the closed-set affinity score for better knowledge distillation from a self-supervised ViT during adaptation. SF-PGL [22] introduces a balanced pseudo-labeling strategy to progressive graph learning for adaptation and distinguishes known and unknown classes by applying the confidence threshold. In addition, the SFDA methods [14, 16, 44, 47] separate known and unknown classes and apply their strategies mentioned in Sec. 2.1 for known knowledge transfer under the open-set setting to tackle the OS-SFDA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.651 + ], + "angle": 0, + "content": "However, under the OS-SFDA setting, where the target domain has a additional abundant class distribution than the source domain, mentioned OS-SFDA methods group unknown classes as one, which restricts the thorough exploration of target class information, hampering knowledge transfer and generalization. Our work advocates exploring unknown classes can better identify known ones. It should be noted that recently some methods have attempted to explore unknown classes in OSDA [11, 23, 49], which requires source data access to realize domain adaptation. On the contrary, our focused OS-SFDA needs no source data for exploring unknown class space that is more realistic and challenging, essentially differing from them. Concretely, their source data access results in the failure to deal with the most realistic OS-SFDA. Moreover, they utilize 'unknown class number prior' to predefine [11, 23] or enumerate [49] unknown class numbers to explore the unknown class space, whose reliance on unknown prior further limits their scalability in real world. Instead, during unknown class exploration, our method requires no source data and unknown class number prior, truly achieving spontaneous and applicable unknown class exploration in OS-SFDA. We present a domain adaptation network for OS-SFDA, which can fully explore the target class space to transfer knowledge on known classes and generalize on unknown classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.17, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.719 + ], + "angle": 0, + "content": "In this section, we will introduce the details of our proposed method, and the framework is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Problem Setting. OS-SFDA aims to solve the knowledge-transferring problem in more realistic scenarios, where the source data have restricted access, and the target domain exhibits wider class distribution than the source domain. In such a task with restricted-accessible source data, we are given an unlabeled target domain \\(\\mathcal{D}^t = \\{(x_i^t)\\}_{i=1}^n\\) with \\(n\\) target samples, and a pre-trained source model. More importantly, the class space \\(\\mathcal{Y}^t\\) of the target domain is broader than the one \\(\\mathcal{Y}^s\\) of the source domain: \\(\\mathcal{Y}^t = \\mathcal{Y}^s \\cup \\mathcal{Y}^{ukn}\\). In other words, in addition to the source domain's class space \\(\\mathcal{Y}^s\\), \\(\\mathcal{Y}^t\\) also contains an additional unknown class space \\(\\mathcal{Y}^{ukn}\\). Here we call \\(\\mathcal{Y}^s\\) and \\(\\mathcal{Y}^{ukn}\\) as the target domain's" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "known class and unknown class spaces, respectively, whose numbers of categories are denoted as \\( K_{n} \\) and \\( K_{un} \\). The source domain pre-trained model is obtained by pre-training on the source domain in \\( \\mathcal{Y}^s \\), mainly consisting of feature extractor \\( f(\\cdot) \\) and classifier \\( g(\\cdot) \\). Therefore, it is expected to enable the source pre-trained model to overcome the class distribution discrepancy between domains, while achieving further knowledge transfer on the known classes and exploration on the unknown classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.235, + 0.828, + 0.252 + ], + "angle": 0, + "content": "3.1. Exploration of the Target Class Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Inspired by the effectiveness of no-parameter clustering methods [1, 25, 32, 48] in exploring the class space, we introduce an unknown diffuser based on the Multilayer Perceptron (MLP), which includes the main network \\( C \\) and the sub-network \\( C_k^{sub} \\) in the current class space, whose goal is to explore the target class space by inferring the target class number \\( K \\). Reasonably, we first initialize the target class space with the source known class space; that is, we assume that the initial value of the target class number \\( K \\) is the number of known classes \\( K_n \\). The unknown diffuser progressively explores a wider class space in the target domain through continuous inferring and optimization of \\( K \\). Note that the value of \\( K \\) changes dynamically with the inference process during training. The unknown diffuser completes the exploration of the target class space by performing two training processes: optimization in the latest class space and hierarchical class exploration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.896, + 0.879 + ], + "angle": 0, + "content": "Optimization in Latest Class Space. Since the target domain has no labels during exploration, we need to use the clusters obtained by clustering the target data to guide us in exploring the target class space. Simultaneously, we transform the exploration of the target class space into the exploration of the target cluster number. In this part, we learn the discriminative cluster assignments through the constraints of the feature-cluster assignment, which is beneficial for exploring the target cluster number. Firstly, the feature is the output of feature extractor \\( f(\\cdot) \\) in the source pre-trained model, denoted as \\( z_{i} = f(\\pmb{x}_{i}^{t})\\in \\mathbb{R}^{d} \\), where \\( d \\) is the dimension of the feature space, \\( i \\) represents the \\( i \\)-th sample. The features of \\( n \\) samples are denoted as \\( Z \\). Then, the main network \\( C \\) in the unknown diffuser maps the feature into the cluster assignment \\( \\widetilde{\\boldsymbol{Z}} = c(\\boldsymbol {Z}) \\), where \\( c(\\cdot) \\) is the mapping operation of \\( C \\), and the cluster assignment of feature \\( z_{i} \\) is denoted as \\( \\widetilde{z}_i\\in \\mathbb{R}^K \\). The discriminative cluster assignments under the current class space \\( (K) \\) are obtained by optimizing the cluster assignment under the guidance of the discriminative feature \\( Z \\). Specifically, to obtain the pseudo cluster-assignment \\( Y^{Z} = \\{y_{i}^{Z}\\}_{i = 1}^{n}, y_{i}^{Z}\\in \\mathbb{R} \\), we first use \\( k \\)-means under current target class space \\( (K) \\) on the feature set \\( Z \\). Then, according to \\( Y^{Z} \\), a teacher cluster soft-distribution \\( \\widetilde{Y}\\in \\mathbb{R}^{n\\times K} \\) is generated based on the feature:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.883, + 0.895, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\widetilde {y} _ {i, k} = \\left\\| \\boldsymbol {z} _ {i} - \\boldsymbol {\\eta} _ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24017" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.873, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.437 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed method. Given the source pre-trained model, we first introduce an unknown diffuser on its initial known class space. The unknown diffuser can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider class space, we improve the reliability of the known class knowledge inside the source model by contrastive learning. In this way, the reliable known class knowledge and clustered pseudo-labels over the wider class space, can jointly serve as supervision during model optimization, leading to powerful knowledge transfer and generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.601 + ], + "angle": 0, + "content": "where \\(\\widetilde{y}_{i,k}\\) is the \\(k\\)-th element in \\(\\widetilde{\\mathbf{y}}_i\\), and \\(\\widetilde{\\mathbf{y}}_i\\) is the teacher cluster soft-distribution of feature \\(\\mathbf{z}_i\\). \\(\\eta_k = \\sum_{i + 1}^{n_k}z_i^k /n_k\\) is the center feature of the \\(k\\)-th center and \\(n_k\\) is the feature number of the \\(k\\)-th master cluster. And \\(\\boldsymbol {z}_i^k\\) is the \\(i\\)-th element in \\(\\mathbf{Z}^k\\in \\mathbb{R}^{n_k\\times d}\\), which is the feature matrix of those features with \\(y_{i}^{Z} = k\\). Note that the teacher cluster soft-distribution of each feature is normalized. Then, we utilize the alignment loss to enforce the consistency of teacher cluster soft-distribution \\(\\widetilde{\\mathbf{Y}}\\) with the cluster assignment \\(\\widetilde{\\mathbf{Z}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.61, + 0.47, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l g} = \\sum_ {i = 1} ^ {n} L _ {a l g} \\left(x _ {i} ^ {t}\\right) = \\sum_ {i = 1} ^ {n} \\widetilde {\\boldsymbol {z}} _ {i} \\cdot \\widetilde {\\boldsymbol {y}} _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.72 + ], + "angle": 0, + "content": "The cluster assignment under the current target class space \\((K)\\) can be promoted through optimization on \\(c(\\cdot)\\) in the training process, so as to promote the exploration of the target class space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Hierarchical Class Exploration. To fully explore the target class space by inferring the precise target class number \\( K \\), inspired by [3], the Metropolis-Hastings framework (M-H) is utilized as the underlying framework in the class diffusion process, which computes the Hasting ratio for changing \\( K \\) through splitting and fusion. Besides that, inspired by the hierarchical clustering [39], we design a hierarchical class exploration method to explore the target class space based on the M-H framework, that performs splitting and merging on more dispersive and coupled clusters, respectively. Firstly, we obtain the discriminative cluster assignment after splitting/merging in all possible split-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.6 + ], + "angle": 0, + "content": "ting/merging cases. We then select the potential clusters to be split or merged based on the assignment distance between clusters. In the end, these selected clusters are put into the candidate set to be split or merged, while we use the M-H framework [8] on those clusters in the candidate set to determine the splitting and merging decisions. We assume that the current cluster distributions and those under all possible splitting/merging cases are called the distributions before and after splitting/merging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Specifically, we first obtain the discriminative cluster distribution after splitting/merging in all possible splitting/merging cases, in which the cluster before splitting/merging is already learned above. Then, the cluster distribution after merging can be acquired directly by merging the assignments of two clusters. However, obtaining the sub-cluster distribution after splitting is apparently unreasonable, by randomly dividing the target data in the master cluster. For subsequent convenience, we define the clusters before/after splitting as the master cluster and sub-cluster, respectively. To gain the discriminative sub-cluster distribution, we introduce the sub-network \\( C_k^{sub} \\) into the unknown diffuser, and \\( C_k^{sub} \\) contains \\( K \\) sub-clustering layers with the output dimension equaling to 2. One of these sub-network \\( C_k^{sub} \\) is applied on each master cluster to generate the sub-cluster assignment \\( \\widetilde{Z}^{sub,k} = c_k^{sub}(Z^k) \\), in which \\( c_k^{sub}(\\cdot) \\) represents the operation of the \\( k \\)-th sub-network applying on the \\( k \\)-th master cluster. Then, to obtain the pseudo-sub-cluster assignment, the \\( k \\)-means first is performed on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24018" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.156 + ], + "angle": 0, + "content": "feature matrix \\( \\mathbf{Z}^k \\) of the \\( k \\)-th master cluster, similar to cluster optimization with Eq. (1), Subsequently, a teacher subcluster assignment \\( \\widetilde{\\mathbf{Y}}^k \\in \\mathbb{R}^{n_k \\times 2} \\) is also generated for the \\( k \\)-th master cluster:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.16, + 0.469, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {i, j} ^ {k} = \\left\\| \\boldsymbol {z} _ {i} ^ {k} - \\boldsymbol {\\eta} _ {j} ^ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.47, + 0.243 + ], + "angle": 0, + "content": "where \\(\\widetilde{y}_{i,j}^{k}\\) is the \\(i\\)-th row and \\(j\\)-th column element in \\(\\widetilde{\\pmb{Y}}^k\\). \\(\\pmb{\\eta}_j^k\\) is the center feature of the \\(j\\)-th sub-cluster in \\(k\\)-th master cluster. And we utilize the alignment loss to optimize them:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.25, + 0.469, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l g} ^ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\widetilde {\\boldsymbol {z}} _ {i} ^ {\\text {s u b}, k} \\cdot \\widetilde {\\boldsymbol {y}} _ {i} ^ {k}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.3, + 0.469, + 0.348 + ], + "angle": 0, + "content": "where \\(\\widetilde{z}_i^{sub,k}\\) is the \\(i\\)-th element in \\(\\widetilde{\\mathbf{Z}}^{sub,k}\\). The training process on the sub-network \\(C_k^{sub}\\) is synchronous with the one on the main network \\(C\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.531 + ], + "angle": 0, + "content": "In addition, hierarchical clustering [39] gradually divides data by merging pairs of clusters with smaller distances or splitting the two most distant sub-clusters in one cluster. According to the assignment distance between clusters, the candidate sets to be split/merged are generated, inspired by the hierarchical clustering. In our method, to explore more different classes, those cluster pairs (sub-cluster pairs) with far assignment distances are considered as those likely to be split. While those with close assignment distances are considered as those likely to be merged due to high similarity between two main clusters. The candidate sets to be split/merged based on this can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.537, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\nC S _ {m e r g} ^ {E D} = \\left\\{\\left(k _ {1}, k _ {2}\\right) \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} E D \\left(\\boldsymbol {\\eta} _ {k _ {1}}, \\boldsymbol {\\eta} _ {k _ {2}}\\right) \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.562, + 0.469, + 0.581 + ], + "angle": 0, + "content": "\\[\nC S _ {s p l} ^ {E D} = \\left\\{k \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} - E D \\left(\\boldsymbol {\\eta} _ {1} ^ {k}, \\boldsymbol {\\eta} _ {2} ^ {k}\\right) \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\( ED(\\cdot, \\cdot) \\) computes the Euclidean distance of the cluster embedding center. \\( n_{if} \\) are the number of cluster(s) with potential to be merged and split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.765 + ], + "angle": 0, + "content": "After obtaining the candidate set, we can infer more accurate \\( K \\) and explore a more accurate target class space, by calculating the Hasting ratio from the M-H framework [8] among (sub)cluster pairs in the candidate set and comparing it with 1. More details are provided in the supplementary material. In this way, we use the inferred latest target class number \\( K \\) to update the unknown diffuser (the output dimension of \\( C \\) and the number of sub-clustering layers) at the next optimization step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "During the optimization in the latest more accurate class space, we learn the discriminative cluster distribution under the latest updated target class number \\( K \\). According to these cluster distributions, we decide the splitting and merging decisions to change the target class number \\( K \\) in the process of hierarchical class exploration. The more precise \\( K \\) and corresponding cluster assignments are generated, through operating these two processes iteratively, which is beneficial to fully explore the target class space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Algorithm 1 The training procedure of our proposed model \nInput: data, source pre-trained model, main network \\( C \\) and sub-network \\( C_k^{sub} \\) of unknown diffuser" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.141, + 0.782, + 0.156 + ], + "angle": 0, + "content": "1: // Exploration of the target class space" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.157, + 0.717, + 0.171 + ], + "angle": 0, + "content": "2: for \\(iter_1 = 1\\) to \\(iter_1^{max}\\) do" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.141, + 0.782, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.172, + 0.891, + 0.2 + ], + "angle": 0, + "content": "3: Obtain clusteredpesudo-distribution \\(\\widetilde{Y}\\) and \\(\\widetilde{Y}^k\\) by \\(k\\)-means" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.202, + 0.891, + 0.233 + ], + "angle": 0, + "content": "4: Optimize \\(C\\) and \\(C_k^{sub}\\) guided by \\(\\widetilde{Y}\\) and \\(\\widetilde{Y}^k\\) in Eqs. (2) and (4)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.234, + 0.891, + 0.263 + ], + "angle": 0, + "content": "5: Determine \\( CS_{slp} / CS_{merg} \\) by similarity measurement in Eqs. (5) and (6)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.264, + 0.892, + 0.294 + ], + "angle": 0, + "content": "6: Diffuse the class space by updating \\( K \\) on \\( CS_{slp} / CS_{merg} \\), based on M-H framework" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.295, + 0.856, + 0.308 + ], + "angle": 0, + "content": "7: Update the unknown diffuser by the updated \\(K\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.31, + 0.584, + 0.322 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.324, + 0.798, + 0.339 + ], + "angle": 0, + "content": "9: // Knowledge transfer and generalization" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.34, + 0.717, + 0.354 + ], + "angle": 0, + "content": "10: for iter2 = 1 to iter2 max do" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.355, + 0.891, + 0.383 + ], + "angle": 0, + "content": "11: Promote reliability of known knowledge by contrastive constraint in Eq. (7)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.385, + 0.891, + 0.415 + ], + "angle": 0, + "content": "12: Jointly optimize \\(C\\) and the pre-trained model in the explored class space via \\(\\mathcal{L}_{kt}\\) and \\(\\mathcal{L}_{alg}\\) in Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.416, + 0.583, + 0.428 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.172, + 0.892, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.457, + 0.844, + 0.473 + ], + "angle": 0, + "content": "3.2. Knowledge Transfer and Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.556 + ], + "angle": 0, + "content": "After exploring the target class space in Sec. 3.1, we aim to transfer knowledge on known classes more effectively and improve the model's generalization on unknown classes. In this subsection, we promote the reliability of known knowledge and further optimize over wider class distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.798 + ], + "angle": 0, + "content": "Promotion of Known Knowledge Reliability. To better utilize known knowledge, we hope to improve the reliability of known knowledge obtained by the classification layer of the pre-trained source model during training. The input of the model is all target samples, including known and unknown samples, from which we need to select the known ones. According to the logit output of known samples having smaller entropy than the unknown ones, we select the top \\(50\\%\\) of samples with the lowest entropy as known samples, which are placed in set \\(\\mathcal{M}\\). After selection, the contrastive constraint \\(\\mathcal{L}_{con}\\) is utilized on \\(\\mathcal{M}\\) to enhance the discrimination of the known samples. Concretely, it brings the known features of similar predictions closer and those of dissimilar predictions farther away, achieving discrimination feature assignment to strengthen the reliability of known knowledge. The contrastive constraint is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.81, + 0.891, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = - \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {+} + \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {-}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.891, + 0.884 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_i = \\sigma(g(\\pmb{z}_i)) \\in \\mathbb{R}^d\\) is the output of classifier \\(g\\) and \\(\\sigma\\) is the softmax function." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Optimization over Wider Class Distribution. After" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.501 + ], + "angle": 0, + "content": "that, we optimize the unknown diffuser over the wider class space by taking advantage of the reliable known knowledge and alignment constraint on the explored space. In such an optimization based on the explored wider class space, we first select the high-confidence known samples as the reliable known knowledge. Then, we hope to use the hard pseudo-labels of the high-confidence known samples as supervision information, thereby leveraging them for better knowledge transfer on the explored class space. Specifically, for a known sample, we take the maximum logit output by the classifier in the source pre-trained model as its confidence. And the known sample is considered as the high-confidence known sample when it is greater than a manually-set confidence threshold, whose hard pseudolabel is represented by the one-hot paradigm \\( \\pmb{y}_i^h (x_i^h\\in \\mathcal{H}) \\). Note that \\( \\mathcal{H} \\) is the set of the high-confidence known samples. Therefore, we can perform better knowledge transfer on the explored class space by leveraging the reliable known knowledge \\( \\pmb{y}_i^h \\) obtained on the source pre-trained model. In addition, we also perform alignment constraint on the explored class space, similar to Eq. (2), to complete the transfer on known samples further and achieve generalization on unknown samples. Finally, we leverage the supervision of the high-confidence known sample labels and the alignment constraint, to jointly achieve superior knowledge transfer on known classes and generalization on unknown classes over the wider class space:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.509, + 0.47, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {k t} + \\mathcal {L} _ {a l g} = \\sum_ {i} ^ {\\mathcal {H}} L _ {k t} (\\widetilde {\\boldsymbol {z}} _ {i}, \\boldsymbol {y} _ {i} ^ {h}) + \\sum_ {i = 1} ^ {n} L _ {a l g} (x _ {i} ^ {t}) \\\\ = \\underbrace {\\sum_ {i} ^ {\\mathcal {H}} L _ {k t} \\left(\\widetilde {\\boldsymbol {z}} _ {i} , \\boldsymbol {y} _ {i} ^ {h}\\right) + \\sum_ {j _ {1} = 1} ^ {n _ {1}} L _ {a l g} \\left(x _ {j _ {1}} ^ {t}\\right)} _ {\\text {k n o w n}} + \\underbrace {\\sum_ {j _ {2} = 1} ^ {n _ {2}} L _ {a l g} \\left(x _ {j _ {2}} ^ {t}\\right)} _ {\\text {u n k n o w n}}, \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.627, + 0.47, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{kt}\\) represents the guidance of high-confidence known samples, which is the cross-entropy loss between their pseudo-labels and soft distributions. \\(\\mathcal{L}_{alg}\\) is the alignment constraint to achieve superior knowledge transfer on known classes and generalization on unknown classes. \\(\\widetilde{z}_i\\) represents the soft-distribution output by the main network \\(C\\) on the explored class space. \\(n_1\\) and \\(n_2\\) represent the number of known and unknown samples, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.21, + 0.778 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.268, + 0.802 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three benchmark datasets for OS-SFDA through experiments, which are Office-31 [28], Office-Home [36], and VisDA-C 2017 [24]. Due to limited space, the results and analysis on VisDA-C 2017 are summarized in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Office-31 [28] consists of three domains (Amazon, We" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "bcam, DSLR), with 31 classes and 4652 images of common objects. Following the protocols [18] on Office-31 and Office-Home, we select the first 20 classes as the shared classes of the source and target domain while the remaining 11 classes are considered unknown." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.258 + ], + "angle": 0, + "content": "- Office-Home [36] is a more challenging dataset. It contains four domains (Art, Product, Clipart, RealWorld) with 65 classes and 15500 images of common objects. Like Office-31, we select the first 25 classes as classes shared by the source and target domain, and the remaining 40 classes are considered unknown." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Baselines. For OS-SFDA, we compare our method with the clustering on Resnet [9], the classic method SHOT [16], and the advanced method AaD [44], which do sufficient experiments under the OS-SFDA setting. In addition, our method uses the unknown diffuser based on clustering (class number-agnostic) to explore the target class space. For a fair comparison, we thoroughly compare our method with the no-parameter clustering methods (FINCH [32], COMIC [25], DenMune [1])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Evaluation. As shown in Tab. 1 and Tab. 2, we report domain adaptation results for \\( OS^{*} \\) and \\( UNK \\), which represent the average accuracy of the known and the unknown class, respectively. Simultaneously, \\( HOS = \\frac{2 \\times OS^{*} \\times UNK}{OS^{*} + UNK} \\) is denoted as the harmonic mean between \\( OS^{*} \\) and \\( UNK \\), which is the key indicator for evaluating performance. Moreover, the inferred target class number \\( K \\) and accuracy ACC are used to evaluate the effectiveness of exploring the target class space in Tab. 3 and Tab. 4. The setting of ACC follows the previous protocols [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Implementation Details. For the source pre-trained model, we utilize the backbone of ResNet-50 [9] for Office-31 and Office-Home. For a fair comparison, we use the same network architecture with SHOT [16] and AaD [44], which includes a backbone and two additional fully connected layers. We adopt SGD with a momentum of 0.9 and a batch size of 64 for all datasets. The learning rate is set to 1e-3 on all datasets for all layers of the backbone, and the last two added fully-connected layers are applied 1e-2. We train 15 epochs for all datasets. The unknown diffuser includes one convolutional layer with 256 hidden dimensions. For a fair comparison, our network and the compared no-parameter clustering methods use the same input features, output by the source pre-trained model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.696, + 0.788 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1 and Tab. 2, the UNK and HOS for our method are much higher than SHOT on all datasets. Our method achieves \\(10.4\\%\\) and \\(2.3\\%\\) relative improvements over the state-of-art AaD for average OS* and HOS on Office-31, respectively. On the relatively small-scale Office-31 with fewer classes, our method demonstrates competitiveness across various indicators, standing in con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24020" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.875, + 0.201 + ], + "angle": 0, + "content": "
methodSFA→DA→WD→AD→WW→AW→DAvg
OS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOS
Cluster✗10.064.017.07.061.013.05.069.010.08.049.014.05.064.010.010.047.017.07.559.013.5
SHOT✓87.250.055.789.038.653.878.244.256.597.472.383.073.547.657.889.675.081.785.854.664.8
AaD✓69.986.777.462.083.971.360.881.869.883.297.489.858.481.568.082.395.288.369.487.877.4
Ours(w/o s)✓78.473.475.887.990.389.172.376.174.192.880.586.270.976.873.785.879.882.781.479.580.3
Ours(w/s)✓82.371.876.790.072.080.072.079.475.579.587.383.274.282.778.280.688.384.379.880.379.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Table 1. Accuracy (%) on Office-31 for OS-SFDA. The three indicators(OS*, UNK, HOS) have been explained in Sec. 4.1. ‘w/o s’ and ‘w/s’ indicate with or without supervision using reliable known knowledge. Note that ‘A→D’ represents that the model per-trained on the source domain Amazon transfers to the target domain DSLR. Bold and ‘_’ represent the best and second results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.266, + 0.879, + 0.485 + ], + "angle": 0, + "content": "
methodSFAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwAvg (first 6 tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX6.063.011.05.055.010.04.062.08.08.059.014.05.055.09.05.056.08.05.056.08.05.558.310.0
SHOT✓67.028.039.581.826.339.887.532.147.066.846.254.677.527.240.280.025.939.176.831.043.4
AaD✓50.567.457.764.066.465.172.269.570.847.180.359.364.768.266.465.071.067.860.670.564.5
Ours(w/o s)✓47.669.256.470.081.875.470.685.677.460.374.766.765.682.773.272.184.277.664.479.771.1
Ours(w/s)✓49.064.855.868.188.076.771.586.678.461.372.666.466.581.673.171.684.877.664.779.771.3
methodSFPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrRw→PrAvg (all tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX8.053.015.06.050.010.05.061.09.08.051.013.05.064.010.05.050.010.05.856.610.6
SHOT✓66.351.157.759.331.040.885.831.646.273.550.659.965.328.940.184.428.242.374.633.945.6
AaD✓46.983.160.045.072.655.669.072.370.656.077.465.048.367.656.467.769.368.558.072.163.6
Ours(w/o s)✓60.076.767.352.666.258.670.785.177.256.583.367.449.170.757.966.182.173.261.878.569.0
Ours(w/s)✓55.985.667.645.470.255.173.983.978.656.784.167.849.374.659.469.580.074.461.679.769.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.495, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Table 2. Accuracy (%) on Office-Home for OS-SFDA. Note that ‘Ar→Cl’ represents that the model per-trained on the source domain Art transfers to the target domain Clipart. The representations of OS*, UNK, HOS, ‘w/o s’, and ‘w/s’ are the same with Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.749 + ], + "angle": 0, + "content": "trast to methods such as AaD and SHOT, where only one indicator exhibits competitive performance. SHOT tends to excel in OS*, while AaD typically performs well on UNK. In contrast, our method displays a well-balanced performance across different measurement dimensions, achieving competitive results overall. For the more challenging Office-Home, our method displays more apparent superiority, which achieves \\(61.6\\%\\), \\(79.7\\%\\), and \\(69.2\\%\\) average accuracy for OS*, UNK, and HOS, with \\(3.6\\%\\), \\(7.6\\%\\), and \\(5.6\\%\\) relative improvement compared with AaD. Furthermore, our method outperforms all no-parameter clustering methods (FINCH, DenMune, and COMIC) for ACC and \\(K\\) on all tasks in Tab. 3 and Tab. 4. The above results demonstrate the superiority and effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Impact of Exploration for the Target Class Space. Compared with OS-SFDA methods (SHOT and AaD) that group all target unknown classes as one, our method tends to explore class space and excavate different unknown classes, which has great application value in more realistic scenarios. The experimental results show that our method achieves excellent improvement over the state-of-the-art method AaD for the known accuracy \\(\\mathrm{OS^{*}}\\) on the more challenging Office Home. In addition, the unknown accuracy UNK for our method is much higher than SHOT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.749 + ], + "angle": 0, + "content": "on Office-31 and Aad on Office-Home. This phenomenon shows that our method can improve knowledge transfer on known classes when exploring class space and simultaneously improve generalization on unknown classes. To more fully demonstrate the effectiveness and advancement of our exploration, we compare our method with classical no-parameter clustering methods for exploring the class space. Concretely, ours provides the more precise inferred target class number \\( K \\) and higher accuracy ACC on all tasks, where the precision of \\( K \\) represents the efficacy in exploring the target class space, as discussed in Sec. 3.1. The above analysis conclusively demonstrates that our method of exploring class space is exceptionally effective, improving knowledge transfer and generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Effect of Reliable Known Knowledge as Supervision. As introduced in Sec. 3.2, we hope to utilize the reliable known knowledge to optimize the unknown diffuser for better knowledge transfer and generalization. Concretely, our method that uses the reliable known knowledge(ours (w/ s)) outperforms all compared methods(including ours(w/o s)) on more challenging Office-Home for average HOS, ACC, and \\(K\\) in Tabs. 1 to 4. Therefore, the more effectiveness brought by reliable known knowledge ('ours (w/ s)') confirms it guides the model for better known and unknown" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "24021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.104, + 0.268, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.231, + 0.192, + 0.245 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.104, + 0.465, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.231, + 0.392, + 0.245 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.104, + 0.663, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.231, + 0.591, + 0.245 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.103, + 0.865, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.231, + 0.794, + 0.245 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.255, + 0.885, + 0.27 + ], + "angle": 0, + "content": "Figure 3. (a) t-SNE of SHOT on \\(\\mathbb{R}\\to \\mathbb{P}\\). (b) t-SNE of AaD on \\(\\mathbb{R}\\to \\mathbb{P}\\). (c) t-SNE of our method on \\(\\mathbb{R}\\to \\mathbb{P}\\). (d) Parameter analysis on \\(\\mathrm{A}\\rightarrow \\mathrm{W}\\)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.293, + 0.468, + 0.392 + ], + "angle": 0, + "content": "
MethodsA→DA→WD→AD→WW→AW→D
FINCH31.0 (8)8.0 (2)23.0 (7)12.0 (3)28.0 (9)23.0 (5)
DenMune20.0(5)52.0(15)58.0(29)46.0(14)65.0(35)22.0(5)
COMIC56.8 (22)47.3 (24)16.1 (40)46.4 (20)12.1 (32)56.6 (22)
Ours(w/o s)69.0(24)80.6(28)60.4(37)82.5(30)59.6(38)68.9(27)
Ours(w/ s)71.2(25)73.3(28)63.1(37)76.3(32)64.2(38)74.5(28)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.445 + ], + "angle": 0, + "content": "Table 3. Accuracy ACC(%) and the inferred target class number \\( K \\) on Office-31. Note that the values before \\( \\left( \\cdot \\right) \\) ’ and inside \\( \\left( \\cdot \\right) \\) ’ represent the performance ACC and \\( K \\) ,respectively." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.459, + 0.462, + 0.653 + ], + "angle": 0, + "content": "
MethodsAr→ClAr→PrAr→RwCl→ArCl→PrCl→Rw
FINCH5.0(4)7.0(3)6.0(3)8.0(3)7.0(4)7.0(4)
DenMune15.0(12)58.0(45)54.0(39)16.0(7)55.0(40)40.0(32)
COMIC6.9(65)50.3(92)9.43(34)4.0(1)10.4(51)2.3(2)
Ours(w/o s)35.5(69)60.5(77)59.1(77)47.3(61)58.9(74)55.6(60)
Ours(w/s)35.9(69)61.4(80)61.3(77)46.9(62)58.9(78)61.0(60)
MethodsPr→ArPr→ClPr→RwRw→ArRw→PrRw→Cl
FINCH12.0(4)5.0(4)4.0(2)11.0(4)12.0(7)7.0(4)
DenMune16.0(7)7.0(6)52.0(37)26.0(13)9.0(8)62.0(43)
COMIC4.2(2)3.7(24)5.7(14)4.1(1)5.2(43)13.3(65)
Ours(w/o s)48.5(51)37.2(56)63.6(75)38.1(28)21.6(50)50.4(75)
Ours(w/s)47.8(51)35.5(56)64.0(68)44.2(35)21.9(50)51.8(75)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.662, + 0.399, + 0.675 + ], + "angle": 0, + "content": "Table 4. ACC(%) and \\(K\\) on Office-Home." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.718 + ], + "angle": 0, + "content": "class discrimination, achieving better knowledge transfer on known classes and generalization on unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Parameter Analysis. We study the impact of supervision information on unknown diffuser optimization during training by applying a variable parameter to the supervision loss. As shown in Fig. 3(d), the chart indicates that when there is no supervision information, the optimization cannot be completed well to obtain a favorable transfer effect on known or unknown classes. As the parameters increase within a certain range, UNK and HOS also improve; but parameters that are too large will lead to poor optimization results. In addition, it can be seen that OS* is not sensitive to this parameter, and its change trend is relatively stable. Therefore, we should select appropriate high" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.341 + ], + "angle": 0, + "content": "confidence known pseudo-labels as supervision information to prevent model overfitting for better guiding the optimization of unknown diffuser." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Effectiveness of the Unknown Diffuser. Tabs. 1 to 4 show our superior performance in exploring unknown class space and domain adaptation in OS-SFDA. The superiority in unknown class exploration indicates that our unknown diffuser can excavate potential unknown classes under the guidance of current discriminative class distributions based on the Metropolis-Hastings (M-H) framework. The impressive domain adaptation effects based on the explored wider space also demonstrate the necessity of using the unknown diffuser for unknown class exploration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Visualization of features. As shown in Fig. 3(a)-(c), t-SNE visualizations indicate that compared with existing methods, our method explores a more accurate and extensive class space, exhibiting remarkable class discrimination." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.619, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In this paper, we propose a domain adaptation network in OS-SFDA for exploring unknown classes, thereby jointly benefiting known knowledge transfer and unknown generalize. We introduce an unknown diffuser into the source pretrained model to excavate a wider target class space. And, the reliable known class knowledge and clustered pseudolabels over the wider class space are captured to serve as supervision during model optimization, realizing impressive knowledge transfer and generalization. Extensive experiments verify the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.764, + 0.688, + 0.781 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This work is supported in part by the National Key Research and Development Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62132016, 62171343, and 62201436), and Fundamental Research Funds for the Central Universities (ZDRC2102)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24022" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Mohamed Abbas, Adel El-Zoghabi, and Amin Shoukry. Denmune: Density peak based clustering using mutual nearest neighbors. Pattern Recognit., 109:107589, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 422-438. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Jason Chang and John W Fisher III. Parallel sampling of dp mixture models using sub-cluster splits. Proc. Adv. Neural Inf. Process. Syst., 26, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[4] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7212-7222, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[6] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE Trans. Image Process., 30:6473-6484, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[7] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. J. Mach. Learn. Res., 17(1):2096-2030, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[8] W Keith Hastings. Monte carlo sampling methods using markov chains and their applications. 1970. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[10] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35:16755-16767, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[11] Taotao Jing, Hongfu Liu, and Zhengming Ding. Towards novel target discovery through open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9322-9331, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[12] Harold W Kuhn. The hungarian method for the assignment problem. Nav. Res. Logist. Q., 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[13] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12376-12385, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[14] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 12365-12377. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[15] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 24110-24119, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.89, + 0.121 + ], + "angle": 0, + "content": "unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 6028-6039. PMLR, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[17] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[18] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2927-2936, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[19] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Proc. Adv. Neural Inf. Process. Syst., 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[20] Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(12):3071-3085, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[21] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[22] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[23] Yingwei Pan, Ting Yao, Yehao Li, Chong-Wah Ngo, and Tao Mei. Exploring category-agnostic clusters for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 13867–13875, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[24] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[25] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In Proc. Int. Conf. Mach. Learn., pages 5092-5101. PMLR, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[26] Masoud Pourreza. Open-set source-free domain adaptation in fundus images analysis. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[27] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 537-555. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[28] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In Proc. Eur. Conf. Comput. Vis., pages 213-226. Springer, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[29] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9000-9009, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[30] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proc. Eur. Conf. Comput. Vis., pages 153-168, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. Proc. Adv. Neural Inf. Process. Syst., 33:16282-16292, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Saquib Sarfraz, Vivek Sharma, and Rainer Stiefelhagen. Efficient parameter-free clustering using first neighbor relations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8934-8943, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[33] Tasfia Shermin, Guojun Lu, Shyh Wei Teng, Manzur Murshed, and Ferdous Sohel. Adversarial network with multiple classifiers for open set domain adaptation. IEEE Trans. Multimed., 23:2732-2744, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proc. AAAI Conf. Artif. Intell., 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[35] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darryll. Adversarial discriminative domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7167-7176, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[36] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 5018-5027, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[37] Guillaume Vray, Devavrat Tomar, Behzad Bozorgtabar, and Jean-Philippe Thiran. Source-free open-set domain adaptation for histopathological images via distilling self-supervised vision transformer. arXiv preprint arXiv:2307.04596, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[38] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[39] Joe H Ward Jr. Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc., 58(301):236-244, 1963. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[40] Kun Wei, Xu Yang, Zhe Xu, and Cheng Deng. Class incremental unsupervised domain adaptation via pseudolabel distillation. IEEE Trans. Image Process., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[41] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9010-9019, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[42] Jiexi Yan, Zhihui Yin, Erkun Yang, Yanhua Yang, and Heng Huang. Learning with diversity: Self-expanded equalization for better generalized deep metric learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 19365-19374, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[43] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8978-8987, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[44] Shiqi Yang, Shangling Jui, Joost van de Weijer, et al. Attracting and dispersing: A simple approach for source-free domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35: 5802-5815, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[45] Shiqi Yang, Yaxing Wang, Luis Herranz, Shangling Jui, and Joost van de Weijer. Casting a bait for offline and online source-free domain adaptation. Comput. Vis. Image Underst., page 103747, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.847, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Xu Yang, Cheng Deng, Tongliang Liu, and Dacheng Tao. Heterogeneous graph attention network for unsupervised multiple-target domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1992-2003, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Yixin Zhang, Zilei Wang, and Weinan He. Class relationship embedded learning for source-free unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7619-7629, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[48] Han Zhao, Xu Yang, and Cheng Deng. Parameter-agnostic deep graph clustering. ACM Trans. Knowl. Discov. Data, 18 (3):1–20, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[49] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Open set domain adaptation by novel class discovery. arXiv preprint arXiv:2203.03329, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24024" + } + ] +] \ No newline at end of file diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_origin.pdf b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a4a54f90230dfc77a696f12eb06e7fdb48ce5804 --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/496a9fbb-032a-4d75-b940-61e6c2422590_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf1e000fd6e2c7fa727f405d574e41df9a2b46b0840cea161cc94021450d01d6 +size 1272088 diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/full.md b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..82430d8ffc485673b5dfd6a1c305b1fdba4fc0b0 --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/full.md @@ -0,0 +1,287 @@ +# Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation + +Fuli Wan*, Han Zhao*, Xu Yang†, and Cheng Deng† +School of Electronic Engineering, Xidian University, Xi'an710071, China +{wfl128.xidian, hzhao1698, xuyang.xd, chdeng.xd}@gmail.com + +# Abstract + +Open-Set Source-Free Domain Adaptation aims to transfer knowledge in realistic scenarios where the target domain has additional unknown classes compared to the limited-access source domain. Due to the absence of information on unknown classes, existing methods mainly transfer knowledge of known classes while roughly grouping unknown classes as one, attenuating the knowledge transfer and generalization. In contrast, this paper advocates that exploring unknown classes can better identify known ones, and proposes a domain adaptation model to transfer knowledge on known and unknown classes jointly. Specifically, given a source pre-trained model, we first introduce an unknown diffuser that can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider space distribution, we enhance the reliability of known class knowledge in the source pre-trained model through contrastive constraint. Finally, various supervision information, including reliable known class knowledge and clustered pseudo-labels, optimize the model for impressive knowledge transfer and generalization. Extensive experiments show that our network can achieve superior exploration and knowledge generalization on unknown classes, while with excellent known class transfer. The code is available at https://github.com/xdwfl/UPUK. + +# 1. Introduction + +By transferring knowledge from label-rich source domains to unlabeled target domains, unsupervised domain adaptation (UDA) has exhibited huge potential to address the challenges of costly data labeling and domain distribution shifts in real-world scenarios. Existing methods mainly learn the domain invariant representation by distribution matching + +![](images/7258fd0de622d507a6a6acaa4714813d3feb5bf1a3e67c9e1a3389df1e1bb8ef.jpg) +Figure 1. Comparison of existing OS-SFDA methods and our proposed method. Existing methods ignore the unknown classes by grouping unknown classes as one, while our method can further explore the unknown class information. + +[19, 20, 34, 40, 46] and adversarial optimization [7, 21, 35], to mitigate domain distribution shift. They can achieve impressive transfer effects by following the close-set and real-time data-accessible assumption, which shares the same class space in source and target domains, and accesses the source data at any time. However, in the real world, the open-set phenomenon, where the target domain has additional unknown class space to the source domain, always appears along with limited access to source data [5, 16]. In such most realistic scenarios, these methods fail to transfer knowledge without accessing source data and shared class distribution, making the knowledge transfer in such scenarios challenging. + +Recently, to tackle such a challenging and realistic task, the open-set source-free domain adaptation (OS-SFDA) has emerged to perform knowledge transfer in the open-set scenario by accessing only the source pre-trained model, not the source data. Considering the lack of information on target unknown classes, existing approaches [6, 22, 31, 37] focus on transferring knowledge of known classes through pseudo-supervision, while grouping unknown classes as + +one, as seen Fig. 1. However, such a grouping way restricts the thorough exploration of target class information, thereby diminishing models' transfer and generalization abilities on known and unknown classes, respectively. + +To this end, we present a domain adaptation network for OS-SFDA, to explore target class space and transfer knowledge on known and unknown classes jointly. Specifically, given the source pre-trained model, an unknown diffuser is introduced to determine to-be-split/merged classes through similarity measures, generating a wider class space distribution, including known and unknown classes. Based on a wider distribution in the feature space, the reliability of the known class knowledge from the source pre-trained model is enhanced through contrastive constraint. Finally, the reliable known class knowledge, along with the clustered pseudo-labels spanning the extensive class space, can collaboratively serve as supervision during model optimization. This collaborative approach results in notable knowledge transfer and enhanced generalization capabilities. Our contributions are summarized as follows: + +- We propose a domain adaptation network that is the first to excavate the unknown class space for boosting knowledge transfer and generalization to the target domain in OS-SFDA. +- We design an unknown diffuser to explore wider and more precise target class space in the target domain, further benefiting known knowledge transfer and unknown generalization. +- To realize effective known knowledge transfer and unknown generalization, we utilize the reliable known knowledge and clustered pseudo-labels in the broader class space, to serve as supervision during optimization. +- Experiments can demonstrate that our method can explore target class space precisely, and yield state-of-the-art results among multiple OS-SFDA benchmarks. + +# 2. Related Work + +# 2.1. Source-Free Domain Adaptation + +In recent years, with accessing only the source pre-trained model for transferring knowledge to the target domain, source-free domain adaptation (SFDA) methods [4, 16, 27, 41, 43-45, 47] have been widely used for overcoming source data privacy to achieve knowledge transfer in real-world scenarios. The existing methods are mainly divided into two categories, including methods based on pseudo-labeling [14, 16] and based on neighborhood clustering [43, 44]. For the former approaches, the classical method - SHOT [16] freezes the source classifier and optimizes the feature extractor, through mutual information maximization and source-target feature alignment based on self-supervised pseudo-labeling. To fully utilize the source and target domain knowledge, CoWA-JWDS [14] intro + +duces a new sample confidence score – JMDS better to exert the importance of pseudo-label confidence in distinguishing samples. For the latter approaches, G-SFDA [43] proposes a method of clustering target features and semantically similar neighborhoods to make the model better adapt to the target domain. AaD [44] has introduced prediction consistency of local neighborhood features and discrete prediction of different potential features, which achieve efficient feature clustering and distribution to better transfer knowledge. In addition, there are also some approaches to tackling SFDA from other aspects. U-SFAN [27] uses predicted quantified uncertainty to guide target adaptation. CRS [47] embeds class relationship similarity to transfer domain-invariant class relationships. + +# 2.2. Open-Set Domain Adaptation + +Recently, open-set domain adaptation (OSDA) tackles the realistic open-set challenge, where the target domain has additional unknown class space to the source domain. Currently, existing methods [2, 10, 15, 18, 29, 30, 33, 38] transfer known class knowledge while grouping unknown classes as one. For example, OSBP [30] uses pseudo labels to guide classifier learning and to construct constraints for the optimization of the model. Later, to better utilize pseudo-labels, STA [18] proposes an end-to-end method to gradually separate known and unknown classes while aligning feature distributions. OVANET [29] introduces the one-vs-all classifier to learn the inter- and intra-class distance, and minimizes class entropy to distinguish known and unknown classes. OSLPP [38] learns a common subspace from the source and target domains while gradually selecting and rejecting pseudo-labeled target data to promote the model's transfer on known classes. By introducing adversarial learning, UADAL [10] aligns the source- and the target-known distribution while segregating the target-unknown distribution, achieving better knowledge transfer. + +# 2.3. Open-Set Source-Free Domain Adaptation + +When the open-set phenomenon appears, along with limited access to source data, the methods without accessing source data and shared class distribution cannot transfer favorable knowledge, causing them to fail in real-world scenarios. OS-SFDA [6, 11, 13, 17, 22, 26, 37, 42] has emerged to perform knowledge transfer in the open-set scenario with accessing only the source pre-trained model, not the source data. Specifically, the existing methods generally transfer known knowledge and group unknown classes as one. Among these methods, without source data, FS [13] designs the self-adaptive model to transfer the source task-specific knowledge to the target domain effectively, and the target unknown classes are grouped into one class under the open-set setting due to the lack of unknown class information. To tackle OS-SFDA, OSHT [6] adopts pseudo + +labeling for adaptation and the entropy-based metric to reject unknown classes. Distill-SODA [37] introduces novel style-based adversarial data augmentation and the closed-set affinity score for better knowledge distillation from a self-supervised ViT during adaptation. SF-PGL [22] introduces a balanced pseudo-labeling strategy to progressive graph learning for adaptation and distinguishes known and unknown classes by applying the confidence threshold. In addition, the SFDA methods [14, 16, 44, 47] separate known and unknown classes and apply their strategies mentioned in Sec. 2.1 for known knowledge transfer under the open-set setting to tackle the OS-SFDA. + +However, under the OS-SFDA setting, where the target domain has a additional abundant class distribution than the source domain, mentioned OS-SFDA methods group unknown classes as one, which restricts the thorough exploration of target class information, hampering knowledge transfer and generalization. Our work advocates exploring unknown classes can better identify known ones. It should be noted that recently some methods have attempted to explore unknown classes in OSDA [11, 23, 49], which requires source data access to realize domain adaptation. On the contrary, our focused OS-SFDA needs no source data for exploring unknown class space that is more realistic and challenging, essentially differing from them. Concretely, their source data access results in the failure to deal with the most realistic OS-SFDA. Moreover, they utilize 'unknown class number prior' to predefine [11, 23] or enumerate [49] unknown class numbers to explore the unknown class space, whose reliance on unknown prior further limits their scalability in real world. Instead, during unknown class exploration, our method requires no source data and unknown class number prior, truly achieving spontaneous and applicable unknown class exploration in OS-SFDA. We present a domain adaptation network for OS-SFDA, which can fully explore the target class space to transfer knowledge on known classes and generalize on unknown classes. + +# 3. Method + +In this section, we will introduce the details of our proposed method, and the framework is depicted in Fig. 2. + +Problem Setting. OS-SFDA aims to solve the knowledge-transferring problem in more realistic scenarios, where the source data have restricted access, and the target domain exhibits wider class distribution than the source domain. In such a task with restricted-accessible source data, we are given an unlabeled target domain $\mathcal{D}^t = \{(x_i^t)\}_{i=1}^n$ with $n$ target samples, and a pre-trained source model. More importantly, the class space $\mathcal{Y}^t$ of the target domain is broader than the one $\mathcal{Y}^s$ of the source domain: $\mathcal{Y}^t = \mathcal{Y}^s \cup \mathcal{Y}^{ukn}$ . In other words, in addition to the source domain's class space $\mathcal{Y}^s$ , $\mathcal{Y}^t$ also contains an additional unknown class space $\mathcal{Y}^{ukn}$ . Here we call $\mathcal{Y}^s$ and $\mathcal{Y}^{ukn}$ as the target domain's + +known class and unknown class spaces, respectively, whose numbers of categories are denoted as $K_{n}$ and $K_{un}$ . The source domain pre-trained model is obtained by pre-training on the source domain in $\mathcal{Y}^s$ , mainly consisting of feature extractor $f(\cdot)$ and classifier $g(\cdot)$ . Therefore, it is expected to enable the source pre-trained model to overcome the class distribution discrepancy between domains, while achieving further knowledge transfer on the known classes and exploration on the unknown classes. + +# 3.1. Exploration of the Target Class Space + +Inspired by the effectiveness of no-parameter clustering methods [1, 25, 32, 48] in exploring the class space, we introduce an unknown diffuser based on the Multilayer Perceptron (MLP), which includes the main network $C$ and the sub-network $C_k^{sub}$ in the current class space, whose goal is to explore the target class space by inferring the target class number $K$ . Reasonably, we first initialize the target class space with the source known class space; that is, we assume that the initial value of the target class number $K$ is the number of known classes $K_n$ . The unknown diffuser progressively explores a wider class space in the target domain through continuous inferring and optimization of $K$ . Note that the value of $K$ changes dynamically with the inference process during training. The unknown diffuser completes the exploration of the target class space by performing two training processes: optimization in the latest class space and hierarchical class exploration. + +Optimization in Latest Class Space. Since the target domain has no labels during exploration, we need to use the clusters obtained by clustering the target data to guide us in exploring the target class space. Simultaneously, we transform the exploration of the target class space into the exploration of the target cluster number. In this part, we learn the discriminative cluster assignments through the constraints of the feature-cluster assignment, which is beneficial for exploring the target cluster number. Firstly, the feature is the output of feature extractor $f(\cdot)$ in the source pre-trained model, denoted as $z_{i} = f(\pmb{x}_{i}^{t})\in \mathbb{R}^{d}$ , where $d$ is the dimension of the feature space, $i$ represents the $i$ -th sample. The features of $n$ samples are denoted as $Z$ . Then, the main network $C$ in the unknown diffuser maps the feature into the cluster assignment $\widetilde{\boldsymbol{Z}} = c(\boldsymbol {Z})$ , where $c(\cdot)$ is the mapping operation of $C$ , and the cluster assignment of feature $z_{i}$ is denoted as $\widetilde{z}_i\in \mathbb{R}^K$ . The discriminative cluster assignments under the current class space $(K)$ are obtained by optimizing the cluster assignment under the guidance of the discriminative feature $Z$ . Specifically, to obtain the pseudo cluster-assignment $Y^{Z} = \{y_{i}^{Z}\}_{i = 1}^{n}, y_{i}^{Z}\in \mathbb{R}$ , we first use $k$ -means under current target class space $(K)$ on the feature set $Z$ . Then, according to $Y^{Z}$ , a teacher cluster soft-distribution $\widetilde{Y}\in \mathbb{R}^{n\times K}$ is generated based on the feature: + +$$ +\widetilde {y} _ {i, k} = \left\| \boldsymbol {z} _ {i} - \boldsymbol {\eta} _ {k} \right\| _ {\ell_ {2}} ^ {2}, \tag {1} +$$ + +![](images/ce7c259dd2bd690febebc3e065781cf69691d9c52954a6b4ea10168bdab24d9a.jpg) +Figure 2. The framework of our proposed method. Given the source pre-trained model, we first introduce an unknown diffuser on its initial known class space. The unknown diffuser can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider class space, we improve the reliability of the known class knowledge inside the source model by contrastive learning. In this way, the reliable known class knowledge and clustered pseudo-labels over the wider class space, can jointly serve as supervision during model optimization, leading to powerful knowledge transfer and generalization capabilities. + +where $\widetilde{y}_{i,k}$ is the $k$ -th element in $\widetilde{\mathbf{y}}_i$ , and $\widetilde{\mathbf{y}}_i$ is the teacher cluster soft-distribution of feature $\mathbf{z}_i$ . $\eta_k = \sum_{i + 1}^{n_k}z_i^k /n_k$ is the center feature of the $k$ -th center and $n_k$ is the feature number of the $k$ -th master cluster. And $\boldsymbol {z}_i^k$ is the $i$ -th element in $\mathbf{Z}^k\in \mathbb{R}^{n_k\times d}$ , which is the feature matrix of those features with $y_{i}^{Z} = k$ . Note that the teacher cluster soft-distribution of each feature is normalized. Then, we utilize the alignment loss to enforce the consistency of teacher cluster soft-distribution $\widetilde{\mathbf{Y}}$ with the cluster assignment $\widetilde{\mathbf{Z}}$ : + +$$ +\mathcal {L} _ {a l g} = \sum_ {i = 1} ^ {n} L _ {a l g} \left(x _ {i} ^ {t}\right) = \sum_ {i = 1} ^ {n} \widetilde {\boldsymbol {z}} _ {i} \cdot \widetilde {\boldsymbol {y}} _ {i}. \tag {2} +$$ + +The cluster assignment under the current target class space $(K)$ can be promoted through optimization on $c(\cdot)$ in the training process, so as to promote the exploration of the target class space. + +Hierarchical Class Exploration. To fully explore the target class space by inferring the precise target class number $K$ , inspired by [3], the Metropolis-Hastings framework (M-H) is utilized as the underlying framework in the class diffusion process, which computes the Hasting ratio for changing $K$ through splitting and fusion. Besides that, inspired by the hierarchical clustering [39], we design a hierarchical class exploration method to explore the target class space based on the M-H framework, that performs splitting and merging on more dispersive and coupled clusters, respectively. Firstly, we obtain the discriminative cluster assignment after splitting/merging in all possible split- + +ting/merging cases. We then select the potential clusters to be split or merged based on the assignment distance between clusters. In the end, these selected clusters are put into the candidate set to be split or merged, while we use the M-H framework [8] on those clusters in the candidate set to determine the splitting and merging decisions. We assume that the current cluster distributions and those under all possible splitting/merging cases are called the distributions before and after splitting/merging. + +Specifically, we first obtain the discriminative cluster distribution after splitting/merging in all possible splitting/merging cases, in which the cluster before splitting/merging is already learned above. Then, the cluster distribution after merging can be acquired directly by merging the assignments of two clusters. However, obtaining the sub-cluster distribution after splitting is apparently unreasonable, by randomly dividing the target data in the master cluster. For subsequent convenience, we define the clusters before/after splitting as the master cluster and sub-cluster, respectively. To gain the discriminative sub-cluster distribution, we introduce the sub-network $C_k^{sub}$ into the unknown diffuser, and $C_k^{sub}$ contains $K$ sub-clustering layers with the output dimension equaling to 2. One of these sub-network $C_k^{sub}$ is applied on each master cluster to generate the sub-cluster assignment $\widetilde{Z}^{sub,k} = c_k^{sub}(Z^k)$ , in which $c_k^{sub}(\cdot)$ represents the operation of the $k$ -th sub-network applying on the $k$ -th master cluster. Then, to obtain the pseudo-sub-cluster assignment, the $k$ -means first is performed on the + +feature matrix $\mathbf{Z}^k$ of the $k$ -th master cluster, similar to cluster optimization with Eq. (1), Subsequently, a teacher subcluster assignment $\widetilde{\mathbf{Y}}^k \in \mathbb{R}^{n_k \times 2}$ is also generated for the $k$ -th master cluster: + +$$ +\tilde {y} _ {i, j} ^ {k} = \left\| \boldsymbol {z} _ {i} ^ {k} - \boldsymbol {\eta} _ {j} ^ {k} \right\| _ {\ell_ {2}} ^ {2}, \tag {3} +$$ + +where $\widetilde{y}_{i,j}^{k}$ is the $i$ -th row and $j$ -th column element in $\widetilde{\pmb{Y}}^k$ . $\pmb{\eta}_j^k$ is the center feature of the $j$ -th sub-cluster in $k$ -th master cluster. And we utilize the alignment loss to optimize them: + +$$ +\mathcal {L} _ {a l g} ^ {\mathrm {s u b}} = \sum_ {k = 1} ^ {K} \sum_ {i = 1} ^ {n _ {k}} \widetilde {\boldsymbol {z}} _ {i} ^ {\text {s u b}, k} \cdot \widetilde {\boldsymbol {y}} _ {i} ^ {k}, \tag {4} +$$ + +where $\widetilde{z}_i^{sub,k}$ is the $i$ -th element in $\widetilde{\mathbf{Z}}^{sub,k}$ . The training process on the sub-network $C_k^{sub}$ is synchronous with the one on the main network $C$ . + +In addition, hierarchical clustering [39] gradually divides data by merging pairs of clusters with smaller distances or splitting the two most distant sub-clusters in one cluster. According to the assignment distance between clusters, the candidate sets to be split/merged are generated, inspired by the hierarchical clustering. In our method, to explore more different classes, those cluster pairs (sub-cluster pairs) with far assignment distances are considered as those likely to be split. While those with close assignment distances are considered as those likely to be merged due to high similarity between two main clusters. The candidate sets to be split/merged based on this can be written as: + +$$ +C S _ {m e r g} ^ {E D} = \left\{\left(k _ {1}, k _ {2}\right) \mid \in \arg \operatorname {s o r t} _ {n _ {i f}} E D \left(\boldsymbol {\eta} _ {k _ {1}}, \boldsymbol {\eta} _ {k _ {2}}\right) \right\}, \tag {5} +$$ + +$$ +C S _ {s p l} ^ {E D} = \left\{k \mid \in \arg \operatorname {s o r t} _ {n _ {i f}} - E D \left(\boldsymbol {\eta} _ {1} ^ {k}, \boldsymbol {\eta} _ {2} ^ {k}\right) \right\}, \tag {6} +$$ + +where $ED(\cdot, \cdot)$ computes the Euclidean distance of the cluster embedding center. $n_{if}$ are the number of cluster(s) with potential to be merged and split. + +After obtaining the candidate set, we can infer more accurate $K$ and explore a more accurate target class space, by calculating the Hasting ratio from the M-H framework [8] among (sub)cluster pairs in the candidate set and comparing it with 1. More details are provided in the supplementary material. In this way, we use the inferred latest target class number $K$ to update the unknown diffuser (the output dimension of $C$ and the number of sub-clustering layers) at the next optimization step. + +During the optimization in the latest more accurate class space, we learn the discriminative cluster distribution under the latest updated target class number $K$ . According to these cluster distributions, we decide the splitting and merging decisions to change the target class number $K$ in the process of hierarchical class exploration. The more precise $K$ and corresponding cluster assignments are generated, through operating these two processes iteratively, which is beneficial to fully explore the target class space. + +Algorithm 1 The training procedure of our proposed model +Input: data, source pre-trained model, main network $C$ and sub-network $C_k^{sub}$ of unknown diffuser + +1: // Exploration of the target class space +2: for $iter_1 = 1$ to $iter_1^{max}$ do + +3: Obtain clusteredpesudo-distribution $\widetilde{Y}$ and $\widetilde{Y}^k$ by $k$ -means +4: Optimize $C$ and $C_k^{sub}$ guided by $\widetilde{Y}$ and $\widetilde{Y}^k$ in Eqs. (2) and (4) +5: Determine $CS_{slp} / CS_{merg}$ by similarity measurement in Eqs. (5) and (6) +6: Diffuse the class space by updating $K$ on $CS_{slp} / CS_{merg}$ , based on M-H framework +7: Update the unknown diffuser by the updated $K$ +8: end for +9: // Knowledge transfer and generalization +10: for iter2 = 1 to iter2 max do +11: Promote reliability of known knowledge by contrastive constraint in Eq. (7) +12: Jointly optimize $C$ and the pre-trained model in the explored class space via $\mathcal{L}_{kt}$ and $\mathcal{L}_{alg}$ in Eq. (8) +13: end for + +# 3.2. Knowledge Transfer and Generalization + +After exploring the target class space in Sec. 3.1, we aim to transfer knowledge on known classes more effectively and improve the model's generalization on unknown classes. In this subsection, we promote the reliability of known knowledge and further optimize over wider class distribution. + +Promotion of Known Knowledge Reliability. To better utilize known knowledge, we hope to improve the reliability of known knowledge obtained by the classification layer of the pre-trained source model during training. The input of the model is all target samples, including known and unknown samples, from which we need to select the known ones. According to the logit output of known samples having smaller entropy than the unknown ones, we select the top $50\%$ of samples with the lowest entropy as known samples, which are placed in set $\mathcal{M}$ . After selection, the contrastive constraint $\mathcal{L}_{con}$ is utilized on $\mathcal{M}$ to enhance the discrimination of the known samples. Concretely, it brings the known features of similar predictions closer and those of dissimilar predictions farther away, achieving discrimination feature assignment to strengthen the reliability of known knowledge. The contrastive constraint is written as: + +$$ +\mathcal {L} _ {c o n} = - \sum_ {i \in \mathcal {M}} \boldsymbol {p} _ {i} ^ {T} \cdot \boldsymbol {p} _ {i} ^ {+} + \sum_ {i \in \mathcal {M}} \boldsymbol {p} _ {i} ^ {T} \cdot \boldsymbol {p} _ {i} ^ {-}, \tag {7} +$$ + +where $\pmb{p}_i = \sigma(g(\pmb{z}_i)) \in \mathbb{R}^d$ is the output of classifier $g$ and $\sigma$ is the softmax function. + +Optimization over Wider Class Distribution. After + +that, we optimize the unknown diffuser over the wider class space by taking advantage of the reliable known knowledge and alignment constraint on the explored space. In such an optimization based on the explored wider class space, we first select the high-confidence known samples as the reliable known knowledge. Then, we hope to use the hard pseudo-labels of the high-confidence known samples as supervision information, thereby leveraging them for better knowledge transfer on the explored class space. Specifically, for a known sample, we take the maximum logit output by the classifier in the source pre-trained model as its confidence. And the known sample is considered as the high-confidence known sample when it is greater than a manually-set confidence threshold, whose hard pseudolabel is represented by the one-hot paradigm $\pmb{y}_i^h (x_i^h\in \mathcal{H})$ . Note that $\mathcal{H}$ is the set of the high-confidence known samples. Therefore, we can perform better knowledge transfer on the explored class space by leveraging the reliable known knowledge $\pmb{y}_i^h$ obtained on the source pre-trained model. In addition, we also perform alignment constraint on the explored class space, similar to Eq. (2), to complete the transfer on known samples further and achieve generalization on unknown samples. Finally, we leverage the supervision of the high-confidence known sample labels and the alignment constraint, to jointly achieve superior knowledge transfer on known classes and generalization on unknown classes over the wider class space: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {k t} + \mathcal {L} _ {a l g} = \sum_ {i} ^ {\mathcal {H}} L _ {k t} (\widetilde {\boldsymbol {z}} _ {i}, \boldsymbol {y} _ {i} ^ {h}) + \sum_ {i = 1} ^ {n} L _ {a l g} (x _ {i} ^ {t}) \\ = \underbrace {\sum_ {i} ^ {\mathcal {H}} L _ {k t} \left(\widetilde {\boldsymbol {z}} _ {i} , \boldsymbol {y} _ {i} ^ {h}\right) + \sum_ {j _ {1} = 1} ^ {n _ {1}} L _ {a l g} \left(x _ {j _ {1}} ^ {t}\right)} _ {\text {k n o w n}} + \underbrace {\sum_ {j _ {2} = 1} ^ {n _ {2}} L _ {a l g} \left(x _ {j _ {2}} ^ {t}\right)} _ {\text {u n k n o w n}}, \tag {8} \\ \end{array} +$$ + +where $\mathcal{L}_{kt}$ represents the guidance of high-confidence known samples, which is the cross-entropy loss between their pseudo-labels and soft distributions. $\mathcal{L}_{alg}$ is the alignment constraint to achieve superior knowledge transfer on known classes and generalization on unknown classes. $\widetilde{z}_i$ represents the soft-distribution output by the main network $C$ on the explored class space. $n_1$ and $n_2$ represent the number of known and unknown samples, respectively. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on three benchmark datasets for OS-SFDA through experiments, which are Office-31 [28], Office-Home [36], and VisDA-C 2017 [24]. Due to limited space, the results and analysis on VisDA-C 2017 are summarized in supplementary material. + +- Office-31 [28] consists of three domains (Amazon, We + +bcam, DSLR), with 31 classes and 4652 images of common objects. Following the protocols [18] on Office-31 and Office-Home, we select the first 20 classes as the shared classes of the source and target domain while the remaining 11 classes are considered unknown. + +- Office-Home [36] is a more challenging dataset. It contains four domains (Art, Product, Clipart, RealWorld) with 65 classes and 15500 images of common objects. Like Office-31, we select the first 25 classes as classes shared by the source and target domain, and the remaining 40 classes are considered unknown. + +Baselines. For OS-SFDA, we compare our method with the clustering on Resnet [9], the classic method SHOT [16], and the advanced method AaD [44], which do sufficient experiments under the OS-SFDA setting. In addition, our method uses the unknown diffuser based on clustering (class number-agnostic) to explore the target class space. For a fair comparison, we thoroughly compare our method with the no-parameter clustering methods (FINCH [32], COMIC [25], DenMune [1]). + +Evaluation. As shown in Tab. 1 and Tab. 2, we report domain adaptation results for $OS^{*}$ and $UNK$ , which represent the average accuracy of the known and the unknown class, respectively. Simultaneously, $HOS = \frac{2 \times OS^{*} \times UNK}{OS^{*} + UNK}$ is denoted as the harmonic mean between $OS^{*}$ and $UNK$ , which is the key indicator for evaluating performance. Moreover, the inferred target class number $K$ and accuracy ACC are used to evaluate the effectiveness of exploring the target class space in Tab. 3 and Tab. 4. The setting of ACC follows the previous protocols [12]. + +Implementation Details. For the source pre-trained model, we utilize the backbone of ResNet-50 [9] for Office-31 and Office-Home. For a fair comparison, we use the same network architecture with SHOT [16] and AaD [44], which includes a backbone and two additional fully connected layers. We adopt SGD with a momentum of 0.9 and a batch size of 64 for all datasets. The learning rate is set to 1e-3 on all datasets for all layers of the backbone, and the last two added fully-connected layers are applied 1e-2. We train 15 epochs for all datasets. The unknown diffuser includes one convolutional layer with 256 hidden dimensions. For a fair comparison, our network and the compared no-parameter clustering methods use the same input features, output by the source pre-trained model. + +# 4.2. Results and Analysis + +As shown in Tab. 1 and Tab. 2, the UNK and HOS for our method are much higher than SHOT on all datasets. Our method achieves $10.4\%$ and $2.3\%$ relative improvements over the state-of-art AaD for average OS* and HOS on Office-31, respectively. On the relatively small-scale Office-31 with fewer classes, our method demonstrates competitiveness across various indicators, standing in con + +
methodSFA→DA→WD→AD→WW→AW→DAvg
OS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOS
Cluster✗10.064.017.07.061.013.05.069.010.08.049.014.05.064.010.010.047.017.07.559.013.5
SHOT✓87.250.055.789.038.653.878.244.256.597.472.383.073.547.657.889.675.081.785.854.664.8
AaD✓69.986.777.462.083.971.360.881.869.883.297.489.858.481.568.082.395.288.369.487.877.4
Ours(w/o s)✓78.473.475.887.990.389.172.376.174.192.880.586.270.976.873.785.879.882.781.479.580.3
Ours(w/s)✓82.371.876.790.072.080.072.079.475.579.587.383.274.282.778.280.688.384.379.880.379.7
+ +Table 1. Accuracy (%) on Office-31 for OS-SFDA. The three indicators(OS*, UNK, HOS) have been explained in Sec. 4.1. ‘w/o s’ and ‘w/s’ indicate with or without supervision using reliable known knowledge. Note that ‘A→D’ represents that the model per-trained on the source domain Amazon transfers to the target domain DSLR. Bold and ‘_’ represent the best and second results, respectively. + +
methodSFAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwAvg (first 6 tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX6.063.011.05.055.010.04.062.08.08.059.014.05.055.09.05.056.08.05.056.08.05.558.310.0
SHOT✓67.028.039.581.826.339.887.532.147.066.846.254.677.527.240.280.025.939.176.831.043.4
AaD✓50.567.457.764.066.465.172.269.570.847.180.359.364.768.266.465.071.067.860.670.564.5
Ours(w/o s)✓47.669.256.470.081.875.470.685.677.460.374.766.765.682.773.272.184.277.664.479.771.1
Ours(w/s)✓49.064.855.868.188.076.771.586.678.461.372.666.466.581.673.171.684.877.664.779.771.3
methodSFPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrRw→PrAvg (all tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX8.053.015.06.050.010.05.061.09.08.051.013.05.064.010.05.050.010.05.856.610.6
SHOT✓66.351.157.759.331.040.885.831.646.273.550.659.965.328.940.184.428.242.374.633.945.6
AaD✓46.983.160.045.072.655.669.072.370.656.077.465.048.367.656.467.769.368.558.072.163.6
Ours(w/o s)✓60.076.767.352.666.258.670.785.177.256.583.367.449.170.757.966.182.173.261.878.569.0
Ours(w/s)✓55.985.667.645.470.255.173.983.978.656.784.167.849.374.659.469.580.074.461.679.769.2
+ +Table 2. Accuracy (%) on Office-Home for OS-SFDA. Note that ‘Ar→Cl’ represents that the model per-trained on the source domain Art transfers to the target domain Clipart. The representations of OS*, UNK, HOS, ‘w/o s’, and ‘w/s’ are the same with Tab. 1. + +trast to methods such as AaD and SHOT, where only one indicator exhibits competitive performance. SHOT tends to excel in OS*, while AaD typically performs well on UNK. In contrast, our method displays a well-balanced performance across different measurement dimensions, achieving competitive results overall. For the more challenging Office-Home, our method displays more apparent superiority, which achieves $61.6\%$ , $79.7\%$ , and $69.2\%$ average accuracy for OS*, UNK, and HOS, with $3.6\%$ , $7.6\%$ , and $5.6\%$ relative improvement compared with AaD. Furthermore, our method outperforms all no-parameter clustering methods (FINCH, DenMune, and COMIC) for ACC and $K$ on all tasks in Tab. 3 and Tab. 4. The above results demonstrate the superiority and effectiveness of our method. + +Impact of Exploration for the Target Class Space. Compared with OS-SFDA methods (SHOT and AaD) that group all target unknown classes as one, our method tends to explore class space and excavate different unknown classes, which has great application value in more realistic scenarios. The experimental results show that our method achieves excellent improvement over the state-of-the-art method AaD for the known accuracy $\mathrm{OS^{*}}$ on the more challenging Office Home. In addition, the unknown accuracy UNK for our method is much higher than SHOT + +on Office-31 and Aad on Office-Home. This phenomenon shows that our method can improve knowledge transfer on known classes when exploring class space and simultaneously improve generalization on unknown classes. To more fully demonstrate the effectiveness and advancement of our exploration, we compare our method with classical no-parameter clustering methods for exploring the class space. Concretely, ours provides the more precise inferred target class number $K$ and higher accuracy ACC on all tasks, where the precision of $K$ represents the efficacy in exploring the target class space, as discussed in Sec. 3.1. The above analysis conclusively demonstrates that our method of exploring class space is exceptionally effective, improving knowledge transfer and generalization. + +Effect of Reliable Known Knowledge as Supervision. As introduced in Sec. 3.2, we hope to utilize the reliable known knowledge to optimize the unknown diffuser for better knowledge transfer and generalization. Concretely, our method that uses the reliable known knowledge(ours (w/ s)) outperforms all compared methods(including ours(w/o s)) on more challenging Office-Home for average HOS, ACC, and $K$ in Tabs. 1 to 4. Therefore, the more effectiveness brought by reliable known knowledge ('ours (w/ s)') confirms it guides the model for better known and unknown + +![](images/1ab423e646ede09859f9823eb36263394f2465b816b76b89757dc15bfdc88c1d.jpg) +(a) + +![](images/af9dd5e01ac10a377c37e314bee6102371cc81e27656a37f85322e057d5c943a.jpg) +(b) +Figure 3. (a) t-SNE of SHOT on $\mathbb{R}\to \mathbb{P}$ . (b) t-SNE of AaD on $\mathbb{R}\to \mathbb{P}$ . (c) t-SNE of our method on $\mathbb{R}\to \mathbb{P}$ . (d) Parameter analysis on $\mathrm{A}\rightarrow \mathrm{W}$ . + +![](images/8ff444d523cf7f5fe0392fbc3fea6ad95c9e56df8187c93cff1f15e2d64751df.jpg) +(c) + +![](images/be0297374d1be915c72b94367dec59dc52cfe1f74e5f50696723cba779430542.jpg) +(d) + +
MethodsA→DA→WD→AD→WW→AW→D
FINCH31.0 (8)8.0 (2)23.0 (7)12.0 (3)28.0 (9)23.0 (5)
DenMune20.0(5)52.0(15)58.0(29)46.0(14)65.0(35)22.0(5)
COMIC56.8 (22)47.3 (24)16.1 (40)46.4 (20)12.1 (32)56.6 (22)
Ours(w/o s)69.0(24)80.6(28)60.4(37)82.5(30)59.6(38)68.9(27)
Ours(w/ s)71.2(25)73.3(28)63.1(37)76.3(32)64.2(38)74.5(28)
+ +Table 3. Accuracy ACC(%) and the inferred target class number $K$ on Office-31. Note that the values before $\left( \cdot \right)$ ’ and inside $\left( \cdot \right)$ ’ represent the performance ACC and $K$ ,respectively. + +
MethodsAr→ClAr→PrAr→RwCl→ArCl→PrCl→Rw
FINCH5.0(4)7.0(3)6.0(3)8.0(3)7.0(4)7.0(4)
DenMune15.0(12)58.0(45)54.0(39)16.0(7)55.0(40)40.0(32)
COMIC6.9(65)50.3(92)9.43(34)4.0(1)10.4(51)2.3(2)
Ours(w/o s)35.5(69)60.5(77)59.1(77)47.3(61)58.9(74)55.6(60)
Ours(w/s)35.9(69)61.4(80)61.3(77)46.9(62)58.9(78)61.0(60)
MethodsPr→ArPr→ClPr→RwRw→ArRw→PrRw→Cl
FINCH12.0(4)5.0(4)4.0(2)11.0(4)12.0(7)7.0(4)
DenMune16.0(7)7.0(6)52.0(37)26.0(13)9.0(8)62.0(43)
COMIC4.2(2)3.7(24)5.7(14)4.1(1)5.2(43)13.3(65)
Ours(w/o s)48.5(51)37.2(56)63.6(75)38.1(28)21.6(50)50.4(75)
Ours(w/s)47.8(51)35.5(56)64.0(68)44.2(35)21.9(50)51.8(75)
+ +Table 4. ACC(%) and $K$ on Office-Home. + +class discrimination, achieving better knowledge transfer on known classes and generalization on unknown classes. + +Parameter Analysis. We study the impact of supervision information on unknown diffuser optimization during training by applying a variable parameter to the supervision loss. As shown in Fig. 3(d), the chart indicates that when there is no supervision information, the optimization cannot be completed well to obtain a favorable transfer effect on known or unknown classes. As the parameters increase within a certain range, UNK and HOS also improve; but parameters that are too large will lead to poor optimization results. In addition, it can be seen that OS* is not sensitive to this parameter, and its change trend is relatively stable. Therefore, we should select appropriate high + +confidence known pseudo-labels as supervision information to prevent model overfitting for better guiding the optimization of unknown diffuser. + +Effectiveness of the Unknown Diffuser. Tabs. 1 to 4 show our superior performance in exploring unknown class space and domain adaptation in OS-SFDA. The superiority in unknown class exploration indicates that our unknown diffuser can excavate potential unknown classes under the guidance of current discriminative class distributions based on the Metropolis-Hastings (M-H) framework. The impressive domain adaptation effects based on the explored wider space also demonstrate the necessity of using the unknown diffuser for unknown class exploration. + +Visualization of features. As shown in Fig. 3(a)-(c), t-SNE visualizations indicate that compared with existing methods, our method explores a more accurate and extensive class space, exhibiting remarkable class discrimination. + +# 5. Conclusion + +In this paper, we propose a domain adaptation network in OS-SFDA for exploring unknown classes, thereby jointly benefiting known knowledge transfer and unknown generalize. We introduce an unknown diffuser into the source pretrained model to excavate a wider target class space. And, the reliable known class knowledge and clustered pseudolabels over the wider class space are captured to serve as supervision during model optimization, realizing impressive knowledge transfer and generalization. Extensive experiments verify the superiority of our method. + +# 6. Acknowledgements + +This work is supported in part by the National Key Research and Development Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62132016, 62171343, and 62201436), and Fundamental Research Funds for the Central Universities (ZDRC2102). + +# References + +[1] Mohamed Abbas, Adel El-Zoghabi, and Amin Shoukry. Denmune: Density peak based clustering using mutual nearest neighbors. Pattern Recognit., 109:107589, 2021. 3, 6 +[2] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 422-438. Springer, 2020. 2 +[3] Jason Chang and John W Fisher III. Parallel sampling of dp mixture models using sub-cluster splits. Proc. Adv. Neural Inf. Process. Syst., 26, 2013. 4 +[4] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7212-7222, 2022. 2 +[5] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 1 +[6] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE Trans. Image Process., 30:6473-6484, 2021. 1, 2 +[7] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. J. Mach. Learn. Res., 17(1):2096-2030, 2016. 1 +[8] W Keith Hastings. Monte carlo sampling methods using markov chains and their applications. 1970. 4, 5 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6 +[10] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35:16755-16767, 2022. 2 +[11] Taotao Jing, Hongfu Liu, and Zhengming Ding. Towards novel target discovery through open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9322-9331, 2021. 2, 3 +[12] Harold W Kuhn. The hungarian method for the assignment problem. Nav. Res. Logist. Q., 2(1-2):83-97, 1955. 6 +[13] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12376-12385, 2020. 2 +[14] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 12365-12377. PMLR, 2022. 2, 3 +[15] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 24110-24119, 2023. 2 +[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for + +unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 6028-6039. PMLR, 2020. 1, 2, 3, 6 +[17] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 2 +[18] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2927-2936, 2019. 2, 6 +[19] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Proc. Adv. Neural Inf. Process. Syst., 29, 2016. 1 +[20] Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(12):3071-3085, 2018. 1 +[21] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 31, 2018. 1 +[22] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 2, 3 +[23] Yingwei Pan, Ting Yao, Yehao Li, Chong-Wah Ngo, and Tao Mei. Exploring category-agnostic clusters for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 13867–13875, 2020. 3 +[24] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017. 6 +[25] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In Proc. Int. Conf. Mach. Learn., pages 5092-5101. PMLR, 2019. 3, 6 +[26] Masoud Pourreza. Open-set source-free domain adaptation in fundus images analysis. 2023. 2 +[27] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 537-555. Springer, 2022. 2 +[28] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In Proc. Eur. Conf. Comput. Vis., pages 213-226. Springer, 2010. 6 +[29] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9000-9009, 2021. 2 +[30] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proc. Eur. Conf. Comput. Vis., pages 153-168, 2018. 2 +[31] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. Proc. Adv. Neural Inf. Process. Syst., 33:16282-16292, 2020. 1 + +[32] Saquib Sarfraz, Vivek Sharma, and Rainer Stiefelhagen. Efficient parameter-free clustering using first neighbor relations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8934-8943, 2019. 3, 6 +[33] Tasfia Shermin, Guojun Lu, Shyh Wei Teng, Manzur Murshed, and Ferdous Sohel. Adversarial network with multiple classifiers for open set domain adaptation. IEEE Trans. Multimed., 23:2732-2744, 2020. 2 +[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proc. AAAI Conf. Artif. Intell., 2016. 1 +[35] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darryll. Adversarial discriminative domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7167-7176, 2017. 1 +[36] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 5018-5027, 2017. 6 +[37] Guillaume Vray, Devavrat Tomar, Behzad Bozorgtabar, and Jean-Philippe Thiran. Source-free open-set domain adaptation for histopathological images via distilling self-supervised vision transformer. arXiv preprint arXiv:2307.04596, 2023. 1, 2, 3 +[38] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2 +[39] Joe H Ward Jr. Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc., 58(301):236-244, 1963. 4, 5 +[40] Kun Wei, Xu Yang, Zhe Xu, and Cheng Deng. Class incremental unsupervised domain adaptation via pseudolabel distillation. IEEE Trans. Image Process., 2024. 1 +[41] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9010-9019, 2021. 2 +[42] Jiexi Yan, Zhihui Yin, Erkun Yang, Yanhua Yang, and Heng Huang. Learning with diversity: Self-expanded equalization for better generalized deep metric learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 19365-19374, 2023. 2 +[43] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8978-8987, 2021. 2 +[44] Shiqi Yang, Shangling Jui, Joost van de Weijer, et al. Attracting and dispersing: A simple approach for source-free domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35: 5802-5815, 2022. 2, 3, 6 +[45] Shiqi Yang, Yaxing Wang, Luis Herranz, Shangling Jui, and Joost van de Weijer. Casting a bait for offline and online source-free domain adaptation. Comput. Vis. Image Underst., page 103747, 2023. 2 +[46] Xu Yang, Cheng Deng, Tongliang Liu, and Dacheng Tao. Heterogeneous graph attention network for unsupervised multiple-target domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1992-2003, 2020. 1 + +[47] Yixin Zhang, Zilei Wang, and Weinan He. Class relationship embedded learning for source-free unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7619-7629, 2023. 2, 3 +[48] Han Zhao, Xu Yang, and Cheng Deng. Parameter-agnostic deep graph clustering. ACM Trans. Knowl. Discov. Data, 18 (3):1–20, 2024. 3 +[49] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Open set domain adaptation by novel class discovery. arXiv preprint arXiv:2203.03329, 2022. 3 \ No newline at end of file diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/images.zip b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48f74e92985416408cd737c2d0411af0d677ae62 --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63217012514ba1ddeae72a9a6deb35cf19602b8465fcdba1a12646f615e53f5f +size 618128 diff --git a/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/layout.json b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc62bf51cc898f0540b879514df9380d2d2a127 --- /dev/null +++ b/2024/Unveiling the Unknown_ Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation/layout.json @@ -0,0 +1,9482 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 514, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 514, + 140 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 514, + 140 + ], + "type": "text", + "content": "Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 160, + 476, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 160, + 476, + 204 + ], + "spans": [ + { + "bbox": [ + 117, + 160, + 476, + 204 + ], + "type": "text", + "content": "Fuli Wan*, Han Zhao*, Xu Yang†, and Cheng Deng† \nSchool of Electronic Engineering, Xidian University, Xi'an710071, China \n{wfl128.xidian, hzhao1698, xuyang.xd, chdeng.xd}@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 568 + ], + "type": "text", + "content": "Open-Set Source-Free Domain Adaptation aims to transfer knowledge in realistic scenarios where the target domain has additional unknown classes compared to the limited-access source domain. Due to the absence of information on unknown classes, existing methods mainly transfer knowledge of known classes while roughly grouping unknown classes as one, attenuating the knowledge transfer and generalization. In contrast, this paper advocates that exploring unknown classes can better identify known ones, and proposes a domain adaptation model to transfer knowledge on known and unknown classes jointly. Specifically, given a source pre-trained model, we first introduce an unknown diffuser that can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider space distribution, we enhance the reliability of known class knowledge in the source pre-trained model through contrastive constraint. Finally, various supervision information, including reliable known class knowledge and clustered pseudo-labels, optimize the model for impressive knowledge transfer and generalization. Extensive experiments show that our network can achieve superior exploration and knowledge generalization on unknown classes, while with excellent known class transfer. The code is available at https://github.com/xdwfl/UPUK." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "content": "By transferring knowledge from label-rich source domains to unlabeled target domains, unsupervised domain adaptation (UDA) has exhibited huge potential to address the challenges of costly data labeling and domain distribution shifts in real-world scenarios. Existing methods mainly learn the domain invariant representation by distribution matching" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 232, + 547, + 387 + ], + "blocks": [ + { + "bbox": [ + 310, + 232, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 310, + 232, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 310, + 232, + 547, + 387 + ], + "type": "image", + "image_path": "7258fd0de622d507a6a6acaa4714813d3feb5bf1a3e67c9e1a3389df1e1bb8ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Comparison of existing OS-SFDA methods and our proposed method. Existing methods ignore the unknown classes by grouping unknown classes as one, while our method can further explore the unknown class information." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "type": "text", + "content": "[19, 20, 34, 40, 46] and adversarial optimization [7, 21, 35], to mitigate domain distribution shift. They can achieve impressive transfer effects by following the close-set and real-time data-accessible assumption, which shares the same class space in source and target domains, and accesses the source data at any time. However, in the real world, the open-set phenomenon, where the target domain has additional unknown class space to the source domain, always appears along with limited access to source data [5, 16]. In such most realistic scenarios, these methods fail to transfer knowledge without accessing source data and shared class distribution, making the knowledge transfer in such scenarios challenging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Recently, to tackle such a challenging and realistic task, the open-set source-free domain adaptation (OS-SFDA) has emerged to perform knowledge transfer in the open-set scenario by accessing only the source pre-trained model, not the source data. Considering the lack of information on target unknown classes, existing approaches [6, 22, 31, 37] focus on transferring knowledge of known classes through pseudo-supervision, while grouping unknown classes as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 151, + 712 + ], + "type": "text", + "content": "†The corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24015" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "one, as seen Fig. 1. However, such a grouping way restricts the thorough exploration of target class information, thereby diminishing models' transfer and generalization abilities on known and unknown classes, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 310 + ], + "type": "text", + "content": "To this end, we present a domain adaptation network for OS-SFDA, to explore target class space and transfer knowledge on known and unknown classes jointly. Specifically, given the source pre-trained model, an unknown diffuser is introduced to determine to-be-split/merged classes through similarity measures, generating a wider class space distribution, including known and unknown classes. Based on a wider distribution in the feature space, the reliability of the known class knowledge from the source pre-trained model is enhanced through contrastive constraint. Finally, the reliable known class knowledge, along with the clustered pseudo-labels spanning the extensive class space, can collaboratively serve as supervision during model optimization. This collaborative approach results in notable knowledge transfer and enhanced generalization capabilities. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 315, + 286, + 493 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 315, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 286, + 361 + ], + "type": "text", + "content": "- We propose a domain adaptation network that is the first to excavate the unknown class space for boosting knowledge transfer and generalization to the target domain in OS-SFDA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 363, + 286, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 286, + 409 + ], + "type": "text", + "content": "- We design an unknown diffuser to explore wider and more precise target class space in the target domain, further benefiting known knowledge transfer and unknown generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 411, + 286, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 286, + 457 + ], + "type": "text", + "content": "- To realize effective known knowledge transfer and unknown generalization, we utilize the reliable known knowledge and clustered pseudo-labels in the broader class space, to serve as supervision during optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 459, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 286, + 493 + ], + "type": "text", + "content": "- Experiments can demonstrate that our method can explore target class space precisely, and yield state-of-the-art results among multiple OS-SFDA benchmarks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 506, + 133, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 133, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 133, + 518 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 527, + 221, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 221, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 221, + 539 + ], + "type": "text", + "content": "2.1. Source-Free Domain Adaptation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 712 + ], + "type": "text", + "content": "In recent years, with accessing only the source pre-trained model for transferring knowledge to the target domain, source-free domain adaptation (SFDA) methods [4, 16, 27, 41, 43-45, 47] have been widely used for overcoming source data privacy to achieve knowledge transfer in real-world scenarios. The existing methods are mainly divided into two categories, including methods based on pseudo-labeling [14, 16] and based on neighborhood clustering [43, 44]. For the former approaches, the classical method - SHOT [16] freezes the source classifier and optimizes the feature extractor, through mutual information maximization and source-target feature alignment based on self-supervised pseudo-labeling. To fully utilize the source and target domain knowledge, CoWA-JWDS [14] intro" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "duces a new sample confidence score – JMDS better to exert the importance of pseudo-label confidence in distinguishing samples. For the latter approaches, G-SFDA [43] proposes a method of clustering target features and semantically similar neighborhoods to make the model better adapt to the target domain. AaD [44] has introduced prediction consistency of local neighborhood features and discrete prediction of different potential features, which achieve efficient feature clustering and distribution to better transfer knowledge. In addition, there are also some approaches to tackling SFDA from other aspects. U-SFAN [27] uses predicted quantified uncertainty to guide target adaptation. CRS [47] embeds class relationship similarity to transfer domain-invariant class relationships." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 248, + 468, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 468, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 468, + 261 + ], + "type": "text", + "content": "2.2. Open-Set Domain Adaptation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 266, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 506 + ], + "type": "text", + "content": "Recently, open-set domain adaptation (OSDA) tackles the realistic open-set challenge, where the target domain has additional unknown class space to the source domain. Currently, existing methods [2, 10, 15, 18, 29, 30, 33, 38] transfer known class knowledge while grouping unknown classes as one. For example, OSBP [30] uses pseudo labels to guide classifier learning and to construct constraints for the optimization of the model. Later, to better utilize pseudo-labels, STA [18] proposes an end-to-end method to gradually separate known and unknown classes while aligning feature distributions. OVANET [29] introduces the one-vs-all classifier to learn the inter- and intra-class distance, and minimizes class entropy to distinguish known and unknown classes. OSLPP [38] learns a common subspace from the source and target domains while gradually selecting and rejecting pseudo-labeled target data to promote the model's transfer on known classes. By introducing adversarial learning, UADAL [10] aligns the source- and the target-known distribution while segregating the target-unknown distribution, achieving better knowledge transfer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 515, + 527, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 527, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 527, + 528 + ], + "type": "text", + "content": "2.3. Open-Set Source-Free Domain Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "When the open-set phenomenon appears, along with limited access to source data, the methods without accessing source data and shared class distribution cannot transfer favorable knowledge, causing them to fail in real-world scenarios. OS-SFDA [6, 11, 13, 17, 22, 26, 37, 42] has emerged to perform knowledge transfer in the open-set scenario with accessing only the source pre-trained model, not the source data. Specifically, the existing methods generally transfer known knowledge and group unknown classes as one. Among these methods, without source data, FS [13] designs the self-adaptive model to transfer the source task-specific knowledge to the target domain effectively, and the target unknown classes are grouped into one class under the open-set setting due to the lack of unknown class information. To tackle OS-SFDA, OSHT [6] adopts pseudo" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24016" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "labeling for adaptation and the entropy-based metric to reject unknown classes. Distill-SODA [37] introduces novel style-based adversarial data augmentation and the closed-set affinity score for better knowledge distillation from a self-supervised ViT during adaptation. SF-PGL [22] introduces a balanced pseudo-labeling strategy to progressive graph learning for adaptation and distinguishes known and unknown classes by applying the confidence threshold. In addition, the SFDA methods [14, 16, 44, 47] separate known and unknown classes and apply their strategies mentioned in Sec. 2.1 for known knowledge transfer under the open-set setting to tackle the OS-SFDA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 515 + ], + "type": "text", + "content": "However, under the OS-SFDA setting, where the target domain has a additional abundant class distribution than the source domain, mentioned OS-SFDA methods group unknown classes as one, which restricts the thorough exploration of target class information, hampering knowledge transfer and generalization. Our work advocates exploring unknown classes can better identify known ones. It should be noted that recently some methods have attempted to explore unknown classes in OSDA [11, 23, 49], which requires source data access to realize domain adaptation. On the contrary, our focused OS-SFDA needs no source data for exploring unknown class space that is more realistic and challenging, essentially differing from them. Concretely, their source data access results in the failure to deal with the most realistic OS-SFDA. Moreover, they utilize 'unknown class number prior' to predefine [11, 23] or enumerate [49] unknown class numbers to explore the unknown class space, whose reliance on unknown prior further limits their scalability in real world. Instead, during unknown class exploration, our method requires no source data and unknown class number prior, truly achieving spontaneous and applicable unknown class exploration in OS-SFDA. We present a domain adaptation network for OS-SFDA, which can fully explore the target class space to transfer knowledge on known classes and generalize on unknown classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 525, + 104, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 104, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 104, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 569 + ], + "type": "text", + "content": "In this section, we will introduce the details of our proposed method, and the framework is depicted in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Problem Setting. OS-SFDA aims to solve the knowledge-transferring problem in more realistic scenarios, where the source data have restricted access, and the target domain exhibits wider class distribution than the source domain. In such a task with restricted-accessible source data, we are given an unlabeled target domain " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t = \\{(x_i^t)\\}_{i=1}^n" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " target samples, and a pre-trained source model. More importantly, the class space " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " of the target domain is broader than the one " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " of the source domain: " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t = \\mathcal{Y}^s \\cup \\mathcal{Y}^{ukn}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ". In other words, in addition to the source domain's class space " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " also contains an additional unknown class space " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{ukn}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ". Here we call " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{ukn}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " as the target domain's" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "known class and unknown class spaces, respectively, whose numbers of categories are denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "K_{n}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "K_{un}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ". The source domain pre-trained model is obtained by pre-training on the source domain in " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^s" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ", mainly consisting of feature extractor " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ". Therefore, it is expected to enable the source pre-trained model to overcome the class distribution discrepancy between domains, while achieving further knowledge transfer on the known classes and exploration on the unknown classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 186, + 506, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 506, + 199 + ], + "type": "text", + "content": "3.1. Exploration of the Target Class Space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": "Inspired by the effectiveness of no-parameter clustering methods [1, 25, 32, 48] in exploring the class space, we introduce an unknown diffuser based on the Multilayer Perceptron (MLP), which includes the main network " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": " and the sub-network " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": " in the current class space, whose goal is to explore the target class space by inferring the target class number " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": ". Reasonably, we first initialize the target class space with the source known class space; that is, we assume that the initial value of the target class number " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": " is the number of known classes " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "K_n" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": ". The unknown diffuser progressively explores a wider class space in the target domain through continuous inferring and optimization of " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": ". Note that the value of " + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 204, + 547, + 407 + ], + "type": "text", + "content": " changes dynamically with the inference process during training. The unknown diffuser completes the exploration of the target class space by performing two training processes: optimization in the latest class space and hierarchical class exploration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": "Optimization in Latest Class Space. Since the target domain has no labels during exploration, we need to use the clusters obtained by clustering the target data to guide us in exploring the target class space. Simultaneously, we transform the exploration of the target class space into the exploration of the target cluster number. In this part, we learn the discriminative cluster assignments through the constraints of the feature-cluster assignment, which is beneficial for exploring the target cluster number. Firstly, the feature is the output of feature extractor " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " in the source pre-trained model, denoted as " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "z_{i} = f(\\pmb{x}_{i}^{t})\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " is the dimension of the feature space, " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": "-th sample. The features of " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " samples are denoted as " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ". Then, the main network " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " in the unknown diffuser maps the feature into the cluster assignment " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{Z}} = c(\\boldsymbol {Z})" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "c(\\cdot)" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " is the mapping operation of " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ", and the cluster assignment of feature " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "\\widetilde{z}_i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ". The discriminative cluster assignments under the current class space " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " are obtained by optimizing the cluster assignment under the guidance of the discriminative feature " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ". Specifically, to obtain the pseudo cluster-assignment " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "Y^{Z} = \\{y_{i}^{Z}\\}_{i = 1}^{n}, y_{i}^{Z}\\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ", we first use " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": "-means under current target class space " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " on the feature set " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ". Then, according to " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "Y^{Z}" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": ", a teacher cluster soft-distribution " + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}\\in \\mathbb{R}^{n\\times K}" + }, + { + "bbox": [ + 304, + 408, + 548, + 696 + ], + "type": "text", + "content": " is generated based on the feature:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 699, + 547, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 699, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 381, + 699, + 547, + 715 + ], + "type": "interline_equation", + "content": "\\widetilde {y} _ {i, k} = \\left\\| \\boldsymbol {z} _ {i} - \\boldsymbol {\\eta} _ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {1}", + "image_path": "449a84dd92cf4c73c72e5919da9b368bfff056e2ae8f423ba571b2fa8bef84ae.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24017" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 534, + 270 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 534, + 270 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 534, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 534, + 270 + ], + "type": "image", + "image_path": "ce7c259dd2bd690febebc3e065781cf69691d9c52954a6b4ea10168bdab24d9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 346 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed method. Given the source pre-trained model, we first introduce an unknown diffuser on its initial known class space. The unknown diffuser can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider class space, we improve the reliability of the known class knowledge inside the source model by contrastive learning. In this way, the reliable known class knowledge and clustered pseudo-labels over the wider class space, can jointly serve as supervision during model optimization, leading to powerful knowledge transfer and generalization capabilities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\widetilde{y}_{i,k}" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": "-th element in " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{y}}_i" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{y}}_i" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " is the teacher cluster soft-distribution of feature " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\eta_k = \\sum_{i + 1}^{n_k}z_i^k /n_k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " is the center feature of the " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": "-th center and " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " is the feature number of the " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": "-th master cluster. And " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\boldsymbol {z}_i^k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": "-th element in " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^k\\in \\mathbb{R}^{n_k\\times d}" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": ", which is the feature matrix of those features with " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "y_{i}^{Z} = k" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": ". Note that the teacher cluster soft-distribution of each feature is normalized. Then, we utilize the alignment loss to enforce the consistency of teacher cluster soft-distribution " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": " with the cluster assignment " + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 366, + 289, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 483, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 483, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 94, + 483, + 287, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l g} = \\sum_ {i = 1} ^ {n} L _ {a l g} \\left(x _ {i} ^ {t}\\right) = \\sum_ {i = 1} ^ {n} \\widetilde {\\boldsymbol {z}} _ {i} \\cdot \\widetilde {\\boldsymbol {y}} _ {i}. \\tag {2}", + "image_path": "ef7d14e53aacad531b6ab420be610b1edc427a2a59599b0b7ca314987364b7eb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "text", + "content": "The cluster assignment under the current target class space " + }, + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "text", + "content": " can be promoted through optimization on " + }, + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "inline_equation", + "content": "c(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 287, + 570 + ], + "type": "text", + "content": " in the training process, so as to promote the exploration of the target class space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Hierarchical Class Exploration. To fully explore the target class space by inferring the precise target class number " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", inspired by [3], the Metropolis-Hastings framework (M-H) is utilized as the underlying framework in the class diffusion process, which computes the Hasting ratio for changing " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " through splitting and fusion. Besides that, inspired by the hierarchical clustering [39], we design a hierarchical class exploration method to explore the target class space based on the M-H framework, that performs splitting and merging on more dispersive and coupled clusters, respectively. Firstly, we obtain the discriminative cluster assignment after splitting/merging in all possible split-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "type": "text", + "content": "ting/merging cases. We then select the potential clusters to be split or merged based on the assignment distance between clusters. In the end, these selected clusters are put into the candidate set to be split or merged, while we use the M-H framework [8] on those clusters in the candidate set to determine the splitting and merging decisions. We assume that the current cluster distributions and those under all possible splitting/merging cases are called the distributions before and after splitting/merging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": "Specifically, we first obtain the discriminative cluster distribution after splitting/merging in all possible splitting/merging cases, in which the cluster before splitting/merging is already learned above. Then, the cluster distribution after merging can be acquired directly by merging the assignments of two clusters. However, obtaining the sub-cluster distribution after splitting is apparently unreasonable, by randomly dividing the target data in the master cluster. For subsequent convenience, we define the clusters before/after splitting as the master cluster and sub-cluster, respectively. To gain the discriminative sub-cluster distribution, we introduce the sub-network " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": " into the unknown diffuser, and " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": " sub-clustering layers with the output dimension equaling to 2. One of these sub-network " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": " is applied on each master cluster to generate the sub-cluster assignment " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}^{sub,k} = c_k^{sub}(Z^k)" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "c_k^{sub}(\\cdot)" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": " represents the operation of the " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": "-th sub-network applying on the " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": "-th master cluster. Then, to obtain the pseudo-sub-cluster assignment, the " + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 482, + 546, + 713 + ], + "type": "text", + "content": "-means first is performed on the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24018" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "content": "feature matrix " + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^k" + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "content": "-th master cluster, similar to cluster optimization with Eq. (1), Subsequently, a teacher subcluster assignment " + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{Y}}^k \\in \\mathbb{R}^{n_k \\times 2}" + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "content": " is also generated for the " + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 71, + 287, + 123 + ], + "type": "text", + "content": "-th master cluster:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 126, + 287, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 126, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 122, + 126, + 287, + 146 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {i, j} ^ {k} = \\left\\| \\boldsymbol {z} _ {i} ^ {k} - \\boldsymbol {\\eta} _ {j} ^ {k} \\right\\| _ {\\ell_ {2}} ^ {2}, \\tag {3}", + "image_path": "b51e629049ae5049b396c19a32510cfc581bf2feeeec8ff681c6c0b60ff05843.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\widetilde{y}_{i,j}^{k}" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": "-th column element in " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{Y}}^k" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}_j^k" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": " is the center feature of the " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": "-th sub-cluster in " + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 152, + 287, + 192 + ], + "type": "text", + "content": "-th master cluster. And we utilize the alignment loss to optimize them:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 198, + 287, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 198, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 110, + 198, + 287, + 231 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l g} ^ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\widetilde {\\boldsymbol {z}} _ {i} ^ {\\text {s u b}, k} \\cdot \\widetilde {\\boldsymbol {y}} _ {i} ^ {k}, \\tag {4}", + "image_path": "2f37e289f4d5335ebe32bc4052ad30b0e745a4970f852ba99c8478c1ba65b585.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\widetilde{z}_i^{sub,k}" + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": "-th element in " + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{Z}}^{sub,k}" + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": ". The training process on the sub-network " + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": " is synchronous with the one on the main network " + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 237, + 287, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 276, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 420 + ], + "type": "text", + "content": "In addition, hierarchical clustering [39] gradually divides data by merging pairs of clusters with smaller distances or splitting the two most distant sub-clusters in one cluster. According to the assignment distance between clusters, the candidate sets to be split/merged are generated, inspired by the hierarchical clustering. In our method, to explore more different classes, those cluster pairs (sub-cluster pairs) with far assignment distances are considered as those likely to be split. While those with close assignment distances are considered as those likely to be merged due to high similarity between two main clusters. The candidate sets to be split/merged based on this can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 287, + 440 + ], + "type": "interline_equation", + "content": "C S _ {m e r g} ^ {E D} = \\left\\{\\left(k _ {1}, k _ {2}\\right) \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} E D \\left(\\boldsymbol {\\eta} _ {k _ {1}}, \\boldsymbol {\\eta} _ {k _ {2}}\\right) \\right\\}, \\tag {5}", + "image_path": "a1782756fc86916958bcf30c7b3035c0cac2acedb012268008778845fbdc0b82.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 445, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 445, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 75, + 445, + 287, + 460 + ], + "type": "interline_equation", + "content": "C S _ {s p l} ^ {E D} = \\left\\{k \\mid \\in \\arg \\operatorname {s o r t} _ {n _ {i f}} - E D \\left(\\boldsymbol {\\eta} _ {1} ^ {k}, \\boldsymbol {\\eta} _ {2} ^ {k}\\right) \\right\\}, \\tag {6}", + "image_path": "ef2945209cff02ffc4730ac1c7bfd78de01ba30f4d6385fb3ed5262005307b16.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "inline_equation", + "content": "ED(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "text", + "content": " computes the Euclidean distance of the cluster embedding center. " + }, + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "inline_equation", + "content": "n_{if}" + }, + { + "bbox": [ + 47, + 462, + 287, + 498 + ], + "type": "text", + "content": " are the number of cluster(s) with potential to be merged and split." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "After obtaining the candidate set, we can infer more accurate " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " and explore a more accurate target class space, by calculating the Hasting ratio from the M-H framework [8] among (sub)cluster pairs in the candidate set and comparing it with 1. More details are provided in the supplementary material. In this way, we use the inferred latest target class number " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " to update the unknown diffuser (the output dimension of " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " and the number of sub-clustering layers) at the next optimization step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "During the optimization in the latest more accurate class space, we learn the discriminative cluster distribution under the latest updated target class number " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ". According to these cluster distributions, we decide the splitting and merging decisions to change the target class number " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " in the process of hierarchical class exploration. The more precise " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " and corresponding cluster assignments are generated, through operating these two processes iteratively, which is beneficial to fully explore the target class space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": "Algorithm 1 The training procedure of our proposed model \nInput: data, source pre-trained model, main network " + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": " and sub-network " + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": " of unknown diffuser" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 111, + 478, + 135 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 312, + 111, + 478, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 111, + 478, + 123 + ], + "spans": [ + { + "bbox": [ + 312, + 111, + 478, + 123 + ], + "type": "text", + "content": "1: // Exploration of the target class space" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "spans": [ + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "inline_equation", + "content": "iter_1 = 1" + }, + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "inline_equation", + "content": "iter_1^{max}" + }, + { + "bbox": [ + 312, + 124, + 438, + 135 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 309, + 136, + 545, + 338 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "text", + "content": "3: Obtain clusteredpesudo-distribution " + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}" + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}^k" + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 312, + 136, + 545, + 158 + ], + "type": "text", + "content": "-means" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "content": "4: Optimize " + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "inline_equation", + "content": "C_k^{sub}" + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "content": " guided by " + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}" + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\widetilde{Y}^k" + }, + { + "bbox": [ + 312, + 159, + 545, + 184 + ], + "type": "text", + "content": " in Eqs. (2) and (4)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 185, + 545, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 545, + 208 + ], + "type": "text", + "content": "5: Determine " + }, + { + "bbox": [ + 312, + 185, + 545, + 208 + ], + "type": "inline_equation", + "content": "CS_{slp} / CS_{merg}" + }, + { + "bbox": [ + 312, + 185, + 545, + 208 + ], + "type": "text", + "content": " by similarity measurement in Eqs. (5) and (6)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "text", + "content": "6: Diffuse the class space by updating " + }, + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "inline_equation", + "content": "CS_{slp} / CS_{merg}" + }, + { + "bbox": [ + 312, + 209, + 545, + 232 + ], + "type": "text", + "content": ", based on M-H framework" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 233, + 523, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 233, + 523, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 233, + 523, + 243 + ], + "type": "text", + "content": "7: Update the unknown diffuser by the updated " + }, + { + "bbox": [ + 312, + 233, + 523, + 243 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 245, + 357, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 245, + 357, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 245, + 357, + 255 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 256, + 488, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 488, + 268 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 488, + 268 + ], + "type": "text", + "content": "9: // Knowledge transfer and generalization" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 309, + 269, + 438, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 269, + 438, + 280 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 438, + 280 + ], + "type": "text", + "content": "10: for iter2 = 1 to iter2 max do" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 281, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 281, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 545, + 303 + ], + "type": "text", + "content": "11: Promote reliability of known knowledge by contrastive constraint in Eq. (7)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "text", + "content": "12: Jointly optimize " + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "text", + "content": " and the pre-trained model in the explored class space via " + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kt}" + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{alg}" + }, + { + "bbox": [ + 309, + 304, + 545, + 328 + ], + "type": "text", + "content": " in Eq. (8)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 309, + 329, + 356, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 329, + 356, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 329, + 356, + 338 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 361, + 516, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 516, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 516, + 374 + ], + "type": "text", + "content": "3.2. Knowledge Transfer and Generalization" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 380, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 440 + ], + "type": "text", + "content": "After exploring the target class space in Sec. 3.1, we aim to transfer knowledge on known classes more effectively and improve the model's generalization on unknown classes. In this subsection, we promote the reliability of known knowledge and further optimize over wider class distribution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "content": "Promotion of Known Knowledge Reliability. To better utilize known knowledge, we hope to improve the reliability of known knowledge obtained by the classification layer of the pre-trained source model during training. The input of the model is all target samples, including known and unknown samples, from which we need to select the known ones. According to the logit output of known samples having smaller entropy than the unknown ones, we select the top " + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "content": " of samples with the lowest entropy as known samples, which are placed in set " + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "content": ". After selection, the contrastive constraint " + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "content": " is utilized on " + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 441, + 545, + 632 + ], + "type": "text", + "content": " to enhance the discrimination of the known samples. Concretely, it brings the known features of similar predictions closer and those of dissimilar predictions farther away, achieving discrimination feature assignment to strengthen the reliability of known knowledge. The contrastive constraint is written as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 343, + 641, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 641, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 343, + 641, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = - \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {+} + \\sum_ {i \\in \\mathcal {M}} \\boldsymbol {p} _ {i} ^ {T} \\cdot \\boldsymbol {p} _ {i} ^ {-}, \\tag {7}", + "image_path": "dc06a9336ae2191906b94286d4dc3fc2d27d206e8c0fa75296c51e4c35df8f09.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i = \\sigma(g(\\pmb{z}_i)) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "content": " is the output of classifier " + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "content": " is the softmax function." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Optimization over Wider Class Distribution. After" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24019" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "that, we optimize the unknown diffuser over the wider class space by taking advantage of the reliable known knowledge and alignment constraint on the explored space. In such an optimization based on the explored wider class space, we first select the high-confidence known samples as the reliable known knowledge. Then, we hope to use the hard pseudo-labels of the high-confidence known samples as supervision information, thereby leveraging them for better knowledge transfer on the explored class space. Specifically, for a known sample, we take the maximum logit output by the classifier in the source pre-trained model as its confidence. And the known sample is considered as the high-confidence known sample when it is greater than a manually-set confidence threshold, whose hard pseudolabel is represented by the one-hot paradigm " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^h (x_i^h\\in \\mathcal{H})" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " is the set of the high-confidence known samples. Therefore, we can perform better knowledge transfer on the explored class space by leveraging the reliable known knowledge " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^h" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " obtained on the source pre-trained model. In addition, we also perform alignment constraint on the explored class space, similar to Eq. (2), to complete the transfer on known samples further and achieve generalization on unknown samples. Finally, we leverage the supervision of the high-confidence known sample labels and the alignment constraint, to jointly achieve superior knowledge transfer on known classes and generalization on unknown classes over the wider class space:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 403, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 403, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 51, + 403, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {k t} + \\mathcal {L} _ {a l g} = \\sum_ {i} ^ {\\mathcal {H}} L _ {k t} (\\widetilde {\\boldsymbol {z}} _ {i}, \\boldsymbol {y} _ {i} ^ {h}) + \\sum_ {i = 1} ^ {n} L _ {a l g} (x _ {i} ^ {t}) \\\\ = \\underbrace {\\sum_ {i} ^ {\\mathcal {H}} L _ {k t} \\left(\\widetilde {\\boldsymbol {z}} _ {i} , \\boldsymbol {y} _ {i} ^ {h}\\right) + \\sum_ {j _ {1} = 1} ^ {n _ {1}} L _ {a l g} \\left(x _ {j _ {1}} ^ {t}\\right)} _ {\\text {k n o w n}} + \\underbrace {\\sum_ {j _ {2} = 1} ^ {n _ {2}} L _ {a l g} \\left(x _ {j _ {2}} ^ {t}\\right)} _ {\\text {u n k n o w n}}, \\tag {8} \\\\ \\end{array}", + "image_path": "58ec5ec067b59220cfc1cc3bda2d6fe90e060c288e24ca515c7cd8451925e170.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kt}" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " represents the guidance of high-confidence known samples, which is the cross-entropy loss between their pseudo-labels and soft distributions. " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{alg}" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " is the alignment constraint to achieve superior knowledge transfer on known classes and generalization on unknown classes. " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\widetilde{z}_i" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " represents the soft-distribution output by the main network " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " on the explored class space. " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 47, + 496, + 287, + 593 + ], + "type": "text", + "content": " represent the number of known and unknown samples, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 621, + 164, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 164, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 164, + 635 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 639, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 286, + 700 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three benchmark datasets for OS-SFDA through experiments, which are Office-31 [28], Office-Home [36], and VisDA-C 2017 [24]. Due to limited space, the results and analysis on VisDA-C 2017 are summarized in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "- Office-31 [28] consists of three domains (Amazon, We" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "type": "text", + "content": "bcam, DSLR), with 31 classes and 4652 images of common objects. Following the protocols [18] on Office-31 and Office-Home, we select the first 20 classes as the shared classes of the source and target domain while the remaining 11 classes are considered unknown." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 204 + ], + "type": "text", + "content": "- Office-Home [36] is a more challenging dataset. It contains four domains (Art, Product, Clipart, RealWorld) with 65 classes and 15500 images of common objects. Like Office-31, we select the first 25 classes as classes shared by the source and target domain, and the remaining 40 classes are considered unknown." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "type": "text", + "content": "Baselines. For OS-SFDA, we compare our method with the clustering on Resnet [9], the classic method SHOT [16], and the advanced method AaD [44], which do sufficient experiments under the OS-SFDA setting. In addition, our method uses the unknown diffuser based on clustering (class number-agnostic) to explore the target class space. For a fair comparison, we thoroughly compare our method with the no-parameter clustering methods (FINCH [32], COMIC [25], DenMune [1])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": "Evaluation. As shown in Tab. 1 and Tab. 2, we report domain adaptation results for " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "OS^{*}" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "UNK" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": ", which represent the average accuracy of the known and the unknown class, respectively. Simultaneously, " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "HOS = \\frac{2 \\times OS^{*} \\times UNK}{OS^{*} + UNK}" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": " is denoted as the harmonic mean between " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "OS^{*}" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "UNK" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": ", which is the key indicator for evaluating performance. Moreover, the inferred target class number " + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": " and accuracy ACC are used to evaluate the effectiveness of exploring the target class space in Tab. 3 and Tab. 4. The setting of ACC follows the previous protocols [12]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 602 + ], + "type": "text", + "content": "Implementation Details. For the source pre-trained model, we utilize the backbone of ResNet-50 [9] for Office-31 and Office-Home. For a fair comparison, we use the same network architecture with SHOT [16] and AaD [44], which includes a backbone and two additional fully connected layers. We adopt SGD with a momentum of 0.9 and a batch size of 64 for all datasets. The learning rate is set to 1e-3 on all datasets for all layers of the backbone, and the last two added fully-connected layers are applied 1e-2. We train 15 epochs for all datasets. The unknown diffuser includes one convolutional layer with 256 hidden dimensions. For a fair comparison, our network and the compared no-parameter clustering methods use the same input features, output by the source pre-trained model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 611, + 425, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 425, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 425, + 624 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1 and Tab. 2, the UNK and HOS for our method are much higher than SHOT on all datasets. Our method achieves " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " relative improvements over the state-of-art AaD for average OS* and HOS on Office-31, respectively. On the relatively small-scale Office-31 with fewer classes, our method demonstrates competitiveness across various indicators, standing in con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24020" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 535, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 535, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 535, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 535, + 159 + ], + "type": "table", + "html": "
methodSFA→DA→WD→AD→WW→AW→DAvg
OS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOSOS*UNK HOS
Cluster✗10.064.017.07.061.013.05.069.010.08.049.014.05.064.010.010.047.017.07.559.013.5
SHOT✓87.250.055.789.038.653.878.244.256.597.472.383.073.547.657.889.675.081.785.854.664.8
AaD✓69.986.777.462.083.971.360.881.869.883.297.489.858.481.568.082.395.288.369.487.877.4
Ours(w/o s)✓78.473.475.887.990.389.172.376.174.192.880.586.270.976.873.785.879.882.781.479.580.3
Ours(w/s)✓82.371.876.790.072.080.072.079.475.579.587.383.274.282.778.280.688.384.379.880.379.7
", + "image_path": "ccdf777ca692021fc6e7c77ae47671ec8314e75128605bb4e3ab481f21219952.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 210, + 537, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "type": "text", + "content": "Table 1. Accuracy (%) on Office-31 for OS-SFDA. The three indicators(OS*, UNK, HOS) have been explained in Sec. 4.1. ‘w/o s’ and ‘w/s’ indicate with or without supervision using reliable known knowledge. Note that ‘A→D’ represents that the model per-trained on the source domain Amazon transfers to the target domain DSLR. Bold and ‘_’ represent the best and second results, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 210, + 537, + 384 + ], + "lines": [ + { + "bbox": [ + 58, + 210, + 537, + 384 + ], + "spans": [ + { + "bbox": [ + 58, + 210, + 537, + 384 + ], + "type": "table", + "html": "
methodSFAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwAvg (first 6 tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX6.063.011.05.055.010.04.062.08.08.059.014.05.055.09.05.056.08.05.056.08.05.558.310.0
SHOT✓67.028.039.581.826.339.887.532.147.066.846.254.677.527.240.280.025.939.176.831.043.4
AaD✓50.567.457.764.066.465.172.269.570.847.180.359.364.768.266.465.071.067.860.670.564.5
Ours(w/o s)✓47.669.256.470.081.875.470.685.677.460.374.766.765.682.773.272.184.277.664.479.771.1
Ours(w/s)✓49.064.855.868.188.076.771.586.678.461.372.666.466.581.673.171.684.877.664.779.771.3
methodSFPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrRw→PrAvg (all tasks)
OS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNK HOSOS* UNKHOS
ClusterX8.053.015.06.050.010.05.061.09.08.051.013.05.064.010.05.050.010.05.856.610.6
SHOT✓66.351.157.759.331.040.885.831.646.273.550.659.965.328.940.184.428.242.374.633.945.6
AaD✓46.983.160.045.072.655.669.072.370.656.077.465.048.367.656.467.769.368.558.072.163.6
Ours(w/o s)✓60.076.767.352.666.258.670.785.177.256.583.367.449.170.757.966.182.173.261.878.569.0
Ours(w/s)✓55.985.667.645.470.255.173.983.978.656.784.167.849.374.659.469.580.074.461.679.769.2
", + "image_path": "13712ff936ebcb14398017864fa06991e683284495c11375b67b6a78bb999802.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "type": "text", + "content": "Table 2. Accuracy (%) on Office-Home for OS-SFDA. Note that ‘Ar→Cl’ represents that the model per-trained on the source domain Art transfers to the target domain Clipart. The representations of OS*, UNK, HOS, ‘w/o s’, and ‘w/s’ are the same with Tab. 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": "trast to methods such as AaD and SHOT, where only one indicator exhibits competitive performance. SHOT tends to excel in OS*, while AaD typically performs well on UNK. In contrast, our method displays a well-balanced performance across different measurement dimensions, achieving competitive results overall. For the more challenging Office-Home, our method displays more apparent superiority, which achieves " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "61.6\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": " average accuracy for OS*, UNK, and HOS, with " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": " relative improvement compared with AaD. Furthermore, our method outperforms all no-parameter clustering methods (FINCH, DenMune, and COMIC) for ACC and " + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 426, + 289, + 593 + ], + "type": "text", + "content": " on all tasks in Tab. 3 and Tab. 4. The above results demonstrate the superiority and effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Impact of Exploration for the Target Class Space. Compared with OS-SFDA methods (SHOT and AaD) that group all target unknown classes as one, our method tends to explore class space and excavate different unknown classes, which has great application value in more realistic scenarios. The experimental results show that our method achieves excellent improvement over the state-of-the-art method AaD for the known accuracy " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{OS^{*}}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " on the more challenging Office Home. In addition, the unknown accuracy UNK for our method is much higher than SHOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "content": "on Office-31 and Aad on Office-Home. This phenomenon shows that our method can improve knowledge transfer on known classes when exploring class space and simultaneously improve generalization on unknown classes. To more fully demonstrate the effectiveness and advancement of our exploration, we compare our method with classical no-parameter clustering methods for exploring the class space. Concretely, ours provides the more precise inferred target class number " + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "content": " and higher accuracy ACC on all tasks, where the precision of " + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "content": " represents the efficacy in exploring the target class space, as discussed in Sec. 3.1. The above analysis conclusively demonstrates that our method of exploring class space is exceptionally effective, improving knowledge transfer and generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Effect of Reliable Known Knowledge as Supervision. As introduced in Sec. 3.2, we hope to utilize the reliable known knowledge to optimize the unknown diffuser for better knowledge transfer and generalization. Concretely, our method that uses the reliable known knowledge(ours (w/ s)) outperforms all compared methods(including ours(w/o s)) on more challenging Office-Home for average HOS, ACC, and " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": " in Tabs. 1 to 4. Therefore, the more effectiveness brought by reliable known knowledge ('ours (w/ s)') confirms it guides the model for better known and unknown" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24021" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 82, + 164, + 175 + ], + "blocks": [ + { + "bbox": [ + 60, + 82, + 164, + 175 + ], + "lines": [ + { + "bbox": [ + 60, + 82, + 164, + 175 + ], + "spans": [ + { + "bbox": [ + 60, + 82, + 164, + 175 + ], + "type": "image", + "image_path": "1ab423e646ede09859f9823eb36263394f2465b816b76b89757dc15bfdc88c1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 182, + 117, + 194 + ], + "lines": [ + { + "bbox": [ + 104, + 182, + 117, + 194 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 117, + 194 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 82, + 284, + 175 + ], + "blocks": [ + { + "bbox": [ + 181, + 82, + 284, + 175 + ], + "lines": [ + { + "bbox": [ + 181, + 82, + 284, + 175 + ], + "spans": [ + { + "bbox": [ + 181, + 82, + 284, + 175 + ], + "type": "image", + "image_path": "af9dd5e01ac10a377c37e314bee6102371cc81e27656a37f85322e057d5c943a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 182, + 239, + 194 + ], + "lines": [ + { + "bbox": [ + 225, + 182, + 239, + 194 + ], + "spans": [ + { + "bbox": [ + 225, + 182, + 239, + 194 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "text", + "content": "Figure 3. (a) t-SNE of SHOT on " + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\to \\mathbb{P}" + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "text", + "content": ". (b) t-SNE of AaD on " + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\to \\mathbb{P}" + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "text", + "content": ". (c) t-SNE of our method on " + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\to \\mathbb{P}" + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "text", + "content": ". (d) Parameter analysis on " + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\rightarrow \\mathrm{W}" + }, + { + "bbox": [ + 50, + 201, + 541, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 82, + 405, + 175 + ], + "blocks": [ + { + "bbox": [ + 304, + 82, + 405, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 82, + 405, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 82, + 405, + 175 + ], + "type": "image", + "image_path": "8ff444d523cf7f5fe0392fbc3fea6ad95c9e56df8187c93cff1f15e2d64751df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 182, + 361, + 194 + ], + "lines": [ + { + "bbox": [ + 349, + 182, + 361, + 194 + ], + "spans": [ + { + "bbox": [ + 349, + 182, + 361, + 194 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 81, + 529, + 175 + ], + "blocks": [ + { + "bbox": [ + 421, + 81, + 529, + 175 + ], + "lines": [ + { + "bbox": [ + 421, + 81, + 529, + 175 + ], + "spans": [ + { + "bbox": [ + 421, + 81, + 529, + 175 + ], + "type": "image", + "image_path": "be0297374d1be915c72b94367dec59dc52cfe1f74e5f50696723cba779430542.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 182, + 485, + 194 + ], + "lines": [ + { + "bbox": [ + 473, + 182, + 485, + 194 + ], + "spans": [ + { + "bbox": [ + 473, + 182, + 485, + 194 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 232, + 286, + 310 + ], + "blocks": [ + { + "bbox": [ + 48, + 232, + 286, + 310 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 286, + 310 + ], + "type": "table", + "html": "
MethodsA→DA→WD→AD→WW→AW→D
FINCH31.0 (8)8.0 (2)23.0 (7)12.0 (3)28.0 (9)23.0 (5)
DenMune20.0(5)52.0(15)58.0(29)46.0(14)65.0(35)22.0(5)
COMIC56.8 (22)47.3 (24)16.1 (40)46.4 (20)12.1 (32)56.6 (22)
Ours(w/o s)69.0(24)80.6(28)60.4(37)82.5(30)59.6(38)68.9(27)
Ours(w/ s)71.2(25)73.3(28)63.1(37)76.3(32)64.2(38)74.5(28)
", + "image_path": "1f333b158594b31c74b2e6ac79c75629b30e86b40d9514760457e764cdb4394a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 54, + 363, + 282, + 517 + ], + "blocks": [ + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "text", + "content": "Table 3. Accuracy ACC(%) and the inferred target class number " + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "text", + "content": " on Office-31. Note that the values before " + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\left( \\cdot \\right)" + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "text", + "content": " ’ and inside " + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\left( \\cdot \\right)" + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "text", + "content": " ’ represent the performance ACC and " + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 318, + 287, + 352 + ], + "type": "text", + "content": " ,respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 363, + 282, + 517 + ], + "lines": [ + { + "bbox": [ + 54, + 363, + 282, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 282, + 517 + ], + "type": "table", + "html": "
MethodsAr→ClAr→PrAr→RwCl→ArCl→PrCl→Rw
FINCH5.0(4)7.0(3)6.0(3)8.0(3)7.0(4)7.0(4)
DenMune15.0(12)58.0(45)54.0(39)16.0(7)55.0(40)40.0(32)
COMIC6.9(65)50.3(92)9.43(34)4.0(1)10.4(51)2.3(2)
Ours(w/o s)35.5(69)60.5(77)59.1(77)47.3(61)58.9(74)55.6(60)
Ours(w/s)35.9(69)61.4(80)61.3(77)46.9(62)58.9(78)61.0(60)
MethodsPr→ArPr→ClPr→RwRw→ArRw→PrRw→Cl
FINCH12.0(4)5.0(4)4.0(2)11.0(4)12.0(7)7.0(4)
DenMune16.0(7)7.0(6)52.0(37)26.0(13)9.0(8)62.0(43)
COMIC4.2(2)3.7(24)5.7(14)4.1(1)5.2(43)13.3(65)
Ours(w/o s)48.5(51)37.2(56)63.6(75)38.1(28)21.6(50)50.4(75)
Ours(w/s)47.8(51)35.5(56)64.0(68)44.2(35)21.9(50)51.8(75)
", + "image_path": "c54e13aa563c0c8ee426c099dd2b6aa1ef23edd436cc2a9fe6282e582d790696.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 524, + 244, + 534 + ], + "lines": [ + { + "bbox": [ + 89, + 524, + 244, + 534 + ], + "spans": [ + { + "bbox": [ + 89, + 524, + 244, + 534 + ], + "type": "text", + "content": "Table 4. ACC(%) and " + }, + { + "bbox": [ + 89, + 524, + 244, + 534 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 89, + 524, + 244, + 534 + ], + "type": "text", + "content": " on Office-Home." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 545, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 568 + ], + "type": "text", + "content": "class discrimination, achieving better knowledge transfer on known classes and generalization on unknown classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Parameter Analysis. We study the impact of supervision information on unknown diffuser optimization during training by applying a variable parameter to the supervision loss. As shown in Fig. 3(d), the chart indicates that when there is no supervision information, the optimization cannot be completed well to obtain a favorable transfer effect on known or unknown classes. As the parameters increase within a certain range, UNK and HOS also improve; but parameters that are too large will lead to poor optimization results. In addition, it can be seen that OS* is not sensitive to this parameter, and its change trend is relatively stable. Therefore, we should select appropriate high" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "text", + "content": "confidence known pseudo-labels as supervision information to prevent model overfitting for better guiding the optimization of unknown diffuser." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 271, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 546, + 390 + ], + "type": "text", + "content": "Effectiveness of the Unknown Diffuser. Tabs. 1 to 4 show our superior performance in exploring unknown class space and domain adaptation in OS-SFDA. The superiority in unknown class exploration indicates that our unknown diffuser can excavate potential unknown classes under the guidance of current discriminative class distributions based on the Metropolis-Hastings (M-H) framework. The impressive domain adaptation effects based on the explored wider space also demonstrate the necessity of using the unknown diffuser for unknown class exploration." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 392, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 439 + ], + "type": "text", + "content": "Visualization of features. As shown in Fig. 3(a)-(c), t-SNE visualizations indicate that compared with existing methods, our method explores a more accurate and extensive class space, exhibiting remarkable class discrimination." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": "In this paper, we propose a domain adaptation network in OS-SFDA for exploring unknown classes, thereby jointly benefiting known knowledge transfer and unknown generalize. We introduce an unknown diffuser into the source pretrained model to excavate a wider target class space. And, the reliable known class knowledge and clustered pseudolabels over the wider class space are captured to serve as supervision during model optimization, realizing impressive knowledge transfer and generalization. Extensive experiments verify the superiority of our method." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 605, + 421, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 421, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 421, + 618 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 624, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 703 + ], + "type": "text", + "content": "This work is supported in part by the National Key Research and Development Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62132016, 62171343, and 62201436), and Fundamental Research Funds for the Central Universities (ZDRC2102)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24022" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Mohamed Abbas, Adel El-Zoghabi, and Amin Shoukry. Denmune: Density peak based clustering using mutual nearest neighbors. Pattern Recognit., 109:107589, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 422-438. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 203 + ], + "type": "text", + "content": "[3] Jason Chang and John W Fisher III. Parallel sampling of dp mixture models using sub-cluster splits. Proc. Adv. Neural Inf. Process. Syst., 26, 2013. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "text", + "content": "[4] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7212-7222, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 280 + ], + "type": "text", + "content": "[5] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 313 + ], + "type": "text", + "content": "[6] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE Trans. Image Process., 30:6473-6484, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 368 + ], + "type": "text", + "content": "[7] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. J. Mach. Learn. Res., 17(1):2096-2030, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 392 + ], + "type": "text", + "content": "[8] W Keith Hastings. Monte carlo sampling methods using markov chains and their applications. 1970. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 434 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 435, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 287, + 490 + ], + "type": "text", + "content": "[10] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35:16755-16767, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[11] Taotao Jing, Hongfu Liu, and Zhengming Ding. Towards novel target discovery through open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9322-9331, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 558 + ], + "type": "text", + "content": "[12] Harold W Kuhn. The hungarian method for the assignment problem. Nav. Res. Logist. Q., 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "text", + "content": "[13] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12376-12385, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[14] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 12365-12377. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 691 + ], + "type": "text", + "content": "[15] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 24110-24119, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "text", + "content": "unsupervised domain adaptation. In Proc. Int. Conf. Mach. Learn., pages 6028-6039. PMLR, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[17] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "text", + "content": "[18] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2927-2936, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "text", + "content": "[19] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Proc. Adv. Neural Inf. Process. Syst., 29, 2016. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[20] Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(12):3071-3085, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 267, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 301 + ], + "type": "text", + "content": "[21] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 31, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[22] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 347, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 392 + ], + "type": "text", + "content": "[23] Yingwei Pan, Ting Yao, Yehao Li, Chong-Wah Ngo, and Tao Mei. Exploring category-agnostic clusters for open-set domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 13867–13875, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[24] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 438, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 483 + ], + "type": "text", + "content": "[25] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In Proc. Int. Conf. Mach. Learn., pages 5092-5101. PMLR, 2019. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 484, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 508 + ], + "type": "text", + "content": "[26] Masoud Pourreza. Open-set source-free domain adaptation in fundus images analysis. 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 553 + ], + "type": "text", + "content": "[27] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In Proc. Eur. Conf. Comput. Vis., pages 537-555. Springer, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 554, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 588 + ], + "type": "text", + "content": "[28] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In Proc. Eur. Conf. Comput. Vis., pages 213-226. Springer, 2010. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 589, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 623 + ], + "type": "text", + "content": "[29] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9000-9009, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "type": "text", + "content": "[30] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proc. Eur. Conf. Comput. Vis., pages 153-168, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[31] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. Proc. Adv. Neural Inf. Process. Syst., 33:16282-16292, 2020. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24023" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Saquib Sarfraz, Vivek Sharma, and Rainer Stiefelhagen. Efficient parameter-free clustering using first neighbor relations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8934-8943, 2019. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[33] Tasfia Shermin, Guojun Lu, Shyh Wei Teng, Manzur Murshed, and Ferdous Sohel. Adversarial network with multiple classifiers for open set domain adaptation. IEEE Trans. Multimed., 23:2732-2744, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "text", + "content": "[34] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proc. AAAI Conf. Artif. Intell., 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "type": "text", + "content": "[35] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darryll. Adversarial discriminative domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7167-7176, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[36] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 5018-5027, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "type": "text", + "content": "[37] Guillaume Vray, Devavrat Tomar, Behzad Bozorgtabar, and Jean-Philippe Thiran. Source-free open-set domain adaptation for histopathological images via distilling self-supervised vision transformer. arXiv preprint arXiv:2307.04596, 2023. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "text", + "content": "[38] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 413 + ], + "type": "text", + "content": "[39] Joe H Ward Jr. Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc., 58(301):236-244, 1963. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[40] Kun Wei, Xu Yang, Zhe Xu, and Cheng Deng. Class incremental unsupervised domain adaptation via pseudolabel distillation. IEEE Trans. Image Process., 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[41] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9010-9019, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "text", + "content": "[42] Jiexi Yan, Zhihui Yin, Erkun Yang, Yanhua Yang, and Heng Huang. Learning with diversity: Self-expanded equalization for better generalized deep metric learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 19365-19374, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "text", + "content": "[43] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8978-8987, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[44] Shiqi Yang, Shangling Jui, Joost van de Weijer, et al. Attracting and dispersing: A simple approach for source-free domain adaptation. Proc. Adv. Neural Inf. Process. Syst., 35: 5802-5815, 2022. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[45] Shiqi Yang, Yaxing Wang, Luis Herranz, Shangling Jui, and Joost van de Weijer. Casting a bait for offline and online source-free domain adaptation. Comput. Vis. Image Underst., page 103747, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[46] Xu Yang, Cheng Deng, Tongliang Liu, and Dacheng Tao. Heterogeneous graph attention network for unsupervised multiple-target domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1992-2003, 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Yixin Zhang, Zilei Wang, and Weinan He. Class relationship embedded learning for source-free unsupervised domain adaptation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7619-7629, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[48] Han Zhao, Xu Yang, and Cheng Deng. Parameter-agnostic deep graph clustering. ACM Trans. Knowl. Discov. Data, 18 (3):1–20, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[49] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Open set domain adaptation by novel class discovery. arXiv preprint arXiv:2203.03329, 2022. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24024" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_content_list.json b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..149c0c2cf59bd3f13ebd179617408e2b89e80262 --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_content_list.json @@ -0,0 +1,2033 @@ +[ + { + "type": "text", + "text": "Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution", + "text_level": 1, + "bbox": [ + 202, + 128, + 772, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shangchen Zhou* Peiqing Yang* Jianyi Wang Yihang Luo Chen Change Loy S-Lab, Nanyang Technological University", + "bbox": [ + 140, + 202, + 830, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://shangchenzhou.com/projects/upscale-a-video", + "bbox": [ + 258, + 241, + 705, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3df4f0b8c0bbe58fd8efa2da63b6b1ada06c32e315115a9bc72d62369fd25ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 276, + 594, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e954c433727a5bf1e749f520a83487b701de7d06eac5702f21aeda33c2930c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 276, + 890, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3cfd7d11cb82a7723e261bfdce2ce4ad291e4f2728b9ec5c30531609ddce7d1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 505, + 217, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3415af2147966bf0abd49be26aad9fc8ba93666e3cd457ea2505bde4a9c8da6.jpg", + "image_caption": [ + "ResShift" + ], + "image_footnote": [], + "bbox": [ + 217, + 505, + 352, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27a3154fefffad72ac0dc582fed48b1712597ef56f9a69c0b3bae079b13cf69c.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 351, + 505, + 486, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0786d235379a2426867ab856975e2b308c9349568aabfe900ed1ebe549c4401.jpg", + "image_caption": [ + "SD $\\times 4$ Upscaler" + ], + "image_footnote": [], + "bbox": [ + 486, + 505, + 617, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6be451dfed78bae9943bac056907db1afb342da10fd65488055acca444e155f.jpg", + "image_caption": [ + "RealBasicVSR", + "Figure 1. Video super-resolution comparisons on both real-world and AI-generated videos. Our proposed Upscale-A-Video showcases excellent upscaling capabilities. By using appropriate text prompts, it achieves impressive results characterized by finer details and heightened visual realism. (Zoom-in for best view)" + ], + "image_footnote": [], + "bbox": [ + 619, + 505, + 753, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8d44988b4926a90ed16197a68632974c3dbdce8b9b2218d608e58c5aa44f070.jpg", + "image_caption": [ + "Upscale-A-Video (Ours)" + ], + "image_footnote": [], + "bbox": [ + 753, + 505, + 888, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 728, + 313, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based diffusion models have exhibited remarkable success in generation and editing, showing great promise for enhancing visual content with their generative prior. However, applying these models to video super-resolution remains challenging due to the high demands for output fidelity and temporal consistency, which is complicated by the inherent randomness in diffusion models. Our study introduces Upscale-A-Video, a text-guided latent diffusion", + "bbox": [ + 75, + 758, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "framework for video upscaling. This framework ensures temporal coherence through two key mechanisms: locally, it integrates temporal layers into U-Net and VAE-Decoder, maintaining consistency within short sequences; globally, without training, a flow-guided recurrent latent propagation module is introduced to enhance overall video stability by propagating and fusing latent across the entire sequences. Thanks to the diffusion paradigm, our model also offers greater flexibility by allowing text prompts to guide texture creation and adjustable noise levels to balance restoration and generation, enabling a trade-off between fidelity and", + "bbox": [ + 500, + 729, + 893, + 895 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2535", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality. Extensive experiments show that Upscale-A-Video surpasses existing methods in both synthetic and real-world benchmarks, as well as in AI-generated videos, showcasing impressive visual realism and temporal consistency.", + "bbox": [ + 76, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 183, + 209, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) in real-world scenarios is a challenging task that aims at enhancing the quality of low-quality videos to produce high-quality results. Unlike previous works that mainly focus on either synthetic degradations [8, 9, 64] or specific camera-related degradations [80], this task is more demanding due to the need for addressing complex and unknown degradations commonly found in low-quality videos, such as downsampling, noise, blur, flickering, and video compression. In addition, maintaining the visual fidelity and ensuring temporal coherence are crucial for the perceptual quality of the output video. Although recent convolutional neural network (CNN)-based networks [10, 73] have shown success in mitigating many forms of degradation, they still fall short in producing realistic textures and details due to their limited generative capabilities, often resulting in over-smoothing, which can be observed in RealBasicVSR results shown in Fig. 1.", + "bbox": [ + 75, + 209, + 468, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [24] have exhibited impressive proficiency in generating high-quality images [49, 50, 52, 85] and videos [6, 11, 17, 25, 87]. Harnessing their generative potential, diffusion-based models have been proposed for image restoration, including both training from scratch [50, 51, 83] and fine-tuning from pretrained Stable Diffusion models [38, 62]. These methods have effectively mitigated the over-smoothing issue often observed in CNN-based models, yielding results with more realistic fine-grained details. However, adapting these diffusion priors to VSR remains a non-trivial challenge. This difficulty stems from the inherent randomness in diffusion sampling, which inevitably introduces unexpected temporal discontinuities in resulting videos. This issue is more pronounced in latent diffusion, where the VAE decoder further introduces flickering in low-level texture details.", + "bbox": [ + 75, + 465, + 468, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent efforts have been made to adapt image diffusion models for video tasks by introducing strategies for temporal consistency, which include: 1) Fine-tuning video models with temporal layers such as 3D convolution [6, 66] and temporal attention [6, 11, 66, 87]; 2) Employing zero-shot mechanisms like cross-frame attention [71, 78] and flow-guided attention [14, 18] in the pretrained models. Although these solutions significantly improve video stability, there remain two primary issues: i) The current methods, operating in the U-Net feature or latent space, struggle to maintain low-level consistency, with issues like texture flickering still present. ii) The existing temporal layers and attention mechanisms can only impose constraints on", + "bbox": [ + 75, + 705, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "short, local input sequences, limiting their capacity to ensure global temporal consistency in longer videos.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these issues, we adopt a local-global strategy for maintaining temporal consistency in video reconstruction, focusing on both fine-grained textures and overall consistency. On a local video clip, we explore finetuning a pretrained image $\\times 4$ upscaling model [2] with additional temporal layers on the video data. Specifically, within a latent diffusion framework, we first finetune the U-Net with integrated 3D convolutions and temporal attention layers, and then tune the VAE-Decoder with video-conditioned inputs and 3D convolutions. The former significantly achieves structure stability in local sequences, while the latter further improves low-level consistency, reducing texture flickering. On a more global scale, we introduce a novel, training-free flow-guided recurrent latent propagation module. Spanning short video segments, it bidirectionally conducts frame-by-frame propagation and latent fusion during inference. Leveraging latent fusion in a long-term sequence, this module encourages overall stability for long videos.", + "bbox": [ + 496, + 122, + 892, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further investigate the generative potential of diffusion models in the VSR task. Following the text-guided fashion, our model can leverage text prompts as an optional condition to steer the model towards producing more realistic and high-quality details, as illustrated in Fig. 1. Furthermore, we enhance the robustness of our model against heavy or unseen degradation by injecting noise into input to dilute its degradation. By adjusting the level of added noise, we can modulate the balance between restoring and generating within the diffusion model. Lower noise levels prioritize the model's restoration capabilities, while higher levels encourage generation of more refined details, thus achieving a trade-off between fidelity and quality of the output.", + "bbox": [ + 496, + 392, + 892, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main contribution of our study is a practical and robust approach to real-world VSR. In particular, we conduct an exhaustive exploration of integrating a local-global temporal strategy within a latent diffusion framework, enhancing temporal coherence and generation quality. We further examine ways to improve versatility by allowing text prompts to guide texture creation, and offer control over noise levels to balance restoration and generation, thereby achieving a trade-off between fidelity and quality. Thanks to our design and pretrained generative prior, our model achieves state-of-the-art performance on existing benchmarks, showing remarkable visual realism and temporal consistency.", + "bbox": [ + 496, + 588, + 890, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 799, + 640, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Super-Resolution. VSR aims to restore a sequence of high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts. Most existing approaches [7-9, 28-31, 36, 37, 64, 75] assume a pre-defined degradation process [39, 45, 75, 82], and their performance", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2536", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "deteriorates significantly in real-world scenarios due to the limited generalizability. To tackle real-world VSR, recent works go beyond the traditional paradigms by assuming inputs with unknown degradations. Due to the lack of real-world paired data for training, Yang et al. [80] propose to collect HR-LR data pairs with iPhone cameras. While the VSR model trained on such data can be effective to videos captured by similar mobile cameras, its generalization capability to other devices is in doubt. Rather than relying on labor-intensive pair data collection, recent studies [10, 73] have shifted towards employing diverse degradations for data augmentation during training, demonstrating better performance in handling real-world cases. Nonetheless, it is still challenging for existing CNN-based approaches [10, 73] to generate photo-realistic textures due to the absence of generative prior. In this study, we set our sights on exploiting robust and extensive generative prior encapsulated in a pretrained image diffusion model, i.e., Stable Diffusion (SD) $\\times 4$ upscaler [2]. By integrating this strong diffusion prior, our approach circumvents the need for exhaustive training from scratch and exhibits improved performance in producing detailed textures.", + "bbox": [ + 76, + 90, + 472, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for Video Tasks. Following the success of text-to-image diffusion models [4, 12, 16, 20, 22, 44, 85], recent studies have ventured into the applications of video diffusion models [15, 25-27, 40-42]. Instead of training from scratch, some methods [14, 19, 48, 71, 78] focus on video generation using off-the-shelf image diffusion models [50, 85] in a zero-shot manner. To keep temporal consistency, cross attention [60] between neighboring frames and optical flow [58] warping are popular solutions adopted in these methods. While being efficient, these methods generally suffer from limited generalizability and the selection of hyperparameters can be tricky. Most recently, Blattmann et al. [6] propose to extend pretrained image diffusion models to the video domain by introducing an additional temporal dimension and fine-tuning the temporal layers only to speed up the training process. Subsequent works [21, 74] following such paradigm are capable of generating impressive video sequences. Inspired by these works, we employ a pretrained image $\\times 4$ upscaling model [2] as generative prior and propose a novel local-global temporal strategy, resulting in temporally coherent outputs with faithful details.", + "bbox": [ + 76, + 421, + 472, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for Restoration Tasks. With notable advancements in diffusion models [24, 46, 55, 77], numerous diffusion-based works have been proposed for image restoration. A straightforward way is to train a diffusion model conditioned on LR images from scratch [50, 51, 53, 72, 83], which, however, demands significant computational resources for training. To avoid such heavy training costs, another prevalent approach [12, 13, 32, 56, 67, 69, 76] is to incorporate constraints into the reverse diffusion process of a pretrained diffusion model. Though efficient, designing these constraints relies on pre-defined image degra", + "bbox": [ + 76, + 736, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dation processes or pretrained SR models as a priori, leading to limited generalizability and inferior performance. Recent works [62, 79] go further by fine-tuning directly on a frozen pretrained diffusion model with several additional trainable layers, demonstrating impressive performance. Inspired by these recent advances, we focus on exploiting the effective diffusion prior for real-world VSR, which remains under-explored and challenging due to the temporal discontinuities caused by inherent randomness during diffusion sampling. With the proposed local-global temporal strategy, our approach is capable of generating temporally coherent results for real-world VSR.", + "bbox": [ + 496, + 90, + 893, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 286, + 635, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to develop a text-guided diffusion framework tailored for real-world VSR. The diffusion denoising process, characterized by its inherent stochastic nature, poses significant challenges when applied to video tasks. These challenges include temporal instability and the emergence of flickering artifacts, which are particularly prominent in VSR tasks involving lengthy video sequences. The complexity of these tasks lies not just in achieving temporal consistency within localized segments, but also in preserving coherence throughout the entire video.", + "bbox": [ + 496, + 311, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our framework incorporates both local and global modules in the latent diffusion model (LDM) to preserve temporal consistency both within and across video segments. Within each diffusion time step $(t = 1,2\\dots T)$ , the video is split into segments and processed with a U-Net that includes temporal layers to ensure consistency within each segment. If the current time step falls within the user-specified global refinement steps $(T^{*})$ , a recurrent latent propagation module is employed to improve consistency across segments during inference. Lastly, a finetuned VAE-Decoder is used to reduce remaining flickering artifacts. Thanks to the diffusion paradigm, our model exhibits remarkable versatility: 1) Input text prompts can enhance video quality further with improved realism and details. 2) User-specified noise levels provide a control over the trade-off between quality and fidelity.", + "bbox": [ + 496, + 462, + 895, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Diffusion Models", + "text_level": 1, + "bbox": [ + 498, + 712, + 774, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pretrained Stable Diffusion Image $\\times 4$ Upscaler. Our Upscale-A-Video is built upon the pretrained text-guided SD $\\times 4$ Upscaler [2]. It employs the LDM framework [50], which uses an autoencoder to transform an image into a latent with an encoder $\\mathcal{E}$ ( $\\times 4$ downsampling), and reconstructs it with a decoder $\\mathcal{D}$ . Conditioned on the low-resolution images $\\mathbf{x}$ , it learns to generate the high-quality counterparts via iterative denoising in the latent space. To achieve this, given latent samples $\\mathbf{z} \\sim p_{data}$ , diffused latents $\\mathbf{z}_t = \\alpha_t\\mathbf{z} + \\sigma_t\\boldsymbol{\\epsilon}$ is generated by introducing Gaussian noise to the latents $\\mathbf{z}$ at each diffusion step $t$ ; where", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2537", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4555426d0f86e56663c3e09a24bb254703a67af42bd4235e40439c4965f8afb8.jpg", + "image_caption": [ + "Figure 2. An overview of Upscale-A-Video. Upscale-A-Video processes long videos using both local and global strategies to maintain temporal coherence. It divides the video into segments and processes them using a U-Net with temporal layers for intra-segment consistency. During user-specified diffusion steps for global refinement, a recurrent latent propagation module is used to enhance inter-segment consistency. Finally, a finetuned VAE-Decoder reduces remaining flickering artifacts for low-level consistency. Our model also allows users to guide texture creation with text prompts and adjust noise levels to balance the effect of restoration and generation." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 888, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ $\\alpha_{t}$ and $\\sigma_t$ define a noise schedule. To enhance the ability of our model to generate fine details, it also applies random noise to input images, i.e., the diffused images $\\mathrm{x}_{\\tau} = \\alpha_{\\tau}\\mathrm{x} + \\sigma_{\\tau}\\epsilon$ where $\\tau$ is the noise level that corresponds to the early steps in noise schedule.", + "bbox": [ + 75, + 441, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adopting v-prediction parameterization [54], the U-Net denoiser $f_{\\theta}$ is trained to make predictions of $\\mathrm{v}_t \\equiv \\alpha_t \\epsilon - \\sigma_t \\mathrm{x}$ . The optimization objective for LDM is as follows:", + "bbox": [ + 75, + 515, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathrm {z}, \\mathrm {x}, \\mathrm {c}, t, \\epsilon} \\left[ \\| \\mathrm {v} - f _ {\\theta} \\left(\\mathrm {z} _ {t}, \\mathrm {x} _ {\\tau}; \\mathrm {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 571, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ serves as an optional set of conditions, including text prompts and noise levels of the diffused image. During inference, the model has the flexibility to involve different text prompts and noise levels to diffusion sampling of $x_0$ , and finally decode it to produce the $\\times 4$ upscaled images.", + "bbox": [ + 75, + 601, + 468, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inflated 2D Convolution. When adapting pretrained 2D diffusion models to video tasks, it is common to inflate their 2D convolutions into the 3D convolutions [6, 17, 71, 78]. This modification allows the diffusion models to smoothly integrate new temporal layers, enabling them to capture and encode temporal information within the pretrained model.", + "bbox": [ + 75, + 676, + 468, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, we explore a new avenue to create a text-guided video diffusion model for VSR tasks, starting with the pretrained $\\mathrm{SD} \\times 4$ Upscaler [2]. To process video data, we first modify its network structure by inflating the 2D convolutions into 3D convolutions, and then initialize our network with this upscaler to inherit its enhancement capabilities. Our goal is to transfer the knowledge learned from image upscaling to video enhancement, enabling more efficient training. In the following, we will describe our", + "bbox": [ + 75, + 766, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local-global strategy within the LDM framework to achieve temporal coherence while harnessing the capability of diffusion for VSR tasks.", + "bbox": [ + 498, + 441, + 892, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Local Consistency within Video Segments", + "text_level": 1, + "bbox": [ + 498, + 498, + 856, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To apply the pretrained text-to-image SD model to video-related tasks, existing video diffusion models typically employ techniques such as 3D convolutions [6, 66], temporal attention [6, 17, 66, 68]sx, and cross-frame attention [71] to ensure temporal consistency.", + "bbox": [ + 496, + 523, + 892, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finetuning Temporal U-Net. Following the existing methods [6, 66], we introduce additional temporal layers into the pretrained image model and learn the local consistency constraint within video segments. As illustrated in Fig. 2, in the modified temporal U-Net, we opt for temporal attention and 3D residual blocks based on 3D convolutions to serve as our temporal layers, and insert them within the pretrained spatial layers. The temporal attention layer performs self-attention along the temporal dimension and focuses on all local frames. Additionally, we add Rotary Position Embedding (RoPE) [59] into the temporal layers to provide the model with positional information for time.", + "bbox": [ + 496, + 599, + 892, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our temporal layers are trained using the same noise schedule as the pretrained image model. Importantly, we keep the pretrained spatial layers fixed during training and only optimize the inserted temporal layers using Eq. (1). An essential benefit of this strategy is that it allows us to leverage the pretrained spatial layers that were learned from a huge, high-quality image dataset. This enables us to concentrate our training efforts on refining the temporal layers.", + "bbox": [ + 496, + 781, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2538", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finetuning Temporal VAE-Decoder. Even after finetuning the U-Net on video data, the VAE-Decoder within the LDM framework, which is trained on images only, still tends to produce flickering artifacts when decoding a latent sequence. To mitigate this issue, we introduce additional temporal 3D residual blocks into the VAE-Decoder to enhance low-level consistency. Furthermore, the diffusion denoising process in the U-Net often introduces color shifts, a problem also encountered by other diffusion-based restoration networks [62, 83]. To tackle this, we condition the input videos with a Spatial Feature Transform (SFT) layer [63], which employs the inputs to transform the features of the first layer of the VAE-Decoder. This allows the input videos to provide low-frequency information, such as color, to strengthen the color fidelity of the output results.", + "bbox": [ + 75, + 90, + 472, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to the training of the Temporal U-Net, we keep the pretrained spatial layers unchanged and only train the newly added temporal layers. These temporal layers are trained on video data using a hybrid loss consisting of L1 loss, LPIPS perceptual loss [86], and an adversarial loss employing a temporal PatchGAN [89] discriminator. As demonstrated by the ablation study in Fig. 6 and Table 2, this step is crucial for achieving favorable results.", + "bbox": [ + 75, + 315, + 472, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Global Consistency cross Video Segments", + "text_level": 1, + "bbox": [ + 76, + 441, + 431, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The trained temporal layers within the LDM are limited to processing local sequences (e.g., eight frames in our U-Net setting), making it impossible to enforce global consistency constraints across video segments. Prior studies [9, 10, 88, 89] have already showcased the benefits of flow-guided long-term propagation in enhancing temporal consistency for video restoration tasks. However, their performance gains are often observed when dealing with long video sequences that provide long-term information. Unfortunately, these methods are not well-suited for diffusion models due to memory constraints, which typically restrict them to processing short video clips.", + "bbox": [ + 75, + 465, + 472, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training-Free Recurrent Latent Propagation. We introduce a training-free flow-guided recurrent propagation module within the latent space. This module ensures global temporal coherence for long input videos, involving bidirectional propagation in the forward and backward directions. Here, we elaborate on the forward propagation, and backward propagation follows the same process.", + "bbox": [ + 75, + 645, + 472, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an input low-resolution video, we first adopt RAFT [58] to estimate optical flow, with its resolution exactly matching the latent resolution, thus no need for resizing. We then check the validity of the estimated flow by evaluating forward-backward consistency error [43]:", + "bbox": [ + 75, + 750, + 472, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i - 1 \\rightarrow i} (p) = \\left\\| f _ {i - 1 \\rightarrow i} (p) + f _ {i \\rightarrow i - 1} (p + f _ {i - 1 \\rightarrow i} (p)) \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 828, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ denotes the position of the last frame latent, $f_{i-1\\rightarrow i}$ and $f_{i\\rightarrow i-1}$ are forward and backward flow respectively. As", + "bbox": [ + 76, + 869, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e554ca14d0af0e594e5c0fffae3a4a8f48768cbeb111915ccd507d03dc193b4d.jpg", + "image_caption": [ + "Figure 3. An illustration of flow-guided recurrent latent propagation. Without requiring any learning, this module can achieve coherence across video segments via long-term latent propagation and aggregation. It relies on optical flow validity determined by forward-backward consistency error [43]. Only latent positions with low consistency errors will be propagated, while those with high errors, marked with a red dot, are not." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 885, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shown in Fig. 3, only latents with small consistency error will be propagated, which can be thought of as an occlusion mask, $M: E_{i-1 \\to i}(p) < \\delta$ , where $\\delta$ is a threshold. Let $z_i$ be the latent feature for the $i$ -th frame at diffusion step $t$ , we update the predicted $\\hat{z}_0$ as $\\widetilde{\\hat{z}_0}$ . The process of recurrent latent propagation and aggregation is expressed as:", + "bbox": [ + 496, + 344, + 893, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i} = \\left[ \\mathcal {W} (\\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i - 1}, f _ {i \\rightarrow i - 1}) * \\beta + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - \\beta) \\right] * M \\\\ + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - M), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 443, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}(\\cdot)$ denotes warping operation (with nearest mode), and $\\beta \\in [0,1]$ serves as a fusion weight for aggregating the warped latent from the previous $(i - 1)^{th}$ frame to the current $i^{th}$ frame, where smaller values tend to preserve current information, while larger values favor the propagated information. We set $\\beta$ to 0.5 by default.", + "bbox": [ + 496, + 494, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is not necessary to apply this module at each diffusion step during the inference process. Instead, we can choose $T^{*}$ steps for latent propagation and aggregation. When dealing with minor video jitter, one can opt to integrate this module early in the diffusion denoising process, while for severe video jitter, such as AIGC videos, it is preferable to execute this module later in the denoising process.", + "bbox": [ + 496, + 585, + 893, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference with Additional Conditions", + "text_level": 1, + "bbox": [ + 500, + 698, + 823, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can further adjust additional conditions of text prompts and noise levels in Upscale-A-Video to influence the diffusion denoising process. Text prompts can guide the generation of texture details, such as animal fur, as shown in Fig. 7. Besides, adjusting the noise level allows us to balance the model's restoration and generation abilities, with smaller values favoring restoration and larger values promoting the generation of more details (see comparisons in Fig. 7). We also adopt Classifier-Free Guidance (CFG) [23] during inference, which can significantly enhance the impact of both text prompts and noise levels, helping produce high-quality videos with finer details.", + "bbox": [ + 496, + 720, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2539", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1f65d759e43ae0346f3e82678727a86eb4b688d650a6803a92cfa5d5255868e1.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on different VSR benchmarks from diverse sources, i.e., synthetic (SPMCS, UDM10, REDS30, YouHQ40), real (VideoLQ), and AIGC (AIGC30) data. The best and second performances are marked in red and orange, respectively. $E_{\\text{warp}}^*$ denotes $E_{\\text{warp}} (\\times 10^{-3})$ ." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsReal-ESRGAN [65]SD × 4 Upscaler [2]ResShift [83]StableSR [62]RealVSR [80]DBVSR [47]RealBasicVSR [10]Ours
SPMCSPSNR ↑22.8923.1923.2722.7123.8824.2824.5125.32
SSIM ↑0.6690.6310.6670.6570.6810.7260.7170.741
LPIPS ↓0.2380.3040.2570.2310.4370.3020.1980.222
E*arp ↓1.3645.0084.9424.8150.2941.3600.5590.367
UDM10PSNR ↑27.1328.0727.6226.4527.3829.6029.1130.79
SSIM ↑0.8430.8110.8270.8250.8250.8800.8760.878
LPIPS ↓0.1900.1860.2220.1810.2780.1550.1720.133
E*arp ↓1.4621.7102.1962.7970.5311.9430.6020.446
REDS30PSNR ↑22.4022.9823.0023.7223.0524.3723.9124.41
SSIM ↑0.5910.5720.5800.6350.6030.6330.6360.631
LPIPS ↓0.3030.3990.3690.3520.6580.5880.2490.335
E*arp ↓3.6583.7534.1311.6450.3789.6591.5571.278
YouHQ40PSNR ↑24.3719.7123.7724.5324.1925.3724.0925.83
SSIM ↑0.7100.5790.6540.7110.6950.7190.6890.733
LPIPS ↓0.2720.4420.3760.2710.4840.4300.3060.268
E*arp ↓1.8563.3994.4261.5290.4851.1491.0520.737
VideoLQCLIP-IQA ↑0.3600.1580.4300.3440.2110.2740.3870.530
MUSIQ ↑49.4826.2140.9544.2324.5229.1555.3357.99
DOVER ↑7.1612.8844.6796.7832.5313.6287.5627.811
AIGC30CLIP-IQA ↑0.4300.3290.5690.4670.2760.2900.5650.674
MUSIQ ↑47.0935.3043.3244.9324.3927.2258.8757.66
DOVER ↑9.7105.6467.0429.6683.2853.52310.6811.67
", + "bbox": [ + 86, + 136, + 883, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 448, + 209, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation", + "text_level": 1, + "bbox": [ + 76, + 473, + 339, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets. We train our Upscale-A-Video using the following datasets: 1) The subset of WebVid10M [5] contains around $335\\mathrm{K}$ video-text pairs with each resolution around $336\\times 596$ , which is commonly used in training video diffusion models [6, 11, 17, 84]; 2) YouHQ dataset. Due to the lack of high-quality video data for training, we additionally collect a large-scale high-definition $(1080\\times 1920)$ dataset from YouTube, containing around $37\\mathrm{K}$ video clips with diverse scenarios, i.e., street view, landscape, animal, human face, static object, nighttime scene, and among others. Training on such high-quality data further enhances the generation ability of our model for real-world VSR. Following the degradation pipeline of RealBasicVSR [10], we generate the LQ-HQ video pairs for training.", + "bbox": [ + 75, + 496, + 468, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing Datasets. We construct four synthetic testing datasets (i.e., SPMCS [81], UDM10 [57], REDS30 [45], and YouHQ40), which follow the same degradation pipeline in training to generate LQ videos. We split YouHQ40 test set from the proposed YouHQ dataset, containing 40 videos. Additionally, we evaluate the models on a real-world dataset (i.e., VideoLQ [10]) and an AIGC dataset (i.e., AIGC30) that collects 30 AI-generated videos by popular text-to-video generation models [1, 3, 6, 17, 25, 41, 68, 84].", + "bbox": [ + 75, + 705, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. Our Upscale-A-Video is trained on 32 NVIDIA A100-80G GPUs with a batch size of 384. The training data is cropped to $80 \\times 80$ with a length of 8. The learning rate is set to $1 \\times 10^{-4}$ using the Adam [34]", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimizer. We first train the U-Net model on both WebVid10M [5] and YouHQ for 70K iterations. Then we train another 10K iterations on YouHQ only. Since there are no text prompts for YouHQ, we use the null prompts during training. In this way, our model can handle VSR with either LQ inputs and prompts or LQ inputs only, leading to a more flexible use in practice. As for VAE-Decoder finetuning, we follow StableSR [62] to first generate 100K synthetic LQ-HQ video pairs on WebVid10M [5] and YouHQ, and the finetuned U-Net model is adopted to generate the corresponding latent codes for the LQ videos.", + "bbox": [ + 496, + 449, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt different metrics to evaluate both the frame quality as well as temporal coherency of the generated results. For synthetic datasets with LQ-HQ pairs, we employ PSNR, SSIM, LPIPS [86], and flow warping error [35] $E_{\\text{warp}}^*$ for evaluation. For real-world and AIGC test data, since no ground-truth videos are available, we conduct our evaluation on commonly used non-reference metrics, i.e., CLIP-IQA [61], MUSIQ [33], and DOVER [70].", + "bbox": [ + 496, + 614, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 743, + 638, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare Upscale-A-Video with several state-of-the-art methods, including Real-ESRGAN [65], SD $\\times 4$ Upscaler [2], ResShift [83], StableSR [62], RealVSR [80], DBVSR [47], and RealBasicVSR [10].", + "bbox": [ + 496, + 765, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation. As shown in Table 1, our approach achieves the highest PSNR across all four synthetic datasets, suggesting its outstanding reconstruction ability. Moreover, our approach achieves the lowest LPIPS scores on both UDM10 and YouHQ40, indicating the high percep", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2540", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/714a240ff7c09fc906f532ce5525969dc40879d00ad16590e2b3f7c7c1afaa89.jpg", + "image_caption": [ + "\"Tourists and brick walls\"" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 263, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eddef5a15146430ba9ddd29f83a0a221ba5a69bde2a4e821d590b8a4a3647c4c.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 264, + 90, + 367, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/212c789eb3a9b7dec78947c794bde3553b778d20c35ca73e0c14abd6d83b6ca3.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 369, + 90, + 470, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05d86810a377a8eaf51e1feebbf934124069df3d4d51dc315fab1b220b81a4b4.jpg", + "image_caption": [ + "SD $\\times 4$ Upscaler" + ], + "image_footnote": [], + "bbox": [ + 472, + 90, + 575, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9b35c85201bf7d73e5f1f20d4ca57f08ed6c1627afe6afdcf12d655ee62a0ce.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 576, + 90, + 678, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dce61d072a26de0e18f29feca9094c4fabce8b95206a513cb9298de879c4d96.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 679, + 90, + 779, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c2379282043e0eb315e1c4f5b42b51479528eb64d1bf478763ff4f65d35ab85.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 781, + 90, + 883, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20dbb699894d42e0912690a79591f0b79737db4609f3e044c01a92d0154ae452.jpg", + "image_caption": [ + "\"A koala on the tree\"" + ], + "image_footnote": [], + "bbox": [ + 83, + 186, + 263, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/827e11544d132b366dcaf230ae35c56713b87de86b5c4743687e552f36bde470.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 264, + 186, + 367, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d122ffb705c49dde6220795713418555d439bd9fefda338e72e89d117925940.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 369, + 186, + 470, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bb9e1f04ef26e24850235a82b8dedc942adb69b967ae3438e696fe17cc64d7f.jpg", + "image_caption": [ + "SD $\\times 4$ Upscaler" + ], + "image_footnote": [], + "bbox": [ + 472, + 186, + 573, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c93d176b8adb15ab50df28fd45c24462ac6ece1567648e60ef7a05656ad08ae8.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 575, + 186, + 676, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ae99125339eac3e6215589bab518d5dd42e0f0e7b27d1fd3e64ca01a46089dc.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 676, + 186, + 779, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3104465d82a45808957055283046a76c8491d6068a3574c288a90b1762fcd827.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 781, + 186, + 883, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/081f2c229fa7023b838f5c7ca09f6efb280225a4d1ac68be0558012272cbc736.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on synthetic low-quality videos from REDS30 [45] and YouHQ40 datasets. Among the tested methods, only our Upscale-A-Video can recover the accurate wall structure and produce detailed koala fur. (Zoom-in for best view)", + "\"A bus and car on the street\"" + ], + "image_footnote": [], + "bbox": [ + 81, + 327, + 305, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6a32ced86f08cd95bf1f6776bc62508001e9d2ed4f77fd06019d238ff463c5d.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 308, + 327, + 452, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5eb6f7f68047ebd623560ef9a8548426f3f1a4b1e24a238acf0db166011005b.jpg", + "image_caption": [ + "Real-ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 454, + 327, + 598, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db4b517bc3358e2093f5e0a8c226e990c8cc14c51285cc0f8c76e72a3ee70dea.jpg", + "image_caption": [ + "ResShift" + ], + "image_footnote": [], + "bbox": [ + 599, + 327, + 743, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8736c44e3c24a2f064c1700ea4729efc5b5fd6b7447b2840677b1751e2469472.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 746, + 327, + 890, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d317f876620dca3768ba11ffa52042eed27fb858227c52609ea3b3515adac9ae.jpg", + "image_caption": [ + "\"A squirrel on a tree\"", + "Figure 5. Qualitative comparisons on real-world test videos in VideoLQ [10] dataset. Our Upscale-A-Video effectively leverages the advantages of the diffusion paradigm in generating high-quality results. When compared to existing methods, it notably excels in its restoration capabilities, successfully recovering the billboard word \"EAT IN or TAKEAWAY\". In particular, when guided by text prompts, Upscale-A-Video showcases promising enhanced results with more details and heightened realism. (Zoom-in for best view)" + ], + "image_footnote": [], + "bbox": [ + 81, + 484, + 305, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95e888a37c95addd25463f052af486590917aafba8aa8e7840a434448f239961.jpg", + "image_caption": [ + "SD $\\times 4$ Upscaler" + ], + "image_footnote": [], + "bbox": [ + 308, + 402, + 452, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b453a197467f9a0f6b4af182cb232bf1d6565eebf4ba2625c7814642852ce71.jpg", + "image_caption": [ + "DBVSR" + ], + "image_footnote": [], + "bbox": [ + 454, + 402, + 598, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fdff56df2f6214a0267e21ab77fb0023e87e69b8a95d1d8114100dc0132ae7cb.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 599, + 402, + 743, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcb755e3c9d39433012712208e02efc973e3f9805242ded33ffcd7cb3eab5ae7.jpg", + "image_caption": [ + "Upscale-A-Video (Ours)" + ], + "image_footnote": [], + "bbox": [ + 746, + 402, + 890, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98b891f86cabf73b2d8563139cb4688db32e2d3468bf5180b53ecdb282707d1c.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 308, + 484, + 382, + 544 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd1dd2900b107266a039550016e66dbf2e9c3e22c5c2687bec815d1c4d9c52d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 484, + 452, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29c8ee582084cd2869576e2ac439b0be44e12bf2b78700b1f394be407e40581a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 484, + 526, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38e561fe0d112776873eb06d66fc19f3679280cdac7c8958ac928c723895a433.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 484, + 598, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29fdf8793b6c9243869fda1625ecdf60d9a59336994eaf2ec4d51d32476a6ad7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 484, + 671, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ef918afa2face933a1ee01da439ebd13da0c0fec52b43a1390391a3f6c42657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 484, + 743, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24762789705e87124e885e779316755afb6a70c93df5bf8bbd2a3312cfbe668e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 484, + 816, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c0e8022ce1fcce978dd0386363f27a19467556b836b1c14b4a8fdcc117a7cb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 484, + 890, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/616e1053c2702cfb8da6153c782d1976d082bfd7f5e2e06fde2fa6154fd6daf3.jpg", + "image_caption": [ + "SD $\\times 4$ Upscaler" + ], + "image_footnote": [], + "bbox": [ + 308, + 556, + 382, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ba2e51027b505b2bbdfe8111cc07e56dffd5f5f6c9bcb7f1c500090bfde9fcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 556, + 452, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb1dfd847b81eef6e9e3cf4a67735a9ec4ef16bc8303d3385a6d2890c459020.jpg", + "image_caption": [ + "DBVSR" + ], + "image_footnote": [], + "bbox": [ + 454, + 556, + 526, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdd2c87d4c2c464d3012f95e72c15a1c6c8ae3716d35b9f7a93a7607c56d0041.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 556, + 598, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/966a5df8a661d33bd8498fbe3cdde694b1fd0c37294b2fb2746c971a64072ed2.jpg", + "image_caption": [ + "ResShift", + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 599, + 556, + 671, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1b21395b8afc2688462996ec5484f6e49c66bb3662758fa92088730dbba463a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 556, + 743, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57edc61f47e511db960555e6de5a08803c415503a02bb642715fd4bdd98c56b.jpg", + "image_caption": [ + "Upscale-A-Video (Ours)" + ], + "image_footnote": [], + "bbox": [ + 754, + 556, + 816, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tual quality of our generated results. In addition, our approach further obtains the highest CLIP-IQA and DOVER scores on both the real dataset [10] and AIGC videos. The superiority across datasets from various sources demonstrates the effectiveness of our approach.", + "bbox": [ + 75, + 709, + 468, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We present visual results on both synthetic [45, 57, 81] and real-world [10] videos in Fig. 4 and Fig. 5, respectively. It is observed that our Upscale-A-Video significantly outperforms existing CNN-and diffusion-based approaches in both artifact removal and details generation. Specifically, Upscale-A-Video is capable of generating more natural details in Fig. 4, and success", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fully recovers the words \"EAT IN or TAKEAWAY\" on the billboard in the first example of Fig. 5, while other methods generate blurry or distorted results. Note that SD $\\times 4$ Upscaler shows less effectiveness for the real-world VSR since it does not consider mixed degradation in its training data.", + "bbox": [ + 496, + 709, + 890, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Consistency. Benefiting from our local-global temporal strategy, Upscale-A-Video achieves the best optical flow error score on UDM10 and the second-best scores on REDS30, SPMCS and YouHQ40, remarkably outperforming other diffusion-based approaches, and even beats the strong CNN-based VSR methods, i.e., RealBasicVSR and DBVSR. We also visualize the temporal profile in", + "bbox": [ + 496, + 795, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2541", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fffd9bb765ccfbb25cae668e028067edea691d067150bedc69d1f045d83d2e8.jpg", + "image_caption": [ + "Figure 6. Comparison of temporal profile. We examine a row and track changes over time. The profile from existing methods and Exp. (a) and (c) (w/o latent propagation module) exhibit noise, suggesting the presence of flickering artifacts. Even with the finetuned decoder, the profile of Exp. (c) still displays some discontinuities. Thanks to long-term latent propagation and aggregation, the profile of our Upscale-A-Video exhibits a more seamless and smoother transition." + ], + "image_footnote": [], + "bbox": [ + 83, + 75, + 267, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad165445da421d911bbb003fe30da0fa97ca0db6ca21c1a6395e881ac543629e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 883, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d2a6717a4dbd376d201101ebaa19c3bd6acb56481cf923643252cd188b9b0e0c.jpg", + "table_caption": [ + "Table 2. Ablation study of finetuned VAE-Decoder and propagation module on YouHQ40." + ], + "table_footnote": [], + "table_body": "
Exp.ft-VAE-Dec.Latent Prop.PSNR↑SSIM↑E*warp↓
(a)23.820.63852.398
(b)✓25.470.72151.815
(c)✓25.750.73280.842
(d)✓✓25.830.73260.737
", + "bbox": [ + 86, + 340, + 452, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 472, + 230, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Finetuned VAE-Decoder. We first investigate the significance of the fine-tuned VAE-Decoder. As shown in Table 2, replacing our finetuned VAE-Decoder with the original decoder leads to worse PSNR, SSIM, and $E_{\\text{warp}}^*$ . Particularly, the increase of $E_{\\text{warp}}^*$ from 0.737 to 1.815 indicates a significant deterioration of temporal coherency. The comparisons in Fig. 6 also suggest inferior temporal consistency without the finetuned VAE-Decoder.", + "bbox": [ + 75, + 494, + 468, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Propagation Module. In addition to fine-tuning VAE-Decoder, our proposed flow-guided recurrent latent propagation module further enhances the stability of long videos. As shown in Table 2, adopting the propagation module can further reduce $E_{\\text{warp}}^*$ error, effectively improving temporal consistency while maintaining high PSNR. Similar phenomena can also be observed in the temporal profile in Fig. 6, showing a more seamless transition.", + "bbox": [ + 75, + 616, + 468, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Prompt. Upscale-A-Video is trained on video data with either labeled prompts or null prompts and thus can handle both scenarios. We examine the use of classifier-free guidance [23] to improve the visual quality during sampling. As shown in Figure 7, compared with the null prompt, adopting proper text prompts can significantly enhance the perceptual quality with more faithful details.", + "bbox": [ + 75, + 736, + 468, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noise Level. It is observed that the level of noise added to the input can affect the performance of our approach. As shown in Fig. 7, with low noise level, the generated results tend to be suboptimal with blurry details. However, a noise", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/048643498f9fe39cf18c8f196b96380e5496c17029ea42efea84550cdf51c160.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 503, + 309, + 627, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cab392bf147f095abf9a3f2520a0c016aeeb9e8218cc513ba0da073f9a4fb69e.jpg", + "image_caption": [ + "w/o Text Prompt" + ], + "image_footnote": [], + "bbox": [ + 632, + 309, + 756, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb750d75bae376c0cbaeecb6b580833beda50d4d131a8fe15d1dc66cf693f4cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 309, + 885, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dba03b83ee73b8209438e44e1a1a3bae2db21bc0eaa790fa509f68f2bfd39e24.jpg", + "image_caption": [ + "Fig. 6. It is observed that our approach achieves superior performance with a more seamless and smoother transition.", + "Noise Level 0", + "Figure 7. Comparison of various conditions of text prompts and different levels of noise. (Zoom-in for best view)" + ], + "image_footnote": [], + "bbox": [ + 503, + 407, + 627, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ec99efab4d96c444171af13754a9a436fb06e11d3a45240827e8d9b47e157d3.jpg", + "image_caption": [ + "Noise Level 50" + ], + "image_footnote": [], + "bbox": [ + 632, + 407, + 756, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eff9969acf979371b7d3e1bdb538422ec584abb85f961a777c264779d1466a3a.jpg", + "image_caption": [ + "w/ Text Prompt", + "Noise Level 150" + ], + "image_footnote": [], + "bbox": [ + 759, + 407, + 885, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "level that is too large may result in oversharpening.", + "bbox": [ + 500, + 546, + 836, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 617, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While diffusion models have achieved impressive performance on a wide range of image tasks, their applications on video tasks, particularly real-world VSR remain challenging and understudied. In this paper, we present Upscale-A-Video, a novel approach to exploit image diffusion prior for real-world VSR while avoiding temporal discontinuities drawn from the inherent randomness during sampling process. Specifically, we enhance the temporal coherence by proposing a novel local-global temporal strategy within the latent diffusion framework. We additionally devote our efforts to achieving a trade-off between fidelity and quality by enabling texture creation with text prompts and noise level control, further facilitating the practical use in real-world scenarios. We believe that our exploration would lay a good foundation for future works.", + "bbox": [ + 496, + 599, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). The project is also supported by Shanghai AI Laboratory.", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2542", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pika Labs. https://www.pika.art/, 2023.6", + "[2] Stable Diffusion x4 Upscaler. https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler,2023.2,3,4,6", + "[3] Zeroscope V2 XL. https://replicate.com/anotherjesse/zeroscope-v2-x1, 2023.6", + "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6", + "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 4, 6", + "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2", + "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2", + "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5", + "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, 2022. 2, 3, 5, 6, 7", + "[11] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 6", + "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 3", + "[13] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In NeurIPS, 2022. 3", + "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical flow-guided attention for consistent text-to-video editing. arXiv preprint arXiv:2310.05922, 2023. 2, 3", + "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3", + "[16] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 3", + "[17] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation:" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A noise prior for video diffusion models. In ICCV, 2023. 2, 4, 6", + "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2", + "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 3", + "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3", + "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPSW, 2022. 5, 8", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3", + "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen Video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 6", + "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022.", + "[27] Yaosi Hu, Zhenzhong Chen, and Chong Luo. LaMD: Latent motion diffusion for video generation. arXiv preprint arXiv:2304.11603, 2023. 3", + "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2", + "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020.", + "[30] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020.", + "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 2", + "[32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 3", + "[33] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In ICCV, 2021. 6", + "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2543", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6", + "[36] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2", + "[37] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. In NeurIPS, 2022. 2", + "[38] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2", + "[39] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. In IEEE TPAMI, 2013. 2", + "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. VDT: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3", + "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 6", + "[42] Kangfu Mei and Vishal Patel. VIDM: Video implicit diffusion models. In AAAI, 2023. 3", + "[43] Simon Meister, Junhwa Hur, and Stefan Roth. UnFlow: Unsupervised learning of optical flow with a bidirectional census loss. In AAAI, 2018. 5", + "[44] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3", + "[45] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 2, 6, 7", + "[46] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3", + "[47] Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. In ICCV, 2021. 6", + "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. FateZero: Fusing attentions for zero-shot text-based video editing. ICCV, 2023. 3", + "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hshmat Sahak, Daniel Watson, Chitwan Sahara, and David Fleet. Denoising diffusion probabilistic models for robust image super-resolution in the wild. arXiv preprint arXiv:2302.07864, 2023. 2, 3", + "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 3", + "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 4", + "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3", + "[56] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In ICLR, 2023. 3", + "[57] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 6, 7", + "[58] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5", + "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3", + "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6", + "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 5, 6", + "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5", + "[64] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2", + "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In ICCVW, 2021. 6", + "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 2, 4", + "[67] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2544", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 4, 6", + "[69] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2023. 3", + "[70] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023. 6", + "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 4", + "[72] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. DiffIR: Efficient diffusion model for image restoration. ICCV, 2023. 3", + "[73] Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao Dong, and Ying Shan. Mitigating artifacts in real-world video super-resolution models. In AAAI, 2023. 2, 3", + "[74] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple diffusion adapter for efficient video generation. arXiv preprint arXiv:2308.09710, 2023. 3", + "[75] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. In ICCV, 2019. 2", + "[76] Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In NeurIPS, 2023. 3", + "[77] S. Yang, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3", + "[78] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender A Video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 2, 3, 4", + "[79] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 3", + "[80] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In ICCV, 2021. 2, 3, 6", + "[81] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 6, 7", + "[82] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. ResShift: Efficient diffusion model for image superresolution by residual shifting. In NeurIPS, 2023. 2, 3, 5, 6", + "[84] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023.6", + "[85] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3", + "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6", + "[87] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2", + "[88] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In ICCV, 2019. 5", + "[89] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. ProPainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 5" + ], + "bbox": [ + 501, + 92, + 890, + 441 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2545", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_model.json b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c54c2d39eaeffaa8566c5cca5f6d261d1da5b54 --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_model.json @@ -0,0 +1,3280 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.13, + 0.773, + 0.177 + ], + "angle": 0, + "content": "Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.203, + 0.831, + 0.24 + ], + "angle": 0, + "content": "Shangchen Zhou* Peiqing Yang* Jianyi Wang Yihang Luo Chen Change Loy S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.242, + 0.706, + 0.256 + ], + "angle": 0, + "content": "https://shangchenzhou.com/projects/upscale-a-video" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.277, + 0.595, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.277, + 0.891, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.506, + 0.218, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.65, + 0.165, + 0.661 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.506, + 0.353, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.65, + 0.307, + 0.66 + ], + "angle": 0, + "content": "ResShift" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.506, + 0.487, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.65, + 0.442, + 0.66 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.506, + 0.619, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.65, + 0.593, + 0.661 + ], + "angle": 0, + "content": "SD \\(\\times 4\\) Upscaler" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.506, + 0.754, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.65, + 0.724, + 0.66 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.506, + 0.889, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.65, + 0.882, + 0.661 + ], + "angle": 0, + "content": "Upscale-A-Video (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.671, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Figure 1. Video super-resolution comparisons on both real-world and AI-generated videos. Our proposed Upscale-A-Video showcases excellent upscaling capabilities. By using appropriate text prompts, it achieves impressive results characterized by finer details and heightened visual realism. (Zoom-in for best view)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.729, + 0.314, + 0.745 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Text-based diffusion models have exhibited remarkable success in generation and editing, showing great promise for enhancing visual content with their generative prior. However, applying these models to video super-resolution remains challenging due to the high demands for output fidelity and temporal consistency, which is complicated by the inherent randomness in diffusion models. Our study introduces Upscale-A-Video, a text-guided latent diffusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.73, + 0.895, + 0.896 + ], + "angle": 0, + "content": "framework for video upscaling. This framework ensures temporal coherence through two key mechanisms: locally, it integrates temporal layers into U-Net and VAE-Decoder, maintaining consistency within short sequences; globally, without training, a flow-guided recurrent latent propagation module is introduced to enhance overall video stability by propagating and fusing latent across the entire sequences. Thanks to the diffusion paradigm, our model also offers greater flexibility by allowing text prompts to guide texture creation and adjustable noise levels to balance restoration and generation, enabling a trade-off between fidelity and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2535" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.152 + ], + "angle": 0, + "content": "quality. Extensive experiments show that Upscale-A-Video surpasses existing methods in both synthetic and real-world benchmarks, as well as in AI-generated videos, showcasing impressive visual realism and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.184, + 0.21, + 0.2 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.47, + 0.465 + ], + "angle": 0, + "content": "Video super-resolution (VSR) in real-world scenarios is a challenging task that aims at enhancing the quality of low-quality videos to produce high-quality results. Unlike previous works that mainly focus on either synthetic degradations [8, 9, 64] or specific camera-related degradations [80], this task is more demanding due to the need for addressing complex and unknown degradations commonly found in low-quality videos, such as downsampling, noise, blur, flickering, and video compression. In addition, maintaining the visual fidelity and ensuring temporal coherence are crucial for the perceptual quality of the output video. Although recent convolutional neural network (CNN)-based networks [10, 73] have shown success in mitigating many forms of degradation, they still fall short in producing realistic textures and details due to their limited generative capabilities, often resulting in over-smoothing, which can be observed in RealBasicVSR results shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Diffusion models [24] have exhibited impressive proficiency in generating high-quality images [49, 50, 52, 85] and videos [6, 11, 17, 25, 87]. Harnessing their generative potential, diffusion-based models have been proposed for image restoration, including both training from scratch [50, 51, 83] and fine-tuning from pretrained Stable Diffusion models [38, 62]. These methods have effectively mitigated the over-smoothing issue often observed in CNN-based models, yielding results with more realistic fine-grained details. However, adapting these diffusion priors to VSR remains a non-trivial challenge. This difficulty stems from the inherent randomness in diffusion sampling, which inevitably introduces unexpected temporal discontinuities in resulting videos. This issue is more pronounced in latent diffusion, where the VAE decoder further introduces flickering in low-level texture details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Recent efforts have been made to adapt image diffusion models for video tasks by introducing strategies for temporal consistency, which include: 1) Fine-tuning video models with temporal layers such as 3D convolution [6, 66] and temporal attention [6, 11, 66, 87]; 2) Employing zero-shot mechanisms like cross-frame attention [71, 78] and flow-guided attention [14, 18] in the pretrained models. Although these solutions significantly improve video stability, there remain two primary issues: i) The current methods, operating in the U-Net feature or latent space, struggle to maintain low-level consistency, with issues like texture flickering still present. ii) The existing temporal layers and attention mechanisms can only impose constraints on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "short, local input sequences, limiting their capacity to ensure global temporal consistency in longer videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.893, + 0.392 + ], + "angle": 0, + "content": "To tackle these issues, we adopt a local-global strategy for maintaining temporal consistency in video reconstruction, focusing on both fine-grained textures and overall consistency. On a local video clip, we explore finetuning a pretrained image \\(\\times 4\\) upscaling model [2] with additional temporal layers on the video data. Specifically, within a latent diffusion framework, we first finetune the U-Net with integrated 3D convolutions and temporal attention layers, and then tune the VAE-Decoder with video-conditioned inputs and 3D convolutions. The former significantly achieves structure stability in local sequences, while the latter further improves low-level consistency, reducing texture flickering. On a more global scale, we introduce a novel, training-free flow-guided recurrent latent propagation module. Spanning short video segments, it bidirectionally conducts frame-by-frame propagation and latent fusion during inference. Leveraging latent fusion in a long-term sequence, this module encourages overall stability for long videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.588 + ], + "angle": 0, + "content": "We further investigate the generative potential of diffusion models in the VSR task. Following the text-guided fashion, our model can leverage text prompts as an optional condition to steer the model towards producing more realistic and high-quality details, as illustrated in Fig. 1. Furthermore, we enhance the robustness of our model against heavy or unseen degradation by injecting noise into input to dilute its degradation. By adjusting the level of added noise, we can modulate the balance between restoring and generating within the diffusion model. Lower noise levels prioritize the model's restoration capabilities, while higher levels encourage generation of more refined details, thus achieving a trade-off between fidelity and quality of the output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.783 + ], + "angle": 0, + "content": "To summarize, the main contribution of our study is a practical and robust approach to real-world VSR. In particular, we conduct an exhaustive exploration of integrating a local-global temporal strategy within a latent diffusion framework, enhancing temporal coherence and generation quality. We further examine ways to improve versatility by allowing text prompts to guide texture creation, and offer control over noise levels to balance restoration and generation, thereby achieving a trade-off between fidelity and quality. Thanks to our design and pretrained generative prior, our model achieves state-of-the-art performance on existing benchmarks, showing remarkable visual realism and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.642, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Video Super-Resolution. VSR aims to restore a sequence of high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts. Most existing approaches [7-9, 28-31, 36, 37, 64, 75] assume a pre-defined degradation process [39, 45, 75, 82], and their performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.42 + ], + "angle": 0, + "content": "deteriorates significantly in real-world scenarios due to the limited generalizability. To tackle real-world VSR, recent works go beyond the traditional paradigms by assuming inputs with unknown degradations. Due to the lack of real-world paired data for training, Yang et al. [80] propose to collect HR-LR data pairs with iPhone cameras. While the VSR model trained on such data can be effective to videos captured by similar mobile cameras, its generalization capability to other devices is in doubt. Rather than relying on labor-intensive pair data collection, recent studies [10, 73] have shifted towards employing diverse degradations for data augmentation during training, demonstrating better performance in handling real-world cases. Nonetheless, it is still challenging for existing CNN-based approaches [10, 73] to generate photo-realistic textures due to the absence of generative prior. In this study, we set our sights on exploiting robust and extensive generative prior encapsulated in a pretrained image diffusion model, i.e., Stable Diffusion (SD) \\(\\times 4\\) upscaler [2]. By integrating this strong diffusion prior, our approach circumvents the need for exhaustive training from scratch and exhibits improved performance in producing detailed textures." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.473, + 0.736 + ], + "angle": 0, + "content": "Diffusion Models for Video Tasks. Following the success of text-to-image diffusion models [4, 12, 16, 20, 22, 44, 85], recent studies have ventured into the applications of video diffusion models [15, 25-27, 40-42]. Instead of training from scratch, some methods [14, 19, 48, 71, 78] focus on video generation using off-the-shelf image diffusion models [50, 85] in a zero-shot manner. To keep temporal consistency, cross attention [60] between neighboring frames and optical flow [58] warping are popular solutions adopted in these methods. While being efficient, these methods generally suffer from limited generalizability and the selection of hyperparameters can be tricky. Most recently, Blattmann et al. [6] propose to extend pretrained image diffusion models to the video domain by introducing an additional temporal dimension and fine-tuning the temporal layers only to speed up the training process. Subsequent works [21, 74] following such paradigm are capable of generating impressive video sequences. Inspired by these works, we employ a pretrained image \\(\\times 4\\) upscaling model [2] as generative prior and propose a novel local-global temporal strategy, resulting in temporally coherent outputs with faithful details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.737, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models for Restoration Tasks. With notable advancements in diffusion models [24, 46, 55, 77], numerous diffusion-based works have been proposed for image restoration. A straightforward way is to train a diffusion model conditioned on LR images from scratch [50, 51, 53, 72, 83], which, however, demands significant computational resources for training. To avoid such heavy training costs, another prevalent approach [12, 13, 32, 56, 67, 69, 76] is to incorporate constraints into the reverse diffusion process of a pretrained diffusion model. Though efficient, designing these constraints relies on pre-defined image degra" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.271 + ], + "angle": 0, + "content": "dation processes or pretrained SR models as a priori, leading to limited generalizability and inferior performance. Recent works [62, 79] go further by fine-tuning directly on a frozen pretrained diffusion model with several additional trainable layers, demonstrating impressive performance. Inspired by these recent advances, we focus on exploiting the effective diffusion prior for real-world VSR, which remains under-explored and challenging due to the temporal discontinuities caused by inherent randomness during diffusion sampling. With the proposed local-global temporal strategy, our approach is capable of generating temporally coherent results for real-world VSR." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.287, + 0.637, + 0.304 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Our objective is to develop a text-guided diffusion framework tailored for real-world VSR. The diffusion denoising process, characterized by its inherent stochastic nature, poses significant challenges when applied to video tasks. These challenges include temporal instability and the emergence of flickering artifacts, which are particularly prominent in VSR tasks involving lengthy video sequences. The complexity of these tasks lies not just in achieving temporal consistency within localized segments, but also in preserving coherence throughout the entire video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.896, + 0.703 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our framework incorporates both local and global modules in the latent diffusion model (LDM) to preserve temporal consistency both within and across video segments. Within each diffusion time step \\((t = 1,2\\dots T)\\), the video is split into segments and processed with a U-Net that includes temporal layers to ensure consistency within each segment. If the current time step falls within the user-specified global refinement steps \\((T^{*})\\), a recurrent latent propagation module is employed to improve consistency across segments during inference. Lastly, a finetuned VAE-Decoder is used to reduce remaining flickering artifacts. Thanks to the diffusion paradigm, our model exhibits remarkable versatility: 1) Input text prompts can enhance video quality further with improved realism and details. 2) User-specified noise levels provide a control over the trade-off between quality and fidelity." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.713, + 0.776, + 0.729 + ], + "angle": 0, + "content": "3.1. Preliminary: Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Pretrained Stable Diffusion Image \\(\\times 4\\) Upscaler. Our Upscale-A-Video is built upon the pretrained text-guided SD \\(\\times 4\\) Upscaler [2]. It employs the LDM framework [50], which uses an autoencoder to transform an image into a latent with an encoder \\(\\mathcal{E}\\) (\\(\\times 4\\) downsampling), and reconstructs it with a decoder \\(\\mathcal{D}\\). Conditioned on the low-resolution images \\(\\mathbf{x}\\), it learns to generate the high-quality counterparts via iterative denoising in the latent space. To achieve this, given latent samples \\(\\mathbf{z} \\sim p_{data}\\), diffused latents \\(\\mathbf{z}_t = \\alpha_t\\mathbf{z} + \\sigma_t\\boldsymbol{\\epsilon}\\) is generated by introducing Gaussian noise to the latents \\(\\mathbf{z}\\) at each diffusion step \\(t\\); where" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "2537" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.889, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.355, + 0.896, + 0.427 + ], + "angle": 0, + "content": "Figure 2. An overview of Upscale-A-Video. Upscale-A-Video processes long videos using both local and global strategies to maintain temporal coherence. It divides the video into segments and processes them using a U-Net with temporal layers for intra-segment consistency. During user-specified diffusion steps for global refinement, a recurrent latent propagation module is used to enhance inter-segment consistency. Finally, a finetuned VAE-Decoder reduces remaining flickering artifacts for low-level consistency. Our model also allows users to guide texture creation with text prompts and adjust noise levels to balance the effect of restoration and generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.516 + ], + "angle": 0, + "content": "\\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \\(\\alpha_{t}\\) and \\(\\sigma_t\\) define a noise schedule. To enhance the ability of our model to generate fine details, it also applies random noise to input images, i.e., the diffused images \\(\\mathrm{x}_{\\tau} = \\alpha_{\\tau}\\mathrm{x} + \\sigma_{\\tau}\\epsilon\\) where \\(\\tau\\) is the noise level that corresponds to the early steps in noise schedule." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Adopting v-prediction parameterization [54], the U-Net denoiser \\( f_{\\theta} \\) is trained to make predictions of \\( \\mathrm{v}_t \\equiv \\alpha_t \\epsilon - \\sigma_t \\mathrm{x} \\). The optimization objective for LDM is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.573, + 0.47, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathrm {z}, \\mathrm {x}, \\mathrm {c}, t, \\epsilon} \\left[ \\| \\mathrm {v} - f _ {\\theta} \\left(\\mathrm {z} _ {t}, \\mathrm {x} _ {\\tau}; \\mathrm {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.469, + 0.676 + ], + "angle": 0, + "content": "where \\( c \\) serves as an optional set of conditions, including text prompts and noise levels of the diffused image. During inference, the model has the flexibility to involve different text prompts and noise levels to diffusion sampling of \\( x_0 \\), and finally decode it to produce the \\( \\times 4 \\) upscaled images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Inflated 2D Convolution. When adapting pretrained 2D diffusion models to video tasks, it is common to inflate their 2D convolutions into the 3D convolutions [6, 17, 71, 78]. This modification allows the diffusion models to smoothly integrate new temporal layers, enabling them to capture and encode temporal information within the pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this study, we explore a new avenue to create a text-guided video diffusion model for VSR tasks, starting with the pretrained \\(\\mathrm{SD} \\times 4\\) Upscaler [2]. To process video data, we first modify its network structure by inflating the 2D convolutions into 3D convolutions, and then initialize our network with this upscaler to inherit its enhancement capabilities. Our goal is to transfer the knowledge learned from image upscaling to video enhancement, enabling more efficient training. In the following, we will describe our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.893, + 0.486 + ], + "angle": 0, + "content": "local-global strategy within the LDM framework to achieve temporal coherence while harnessing the capability of diffusion for VSR tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.857, + 0.517 + ], + "angle": 0, + "content": "3.2. Local Consistency within Video Segments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.893, + 0.599 + ], + "angle": 0, + "content": "To apply the pretrained text-to-image SD model to video-related tasks, existing video diffusion models typically employ techniques such as 3D convolutions [6, 66], temporal attention [6, 17, 66, 68]sx, and cross-frame attention [71] to ensure temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Finetuning Temporal U-Net. Following the existing methods [6, 66], we introduce additional temporal layers into the pretrained image model and learn the local consistency constraint within video segments. As illustrated in Fig. 2, in the modified temporal U-Net, we opt for temporal attention and 3D residual blocks based on 3D convolutions to serve as our temporal layers, and insert them within the pretrained spatial layers. The temporal attention layer performs self-attention along the temporal dimension and focuses on all local frames. Additionally, we add Rotary Position Embedding (RoPE) [59] into the temporal layers to provide the model with positional information for time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our temporal layers are trained using the same noise schedule as the pretrained image model. Importantly, we keep the pretrained spatial layers fixed during training and only optimize the inserted temporal layers using Eq. (1). An essential benefit of this strategy is that it allows us to leverage the pretrained spatial layers that were learned from a huge, high-quality image dataset. This enables us to concentrate our training efforts on refining the temporal layers." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.315 + ], + "angle": 0, + "content": "Finetuning Temporal VAE-Decoder. Even after finetuning the U-Net on video data, the VAE-Decoder within the LDM framework, which is trained on images only, still tends to produce flickering artifacts when decoding a latent sequence. To mitigate this issue, we introduce additional temporal 3D residual blocks into the VAE-Decoder to enhance low-level consistency. Furthermore, the diffusion denoising process in the U-Net often introduces color shifts, a problem also encountered by other diffusion-based restoration networks [62, 83]. To tackle this, we condition the input videos with a Spatial Feature Transform (SFT) layer [63], which employs the inputs to transform the features of the first layer of the VAE-Decoder. This allows the input videos to provide low-frequency information, such as color, to strengthen the color fidelity of the output results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.436 + ], + "angle": 0, + "content": "Similar to the training of the Temporal U-Net, we keep the pretrained spatial layers unchanged and only train the newly added temporal layers. These temporal layers are trained on video data using a hybrid loss consisting of L1 loss, LPIPS perceptual loss [86], and an adversarial loss employing a temporal PatchGAN [89] discriminator. As demonstrated by the ablation study in Fig. 6 and Table 2, this step is crucial for achieving favorable results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.433, + 0.459 + ], + "angle": 0, + "content": "3.3. Global Consistency cross Video Segments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.645 + ], + "angle": 0, + "content": "The trained temporal layers within the LDM are limited to processing local sequences (e.g., eight frames in our U-Net setting), making it impossible to enforce global consistency constraints across video segments. Prior studies [9, 10, 88, 89] have already showcased the benefits of flow-guided long-term propagation in enhancing temporal consistency for video restoration tasks. However, their performance gains are often observed when dealing with long video sequences that provide long-term information. Unfortunately, these methods are not well-suited for diffusion models due to memory constraints, which typically restrict them to processing short video clips." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.473, + 0.749 + ], + "angle": 0, + "content": "Training-Free Recurrent Latent Propagation. We introduce a training-free flow-guided recurrent propagation module within the latent space. This module ensures global temporal coherence for long input videos, involving bidirectional propagation in the forward and backward directions. Here, we elaborate on the forward propagation, and backward propagation follows the same process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.825 + ], + "angle": 0, + "content": "Given an input low-resolution video, we first adopt RAFT [58] to estimate optical flow, with its resolution exactly matching the latent resolution, thus no need for resizing. We then check the validity of the estimated flow by evaluating forward-backward consistency error [43]:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.829, + 0.47, + 0.87 + ], + "angle": 0, + "content": "\\[\nE _ {i - 1 \\rightarrow i} (p) = \\left\\| f _ {i - 1 \\rightarrow i} (p) + f _ {i \\rightarrow i - 1} (p + f _ {i - 1 \\rightarrow i} (p)) \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\( p \\) denotes the position of the last frame latent, \\( f_{i-1\\rightarrow i} \\) and \\( f_{i\\rightarrow i-1} \\) are forward and backward flow respectively. As" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.887, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 3. An illustration of flow-guided recurrent latent propagation. Without requiring any learning, this module can achieve coherence across video segments via long-term latent propagation and aggregation. It relies on optical flow validity determined by forward-backward consistency error [43]. Only latent positions with low consistency errors will be propagated, while those with high errors, marked with a red dot, are not." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.438 + ], + "angle": 0, + "content": "shown in Fig. 3, only latents with small consistency error will be propagated, which can be thought of as an occlusion mask, \\( M: E_{i-1 \\to i}(p) < \\delta \\), where \\( \\delta \\) is a threshold. Let \\( z_i \\) be the latent feature for the \\( i \\)-th frame at diffusion step \\( t \\), we update the predicted \\( \\hat{z}_0 \\) as \\( \\widetilde{\\hat{z}_0} \\). The process of recurrent latent propagation and aggregation is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.444, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i} = \\left[ \\mathcal {W} (\\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i - 1}, f _ {i \\rightarrow i - 1}) * \\beta + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - \\beta) \\right] * M \\\\ + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - M), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.586 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}(\\cdot)\\) denotes warping operation (with nearest mode), and \\(\\beta \\in [0,1]\\) serves as a fusion weight for aggregating the warped latent from the previous \\((i - 1)^{th}\\) frame to the current \\(i^{th}\\) frame, where smaller values tend to preserve current information, while larger values favor the propagated information. We set \\(\\beta\\) to 0.5 by default." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.691 + ], + "angle": 0, + "content": "It is not necessary to apply this module at each diffusion step during the inference process. Instead, we can choose \\( T^{*} \\) steps for latent propagation and aggregation. When dealing with minor video jitter, one can opt to integrate this module early in the diffusion denoising process, while for severe video jitter, such as AIGC videos, it is preferable to execute this module later in the denoising process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.699, + 0.824, + 0.714 + ], + "angle": 0, + "content": "3.4. Inference with Additional Conditions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We can further adjust additional conditions of text prompts and noise levels in Upscale-A-Video to influence the diffusion denoising process. Text prompts can guide the generation of texture details, such as animal fur, as shown in Fig. 7. Besides, adjusting the noise level allows us to balance the model's restoration and generation abilities, with smaller values favoring restoration and larger values promoting the generation of more details (see comparisons in Fig. 7). We also adopt Classifier-Free Guidance (CFG) [23] during inference, which can significantly enhance the impact of both text prompts and noise levels, helping produce high-quality videos with finer details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2539" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.135 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on different VSR benchmarks from diverse sources, i.e., synthetic (SPMCS, UDM10, REDS30, YouHQ40), real (VideoLQ), and AIGC (AIGC30) data. The best and second performances are marked in red and orange, respectively. \\( E_{\\text{warp}}^* \\) denotes \\( E_{\\text{warp}} (\\times 10^{-3}) \\)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.137, + 0.885, + 0.431 + ], + "angle": 0, + "content": "
DatasetsMetricsReal-ESRGAN [65]SD × 4 Upscaler [2]ResShift [83]StableSR [62]RealVSR [80]DBVSR [47]RealBasicVSR [10]Ours
SPMCSPSNR ↑22.8923.1923.2722.7123.8824.2824.5125.32
SSIM ↑0.6690.6310.6670.6570.6810.7260.7170.741
LPIPS ↓0.2380.3040.2570.2310.4370.3020.1980.222
E*arp ↓1.3645.0084.9424.8150.2941.3600.5590.367
UDM10PSNR ↑27.1328.0727.6226.4527.3829.6029.1130.79
SSIM ↑0.8430.8110.8270.8250.8250.8800.8760.878
LPIPS ↓0.1900.1860.2220.1810.2780.1550.1720.133
E*arp ↓1.4621.7102.1962.7970.5311.9430.6020.446
REDS30PSNR ↑22.4022.9823.0023.7223.0524.3723.9124.41
SSIM ↑0.5910.5720.5800.6350.6030.6330.6360.631
LPIPS ↓0.3030.3990.3690.3520.6580.5880.2490.335
E*arp ↓3.6583.7534.1311.6450.3789.6591.5571.278
YouHQ40PSNR ↑24.3719.7123.7724.5324.1925.3724.0925.83
SSIM ↑0.7100.5790.6540.7110.6950.7190.6890.733
LPIPS ↓0.2720.4420.3760.2710.4840.4300.3060.268
E*arp ↓1.8563.3994.4261.5290.4851.1491.0520.737
VideoLQCLIP-IQA ↑0.3600.1580.4300.3440.2110.2740.3870.530
MUSIQ ↑49.4826.2140.9544.2324.5229.1555.3357.99
DOVER ↑7.1612.8844.6796.7832.5313.6287.5627.811
AIGC30CLIP-IQA ↑0.4300.3290.5690.4670.2760.2900.5650.674
MUSIQ ↑47.0935.3043.3244.9324.3927.2258.8757.66
DOVER ↑9.7105.6467.0429.6683.2853.52310.6811.67
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.21, + 0.466 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.474, + 0.34, + 0.49 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Training Datasets. We train our Upscale-A-Video using the following datasets: 1) The subset of WebVid10M [5] contains around \\(335\\mathrm{K}\\) video-text pairs with each resolution around \\(336\\times 596\\), which is commonly used in training video diffusion models [6, 11, 17, 84]; 2) YouHQ dataset. Due to the lack of high-quality video data for training, we additionally collect a large-scale high-definition \\((1080\\times 1920)\\) dataset from YouTube, containing around \\(37\\mathrm{K}\\) video clips with diverse scenarios, i.e., street view, landscape, animal, human face, static object, nighttime scene, and among others. Training on such high-quality data further enhances the generation ability of our model for real-world VSR. Following the degradation pipeline of RealBasicVSR [10], we generate the LQ-HQ video pairs for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Testing Datasets. We construct four synthetic testing datasets (i.e., SPMCS [81], UDM10 [57], REDS30 [45], and YouHQ40), which follow the same degradation pipeline in training to generate LQ videos. We split YouHQ40 test set from the proposed YouHQ dataset, containing 40 videos. Additionally, we evaluate the models on a real-world dataset (i.e., VideoLQ [10]) and an AIGC dataset (i.e., AIGC30) that collects 30 AI-generated videos by popular text-to-video generation models [1, 3, 6, 17, 25, 41, 68, 84]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Training Details. Our Upscale-A-Video is trained on 32 NVIDIA A100-80G GPUs with a batch size of 384. The training data is cropped to \\(80 \\times 80\\) with a length of 8. The learning rate is set to \\(1 \\times 10^{-4}\\) using the Adam [34]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.614 + ], + "angle": 0, + "content": "optimizer. We first train the U-Net model on both WebVid10M [5] and YouHQ for 70K iterations. Then we train another 10K iterations on YouHQ only. Since there are no text prompts for YouHQ, we use the null prompts during training. In this way, our model can handle VSR with either LQ inputs and prompts or LQ inputs only, leading to a more flexible use in practice. As for VAE-Decoder finetuning, we follow StableSR [62] to first generate 100K synthetic LQ-HQ video pairs on WebVid10M [5] and YouHQ, and the finetuned U-Net model is adopted to generate the corresponding latent codes for the LQ videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt different metrics to evaluate both the frame quality as well as temporal coherency of the generated results. For synthetic datasets with LQ-HQ pairs, we employ PSNR, SSIM, LPIPS [86], and flow warping error [35] \\( E_{\\text{warp}}^* \\) for evaluation. For real-world and AIGC test data, since no ground-truth videos are available, we conduct our evaluation on commonly used non-reference metrics, i.e., CLIP-IQA [61], MUSIQ [33], and DOVER [70]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.744, + 0.64, + 0.76 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.826 + ], + "angle": 0, + "content": "We compare Upscale-A-Video with several state-of-the-art methods, including Real-ESRGAN [65], SD \\(\\times 4\\) Upscaler [2], ResShift [83], StableSR [62], RealVSR [80], DBVSR [47], and RealBasicVSR [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Quantitative Evaluation. As shown in Table 1, our approach achieves the highest PSNR across all four synthetic datasets, suggesting its outstanding reconstruction ability. Moreover, our approach achieves the lowest LPIPS scores on both UDM10 and YouHQ40, indicating the high percep" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2540" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.264, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.172, + 0.258, + 0.184 + ], + "angle": 0, + "content": "\"Tourists and brick walls\"" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.091, + 0.369, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.172, + 0.334, + 0.183 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.091, + 0.472, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.172, + 0.441, + 0.183 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.091, + 0.576, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.172, + 0.56, + 0.183 + ], + "angle": 0, + "content": "SD \\(\\times 4\\) Upscaler" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.091, + 0.679, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.172, + 0.661, + 0.183 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.091, + 0.781, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.172, + 0.742, + 0.183 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.091, + 0.885, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.172, + 0.84, + 0.183 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.188, + 0.264, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.269, + 0.242, + 0.28 + ], + "angle": 0, + "content": "\"A koala on the tree\"" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.188, + 0.369, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.27, + 0.338, + 0.28 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.188, + 0.472, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.27, + 0.445, + 0.28 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.188, + 0.575, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.27, + 0.564, + 0.281 + ], + "angle": 0, + "content": "SD \\(\\times 4\\) Upscaler" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.188, + 0.678, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.27, + 0.665, + 0.28 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.188, + 0.781, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.27, + 0.746, + 0.28 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.188, + 0.885, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.27, + 0.843, + 0.279 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on synthetic low-quality videos from REDS30 [45] and YouHQ40 datasets. Among the tested methods, only our Upscale-A-Video can recover the accurate wall structure and produce detailed koala fur. (Zoom-in for best view)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.328, + 0.306, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.466, + 0.295, + 0.478 + ], + "angle": 0, + "content": "\"A bus and car on the street\"" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.328, + 0.453, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.391, + 0.402, + 0.401 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.328, + 0.599, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.391, + 0.566, + 0.401 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.328, + 0.745, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.391, + 0.694, + 0.401 + ], + "angle": 0, + "content": "ResShift" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.328, + 0.892, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.391, + 0.842, + 0.401 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.486, + 0.306, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.619, + 0.267, + 0.63 + ], + "angle": 0, + "content": "\"A squirrel on a tree\"" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.403, + 0.453, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.466, + 0.422, + 0.478 + ], + "angle": 0, + "content": "SD \\(\\times 4\\) Upscaler" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.403, + 0.599, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.466, + 0.548, + 0.477 + ], + "angle": 0, + "content": "DBVSR" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.403, + 0.745, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.466, + 0.711, + 0.477 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.403, + 0.891, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.466, + 0.88, + 0.478 + ], + "angle": 0, + "content": "Upscale-A-Video (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.486, + 0.383, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.546, + 0.4, + 0.556 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.486, + 0.453, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.486, + 0.527, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.486, + 0.599, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.486, + 0.672, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.546, + 0.694, + 0.556 + ], + "angle": 0, + "content": "ResShift" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.486, + 0.745, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.486, + 0.818, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.486, + 0.891, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.557, + 0.383, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.619, + 0.422, + 0.63 + ], + "angle": 0, + "content": "SD \\(\\times 4\\) Upscaler" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.557, + 0.453, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.557, + 0.527, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.557, + 0.599, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.619, + 0.548, + 0.629 + ], + "angle": 0, + "content": "DBVSR" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.557, + 0.672, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.619, + 0.71, + 0.629 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.557, + 0.744, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.557, + 0.818, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.619, + 0.88, + 0.63 + ], + "angle": 0, + "content": "Upscale-A-Video (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons on real-world test videos in VideoLQ [10] dataset. Our Upscale-A-Video effectively leverages the advantages of the diffusion paradigm in generating high-quality results. When compared to existing methods, it notably excels in its restoration capabilities, successfully recovering the billboard word \"EAT IN or TAKEAWAY\". In particular, when guided by text prompts, Upscale-A-Video showcases promising enhanced results with more details and heightened realism. (Zoom-in for best view)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.469, + 0.785 + ], + "angle": 0, + "content": "tual quality of our generated results. In addition, our approach further obtains the highest CLIP-IQA and DOVER scores on both the real dataset [10] and AIGC videos. The superiority across datasets from various sources demonstrates the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Qualitative Evaluation. We present visual results on both synthetic [45, 57, 81] and real-world [10] videos in Fig. 4 and Fig. 5, respectively. It is observed that our Upscale-A-Video significantly outperforms existing CNN-and diffusion-based approaches in both artifact removal and details generation. Specifically, Upscale-A-Video is capable of generating more natural details in Fig. 4, and success" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.892, + 0.785 + ], + "angle": 0, + "content": "fully recovers the words \"EAT IN or TAKEAWAY\" on the billboard in the first example of Fig. 5, while other methods generate blurry or distorted results. Note that SD \\(\\times 4\\) Upscaler shows less effectiveness for the real-world VSR since it does not consider mixed degradation in its training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Temporal Consistency. Benefiting from our local-global temporal strategy, Upscale-A-Video achieves the best optical flow error score on UDM10 and the second-best scores on REDS30, SPMCS and YouHQ40, remarkably outperforming other diffusion-based approaches, and even beats the strong CNN-based VSR methods, i.e., RealBasicVSR and DBVSR. We also visualize the temporal profile in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.076, + 0.268, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.885, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 6. Comparison of temporal profile. We examine a row and track changes over time. The profile from existing methods and Exp. (a) and (c) (w/o latent propagation module) exhibit noise, suggesting the presence of flickering artifacts. Even with the finetuned decoder, the profile of Exp. (c) still displays some discontinuities. Thanks to long-term latent propagation and aggregation, the profile of our Upscale-A-Video exhibits a more seamless and smoother transition." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.31, + 0.47, + 0.338 + ], + "angle": 0, + "content": "Table 2. Ablation study of finetuned VAE-Decoder and propagation module on YouHQ40." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.342, + 0.454, + 0.426 + ], + "angle": 0, + "content": "
Exp.ft-VAE-Dec.Latent Prop.PSNR↑SSIM↑E*warp↓
(a)23.820.63852.398
(b)✓25.470.72151.815
(c)✓25.750.73280.842
(d)✓✓25.830.73260.737
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.464 + ], + "angle": 0, + "content": "Fig. 6. It is observed that our approach achieves superior performance with a more seamless and smoother transition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.232, + 0.489 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.616 + ], + "angle": 0, + "content": "Effectiveness of Finetuned VAE-Decoder. We first investigate the significance of the fine-tuned VAE-Decoder. As shown in Table 2, replacing our finetuned VAE-Decoder with the original decoder leads to worse PSNR, SSIM, and \\( E_{\\text{warp}}^* \\). Particularly, the increase of \\( E_{\\text{warp}}^* \\) from 0.737 to 1.815 indicates a significant deterioration of temporal coherency. The comparisons in Fig. 6 also suggest inferior temporal consistency without the finetuned VAE-Decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Effectiveness of Propagation Module. In addition to fine-tuning VAE-Decoder, our proposed flow-guided recurrent latent propagation module further enhances the stability of long videos. As shown in Table 2, adopting the propagation module can further reduce \\( E_{\\text{warp}}^* \\) error, effectively improving temporal consistency while maintaining high PSNR. Similar phenomena can also be observed in the temporal profile in Fig. 6, showing a more seamless transition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.469, + 0.841 + ], + "angle": 0, + "content": "Text Prompt. Upscale-A-Video is trained on video data with either labeled prompts or null prompts and thus can handle both scenarios. We examine the use of classifier-free guidance [23] to improve the visual quality during sampling. As shown in Figure 7, compared with the null prompt, adopting proper text prompts can significantly enhance the perceptual quality with more faithful details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Noise Level. It is observed that the level of noise added to the input can affect the performance of our approach. As shown in Fig. 7, with low noise level, the generated results tend to be suboptimal with blurry details. However, a noise" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.31, + 0.629, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.397, + 0.58, + 0.407 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.31, + 0.757, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.397, + 0.735, + 0.407 + ], + "angle": 0, + "content": "w/o Text Prompt" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.31, + 0.886, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.397, + 0.86, + 0.407 + ], + "angle": 0, + "content": "w/ Text Prompt" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.409, + 0.629, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.495, + 0.599, + 0.504 + ], + "angle": 0, + "content": "Noise Level 0" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.409, + 0.757, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.495, + 0.732, + 0.504 + ], + "angle": 0, + "content": "Noise Level 50" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.409, + 0.886, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.495, + 0.863, + 0.504 + ], + "angle": 0, + "content": "Noise Level 150" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.505, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Figure 7. Comparison of various conditions of text prompts and different levels of noise. (Zoom-in for best view)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.547, + 0.838, + 0.563 + ], + "angle": 0, + "content": "level that is too large may result in oversharpening." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.619, + 0.592 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.825 + ], + "angle": 0, + "content": "While diffusion models have achieved impressive performance on a wide range of image tasks, their applications on video tasks, particularly real-world VSR remain challenging and understudied. In this paper, we present Upscale-A-Video, a novel approach to exploit image diffusion prior for real-world VSR while avoiding temporal discontinuities drawn from the inherent randomness during sampling process. Specifically, we enhance the temporal coherence by proposing a novel local-global temporal strategy within the latent diffusion framework. We additionally devote our efforts to achieving a trade-off between fidelity and quality by enabling texture creation with text prompts and noise level control, further facilitating the practical use in real-world scenarios. We believe that our exploration would lay a good foundation for future works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). The project is also supported by Shanghai AI Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2542" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.42, + 0.128 + ], + "angle": 0, + "content": "[1] Pika Labs. https://www.pika.art/, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.468, + 0.17 + ], + "angle": 0, + "content": "[2] Stable Diffusion x4 Upscaler. https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler,2023.2,3,4,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.468, + 0.198 + ], + "angle": 0, + "content": "[3] Zeroscope V2 XL. https://replicate.com/anotherjesse/zeroscope-v2-x1, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.468, + 0.237 + ], + "angle": 0, + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.24, + 0.468, + 0.279 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.282, + 0.468, + 0.335 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.337, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.378, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.468, + 0.528 + ], + "angle": 0, + "content": "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.53, + 0.468, + 0.597 + ], + "angle": 0, + "content": "[11] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.468, + 0.639 + ], + "angle": 0, + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.64, + 0.468, + 0.68 + ], + "angle": 0, + "content": "[13] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.681, + 0.468, + 0.748 + ], + "angle": 0, + "content": "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical flow-guided attention for consistent text-to-video editing. arXiv preprint arXiv:2310.05922, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[16] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.861, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "A noise prior for video diffusion models. In ICCV, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPSW, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen Video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[27] Yaosi Hu, Zhenzhong Chen, and Chong Luo. LaMD: Latent motion diffusion for video generation. arXiv preprint arXiv:2304.11603, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[30] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[33] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2543" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[36] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[37] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.259, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[38] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[39] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. In IEEE TPAMI, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. VDT: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[42] Kangfu Mei and Vishal Patel. VIDM: Video implicit diffusion models. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[43] Simon Meister, Junhwa Hur, and Stefan Roth. UnFlow: Unsupervised learning of optical flow with a bidirectional census loss. In AAAI, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[44] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[45] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[46] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[47] Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. FateZero: Fusing attentions for zero-shot text-based video editing. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[51] Hshmat Sahak, Daniel Watson, Chitwan Sahara, and David Fleet. Denoising diffusion probabilistic models for robust image super-resolution in the wild. arXiv preprint arXiv:2302.07864, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[56] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[57] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[58] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[64] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In ICCVW, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[67] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2544" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[68] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[69] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[70] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.345 + ], + "angle": 0, + "content": "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[72] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. DiffIR: Efficient diffusion model for image restoration. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[73] Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao Dong, and Ying Shan. Mitigating artifacts in real-world video super-resolution models. In AAAI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[74] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple diffusion adapter for efficient video generation. arXiv preprint arXiv:2308.09710, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[75] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[76] Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[77] S. Yang, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[78] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender A Video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[79] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[80] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In ICCV, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[81] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[82] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[83] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. ResShift: Efficient diffusion model for image superresolution by residual shifting. In NeurIPS, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[84] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[85] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[87] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[88] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[89] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. ProPainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2545" + } + ] +] \ No newline at end of file diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_origin.pdf b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9298cc088088ee42e4083dd827e80b932eac73fb --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/415b5691-a710-48d9-b0ac-0daaf56a288f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52d7887c4518f762cdfcb9c0e6932b1fd40b0e336a831d264019a66b0a7db1ca +size 8496800 diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/full.md b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7705fe676a192e14ddbcb012327de6297ba0492 --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/full.md @@ -0,0 +1,423 @@ +# Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution + +Shangchen Zhou* Peiqing Yang* Jianyi Wang Yihang Luo Chen Change Loy S-Lab, Nanyang Technological University + +https://shangchenzhou.com/projects/upscale-a-video + +![](images/e3df4f0b8c0bbe58fd8efa2da63b6b1ada06c32e315115a9bc72d62369fd25ba.jpg) + +![](images/4e954c433727a5bf1e749f520a83487b701de7d06eac5702f21aeda33c2930c8.jpg) + +![](images/d3cfd7d11cb82a7723e261bfdce2ce4ad291e4f2728b9ec5c30531609ddce7d1.jpg) +Input + +![](images/e3415af2147966bf0abd49be26aad9fc8ba93666e3cd457ea2505bde4a9c8da6.jpg) +ResShift + +![](images/27a3154fefffad72ac0dc582fed48b1712597ef56f9a69c0b3bae079b13cf69c.jpg) +StableSR + +![](images/b0786d235379a2426867ab856975e2b308c9349568aabfe900ed1ebe549c4401.jpg) +SD $\times 4$ Upscaler + +![](images/a6be451dfed78bae9943bac056907db1afb342da10fd65488055acca444e155f.jpg) +RealBasicVSR +Figure 1. Video super-resolution comparisons on both real-world and AI-generated videos. Our proposed Upscale-A-Video showcases excellent upscaling capabilities. By using appropriate text prompts, it achieves impressive results characterized by finer details and heightened visual realism. (Zoom-in for best view) + +![](images/e8d44988b4926a90ed16197a68632974c3dbdce8b9b2218d608e58c5aa44f070.jpg) +Upscale-A-Video (Ours) + +# Abstract + +Text-based diffusion models have exhibited remarkable success in generation and editing, showing great promise for enhancing visual content with their generative prior. However, applying these models to video super-resolution remains challenging due to the high demands for output fidelity and temporal consistency, which is complicated by the inherent randomness in diffusion models. Our study introduces Upscale-A-Video, a text-guided latent diffusion + +framework for video upscaling. This framework ensures temporal coherence through two key mechanisms: locally, it integrates temporal layers into U-Net and VAE-Decoder, maintaining consistency within short sequences; globally, without training, a flow-guided recurrent latent propagation module is introduced to enhance overall video stability by propagating and fusing latent across the entire sequences. Thanks to the diffusion paradigm, our model also offers greater flexibility by allowing text prompts to guide texture creation and adjustable noise levels to balance restoration and generation, enabling a trade-off between fidelity and + +quality. Extensive experiments show that Upscale-A-Video surpasses existing methods in both synthetic and real-world benchmarks, as well as in AI-generated videos, showcasing impressive visual realism and temporal consistency. + +# 1. Introduction + +Video super-resolution (VSR) in real-world scenarios is a challenging task that aims at enhancing the quality of low-quality videos to produce high-quality results. Unlike previous works that mainly focus on either synthetic degradations [8, 9, 64] or specific camera-related degradations [80], this task is more demanding due to the need for addressing complex and unknown degradations commonly found in low-quality videos, such as downsampling, noise, blur, flickering, and video compression. In addition, maintaining the visual fidelity and ensuring temporal coherence are crucial for the perceptual quality of the output video. Although recent convolutional neural network (CNN)-based networks [10, 73] have shown success in mitigating many forms of degradation, they still fall short in producing realistic textures and details due to their limited generative capabilities, often resulting in over-smoothing, which can be observed in RealBasicVSR results shown in Fig. 1. + +Diffusion models [24] have exhibited impressive proficiency in generating high-quality images [49, 50, 52, 85] and videos [6, 11, 17, 25, 87]. Harnessing their generative potential, diffusion-based models have been proposed for image restoration, including both training from scratch [50, 51, 83] and fine-tuning from pretrained Stable Diffusion models [38, 62]. These methods have effectively mitigated the over-smoothing issue often observed in CNN-based models, yielding results with more realistic fine-grained details. However, adapting these diffusion priors to VSR remains a non-trivial challenge. This difficulty stems from the inherent randomness in diffusion sampling, which inevitably introduces unexpected temporal discontinuities in resulting videos. This issue is more pronounced in latent diffusion, where the VAE decoder further introduces flickering in low-level texture details. + +Recent efforts have been made to adapt image diffusion models for video tasks by introducing strategies for temporal consistency, which include: 1) Fine-tuning video models with temporal layers such as 3D convolution [6, 66] and temporal attention [6, 11, 66, 87]; 2) Employing zero-shot mechanisms like cross-frame attention [71, 78] and flow-guided attention [14, 18] in the pretrained models. Although these solutions significantly improve video stability, there remain two primary issues: i) The current methods, operating in the U-Net feature or latent space, struggle to maintain low-level consistency, with issues like texture flickering still present. ii) The existing temporal layers and attention mechanisms can only impose constraints on + +short, local input sequences, limiting their capacity to ensure global temporal consistency in longer videos. + +To tackle these issues, we adopt a local-global strategy for maintaining temporal consistency in video reconstruction, focusing on both fine-grained textures and overall consistency. On a local video clip, we explore finetuning a pretrained image $\times 4$ upscaling model [2] with additional temporal layers on the video data. Specifically, within a latent diffusion framework, we first finetune the U-Net with integrated 3D convolutions and temporal attention layers, and then tune the VAE-Decoder with video-conditioned inputs and 3D convolutions. The former significantly achieves structure stability in local sequences, while the latter further improves low-level consistency, reducing texture flickering. On a more global scale, we introduce a novel, training-free flow-guided recurrent latent propagation module. Spanning short video segments, it bidirectionally conducts frame-by-frame propagation and latent fusion during inference. Leveraging latent fusion in a long-term sequence, this module encourages overall stability for long videos. + +We further investigate the generative potential of diffusion models in the VSR task. Following the text-guided fashion, our model can leverage text prompts as an optional condition to steer the model towards producing more realistic and high-quality details, as illustrated in Fig. 1. Furthermore, we enhance the robustness of our model against heavy or unseen degradation by injecting noise into input to dilute its degradation. By adjusting the level of added noise, we can modulate the balance between restoring and generating within the diffusion model. Lower noise levels prioritize the model's restoration capabilities, while higher levels encourage generation of more refined details, thus achieving a trade-off between fidelity and quality of the output. + +To summarize, the main contribution of our study is a practical and robust approach to real-world VSR. In particular, we conduct an exhaustive exploration of integrating a local-global temporal strategy within a latent diffusion framework, enhancing temporal coherence and generation quality. We further examine ways to improve versatility by allowing text prompts to guide texture creation, and offer control over noise levels to balance restoration and generation, thereby achieving a trade-off between fidelity and quality. Thanks to our design and pretrained generative prior, our model achieves state-of-the-art performance on existing benchmarks, showing remarkable visual realism and temporal consistency. + +# 2. Related Work + +Video Super-Resolution. VSR aims to restore a sequence of high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts. Most existing approaches [7-9, 28-31, 36, 37, 64, 75] assume a pre-defined degradation process [39, 45, 75, 82], and their performance + +deteriorates significantly in real-world scenarios due to the limited generalizability. To tackle real-world VSR, recent works go beyond the traditional paradigms by assuming inputs with unknown degradations. Due to the lack of real-world paired data for training, Yang et al. [80] propose to collect HR-LR data pairs with iPhone cameras. While the VSR model trained on such data can be effective to videos captured by similar mobile cameras, its generalization capability to other devices is in doubt. Rather than relying on labor-intensive pair data collection, recent studies [10, 73] have shifted towards employing diverse degradations for data augmentation during training, demonstrating better performance in handling real-world cases. Nonetheless, it is still challenging for existing CNN-based approaches [10, 73] to generate photo-realistic textures due to the absence of generative prior. In this study, we set our sights on exploiting robust and extensive generative prior encapsulated in a pretrained image diffusion model, i.e., Stable Diffusion (SD) $\times 4$ upscaler [2]. By integrating this strong diffusion prior, our approach circumvents the need for exhaustive training from scratch and exhibits improved performance in producing detailed textures. + +Diffusion Models for Video Tasks. Following the success of text-to-image diffusion models [4, 12, 16, 20, 22, 44, 85], recent studies have ventured into the applications of video diffusion models [15, 25-27, 40-42]. Instead of training from scratch, some methods [14, 19, 48, 71, 78] focus on video generation using off-the-shelf image diffusion models [50, 85] in a zero-shot manner. To keep temporal consistency, cross attention [60] between neighboring frames and optical flow [58] warping are popular solutions adopted in these methods. While being efficient, these methods generally suffer from limited generalizability and the selection of hyperparameters can be tricky. Most recently, Blattmann et al. [6] propose to extend pretrained image diffusion models to the video domain by introducing an additional temporal dimension and fine-tuning the temporal layers only to speed up the training process. Subsequent works [21, 74] following such paradigm are capable of generating impressive video sequences. Inspired by these works, we employ a pretrained image $\times 4$ upscaling model [2] as generative prior and propose a novel local-global temporal strategy, resulting in temporally coherent outputs with faithful details. + +Diffusion Models for Restoration Tasks. With notable advancements in diffusion models [24, 46, 55, 77], numerous diffusion-based works have been proposed for image restoration. A straightforward way is to train a diffusion model conditioned on LR images from scratch [50, 51, 53, 72, 83], which, however, demands significant computational resources for training. To avoid such heavy training costs, another prevalent approach [12, 13, 32, 56, 67, 69, 76] is to incorporate constraints into the reverse diffusion process of a pretrained diffusion model. Though efficient, designing these constraints relies on pre-defined image degra + +dation processes or pretrained SR models as a priori, leading to limited generalizability and inferior performance. Recent works [62, 79] go further by fine-tuning directly on a frozen pretrained diffusion model with several additional trainable layers, demonstrating impressive performance. Inspired by these recent advances, we focus on exploiting the effective diffusion prior for real-world VSR, which remains under-explored and challenging due to the temporal discontinuities caused by inherent randomness during diffusion sampling. With the proposed local-global temporal strategy, our approach is capable of generating temporally coherent results for real-world VSR. + +# 3. Methodology + +Our objective is to develop a text-guided diffusion framework tailored for real-world VSR. The diffusion denoising process, characterized by its inherent stochastic nature, poses significant challenges when applied to video tasks. These challenges include temporal instability and the emergence of flickering artifacts, which are particularly prominent in VSR tasks involving lengthy video sequences. The complexity of these tasks lies not just in achieving temporal consistency within localized segments, but also in preserving coherence throughout the entire video. + +As illustrated in Fig. 2, our framework incorporates both local and global modules in the latent diffusion model (LDM) to preserve temporal consistency both within and across video segments. Within each diffusion time step $(t = 1,2\dots T)$ , the video is split into segments and processed with a U-Net that includes temporal layers to ensure consistency within each segment. If the current time step falls within the user-specified global refinement steps $(T^{*})$ , a recurrent latent propagation module is employed to improve consistency across segments during inference. Lastly, a finetuned VAE-Decoder is used to reduce remaining flickering artifacts. Thanks to the diffusion paradigm, our model exhibits remarkable versatility: 1) Input text prompts can enhance video quality further with improved realism and details. 2) User-specified noise levels provide a control over the trade-off between quality and fidelity. + +# 3.1. Preliminary: Diffusion Models + +Pretrained Stable Diffusion Image $\times 4$ Upscaler. Our Upscale-A-Video is built upon the pretrained text-guided SD $\times 4$ Upscaler [2]. It employs the LDM framework [50], which uses an autoencoder to transform an image into a latent with an encoder $\mathcal{E}$ ( $\times 4$ downsampling), and reconstructs it with a decoder $\mathcal{D}$ . Conditioned on the low-resolution images $\mathbf{x}$ , it learns to generate the high-quality counterparts via iterative denoising in the latent space. To achieve this, given latent samples $\mathbf{z} \sim p_{data}$ , diffused latents $\mathbf{z}_t = \alpha_t\mathbf{z} + \sigma_t\boldsymbol{\epsilon}$ is generated by introducing Gaussian noise to the latents $\mathbf{z}$ at each diffusion step $t$ ; where + +![](images/4555426d0f86e56663c3e09a24bb254703a67af42bd4235e40439c4965f8afb8.jpg) +Figure 2. An overview of Upscale-A-Video. Upscale-A-Video processes long videos using both local and global strategies to maintain temporal coherence. It divides the video into segments and processes them using a U-Net with temporal layers for intra-segment consistency. During user-specified diffusion steps for global refinement, a recurrent latent propagation module is used to enhance inter-segment consistency. Finally, a finetuned VAE-Decoder reduces remaining flickering artifacts for low-level consistency. Our model also allows users to guide texture creation with text prompts and adjust noise levels to balance the effect of restoration and generation. + +$\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ $\alpha_{t}$ and $\sigma_t$ define a noise schedule. To enhance the ability of our model to generate fine details, it also applies random noise to input images, i.e., the diffused images $\mathrm{x}_{\tau} = \alpha_{\tau}\mathrm{x} + \sigma_{\tau}\epsilon$ where $\tau$ is the noise level that corresponds to the early steps in noise schedule. + +Adopting v-prediction parameterization [54], the U-Net denoiser $f_{\theta}$ is trained to make predictions of $\mathrm{v}_t \equiv \alpha_t \epsilon - \sigma_t \mathrm{x}$ . The optimization objective for LDM is as follows: + +$$ +\mathbb {E} _ {\mathrm {z}, \mathrm {x}, \mathrm {c}, t, \epsilon} \left[ \| \mathrm {v} - f _ {\theta} \left(\mathrm {z} _ {t}, \mathrm {x} _ {\tau}; \mathrm {c}, t\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $c$ serves as an optional set of conditions, including text prompts and noise levels of the diffused image. During inference, the model has the flexibility to involve different text prompts and noise levels to diffusion sampling of $x_0$ , and finally decode it to produce the $\times 4$ upscaled images. + +Inflated 2D Convolution. When adapting pretrained 2D diffusion models to video tasks, it is common to inflate their 2D convolutions into the 3D convolutions [6, 17, 71, 78]. This modification allows the diffusion models to smoothly integrate new temporal layers, enabling them to capture and encode temporal information within the pretrained model. + +In this study, we explore a new avenue to create a text-guided video diffusion model for VSR tasks, starting with the pretrained $\mathrm{SD} \times 4$ Upscaler [2]. To process video data, we first modify its network structure by inflating the 2D convolutions into 3D convolutions, and then initialize our network with this upscaler to inherit its enhancement capabilities. Our goal is to transfer the knowledge learned from image upscaling to video enhancement, enabling more efficient training. In the following, we will describe our + +local-global strategy within the LDM framework to achieve temporal coherence while harnessing the capability of diffusion for VSR tasks. + +# 3.2. Local Consistency within Video Segments + +To apply the pretrained text-to-image SD model to video-related tasks, existing video diffusion models typically employ techniques such as 3D convolutions [6, 66], temporal attention [6, 17, 66, 68]sx, and cross-frame attention [71] to ensure temporal consistency. + +Finetuning Temporal U-Net. Following the existing methods [6, 66], we introduce additional temporal layers into the pretrained image model and learn the local consistency constraint within video segments. As illustrated in Fig. 2, in the modified temporal U-Net, we opt for temporal attention and 3D residual blocks based on 3D convolutions to serve as our temporal layers, and insert them within the pretrained spatial layers. The temporal attention layer performs self-attention along the temporal dimension and focuses on all local frames. Additionally, we add Rotary Position Embedding (RoPE) [59] into the temporal layers to provide the model with positional information for time. + +Our temporal layers are trained using the same noise schedule as the pretrained image model. Importantly, we keep the pretrained spatial layers fixed during training and only optimize the inserted temporal layers using Eq. (1). An essential benefit of this strategy is that it allows us to leverage the pretrained spatial layers that were learned from a huge, high-quality image dataset. This enables us to concentrate our training efforts on refining the temporal layers. + +Finetuning Temporal VAE-Decoder. Even after finetuning the U-Net on video data, the VAE-Decoder within the LDM framework, which is trained on images only, still tends to produce flickering artifacts when decoding a latent sequence. To mitigate this issue, we introduce additional temporal 3D residual blocks into the VAE-Decoder to enhance low-level consistency. Furthermore, the diffusion denoising process in the U-Net often introduces color shifts, a problem also encountered by other diffusion-based restoration networks [62, 83]. To tackle this, we condition the input videos with a Spatial Feature Transform (SFT) layer [63], which employs the inputs to transform the features of the first layer of the VAE-Decoder. This allows the input videos to provide low-frequency information, such as color, to strengthen the color fidelity of the output results. + +Similar to the training of the Temporal U-Net, we keep the pretrained spatial layers unchanged and only train the newly added temporal layers. These temporal layers are trained on video data using a hybrid loss consisting of L1 loss, LPIPS perceptual loss [86], and an adversarial loss employing a temporal PatchGAN [89] discriminator. As demonstrated by the ablation study in Fig. 6 and Table 2, this step is crucial for achieving favorable results. + +# 3.3. Global Consistency cross Video Segments + +The trained temporal layers within the LDM are limited to processing local sequences (e.g., eight frames in our U-Net setting), making it impossible to enforce global consistency constraints across video segments. Prior studies [9, 10, 88, 89] have already showcased the benefits of flow-guided long-term propagation in enhancing temporal consistency for video restoration tasks. However, their performance gains are often observed when dealing with long video sequences that provide long-term information. Unfortunately, these methods are not well-suited for diffusion models due to memory constraints, which typically restrict them to processing short video clips. + +Training-Free Recurrent Latent Propagation. We introduce a training-free flow-guided recurrent propagation module within the latent space. This module ensures global temporal coherence for long input videos, involving bidirectional propagation in the forward and backward directions. Here, we elaborate on the forward propagation, and backward propagation follows the same process. + +Given an input low-resolution video, we first adopt RAFT [58] to estimate optical flow, with its resolution exactly matching the latent resolution, thus no need for resizing. We then check the validity of the estimated flow by evaluating forward-backward consistency error [43]: + +$$ +E _ {i - 1 \rightarrow i} (p) = \left\| f _ {i - 1 \rightarrow i} (p) + f _ {i \rightarrow i - 1} (p + f _ {i - 1 \rightarrow i} (p)) \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $p$ denotes the position of the last frame latent, $f_{i-1\rightarrow i}$ and $f_{i\rightarrow i-1}$ are forward and backward flow respectively. As + +![](images/e554ca14d0af0e594e5c0fffae3a4a8f48768cbeb111915ccd507d03dc193b4d.jpg) +Figure 3. An illustration of flow-guided recurrent latent propagation. Without requiring any learning, this module can achieve coherence across video segments via long-term latent propagation and aggregation. It relies on optical flow validity determined by forward-backward consistency error [43]. Only latent positions with low consistency errors will be propagated, while those with high errors, marked with a red dot, are not. + +shown in Fig. 3, only latents with small consistency error will be propagated, which can be thought of as an occlusion mask, $M: E_{i-1 \to i}(p) < \delta$ , where $\delta$ is a threshold. Let $z_i$ be the latent feature for the $i$ -th frame at diffusion step $t$ , we update the predicted $\hat{z}_0$ as $\widetilde{\hat{z}_0}$ . The process of recurrent latent propagation and aggregation is expressed as: + +$$ +\begin{array}{l} \widetilde {\hat {\mathbf {z}}} _ {0} ^ {i} = \left[ \mathcal {W} (\widetilde {\hat {\mathbf {z}}} _ {0} ^ {i - 1}, f _ {i \rightarrow i - 1}) * \beta + \hat {\mathbf {z}} _ {0} ^ {i} * (1 - \beta) \right] * M \\ + \hat {\mathbf {z}} _ {0} ^ {i} * (1 - M), \tag {3} \\ \end{array} +$$ + +where $\mathcal{W}(\cdot)$ denotes warping operation (with nearest mode), and $\beta \in [0,1]$ serves as a fusion weight for aggregating the warped latent from the previous $(i - 1)^{th}$ frame to the current $i^{th}$ frame, where smaller values tend to preserve current information, while larger values favor the propagated information. We set $\beta$ to 0.5 by default. + +It is not necessary to apply this module at each diffusion step during the inference process. Instead, we can choose $T^{*}$ steps for latent propagation and aggregation. When dealing with minor video jitter, one can opt to integrate this module early in the diffusion denoising process, while for severe video jitter, such as AIGC videos, it is preferable to execute this module later in the denoising process. + +# 3.4. Inference with Additional Conditions + +We can further adjust additional conditions of text prompts and noise levels in Upscale-A-Video to influence the diffusion denoising process. Text prompts can guide the generation of texture details, such as animal fur, as shown in Fig. 7. Besides, adjusting the noise level allows us to balance the model's restoration and generation abilities, with smaller values favoring restoration and larger values promoting the generation of more details (see comparisons in Fig. 7). We also adopt Classifier-Free Guidance (CFG) [23] during inference, which can significantly enhance the impact of both text prompts and noise levels, helping produce high-quality videos with finer details. + +Table 1. Quantitative comparisons on different VSR benchmarks from diverse sources, i.e., synthetic (SPMCS, UDM10, REDS30, YouHQ40), real (VideoLQ), and AIGC (AIGC30) data. The best and second performances are marked in red and orange, respectively. $E_{\text{warp}}^*$ denotes $E_{\text{warp}} (\times 10^{-3})$ . + +
DatasetsMetricsReal-ESRGAN [65]SD × 4 Upscaler [2]ResShift [83]StableSR [62]RealVSR [80]DBVSR [47]RealBasicVSR [10]Ours
SPMCSPSNR ↑22.8923.1923.2722.7123.8824.2824.5125.32
SSIM ↑0.6690.6310.6670.6570.6810.7260.7170.741
LPIPS ↓0.2380.3040.2570.2310.4370.3020.1980.222
E*arp ↓1.3645.0084.9424.8150.2941.3600.5590.367
UDM10PSNR ↑27.1328.0727.6226.4527.3829.6029.1130.79
SSIM ↑0.8430.8110.8270.8250.8250.8800.8760.878
LPIPS ↓0.1900.1860.2220.1810.2780.1550.1720.133
E*arp ↓1.4621.7102.1962.7970.5311.9430.6020.446
REDS30PSNR ↑22.4022.9823.0023.7223.0524.3723.9124.41
SSIM ↑0.5910.5720.5800.6350.6030.6330.6360.631
LPIPS ↓0.3030.3990.3690.3520.6580.5880.2490.335
E*arp ↓3.6583.7534.1311.6450.3789.6591.5571.278
YouHQ40PSNR ↑24.3719.7123.7724.5324.1925.3724.0925.83
SSIM ↑0.7100.5790.6540.7110.6950.7190.6890.733
LPIPS ↓0.2720.4420.3760.2710.4840.4300.3060.268
E*arp ↓1.8563.3994.4261.5290.4851.1491.0520.737
VideoLQCLIP-IQA ↑0.3600.1580.4300.3440.2110.2740.3870.530
MUSIQ ↑49.4826.2140.9544.2324.5229.1555.3357.99
DOVER ↑7.1612.8844.6796.7832.5313.6287.5627.811
AIGC30CLIP-IQA ↑0.4300.3290.5690.4670.2760.2900.5650.674
MUSIQ ↑47.0935.3043.3244.9324.3927.2258.8757.66
DOVER ↑9.7105.6467.0429.6683.2853.52310.6811.67
+ +# 4. Experiments + +# 4.1. Datasets and Implementation + +Training Datasets. We train our Upscale-A-Video using the following datasets: 1) The subset of WebVid10M [5] contains around $335\mathrm{K}$ video-text pairs with each resolution around $336\times 596$ , which is commonly used in training video diffusion models [6, 11, 17, 84]; 2) YouHQ dataset. Due to the lack of high-quality video data for training, we additionally collect a large-scale high-definition $(1080\times 1920)$ dataset from YouTube, containing around $37\mathrm{K}$ video clips with diverse scenarios, i.e., street view, landscape, animal, human face, static object, nighttime scene, and among others. Training on such high-quality data further enhances the generation ability of our model for real-world VSR. Following the degradation pipeline of RealBasicVSR [10], we generate the LQ-HQ video pairs for training. + +Testing Datasets. We construct four synthetic testing datasets (i.e., SPMCS [81], UDM10 [57], REDS30 [45], and YouHQ40), which follow the same degradation pipeline in training to generate LQ videos. We split YouHQ40 test set from the proposed YouHQ dataset, containing 40 videos. Additionally, we evaluate the models on a real-world dataset (i.e., VideoLQ [10]) and an AIGC dataset (i.e., AIGC30) that collects 30 AI-generated videos by popular text-to-video generation models [1, 3, 6, 17, 25, 41, 68, 84]. + +Training Details. Our Upscale-A-Video is trained on 32 NVIDIA A100-80G GPUs with a batch size of 384. The training data is cropped to $80 \times 80$ with a length of 8. The learning rate is set to $1 \times 10^{-4}$ using the Adam [34] + +optimizer. We first train the U-Net model on both WebVid10M [5] and YouHQ for 70K iterations. Then we train another 10K iterations on YouHQ only. Since there are no text prompts for YouHQ, we use the null prompts during training. In this way, our model can handle VSR with either LQ inputs and prompts or LQ inputs only, leading to a more flexible use in practice. As for VAE-Decoder finetuning, we follow StableSR [62] to first generate 100K synthetic LQ-HQ video pairs on WebVid10M [5] and YouHQ, and the finetuned U-Net model is adopted to generate the corresponding latent codes for the LQ videos. + +Evaluation Metrics. We adopt different metrics to evaluate both the frame quality as well as temporal coherency of the generated results. For synthetic datasets with LQ-HQ pairs, we employ PSNR, SSIM, LPIPS [86], and flow warping error [35] $E_{\text{warp}}^*$ for evaluation. For real-world and AIGC test data, since no ground-truth videos are available, we conduct our evaluation on commonly used non-reference metrics, i.e., CLIP-IQA [61], MUSIQ [33], and DOVER [70]. + +# 4.2. Comparisons + +We compare Upscale-A-Video with several state-of-the-art methods, including Real-ESRGAN [65], SD $\times 4$ Upscaler [2], ResShift [83], StableSR [62], RealVSR [80], DBVSR [47], and RealBasicVSR [10]. + +Quantitative Evaluation. As shown in Table 1, our approach achieves the highest PSNR across all four synthetic datasets, suggesting its outstanding reconstruction ability. Moreover, our approach achieves the lowest LPIPS scores on both UDM10 and YouHQ40, indicating the high percep + +![](images/714a240ff7c09fc906f532ce5525969dc40879d00ad16590e2b3f7c7c1afaa89.jpg) +"Tourists and brick walls" + +![](images/eddef5a15146430ba9ddd29f83a0a221ba5a69bde2a4e821d590b8a4a3647c4c.jpg) +Bicubic + +![](images/212c789eb3a9b7dec78947c794bde3553b778d20c35ca73e0c14abd6d83b6ca3.jpg) +StableSR + +![](images/05d86810a377a8eaf51e1feebbf934124069df3d4d51dc315fab1b220b81a4b4.jpg) +SD $\times 4$ Upscaler + +![](images/d9b35c85201bf7d73e5f1f20d4ca57f08ed6c1627afe6afdcf12d655ee62a0ce.jpg) +RealBasicVSR + +![](images/5dce61d072a26de0e18f29feca9094c4fabce8b95206a513cb9298de879c4d96.jpg) +Ours + +![](images/9c2379282043e0eb315e1c4f5b42b51479528eb64d1bf478763ff4f65d35ab85.jpg) +GT + +![](images/20dbb699894d42e0912690a79591f0b79737db4609f3e044c01a92d0154ae452.jpg) +"A koala on the tree" + +![](images/827e11544d132b366dcaf230ae35c56713b87de86b5c4743687e552f36bde470.jpg) +Bicubic + +![](images/8d122ffb705c49dde6220795713418555d439bd9fefda338e72e89d117925940.jpg) +StableSR + +![](images/9bb9e1f04ef26e24850235a82b8dedc942adb69b967ae3438e696fe17cc64d7f.jpg) +SD $\times 4$ Upscaler + +![](images/c93d176b8adb15ab50df28fd45c24462ac6ece1567648e60ef7a05656ad08ae8.jpg) +RealBasicVSR + +![](images/4ae99125339eac3e6215589bab518d5dd42e0f0e7b27d1fd3e64ca01a46089dc.jpg) +Ours + +![](images/3104465d82a45808957055283046a76c8491d6068a3574c288a90b1762fcd827.jpg) +GT + +![](images/081f2c229fa7023b838f5c7ca09f6efb280225a4d1ac68be0558012272cbc736.jpg) +Figure 4. Qualitative comparisons on synthetic low-quality videos from REDS30 [45] and YouHQ40 datasets. Among the tested methods, only our Upscale-A-Video can recover the accurate wall structure and produce detailed koala fur. (Zoom-in for best view) +"A bus and car on the street" + +![](images/d6a32ced86f08cd95bf1f6776bc62508001e9d2ed4f77fd06019d238ff463c5d.jpg) +Bicubic + +![](images/c5eb6f7f68047ebd623560ef9a8548426f3f1a4b1e24a238acf0db166011005b.jpg) +Real-ESRGAN + +![](images/db4b517bc3358e2093f5e0a8c226e990c8cc14c51285cc0f8c76e72a3ee70dea.jpg) +ResShift + +![](images/8736c44e3c24a2f064c1700ea4729efc5b5fd6b7447b2840677b1751e2469472.jpg) +StableSR + +![](images/d317f876620dca3768ba11ffa52042eed27fb858227c52609ea3b3515adac9ae.jpg) +"A squirrel on a tree" +Figure 5. Qualitative comparisons on real-world test videos in VideoLQ [10] dataset. Our Upscale-A-Video effectively leverages the advantages of the diffusion paradigm in generating high-quality results. When compared to existing methods, it notably excels in its restoration capabilities, successfully recovering the billboard word "EAT IN or TAKEAWAY". In particular, when guided by text prompts, Upscale-A-Video showcases promising enhanced results with more details and heightened realism. (Zoom-in for best view) + +![](images/95e888a37c95addd25463f052af486590917aafba8aa8e7840a434448f239961.jpg) +SD $\times 4$ Upscaler + +![](images/7b453a197467f9a0f6b4af182cb232bf1d6565eebf4ba2625c7814642852ce71.jpg) +DBVSR + +![](images/fdff56df2f6214a0267e21ab77fb0023e87e69b8a95d1d8114100dc0132ae7cb.jpg) +RealBasicVSR + +![](images/dcb755e3c9d39433012712208e02efc973e3f9805242ded33ffcd7cb3eab5ae7.jpg) +Upscale-A-Video (Ours) + +![](images/98b891f86cabf73b2d8563139cb4688db32e2d3468bf5180b53ecdb282707d1c.jpg) +Bicubic + +![](images/dd1dd2900b107266a039550016e66dbf2e9c3e22c5c2687bec815d1c4d9c52d3.jpg) + +![](images/29c8ee582084cd2869576e2ac439b0be44e12bf2b78700b1f394be407e40581a.jpg) + +![](images/38e561fe0d112776873eb06d66fc19f3679280cdac7c8958ac928c723895a433.jpg) + +![](images/29fdf8793b6c9243869fda1625ecdf60d9a59336994eaf2ec4d51d32476a6ad7.jpg) + +![](images/8ef918afa2face933a1ee01da439ebd13da0c0fec52b43a1390391a3f6c42657.jpg) + +![](images/24762789705e87124e885e779316755afb6a70c93df5bf8bbd2a3312cfbe668e.jpg) + +![](images/2c0e8022ce1fcce978dd0386363f27a19467556b836b1c14b4a8fdcc117a7cb0.jpg) + +![](images/616e1053c2702cfb8da6153c782d1976d082bfd7f5e2e06fde2fa6154fd6daf3.jpg) +SD $\times 4$ Upscaler + +![](images/3ba2e51027b505b2bbdfe8111cc07e56dffd5f5f6c9bcb7f1c500090bfde9fcd.jpg) + +![](images/ddb1dfd847b81eef6e9e3cf4a67735a9ec4ef16bc8303d3385a6d2890c459020.jpg) +DBVSR + +![](images/bdd2c87d4c2c464d3012f95e72c15a1c6c8ae3716d35b9f7a93a7607c56d0041.jpg) + +![](images/966a5df8a661d33bd8498fbe3cdde694b1fd0c37294b2fb2746c971a64072ed2.jpg) +ResShift +RealBasicVSR + +![](images/e1b21395b8afc2688462996ec5484f6e49c66bb3662758fa92088730dbba463a.jpg) + +![](images/b57edc61f47e511db960555e6de5a08803c415503a02bb642715fd4bdd98c56b.jpg) +Upscale-A-Video (Ours) + +tual quality of our generated results. In addition, our approach further obtains the highest CLIP-IQA and DOVER scores on both the real dataset [10] and AIGC videos. The superiority across datasets from various sources demonstrates the effectiveness of our approach. + +Qualitative Evaluation. We present visual results on both synthetic [45, 57, 81] and real-world [10] videos in Fig. 4 and Fig. 5, respectively. It is observed that our Upscale-A-Video significantly outperforms existing CNN-and diffusion-based approaches in both artifact removal and details generation. Specifically, Upscale-A-Video is capable of generating more natural details in Fig. 4, and success + +fully recovers the words "EAT IN or TAKEAWAY" on the billboard in the first example of Fig. 5, while other methods generate blurry or distorted results. Note that SD $\times 4$ Upscaler shows less effectiveness for the real-world VSR since it does not consider mixed degradation in its training data. + +Temporal Consistency. Benefiting from our local-global temporal strategy, Upscale-A-Video achieves the best optical flow error score on UDM10 and the second-best scores on REDS30, SPMCS and YouHQ40, remarkably outperforming other diffusion-based approaches, and even beats the strong CNN-based VSR methods, i.e., RealBasicVSR and DBVSR. We also visualize the temporal profile in + +![](images/9fffd9bb765ccfbb25cae668e028067edea691d067150bedc69d1f045d83d2e8.jpg) +Figure 6. Comparison of temporal profile. We examine a row and track changes over time. The profile from existing methods and Exp. (a) and (c) (w/o latent propagation module) exhibit noise, suggesting the presence of flickering artifacts. Even with the finetuned decoder, the profile of Exp. (c) still displays some discontinuities. Thanks to long-term latent propagation and aggregation, the profile of our Upscale-A-Video exhibits a more seamless and smoother transition. + +![](images/ad165445da421d911bbb003fe30da0fa97ca0db6ca21c1a6395e881ac543629e.jpg) + +Table 2. Ablation study of finetuned VAE-Decoder and propagation module on YouHQ40. + +
Exp.ft-VAE-Dec.Latent Prop.PSNR↑SSIM↑E*warp↓
(a)23.820.63852.398
(b)✓25.470.72151.815
(c)✓25.750.73280.842
(d)✓✓25.830.73260.737
+ +# 4.3. Ablation Study + +Effectiveness of Finetuned VAE-Decoder. We first investigate the significance of the fine-tuned VAE-Decoder. As shown in Table 2, replacing our finetuned VAE-Decoder with the original decoder leads to worse PSNR, SSIM, and $E_{\text{warp}}^*$ . Particularly, the increase of $E_{\text{warp}}^*$ from 0.737 to 1.815 indicates a significant deterioration of temporal coherency. The comparisons in Fig. 6 also suggest inferior temporal consistency without the finetuned VAE-Decoder. + +Effectiveness of Propagation Module. In addition to fine-tuning VAE-Decoder, our proposed flow-guided recurrent latent propagation module further enhances the stability of long videos. As shown in Table 2, adopting the propagation module can further reduce $E_{\text{warp}}^*$ error, effectively improving temporal consistency while maintaining high PSNR. Similar phenomena can also be observed in the temporal profile in Fig. 6, showing a more seamless transition. + +Text Prompt. Upscale-A-Video is trained on video data with either labeled prompts or null prompts and thus can handle both scenarios. We examine the use of classifier-free guidance [23] to improve the visual quality during sampling. As shown in Figure 7, compared with the null prompt, adopting proper text prompts can significantly enhance the perceptual quality with more faithful details. + +Noise Level. It is observed that the level of noise added to the input can affect the performance of our approach. As shown in Fig. 7, with low noise level, the generated results tend to be suboptimal with blurry details. However, a noise + +![](images/048643498f9fe39cf18c8f196b96380e5496c17029ea42efea84550cdf51c160.jpg) +Input + +![](images/cab392bf147f095abf9a3f2520a0c016aeeb9e8218cc513ba0da073f9a4fb69e.jpg) +w/o Text Prompt + +![](images/fb750d75bae376c0cbaeecb6b580833beda50d4d131a8fe15d1dc66cf693f4cb.jpg) + +![](images/dba03b83ee73b8209438e44e1a1a3bae2db21bc0eaa790fa509f68f2bfd39e24.jpg) +Fig. 6. It is observed that our approach achieves superior performance with a more seamless and smoother transition. +Noise Level 0 +Figure 7. Comparison of various conditions of text prompts and different levels of noise. (Zoom-in for best view) + +![](images/6ec99efab4d96c444171af13754a9a436fb06e11d3a45240827e8d9b47e157d3.jpg) +Noise Level 50 + +![](images/eff9969acf979371b7d3e1bdb538422ec584abb85f961a777c264779d1466a3a.jpg) +w/ Text Prompt +Noise Level 150 + +level that is too large may result in oversharpening. + +# 5. Conclusion + +While diffusion models have achieved impressive performance on a wide range of image tasks, their applications on video tasks, particularly real-world VSR remain challenging and understudied. In this paper, we present Upscale-A-Video, a novel approach to exploit image diffusion prior for real-world VSR while avoiding temporal discontinuities drawn from the inherent randomness during sampling process. Specifically, we enhance the temporal coherence by proposing a novel local-global temporal strategy within the latent diffusion framework. We additionally devote our efforts to achieving a trade-off between fidelity and quality by enabling texture creation with text prompts and noise level control, further facilitating the practical use in real-world scenarios. We believe that our exploration would lay a good foundation for future works. + +Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). The project is also supported by Shanghai AI Laboratory. + +# References + +[1] Pika Labs. https://www.pika.art/, 2023.6 +[2] Stable Diffusion x4 Upscaler. https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler,2023.2,3,4,6 +[3] Zeroscope V2 XL. https://replicate.com/anotherjesse/zeroscope-v2-x1, 2023.6 +[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6 +[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 4, 6 +[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2 +[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2 +[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5 +[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, 2022. 2, 3, 5, 6, 7 +[11] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 6 +[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 3 +[13] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In NeurIPS, 2022. 3 +[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical flow-guided attention for consistent text-to-video editing. arXiv preprint arXiv:2310.05922, 2023. 2, 3 +[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3 +[16] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 3 +[17] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation: + +A noise prior for video diffusion models. In ICCV, 2023. 2, 4, 6 +[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2 +[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 3 +[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3 +[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPSW, 2022. 5, 8 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 +[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen Video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 6 +[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022. +[27] Yaosi Hu, Zhenzhong Chen, and Chong Luo. LaMD: Latent motion diffusion for video generation. arXiv preprint arXiv:2304.11603, 2023. 3 +[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2 +[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020. +[30] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020. +[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 2 +[32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 3 +[33] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In ICCV, 2021. 6 +[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 + +[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6 +[36] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2 +[37] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. In NeurIPS, 2022. 2 +[38] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2 +[39] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. In IEEE TPAMI, 2013. 2 +[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. VDT: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3 +[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 6 +[42] Kangfu Mei and Vishal Patel. VIDM: Video implicit diffusion models. In AAAI, 2023. 3 +[43] Simon Meister, Junhwa Hur, and Stefan Roth. UnFlow: Unsupervised learning of optical flow with a bidirectional census loss. In AAAI, 2018. 5 +[44] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3 +[45] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 2, 6, 7 +[46] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3 +[47] Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. In ICCV, 2021. 6 +[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. FateZero: Fusing attentions for zero-shot text-based video editing. ICCV, 2023. 3 +[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 + +[51] Hshmat Sahak, Daniel Watson, Chitwan Sahara, and David Fleet. Denoising diffusion probabilistic models for robust image super-resolution in the wild. arXiv preprint arXiv:2302.07864, 2023. 2, 3 +[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 3 +[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 4 +[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3 +[56] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In ICLR, 2023. 3 +[57] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 6, 7 +[58] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5 +[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3 +[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6 +[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 5, 6 +[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5 +[64] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2 +[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In ICCVW, 2021. 6 +[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 2, 4 +[67] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2022. 3 + +[68] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 4, 6 +[69] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2023. 3 +[70] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023. 6 +[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 4 +[72] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. DiffIR: Efficient diffusion model for image restoration. ICCV, 2023. 3 +[73] Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao Dong, and Ying Shan. Mitigating artifacts in real-world video super-resolution models. In AAAI, 2023. 2, 3 +[74] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple diffusion adapter for efficient video generation. arXiv preprint arXiv:2308.09710, 2023. 3 +[75] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. In ICCV, 2019. 2 +[76] Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In NeurIPS, 2023. 3 +[77] S. Yang, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3 +[78] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender A Video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 2, 3, 4 +[79] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 3 +[80] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In ICCV, 2021. 2, 3, 6 +[81] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 6, 7 +[82] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2 + +[83] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. ResShift: Efficient diffusion model for image superresolution by residual shifting. In NeurIPS, 2023. 2, 3, 5, 6 +[84] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023.6 +[85] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3 +[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6 +[87] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 +[88] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In ICCV, 2019. 5 +[89] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. ProPainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 5 \ No newline at end of file diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/images.zip b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..730b89996471702e591ead4b2ac059a5cd9e09d7 --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e77340180c5eb6f32706a83594f3159ad9ecadc594a02c1ee3606e3ca52b594 +size 910578 diff --git a/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/layout.json b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de49745eac2e57b3897b4f11c13f5e3261f17c87 --- /dev/null +++ b/2024/Upscale-A-Video_ Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution/layout.json @@ -0,0 +1,11694 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 102, + 473, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 102, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 102, + 473, + 140 + ], + "type": "text", + "content": "Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 160, + 508, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 160, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 86, + 160, + 508, + 190 + ], + "type": "text", + "content": "Shangchen Zhou* Peiqing Yang* Jianyi Wang Yihang Luo Chen Change Loy S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 158, + 191, + 432, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 191, + 432, + 202 + ], + "spans": [ + { + "bbox": [ + 158, + 191, + 432, + 202 + ], + "type": "text", + "content": "https://shangchenzhou.com/projects/upscale-a-video" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 219, + 364, + 399 + ], + "blocks": [ + { + "bbox": [ + 50, + 219, + 364, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 219, + 364, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 219, + 364, + 399 + ], + "type": "image", + "image_path": "e3df4f0b8c0bbe58fd8efa2da63b6b1ada06c32e315115a9bc72d62369fd25ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 219, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 365, + 219, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 365, + 219, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 365, + 219, + 545, + 399 + ], + "type": "image", + "image_path": "4e954c433727a5bf1e749f520a83487b701de7d06eac5702f21aeda33c2930c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 400, + 133, + 513 + ], + "blocks": [ + { + "bbox": [ + 50, + 400, + 133, + 513 + ], + "lines": [ + { + "bbox": [ + 50, + 400, + 133, + 513 + ], + "spans": [ + { + "bbox": [ + 50, + 400, + 133, + 513 + ], + "type": "image", + "image_path": "d3cfd7d11cb82a7723e261bfdce2ce4ad291e4f2728b9ec5c30531609ddce7d1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 514, + 100, + 523 + ], + "lines": [ + { + "bbox": [ + 83, + 514, + 100, + 523 + ], + "spans": [ + { + "bbox": [ + 83, + 514, + 100, + 523 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 400, + 216, + 513 + ], + "blocks": [ + { + "bbox": [ + 133, + 400, + 216, + 513 + ], + "lines": [ + { + "bbox": [ + 133, + 400, + 216, + 513 + ], + "spans": [ + { + "bbox": [ + 133, + 400, + 216, + 513 + ], + "type": "image", + "image_path": "e3415af2147966bf0abd49be26aad9fc8ba93666e3cd457ea2505bde4a9c8da6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 514, + 187, + 522 + ], + "lines": [ + { + "bbox": [ + 160, + 514, + 187, + 522 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 187, + 522 + ], + "type": "text", + "content": "ResShift" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 215, + 400, + 298, + 513 + ], + "blocks": [ + { + "bbox": [ + 215, + 400, + 298, + 513 + ], + "lines": [ + { + "bbox": [ + 215, + 400, + 298, + 513 + ], + "spans": [ + { + "bbox": [ + 215, + 400, + 298, + 513 + ], + "type": "image", + "image_path": "27a3154fefffad72ac0dc582fed48b1712597ef56f9a69c0b3bae079b13cf69c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 514, + 270, + 522 + ], + "lines": [ + { + "bbox": [ + 241, + 514, + 270, + 522 + ], + "spans": [ + { + "bbox": [ + 241, + 514, + 270, + 522 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 400, + 378, + 513 + ], + "blocks": [ + { + "bbox": [ + 298, + 400, + 378, + 513 + ], + "lines": [ + { + "bbox": [ + 298, + 400, + 378, + 513 + ], + "spans": [ + { + "bbox": [ + 298, + 400, + 378, + 513 + ], + "type": "image", + "image_path": "b0786d235379a2426867ab856975e2b308c9349568aabfe900ed1ebe549c4401.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 514, + 362, + 523 + ], + "lines": [ + { + "bbox": [ + 313, + 514, + 362, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 362, + 523 + ], + "type": "text", + "content": "SD " + }, + { + "bbox": [ + 313, + 514, + 362, + 523 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 514, + 362, + 523 + ], + "type": "text", + "content": " Upscaler" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 379, + 400, + 461, + 513 + ], + "blocks": [ + { + "bbox": [ + 379, + 400, + 461, + 513 + ], + "lines": [ + { + "bbox": [ + 379, + 400, + 461, + 513 + ], + "spans": [ + { + "bbox": [ + 379, + 400, + 461, + 513 + ], + "type": "image", + "image_path": "a6be451dfed78bae9943bac056907db1afb342da10fd65488055acca444e155f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 514, + 443, + 522 + ], + "lines": [ + { + "bbox": [ + 397, + 514, + 443, + 522 + ], + "spans": [ + { + "bbox": [ + 397, + 514, + 443, + 522 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 531, + 547, + 564 + ], + "lines": [ + { + "bbox": [ + 46, + 531, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 547, + 564 + ], + "type": "text", + "content": "Figure 1. Video super-resolution comparisons on both real-world and AI-generated videos. Our proposed Upscale-A-Video showcases excellent upscaling capabilities. By using appropriate text prompts, it achieves impressive results characterized by finer details and heightened visual realism. (Zoom-in for best view)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 461, + 400, + 544, + 513 + ], + "blocks": [ + { + "bbox": [ + 461, + 400, + 544, + 513 + ], + "lines": [ + { + "bbox": [ + 461, + 400, + 544, + 513 + ], + "spans": [ + { + "bbox": [ + 461, + 400, + 544, + 513 + ], + "type": "image", + "image_path": "e8d44988b4926a90ed16197a68632974c3dbdce8b9b2218d608e58c5aa44f070.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 514, + 539, + 523 + ], + "lines": [ + { + "bbox": [ + 466, + 514, + 539, + 523 + ], + "spans": [ + { + "bbox": [ + 466, + 514, + 539, + 523 + ], + "type": "text", + "content": "Upscale-A-Video (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "spans": [ + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 601, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 696 + ], + "type": "text", + "content": "Text-based diffusion models have exhibited remarkable success in generation and editing, showing great promise for enhancing visual content with their generative prior. However, applying these models to video super-resolution remains challenging due to the high demands for output fidelity and temporal consistency, which is complicated by the inherent randomness in diffusion models. Our study introduces Upscale-A-Video, a text-guided latent diffusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "type": "text", + "content": "framework for video upscaling. This framework ensures temporal coherence through two key mechanisms: locally, it integrates temporal layers into U-Net and VAE-Decoder, maintaining consistency within short sequences; globally, without training, a flow-guided recurrent latent propagation module is introduced to enhance overall video stability by propagating and fusing latent across the entire sequences. Thanks to the diffusion paradigm, our model also offers greater flexibility by allowing text prompts to guide texture creation and adjustable noise levels to balance restoration and generation, enabling a trade-off between fidelity and" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2535" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": "quality. Extensive experiments show that Upscale-A-Video surpasses existing methods in both synthetic and real-world benchmarks, as well as in AI-generated videos, showcasing impressive visual realism and temporal consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 145, + 128, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 145, + 128, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 128, + 158 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 166, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 368 + ], + "type": "text", + "content": "Video super-resolution (VSR) in real-world scenarios is a challenging task that aims at enhancing the quality of low-quality videos to produce high-quality results. Unlike previous works that mainly focus on either synthetic degradations [8, 9, 64] or specific camera-related degradations [80], this task is more demanding due to the need for addressing complex and unknown degradations commonly found in low-quality videos, such as downsampling, noise, blur, flickering, and video compression. In addition, maintaining the visual fidelity and ensuring temporal coherence are crucial for the perceptual quality of the output video. Although recent convolutional neural network (CNN)-based networks [10, 73] have shown success in mitigating many forms of degradation, they still fall short in producing realistic textures and details due to their limited generative capabilities, often resulting in over-smoothing, which can be observed in RealBasicVSR results shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 559 + ], + "type": "text", + "content": "Diffusion models [24] have exhibited impressive proficiency in generating high-quality images [49, 50, 52, 85] and videos [6, 11, 17, 25, 87]. Harnessing their generative potential, diffusion-based models have been proposed for image restoration, including both training from scratch [50, 51, 83] and fine-tuning from pretrained Stable Diffusion models [38, 62]. These methods have effectively mitigated the over-smoothing issue often observed in CNN-based models, yielding results with more realistic fine-grained details. However, adapting these diffusion priors to VSR remains a non-trivial challenge. This difficulty stems from the inherent randomness in diffusion sampling, which inevitably introduces unexpected temporal discontinuities in resulting videos. This issue is more pronounced in latent diffusion, where the VAE decoder further introduces flickering in low-level texture details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "type": "text", + "content": "Recent efforts have been made to adapt image diffusion models for video tasks by introducing strategies for temporal consistency, which include: 1) Fine-tuning video models with temporal layers such as 3D convolution [6, 66] and temporal attention [6, 11, 66, 87]; 2) Employing zero-shot mechanisms like cross-frame attention [71, 78] and flow-guided attention [14, 18] in the pretrained models. Although these solutions significantly improve video stability, there remain two primary issues: i) The current methods, operating in the U-Net feature or latent space, struggle to maintain low-level consistency, with issues like texture flickering still present. ii) The existing temporal layers and attention mechanisms can only impose constraints on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "short, local input sequences, limiting their capacity to ensure global temporal consistency in longer videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 97, + 546, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 546, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 546, + 310 + ], + "type": "text", + "content": "To tackle these issues, we adopt a local-global strategy for maintaining temporal consistency in video reconstruction, focusing on both fine-grained textures and overall consistency. On a local video clip, we explore finetuning a pretrained image " + }, + { + "bbox": [ + 304, + 97, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 97, + 546, + 310 + ], + "type": "text", + "content": " upscaling model [2] with additional temporal layers on the video data. Specifically, within a latent diffusion framework, we first finetune the U-Net with integrated 3D convolutions and temporal attention layers, and then tune the VAE-Decoder with video-conditioned inputs and 3D convolutions. The former significantly achieves structure stability in local sequences, while the latter further improves low-level consistency, reducing texture flickering. On a more global scale, we introduce a novel, training-free flow-guided recurrent latent propagation module. Spanning short video segments, it bidirectionally conducts frame-by-frame propagation and latent fusion during inference. Leveraging latent fusion in a long-term sequence, this module encourages overall stability for long videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 311, + 546, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 546, + 465 + ], + "type": "text", + "content": "We further investigate the generative potential of diffusion models in the VSR task. Following the text-guided fashion, our model can leverage text prompts as an optional condition to steer the model towards producing more realistic and high-quality details, as illustrated in Fig. 1. Furthermore, we enhance the robustness of our model against heavy or unseen degradation by injecting noise into input to dilute its degradation. By adjusting the level of added noise, we can modulate the balance between restoring and generating within the diffusion model. Lower noise levels prioritize the model's restoration capabilities, while higher levels encourage generation of more refined details, thus achieving a trade-off between fidelity and quality of the output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 466, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 620 + ], + "type": "text", + "content": "To summarize, the main contribution of our study is a practical and robust approach to real-world VSR. In particular, we conduct an exhaustive exploration of integrating a local-global temporal strategy within a latent diffusion framework, enhancing temporal coherence and generation quality. We further examine ways to improve versatility by allowing text prompts to guide texture creation, and offer control over noise levels to balance restoration and generation, thereby achieving a trade-off between fidelity and quality. Thanks to our design and pretrained generative prior, our model achieves state-of-the-art performance on existing benchmarks, showing remarkable visual realism and temporal consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 633, + 392, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 392, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 392, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Video Super-Resolution. VSR aims to restore a sequence of high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts. Most existing approaches [7-9, 28-31, 36, 37, 64, 75] assume a pre-defined degradation process [39, 45, 75, 82], and their performance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2536" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 332 + ], + "type": "text", + "content": "deteriorates significantly in real-world scenarios due to the limited generalizability. To tackle real-world VSR, recent works go beyond the traditional paradigms by assuming inputs with unknown degradations. Due to the lack of real-world paired data for training, Yang et al. [80] propose to collect HR-LR data pairs with iPhone cameras. While the VSR model trained on such data can be effective to videos captured by similar mobile cameras, its generalization capability to other devices is in doubt. Rather than relying on labor-intensive pair data collection, recent studies [10, 73] have shifted towards employing diverse degradations for data augmentation during training, demonstrating better performance in handling real-world cases. Nonetheless, it is still challenging for existing CNN-based approaches [10, 73] to generate photo-realistic textures due to the absence of generative prior. In this study, we set our sights on exploiting robust and extensive generative prior encapsulated in a pretrained image diffusion model, i.e., Stable Diffusion (SD) " + }, + { + "bbox": [ + 47, + 72, + 289, + 332 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 47, + 72, + 289, + 332 + ], + "type": "text", + "content": " upscaler [2]. By integrating this strong diffusion prior, our approach circumvents the need for exhaustive training from scratch and exhibits improved performance in producing detailed textures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 334, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 289, + 582 + ], + "type": "text", + "content": "Diffusion Models for Video Tasks. Following the success of text-to-image diffusion models [4, 12, 16, 20, 22, 44, 85], recent studies have ventured into the applications of video diffusion models [15, 25-27, 40-42]. Instead of training from scratch, some methods [14, 19, 48, 71, 78] focus on video generation using off-the-shelf image diffusion models [50, 85] in a zero-shot manner. To keep temporal consistency, cross attention [60] between neighboring frames and optical flow [58] warping are popular solutions adopted in these methods. While being efficient, these methods generally suffer from limited generalizability and the selection of hyperparameters can be tricky. Most recently, Blattmann et al. [6] propose to extend pretrained image diffusion models to the video domain by introducing an additional temporal dimension and fine-tuning the temporal layers only to speed up the training process. Subsequent works [21, 74] following such paradigm are capable of generating impressive video sequences. Inspired by these works, we employ a pretrained image " + }, + { + "bbox": [ + 47, + 334, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 47, + 334, + 289, + 582 + ], + "type": "text", + "content": " upscaling model [2] as generative prior and propose a novel local-global temporal strategy, resulting in temporally coherent outputs with faithful details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 583, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 289, + 715 + ], + "type": "text", + "content": "Diffusion Models for Restoration Tasks. With notable advancements in diffusion models [24, 46, 55, 77], numerous diffusion-based works have been proposed for image restoration. A straightforward way is to train a diffusion model conditioned on LR images from scratch [50, 51, 53, 72, 83], which, however, demands significant computational resources for training. To avoid such heavy training costs, another prevalent approach [12, 13, 32, 56, 67, 69, 76] is to incorporate constraints into the reverse diffusion process of a pretrained diffusion model. Though efficient, designing these constraints relies on pre-defined image degra" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 214 + ], + "type": "text", + "content": "dation processes or pretrained SR models as a priori, leading to limited generalizability and inferior performance. Recent works [62, 79] go further by fine-tuning directly on a frozen pretrained diffusion model with several additional trainable layers, demonstrating impressive performance. Inspired by these recent advances, we focus on exploiting the effective diffusion prior for real-world VSR, which remains under-explored and challenging due to the temporal discontinuities caused by inherent randomness during diffusion sampling. With the proposed local-global temporal strategy, our approach is capable of generating temporally coherent results for real-world VSR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 227, + 389, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 227, + 389, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 389, + 240 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 247, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 365 + ], + "type": "text", + "content": "Our objective is to develop a text-guided diffusion framework tailored for real-world VSR. The diffusion denoising process, characterized by its inherent stochastic nature, poses significant challenges when applied to video tasks. These challenges include temporal instability and the emergence of flickering artifacts, which are particularly prominent in VSR tasks involving lengthy video sequences. The complexity of these tasks lies not just in achieving temporal consistency within localized segments, but also in preserving coherence throughout the entire video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our framework incorporates both local and global modules in the latent diffusion model (LDM) to preserve temporal consistency both within and across video segments. Within each diffusion time step " + }, + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "inline_equation", + "content": "(t = 1,2\\dots T)" + }, + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "text", + "content": ", the video is split into segments and processed with a U-Net that includes temporal layers to ensure consistency within each segment. If the current time step falls within the user-specified global refinement steps " + }, + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "inline_equation", + "content": "(T^{*})" + }, + { + "bbox": [ + 304, + 366, + 548, + 556 + ], + "type": "text", + "content": ", a recurrent latent propagation module is employed to improve consistency across segments during inference. Lastly, a finetuned VAE-Decoder is used to reduce remaining flickering artifacts. Thanks to the diffusion paradigm, our model exhibits remarkable versatility: 1) Input text prompts can enhance video quality further with improved realism and details. 2) User-specified noise levels provide a control over the trade-off between quality and fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 564, + 474, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 564, + 474, + 577 + ], + "spans": [ + { + "bbox": [ + 305, + 564, + 474, + 577 + ], + "type": "text", + "content": "3.1. Preliminary: Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Pretrained Stable Diffusion Image " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " Upscaler. Our Upscale-A-Video is built upon the pretrained text-guided SD " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " Upscaler [2]. It employs the LDM framework [50], which uses an autoencoder to transform an image into a latent with an encoder " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " downsampling), and reconstructs it with a decoder " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". Conditioned on the low-resolution images " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", it learns to generate the high-quality counterparts via iterative denoising in the latent space. To achieve this, given latent samples " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim p_{data}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", diffused latents " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\alpha_t\\mathbf{z} + \\sigma_t\\boldsymbol{\\epsilon}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " is generated by introducing Gaussian noise to the latents " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " at each diffusion step " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "; where" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "2537" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 544, + 276 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 276 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 276 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 276 + ], + "type": "image", + "image_path": "4555426d0f86e56663c3e09a24bb254703a67af42bd4235e40439c4965f8afb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 548, + 338 + ], + "type": "text", + "content": "Figure 2. An overview of Upscale-A-Video. Upscale-A-Video processes long videos using both local and global strategies to maintain temporal coherence. It divides the video into segments and processes them using a U-Net with temporal layers for intra-segment consistency. During user-specified diffusion steps for global refinement, a recurrent latent propagation module is used to enhance inter-segment consistency. Finally, a finetuned VAE-Decoder reduces remaining flickering artifacts for low-level consistency. Our model also allows users to guide texture creation with text prompts and adjust noise levels to balance the effect of restoration and generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "text", + "content": " define a noise schedule. To enhance the ability of our model to generate fine details, it also applies random noise to input images, i.e., the diffused images " + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\mathrm{x}_{\\tau} = \\alpha_{\\tau}\\mathrm{x} + \\sigma_{\\tau}\\epsilon" + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 350, + 287, + 408 + ], + "type": "text", + "content": " is the noise level that corresponds to the early steps in noise schedule." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": "Adopting v-prediction parameterization [54], the U-Net denoiser " + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": " is trained to make predictions of " + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_t \\equiv \\alpha_t \\epsilon - \\sigma_t \\mathrm{x}" + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": ". The optimization objective for LDM is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 453, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 453, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 97, + 453, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathrm {z}, \\mathrm {x}, \\mathrm {c}, t, \\epsilon} \\left[ \\| \\mathrm {v} - f _ {\\theta} \\left(\\mathrm {z} _ {t}, \\mathrm {x} _ {\\tau}; \\mathrm {c}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "e1d9970dddc22b2fd0227b2c2bb04a69c4d64b8f79bc9104871e69bf7e77b8d4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "text", + "content": " serves as an optional set of conditions, including text prompts and noise levels of the diffused image. During inference, the model has the flexibility to involve different text prompts and noise levels to diffusion sampling of " + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "text", + "content": ", and finally decode it to produce the " + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 476, + 287, + 535 + ], + "type": "text", + "content": " upscaled images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 536, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 606 + ], + "type": "text", + "content": "Inflated 2D Convolution. When adapting pretrained 2D diffusion models to video tasks, it is common to inflate their 2D convolutions into the 3D convolutions [6, 17, 71, 78]. This modification allows the diffusion models to smoothly integrate new temporal layers, enabling them to capture and encode temporal information within the pretrained model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "content": "In this study, we explore a new avenue to create a text-guided video diffusion model for VSR tasks, starting with the pretrained " + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SD} \\times 4" + }, + { + "bbox": [ + 46, + 607, + 288, + 714 + ], + "type": "text", + "content": " Upscaler [2]. To process video data, we first modify its network structure by inflating the 2D convolutions into 3D convolutions, and then initialize our network with this upscaler to inherit its enhancement capabilities. Our goal is to transfer the knowledge learned from image upscaling to video enhancement, enabling more efficient training. In the following, we will describe our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 546, + 384 + ], + "type": "text", + "content": "local-global strategy within the LDM framework to achieve temporal coherence while harnessing the capability of diffusion for VSR tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 395, + 524, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 524, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 524, + 409 + ], + "type": "text", + "content": "3.2. Local Consistency within Video Segments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 546, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 474 + ], + "type": "text", + "content": "To apply the pretrained text-to-image SD model to video-related tasks, existing video diffusion models typically employ techniques such as 3D convolutions [6, 66], temporal attention [6, 17, 66, 68]sx, and cross-frame attention [71] to ensure temporal consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 475, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 546, + 617 + ], + "type": "text", + "content": "Finetuning Temporal U-Net. Following the existing methods [6, 66], we introduce additional temporal layers into the pretrained image model and learn the local consistency constraint within video segments. As illustrated in Fig. 2, in the modified temporal U-Net, we opt for temporal attention and 3D residual blocks based on 3D convolutions to serve as our temporal layers, and insert them within the pretrained spatial layers. The temporal attention layer performs self-attention along the temporal dimension and focuses on all local frames. Additionally, we add Rotary Position Embedding (RoPE) [59] into the temporal layers to provide the model with positional information for time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 619, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 547, + 715 + ], + "type": "text", + "content": "Our temporal layers are trained using the same noise schedule as the pretrained image model. Importantly, we keep the pretrained spatial layers fixed during training and only optimize the inserted temporal layers using Eq. (1). An essential benefit of this strategy is that it allows us to leverage the pretrained spatial layers that were learned from a huge, high-quality image dataset. This enables us to concentrate our training efforts on refining the temporal layers." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2538" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 249 + ], + "type": "text", + "content": "Finetuning Temporal VAE-Decoder. Even after finetuning the U-Net on video data, the VAE-Decoder within the LDM framework, which is trained on images only, still tends to produce flickering artifacts when decoding a latent sequence. To mitigate this issue, we introduce additional temporal 3D residual blocks into the VAE-Decoder to enhance low-level consistency. Furthermore, the diffusion denoising process in the U-Net often introduces color shifts, a problem also encountered by other diffusion-based restoration networks [62, 83]. To tackle this, we condition the input videos with a Spatial Feature Transform (SFT) layer [63], which employs the inputs to transform the features of the first layer of the VAE-Decoder. This allows the input videos to provide low-frequency information, such as color, to strengthen the color fidelity of the output results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": "Similar to the training of the Temporal U-Net, we keep the pretrained spatial layers unchanged and only train the newly added temporal layers. These temporal layers are trained on video data using a hybrid loss consisting of L1 loss, LPIPS perceptual loss [86], and an adversarial loss employing a temporal PatchGAN [89] discriminator. As demonstrated by the ablation study in Fig. 6 and Table 2, this step is crucial for achieving favorable results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 350, + 264, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 264, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 264, + 363 + ], + "type": "text", + "content": "3.3. Global Consistency cross Video Segments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 510 + ], + "type": "text", + "content": "The trained temporal layers within the LDM are limited to processing local sequences (e.g., eight frames in our U-Net setting), making it impossible to enforce global consistency constraints across video segments. Prior studies [9, 10, 88, 89] have already showcased the benefits of flow-guided long-term propagation in enhancing temporal consistency for video restoration tasks. However, their performance gains are often observed when dealing with long video sequences that provide long-term information. Unfortunately, these methods are not well-suited for diffusion models due to memory constraints, which typically restrict them to processing short video clips." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 511, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 289, + 593 + ], + "type": "text", + "content": "Training-Free Recurrent Latent Propagation. We introduce a training-free flow-guided recurrent propagation module within the latent space. This module ensures global temporal coherence for long input videos, involving bidirectional propagation in the forward and backward directions. Here, we elaborate on the forward propagation, and backward propagation follows the same process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 653 + ], + "type": "text", + "content": "Given an input low-resolution video, we first adopt RAFT [58] to estimate optical flow, with its resolution exactly matching the latent resolution, thus no need for resizing. We then check the validity of the estimated flow by evaluating forward-backward consistency error [43]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 656, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 656, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 50, + 656, + 287, + 689 + ], + "type": "interline_equation", + "content": "E _ {i - 1 \\rightarrow i} (p) = \\left\\| f _ {i - 1 \\rightarrow i} (p) + f _ {i \\rightarrow i - 1} (p + f _ {i - 1 \\rightarrow i} (p)) \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "5f33f7e5af1a2218c72665cc726b2556fe9ce20c1a63123c57c0497bf70adb7b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": " denotes the position of the last frame latent, " + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{i-1\\rightarrow i}" + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{i\\rightarrow i-1}" + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": " are forward and backward flow respectively. As" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 542, + 175 + ], + "type": "image", + "image_path": "e554ca14d0af0e594e5c0fffae3a4a8f48768cbeb111915ccd507d03dc193b4d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "type": "text", + "content": "Figure 3. An illustration of flow-guided recurrent latent propagation. Without requiring any learning, this module can achieve coherence across video segments via long-term latent propagation and aggregation. It relies on optical flow validity determined by forward-backward consistency error [43]. Only latent positions with low consistency errors will be propagated, while those with high errors, marked with a red dot, are not." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": "shown in Fig. 3, only latents with small consistency error will be propagated, which can be thought of as an occlusion mask, " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "M: E_{i-1 \\to i}(p) < \\delta" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": " is a threshold. Let " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": " be the latent feature for the " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": "-th frame at diffusion step " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": ", we update the predicted " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\widetilde{\\hat{z}_0}" + }, + { + "bbox": [ + 304, + 273, + 547, + 346 + ], + "type": "text", + "content": ". The process of recurrent latent propagation and aggregation is expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 351, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 351, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 324, + 351, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i} = \\left[ \\mathcal {W} (\\widetilde {\\hat {\\mathbf {z}}} _ {0} ^ {i - 1}, f _ {i \\rightarrow i - 1}) * \\beta + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - \\beta) \\right] * M \\\\ + \\hat {\\mathbf {z}} _ {0} ^ {i} * (1 - M), \\tag {3} \\\\ \\end{array}", + "image_path": "5be43efc82b38583a85bff50198e07bc4a94537a7cea640074051760acd831b9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot)" + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": " denotes warping operation (with nearest mode), and " + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": " serves as a fusion weight for aggregating the warped latent from the previous " + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "inline_equation", + "content": "(i - 1)^{th}" + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": " frame to the current " + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": " frame, where smaller values tend to preserve current information, while larger values favor the propagated information. We set " + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 392, + 545, + 464 + ], + "type": "text", + "content": " to 0.5 by default." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 464, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 547 + ], + "type": "text", + "content": "It is not necessary to apply this module at each diffusion step during the inference process. Instead, we can choose " + }, + { + "bbox": [ + 304, + 464, + 547, + 547 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 304, + 464, + 547, + 547 + ], + "type": "text", + "content": " steps for latent propagation and aggregation. When dealing with minor video jitter, one can opt to integrate this module early in the diffusion denoising process, while for severe video jitter, such as AIGC videos, it is preferable to execute this module later in the denoising process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 553, + 504, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 504, + 565 + ], + "type": "text", + "content": "3.4. Inference with Additional Conditions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 571, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 713 + ], + "type": "text", + "content": "We can further adjust additional conditions of text prompts and noise levels in Upscale-A-Video to influence the diffusion denoising process. Text prompts can guide the generation of texture details, such as animal fur, as shown in Fig. 7. Besides, adjusting the noise level allows us to balance the model's restoration and generation abilities, with smaller values favoring restoration and larger values promoting the generation of more details (see comparisons in Fig. 7). We also adopt Classifier-Free Guidance (CFG) [23] during inference, which can significantly enhance the impact of both text prompts and noise levels, helping produce high-quality videos with finer details." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2539" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 108, + 541, + 341 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on different VSR benchmarks from diverse sources, i.e., synthetic (SPMCS, UDM10, REDS30, YouHQ40), real (VideoLQ), and AIGC (AIGC30) data. The best and second performances are marked in red and orange, respectively. " + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}} (\\times 10^{-3})" + }, + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 108, + 541, + 341 + ], + "lines": [ + { + "bbox": [ + 53, + 108, + 541, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 541, + 341 + ], + "type": "table", + "html": "
DatasetsMetricsReal-ESRGAN [65]SD × 4 Upscaler [2]ResShift [83]StableSR [62]RealVSR [80]DBVSR [47]RealBasicVSR [10]Ours
SPMCSPSNR ↑22.8923.1923.2722.7123.8824.2824.5125.32
SSIM ↑0.6690.6310.6670.6570.6810.7260.7170.741
LPIPS ↓0.2380.3040.2570.2310.4370.3020.1980.222
E*arp ↓1.3645.0084.9424.8150.2941.3600.5590.367
UDM10PSNR ↑27.1328.0727.6226.4527.3829.6029.1130.79
SSIM ↑0.8430.8110.8270.8250.8250.8800.8760.878
LPIPS ↓0.1900.1860.2220.1810.2780.1550.1720.133
E*arp ↓1.4621.7102.1962.7970.5311.9430.6020.446
REDS30PSNR ↑22.4022.9823.0023.7223.0524.3723.9124.41
SSIM ↑0.5910.5720.5800.6350.6030.6330.6360.631
LPIPS ↓0.3030.3990.3690.3520.6580.5880.2490.335
E*arp ↓3.6583.7534.1311.6450.3789.6591.5571.278
YouHQ40PSNR ↑24.3719.7123.7724.5324.1925.3724.0925.83
SSIM ↑0.7100.5790.6540.7110.6950.7190.6890.733
LPIPS ↓0.2720.4420.3760.2710.4840.4300.3060.268
E*arp ↓1.8563.3994.4261.5290.4851.1491.0520.737
VideoLQCLIP-IQA ↑0.3600.1580.4300.3440.2110.2740.3870.530
MUSIQ ↑49.4826.2140.9544.2324.5229.1555.3357.99
DOVER ↑7.1612.8844.6796.7832.5313.6287.5627.811
AIGC30CLIP-IQA ↑0.4300.3290.5690.4670.2760.2900.5650.674
MUSIQ ↑47.0935.3043.3244.9324.3927.2258.8757.66
DOVER ↑9.7105.6467.0429.6683.2853.52310.6811.67
", + "image_path": "1f65d759e43ae0346f3e82678727a86eb4b688d650a6803a92cfa5d5255868e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 355, + 128, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 128, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 128, + 369 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 375, + 208, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 208, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 208, + 388 + ], + "type": "text", + "content": "4.1. Datasets and Implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "content": "Training Datasets. We train our Upscale-A-Video using the following datasets: 1) The subset of WebVid10M [5] contains around " + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "inline_equation", + "content": "335\\mathrm{K}" + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "content": " video-text pairs with each resolution around " + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "inline_equation", + "content": "336\\times 596" + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "content": ", which is commonly used in training video diffusion models [6, 11, 17, 84]; 2) YouHQ dataset. Due to the lack of high-quality video data for training, we additionally collect a large-scale high-definition " + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "inline_equation", + "content": "(1080\\times 1920)" + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "content": " dataset from YouTube, containing around " + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "inline_equation", + "content": "37\\mathrm{K}" + }, + { + "bbox": [ + 46, + 393, + 287, + 559 + ], + "type": "text", + "content": " video clips with diverse scenarios, i.e., street view, landscape, animal, human face, static object, nighttime scene, and among others. Training on such high-quality data further enhances the generation ability of our model for real-world VSR. Following the degradation pipeline of RealBasicVSR [10], we generate the LQ-HQ video pairs for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 666 + ], + "type": "text", + "content": "Testing Datasets. We construct four synthetic testing datasets (i.e., SPMCS [81], UDM10 [57], REDS30 [45], and YouHQ40), which follow the same degradation pipeline in training to generate LQ videos. We split YouHQ40 test set from the proposed YouHQ dataset, containing 40 videos. Additionally, we evaluate the models on a real-world dataset (i.e., VideoLQ [10]) and an AIGC dataset (i.e., AIGC30) that collects 30 AI-generated videos by popular text-to-video generation models [1, 3, 6, 17, 25, 41, 68, 84]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "Training Details. Our Upscale-A-Video is trained on 32 NVIDIA A100-80G GPUs with a batch size of 384. The training data is cropped to " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "80 \\times 80" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " with a length of 8. The learning rate is set to " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " using the Adam [34]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 356, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 486 + ], + "type": "text", + "content": "optimizer. We first train the U-Net model on both WebVid10M [5] and YouHQ for 70K iterations. Then we train another 10K iterations on YouHQ only. Since there are no text prompts for YouHQ, we use the null prompts during training. In this way, our model can handle VSR with either LQ inputs and prompts or LQ inputs only, leading to a more flexible use in practice. As for VAE-Decoder finetuning, we follow StableSR [62] to first generate 100K synthetic LQ-HQ video pairs on WebVid10M [5] and YouHQ, and the finetuned U-Net model is adopted to generate the corresponding latent codes for the LQ videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt different metrics to evaluate both the frame quality as well as temporal coherency of the generated results. For synthetic datasets with LQ-HQ pairs, we employ PSNR, SSIM, LPIPS [86], and flow warping error [35] " + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "content": " for evaluation. For real-world and AIGC test data, since no ground-truth videos are available, we conduct our evaluation on commonly used non-reference metrics, i.e., CLIP-IQA [61], MUSIQ [33], and DOVER [70]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 589, + 391, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 589, + 391, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 391, + 601 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 606, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 654 + ], + "type": "text", + "content": "We compare Upscale-A-Video with several state-of-the-art methods, including Real-ESRGAN [65], SD " + }, + { + "bbox": [ + 304, + 606, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 606, + 545, + 654 + ], + "type": "text", + "content": " Upscaler [2], ResShift [83], StableSR [62], RealVSR [80], DBVSR [47], and RealBasicVSR [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "Quantitative Evaluation. As shown in Table 1, our approach achieves the highest PSNR across all four synthetic datasets, suggesting its outstanding reconstruction ability. Moreover, our approach achieves the lowest LPIPS scores on both UDM10 and YouHQ40, indicating the high percep" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2540" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 161, + 134 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 161, + 134 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 161, + 134 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 161, + 134 + ], + "type": "image", + "image_path": "714a240ff7c09fc906f532ce5525969dc40879d00ad16590e2b3f7c7c1afaa89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 136, + 157, + 145 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 157, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 157, + 145 + ], + "type": "text", + "content": "\"Tourists and brick walls\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 225, + 134 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 225, + 134 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 225, + 134 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 225, + 134 + ], + "type": "image", + "image_path": "eddef5a15146430ba9ddd29f83a0a221ba5a69bde2a4e821d590b8a4a3647c4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 136, + 204, + 144 + ], + "lines": [ + { + "bbox": [ + 181, + 136, + 204, + 144 + ], + "spans": [ + { + "bbox": [ + 181, + 136, + 204, + 144 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 226, + 72, + 288, + 134 + ], + "blocks": [ + { + "bbox": [ + 226, + 72, + 288, + 134 + ], + "lines": [ + { + "bbox": [ + 226, + 72, + 288, + 134 + ], + "spans": [ + { + "bbox": [ + 226, + 72, + 288, + 134 + ], + "type": "image", + "image_path": "212c789eb3a9b7dec78947c794bde3553b778d20c35ca73e0c14abd6d83b6ca3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 136, + 269, + 144 + ], + "lines": [ + { + "bbox": [ + 242, + 136, + 269, + 144 + ], + "spans": [ + { + "bbox": [ + 242, + 136, + 269, + 144 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 289, + 72, + 352, + 134 + ], + "blocks": [ + { + "bbox": [ + 289, + 72, + 352, + 134 + ], + "lines": [ + { + "bbox": [ + 289, + 72, + 352, + 134 + ], + "spans": [ + { + "bbox": [ + 289, + 72, + 352, + 134 + ], + "type": "image", + "image_path": "05d86810a377a8eaf51e1feebbf934124069df3d4d51dc315fab1b220b81a4b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 295, + 136, + 342, + 144 + ], + "lines": [ + { + "bbox": [ + 295, + 136, + 342, + 144 + ], + "spans": [ + { + "bbox": [ + 295, + 136, + 342, + 144 + ], + "type": "text", + "content": "SD " + }, + { + "bbox": [ + 295, + 136, + 342, + 144 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 295, + 136, + 342, + 144 + ], + "type": "text", + "content": " Upscaler" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 415, + 134 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 415, + 134 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 415, + 134 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 415, + 134 + ], + "type": "image", + "image_path": "d9b35c85201bf7d73e5f1f20d4ca57f08ed6c1627afe6afdcf12d655ee62a0ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 136, + 404, + 144 + ], + "lines": [ + { + "bbox": [ + 360, + 136, + 404, + 144 + ], + "spans": [ + { + "bbox": [ + 360, + 136, + 404, + 144 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 477, + 134 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 477, + 134 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 477, + 134 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 477, + 134 + ], + "type": "image", + "image_path": "5dce61d072a26de0e18f29feca9094c4fabce8b95206a513cb9298de879c4d96.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 136, + 454, + 144 + ], + "lines": [ + { + "bbox": [ + 437, + 136, + 454, + 144 + ], + "spans": [ + { + "bbox": [ + 437, + 136, + 454, + 144 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 478, + 72, + 541, + 134 + ], + "blocks": [ + { + "bbox": [ + 478, + 72, + 541, + 134 + ], + "lines": [ + { + "bbox": [ + 478, + 72, + 541, + 134 + ], + "spans": [ + { + "bbox": [ + 478, + 72, + 541, + 134 + ], + "type": "image", + "image_path": "9c2379282043e0eb315e1c4f5b42b51479528eb64d1bf478763ff4f65d35ab85.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 136, + 514, + 144 + ], + "lines": [ + { + "bbox": [ + 503, + 136, + 514, + 144 + ], + "spans": [ + { + "bbox": [ + 503, + 136, + 514, + 144 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 148, + 161, + 212 + ], + "blocks": [ + { + "bbox": [ + 51, + 148, + 161, + 212 + ], + "lines": [ + { + "bbox": [ + 51, + 148, + 161, + 212 + ], + "spans": [ + { + "bbox": [ + 51, + 148, + 161, + 212 + ], + "type": "image", + "image_path": "20dbb699894d42e0912690a79591f0b79737db4609f3e044c01a92d0154ae452.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 213, + 148, + 221 + ], + "lines": [ + { + "bbox": [ + 65, + 213, + 148, + 221 + ], + "spans": [ + { + "bbox": [ + 65, + 213, + 148, + 221 + ], + "type": "text", + "content": "\"A koala on the tree\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 162, + 148, + 225, + 212 + ], + "blocks": [ + { + "bbox": [ + 162, + 148, + 225, + 212 + ], + "lines": [ + { + "bbox": [ + 162, + 148, + 225, + 212 + ], + "spans": [ + { + "bbox": [ + 162, + 148, + 225, + 212 + ], + "type": "image", + "image_path": "827e11544d132b366dcaf230ae35c56713b87de86b5c4743687e552f36bde470.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 213, + 206, + 221 + ], + "lines": [ + { + "bbox": [ + 182, + 213, + 206, + 221 + ], + "spans": [ + { + "bbox": [ + 182, + 213, + 206, + 221 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 226, + 148, + 288, + 212 + ], + "blocks": [ + { + "bbox": [ + 226, + 148, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 226, + 148, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 226, + 148, + 288, + 212 + ], + "type": "image", + "image_path": "8d122ffb705c49dde6220795713418555d439bd9fefda338e72e89d117925940.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 213, + 272, + 221 + ], + "lines": [ + { + "bbox": [ + 244, + 213, + 272, + 221 + ], + "spans": [ + { + "bbox": [ + 244, + 213, + 272, + 221 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 289, + 148, + 351, + 212 + ], + "blocks": [ + { + "bbox": [ + 289, + 148, + 351, + 212 + ], + "lines": [ + { + "bbox": [ + 289, + 148, + 351, + 212 + ], + "spans": [ + { + "bbox": [ + 289, + 148, + 351, + 212 + ], + "type": "image", + "image_path": "9bb9e1f04ef26e24850235a82b8dedc942adb69b967ae3438e696fe17cc64d7f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 213, + 345, + 222 + ], + "lines": [ + { + "bbox": [ + 297, + 213, + 345, + 222 + ], + "spans": [ + { + "bbox": [ + 297, + 213, + 345, + 222 + ], + "type": "text", + "content": "SD " + }, + { + "bbox": [ + 297, + 213, + 345, + 222 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 297, + 213, + 345, + 222 + ], + "type": "text", + "content": " Upscaler" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 352, + 148, + 414, + 212 + ], + "blocks": [ + { + "bbox": [ + 352, + 148, + 414, + 212 + ], + "lines": [ + { + "bbox": [ + 352, + 148, + 414, + 212 + ], + "spans": [ + { + "bbox": [ + 352, + 148, + 414, + 212 + ], + "type": "image", + "image_path": "c93d176b8adb15ab50df28fd45c24462ac6ece1567648e60ef7a05656ad08ae8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 213, + 406, + 221 + ], + "lines": [ + { + "bbox": [ + 362, + 213, + 406, + 221 + ], + "spans": [ + { + "bbox": [ + 362, + 213, + 406, + 221 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 414, + 148, + 477, + 212 + ], + "blocks": [ + { + "bbox": [ + 414, + 148, + 477, + 212 + ], + "lines": [ + { + "bbox": [ + 414, + 148, + 477, + 212 + ], + "spans": [ + { + "bbox": [ + 414, + 148, + 477, + 212 + ], + "type": "image", + "image_path": "4ae99125339eac3e6215589bab518d5dd42e0f0e7b27d1fd3e64ca01a46089dc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 213, + 456, + 221 + ], + "lines": [ + { + "bbox": [ + 439, + 213, + 456, + 221 + ], + "spans": [ + { + "bbox": [ + 439, + 213, + 456, + 221 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 478, + 148, + 541, + 212 + ], + "blocks": [ + { + "bbox": [ + 478, + 148, + 541, + 212 + ], + "lines": [ + { + "bbox": [ + 478, + 148, + 541, + 212 + ], + "spans": [ + { + "bbox": [ + 478, + 148, + 541, + 212 + ], + "type": "image", + "image_path": "3104465d82a45808957055283046a76c8491d6068a3574c288a90b1762fcd827.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 213, + 515, + 220 + ], + "lines": [ + { + "bbox": [ + 504, + 213, + 515, + 220 + ], + "spans": [ + { + "bbox": [ + 504, + 213, + 515, + 220 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 50, + 259, + 187, + 367 + ], + "blocks": [ + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on synthetic low-quality videos from REDS30 [45] and YouHQ40 datasets. Among the tested methods, only our Upscale-A-Video can recover the accurate wall structure and produce detailed koala fur. (Zoom-in for best view)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 259, + 187, + 367 + ], + "lines": [ + { + "bbox": [ + 50, + 259, + 187, + 367 + ], + "spans": [ + { + "bbox": [ + 50, + 259, + 187, + 367 + ], + "type": "image", + "image_path": "081f2c229fa7023b838f5c7ca09f6efb280225a4d1ac68be0558012272cbc736.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 369, + 180, + 378 + ], + "lines": [ + { + "bbox": [ + 57, + 369, + 180, + 378 + ], + "spans": [ + { + "bbox": [ + 57, + 369, + 180, + 378 + ], + "type": "text", + "content": "\"A bus and car on the street\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 189, + 259, + 277, + 308 + ], + "blocks": [ + { + "bbox": [ + 189, + 259, + 277, + 308 + ], + "lines": [ + { + "bbox": [ + 189, + 259, + 277, + 308 + ], + "spans": [ + { + "bbox": [ + 189, + 259, + 277, + 308 + ], + "type": "image", + "image_path": "d6a32ced86f08cd95bf1f6776bc62508001e9d2ed4f77fd06019d238ff463c5d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 309, + 246, + 317 + ], + "lines": [ + { + "bbox": [ + 220, + 309, + 246, + 317 + ], + "spans": [ + { + "bbox": [ + 220, + 309, + 246, + 317 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 278, + 259, + 366, + 308 + ], + "blocks": [ + { + "bbox": [ + 278, + 259, + 366, + 308 + ], + "lines": [ + { + "bbox": [ + 278, + 259, + 366, + 308 + ], + "spans": [ + { + "bbox": [ + 278, + 259, + 366, + 308 + ], + "type": "image", + "image_path": "c5eb6f7f68047ebd623560ef9a8548426f3f1a4b1e24a238acf0db166011005b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 309, + 346, + 317 + ], + "lines": [ + { + "bbox": [ + 298, + 309, + 346, + 317 + ], + "spans": [ + { + "bbox": [ + 298, + 309, + 346, + 317 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 367, + 259, + 455, + 308 + ], + "blocks": [ + { + "bbox": [ + 367, + 259, + 455, + 308 + ], + "lines": [ + { + "bbox": [ + 367, + 259, + 455, + 308 + ], + "spans": [ + { + "bbox": [ + 367, + 259, + 455, + 308 + ], + "type": "image", + "image_path": "db4b517bc3358e2093f5e0a8c226e990c8cc14c51285cc0f8c76e72a3ee70dea.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 309, + 424, + 317 + ], + "lines": [ + { + "bbox": [ + 398, + 309, + 424, + 317 + ], + "spans": [ + { + "bbox": [ + 398, + 309, + 424, + 317 + ], + "type": "text", + "content": "ResShift" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 457, + 259, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 457, + 259, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 457, + 259, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 457, + 259, + 545, + 308 + ], + "type": "image", + "image_path": "8736c44e3c24a2f064c1700ea4729efc5b5fd6b7447b2840677b1751e2469472.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 309, + 515, + 317 + ], + "lines": [ + { + "bbox": [ + 485, + 309, + 515, + 317 + ], + "spans": [ + { + "bbox": [ + 485, + 309, + 515, + 317 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 50, + 384, + 187, + 488 + ], + "blocks": [ + { + "bbox": [ + 50, + 384, + 187, + 488 + ], + "lines": [ + { + "bbox": [ + 50, + 384, + 187, + 488 + ], + "spans": [ + { + "bbox": [ + 50, + 384, + 187, + 488 + ], + "type": "image", + "image_path": "d317f876620dca3768ba11ffa52042eed27fb858227c52609ea3b3515adac9ae.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 490, + 163, + 498 + ], + "lines": [ + { + "bbox": [ + 74, + 490, + 163, + 498 + ], + "spans": [ + { + "bbox": [ + 74, + 490, + 163, + 498 + ], + "type": "text", + "content": "\"A squirrel on a tree\"" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 504, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 545, + 548 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons on real-world test videos in VideoLQ [10] dataset. Our Upscale-A-Video effectively leverages the advantages of the diffusion paradigm in generating high-quality results. When compared to existing methods, it notably excels in its restoration capabilities, successfully recovering the billboard word \"EAT IN or TAKEAWAY\". In particular, when guided by text prompts, Upscale-A-Video showcases promising enhanced results with more details and heightened realism. (Zoom-in for best view)" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 189, + 319, + 277, + 367 + ], + "blocks": [ + { + "bbox": [ + 189, + 319, + 277, + 367 + ], + "lines": [ + { + "bbox": [ + 189, + 319, + 277, + 367 + ], + "spans": [ + { + "bbox": [ + 189, + 319, + 277, + 367 + ], + "type": "image", + "image_path": "95e888a37c95addd25463f052af486590917aafba8aa8e7840a434448f239961.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 369, + 258, + 378 + ], + "lines": [ + { + "bbox": [ + 208, + 369, + 258, + 378 + ], + "spans": [ + { + "bbox": [ + 208, + 369, + 258, + 378 + ], + "type": "text", + "content": "SD " + }, + { + "bbox": [ + 208, + 369, + 258, + 378 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 208, + 369, + 258, + 378 + ], + "type": "text", + "content": " Upscaler" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 278, + 319, + 366, + 367 + ], + "blocks": [ + { + "bbox": [ + 278, + 319, + 366, + 367 + ], + "lines": [ + { + "bbox": [ + 278, + 319, + 366, + 367 + ], + "spans": [ + { + "bbox": [ + 278, + 319, + 366, + 367 + ], + "type": "image", + "image_path": "7b453a197467f9a0f6b4af182cb232bf1d6565eebf4ba2625c7814642852ce71.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 369, + 335, + 377 + ], + "lines": [ + { + "bbox": [ + 309, + 369, + 335, + 377 + ], + "spans": [ + { + "bbox": [ + 309, + 369, + 335, + 377 + ], + "type": "text", + "content": "DBVSR" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 367, + 319, + 455, + 367 + ], + "blocks": [ + { + "bbox": [ + 367, + 319, + 455, + 367 + ], + "lines": [ + { + "bbox": [ + 367, + 319, + 455, + 367 + ], + "spans": [ + { + "bbox": [ + 367, + 319, + 455, + 367 + ], + "type": "image", + "image_path": "fdff56df2f6214a0267e21ab77fb0023e87e69b8a95d1d8114100dc0132ae7cb.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 369, + 435, + 377 + ], + "lines": [ + { + "bbox": [ + 389, + 369, + 435, + 377 + ], + "spans": [ + { + "bbox": [ + 389, + 369, + 435, + 377 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 457, + 319, + 545, + 367 + ], + "blocks": [ + { + "bbox": [ + 457, + 319, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 457, + 319, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 457, + 319, + 545, + 367 + ], + "type": "image", + "image_path": "dcb755e3c9d39433012712208e02efc973e3f9805242ded33ffcd7cb3eab5ae7.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 369, + 538, + 378 + ], + "lines": [ + { + "bbox": [ + 463, + 369, + 538, + 378 + ], + "spans": [ + { + "bbox": [ + 463, + 369, + 538, + 378 + ], + "type": "text", + "content": "Upscale-A-Video (Ours)" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 189, + 384, + 234, + 431 + ], + "blocks": [ + { + "bbox": [ + 189, + 384, + 234, + 431 + ], + "lines": [ + { + "bbox": [ + 189, + 384, + 234, + 431 + ], + "spans": [ + { + "bbox": [ + 189, + 384, + 234, + 431 + ], + "type": "image", + "image_path": "98b891f86cabf73b2d8563139cb4688db32e2d3468bf5180b53ecdb282707d1c.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 432, + 244, + 440 + ], + "lines": [ + { + "bbox": [ + 220, + 432, + 244, + 440 + ], + "spans": [ + { + "bbox": [ + 220, + 432, + 244, + 440 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 235, + 384, + 277, + 430 + ], + "blocks": [ + { + "bbox": [ + 235, + 384, + 277, + 430 + ], + "lines": [ + { + "bbox": [ + 235, + 384, + 277, + 430 + ], + "spans": [ + { + "bbox": [ + 235, + 384, + 277, + 430 + ], + "type": "image", + "image_path": "dd1dd2900b107266a039550016e66dbf2e9c3e22c5c2687bec815d1c4d9c52d3.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 278, + 384, + 322, + 430 + ], + "blocks": [ + { + "bbox": [ + 278, + 384, + 322, + 430 + ], + "lines": [ + { + "bbox": [ + 278, + 384, + 322, + 430 + ], + "spans": [ + { + "bbox": [ + 278, + 384, + 322, + 430 + ], + "type": "image", + "image_path": "29c8ee582084cd2869576e2ac439b0be44e12bf2b78700b1f394be407e40581a.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 323, + 384, + 366, + 430 + ], + "blocks": [ + { + "bbox": [ + 323, + 384, + 366, + 430 + ], + "lines": [ + { + "bbox": [ + 323, + 384, + 366, + 430 + ], + "spans": [ + { + "bbox": [ + 323, + 384, + 366, + 430 + ], + "type": "image", + "image_path": "38e561fe0d112776873eb06d66fc19f3679280cdac7c8958ac928c723895a433.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 367, + 384, + 411, + 430 + ], + "blocks": [ + { + "bbox": [ + 367, + 384, + 411, + 430 + ], + "lines": [ + { + "bbox": [ + 367, + 384, + 411, + 430 + ], + "spans": [ + { + "bbox": [ + 367, + 384, + 411, + 430 + ], + "type": "image", + "image_path": "29fdf8793b6c9243869fda1625ecdf60d9a59336994eaf2ec4d51d32476a6ad7.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 411, + 384, + 455, + 430 + ], + "blocks": [ + { + "bbox": [ + 411, + 384, + 455, + 430 + ], + "lines": [ + { + "bbox": [ + 411, + 384, + 455, + 430 + ], + "spans": [ + { + "bbox": [ + 411, + 384, + 455, + 430 + ], + "type": "image", + "image_path": "8ef918afa2face933a1ee01da439ebd13da0c0fec52b43a1390391a3f6c42657.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 456, + 384, + 500, + 430 + ], + "blocks": [ + { + "bbox": [ + 456, + 384, + 500, + 430 + ], + "lines": [ + { + "bbox": [ + 456, + 384, + 500, + 430 + ], + "spans": [ + { + "bbox": [ + 456, + 384, + 500, + 430 + ], + "type": "image", + "image_path": "24762789705e87124e885e779316755afb6a70c93df5bf8bbd2a3312cfbe668e.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 501, + 384, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 501, + 384, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 501, + 384, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 501, + 384, + 545, + 430 + ], + "type": "image", + "image_path": "2c0e8022ce1fcce978dd0386363f27a19467556b836b1c14b4a8fdcc117a7cb0.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 189, + 441, + 234, + 488 + ], + "blocks": [ + { + "bbox": [ + 189, + 441, + 234, + 488 + ], + "lines": [ + { + "bbox": [ + 189, + 441, + 234, + 488 + ], + "spans": [ + { + "bbox": [ + 189, + 441, + 234, + 488 + ], + "type": "image", + "image_path": "616e1053c2702cfb8da6153c782d1976d082bfd7f5e2e06fde2fa6154fd6daf3.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 490, + 258, + 498 + ], + "lines": [ + { + "bbox": [ + 208, + 490, + 258, + 498 + ], + "spans": [ + { + "bbox": [ + 208, + 490, + 258, + 498 + ], + "type": "text", + "content": "SD " + }, + { + "bbox": [ + 208, + 490, + 258, + 498 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 208, + 490, + 258, + 498 + ], + "type": "text", + "content": " Upscaler" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 235, + 441, + 277, + 488 + ], + "blocks": [ + { + "bbox": [ + 235, + 441, + 277, + 488 + ], + "lines": [ + { + "bbox": [ + 235, + 441, + 277, + 488 + ], + "spans": [ + { + "bbox": [ + 235, + 441, + 277, + 488 + ], + "type": "image", + "image_path": "3ba2e51027b505b2bbdfe8111cc07e56dffd5f5f6c9bcb7f1c500090bfde9fcd.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 278, + 441, + 322, + 488 + ], + "blocks": [ + { + "bbox": [ + 278, + 441, + 322, + 488 + ], + "lines": [ + { + "bbox": [ + 278, + 441, + 322, + 488 + ], + "spans": [ + { + "bbox": [ + 278, + 441, + 322, + 488 + ], + "type": "image", + "image_path": "ddb1dfd847b81eef6e9e3cf4a67735a9ec4ef16bc8303d3385a6d2890c459020.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 490, + 335, + 498 + ], + "lines": [ + { + "bbox": [ + 309, + 490, + 335, + 498 + ], + "spans": [ + { + "bbox": [ + 309, + 490, + 335, + 498 + ], + "type": "text", + "content": "DBVSR" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 323, + 441, + 366, + 488 + ], + "blocks": [ + { + "bbox": [ + 323, + 441, + 366, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 441, + 366, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 441, + 366, + 488 + ], + "type": "image", + "image_path": "bdd2c87d4c2c464d3012f95e72c15a1c6c8ae3716d35b9f7a93a7607c56d0041.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 367, + 441, + 411, + 488 + ], + "blocks": [ + { + "bbox": [ + 397, + 432, + 424, + 440 + ], + "lines": [ + { + "bbox": [ + 397, + 432, + 424, + 440 + ], + "spans": [ + { + "bbox": [ + 397, + 432, + 424, + 440 + ], + "type": "text", + "content": "ResShift" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 441, + 411, + 488 + ], + "lines": [ + { + "bbox": [ + 367, + 441, + 411, + 488 + ], + "spans": [ + { + "bbox": [ + 367, + 441, + 411, + 488 + ], + "type": "image", + "image_path": "966a5df8a661d33bd8498fbe3cdde694b1fd0c37294b2fb2746c971a64072ed2.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 490, + 434, + 498 + ], + "lines": [ + { + "bbox": [ + 388, + 490, + 434, + 498 + ], + "spans": [ + { + "bbox": [ + 388, + 490, + 434, + 498 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 411, + 441, + 455, + 488 + ], + "blocks": [ + { + "bbox": [ + 411, + 441, + 455, + 488 + ], + "lines": [ + { + "bbox": [ + 411, + 441, + 455, + 488 + ], + "spans": [ + { + "bbox": [ + 411, + 441, + 455, + 488 + ], + "type": "image", + "image_path": "e1b21395b8afc2688462996ec5484f6e49c66bb3662758fa92088730dbba463a.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 462, + 441, + 500, + 488 + ], + "blocks": [ + { + "bbox": [ + 462, + 441, + 500, + 488 + ], + "lines": [ + { + "bbox": [ + 462, + 441, + 500, + 488 + ], + "spans": [ + { + "bbox": [ + 462, + 441, + 500, + 488 + ], + "type": "image", + "image_path": "b57edc61f47e511db960555e6de5a08803c415503a02bb642715fd4bdd98c56b.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 490, + 538, + 498 + ], + "lines": [ + { + "bbox": [ + 463, + 490, + 538, + 498 + ], + "spans": [ + { + "bbox": [ + 463, + 490, + 538, + 498 + ], + "type": "text", + "content": "Upscale-A-Video (Ours)" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "text", + "content": "tual quality of our generated results. In addition, our approach further obtains the highest CLIP-IQA and DOVER scores on both the real dataset [10] and AIGC videos. The superiority across datasets from various sources demonstrates the effectiveness of our approach." + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Qualitative Evaluation. We present visual results on both synthetic [45, 57, 81] and real-world [10] videos in Fig. 4 and Fig. 5, respectively. It is observed that our Upscale-A-Video significantly outperforms existing CNN-and diffusion-based approaches in both artifact removal and details generation. Specifically, Upscale-A-Video is capable of generating more natural details in Fig. 4, and success" + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 562, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 545, + 621 + ], + "type": "text", + "content": "fully recovers the words \"EAT IN or TAKEAWAY\" on the billboard in the first example of Fig. 5, while other methods generate blurry or distorted results. Note that SD " + }, + { + "bbox": [ + 304, + 562, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 562, + 545, + 621 + ], + "type": "text", + "content": " Upscaler shows less effectiveness for the real-world VSR since it does not consider mixed degradation in its training data." + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": "Temporal Consistency. Benefiting from our local-global temporal strategy, Upscale-A-Video achieves the best optical flow error score on UDM10 and the second-best scores on REDS30, SPMCS and YouHQ40, remarkably outperforming other diffusion-based approaches, and even beats the strong CNN-based VSR methods, i.e., RealBasicVSR and DBVSR. We also visualize the temporal profile in" + } + ] + } + ], + "index": 74 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2541" + } + ] + } + ], + "index": 75 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 60, + 164, + 186 + ], + "blocks": [ + { + "bbox": [ + 51, + 60, + 164, + 186 + ], + "lines": [ + { + "bbox": [ + 51, + 60, + 164, + 186 + ], + "spans": [ + { + "bbox": [ + 51, + 60, + 164, + 186 + ], + "type": "image", + "image_path": "9fffd9bb765ccfbb25cae668e028067edea691d067150bedc69d1f045d83d2e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "type": "text", + "content": "Figure 6. Comparison of temporal profile. We examine a row and track changes over time. The profile from existing methods and Exp. (a) and (c) (w/o latent propagation module) exhibit noise, suggesting the presence of flickering artifacts. Even with the finetuned decoder, the profile of Exp. (c) still displays some discontinuities. Thanks to long-term latent propagation and aggregation, the profile of our Upscale-A-Video exhibits a more seamless and smoother transition." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 541, + 186 + ], + "type": "image", + "image_path": "ad165445da421d911bbb003fe30da0fa97ca0db6ca21c1a6395e881ac543629e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 270, + 277, + 337 + ], + "blocks": [ + { + "bbox": [ + 47, + 245, + 287, + 267 + ], + "lines": [ + { + "bbox": [ + 47, + 245, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 287, + 267 + ], + "type": "text", + "content": "Table 2. Ablation study of finetuned VAE-Decoder and propagation module on YouHQ40." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 270, + 277, + 337 + ], + "lines": [ + { + "bbox": [ + 53, + 270, + 277, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 277, + 337 + ], + "type": "table", + "html": "
Exp.ft-VAE-Dec.Latent Prop.PSNR↑SSIM↑E*warp↓
(a)23.820.63852.398
(b)✓25.470.72151.815
(c)✓25.750.73280.842
(d)✓✓25.830.73260.737
", + "image_path": "d2a6717a4dbd376d201101ebaa19c3bd6acb56481cf923643252cd188b9b0e0c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 374, + 141, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 141, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 141, + 387 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "content": "Effectiveness of Finetuned VAE-Decoder. We first investigate the significance of the fine-tuned VAE-Decoder. As shown in Table 2, replacing our finetuned VAE-Decoder with the original decoder leads to worse PSNR, SSIM, and " + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "content": ". Particularly, the increase of " + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "content": " from 0.737 to 1.815 indicates a significant deterioration of temporal coherency. The comparisons in Fig. 6 also suggest inferior temporal consistency without the finetuned VAE-Decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 488, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 582 + ], + "type": "text", + "content": "Effectiveness of Propagation Module. In addition to fine-tuning VAE-Decoder, our proposed flow-guided recurrent latent propagation module further enhances the stability of long videos. As shown in Table 2, adopting the propagation module can further reduce " + }, + { + "bbox": [ + 46, + 488, + 287, + 582 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 46, + 488, + 287, + 582 + ], + "type": "text", + "content": " error, effectively improving temporal consistency while maintaining high PSNR. Similar phenomena can also be observed in the temporal profile in Fig. 6, showing a more seamless transition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 583, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 666 + ], + "type": "text", + "content": "Text Prompt. Upscale-A-Video is trained on video data with either labeled prompts or null prompts and thus can handle both scenarios. We examine the use of classifier-free guidance [23] to improve the visual quality during sampling. As shown in Figure 7, compared with the null prompt, adopting proper text prompts can significantly enhance the perceptual quality with more faithful details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Noise Level. It is observed that the level of noise added to the input can affect the performance of our approach. As shown in Fig. 7, with low noise level, the generated results tend to be suboptimal with blurry details. However, a noise" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 245, + 384, + 312 + ], + "blocks": [ + { + "bbox": [ + 308, + 245, + 384, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 245, + 384, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 384, + 312 + ], + "type": "image", + "image_path": "048643498f9fe39cf18c8f196b96380e5496c17029ea42efea84550cdf51c160.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 314, + 354, + 322 + ], + "lines": [ + { + "bbox": [ + 337, + 314, + 354, + 322 + ], + "spans": [ + { + "bbox": [ + 337, + 314, + 354, + 322 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 387, + 245, + 463, + 312 + ], + "blocks": [ + { + "bbox": [ + 387, + 245, + 463, + 312 + ], + "lines": [ + { + "bbox": [ + 387, + 245, + 463, + 312 + ], + "spans": [ + { + "bbox": [ + 387, + 245, + 463, + 312 + ], + "type": "image", + "image_path": "cab392bf147f095abf9a3f2520a0c016aeeb9e8218cc513ba0da073f9a4fb69e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 314, + 449, + 322 + ], + "lines": [ + { + "bbox": [ + 400, + 314, + 449, + 322 + ], + "spans": [ + { + "bbox": [ + 400, + 314, + 449, + 322 + ], + "type": "text", + "content": "w/o Text Prompt" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 465, + 245, + 542, + 312 + ], + "blocks": [ + { + "bbox": [ + 465, + 245, + 542, + 312 + ], + "lines": [ + { + "bbox": [ + 465, + 245, + 542, + 312 + ], + "spans": [ + { + "bbox": [ + 465, + 245, + 542, + 312 + ], + "type": "image", + "image_path": "fb750d75bae376c0cbaeecb6b580833beda50d4d131a8fe15d1dc66cf693f4cb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 323, + 384, + 390 + ], + "blocks": [ + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "type": "text", + "content": "Fig. 6. It is observed that our approach achieves superior performance with a more seamless and smoother transition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 323, + 384, + 390 + ], + "lines": [ + { + "bbox": [ + 308, + 323, + 384, + 390 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 384, + 390 + ], + "type": "image", + "image_path": "dba03b83ee73b8209438e44e1a1a3bae2db21bc0eaa790fa509f68f2bfd39e24.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 392, + 366, + 399 + ], + "lines": [ + { + "bbox": [ + 326, + 392, + 366, + 399 + ], + "spans": [ + { + "bbox": [ + 326, + 392, + 366, + 399 + ], + "type": "text", + "content": "Noise Level 0" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 399, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 421 + ], + "type": "text", + "content": "Figure 7. Comparison of various conditions of text prompts and different levels of noise. (Zoom-in for best view)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 387, + 323, + 463, + 390 + ], + "blocks": [ + { + "bbox": [ + 387, + 323, + 463, + 390 + ], + "lines": [ + { + "bbox": [ + 387, + 323, + 463, + 390 + ], + "spans": [ + { + "bbox": [ + 387, + 323, + 463, + 390 + ], + "type": "image", + "image_path": "6ec99efab4d96c444171af13754a9a436fb06e11d3a45240827e8d9b47e157d3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 392, + 447, + 399 + ], + "lines": [ + { + "bbox": [ + 403, + 392, + 447, + 399 + ], + "spans": [ + { + "bbox": [ + 403, + 392, + 447, + 399 + ], + "type": "text", + "content": "Noise Level 50" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 465, + 323, + 542, + 390 + ], + "blocks": [ + { + "bbox": [ + 481, + 314, + 526, + 322 + ], + "lines": [ + { + "bbox": [ + 481, + 314, + 526, + 322 + ], + "spans": [ + { + "bbox": [ + 481, + 314, + 526, + 322 + ], + "type": "text", + "content": "w/ Text Prompt" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 323, + 542, + 390 + ], + "lines": [ + { + "bbox": [ + 465, + 323, + 542, + 390 + ], + "spans": [ + { + "bbox": [ + 465, + 323, + 542, + 390 + ], + "type": "image", + "image_path": "eff9969acf979371b7d3e1bdb538422ec584abb85f961a777c264779d1466a3a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 392, + 528, + 399 + ], + "lines": [ + { + "bbox": [ + 480, + 392, + 528, + 399 + ], + "spans": [ + { + "bbox": [ + 480, + 392, + 528, + 399 + ], + "type": "text", + "content": "Noise Level 150" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 433, + 512, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 512, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 512, + 445 + ], + "type": "text", + "content": "level that is too large may result in oversharpening." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 475, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 653 + ], + "type": "text", + "content": "While diffusion models have achieved impressive performance on a wide range of image tasks, their applications on video tasks, particularly real-world VSR remain challenging and understudied. In this paper, we present Upscale-A-Video, a novel approach to exploit image diffusion prior for real-world VSR while avoiding temporal discontinuities drawn from the inherent randomness during sampling process. Specifically, we enhance the temporal coherence by proposing a novel local-global temporal strategy within the latent diffusion framework. We additionally devote our efforts to achieving a trade-off between fidelity and quality by enabling texture creation with text prompts and noise level control, further facilitating the practical use in real-world scenarios. We believe that our exploration would lay a good foundation for future works." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). The project is also supported by Shanghai AI Laboratory." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2542" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 257, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 257, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 257, + 101 + ], + "type": "text", + "content": "[1] Pika Labs. https://www.pika.art/, 2023.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "type": "text", + "content": "[2] Stable Diffusion x4 Upscaler. https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler,2023.2,3,4,6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 135, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 286, + 156 + ], + "type": "text", + "content": "[3] Zeroscope V2 XL. https://replicate.com/anotherjesse/zeroscope-v2-x1, 2023.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 157, + 286, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 286, + 187 + ], + "type": "text", + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 190, + 286, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 286, + 220 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 223, + 286, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 223, + 286, + 265 + ], + "spans": [ + { + "bbox": [ + 53, + 223, + 286, + 265 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "text", + "content": "[7] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 299, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 286, + 341 + ], + "type": "text", + "content": "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "type": "text", + "content": "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 386, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 286, + 418 + ], + "type": "text", + "content": "[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 419, + 286, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 419, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 419, + 286, + 472 + ], + "type": "text", + "content": "[11] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 474, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 286, + 506 + ], + "type": "text", + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 506, + 286, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 286, + 538 + ], + "type": "text", + "content": "[13] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 539, + 286, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 539, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 539, + 286, + 592 + ], + "type": "text", + "content": "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical flow-guided attention for consistent text-to-video editing. arXiv preprint arXiv:2310.05922, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 594, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 594, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 286, + 635 + ], + "type": "text", + "content": "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 638, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 286, + 680 + ], + "type": "text", + "content": "[16] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[17] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "A noise prior for video diffusion models. In ICCV, 2023. 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 163 + ], + "type": "text", + "content": "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 206 + ], + "type": "text", + "content": "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPSW, 2022. 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 321, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 342 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "text", + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen Video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "text", + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 464 + ], + "type": "text", + "content": "[27] Yaosi Hu, Zhenzhong Chen, and Chong Luo. LaMD: Latent motion diffusion for video generation. arXiv preprint arXiv:2304.11603, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "text", + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 501, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 543 + ], + "type": "text", + "content": "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 545, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 576 + ], + "type": "text", + "content": "[30] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "text", + "content": "[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 624, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 654 + ], + "type": "text", + "content": "[32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[33] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In ICCV, 2021. 6" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "text", + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2543" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "text", + "content": "[36] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 203 + ], + "type": "text", + "content": "[37] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 249 + ], + "type": "text", + "content": "[38] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 270 + ], + "type": "text", + "content": "[39] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. In IEEE TPAMI, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "type": "text", + "content": "[40] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. VDT: An empirical study on video diffusion with transformers. arXiv preprint arXiv:2305.13311, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "text", + "content": "[41] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 380 + ], + "type": "text", + "content": "[42] Kangfu Mei and Vishal Patel. VIDM: Video implicit diffusion models. In AAAI, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 414 + ], + "type": "text", + "content": "[43] Simon Meister, Junhwa Hur, and Stefan Roth. UnFlow: Unsupervised learning of optical flow with a bidirectional census loss. In AAAI, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[44] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "text", + "content": "[45] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 504, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 556 + ], + "type": "text", + "content": "[46] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[47] Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. In ICCV, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "type": "text", + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. FateZero: Fusing attentions for zero-shot text-based video editing. ICCV, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[51] Hshmat Sahak, Daniel Watson, Chitwan Sahara, and David Fleet. Denoising diffusion probabilistic models for robust image super-resolution in the wild. arXiv preprint arXiv:2302.07864, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 207, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 228 + ], + "type": "text", + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "text", + "content": "[56] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In ICLR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "type": "text", + "content": "[57] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "text", + "content": "[58] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "text", + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "text", + "content": "[61] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[62] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[63] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "text", + "content": "[64] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[65] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In ICCVW, 2021. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 2, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[67] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2022. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2544" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[68] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[69] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[70] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "type": "text", + "content": "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 317 + ], + "type": "text", + "content": "[72] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. DiffIR: Efficient diffusion model for image restoration. ICCV, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "text", + "content": "[73] Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao Dong, and Ying Shan. Mitigating artifacts in real-world video super-resolution models. In AAAI, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[74] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple diffusion adapter for efficient video generation. arXiv preprint arXiv:2308.09710, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[75] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. In ICCV, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[76] Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. PGDiff: Guiding diffusion models for versatile face restoration via partial guidance. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[77] S. Yang, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[78] Shuai Yang, Yifan Zhou, Ziwei Liu, , and Chen Change Loy. Rerender A Video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia, 2023. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[79] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "type": "text", + "content": "[80] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In ICCV, 2021. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[81] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[82] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 350 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[83] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. ResShift: Efficient diffusion model for image superresolution by residual shifting. In NeurIPS, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[84] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[85] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "text", + "content": "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 240, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 282 + ], + "type": "text", + "content": "[87] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 285, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 317 + ], + "type": "text", + "content": "[88] Shangchen Zhou, Jiawei Zhang, Jinshan Pan, Haozhe Xie, Wangmeng Zuo, and Jimmy Ren. Spatio-temporal filter adaptive network for video deblurring. In ICCV, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "text", + "content": "[89] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. ProPainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 5" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2545" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_content_list.json b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa78eec4a9c8a76b462f96759c8b197203594b5e --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_content_list.json @@ -0,0 +1,1644 @@ +[ + { + "type": "text", + "text": "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model", + "text_level": 1, + "bbox": [ + 210, + 130, + 759, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Yang $^{1*}$ Jian Tao $^{1*}$ Jiafei Lyu $^{1\\dagger}$ Chunjiang Ge $^{2}$", + "bbox": [ + 264, + 202, + 702, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxin Chen3 Weihan Shen3 Xiaolong Zhu3 Xiu Li1†", + "bbox": [ + 248, + 220, + 718, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University", + "bbox": [ + 192, + 239, + 772, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Automation, Tsinghua University $^{3}$ Parametrix Technology Company Ltd.", + "bbox": [ + 117, + 256, + 849, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yk22,tj22,lvjf20}@mails.tsinghua.edu.cn li.xiu@sz.tsinghua.edu.cn", + "bbox": [ + 183, + 276, + 779, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 312, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO.", + "bbox": [ + 76, + 358, + 472, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in image generation models have yielded unprecedented success in producing high-quality images", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from textual prompts [30, 44, 49]. Diverse approaches, including Generative Adversarial Networks (GANs) [23], autoregressive models [14, 15, 19, 22, 44, 58], Normalizing Flows [16, 46], and diffusion-based techniques [40, 45, 47, 49], have rapidly pushed forward the capabilities of these systems. With the proper textual inputs, such models are now adept at crafting images that are not only visually compelling but also semantically coherent, garnering widespread interest for their potential applications and implications.", + "bbox": [ + 500, + 327, + 890, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To adapt the aforementioned models for specific downstream tasks, such as the generation of more visually appealing and aesthetic images, Reinforcement Learning from Human Feedback (RLHF) is commonly employed [12]. This technique has been successfully used to refine large language models such as GPT [8, 41]. Now, efforts are being made to extend this method to diffusion models to enhance their performance. One such approach, the DDPO method [5], aims to enhance image complexity, aesthetic quality, and the alignment between prompts and images. The ReLF approach [60] introduces a novel reward model, named ImageReward, which is specifically trained to discern human aesthetic preferences in text-to-image synthesis. This model is then utilized to fine-tune diffusion models to produce images that align more closely with human preferences. Nonetheless, developing a robust reward model for various tasks can be challenging, often necessitating a vast collection of images and abundant training resources. For example, to diminish the rate of deformities in character images, one must amass a substantial dataset of deformed and non-deformed images generated from identical prompts. Subsequently, a network is constructed to discern and learn the human preference for non-deformed imagery, serving as the reward model.", + "bbox": [ + 496, + 484, + 892, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of natural language processing, Direct Preference Optimization (DPO) has been proposed to reduce training costs [43]. This method forgoes the training of a", + "bbox": [ + 498, + 854, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding authors.", + "bbox": [ + 94, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef3b85de653712224e547e7d86884b6bac70ef933b10322e1f6a5616f8fb4c4a.jpg", + "image_caption": [ + "Figure 1. Overview of D3PO. The diffusion model generates two corresponding images based on the provided prompts. Guided by specific task requirements—such as improving prompt-image alignment, enhancing image incompressibility, or refining aesthetic quality—human evaluators select the preferred image. Leveraging this human feedback, our method directly updates the diffusion model's parameters without necessitating the training of a reward model." + ], + "image_footnote": [], + "bbox": [ + 122, + 92, + 848, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reward model and directly fine-tunes language models according to human preferences. However, this straightforward and easy-to-train method encounters challenges when applied to fine-tune diffusion models. During the DPO training process, the complete sentence generated by the language model is treated as a single output, necessitating the storage of gradients from multiple forward passes. With diffusion models, one must store the gradients across multiple latent image representations, which are significantly larger than word embeddings, leading to memory consumption that is typically unsustainable.", + "bbox": [ + 75, + 472, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the issue of high computational overhead and enable the use of the DPO method to fine-tune diffusion models directly with human feedback, we conceptualize the denoising process as a multi-step MDP, which utilizes a pre-trained model to represent an action value function $Q$ that is commonly estimated in RL. We extend the theoretical framework of DPO into the formulated MDP, which allows for direct parameter updates at each step of the denoising process based on human feedback, thereby circumventing the significant computational costs and eliminating the need for a reward model. To the best of our knowledge, this is the first work that fine-tune diffusion models without reward models. Our main contributions are as follows:", + "bbox": [ + 75, + 641, + 470, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce an innovative approach for fine-tuning diffusion models that could significantly modify the current RLHF framework for fine-tuning diffusion models. This method bypasses resource-intensive reward model train", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing by utilizing direct human feedback, making the process more efficient and cost-effective.", + "bbox": [ + 511, + 472, + 888, + 501 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We expand the theoretical framework of DPO into a multi-step MDP, demonstrating that directly updating the policy based on human preferences within an MDP is equivalent to learning the optimal reward model first and then using it to guide policy updates. This establishes a robust theoretical foundation and provides assurance for our proposed method.", + "- In our experiments, we have demonstrated the effectiveness of our method by using human feedback to successfully address issues of hand and full-body deformities, enhance the safety of generated images, and improve prompt-image alignment." + ], + "bbox": [ + 500, + 503, + 890, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 704, + 640, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models. Denoising diffusion probabilistic models, introduced in [54] and further advanced by [26], have emerged as powerful tools for generating diverse data types. They have been successfully applied in various domains such as image synthesis [44, 49], video generation [27, 36, 53], and robotics control systems [1, 11, 29]. Notably, test-to-image diffusion models have enabled the creation of highly realistic images from textual descriptions [44, 49], opening new avenues in digital art and design.", + "bbox": [ + 496, + 731, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies have focused on refining the guidance of diffusion models for more precise manipulation over the", + "bbox": [ + 498, + 869, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generative process. Techniques such as adapters [63] and compositional approaches [17, 38] have been introduced to incorporate additional input constraints and blend multiple models, respectively, enhancing image quality and generation control. The implementation of classifier-based [13] and classifier-free guidance [25] has also contributed significantly to achieving greater autonomy in the generation process, resulting in outputs that closely align with user intentions. In our work, we utilize Stable Diffusion [45] to generate images based on specific prompts.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RLHF. RLHF stands as a salient strategy in the domain of machine learning when objectives are complex or difficult to define explicitly. This technique has been instrumental across various applications, from gaming, as demonstrated with Atari [4, 12], to more intricate tasks in robotics [10, 64]. The integration of RLHF into the development of large language models (LLMs) has marked a significant milestone in the field, with notable models like OpenAI's GPT-4 [41], Anthropic's Claude [2], Google's Bard [24], and Meta's Llama 2-Chat [57] leveraging this approach to enhance their performance and relevance. The effectiveness of RLHF in refining the behavior of LLMs to be more aligned with human values, such as helpfulness and harmlessness, has been extensively studied [4, 64]. The technique has also proven beneficial in more focused tasks, such as summarization, where models are trained to distill extensive information into concise representations [55]. Some recent research utilizes Reinforcement Learning from AI Feedback (RLAIF) [32, 62] as an alternative to RLHF for model fine-tuning. RLAIF offers convenience and efficiency by replacing human feedback with AI-generated feedback. However, for tasks like assessing hand generation normality or image aesthetic appeal, reliable judgment models are currently lacking. Hence, this paper still relies on human feedback for evaluation.", + "bbox": [ + 75, + 246, + 472, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tune Diffusion Models with RL. Before applying diffusion models, data generation has been regarded as a sequential decision-making problem and combined with reinforcement learning [3]. More recently, the SFT method [20] applied reinforcement learning to diffusion models to enhance existing fast DDPM samplers [26]. Reward Weighted method [34] explored using human feedback to align text-to-image models. It uses the reward model for the coefficients of the loss function instead of using reinforcement learning optimization objectives. ReFL [60] employs the framework of RLHF. It begins by training a model called ImageReward based on human preferences and then finetunes the diffusion model using reinforcement learning. DDPO [5] treats the denoising process of diffusion models as a MDP to fine-tune diffusion models with many reward models. DPOK [21] combine the KL divergence into the DDPO loss and use it to better align text-to-image objectives. All these models need a robust reward model, which", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "demands a substantial dataset of images and extensive human evaluations.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct Preference Optimization. In the realm of reinforcement learning, exploring policies derived from preferences rather than explicit rewards has gained attention through various methods. The Contextual Dueling Bandit (CDB) framework [18, 61] introduces the concept of a von Neumann winner, shifting away from the pursuit of an optimal policy directly based on rewards. Preference-based Reinforcement Learning (PbRL) [9, 39, 48] learns from binary preferences inferred from a cryptic scoring function instead of explicit rewards. Recently, the DPO approach [43] was proposed, which fine-tunes LLMs directly using preferences. DPO leverages the correlation between reward functions and optimal policies, effectively addressing the challenge of constrained reward maximization in a single phase of policy training.", + "bbox": [ + 496, + 121, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 359, + 638, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MDP. We consider the MDP formulation described in [56]. In this setting, an agent perceives a state $s \\in S$ and executes an action, where $S, A$ denote state space and action space, respectively. The transition probability function, denoted by $P(s'|s, a)$ , governs the progression from the current state $s$ to the subsequent state $s'$ upon the agent's action $a$ . Concurrently, the agent is awarded a scalar reward $r$ , determined by the reward function $r: S \\times A \\to \\mathbb{R}$ . The agent's objective is to ascertain a policy $\\pi(a|s)$ that maximizes the cumulative returns of trajectories $\\tau = (s_0, a_0, s_1, a_1, \\dots, s_{T-1}, a_{T-1})$ , which can be represented as: $\\mathcal{J}(\\pi) = \\mathbb{E}_{\\tau}[\\sum_{t=0}^{T-1} r(s_t, a_t)]$ .", + "bbox": [ + 496, + 383, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models learn to model a probability distribution $p(x)$ by inverting a Markovian forward process $q(\\pmb{x}_t|\\pmb{x}_{t-1})$ which adds noise to the data. The denoising process is modeled by a neural network to predict the mean of $\\pmb{x}_{t-1}$ or the noise $\\epsilon_{t-1}$ of the forward process. In our work, we use network $\\mu_\\theta(\\pmb{x}_t; t)$ to predict the mean of $\\pmb{x}_{t-1}$ instead of predicting the noise. Using the mean squared error (MSE) as a measure, the objective of this network can be written as:", + "bbox": [ + 496, + 551, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1, T ], \\boldsymbol {x} _ {0} \\sim p (\\boldsymbol {x} _ {0}), \\boldsymbol {x} _ {t} \\sim q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {0})} \\left[ \\| \\tilde {\\boldsymbol {\\mu}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) - \\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 681, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mu}_{\\theta}(\\boldsymbol{x}_t,\\boldsymbol{x}_0)$ is the forward process posterior mean. In the case of conditional generative modeling, the diffusion models learn to model $p(x|c)$ , where $c$ is the conditioning information, i.e., image category and image caption. This is done by adding an additional input $c$ to the denoising neural network, as in $\\mu_{\\theta}(\\boldsymbol{x}_t,t;\\boldsymbol{c})$ . To generate a sample from the learned distribution $p_{\\theta}(x|c)$ , we start by drawing a sample $\\boldsymbol{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and then progressively denoise the sample by iterated application of $\\epsilon_{\\theta}$ according to a specific sampler [26]. Given the noise-related parameter $\\sigma_t$ , the reverse process can be written as:", + "bbox": [ + 496, + 712, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reward learning for preferences. The basic framework to model preferences is to learn a reward function $r^{*}(s,a)$ from human feedback [12, 33, 59]. The segment $\\sigma = \\{s_k,a_k,s_{k + 1},a_{k + 1},\\dots ,s_m,a_m\\}$ is a sequence of observations and actions. By following the Bradley-Terry model [6], the human preference distribution $p^*$ by using the reward function can be expressed as:", + "bbox": [ + 76, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {*} (\\sigma_ {1} \\succ \\sigma_ {0}) = \\frac {\\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {1} , a _ {t} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {i} , a _ {t} ^ {i}\\right)\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 205, + 468, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{i} \\succ \\sigma_{j}$ denotes that segment $i$ is preferable to segment $j$ . Now we have the preference distribution of human feedback, and we want to use a network $r_{\\phi}$ to approximate $r^{*}$ . Given the human preference $y \\in \\{(1,0), (0,1)\\}$ which is recorded in dataset $\\mathcal{D}$ as a triple $(\\sigma_{0}, \\sigma_{1}, y)$ , framing the problem as a binary classification, the reward function modeled as a network is updated by minimizing:", + "bbox": [ + 76, + 253, + 472, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\phi) = - \\mathbb {E} _ {\\left(\\sigma_ {1}, \\sigma_ {0}, y\\right) \\sim \\mathcal {D}} [ y (0) \\log p _ {\\phi} \\left(\\sigma_ {0} \\succ \\sigma_ {1}\\right) \\\\ \\left. + y (1) \\log p _ {\\phi} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 368, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 412, + 168, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe a method to directly fine-tune diffusion models using human feedback, bypassing the conventional requirement for a reward model. Initially, we reinterpret the denoising process inherent in diffusion models as a multi-step MDP. Then we extend the theory of DPO to MDP, which allows us to apply the principles to effectively translate human preferences into policy improvements in diffusion models.", + "bbox": [ + 75, + 439, + 468, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Denoising process as a multi-step MDP", + "text_level": 1, + "bbox": [ + 76, + 568, + 413, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conceptualize the denoising process within the diffusion model as a multi-step MDP, which varies slightly from the approach outlined in [5]. To enhance clarity, we have redefined the states, transition probabilities, and policy functions. The correspondence between notations in the diffusion model and the MDP is established as follows:", + "bbox": [ + 75, + 590, + 468, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}) \\quad P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\boldsymbol {c}}, \\delta_ {t + 1}, \\delta_ {\\boldsymbol {x} _ {T - 1 - t}}\\right) \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - 1 - t} \\quad \\pi (\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}) \\triangleq p _ {\\theta} (\\mathbf {x} _ {T - 1 - t} \\mid \\mathbf {c}, t, \\mathbf {x} _ {T - t}) \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\boldsymbol {c}), \\delta_ {0}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ r \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq r \\left(\\left(\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}\\right), \\boldsymbol {x} _ {T - t - 1}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 681, + 452, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{x}$ represents the Dirac delta distribution, and $T$ denotes the maximize denoising timesteps. Leveraging this mapping, we can employ RL techniques to fine-tune diffusion models by maximizing returns. However, this approach requires a proficient reward model capable of adequately rewarding the noisy images. The task becomes exceptionally challenging, particularly when $t$ is low, and $\\pmb{x}_{T - t}$ closely resembles Gaussian noise, even for an experienced expert.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Direct Preference Optimization for MDP", + "text_level": 1, + "bbox": [ + 500, + 90, + 849, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The DPO method does not train a separate reward model but instead directly optimizes the LLMs with the preference data. Given a prompt $x$ and a pair of answers $(y_1, y_0) \\sim \\pi_{\\mathrm{ref}}(y|x)$ , where $\\pi_{\\mathrm{ref}}$ represents the reference (pre-trained) model, these responses are ranked and stored in $\\mathcal{D}$ as a tuple $(x, y_w, y_l)$ , where $y_w$ denotes the preferred answer and $y_l$ indicates the inferior one. DPO optimizes $\\pi_\\theta$ with the human preference dataset by using the following loss:", + "bbox": [ + 496, + 113, + 893, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\rho \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {\\text {r e f}} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {\\text {r e f}} (y _ {l} | x)}\\right) \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 258, + 893, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\rho$ is the logistic function, and $\\beta$ is the parameter controlling the deviation from the $\\pi_{\\theta}$ and $\\pi_{\\mathrm{ref}}$ . In our framework, we treat segments $\\sigma^1, \\sigma^0$ as $y_1, y_0$ and use DPO to fine-tune diffusion models. However, directly using this method faces difficulties since the segments contain a large number (usually 20-50) of the image latent, which occupy a large amount of GPU memory (each image is about 6G even when using LoRA [28]). Since we can only get human preferences for the final image $x_0$ , if we want to update $\\pi_{\\theta}(\\sigma) = \\prod_{t=k}^{T} \\pi_{\\theta}(s_t, a_t)$ , it will consume more than 100G GPU memory, which makes the fine-tuning process nearly impossible.", + "bbox": [ + 496, + 306, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this problem, we extend the DPO theory to MDP. Firstly, we need to reconsider the objective of the RL method. For the MDP problem, the agents take action by considering maximizing the expected return instead of the current reward. For actor-critic methods such as DDPG [37], the optimization objective of policy $\\pi$ gives:", + "bbox": [ + 496, + 489, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} [ Q ^ {*} (s, a) ]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 590, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $d^{\\pi} = (1 - \\gamma)\\sum_{t = 0}^{\\infty}\\gamma^{t}P_{t}^{\\pi}(s)$ represents the state visitation distribution, where $P_{t}^{\\pi}(s)$ denotes the probability of being in state $s$ at timestep $t$ given policy $\\pi$ . Additionally, $Q^{*}(s,a)$ denotes the optimal action-value function. The optimal policy can be written as:", + "bbox": [ + 496, + 623, + 893, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {*} (a | s) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} a = \\arg \\max _ {\\hat {a}} Q ^ {*} (s, \\hat {a}), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 710, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to some popular methods, we use KL-divergence to prevent the fine-tuned policy from deviating from the reference policy, hence relieving the out-of-distribution (OOD) issue. By incorporating this constraint, the RL objective can be rewritten as:", + "bbox": [ + 496, + 761, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} \\left[ Q ^ {*} (s, a) \\right] - \\beta \\mathbb {D} _ {K L} \\left[ \\pi (a | s) \\| \\pi_ {\\text {r e f}} (a | s) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 840, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\beta$ is the temperature parameter that controls the deviation of $\\pi_{\\theta}(a|s)$ and $\\pi_{\\mathrm{ref}}(a|s)$ .", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compressibility", + "text_level": 1, + "bbox": [ + 122, + 119, + 263, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Incompressibility", + "text_level": 1, + "bbox": [ + 120, + 186, + 267, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aesthetic Quality", + "text_level": 1, + "bbox": [ + 120, + 258, + 272, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e2f8ad454545b73bf2416007ff54ab2e1a1912a35c6b5527e66962f40da21a8d.jpg", + "image_caption": [ + "Figure 2. Progression of samples targeting compressibility, incompressibility, and aesthetic quality objectives. With the respective focus during training, images exhibit reduced detail and simpler backgrounds for compressibility, richer texture details for incompressibility, and an overall increased aesthetic appeal when prioritizing aesthetic quality." + ], + "image_footnote": [], + "bbox": [ + 289, + 90, + 851, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1 Given the objective of Eq. (8), the optimal policy $\\pi^{*}(a|s)$ has the following expression:", + "bbox": [ + 75, + 446, + 470, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {*} (a | s) = \\pi_ {\\text {r e f}} (a | s) \\exp \\left(\\frac {1}{\\beta} Q ^ {*} (s, a)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 484, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof can be seen in Appendix B.1. By rearranging the formula of Eq. (9), we can obtain that:", + "bbox": [ + 76, + 518, + 468, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {*} (s, a) = \\beta \\log \\frac {\\pi^ {*} (a | s)}{\\pi_ {\\mathrm {r e f}} (a | s)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 556, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, considering Eq. (3) and noticing that $Q^{*}(s_{t},a_{t}) = \\mathbb{E}\\left[\\sum_{t = k}^{T}r^{*}(s_{t},a_{t})\\right]$ under the policy $\\pi^* (a|s)$ , we make a substitution. By replacing $\\sum_{t = k}^{T}r^{*}(s_{t},a_{t})$ with $Q^{*}(s_{t},a_{t})$ we define a new distribution that can be rewritten as:", + "bbox": [ + 76, + 594, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) = \\frac {\\exp \\left(Q ^ {*} \\left(s _ {k} ^ {1} , a _ {k} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(Q ^ {*} \\left(s _ {k} ^ {i} , a _ {k} ^ {i}\\right)\\right)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 667, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We suppose $\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})$ is sampled from a normal distribution with mean $\\mathbb{E}[\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})]$ and standard deviation $\\sigma^2$ . From a statistical perspective, we can establish the relationship between the new distribution $\\tilde{p}^{*}(\\sigma_{1}\\succ \\sigma_{0})$ and the raw distribution $p^{*}(\\sigma_{1}\\succ \\sigma_{0})$ .", + "bbox": [ + 76, + 710, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2 For $i \\in \\{0,1\\}$ , suppose the expected return satisfies a normal distribution, i.e., $\\sum_{t=0}^{T} r^*(s_t^i, a_t^i) \\sim \\mathcal{N}(Q^*(s_0^i, a_0^i), \\sigma^2)$ . Given $Q^*(s,a) \\in [Q_{\\min}, Q_{\\max}]$ where $Q_{\\min}$ and $Q_{\\max}$ represent the lower and upper bounds of the values, then", + "bbox": [ + 76, + 792, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| p ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) - \\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right| < \\frac {\\left(\\xi^ {2} + 1\\right) \\left(\\exp \\left(\\sigma^ {2}\\right) - 1\\right)}{1 6 \\xi \\delta}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 875, + 449, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with probability at least $1 - \\delta$ , where $\\xi = \\frac{\\exp(Q_{\\mathrm{max}})}{\\exp(Q_{\\mathrm{min}})}$ .", + "bbox": [ + 503, + 444, + 854, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof can be seen in Appendix B.2. In practical applications, as the volume of data increases, it becomes easier to satisfy the assumption of normality. Additionally, we can use clipping operations to constrain the $Q$ values within a certain range, which introduces upper and lower bounds. Therefore, the aforementioned assumption is reasonable. As shown in proposition 2, their deviation can be bounded at the scale of $\\mathcal{O}\\left(\\frac{\\xi}{\\delta} (\\exp (\\sigma^2) - 1)\\right)$ . It is clear that this bound can approach 0 if the $\\sigma^2$ is close to 0. In practice, $\\sigma^2$ approaches 0 if the $Q$ function can be estimated with a small standard deviation.", + "bbox": [ + 496, + 470, + 890, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By combining Eq. (11), Eq. (4), and Eq. (10), replacing $p^*(\\sigma_1 \\succ \\sigma_0)$ with $\\tilde{p}^*(\\sigma_1 \\succ \\sigma_0)$ , and substituting $\\pi^*(s, a)$ with the policy network $\\pi_\\theta(s, a)$ that requires learning, we derive the following loss function for updating $\\pi_\\theta(a|s)$ :", + "bbox": [ + 496, + 647, + 890, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(s _ {k}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}) ], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 713, + 893, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_w = \\{s_k^w, a_k^w, s_{k+1}^w, a_{k+1}^w, \\dots, s_T^w, a_T^w\\}$ denotes the segment preferred over another segment $\\sigma_l = \\{s_k^l, a_k^l, s_{k+1}^l, a_{k+1}^l, \\dots, s_T^l, a_T^l\\}$ .", + "bbox": [ + 496, + 750, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Direct preference for Denoising Diffusion Policy Optimization", + "text_level": 1, + "bbox": [ + 500, + 801, + 890, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the denoising process as a multi-step MDP and using the mapping relationship depicted in Section 4.1, we can use DPO to directly update diffusion models by using Eq. (12). In the denoising process, we set $k = 0$ and", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ad20a06d7db5be20dee43c8ce0c6c281c4b0fca3f2e0414b22f03637ad7b813.jpg", + "image_caption": [ + "Figure 3. Comparison of D3PO against existing methods. The horizontal axis represents the number of image sample pairs generated for updating parameters. The rewards denote image size for incompressity objective, negative image size for the compressivity objective, and the LAION aesthetic score for the aesthetic objective. All experiments are conducted with 5 different seeds." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 356, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57f5e874c62260282f9da65854f1439ecb81eb556c5754d9caeff260f0ce439f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 88, + 612, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdcff313aeee8d751195d901691896162429675553edf98901dccba206f7d114.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 872, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$T = 20$ . We first sample an initial state $s_0^w = s_0^l = s_0$ and then use Eq. (2) to generate two segments. After manually choosing which segment is better, the probability of $\\pi_{\\theta}(a_0^w | s_0^w)$ is gradually increasing and $\\pi_{\\theta}(a_0^l | s_0^l)$ is decreasing, which guides the diffusion model to generate images of human preference. However, the approach of only updating $\\pi_{\\theta}(\\cdot | s_0)$ does not fully utilize the information within the segment.", + "bbox": [ + 75, + 306, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the middle states of the segment are noises and semi-finished images, it is hard for humans to judge which segment is better by observing the whole segment. But we can conveniently compare the final image $\\mathbf{x}_0$ . Like many RL methods [7, 51, 52] which give rewards by $\\forall s_t, a_t \\in \\sigma, r(s_t, a_t) = 1$ for winning the game and $\\forall t \\in \\sigma, r(s_t, a_t) = -1$ for losing the game, we also assume that if the segment is preferred, then any state-action pair of the segment is better than the other segment. By using this assumption, we construct $T$ sub-segments for the agent to learn, which can be written as:", + "bbox": [ + 75, + 429, + 470, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} = \\left\\{s _ {i}, a _ {i}, s _ {i + 1}, a _ {i + 1}, \\dots , s _ {T - 1}, a _ {T - 1} \\right\\}, \\quad 0 \\leq i \\leq T - 1 \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 604, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using these sub-segments, the overall loss of the D3PO algorithm gives:", + "bbox": [ + 76, + 641, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} (\\theta) = - E _ {\\left(s _ {i}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}) ], \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 676, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i \\in [0, T - 1]$ . Compared to Eq. (12), Eq. (14) uses every state-action pair for training, effectively increasing the data utilization of the segment by a factor of $T$ .", + "bbox": [ + 75, + 712, + 468, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overview of our method is shown in Fig. 1. The pseudocode of D3PO can be seen in Appendix A.", + "bbox": [ + 75, + 757, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 800, + 200, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we evaluate the effectiveness of D3PO in fine-tuning diffusion models. Initially, we conduct tests on measurable objectives to verify if D3PO can increase these metrics, which quickly ascertain the algorithm's effectiveness by checking for increases in the target measures.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we apply D3PO to experiments aimed at lowering the rate of deformities in hands and full-body images generated by diffusion models. Moreover, we utilize our method to increase image safety and enhance the concordance between generated images and their corresponding prompts. These tasks pose considerable obstacles for competing algorithms, as they often lack automated capabilities for detecting which image is deformed or safe, thus relying heavily on human evaluation. We use Stable Diffusion v1.5 [47] to generate images in most of the experiments.", + "bbox": [ + 496, + 308, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Pre-defined Quantifiable Objectives Results", + "text_level": 1, + "bbox": [ + 498, + 469, + 870, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initially conduct experiments with D3PO using quantitative objectives (alternatively referred to as optimal reward models). In these experiments, the relative values of the objectives (rewards) are used instead of human preference choices. Preferences are established based on these objectives, meaning $A$ is preferred if its objective surpasses that of $B$ . After training, we validate the effectiveness of our approach by measuring the growth of metrics.", + "bbox": [ + 496, + 492, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experimental setup, we benchmark against several popular methods: DDPO [5], DPOK [21], and Reward Weighted [34], all of which necessitate a reward model. Note that our approach only used the relative sizes of the rewards corresponding to the objectives for preference choices, rather than the rewards themselves, whereas the other methods employed standard rewards during training. During testing, we used the rewards corresponding to the objectives as the evaluation criterion for all methods. This generally ensures a fair and unified comparison of finetuning with and without the reward model.", + "bbox": [ + 496, + 614, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first use the size of the images to measure the preference relationship between two pictures. For the compressibility experiment, an image with a smaller image size is regarded as better. Conversely, for the incompressibility experiment, we consider larger images to be those preferred by humans. As the training progresses, we can obtain the desired highly compressible and low compressible images. We then utilize the LAION aesthetics predictor [50] to pre", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8946", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dict the aesthetic rating of images. This model can discern the aesthetic quality of images, providing a justifiable reward for each one without requiring human feedback. The model can generate more aesthetic images after fine-tuning. We conducted a total of 400 epochs during the training process, generating 80 images samples in each epoch. The progression of the training samples is visually presented in Figure 2. More quantitative samples are shown in Figure 8. The testing curves of D3PO and other methods are shown in Figure 3. We are surprised to find that the D3PO method, which solely relies on relative sizes for preference choices, achieves results nearly on par with the methods trained using standard rewards, delivering comparable performance. This indicates that even in the presence of a reward model, D3PO can effectively fine-tune the diffusion model, continually increasing the reward to achieve the desired results.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experiments without Any Reward Model", + "text_level": 1, + "bbox": [ + 76, + 345, + 429, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments for some objectives without any reward model. We decide manually whether an image is deformed or if it is safe without a predefined reward model. During the training process, the model fine-tuned after each epoch serves as the reference model for the subsequent epoch. For each prompt, we generate 7 images with the same initial state $x_{T} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ .", + "bbox": [ + 75, + 369, + 470, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1 Reduce Image Distortion", + "text_level": 1, + "bbox": [ + 76, + 500, + 305, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the prompt \"1 hand\" to generate images, manually selecting those that are deformed. Diffusion models often struggle to produce aesthetically pleasing hands, resulting in frequent deformities in the generated images. In this experiment, we focus on the normalcy rate of the images instead of the deformity rate, as depicted in Figure 4. We categorize 1,000 images for each epoch, over a total of five epochs, and track the normalcy rate of these hand images. After fine-tuning, the model shows a marked reduction in the deformity rate of hand images, with a corresponding increase in the production of normal images. Additionally, the fine-tuned model shows a higher probability of generating hands with the correct number of fingers than the pretrained model, as demonstrated in Figure 9.", + "bbox": [ + 75, + 525, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the generality of our method, we generated images with the Anything v5 model ${}^{1}$ , renowned for creating anime character images. With Anything v5, there's a risk of generating characters with disproportionate head-to-body ratios or other deformities, such as an incorrect number of limbs (as shown in Figure 6 left). We categorize such outputs as deformed. We assume that non-selected images are more favorable than the deformed ones, though we do not rank preferences within the deformed or non-deformed", + "bbox": [ + 75, + 737, + 470, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sets. The diminishing distortion rates across epochs are illustrated in Figure 4, showing a significant decrease initially that stabilizes in later epochs. The visual examples are provided in Figure 6.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15638d63e4b751f1e53b0d42ab5d1f5d75d5436dce644bbb98cb668c2598ee5d.jpg", + "image_caption": [ + "Figure 4. Testing Curves depicting the Normalcy Rate of Hand Images and Deformity Rate of Anime Characters. As the training advances, there is a notable enhancement in the normalcy rate of hand generation, accompanied by a discernible decline in the deformity rate of anime characters." + ], + "image_footnote": [], + "bbox": [ + 519, + 162, + 874, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2 Enhance Image Safety", + "text_level": 1, + "bbox": [ + 500, + 425, + 709, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we utilized unsafe prompts to generate images using a diffusion model. These prompts contained edgy terms that could lead to the creation of both normal and Not Safe for Work (NSFW) images, examples being 'ambiguous beauty' and 'provocative art'. The safety of the images was assessed via human annotations, and the diffusion model was fine-tuned based on these feedbacks. Across 10 epochs, we generated 1,000 images per epoch. Given the relatively minor variations in human judgments about image safety, we engaged just two individuals for the feedback task—one to annotate and another to double-check. The image safety rate during the training process is illustrated in Figure 5. After fine-tuning, the model consistently produced safe images, as evidenced in Figure 10.", + "bbox": [ + 496, + 449, + 893, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ce5b477a552d5771aa5ead7de7b068fb32209b1d77b2032c0f48fe197f60f4c.jpg", + "image_caption": [ + "Figure 5. Safety rate curves of the training procession." + ], + "image_footnote": [], + "bbox": [ + 547, + 672, + 844, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3 Prompt-Image Alignment", + "text_level": 1, + "bbox": [ + 500, + 845, + 735, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ human feedback to evaluate the alignment preference between two images generated from each prompt.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/stablediffusionapi/anything-v5", + "bbox": [ + 94, + 886, + 383, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c4009da913d0a9a7446270fa01b84a9fb70316a54549d285dcd10050220ac30.jpg", + "image_caption": [ + "Figure 6. Image samples of the pre-trained model and the fine-tuned model. The images on the left side of the arrows are distorted images (such as having 3 legs in the image) generated by the pre-trained model, while the images on the right side are normal images generated after fine-tuning the model. Both sets of images used the same initial Gaussian noise, prompts, and seeds." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 415, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b95f3a01561f2c5011769c87219c3eb4c077a1be33496906b4f9d8281381885d.jpg", + "image_caption": [ + "finetuning" + ], + "image_footnote": [], + "bbox": [ + 555, + 88, + 852, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0a8bc25c6dbdcb81331fe34180c03e75e1e399413d5f3703b07355dbf23631d.jpg", + "image_caption": [ + "Figure 7. Comparative evaluation of 300 text prompts: Our study involved generating images from the pre-trained diffusion model, the fine-tuned models using other methods, and our fine-tuned model—using the same text prompts. For each prompt, human evaluators are tasked with determining which image is better aligned with the text. Each image was assessed by 5 human raters, and we report the percentage of images that received favorable evaluations. We also highlight the percentage with more than half vote in the black box." + ], + "image_footnote": [], + "bbox": [ + 91, + 390, + 455, + 545 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e1ec9b08c355219ec6880d2aae27646e3a77cd88ac4b6fceeb60532b88d7699.jpg", + "table_caption": [ + "Table 1. Evaluations of prompt-image alignment. We compare D3PO with those that (a) do not fine-tune, (b) use the preferred images to fine-tune, and (c) use Reward Weighted method to fine-tune. D3PO achieve the best performance. D3PO demonstrated superior performance across all comparisons." + ], + "table_footnote": [], + "table_body": "
MethodsCLIP score (↑)BLIP score (↑)ImageReward score (↑)Human preference (↑)
No fine-tune30.71.950.0414.7%
Preferred images31.01.970.0811%
Reward Weighted31.52.010.1718.7%
D3PO31.92.060.2755.7%
", + "bbox": [ + 78, + 777, + 467, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each epoch, we use 4,000 prompts, generate two images per prompt, and assess preferences with feedback from 16 different evaluators. The training spans 10 epochs.", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The preference comparisons between images from the pretrained and fine-tuned models are conducted by an additional 5 evaluators, with the comparative preferences depicted in Figure 7. We also execute quantitative evaluations of models using metrics that measure the congruence between prompts and images, including CLIP [42], BLIP [35], and ImageReward [60], as presented in Table 1.", + "bbox": [ + 496, + 392, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 513, + 619, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a direct preference denoising diffusion policy optimization method, named D3PO, to fine-tune diffusion models purely from human feedback without learning a separate reward model. D3PO views the denoising process as a multi-step MDP, making it possible to utilize the DPO-style optimization formula by formulating the action-value function $Q$ with the reference model and the fine-tuned model. D3PO updates parameters at each step of denoising and consumes much fewer GPU memory overheads than directly applying the DPO algorithm. The empirical experiments illustrate that our method achieves competitive or even better performance compared with a diffusion model fine-tuned with a reward model that is trained with a large amount of images and human preferences in terms of image compressibility, image compressibility and aesthetic quality. We further show that D3PO can also benefit challenging tasks such as reducing image distortion rates, enhancing the safety of the generated images, and aligning prompts and images.", + "bbox": [ + 496, + 540, + 893, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 844, + 668, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the STI 2030-Major Projects under Grant 2021ZD0201404.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 2", + "[2] Anthropic. Introducing claude, 2023. 3", + "[3] Philip Bachman and Doina Precup. Data generation as sequential decision making. Advances in Neural Information Processing Systems, 28, 2015. 3", + "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3", + "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 3, 4, 6", + "[6] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4", + "[7] Noam Brown and Tuomas Sandholm. Superhuman ai for multiplayer poker. Science, 365(6456):885-890, 2019. 6", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[9] Robert Busa-Fekete, Balázs Szörényi, Paul Weng, Weiwei Cheng, and Eyke Hüllermeier. Preference-based reinforcement learning: evolutionary direct policy search using a preference-based racing algorithm. Machine learning, 97: 327-351, 2014. 3", + "[10] Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023. 3", + "[11] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2", + "[12] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 1, 3, 4", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3", + "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890-16902, 2022. 1", + "[16] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1", + "[17] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 3", + "[18] Miroslav Dudík, Katja Hofmann, Robert E Schapire, Aleksandris Slivkins, and Masrour Zoghi. Contextual dueling bandits. In Conference on Learning Theory, pages 563-587. PMLR, 2015. 3", + "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1", + "[20] Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362, 2023. 3", + "[21] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381, 2023. 3, 6", + "[22] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 1", + "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[24] Google. Bard, 2023. 3", + "[25] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[28] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[29] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1", + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16", + "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3", + "[33] Kimin Lee, Laura Smith, and Pieter Abbeel. Pebble: Feedback-efficient interactive reinforcement learning via rel-labeling experience and unsupervised pre-training. arXiv preprint arXiv:2106.05091, 2021. 4", + "[34] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3, 6", + "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8", + "[36] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 2", + "[37] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 4", + "[38] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3", + "[39] Runze Liu, Yali Du, Fengshuo Bai, Jiafei Lyu, and Xiu Li. Zero-shot preference learning for offline rl via optimal transport. arXiv preprint arXiv:2306.03615, 2023. 3", + "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[41] OpenAI. Gpt-4 technical report, 2023. 1, 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8", + "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "preprint arXiv:2305.18290, 2023. 1, 3", + "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2", + "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3", + "[46] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 6", + "[48] Aadirupa Saha, Aldo Pacchiano, and Jonathan Lee. Dueling rl: Reinforcement learning with trajectory preferences. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 6263-6289. PMLR, 2023. 3", + "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[50] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6", + "[51] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 6", + "[52] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676): 354-359, 2017. 6", + "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8950", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 3", + "[56] Richard S Sutton, Andrew G Barto, et al. Introduction to reinforcement learning. 1998. 3", + "[57] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[58] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1", + "[59] Aaron Wilson, Alan Fern, and Prasad Tadepalli. A bayesian approach for policy learning from trajectory preference queries. Advances in neural information processing systems, 25, 2012. 4", + "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.1,3,8,14", + "[61] Yisong Yue, Josef Broder, Robert Kleinberg, and Thorsten Joachims. The k-armed dueling bandits problem. Journal of Computer and System Sciences, 78(5):1538-1556, 2012. JCSS Special Issue: Cloud Computing 2011. 3", + "[62] Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. Huatuogpt, towards taming language model to be a doctor. arXiv preprint arXiv:2305.15075, 2023. 3", + "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3", + "[64] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 685 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8951", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_model.json b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..90fbc3c2f4d7d9b45c0d5ddfa93841cf16abbf0e --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.211, + 0.131, + 0.761, + 0.175 + ], + "angle": 0, + "content": "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.203, + 0.704, + 0.221 + ], + "angle": 0, + "content": "Kai Yang\\(^{1*}\\) Jian Tao\\(^{1*}\\) Jiafei Lyu\\(^{1\\dagger}\\) Chunjiang Ge\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.222, + 0.72, + 0.239 + ], + "angle": 0, + "content": "Jiaxin Chen3 Weihan Shen3 Xiaolong Zhu3 Xiu Li1†" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.24, + 0.774, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua Shenzhen International Graduate School, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.257, + 0.851, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Automation, Tsinghua University \\(^{3}\\) Parametrix Technology Company Ltd." + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.277, + 0.78, + 0.292 + ], + "angle": 0, + "content": "{yk22,tj22,lvjf20}@mails.tsinghua.edu.cn li.xiu@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.313, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.473, + 0.797 + ], + "angle": 0, + "content": "Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Recent advances in image generation models have yielded unprecedented success in producing high-quality images" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.892, + 0.48 + ], + "angle": 0, + "content": "from textual prompts [30, 44, 49]. Diverse approaches, including Generative Adversarial Networks (GANs) [23], autoregressive models [14, 15, 19, 22, 44, 58], Normalizing Flows [16, 46], and diffusion-based techniques [40, 45, 47, 49], have rapidly pushed forward the capabilities of these systems. With the proper textual inputs, such models are now adept at crafting images that are not only visually compelling but also semantically coherent, garnering widespread interest for their potential applications and implications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.849 + ], + "angle": 0, + "content": "To adapt the aforementioned models for specific downstream tasks, such as the generation of more visually appealing and aesthetic images, Reinforcement Learning from Human Feedback (RLHF) is commonly employed [12]. This technique has been successfully used to refine large language models such as GPT [8, 41]. Now, efforts are being made to extend this method to diffusion models to enhance their performance. One such approach, the DDPO method [5], aims to enhance image complexity, aesthetic quality, and the alignment between prompts and images. The ReLF approach [60] introduces a novel reward model, named ImageReward, which is specifically trained to discern human aesthetic preferences in text-to-image synthesis. This model is then utilized to fine-tune diffusion models to produce images that align more closely with human preferences. Nonetheless, developing a robust reward model for various tasks can be challenging, often necessitating a vast collection of images and abundant training resources. For example, to diminish the rate of deformities in character images, one must amass a substantial dataset of deformed and non-deformed images generated from identical prompts. Subsequently, a network is constructed to discern and learn the human preference for non-deformed imagery, serving as the reward model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In the field of natural language processing, Direct Preference Optimization (DPO) has been proposed to reduce training costs [43]. This method forgoes the training of a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.093, + 0.849, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.39, + 0.893, + 0.447 + ], + "angle": 0, + "content": "Figure 1. Overview of D3PO. The diffusion model generates two corresponding images based on the provided prompts. Guided by specific task requirements—such as improving prompt-image alignment, enhancing image incompressibility, or refining aesthetic quality—human evaluators select the preferred image. Leveraging this human feedback, our method directly updates the diffusion model's parameters without necessitating the training of a reward model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.639 + ], + "angle": 0, + "content": "reward model and directly fine-tunes language models according to human preferences. However, this straightforward and easy-to-train method encounters challenges when applied to fine-tune diffusion models. During the DPO training process, the complete sentence generated by the language model is treated as a single output, necessitating the storage of gradients from multiple forward passes. With diffusion models, one must store the gradients across multiple latent image representations, which are significantly larger than word embeddings, leading to memory consumption that is typically unsustainable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.837 + ], + "angle": 0, + "content": "To address the issue of high computational overhead and enable the use of the DPO method to fine-tune diffusion models directly with human feedback, we conceptualize the denoising process as a multi-step MDP, which utilizes a pre-trained model to represent an action value function \\( Q \\) that is commonly estimated in RL. We extend the theoretical framework of DPO into the formulated MDP, which allows for direct parameter updates at each step of the denoising process based on human feedback, thereby circumventing the significant computational costs and eliminating the need for a reward model. To the best of our knowledge, this is the first work that fine-tune diffusion models without reward models. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We introduce an innovative approach for fine-tuning diffusion models that could significantly modify the current RLHF framework for fine-tuning diffusion models. This method bypasses resource-intensive reward model train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.89, + 0.502 + ], + "angle": 0, + "content": "ing by utilizing direct human feedback, making the process more efficient and cost-effective." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.609 + ], + "angle": 0, + "content": "- We expand the theoretical framework of DPO into a multi-step MDP, demonstrating that directly updating the policy based on human preferences within an MDP is equivalent to learning the optimal reward model first and then using it to guide policy updates. This establishes a robust theoretical foundation and provides assurance for our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.61, + 0.892, + 0.685 + ], + "angle": 0, + "content": "- In our experiments, we have demonstrated the effectiveness of our method by using human feedback to successfully address issues of hand and full-body deformities, enhance the safety of generated images, and improve prompt-image alignment." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.642, + 0.721 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Diffusion models. Denoising diffusion probabilistic models, introduced in [54] and further advanced by [26], have emerged as powerful tools for generating diverse data types. They have been successfully applied in various domains such as image synthesis [44, 49], video generation [27, 36, 53], and robotics control systems [1, 11, 29]. Notably, test-to-image diffusion models have enabled the creation of highly realistic images from textual descriptions [44, 49], opening new avenues in digital art and design." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Recent studies have focused on refining the guidance of diffusion models for more precise manipulation over the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "generative process. Techniques such as adapters [63] and compositional approaches [17, 38] have been introduced to incorporate additional input constraints and blend multiple models, respectively, enhancing image quality and generation control. The implementation of classifier-based [13] and classifier-free guidance [25] has also contributed significantly to achieving greater autonomy in the generation process, resulting in outputs that closely align with user intentions. In our work, we utilize Stable Diffusion [45] to generate images based on specific prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.626 + ], + "angle": 0, + "content": "RLHF. RLHF stands as a salient strategy in the domain of machine learning when objectives are complex or difficult to define explicitly. This technique has been instrumental across various applications, from gaming, as demonstrated with Atari [4, 12], to more intricate tasks in robotics [10, 64]. The integration of RLHF into the development of large language models (LLMs) has marked a significant milestone in the field, with notable models like OpenAI's GPT-4 [41], Anthropic's Claude [2], Google's Bard [24], and Meta's Llama 2-Chat [57] leveraging this approach to enhance their performance and relevance. The effectiveness of RLHF in refining the behavior of LLMs to be more aligned with human values, such as helpfulness and harmlessness, has been extensively studied [4, 64]. The technique has also proven beneficial in more focused tasks, such as summarization, where models are trained to distill extensive information into concise representations [55]. Some recent research utilizes Reinforcement Learning from AI Feedback (RLAIF) [32, 62] as an alternative to RLHF for model fine-tuning. RLAIF offers convenience and efficiency by replacing human feedback with AI-generated feedback. However, for tasks like assessing hand generation normality or image aesthetic appeal, reliable judgment models are currently lacking. Hence, this paper still relies on human feedback for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Fine-tune Diffusion Models with RL. Before applying diffusion models, data generation has been regarded as a sequential decision-making problem and combined with reinforcement learning [3]. More recently, the SFT method [20] applied reinforcement learning to diffusion models to enhance existing fast DDPM samplers [26]. Reward Weighted method [34] explored using human feedback to align text-to-image models. It uses the reward model for the coefficients of the loss function instead of using reinforcement learning optimization objectives. ReFL [60] employs the framework of RLHF. It begins by training a model called ImageReward based on human preferences and then finetunes the diffusion model using reinforcement learning. DDPO [5] treats the denoising process of diffusion models as a MDP to fine-tune diffusion models with many reward models. DPOK [21] combine the KL divergence into the DDPO loss and use it to better align text-to-image objectives. All these models need a robust reward model, which" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "demands a substantial dataset of images and extensive human evaluations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.35 + ], + "angle": 0, + "content": "Direct Preference Optimization. In the realm of reinforcement learning, exploring policies derived from preferences rather than explicit rewards has gained attention through various methods. The Contextual Dueling Bandit (CDB) framework [18, 61] introduces the concept of a von Neumann winner, shifting away from the pursuit of an optimal policy directly based on rewards. Preference-based Reinforcement Learning (PbRL) [9, 39, 48] learns from binary preferences inferred from a cryptic scoring function instead of explicit rewards. Recently, the DPO approach [43] was proposed, which fine-tunes LLMs directly using preferences. DPO leverages the correlation between reward functions and optimal policies, effectively addressing the challenge of constrained reward maximization in a single phase of policy training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.64, + 0.377 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.553 + ], + "angle": 0, + "content": "MDP. We consider the MDP formulation described in [56]. In this setting, an agent perceives a state \\( s \\in S \\) and executes an action, where \\( S, A \\) denote state space and action space, respectively. The transition probability function, denoted by \\( P(s'|s, a) \\), governs the progression from the current state \\( s \\) to the subsequent state \\( s' \\) upon the agent's action \\( a \\). Concurrently, the agent is awarded a scalar reward \\( r \\), determined by the reward function \\( r: S \\times A \\to \\mathbb{R} \\). The agent's objective is to ascertain a policy \\( \\pi(a|s) \\) that maximizes the cumulative returns of trajectories \\( \\tau = (s_0, a_0, s_1, a_1, \\dots, s_{T-1}, a_{T-1}) \\), which can be represented as: \\( \\mathcal{J}(\\pi) = \\mathbb{E}_{\\tau}[\\sum_{t=0}^{T-1} r(s_t, a_t)] \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models learn to model a probability distribution \\( p(x) \\) by inverting a Markovian forward process \\( q(\\pmb{x}_t|\\pmb{x}_{t-1}) \\) which adds noise to the data. The denoising process is modeled by a neural network to predict the mean of \\( \\pmb{x}_{t-1} \\) or the noise \\( \\epsilon_{t-1} \\) of the forward process. In our work, we use network \\( \\mu_\\theta(\\pmb{x}_t; t) \\) to predict the mean of \\( \\pmb{x}_{t-1} \\) instead of predicting the noise. Using the mean squared error (MSE) as a measure, the objective of this network can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.683, + 0.891, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1, T ], \\boldsymbol {x} _ {0} \\sim p (\\boldsymbol {x} _ {0}), \\boldsymbol {x} _ {t} \\sim q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {0})} \\left[ \\| \\tilde {\\boldsymbol {\\mu}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) - \\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.895, + 0.879 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mu}_{\\theta}(\\boldsymbol{x}_t,\\boldsymbol{x}_0)\\) is the forward process posterior mean. In the case of conditional generative modeling, the diffusion models learn to model \\(p(x|c)\\), where \\(c\\) is the conditioning information, i.e., image category and image caption. This is done by adding an additional input \\(c\\) to the denoising neural network, as in \\(\\mu_{\\theta}(\\boldsymbol{x}_t,t;\\boldsymbol{c})\\). To generate a sample from the learned distribution \\(p_{\\theta}(x|c)\\), we start by drawing a sample \\(\\boldsymbol{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and then progressively denoise the sample by iterated application of \\(\\epsilon_{\\theta}\\) according to a specific sampler [26]. Given the noise-related parameter \\(\\sigma_t\\), the reverse process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Reward learning for preferences. The basic framework to model preferences is to learn a reward function \\( r^{*}(s,a) \\) from human feedback [12, 33, 59]. The segment \\( \\sigma = \\{s_k,a_k,s_{k + 1},a_{k + 1},\\dots ,s_m,a_m\\} \\) is a sequence of observations and actions. By following the Bradley-Terry model [6], the human preference distribution \\( p^* \\) by using the reward function can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.206, + 0.47, + 0.248 + ], + "angle": 0, + "content": "\\[\np ^ {*} (\\sigma_ {1} \\succ \\sigma_ {0}) = \\frac {\\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {1} , a _ {t} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {i} , a _ {t} ^ {i}\\right)\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.255, + 0.473, + 0.362 + ], + "angle": 0, + "content": "where \\(\\sigma_{i} \\succ \\sigma_{j}\\) denotes that segment \\(i\\) is preferable to segment \\(j\\). Now we have the preference distribution of human feedback, and we want to use a network \\(r_{\\phi}\\) to approximate \\(r^{*}\\). Given the human preference \\(y \\in \\{(1,0), (0,1)\\}\\) which is recorded in dataset \\(\\mathcal{D}\\) as a triple \\((\\sigma_{0}, \\sigma_{1}, y)\\), framing the problem as a binary classification, the reward function modeled as a network is updated by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.369, + 0.47, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\phi) = - \\mathbb {E} _ {\\left(\\sigma_ {1}, \\sigma_ {0}, y\\right) \\sim \\mathcal {D}} [ y (0) \\log p _ {\\phi} \\left(\\sigma_ {0} \\succ \\sigma_ {1}\\right) \\\\ \\left. + y (1) \\log p _ {\\phi} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.169, + 0.43 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.56 + ], + "angle": 0, + "content": "In this section, we describe a method to directly fine-tune diffusion models using human feedback, bypassing the conventional requirement for a reward model. Initially, we reinterpret the denoising process inherent in diffusion models as a multi-step MDP. Then we extend the theory of DPO to MDP, which allows us to apply the principles to effectively translate human preferences into policy improvements in diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.414, + 0.586 + ], + "angle": 0, + "content": "4.1. Denoising process as a multi-step MDP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.676 + ], + "angle": 0, + "content": "We conceptualize the denoising process within the diffusion model as a multi-step MDP, which varies slightly from the approach outlined in [5]. To enhance clarity, we have redefined the states, transition probabilities, and policy functions. The correspondence between notations in the diffusion model and the MDP is established as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.682, + 0.454, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}) \\quad P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\boldsymbol {c}}, \\delta_ {t + 1}, \\delta_ {\\boldsymbol {x} _ {T - 1 - t}}\\right) \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - 1 - t} \\quad \\pi (\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}) \\triangleq p _ {\\theta} (\\mathbf {x} _ {T - 1 - t} \\mid \\mathbf {c}, t, \\mathbf {x} _ {T - t}) \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\boldsymbol {c}), \\delta_ {0}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ r \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq r \\left(\\left(\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}\\right), \\boldsymbol {x} _ {T - t - 1}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\delta_{x}\\) represents the Dirac delta distribution, and \\(T\\) denotes the maximize denoising timesteps. Leveraging this mapping, we can employ RL techniques to fine-tune diffusion models by maximizing returns. However, this approach requires a proficient reward model capable of adequately rewarding the noisy images. The task becomes exceptionally challenging, particularly when \\(t\\) is low, and \\(\\pmb{x}_{T - t}\\) closely resembles Gaussian noise, even for an experienced expert." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.85, + 0.107 + ], + "angle": 0, + "content": "4.2. Direct Preference Optimization for MDP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.894, + 0.237 + ], + "angle": 0, + "content": "The DPO method does not train a separate reward model but instead directly optimizes the LLMs with the preference data. Given a prompt \\( x \\) and a pair of answers \\( (y_1, y_0) \\sim \\pi_{\\mathrm{ref}}(y|x) \\), where \\( \\pi_{\\mathrm{ref}} \\) represents the reference (pre-trained) model, these responses are ranked and stored in \\( \\mathcal{D} \\) as a tuple \\( (x, y_w, y_l) \\), where \\( y_w \\) denotes the preferred answer and \\( y_l \\) indicates the inferior one. DPO optimizes \\( \\pi_\\theta \\) with the human preference dataset by using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.26, + 0.894, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\rho \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {\\text {r e f}} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {\\text {r e f}} (y _ {l} | x)}\\right) \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Here \\(\\rho\\) is the logistic function, and \\(\\beta\\) is the parameter controlling the deviation from the \\(\\pi_{\\theta}\\) and \\(\\pi_{\\mathrm{ref}}\\). In our framework, we treat segments \\(\\sigma^1, \\sigma^0\\) as \\(y_1, y_0\\) and use DPO to fine-tune diffusion models. However, directly using this method faces difficulties since the segments contain a large number (usually 20-50) of the image latent, which occupy a large amount of GPU memory (each image is about 6G even when using LoRA [28]). Since we can only get human preferences for the final image \\(x_0\\), if we want to update \\(\\pi_{\\theta}(\\sigma) = \\prod_{t=k}^{T} \\pi_{\\theta}(s_t, a_t)\\), it will consume more than 100G GPU memory, which makes the fine-tuning process nearly impossible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.581 + ], + "angle": 0, + "content": "To address this problem, we extend the DPO theory to MDP. Firstly, we need to reconsider the objective of the RL method. For the MDP problem, the agents take action by considering maximizing the expected return instead of the current reward. For actor-critic methods such as DDPG [37], the optimization objective of policy \\(\\pi\\) gives:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.592, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} [ Q ^ {*} (s, a) ]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.894, + 0.702 + ], + "angle": 0, + "content": "Here, \\( d^{\\pi} = (1 - \\gamma)\\sum_{t = 0}^{\\infty}\\gamma^{t}P_{t}^{\\pi}(s) \\) represents the state visitation distribution, where \\( P_{t}^{\\pi}(s) \\) denotes the probability of being in state \\( s \\) at timestep \\( t \\) given policy \\( \\pi \\). Additionally, \\( Q^{*}(s,a) \\) denotes the optimal action-value function. The optimal policy can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.712, + 0.892, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\pi^ {*} (a | s) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} a = \\arg \\max _ {\\hat {a}} Q ^ {*} (s, \\hat {a}), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Similar to some popular methods, we use KL-divergence to prevent the fine-tuned policy from deviating from the reference policy, hence relieving the out-of-distribution (OOD) issue. By incorporating this constraint, the RL objective can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.842, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} \\left[ Q ^ {*} (s, a) \\right] - \\beta \\mathbb {D} _ {K L} \\left[ \\pi (a | s) \\| \\pi_ {\\text {r e f}} (a | s) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\beta\\) is the temperature parameter that controls the deviation of \\(\\pi_{\\theta}(a|s)\\) and \\(\\pi_{\\mathrm{ref}}(a|s)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8944" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.123, + 0.12, + 0.264, + 0.14 + ], + "angle": 0, + "content": "Compressibility" + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.187, + 0.268, + 0.207 + ], + "angle": 0, + "content": "Incompressibility" + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.26, + 0.274, + 0.28 + ], + "angle": 0, + "content": "Aesthetic Quality" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.091, + 0.852, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.895, + 0.422 + ], + "angle": 0, + "content": "Figure 2. Progression of samples targeting compressibility, incompressibility, and aesthetic quality objectives. With the respective focus during training, images exhibit reduced detail and simpler backgrounds for compressibility, richer texture details for incompressibility, and an overall increased aesthetic appeal when prioritizing aesthetic quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.479 + ], + "angle": 0, + "content": "Proposition 1 Given the objective of Eq. (8), the optimal policy \\(\\pi^{*}(a|s)\\) has the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.485, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\pi^ {*} (a | s) = \\pi_ {\\text {r e f}} (a | s) \\exp \\left(\\frac {1}{\\beta} Q ^ {*} (s, a)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.551 + ], + "angle": 0, + "content": "The proof can be seen in Appendix B.1. By rearranging the formula of Eq. (9), we can obtain that:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.557, + 0.469, + 0.589 + ], + "angle": 0, + "content": "\\[\nQ ^ {*} (s, a) = \\beta \\log \\frac {\\pi^ {*} (a | s)}{\\pi_ {\\mathrm {r e f}} (a | s)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Now, considering Eq. (3) and noticing that \\( Q^{*}(s_{t},a_{t}) = \\mathbb{E}\\left[\\sum_{t = k}^{T}r^{*}(s_{t},a_{t})\\right] \\) under the policy \\( \\pi^* (a|s) \\), we make a substitution. By replacing \\( \\sum_{t = k}^{T}r^{*}(s_{t},a_{t}) \\) with \\( Q^{*}(s_{t},a_{t}) \\) we define a new distribution that can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.669, + 0.47, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) = \\frac {\\exp \\left(Q ^ {*} \\left(s _ {k} ^ {1} , a _ {k} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(Q ^ {*} \\left(s _ {k} ^ {i} , a _ {k} ^ {i}\\right)\\right)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.789 + ], + "angle": 0, + "content": "We suppose \\(\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})\\) is sampled from a normal distribution with mean \\(\\mathbb{E}[\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})]\\) and standard deviation \\(\\sigma^2\\). From a statistical perspective, we can establish the relationship between the new distribution \\(\\tilde{p}^{*}(\\sigma_{1}\\succ \\sigma_{0})\\) and the raw distribution \\(p^{*}(\\sigma_{1}\\succ \\sigma_{0})\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.794, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Proposition 2 For \\( i \\in \\{0,1\\} \\), suppose the expected return satisfies a normal distribution, i.e., \\( \\sum_{t=0}^{T} r^*(s_t^i, a_t^i) \\sim \\mathcal{N}(Q^*(s_0^i, a_0^i), \\sigma^2) \\). Given \\( Q^*(s,a) \\in [Q_{\\min}, Q_{\\max}] \\) where \\( Q_{\\min} \\) and \\( Q_{\\max} \\) represent the lower and upper bounds of the values, then" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.875, + 0.45, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left| p ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) - \\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right| < \\frac {\\left(\\xi^ {2} + 1\\right) \\left(\\exp \\left(\\sigma^ {2}\\right) - 1\\right)}{1 6 \\xi \\delta}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.445, + 0.855, + 0.468 + ], + "angle": 0, + "content": "with probability at least \\(1 - \\delta\\), where \\(\\xi = \\frac{\\exp(Q_{\\mathrm{max}})}{\\exp(Q_{\\mathrm{min}})}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.647 + ], + "angle": 0, + "content": "The proof can be seen in Appendix B.2. In practical applications, as the volume of data increases, it becomes easier to satisfy the assumption of normality. Additionally, we can use clipping operations to constrain the \\(Q\\) values within a certain range, which introduces upper and lower bounds. Therefore, the aforementioned assumption is reasonable. As shown in proposition 2, their deviation can be bounded at the scale of \\(\\mathcal{O}\\left(\\frac{\\xi}{\\delta} (\\exp (\\sigma^2) - 1)\\right)\\). It is clear that this bound can approach 0 if the \\(\\sigma^2\\) is close to 0. In practice, \\(\\sigma^2\\) approaches 0 if the \\(Q\\) function can be estimated with a small standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.71 + ], + "angle": 0, + "content": "By combining Eq. (11), Eq. (4), and Eq. (10), replacing \\( p^*(\\sigma_1 \\succ \\sigma_0) \\) with \\( \\tilde{p}^*(\\sigma_1 \\succ \\sigma_0) \\), and substituting \\( \\pi^*(s, a) \\) with the policy network \\( \\pi_\\theta(s, a) \\) that requires learning, we derive the following loss function for updating \\( \\pi_\\theta(a|s) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.714, + 0.895, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(s _ {k}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}) ], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.798 + ], + "angle": 0, + "content": "where \\(\\sigma_w = \\{s_k^w, a_k^w, s_{k+1}^w, a_{k+1}^w, \\dots, s_T^w, a_T^w\\}\\) denotes the segment preferred over another segment \\(\\sigma_l = \\{s_k^l, a_k^l, s_{k+1}^l, a_{k+1}^l, \\dots, s_T^l, a_T^l\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.892, + 0.834 + ], + "angle": 0, + "content": "4.3. Direct preference for Denoising Diffusion Policy Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Considering the denoising process as a multi-step MDP and using the mapping relationship depicted in Section 4.1, we can use DPO to directly update diffusion models by using Eq. (12). In the denoising process, we set \\( k = 0 \\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.357, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.089, + 0.613, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.089, + 0.874, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Figure 3. Comparison of D3PO against existing methods. The horizontal axis represents the number of image sample pairs generated for updating parameters. The rewards denote image size for incompressity objective, negative image size for the compressivity objective, and the LAION aesthetic score for the aesthetic objective. All experiments are conducted with 5 different seeds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.429 + ], + "angle": 0, + "content": "\\(T = 20\\). We first sample an initial state \\(s_0^w = s_0^l = s_0\\) and then use Eq. (2) to generate two segments. After manually choosing which segment is better, the probability of \\(\\pi_{\\theta}(a_0^w | s_0^w)\\) is gradually increasing and \\(\\pi_{\\theta}(a_0^l | s_0^l)\\) is decreasing, which guides the diffusion model to generate images of human preference. However, the approach of only updating \\(\\pi_{\\theta}(\\cdot | s_0)\\) does not fully utilize the information within the segment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.596 + ], + "angle": 0, + "content": "Since the middle states of the segment are noises and semi-finished images, it is hard for humans to judge which segment is better by observing the whole segment. But we can conveniently compare the final image \\(\\mathbf{x}_0\\). Like many RL methods [7, 51, 52] which give rewards by \\(\\forall s_t, a_t \\in \\sigma, r(s_t, a_t) = 1\\) for winning the game and \\(\\forall t \\in \\sigma, r(s_t, a_t) = -1\\) for losing the game, we also assume that if the segment is preferred, then any state-action pair of the segment is better than the other segment. By using this assumption, we construct \\(T\\) sub-segments for the agent to learn, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.605, + 0.469, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} = \\left\\{s _ {i}, a _ {i}, s _ {i + 1}, a _ {i + 1}, \\dots , s _ {T - 1}, a _ {T - 1} \\right\\}, \\quad 0 \\leq i \\leq T - 1 \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Using these sub-segments, the overall loss of the D3PO algorithm gives:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.678, + 0.469, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} (\\theta) = - E _ {\\left(s _ {i}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}) ], \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.469, + 0.758 + ], + "angle": 0, + "content": "where \\( i \\in [0, T - 1] \\). Compared to Eq. (12), Eq. (14) uses every state-action pair for training, effectively increasing the data utilization of the segment by a factor of \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.469, + 0.789 + ], + "angle": 0, + "content": "The overview of our method is shown in Fig. 1. The pseudocode of D3PO can be seen in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.202, + 0.817 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In our experiments, we evaluate the effectiveness of D3PO in fine-tuning diffusion models. Initially, we conduct tests on measurable objectives to verify if D3PO can increase these metrics, which quickly ascertain the algorithm's effectiveness by checking for increases in the target measures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Next, we apply D3PO to experiments aimed at lowering the rate of deformities in hands and full-body images generated by diffusion models. Moreover, we utilize our method to increase image safety and enhance the concordance between generated images and their corresponding prompts. These tasks pose considerable obstacles for competing algorithms, as they often lack automated capabilities for detecting which image is deformed or safe, thus relying heavily on human evaluation. We use Stable Diffusion v1.5 [47] to generate images in most of the experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.47, + 0.871, + 0.485 + ], + "angle": 0, + "content": "5.1. Pre-defined Quantifiable Objectives Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.614 + ], + "angle": 0, + "content": "We initially conduct experiments with D3PO using quantitative objectives (alternatively referred to as optimal reward models). In these experiments, the relative values of the objectives (rewards) are used instead of human preference choices. Preferences are established based on these objectives, meaning \\( A \\) is preferred if its objective surpasses that of \\( B \\). After training, we validate the effectiveness of our approach by measuring the growth of metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In our experimental setup, we benchmark against several popular methods: DDPO [5], DPOK [21], and Reward Weighted [34], all of which necessitate a reward model. Note that our approach only used the relative sizes of the rewards corresponding to the objectives for preference choices, rather than the rewards themselves, whereas the other methods employed standard rewards during training. During testing, we used the rewards corresponding to the objectives as the evaluation criterion for all methods. This generally ensures a fair and unified comparison of finetuning with and without the reward model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first use the size of the images to measure the preference relationship between two pictures. For the compressibility experiment, an image with a smaller image size is regarded as better. Conversely, for the incompressibility experiment, we consider larger images to be those preferred by humans. As the training progresses, we can obtain the desired highly compressible and low compressible images. We then utilize the LAION aesthetics predictor [50] to pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "dict the aesthetic rating of images. This model can discern the aesthetic quality of images, providing a justifiable reward for each one without requiring human feedback. The model can generate more aesthetic images after fine-tuning. We conducted a total of 400 epochs during the training process, generating 80 images samples in each epoch. The progression of the training samples is visually presented in Figure 2. More quantitative samples are shown in Figure 8. The testing curves of D3PO and other methods are shown in Figure 3. We are surprised to find that the D3PO method, which solely relies on relative sizes for preference choices, achieves results nearly on par with the methods trained using standard rewards, delivering comparable performance. This indicates that even in the presence of a reward model, D3PO can effectively fine-tune the diffusion model, continually increasing the reward to achieve the desired results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.346, + 0.431, + 0.362 + ], + "angle": 0, + "content": "5.2. Experiments without Any Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.471, + 0.477 + ], + "angle": 0, + "content": "We conduct experiments for some objectives without any reward model. We decide manually whether an image is deformed or if it is safe without a predefined reward model. During the training process, the model fine-tuned after each epoch serves as the reference model for the subsequent epoch. For each prompt, we generate 7 images with the same initial state \\( x_{T} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.307, + 0.515 + ], + "angle": 0, + "content": "5.2.1 Reduce Image Distortion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.736 + ], + "angle": 0, + "content": "We use the prompt \"1 hand\" to generate images, manually selecting those that are deformed. Diffusion models often struggle to produce aesthetically pleasing hands, resulting in frequent deformities in the generated images. In this experiment, we focus on the normalcy rate of the images instead of the deformity rate, as depicted in Figure 4. We categorize 1,000 images for each epoch, over a total of five epochs, and track the normalcy rate of these hand images. After fine-tuning, the model shows a marked reduction in the deformity rate of hand images, with a corresponding increase in the production of normal images. Additionally, the fine-tuned model shows a higher probability of generating hands with the correct number of fingers than the pretrained model, as demonstrated in Figure 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.471, + 0.875 + ], + "angle": 0, + "content": "To assess the generality of our method, we generated images with the Anything v5 model \\( {}^{1} \\) , renowned for creating anime character images. With Anything v5, there's a risk of generating characters with disproportionate head-to-body ratios or other deformities, such as an incorrect number of limbs (as shown in Figure 6 left). We categorize such outputs as deformed. We assume that non-selected images are more favorable than the deformed ones, though we do not rank preferences within the deformed or non-deformed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "sets. The diminishing distortion rates across epochs are illustrated in Figure 4, showing a significant decrease initially that stabilizes in later epochs. The visual examples are provided in Figure 6." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.164, + 0.875, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Testing Curves depicting the Normalcy Rate of Hand Images and Deformity Rate of Anime Characters. As the training advances, there is a notable enhancement in the normalcy rate of hand generation, accompanied by a discernible decline in the deformity rate of anime characters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.426, + 0.71, + 0.442 + ], + "angle": 0, + "content": "5.2.2 Enhance Image Safety" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.662 + ], + "angle": 0, + "content": "In this experiment, we utilized unsafe prompts to generate images using a diffusion model. These prompts contained edgy terms that could lead to the creation of both normal and Not Safe for Work (NSFW) images, examples being 'ambiguous beauty' and 'provocative art'. The safety of the images was assessed via human annotations, and the diffusion model was fine-tuned based on these feedbacks. Across 10 epochs, we generated 1,000 images per epoch. Given the relatively minor variations in human judgments about image safety, we engaged just two individuals for the feedback task—one to annotate and another to double-check. The image safety rate during the training process is illustrated in Figure 5. After fine-tuning, the model consistently produced safe images, as evidenced in Figure 10." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.673, + 0.845, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.824, + 0.858, + 0.838 + ], + "angle": 0, + "content": "Figure 5. Safety rate curves of the training procession." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.736, + 0.864 + ], + "angle": 0, + "content": "5.2.3 Prompt-Image Alignment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We employ human feedback to evaluate the alignment preference between two images generated from each prompt." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.385, + 0.902 + ], + "angle": 0, + "content": "1https://huggingface.co/stablediffusionapi/anything-v5" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.416, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.181, + 0.532, + 0.218 + ], + "angle": 0, + "content": "finetuning" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.089, + 0.853, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 6. Image samples of the pre-trained model and the fine-tuned model. The images on the left side of the arrows are distorted images (such as having 3 legs in the image) generated by the pre-trained model, while the images on the right side are normal images generated after fine-tuning the model. Both sets of images used the same initial Gaussian noise, prompts, and seeds." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.391, + 0.456, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.555, + 0.473, + 0.68 + ], + "angle": 0, + "content": "Figure 7. Comparative evaluation of 300 text prompts: Our study involved generating images from the pre-trained diffusion model, the fine-tuned models using other methods, and our fine-tuned model—using the same text prompts. For each prompt, human evaluators are tasked with determining which image is better aligned with the text. Each image was assessed by 5 human raters, and we report the percentage of images that received favorable evaluations. We also highlight the percentage with more than half vote in the black box." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.768 + ], + "angle": 0, + "content": "Table 1. Evaluations of prompt-image alignment. We compare D3PO with those that (a) do not fine-tune, (b) use the preferred images to fine-tune, and (c) use Reward Weighted method to fine-tune. D3PO achieve the best performance. D3PO demonstrated superior performance across all comparisons." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.778, + 0.468, + 0.829 + ], + "angle": 0, + "content": "
MethodsCLIP score (↑)BLIP score (↑)ImageReward score (↑)Human preference (↑)
No fine-tune30.71.950.0414.7%
Preferred images31.01.970.0811%
Reward Weighted31.52.010.1718.7%
D3PO31.92.060.2755.7%
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For each epoch, we use 4,000 prompts, generate two images per prompt, and assess preferences with feedback from 16 different evaluators. The training spans 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.501 + ], + "angle": 0, + "content": "The preference comparisons between images from the pretrained and fine-tuned models are conducted by an additional 5 evaluators, with the comparative preferences depicted in Figure 7. We also execute quantitative evaluations of models using metrics that measure the congruence between prompts and images, including CLIP [42], BLIP [35], and ImageReward [60], as presented in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.62, + 0.531 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.895, + 0.83 + ], + "angle": 0, + "content": "In this paper, we propose a direct preference denoising diffusion policy optimization method, named D3PO, to fine-tune diffusion models purely from human feedback without learning a separate reward model. D3PO views the denoising process as a multi-step MDP, making it possible to utilize the DPO-style optimization formula by formulating the action-value function \\( Q \\) with the reference model and the fine-tuned model. D3PO updates parameters at each step of denoising and consumes much fewer GPU memory overheads than directly applying the DPO algorithm. The empirical experiments illustrate that our method achieves competitive or even better performance compared with a diffusion model fine-tuned with a reward model that is trained with a large amount of images and human preferences in terms of image compressibility, image compressibility and aesthetic quality. We further show that D3PO can also benefit challenging tasks such as reducing image distortion rates, enhancing the safety of the generated images, and aligning prompts and images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.669, + 0.862 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the STI 2030-Major Projects under Grant 2021ZD0201404." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8948" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.348, + 0.2 + ], + "angle": 0, + "content": "[2] Anthropic. Introducing claude, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[3] Philip Bachman and Doina Precup. Data generation as sequential decision making. Advances in Neural Information Processing Systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[6] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[7] Noam Brown and Tuomas Sandholm. Superhuman ai for multiplayer poker. Science, 365(6456):885-890, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[9] Robert Busa-Fekete, Balázs Szörényi, Paul Weng, Weiwei Cheng, and Eyke Hüllermeier. Preference-based reinforcement learning: evolutionary direct policy search using a preference-based racing algorithm. Machine learning, 97: 327-351, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[10] Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[11] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[12] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890-16902, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[16] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[17] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[18] Miroslav Dudík, Katja Hofmann, Robert E Schapire, Aleksandris Slivkins, and Masrour Zoghi. Contextual dueling bandits. In Conference on Learning Theory, pages 563-587. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[20] Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.671, + 0.636 + ], + "angle": 0, + "content": "[24] Google. Bard, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[25] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8949" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[33] Kimin Lee, Laura Smith, and Pieter Abbeel. Pebble: Feedback-efficient interactive reinforcement learning via rel-labeling experience and unsupervised pre-training. arXiv preprint arXiv:2106.05091, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[34] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[36] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[37] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[38] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[39] Runze Liu, Yali Du, Fengshuo Bai, Jiafei Lyu, and Xiu Li. Zero-shot preference learning for offline rl via optimal transport. arXiv preprint arXiv:2306.03615, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.369, + 0.789 + ], + "angle": 0, + "content": "[41] OpenAI. Gpt-4 technical report, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "preprint arXiv:2305.18290, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[46] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.365, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[48] Aadirupa Saha, Aldo Pacchiano, and Jonathan Lee. Dueling rl: Reinforcement learning with trajectory preferences. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 6263-6289. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.521, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[50] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[51] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[52] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676): 354-359, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[55] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[56] Richard S Sutton, Andrew G Barto, et al. Introduction to reinforcement learning. 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[57] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[58] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[59] Aaron Wilson, Alan Fern, and Prasad Tadepalli. A bayesian approach for policy learning from trajectory preference queries. Advances in neural information processing systems, 25, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.1,3,8,14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[61] Yisong Yue, Josef Broder, Robert Kleinberg, and Thorsten Joachims. The k-armed dueling bandits problem. Journal of Computer and System Sciences, 78(5):1538-1556, 2012. JCSS Special Issue: Cloud Computing 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[62] Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. Huatuogpt, towards taming language model to be a doctor. arXiv preprint arXiv:2305.15075, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[64] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8951" + } + ] +] \ No newline at end of file diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_origin.pdf b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a0af8cd2cf67f07cce2436cff96030f59227dea0 --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/df1c01a2-2342-4d01-85c2-b06c392b4ad2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c40f7b3644a25ecc77b37b166b0ef4ec15844bd7265bd1b2f919f922d96b609f +size 9870652 diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/full.md b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..028af9ed9c439eee9c5505ff4d18eff0539892ee --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/full.md @@ -0,0 +1,342 @@ +# Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model + +Kai Yang $^{1*}$ Jian Tao $^{1*}$ Jiafei Lyu $^{1\dagger}$ Chunjiang Ge $^{2}$ + +Jiaxin Chen3 Weihan Shen3 Xiaolong Zhu3 Xiu Li1† + +$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University + +$^{2}$ Department of Automation, Tsinghua University $^{3}$ Parametrix Technology Company Ltd. + +{yk22,tj22,lvjf20}@mails.tsinghua.edu.cn li.xiu@sz.tsinghua.edu.cn + +# Abstract + +Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO. + +# 1. Introduction + +Recent advances in image generation models have yielded unprecedented success in producing high-quality images + +from textual prompts [30, 44, 49]. Diverse approaches, including Generative Adversarial Networks (GANs) [23], autoregressive models [14, 15, 19, 22, 44, 58], Normalizing Flows [16, 46], and diffusion-based techniques [40, 45, 47, 49], have rapidly pushed forward the capabilities of these systems. With the proper textual inputs, such models are now adept at crafting images that are not only visually compelling but also semantically coherent, garnering widespread interest for their potential applications and implications. + +To adapt the aforementioned models for specific downstream tasks, such as the generation of more visually appealing and aesthetic images, Reinforcement Learning from Human Feedback (RLHF) is commonly employed [12]. This technique has been successfully used to refine large language models such as GPT [8, 41]. Now, efforts are being made to extend this method to diffusion models to enhance their performance. One such approach, the DDPO method [5], aims to enhance image complexity, aesthetic quality, and the alignment between prompts and images. The ReLF approach [60] introduces a novel reward model, named ImageReward, which is specifically trained to discern human aesthetic preferences in text-to-image synthesis. This model is then utilized to fine-tune diffusion models to produce images that align more closely with human preferences. Nonetheless, developing a robust reward model for various tasks can be challenging, often necessitating a vast collection of images and abundant training resources. For example, to diminish the rate of deformities in character images, one must amass a substantial dataset of deformed and non-deformed images generated from identical prompts. Subsequently, a network is constructed to discern and learn the human preference for non-deformed imagery, serving as the reward model. + +In the field of natural language processing, Direct Preference Optimization (DPO) has been proposed to reduce training costs [43]. This method forgoes the training of a + +![](images/ef3b85de653712224e547e7d86884b6bac70ef933b10322e1f6a5616f8fb4c4a.jpg) +Figure 1. Overview of D3PO. The diffusion model generates two corresponding images based on the provided prompts. Guided by specific task requirements—such as improving prompt-image alignment, enhancing image incompressibility, or refining aesthetic quality—human evaluators select the preferred image. Leveraging this human feedback, our method directly updates the diffusion model's parameters without necessitating the training of a reward model. + +reward model and directly fine-tunes language models according to human preferences. However, this straightforward and easy-to-train method encounters challenges when applied to fine-tune diffusion models. During the DPO training process, the complete sentence generated by the language model is treated as a single output, necessitating the storage of gradients from multiple forward passes. With diffusion models, one must store the gradients across multiple latent image representations, which are significantly larger than word embeddings, leading to memory consumption that is typically unsustainable. + +To address the issue of high computational overhead and enable the use of the DPO method to fine-tune diffusion models directly with human feedback, we conceptualize the denoising process as a multi-step MDP, which utilizes a pre-trained model to represent an action value function $Q$ that is commonly estimated in RL. We extend the theoretical framework of DPO into the formulated MDP, which allows for direct parameter updates at each step of the denoising process based on human feedback, thereby circumventing the significant computational costs and eliminating the need for a reward model. To the best of our knowledge, this is the first work that fine-tune diffusion models without reward models. Our main contributions are as follows: + +- We introduce an innovative approach for fine-tuning diffusion models that could significantly modify the current RLHF framework for fine-tuning diffusion models. This method bypasses resource-intensive reward model train + +ing by utilizing direct human feedback, making the process more efficient and cost-effective. + +- We expand the theoretical framework of DPO into a multi-step MDP, demonstrating that directly updating the policy based on human preferences within an MDP is equivalent to learning the optimal reward model first and then using it to guide policy updates. This establishes a robust theoretical foundation and provides assurance for our proposed method. +- In our experiments, we have demonstrated the effectiveness of our method by using human feedback to successfully address issues of hand and full-body deformities, enhance the safety of generated images, and improve prompt-image alignment. + +# 2. Related Work + +Diffusion models. Denoising diffusion probabilistic models, introduced in [54] and further advanced by [26], have emerged as powerful tools for generating diverse data types. They have been successfully applied in various domains such as image synthesis [44, 49], video generation [27, 36, 53], and robotics control systems [1, 11, 29]. Notably, test-to-image diffusion models have enabled the creation of highly realistic images from textual descriptions [44, 49], opening new avenues in digital art and design. + +Recent studies have focused on refining the guidance of diffusion models for more precise manipulation over the + +generative process. Techniques such as adapters [63] and compositional approaches [17, 38] have been introduced to incorporate additional input constraints and blend multiple models, respectively, enhancing image quality and generation control. The implementation of classifier-based [13] and classifier-free guidance [25] has also contributed significantly to achieving greater autonomy in the generation process, resulting in outputs that closely align with user intentions. In our work, we utilize Stable Diffusion [45] to generate images based on specific prompts. + +RLHF. RLHF stands as a salient strategy in the domain of machine learning when objectives are complex or difficult to define explicitly. This technique has been instrumental across various applications, from gaming, as demonstrated with Atari [4, 12], to more intricate tasks in robotics [10, 64]. The integration of RLHF into the development of large language models (LLMs) has marked a significant milestone in the field, with notable models like OpenAI's GPT-4 [41], Anthropic's Claude [2], Google's Bard [24], and Meta's Llama 2-Chat [57] leveraging this approach to enhance their performance and relevance. The effectiveness of RLHF in refining the behavior of LLMs to be more aligned with human values, such as helpfulness and harmlessness, has been extensively studied [4, 64]. The technique has also proven beneficial in more focused tasks, such as summarization, where models are trained to distill extensive information into concise representations [55]. Some recent research utilizes Reinforcement Learning from AI Feedback (RLAIF) [32, 62] as an alternative to RLHF for model fine-tuning. RLAIF offers convenience and efficiency by replacing human feedback with AI-generated feedback. However, for tasks like assessing hand generation normality or image aesthetic appeal, reliable judgment models are currently lacking. Hence, this paper still relies on human feedback for evaluation. + +Fine-tune Diffusion Models with RL. Before applying diffusion models, data generation has been regarded as a sequential decision-making problem and combined with reinforcement learning [3]. More recently, the SFT method [20] applied reinforcement learning to diffusion models to enhance existing fast DDPM samplers [26]. Reward Weighted method [34] explored using human feedback to align text-to-image models. It uses the reward model for the coefficients of the loss function instead of using reinforcement learning optimization objectives. ReFL [60] employs the framework of RLHF. It begins by training a model called ImageReward based on human preferences and then finetunes the diffusion model using reinforcement learning. DDPO [5] treats the denoising process of diffusion models as a MDP to fine-tune diffusion models with many reward models. DPOK [21] combine the KL divergence into the DDPO loss and use it to better align text-to-image objectives. All these models need a robust reward model, which + +demands a substantial dataset of images and extensive human evaluations. + +Direct Preference Optimization. In the realm of reinforcement learning, exploring policies derived from preferences rather than explicit rewards has gained attention through various methods. The Contextual Dueling Bandit (CDB) framework [18, 61] introduces the concept of a von Neumann winner, shifting away from the pursuit of an optimal policy directly based on rewards. Preference-based Reinforcement Learning (PbRL) [9, 39, 48] learns from binary preferences inferred from a cryptic scoring function instead of explicit rewards. Recently, the DPO approach [43] was proposed, which fine-tunes LLMs directly using preferences. DPO leverages the correlation between reward functions and optimal policies, effectively addressing the challenge of constrained reward maximization in a single phase of policy training. + +# 3. Preliminaries + +MDP. We consider the MDP formulation described in [56]. In this setting, an agent perceives a state $s \in S$ and executes an action, where $S, A$ denote state space and action space, respectively. The transition probability function, denoted by $P(s'|s, a)$ , governs the progression from the current state $s$ to the subsequent state $s'$ upon the agent's action $a$ . Concurrently, the agent is awarded a scalar reward $r$ , determined by the reward function $r: S \times A \to \mathbb{R}$ . The agent's objective is to ascertain a policy $\pi(a|s)$ that maximizes the cumulative returns of trajectories $\tau = (s_0, a_0, s_1, a_1, \dots, s_{T-1}, a_{T-1})$ , which can be represented as: $\mathcal{J}(\pi) = \mathbb{E}_{\tau}[\sum_{t=0}^{T-1} r(s_t, a_t)]$ . + +Diffusion models. Diffusion models learn to model a probability distribution $p(x)$ by inverting a Markovian forward process $q(\pmb{x}_t|\pmb{x}_{t-1})$ which adds noise to the data. The denoising process is modeled by a neural network to predict the mean of $\pmb{x}_{t-1}$ or the noise $\epsilon_{t-1}$ of the forward process. In our work, we use network $\mu_\theta(\pmb{x}_t; t)$ to predict the mean of $\pmb{x}_{t-1}$ instead of predicting the noise. Using the mean squared error (MSE) as a measure, the objective of this network can be written as: + +$$ +\mathcal {L} = \mathbb {E} _ {t \sim [ 1, T ], \boldsymbol {x} _ {0} \sim p (\boldsymbol {x} _ {0}), \boldsymbol {x} _ {t} \sim q (\boldsymbol {x} _ {t} | \boldsymbol {x} _ {0})} \left[ \| \tilde {\boldsymbol {\mu}} (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}) - \boldsymbol {\mu} _ {\theta} (\boldsymbol {x} _ {t}, t) \| ^ {2} \right], \tag {1} +$$ + +where $\tilde{\mu}_{\theta}(\boldsymbol{x}_t,\boldsymbol{x}_0)$ is the forward process posterior mean. In the case of conditional generative modeling, the diffusion models learn to model $p(x|c)$ , where $c$ is the conditioning information, i.e., image category and image caption. This is done by adding an additional input $c$ to the denoising neural network, as in $\mu_{\theta}(\boldsymbol{x}_t,t;\boldsymbol{c})$ . To generate a sample from the learned distribution $p_{\theta}(x|c)$ , we start by drawing a sample $\boldsymbol{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and then progressively denoise the sample by iterated application of $\epsilon_{\theta}$ according to a specific sampler [26]. Given the noise-related parameter $\sigma_t$ , the reverse process can be written as: + +$$ +p _ {\theta} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {c}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(\boldsymbol {x} _ {t}, \boldsymbol {c}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right). \tag {2} +$$ + +Reward learning for preferences. The basic framework to model preferences is to learn a reward function $r^{*}(s,a)$ from human feedback [12, 33, 59]. The segment $\sigma = \{s_k,a_k,s_{k + 1},a_{k + 1},\dots ,s_m,a_m\}$ is a sequence of observations and actions. By following the Bradley-Terry model [6], the human preference distribution $p^*$ by using the reward function can be expressed as: + +$$ +p ^ {*} (\sigma_ {1} \succ \sigma_ {0}) = \frac {\exp \left(\sum_ {t = k} ^ {T} r ^ {*} \left(s _ {t} ^ {1} , a _ {t} ^ {1}\right)\right)}{\sum_ {i \in \{0 , 1 \}} \exp \left(\sum_ {t = k} ^ {T} r ^ {*} \left(s _ {t} ^ {i} , a _ {t} ^ {i}\right)\right)}, \tag {3} +$$ + +where $\sigma_{i} \succ \sigma_{j}$ denotes that segment $i$ is preferable to segment $j$ . Now we have the preference distribution of human feedback, and we want to use a network $r_{\phi}$ to approximate $r^{*}$ . Given the human preference $y \in \{(1,0), (0,1)\}$ which is recorded in dataset $\mathcal{D}$ as a triple $(\sigma_{0}, \sigma_{1}, y)$ , framing the problem as a binary classification, the reward function modeled as a network is updated by minimizing: + +$$ +\begin{array}{l} \mathcal {L} (\phi) = - \mathbb {E} _ {\left(\sigma_ {1}, \sigma_ {0}, y\right) \sim \mathcal {D}} [ y (0) \log p _ {\phi} \left(\sigma_ {0} \succ \sigma_ {1}\right) \\ \left. + y (1) \log p _ {\phi} \left(\sigma_ {1} \succ \sigma_ {0}\right) \right]. \tag {4} \\ \end{array} +$$ + +# 4. Method + +In this section, we describe a method to directly fine-tune diffusion models using human feedback, bypassing the conventional requirement for a reward model. Initially, we reinterpret the denoising process inherent in diffusion models as a multi-step MDP. Then we extend the theory of DPO to MDP, which allows us to apply the principles to effectively translate human preferences into policy improvements in diffusion models. + +# 4.1. Denoising process as a multi-step MDP + +We conceptualize the denoising process within the diffusion model as a multi-step MDP, which varies slightly from the approach outlined in [5]. To enhance clarity, we have redefined the states, transition probabilities, and policy functions. The correspondence between notations in the diffusion model and the MDP is established as follows: + +$$ +\begin{array}{l} \mathbf {s} _ {t} \triangleq (\boldsymbol {c}, t, \boldsymbol {x} _ {T - t}) \quad P \left(\mathbf {s} _ {t + 1} \mid \mathbf {s} _ {t}, \mathbf {a} _ {t}\right) \triangleq \left(\delta_ {\boldsymbol {c}}, \delta_ {t + 1}, \delta_ {\boldsymbol {x} _ {T - 1 - t}}\right) \\ \mathbf {a} _ {t} \triangleq \mathbf {x} _ {T - 1 - t} \quad \pi (\mathbf {a} _ {t} \mid \mathbf {s} _ {t}) \triangleq p _ {\theta} (\mathbf {x} _ {T - 1 - t} \mid \mathbf {c}, t, \mathbf {x} _ {T - t}) \\ \rho_ {0} \left(\mathbf {s} _ {0}\right) \triangleq (p (\boldsymbol {c}), \delta_ {0}, \mathcal {N} (\mathbf {0}, \mathbf {I})) \\ r \left(\mathbf {s} _ {t}, \mathbf {a} _ {t}\right) \triangleq r \left(\left(\boldsymbol {c}, t, \boldsymbol {x} _ {T - t}\right), \boldsymbol {x} _ {T - t - 1}\right) \\ \end{array} +$$ + +where $\delta_{x}$ represents the Dirac delta distribution, and $T$ denotes the maximize denoising timesteps. Leveraging this mapping, we can employ RL techniques to fine-tune diffusion models by maximizing returns. However, this approach requires a proficient reward model capable of adequately rewarding the noisy images. The task becomes exceptionally challenging, particularly when $t$ is low, and $\pmb{x}_{T - t}$ closely resembles Gaussian noise, even for an experienced expert. + +# 4.2. Direct Preference Optimization for MDP + +The DPO method does not train a separate reward model but instead directly optimizes the LLMs with the preference data. Given a prompt $x$ and a pair of answers $(y_1, y_0) \sim \pi_{\mathrm{ref}}(y|x)$ , where $\pi_{\mathrm{ref}}$ represents the reference (pre-trained) model, these responses are ranked and stored in $\mathcal{D}$ as a tuple $(x, y_w, y_l)$ , where $y_w$ denotes the preferred answer and $y_l$ indicates the inferior one. DPO optimizes $\pi_\theta$ with the human preference dataset by using the following loss: + +$$ +\mathcal {L} _ {\mathrm {D P O}} (\theta) = - \mathbb {E} _ {(x, y _ {w}, y _ {l}) \sim \mathcal {D}} \left[ \log \rho \left(\beta \log \frac {\pi_ {\theta} (y _ {w} | x)}{\pi_ {\text {r e f}} (y _ {w} | x)} - \beta \log \frac {\pi_ {\theta} (y _ {l} | x)}{\pi_ {\text {r e f}} (y _ {l} | x)}\right) \right]. \tag {5} +$$ + +Here $\rho$ is the logistic function, and $\beta$ is the parameter controlling the deviation from the $\pi_{\theta}$ and $\pi_{\mathrm{ref}}$ . In our framework, we treat segments $\sigma^1, \sigma^0$ as $y_1, y_0$ and use DPO to fine-tune diffusion models. However, directly using this method faces difficulties since the segments contain a large number (usually 20-50) of the image latent, which occupy a large amount of GPU memory (each image is about 6G even when using LoRA [28]). Since we can only get human preferences for the final image $x_0$ , if we want to update $\pi_{\theta}(\sigma) = \prod_{t=k}^{T} \pi_{\theta}(s_t, a_t)$ , it will consume more than 100G GPU memory, which makes the fine-tuning process nearly impossible. + +To address this problem, we extend the DPO theory to MDP. Firstly, we need to reconsider the objective of the RL method. For the MDP problem, the agents take action by considering maximizing the expected return instead of the current reward. For actor-critic methods such as DDPG [37], the optimization objective of policy $\pi$ gives: + +$$ +\max _ {\pi} \mathbb {E} _ {s \sim d ^ {\pi}, a \sim \pi (\cdot | s)} [ Q ^ {*} (s, a) ]. \tag {6} +$$ + +Here, $d^{\pi} = (1 - \gamma)\sum_{t = 0}^{\infty}\gamma^{t}P_{t}^{\pi}(s)$ represents the state visitation distribution, where $P_{t}^{\pi}(s)$ denotes the probability of being in state $s$ at timestep $t$ given policy $\pi$ . Additionally, $Q^{*}(s,a)$ denotes the optimal action-value function. The optimal policy can be written as: + +$$ +\pi^ {*} (a | s) = \left\{ \begin{array}{l l} 1, & \text {i f} a = \arg \max _ {\hat {a}} Q ^ {*} (s, \hat {a}), \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +Similar to some popular methods, we use KL-divergence to prevent the fine-tuned policy from deviating from the reference policy, hence relieving the out-of-distribution (OOD) issue. By incorporating this constraint, the RL objective can be rewritten as: + +$$ +\max _ {\pi} \mathbb {E} _ {s \sim d ^ {\pi}, a \sim \pi (\cdot | s)} \left[ Q ^ {*} (s, a) \right] - \beta \mathbb {D} _ {K L} \left[ \pi (a | s) \| \pi_ {\text {r e f}} (a | s) \right]. \tag {8} +$$ + +Here, $\beta$ is the temperature parameter that controls the deviation of $\pi_{\theta}(a|s)$ and $\pi_{\mathrm{ref}}(a|s)$ . + +# Compressibility + +# Incompressibility + +# Aesthetic Quality + +![](images/e2f8ad454545b73bf2416007ff54ab2e1a1912a35c6b5527e66962f40da21a8d.jpg) +Figure 2. Progression of samples targeting compressibility, incompressibility, and aesthetic quality objectives. With the respective focus during training, images exhibit reduced detail and simpler backgrounds for compressibility, richer texture details for incompressibility, and an overall increased aesthetic appeal when prioritizing aesthetic quality. + +Proposition 1 Given the objective of Eq. (8), the optimal policy $\pi^{*}(a|s)$ has the following expression: + +$$ +\pi^ {*} (a | s) = \pi_ {\text {r e f}} (a | s) \exp \left(\frac {1}{\beta} Q ^ {*} (s, a)\right). \tag {9} +$$ + +The proof can be seen in Appendix B.1. By rearranging the formula of Eq. (9), we can obtain that: + +$$ +Q ^ {*} (s, a) = \beta \log \frac {\pi^ {*} (a | s)}{\pi_ {\mathrm {r e f}} (a | s)}. \tag {10} +$$ + +Now, considering Eq. (3) and noticing that $Q^{*}(s_{t},a_{t}) = \mathbb{E}\left[\sum_{t = k}^{T}r^{*}(s_{t},a_{t})\right]$ under the policy $\pi^* (a|s)$ , we make a substitution. By replacing $\sum_{t = k}^{T}r^{*}(s_{t},a_{t})$ with $Q^{*}(s_{t},a_{t})$ we define a new distribution that can be rewritten as: + +$$ +\tilde {p} ^ {*} \left(\sigma_ {1} \succ \sigma_ {0}\right) = \frac {\exp \left(Q ^ {*} \left(s _ {k} ^ {1} , a _ {k} ^ {1}\right)\right)}{\sum_ {i \in \{0 , 1 \}} \exp \left(Q ^ {*} \left(s _ {k} ^ {i} , a _ {k} ^ {i}\right)\right)}. \tag {11} +$$ + +We suppose $\sum_{t = k}^{m}r^{*}(s_{t},a_{t})$ is sampled from a normal distribution with mean $\mathbb{E}[\sum_{t = k}^{m}r^{*}(s_{t},a_{t})]$ and standard deviation $\sigma^2$ . From a statistical perspective, we can establish the relationship between the new distribution $\tilde{p}^{*}(\sigma_{1}\succ \sigma_{0})$ and the raw distribution $p^{*}(\sigma_{1}\succ \sigma_{0})$ . + +Proposition 2 For $i \in \{0,1\}$ , suppose the expected return satisfies a normal distribution, i.e., $\sum_{t=0}^{T} r^*(s_t^i, a_t^i) \sim \mathcal{N}(Q^*(s_0^i, a_0^i), \sigma^2)$ . Given $Q^*(s,a) \in [Q_{\min}, Q_{\max}]$ where $Q_{\min}$ and $Q_{\max}$ represent the lower and upper bounds of the values, then + +$$ +\left| p ^ {*} \left(\sigma_ {1} \succ \sigma_ {0}\right) - \tilde {p} ^ {*} \left(\sigma_ {1} \succ \sigma_ {0}\right) \right| < \frac {\left(\xi^ {2} + 1\right) \left(\exp \left(\sigma^ {2}\right) - 1\right)}{1 6 \xi \delta} +$$ + +with probability at least $1 - \delta$ , where $\xi = \frac{\exp(Q_{\mathrm{max}})}{\exp(Q_{\mathrm{min}})}$ . + +The proof can be seen in Appendix B.2. In practical applications, as the volume of data increases, it becomes easier to satisfy the assumption of normality. Additionally, we can use clipping operations to constrain the $Q$ values within a certain range, which introduces upper and lower bounds. Therefore, the aforementioned assumption is reasonable. As shown in proposition 2, their deviation can be bounded at the scale of $\mathcal{O}\left(\frac{\xi}{\delta} (\exp (\sigma^2) - 1)\right)$ . It is clear that this bound can approach 0 if the $\sigma^2$ is close to 0. In practice, $\sigma^2$ approaches 0 if the $Q$ function can be estimated with a small standard deviation. + +By combining Eq. (11), Eq. (4), and Eq. (10), replacing $p^*(\sigma_1 \succ \sigma_0)$ with $\tilde{p}^*(\sigma_1 \succ \sigma_0)$ , and substituting $\pi^*(s, a)$ with the policy network $\pi_\theta(s, a)$ that requires learning, we derive the following loss function for updating $\pi_\theta(a|s)$ : + +$$ +\mathcal {L} (\theta) = - \mathbb {E} _ {\left(s _ {k}, \sigma_ {w}, \sigma_ {l}\right)} [ \log \rho (\beta \log \frac {\pi_ {\theta} \left(a _ {k} ^ {w} \mid s _ {k} ^ {w}\right)}{\pi_ {\text {r e f}} \left(a _ {k} ^ {w} \mid s _ {k} ^ {w}\right)} - \beta \log \frac {\pi_ {\theta} \left(a _ {k} ^ {l} \mid s _ {k} ^ {l}\right)}{\pi_ {\text {r e f}} \left(a _ {k} ^ {l} \mid s _ {k} ^ {l}\right)}) ], \tag {12} +$$ + +where $\sigma_w = \{s_k^w, a_k^w, s_{k+1}^w, a_{k+1}^w, \dots, s_T^w, a_T^w\}$ denotes the segment preferred over another segment $\sigma_l = \{s_k^l, a_k^l, s_{k+1}^l, a_{k+1}^l, \dots, s_T^l, a_T^l\}$ . + +# 4.3. Direct preference for Denoising Diffusion Policy Optimization + +Considering the denoising process as a multi-step MDP and using the mapping relationship depicted in Section 4.1, we can use DPO to directly update diffusion models by using Eq. (12). In the denoising process, we set $k = 0$ and + +![](images/6ad20a06d7db5be20dee43c8ce0c6c281c4b0fca3f2e0414b22f03637ad7b813.jpg) +Figure 3. Comparison of D3PO against existing methods. The horizontal axis represents the number of image sample pairs generated for updating parameters. The rewards denote image size for incompressity objective, negative image size for the compressivity objective, and the LAION aesthetic score for the aesthetic objective. All experiments are conducted with 5 different seeds. + +![](images/57f5e874c62260282f9da65854f1439ecb81eb556c5754d9caeff260f0ce439f.jpg) + +![](images/fdcff313aeee8d751195d901691896162429675553edf98901dccba206f7d114.jpg) + +$T = 20$ . We first sample an initial state $s_0^w = s_0^l = s_0$ and then use Eq. (2) to generate two segments. After manually choosing which segment is better, the probability of $\pi_{\theta}(a_0^w | s_0^w)$ is gradually increasing and $\pi_{\theta}(a_0^l | s_0^l)$ is decreasing, which guides the diffusion model to generate images of human preference. However, the approach of only updating $\pi_{\theta}(\cdot | s_0)$ does not fully utilize the information within the segment. + +Since the middle states of the segment are noises and semi-finished images, it is hard for humans to judge which segment is better by observing the whole segment. But we can conveniently compare the final image $\mathbf{x}_0$ . Like many RL methods [7, 51, 52] which give rewards by $\forall s_t, a_t \in \sigma, r(s_t, a_t) = 1$ for winning the game and $\forall t \in \sigma, r(s_t, a_t) = -1$ for losing the game, we also assume that if the segment is preferred, then any state-action pair of the segment is better than the other segment. By using this assumption, we construct $T$ sub-segments for the agent to learn, which can be written as: + +$$ +\sigma_ {i} = \left\{s _ {i}, a _ {i}, s _ {i + 1}, a _ {i + 1}, \dots , s _ {T - 1}, a _ {T - 1} \right\}, \quad 0 \leq i \leq T - 1 \tag {13} +$$ + +Using these sub-segments, the overall loss of the D3PO algorithm gives: + +$$ +\mathcal {L} _ {i} (\theta) = - E _ {\left(s _ {i}, \sigma_ {w}, \sigma_ {l}\right)} [ \log \rho (\beta \log \frac {\pi_ {\theta} \left(a _ {i} ^ {w} \mid s _ {i} ^ {w}\right)}{\pi_ {\text {r e f}} \left(a _ {i} ^ {w} \mid s _ {i} ^ {w}\right)} - \beta \log \frac {\pi_ {\theta} \left(a _ {i} ^ {l} \mid s _ {i} ^ {l}\right)}{\pi_ {\text {r e f}} \left(a _ {i} ^ {l} \mid s _ {i} ^ {l}\right)}) ], \tag {14} +$$ + +where $i \in [0, T - 1]$ . Compared to Eq. (12), Eq. (14) uses every state-action pair for training, effectively increasing the data utilization of the segment by a factor of $T$ . + +The overview of our method is shown in Fig. 1. The pseudocode of D3PO can be seen in Appendix A. + +# 5. Experiment + +In our experiments, we evaluate the effectiveness of D3PO in fine-tuning diffusion models. Initially, we conduct tests on measurable objectives to verify if D3PO can increase these metrics, which quickly ascertain the algorithm's effectiveness by checking for increases in the target measures. + +Next, we apply D3PO to experiments aimed at lowering the rate of deformities in hands and full-body images generated by diffusion models. Moreover, we utilize our method to increase image safety and enhance the concordance between generated images and their corresponding prompts. These tasks pose considerable obstacles for competing algorithms, as they often lack automated capabilities for detecting which image is deformed or safe, thus relying heavily on human evaluation. We use Stable Diffusion v1.5 [47] to generate images in most of the experiments. + +# 5.1. Pre-defined Quantifiable Objectives Results + +We initially conduct experiments with D3PO using quantitative objectives (alternatively referred to as optimal reward models). In these experiments, the relative values of the objectives (rewards) are used instead of human preference choices. Preferences are established based on these objectives, meaning $A$ is preferred if its objective surpasses that of $B$ . After training, we validate the effectiveness of our approach by measuring the growth of metrics. + +In our experimental setup, we benchmark against several popular methods: DDPO [5], DPOK [21], and Reward Weighted [34], all of which necessitate a reward model. Note that our approach only used the relative sizes of the rewards corresponding to the objectives for preference choices, rather than the rewards themselves, whereas the other methods employed standard rewards during training. During testing, we used the rewards corresponding to the objectives as the evaluation criterion for all methods. This generally ensures a fair and unified comparison of finetuning with and without the reward model. + +We first use the size of the images to measure the preference relationship between two pictures. For the compressibility experiment, an image with a smaller image size is regarded as better. Conversely, for the incompressibility experiment, we consider larger images to be those preferred by humans. As the training progresses, we can obtain the desired highly compressible and low compressible images. We then utilize the LAION aesthetics predictor [50] to pre + +dict the aesthetic rating of images. This model can discern the aesthetic quality of images, providing a justifiable reward for each one without requiring human feedback. The model can generate more aesthetic images after fine-tuning. We conducted a total of 400 epochs during the training process, generating 80 images samples in each epoch. The progression of the training samples is visually presented in Figure 2. More quantitative samples are shown in Figure 8. The testing curves of D3PO and other methods are shown in Figure 3. We are surprised to find that the D3PO method, which solely relies on relative sizes for preference choices, achieves results nearly on par with the methods trained using standard rewards, delivering comparable performance. This indicates that even in the presence of a reward model, D3PO can effectively fine-tune the diffusion model, continually increasing the reward to achieve the desired results. + +# 5.2. Experiments without Any Reward Model + +We conduct experiments for some objectives without any reward model. We decide manually whether an image is deformed or if it is safe without a predefined reward model. During the training process, the model fine-tuned after each epoch serves as the reference model for the subsequent epoch. For each prompt, we generate 7 images with the same initial state $x_{T} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . + +# 5.2.1 Reduce Image Distortion + +We use the prompt "1 hand" to generate images, manually selecting those that are deformed. Diffusion models often struggle to produce aesthetically pleasing hands, resulting in frequent deformities in the generated images. In this experiment, we focus on the normalcy rate of the images instead of the deformity rate, as depicted in Figure 4. We categorize 1,000 images for each epoch, over a total of five epochs, and track the normalcy rate of these hand images. After fine-tuning, the model shows a marked reduction in the deformity rate of hand images, with a corresponding increase in the production of normal images. Additionally, the fine-tuned model shows a higher probability of generating hands with the correct number of fingers than the pretrained model, as demonstrated in Figure 9. + +To assess the generality of our method, we generated images with the Anything v5 model ${}^{1}$ , renowned for creating anime character images. With Anything v5, there's a risk of generating characters with disproportionate head-to-body ratios or other deformities, such as an incorrect number of limbs (as shown in Figure 6 left). We categorize such outputs as deformed. We assume that non-selected images are more favorable than the deformed ones, though we do not rank preferences within the deformed or non-deformed + +sets. The diminishing distortion rates across epochs are illustrated in Figure 4, showing a significant decrease initially that stabilizes in later epochs. The visual examples are provided in Figure 6. + +![](images/15638d63e4b751f1e53b0d42ab5d1f5d75d5436dce644bbb98cb668c2598ee5d.jpg) +Figure 4. Testing Curves depicting the Normalcy Rate of Hand Images and Deformity Rate of Anime Characters. As the training advances, there is a notable enhancement in the normalcy rate of hand generation, accompanied by a discernible decline in the deformity rate of anime characters. + +# 5.2.2 Enhance Image Safety + +In this experiment, we utilized unsafe prompts to generate images using a diffusion model. These prompts contained edgy terms that could lead to the creation of both normal and Not Safe for Work (NSFW) images, examples being 'ambiguous beauty' and 'provocative art'. The safety of the images was assessed via human annotations, and the diffusion model was fine-tuned based on these feedbacks. Across 10 epochs, we generated 1,000 images per epoch. Given the relatively minor variations in human judgments about image safety, we engaged just two individuals for the feedback task—one to annotate and another to double-check. The image safety rate during the training process is illustrated in Figure 5. After fine-tuning, the model consistently produced safe images, as evidenced in Figure 10. + +![](images/6ce5b477a552d5771aa5ead7de7b068fb32209b1d77b2032c0f48fe197f60f4c.jpg) +Figure 5. Safety rate curves of the training procession. + +# 5.2.3 Prompt-Image Alignment + +We employ human feedback to evaluate the alignment preference between two images generated from each prompt. + +![](images/6c4009da913d0a9a7446270fa01b84a9fb70316a54549d285dcd10050220ac30.jpg) +Figure 6. Image samples of the pre-trained model and the fine-tuned model. The images on the left side of the arrows are distorted images (such as having 3 legs in the image) generated by the pre-trained model, while the images on the right side are normal images generated after fine-tuning the model. Both sets of images used the same initial Gaussian noise, prompts, and seeds. + +![](images/b95f3a01561f2c5011769c87219c3eb4c077a1be33496906b4f9d8281381885d.jpg) +finetuning + +![](images/f0a8bc25c6dbdcb81331fe34180c03e75e1e399413d5f3703b07355dbf23631d.jpg) +Figure 7. Comparative evaluation of 300 text prompts: Our study involved generating images from the pre-trained diffusion model, the fine-tuned models using other methods, and our fine-tuned model—using the same text prompts. For each prompt, human evaluators are tasked with determining which image is better aligned with the text. Each image was assessed by 5 human raters, and we report the percentage of images that received favorable evaluations. We also highlight the percentage with more than half vote in the black box. + +Table 1. Evaluations of prompt-image alignment. We compare D3PO with those that (a) do not fine-tune, (b) use the preferred images to fine-tune, and (c) use Reward Weighted method to fine-tune. D3PO achieve the best performance. D3PO demonstrated superior performance across all comparisons. + +
MethodsCLIP score (↑)BLIP score (↑)ImageReward score (↑)Human preference (↑)
No fine-tune30.71.950.0414.7%
Preferred images31.01.970.0811%
Reward Weighted31.52.010.1718.7%
D3PO31.92.060.2755.7%
+ +For each epoch, we use 4,000 prompts, generate two images per prompt, and assess preferences with feedback from 16 different evaluators. The training spans 10 epochs. + +The preference comparisons between images from the pretrained and fine-tuned models are conducted by an additional 5 evaluators, with the comparative preferences depicted in Figure 7. We also execute quantitative evaluations of models using metrics that measure the congruence between prompts and images, including CLIP [42], BLIP [35], and ImageReward [60], as presented in Table 1. + +# 6. Conclusion + +In this paper, we propose a direct preference denoising diffusion policy optimization method, named D3PO, to fine-tune diffusion models purely from human feedback without learning a separate reward model. D3PO views the denoising process as a multi-step MDP, making it possible to utilize the DPO-style optimization formula by formulating the action-value function $Q$ with the reference model and the fine-tuned model. D3PO updates parameters at each step of denoising and consumes much fewer GPU memory overheads than directly applying the DPO algorithm. The empirical experiments illustrate that our method achieves competitive or even better performance compared with a diffusion model fine-tuned with a reward model that is trained with a large amount of images and human preferences in terms of image compressibility, image compressibility and aesthetic quality. We further show that D3PO can also benefit challenging tasks such as reducing image distortion rates, enhancing the safety of the generated images, and aligning prompts and images. + +# Acknowledgements + +This work was supported by the STI 2030-Major Projects under Grant 2021ZD0201404. + +# References + +[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 2 +[2] Anthropic. Introducing claude, 2023. 3 +[3] Philip Bachman and Doina Precup. Data generation as sequential decision making. Advances in Neural Information Processing Systems, 28, 2015. 3 +[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3 +[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 3, 4, 6 +[6] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4 +[7] Noam Brown and Tuomas Sandholm. Superhuman ai for multiplayer poker. Science, 365(6456):885-890, 2019. 6 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[9] Robert Busa-Fekete, Balázs Szörényi, Paul Weng, Weiwei Cheng, and Eyke Hüllermeier. Preference-based reinforcement learning: evolutionary direct policy search using a preference-based racing algorithm. Machine learning, 97: 327-351, 2014. 3 +[10] Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023. 3 +[11] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2 +[12] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 1, 3, 4 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3 +[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1 + +[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890-16902, 2022. 1 +[16] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1 +[17] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 3 +[18] Miroslav Dudík, Katja Hofmann, Robert E Schapire, Aleksandris Slivkins, and Masrour Zoghi. Contextual dueling bandits. In Conference on Learning Theory, pages 563-587. PMLR, 2015. 3 +[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1 +[20] Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362, 2023. 3 +[21] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381, 2023. 3, 6 +[22] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 1 +[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[24] Google. Bard, 2023. 3 +[25] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[28] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[29] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 2 + +[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1 +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16 +[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3 +[33] Kimin Lee, Laura Smith, and Pieter Abbeel. Pebble: Feedback-efficient interactive reinforcement learning via rel-labeling experience and unsupervised pre-training. arXiv preprint arXiv:2106.05091, 2021. 4 +[34] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3, 6 +[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8 +[36] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 2 +[37] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 4 +[38] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3 +[39] Runze Liu, Yali Du, Fengshuo Bai, Jiafei Lyu, and Xiu Li. Zero-shot preference learning for offline rl via optimal transport. arXiv preprint arXiv:2306.03615, 2023. 3 +[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[41] OpenAI. Gpt-4 technical report, 2023. 1, 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8 +[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct + +preprint arXiv:2305.18290, 2023. 1, 3 +[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2 +[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3 +[46] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 6 +[48] Aadirupa Saha, Aldo Pacchiano, and Jonathan Lee. Dueling rl: Reinforcement learning with trajectory preferences. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 6263-6289. PMLR, 2023. 3 +[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[50] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6 +[51] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 6 +[52] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676): 354-359, 2017. 6 +[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 + +[55] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 3 +[56] Richard S Sutton, Andrew G Barto, et al. Introduction to reinforcement learning. 1998. 3 +[57] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[58] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1 +[59] Aaron Wilson, Alan Fern, and Prasad Tadepalli. A bayesian approach for policy learning from trajectory preference queries. Advances in neural information processing systems, 25, 2012. 4 +[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.1,3,8,14 +[61] Yisong Yue, Josef Broder, Robert Kleinberg, and Thorsten Joachims. The k-armed dueling bandits problem. Journal of Computer and System Sciences, 78(5):1538-1556, 2012. JCSS Special Issue: Cloud Computing 2011. 3 +[62] Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. Huatuogpt, towards taming language model to be a doctor. arXiv preprint arXiv:2305.15075, 2023. 3 +[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3 +[64] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 3 \ No newline at end of file diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/images.zip b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fd2acaf09ed7e3b8609ee5e997c134ef813175a8 --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae00ada8de5b4e0761a50dbbb2e0702fdccaca3850cfb0e59792276a941c24be +size 506253 diff --git a/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/layout.json b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..06934a34f876fe602f633587e546312d27872fa0 --- /dev/null +++ b/2024/Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model/layout.json @@ -0,0 +1,9505 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 465, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 465, + 138 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 465, + 138 + ], + "type": "text", + "content": "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "spans": [ + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": "Kai Yang" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": " Jian Tao" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": " Jiafei Lyu" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "text", + "content": " Chunjiang Ge" + }, + { + "bbox": [ + 162, + 160, + 430, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 175, + 440, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 175, + 440, + 189 + ], + "spans": [ + { + "bbox": [ + 152, + 175, + 440, + 189 + ], + "type": "text", + "content": "Jiaxin Chen3 Weihan Shen3 Xiaolong Zhu3 Xiu Li1†" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 190, + 473, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 190, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 118, + 190, + 473, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 190, + 473, + 203 + ], + "type": "text", + "content": " Tsinghua Shenzhen International Graduate School, Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "spans": [ + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 203, + 520, + 217 + ], + "type": "text", + "content": " Parametrix Technology Company Ltd." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 219, + 477, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 219, + 477, + 231 + ], + "spans": [ + { + "bbox": [ + 112, + 219, + 477, + 231 + ], + "type": "text", + "content": "{yk22,tj22,lvjf20}@mails.tsinghua.edu.cn li.xiu@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 284, + 289, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 289, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 289, + 631 + ], + "type": "text", + "content": "Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "Recent advances in image generation models have yielded unprecedented success in producing high-quality images" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 259, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 545, + 380 + ], + "type": "text", + "content": "from textual prompts [30, 44, 49]. Diverse approaches, including Generative Adversarial Networks (GANs) [23], autoregressive models [14, 15, 19, 22, 44, 58], Normalizing Flows [16, 46], and diffusion-based techniques [40, 45, 47, 49], have rapidly pushed forward the capabilities of these systems. With the proper textual inputs, such models are now adept at crafting images that are not only visually compelling but also semantically coherent, garnering widespread interest for their potential applications and implications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 546, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 672 + ], + "type": "text", + "content": "To adapt the aforementioned models for specific downstream tasks, such as the generation of more visually appealing and aesthetic images, Reinforcement Learning from Human Feedback (RLHF) is commonly employed [12]. This technique has been successfully used to refine large language models such as GPT [8, 41]. Now, efforts are being made to extend this method to diffusion models to enhance their performance. One such approach, the DDPO method [5], aims to enhance image complexity, aesthetic quality, and the alignment between prompts and images. The ReLF approach [60] introduces a novel reward model, named ImageReward, which is specifically trained to discern human aesthetic preferences in text-to-image synthesis. This model is then utilized to fine-tune diffusion models to produce images that align more closely with human preferences. Nonetheless, developing a robust reward model for various tasks can be challenging, often necessitating a vast collection of images and abundant training resources. For example, to diminish the rate of deformities in character images, one must amass a substantial dataset of deformed and non-deformed images generated from identical prompts. Subsequently, a network is constructed to discern and learn the human preference for non-deformed imagery, serving as the reward model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": "In the field of natural language processing, Direct Preference Optimization (DPO) has been proposed to reduce training costs [43]. This method forgoes the training of a" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8941" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 73, + 519, + 300 + ], + "blocks": [ + { + "bbox": [ + 75, + 73, + 519, + 300 + ], + "lines": [ + { + "bbox": [ + 75, + 73, + 519, + 300 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 519, + 300 + ], + "type": "image", + "image_path": "ef3b85de653712224e547e7d86884b6bac70ef933b10322e1f6a5616f8fb4c4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 354 + ], + "type": "text", + "content": "Figure 1. Overview of D3PO. The diffusion model generates two corresponding images based on the provided prompts. Guided by specific task requirements—such as improving prompt-image alignment, enhancing image incompressibility, or refining aesthetic quality—human evaluators select the preferred image. Leveraging this human feedback, our method directly updates the diffusion model's parameters without necessitating the training of a reward model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": "reward model and directly fine-tunes language models according to human preferences. However, this straightforward and easy-to-train method encounters challenges when applied to fine-tune diffusion models. During the DPO training process, the complete sentence generated by the language model is treated as a single output, necessitating the storage of gradients from multiple forward passes. With diffusion models, one must store the gradients across multiple latent image representations, which are significantly larger than word embeddings, leading to memory consumption that is typically unsustainable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 662 + ], + "type": "text", + "content": "To address the issue of high computational overhead and enable the use of the DPO method to fine-tune diffusion models directly with human feedback, we conceptualize the denoising process as a multi-step MDP, which utilizes a pre-trained model to represent an action value function " + }, + { + "bbox": [ + 46, + 508, + 288, + 662 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 508, + 288, + 662 + ], + "type": "text", + "content": " that is commonly estimated in RL. We extend the theoretical framework of DPO into the formulated MDP, which allows for direct parameter updates at each step of the denoising process based on human feedback, thereby circumventing the significant computational costs and eliminating the need for a reward model. To the best of our knowledge, this is the first work that fine-tune diffusion models without reward models. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "- We introduce an innovative approach for fine-tuning diffusion models that could significantly modify the current RLHF framework for fine-tuning diffusion models. This method bypasses resource-intensive reward model train" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 374, + 544, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 544, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 544, + 397 + ], + "type": "text", + "content": "ing by utilizing direct human feedback, making the process more efficient and cost-effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 399, + 545, + 542 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 399, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 482 + ], + "type": "text", + "content": "- We expand the theoretical framework of DPO into a multi-step MDP, demonstrating that directly updating the policy based on human preferences within an MDP is equivalent to learning the optimal reward model first and then using it to guide policy updates. This establishes a robust theoretical foundation and provides assurance for our proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 483, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 545, + 542 + ], + "type": "text", + "content": "- In our experiments, we have demonstrated the effectiveness of our method by using human feedback to successfully address issues of hand and full-body deformities, enhance the safety of generated images, and improve prompt-image alignment." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 392, + 571 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 579, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 688 + ], + "type": "text", + "content": "Diffusion models. Denoising diffusion probabilistic models, introduced in [54] and further advanced by [26], have emerged as powerful tools for generating diverse data types. They have been successfully applied in various domains such as image synthesis [44, 49], video generation [27, 36, 53], and robotics control systems [1, 11, 29]. Notably, test-to-image diffusion models have enabled the creation of highly realistic images from textual descriptions [44, 49], opening new avenues in digital art and design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": "Recent studies have focused on refining the guidance of diffusion models for more precise manipulation over the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8942" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "generative process. Techniques such as adapters [63] and compositional approaches [17, 38] have been introduced to incorporate additional input constraints and blend multiple models, respectively, enhancing image quality and generation control. The implementation of classifier-based [13] and classifier-free guidance [25] has also contributed significantly to achieving greater autonomy in the generation process, resulting in outputs that closely align with user intentions. In our work, we utilize Stable Diffusion [45] to generate images based on specific prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 495 + ], + "type": "text", + "content": "RLHF. RLHF stands as a salient strategy in the domain of machine learning when objectives are complex or difficult to define explicitly. This technique has been instrumental across various applications, from gaming, as demonstrated with Atari [4, 12], to more intricate tasks in robotics [10, 64]. The integration of RLHF into the development of large language models (LLMs) has marked a significant milestone in the field, with notable models like OpenAI's GPT-4 [41], Anthropic's Claude [2], Google's Bard [24], and Meta's Llama 2-Chat [57] leveraging this approach to enhance their performance and relevance. The effectiveness of RLHF in refining the behavior of LLMs to be more aligned with human values, such as helpfulness and harmlessness, has been extensively studied [4, 64]. The technique has also proven beneficial in more focused tasks, such as summarization, where models are trained to distill extensive information into concise representations [55]. Some recent research utilizes Reinforcement Learning from AI Feedback (RLAIF) [32, 62] as an alternative to RLHF for model fine-tuning. RLAIF offers convenience and efficiency by replacing human feedback with AI-generated feedback. However, for tasks like assessing hand generation normality or image aesthetic appeal, reliable judgment models are currently lacking. Hence, this paper still relies on human feedback for evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Fine-tune Diffusion Models with RL. Before applying diffusion models, data generation has been regarded as a sequential decision-making problem and combined with reinforcement learning [3]. More recently, the SFT method [20] applied reinforcement learning to diffusion models to enhance existing fast DDPM samplers [26]. Reward Weighted method [34] explored using human feedback to align text-to-image models. It uses the reward model for the coefficients of the loss function instead of using reinforcement learning optimization objectives. ReFL [60] employs the framework of RLHF. It begins by training a model called ImageReward based on human preferences and then finetunes the diffusion model using reinforcement learning. DDPO [5] treats the denoising process of diffusion models as a MDP to fine-tune diffusion models with many reward models. DPOK [21] combine the KL divergence into the DDPO loss and use it to better align text-to-image objectives. All these models need a robust reward model, which" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "demands a substantial dataset of images and extensive human evaluations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 96, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 277 + ], + "type": "text", + "content": "Direct Preference Optimization. In the realm of reinforcement learning, exploring policies derived from preferences rather than explicit rewards has gained attention through various methods. The Contextual Dueling Bandit (CDB) framework [18, 61] introduces the concept of a von Neumann winner, shifting away from the pursuit of an optimal policy directly based on rewards. Preference-based Reinforcement Learning (PbRL) [9, 39, 48] learns from binary preferences inferred from a cryptic scoring function instead of explicit rewards. Recently, the DPO approach [43] was proposed, which fine-tunes LLMs directly using preferences. DPO leverages the correlation between reward functions and optimal policies, effectively addressing the challenge of constrained reward maximization in a single phase of policy training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 285, + 391, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 391, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 391, + 298 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": "MDP. We consider the MDP formulation described in [56]. In this setting, an agent perceives a state " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": " and executes an action, where " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "S, A" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": " denote state space and action space, respectively. The transition probability function, denoted by " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "P(s'|s, a)" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": ", governs the progression from the current state " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": " to the subsequent state " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": " upon the agent's action " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": ". Concurrently, the agent is awarded a scalar reward " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": ", determined by the reward function " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "r: S \\times A \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": ". The agent's objective is to ascertain a policy " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\pi(a|s)" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": " that maximizes the cumulative returns of trajectories " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\tau = (s_0, a_0, s_1, a_1, \\dots, s_{T-1}, a_{T-1})" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": ", which can be represented as: " + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\pi) = \\mathbb{E}_{\\tau}[\\sum_{t=0}^{T-1} r(s_t, a_t)]" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": "Diffusion models. Diffusion models learn to model a probability distribution " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " by inverting a Markovian forward process " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t|\\pmb{x}_{t-1})" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " which adds noise to the data. The denoising process is modeled by a neural network to predict the mean of " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " or the noise " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\epsilon_{t-1}" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " of the forward process. In our work, we use network " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mu_\\theta(\\pmb{x}_t; t)" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " to predict the mean of " + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 304, + 437, + 545, + 536 + ], + "type": "text", + "content": " instead of predicting the noise. Using the mean squared error (MSE) as a measure, the objective of this network can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 540, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 540, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 540, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {t \\sim [ 1, T ], \\boldsymbol {x} _ {0} \\sim p (\\boldsymbol {x} _ {0}), \\boldsymbol {x} _ {t} \\sim q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {0})} \\left[ \\| \\tilde {\\boldsymbol {\\mu}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) - \\boldsymbol {\\mu} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| ^ {2} \\right], \\tag {1}", + "image_path": "ccfacb46609417fa872b99668840aef1048821b11ede7986dfacb217d1162e93.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_{\\theta}(\\boldsymbol{x}_t,\\boldsymbol{x}_0)" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": " is the forward process posterior mean. In the case of conditional generative modeling, the diffusion models learn to model " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "p(x|c)" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": " is the conditioning information, i.e., image category and image caption. This is done by adding an additional input " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": " to the denoising neural network, as in " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\boldsymbol{x}_t,t;\\boldsymbol{c})" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": ". To generate a sample from the learned distribution " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x|c)" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": ", we start by drawing a sample " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": " and then progressively denoise the sample by iterated application of " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": " according to a specific sampler [26]. Given the noise-related parameter " + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 304, + 564, + 547, + 696 + ], + "type": "text", + "content": ", the reverse process can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {2}", + "image_path": "8f985ba612a10b3f41e263139d6c47ea330cd92de2101afaaf0be14ebc9ddfa5.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8943" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": "Reward learning for preferences. The basic framework to model preferences is to learn a reward function " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "r^{*}(s,a)" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " from human feedback [12, 33, 59]. The segment " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\sigma = \\{s_k,a_k,s_{k + 1},a_{k + 1},\\dots ,s_m,a_m\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " is a sequence of observations and actions. By following the Bradley-Terry model [6], the human preference distribution " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " by using the reward function can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 60, + 163, + 287, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 60, + 163, + 287, + 196 + ], + "type": "interline_equation", + "content": "p ^ {*} (\\sigma_ {1} \\succ \\sigma_ {0}) = \\frac {\\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {1} , a _ {t} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(\\sum_ {t = k} ^ {T} r ^ {*} \\left(s _ {t} ^ {i} , a _ {t} ^ {i}\\right)\\right)}, \\tag {3}", + "image_path": "66b5ad7337325dc6e85cfb0cdd61cf925309f6f106c37d453a0af7eb68fc73ed.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\sigma_{i} \\succ \\sigma_{j}" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": " denotes that segment " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": " is preferable to segment " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": ". Now we have the preference distribution of human feedback, and we want to use a network " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "r^{*}" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": ". Given the human preference " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "y \\in \\{(1,0), (0,1)\\}" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": " which is recorded in dataset " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": " as a triple " + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "inline_equation", + "content": "(\\sigma_{0}, \\sigma_{1}, y)" + }, + { + "bbox": [ + 47, + 201, + 289, + 286 + ], + "type": "text", + "content": ", framing the problem as a binary classification, the reward function modeled as a network is updated by minimizing:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 292, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 287, + 321 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\phi) = - \\mathbb {E} _ {\\left(\\sigma_ {1}, \\sigma_ {0}, y\\right) \\sim \\mathcal {D}} [ y (0) \\log p _ {\\phi} \\left(\\sigma_ {0} \\succ \\sigma_ {1}\\right) \\\\ \\left. + y (1) \\log p _ {\\phi} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "4818ea64930a0b8ae76626d30963b80424b821cf8424900acfeeb08e45fed8ff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 327, + 103, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 103, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 103, + 340 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "content": "In this section, we describe a method to directly fine-tune diffusion models using human feedback, bypassing the conventional requirement for a reward model. Initially, we reinterpret the denoising process inherent in diffusion models as a multi-step MDP. Then we extend the theory of DPO to MDP, which allows us to apply the principles to effectively translate human preferences into policy improvements in diffusion models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 450, + 253, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 253, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 253, + 464 + ], + "type": "text", + "content": "4.1. Denoising process as a multi-step MDP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 468, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 535 + ], + "type": "text", + "content": "We conceptualize the denoising process within the diffusion model as a multi-step MDP, which varies slightly from the approach outlined in [5]. To enhance clarity, we have redefined the states, transition probabilities, and policy functions. The correspondence between notations in the diffusion model and the MDP is established as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 540, + 277, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 277, + 600 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 277, + 600 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}) \\quad P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\boldsymbol {c}}, \\delta_ {t + 1}, \\delta_ {\\boldsymbol {x} _ {T - 1 - t}}\\right) \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - 1 - t} \\quad \\pi (\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}) \\triangleq p _ {\\theta} (\\mathbf {x} _ {T - 1 - t} \\mid \\mathbf {c}, t, \\mathbf {x} _ {T - t}) \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\boldsymbol {c}), \\delta_ {0}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ r \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq r \\left(\\left(\\boldsymbol {c}, t, \\boldsymbol {x} _ {T - t}\\right), \\boldsymbol {x} _ {T - t - 1}\\right) \\\\ \\end{array}", + "image_path": "ace515d0dc955343c8ac5d80cef214a4a8351aee1c6ff6343ff50ed5fe3c6908.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{x}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " represents the Dirac delta distribution, and " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " denotes the maximize denoising timesteps. Leveraging this mapping, we can employ RL techniques to fine-tune diffusion models by maximizing returns. However, this approach requires a proficient reward model capable of adequately rewarding the noisy images. The task becomes exceptionally challenging, particularly when " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is low, and " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{T - t}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " closely resembles Gaussian noise, even for an experienced expert." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "text", + "content": "4.2. Direct Preference Optimization for MDP" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": "The DPO method does not train a separate reward model but instead directly optimizes the LLMs with the preference data. Given a prompt " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " and a pair of answers " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "(y_1, y_0) \\sim \\pi_{\\mathrm{ref}}(y|x)" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " represents the reference (pre-trained) model, these responses are ranked and stored in " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " as a tuple " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "(x, y_w, y_l)" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "y_w" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " denotes the preferred answer and " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "y_l" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " indicates the inferior one. DPO optimizes " + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\pi_\\theta" + }, + { + "bbox": [ + 304, + 90, + 547, + 187 + ], + "type": "text", + "content": " with the human preference dataset by using the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 205, + 547, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 547, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\rho \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {w} | x)}{\\pi_ {\\text {r e f}} (y _ {w} | x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {l} | x)}{\\pi_ {\\text {r e f}} (y _ {l} | x)}\\right) \\right]. \\tag {5}", + "image_path": "84d6c07400f8fd466a3fd11d5089839c710044e2d627b1b0df460efc143e1b2d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": " is the logistic function, and " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": " is the parameter controlling the deviation from the " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": ". In our framework, we treat segments " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\sigma^1, \\sigma^0" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "y_1, y_0" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": " and use DPO to fine-tune diffusion models. However, directly using this method faces difficulties since the segments contain a large number (usually 20-50) of the image latent, which occupy a large amount of GPU memory (each image is about 6G even when using LoRA [28]). Since we can only get human preferences for the final image " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": ", if we want to update " + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\sigma) = \\prod_{t=k}^{T} \\pi_{\\theta}(s_t, a_t)" + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": ", it will consume more than 100G GPU memory, which makes the fine-tuning process nearly impossible." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 388, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 460 + ], + "type": "text", + "content": "To address this problem, we extend the DPO theory to MDP. Firstly, we need to reconsider the objective of the RL method. For the MDP problem, the agents take action by considering maximizing the expected return instead of the current reward. For actor-critic methods such as DDPG [37], the optimization objective of policy " + }, + { + "bbox": [ + 304, + 388, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 388, + 545, + 460 + ], + "type": "text", + "content": " gives:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 468, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 468, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 363, + 468, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} [ Q ^ {*} (s, a) ]. \\tag {6}", + "image_path": "07a08130ecc7f2617d165967ad9beb5d7e1f4d51ecfbf85bad9af97be9b87b39.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "d^{\\pi} = (1 - \\gamma)\\sum_{t = 0}^{\\infty}\\gamma^{t}P_{t}^{\\pi}(s)" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": " represents the state visitation distribution, where " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "P_{t}^{\\pi}(s)" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": " denotes the probability of being in state " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": " given policy " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "inline_equation", + "content": "Q^{*}(s,a)" + }, + { + "bbox": [ + 304, + 494, + 547, + 555 + ], + "type": "text", + "content": " denotes the optimal action-value function. The optimal policy can be written as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 335, + 563, + 545, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 563, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 335, + 563, + 545, + 596 + ], + "type": "interline_equation", + "content": "\\pi^ {*} (a | s) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} a = \\arg \\max _ {\\hat {a}} Q ^ {*} (s, \\hat {a}), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}", + "image_path": "8615a4bd77bbe94ebbdfb46bd33da6e74cb1c9bc692fe1999839e12002007ca6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 603, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 659 + ], + "type": "text", + "content": "Similar to some popular methods, we use KL-divergence to prevent the fine-tuned policy from deviating from the reference policy, hence relieving the out-of-distribution (OOD) issue. By incorporating this constraint, the RL objective can be rewritten as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 666, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 666, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 310, + 666, + 545, + 683 + ], + "type": "interline_equation", + "content": "\\max _ {\\pi} \\mathbb {E} _ {s \\sim d ^ {\\pi}, a \\sim \\pi (\\cdot | s)} \\left[ Q ^ {*} (s, a) \\right] - \\beta \\mathbb {D} _ {K L} \\left[ \\pi (a | s) \\| \\pi_ {\\text {r e f}} (a | s) \\right]. \\tag {8}", + "image_path": "a9c2c32d2a60137ff13fd19c8b64d680d3fe6e081cfad356f14a5d4049477bc3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the temperature parameter that controls the deviation of " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(a|s)" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}(a|s)" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8944" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 95, + 161, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 95, + 161, + 110 + ], + "spans": [ + { + "bbox": [ + 75, + 95, + 161, + 110 + ], + "type": "text", + "content": "Compressibility" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 148, + 164, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 148, + 164, + 163 + ], + "spans": [ + { + "bbox": [ + 74, + 148, + 164, + 163 + ], + "type": "text", + "content": "Incompressibility" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 205, + 167, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 205, + 167, + 221 + ], + "spans": [ + { + "bbox": [ + 74, + 205, + 167, + 221 + ], + "type": "text", + "content": "Aesthetic Quality" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 72, + 521, + 290 + ], + "blocks": [ + { + "bbox": [ + 177, + 72, + 521, + 290 + ], + "lines": [ + { + "bbox": [ + 177, + 72, + 521, + 290 + ], + "spans": [ + { + "bbox": [ + 177, + 72, + 521, + 290 + ], + "type": "image", + "image_path": "e2f8ad454545b73bf2416007ff54ab2e1a1912a35c6b5527e66962f40da21a8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "type": "text", + "content": "Figure 2. Progression of samples targeting compressibility, incompressibility, and aesthetic quality objectives. With the respective focus during training, images exhibit reduced detail and simpler backgrounds for compressibility, richer texture details for incompressibility, and an overall increased aesthetic appeal when prioritizing aesthetic quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 288, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 379 + ], + "type": "text", + "content": "Proposition 1 Given the objective of Eq. (8), the optimal policy " + }, + { + "bbox": [ + 46, + 354, + 288, + 379 + ], + "type": "inline_equation", + "content": "\\pi^{*}(a|s)" + }, + { + "bbox": [ + 46, + 354, + 288, + 379 + ], + "type": "text", + "content": " has the following expression:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 384, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 384, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 89, + 384, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\pi^ {*} (a | s) = \\pi_ {\\text {r e f}} (a | s) \\exp \\left(\\frac {1}{\\beta} Q ^ {*} (s, a)\\right). \\tag {9}", + "image_path": "65f2f953580f326d0faebe2f883fb4d0a08d0619dd800b51001799e1f922e2d2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "type": "text", + "content": "The proof can be seen in Appendix B.1. By rearranging the formula of Eq. (9), we can obtain that:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 441, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 441, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 109, + 441, + 287, + 466 + ], + "type": "interline_equation", + "content": "Q ^ {*} (s, a) = \\beta \\log \\frac {\\pi^ {*} (a | s)}{\\pi_ {\\mathrm {r e f}} (a | s)}. \\tag {10}", + "image_path": "9fc70e5cebf7a5d08b61524405c9910713ba5f224db9b0cd4ae0d8b32ec97cdf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "content": "Now, considering Eq. (3) and noticing that " + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "inline_equation", + "content": "Q^{*}(s_{t},a_{t}) = \\mathbb{E}\\left[\\sum_{t = k}^{T}r^{*}(s_{t},a_{t})\\right]" + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "content": " under the policy " + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\pi^* (a|s)" + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "content": ", we make a substitution. By replacing " + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\sum_{t = k}^{T}r^{*}(s_{t},a_{t})" + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "inline_equation", + "content": "Q^{*}(s_{t},a_{t})" + }, + { + "bbox": [ + 47, + 471, + 287, + 526 + ], + "type": "text", + "content": " we define a new distribution that can be rewritten as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 529, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 529, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 67, + 529, + 287, + 559 + ], + "type": "interline_equation", + "content": "\\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) = \\frac {\\exp \\left(Q ^ {*} \\left(s _ {k} ^ {1} , a _ {k} ^ {1}\\right)\\right)}{\\sum_ {i \\in \\{0 , 1 \\}} \\exp \\left(Q ^ {*} \\left(s _ {k} ^ {i} , a _ {k} ^ {i}\\right)\\right)}. \\tag {11}", + "image_path": "9a69615f24d314c7e8e66823476c48e5eb55016e8b1e569b129078a62aefbb37.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": "We suppose " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " is sampled from a normal distribution with mean " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\sum_{t = k}^{m}r^{*}(s_{t},a_{t})]" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": ". From a statistical perspective, we can establish the relationship between the new distribution " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\tilde{p}^{*}(\\sigma_{1}\\succ \\sigma_{0})" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": " and the raw distribution " + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "inline_equation", + "content": "p^{*}(\\sigma_{1}\\succ \\sigma_{0})" + }, + { + "bbox": [ + 47, + 563, + 287, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": "Proposition 2 For " + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "inline_equation", + "content": "i \\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": ", suppose the expected return satisfies a normal distribution, i.e., " + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\sum_{t=0}^{T} r^*(s_t^i, a_t^i) \\sim \\mathcal{N}(Q^*(s_0^i, a_0^i), \\sigma^2)" + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "inline_equation", + "content": "Q^*(s,a) \\in [Q_{\\min}, Q_{\\max}]" + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "inline_equation", + "content": "Q_{\\min}" + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "inline_equation", + "content": "Q_{\\max}" + }, + { + "bbox": [ + 47, + 628, + 287, + 689 + ], + "type": "text", + "content": " represent the lower and upper bounds of the values, then" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 693, + 275, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 275, + 716 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 275, + 716 + ], + "type": "interline_equation", + "content": "\\left| p ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) - \\tilde {p} ^ {*} \\left(\\sigma_ {1} \\succ \\sigma_ {0}\\right) \\right| < \\frac {\\left(\\xi^ {2} + 1\\right) \\left(\\exp \\left(\\sigma^ {2}\\right) - 1\\right)}{1 6 \\xi \\delta}", + "image_path": "ceb4f351c7d9062cbdc774fed3aaf711e0fc7038d912379961232f5ea07dee8f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "text", + "content": "with probability at least " + }, + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "inline_equation", + "content": "1 - \\delta" + }, + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "inline_equation", + "content": "\\xi = \\frac{\\exp(Q_{\\mathrm{max}})}{\\exp(Q_{\\mathrm{min}})}" + }, + { + "bbox": [ + 308, + 352, + 523, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": "The proof can be seen in Appendix B.2. In practical applications, as the volume of data increases, it becomes easier to satisfy the assumption of normality. Additionally, we can use clipping operations to constrain the " + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": " values within a certain range, which introduces upper and lower bounds. Therefore, the aforementioned assumption is reasonable. As shown in proposition 2, their deviation can be bounded at the scale of " + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(\\frac{\\xi}{\\delta} (\\exp (\\sigma^2) - 1)\\right)" + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": ". It is clear that this bound can approach 0 if the " + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": " is close to 0. In practice, " + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": " approaches 0 if the " + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 373, + 545, + 512 + ], + "type": "text", + "content": " function can be estimated with a small standard deviation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": "By combining Eq. (11), Eq. (4), and Eq. (10), replacing " + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "inline_equation", + "content": "p^*(\\sigma_1 \\succ \\sigma_0)" + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\tilde{p}^*(\\sigma_1 \\succ \\sigma_0)" + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": ", and substituting " + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\pi^*(s, a)" + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": " with the policy network " + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(s, a)" + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": " that requires learning, we derive the following loss function for updating " + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(a|s)" + }, + { + "bbox": [ + 304, + 513, + 545, + 562 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 565, + 547, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 565, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 565, + 547, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(s _ {k}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {w} \\mid s _ {k} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {k} ^ {l} \\mid s _ {k} ^ {l}\\right)}) ], \\tag {12}", + "image_path": "552fd784273db3c533e3bb08b916802f27bd9e846f000537e9f89721ed9c97fd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\sigma_w = \\{s_k^w, a_k^w, s_{k+1}^w, a_{k+1}^w, \\dots, s_T^w, a_T^w\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "text", + "content": " denotes the segment preferred over another segment " + }, + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\sigma_l = \\{s_k^l, a_k^l, s_{k+1}^l, a_{k+1}^l, \\dots, s_T^l, a_T^l\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 632 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 635, + 545, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 545, + 660 + ], + "type": "text", + "content": "4.3. Direct preference for Denoising Diffusion Policy Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Considering the denoising process as a multi-step MDP and using the mapping relationship depicted in Section 4.1, we can use DPO to directly update diffusion models by using Eq. (12). In the denoising process, we set " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8945" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 218, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 218, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 218, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 218, + 182 + ], + "type": "image", + "image_path": "6ad20a06d7db5be20dee43c8ce0c6c281c4b0fca3f2e0414b22f03637ad7b813.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "type": "text", + "content": "Figure 3. Comparison of D3PO against existing methods. The horizontal axis represents the number of image sample pairs generated for updating parameters. The rewards denote image size for incompressity objective, negative image size for the compressivity objective, and the LAION aesthetic score for the aesthetic objective. All experiments are conducted with 5 different seeds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 70, + 375, + 182 + ], + "blocks": [ + { + "bbox": [ + 218, + 70, + 375, + 182 + ], + "lines": [ + { + "bbox": [ + 218, + 70, + 375, + 182 + ], + "spans": [ + { + "bbox": [ + 218, + 70, + 375, + 182 + ], + "type": "image", + "image_path": "57f5e874c62260282f9da65854f1439ecb81eb556c5754d9caeff260f0ce439f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 534, + 182 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 534, + 182 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 534, + 182 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 534, + 182 + ], + "type": "image", + "image_path": "fdcff313aeee8d751195d901691896162429675553edf98901dccba206f7d114.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": ". We first sample an initial state " + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "s_0^w = s_0^l = s_0" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": " and then use Eq. (2) to generate two segments. After manually choosing which segment is better, the probability of " + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(a_0^w | s_0^w)" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": " is gradually increasing and " + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(a_0^l | s_0^l)" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": " is decreasing, which guides the diffusion model to generate images of human preference. However, the approach of only updating " + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\cdot | s_0)" + }, + { + "bbox": [ + 46, + 243, + 287, + 339 + ], + "type": "text", + "content": " does not fully utilize the information within the segment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "content": "Since the middle states of the segment are noises and semi-finished images, it is hard for humans to judge which segment is better by observing the whole segment. But we can conveniently compare the final image " + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "content": ". Like many RL methods [7, 51, 52] which give rewards by " + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\forall s_t, a_t \\in \\sigma, r(s_t, a_t) = 1" + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "content": " for winning the game and " + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\forall t \\in \\sigma, r(s_t, a_t) = -1" + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "content": " for losing the game, we also assume that if the segment is preferred, then any state-action pair of the segment is better than the other segment. By using this assumption, we construct " + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 340, + 288, + 472 + ], + "type": "text", + "content": " sub-segments for the agent to learn, which can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 479, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 479, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 51, + 479, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} = \\left\\{s _ {i}, a _ {i}, s _ {i + 1}, a _ {i + 1}, \\dots , s _ {T - 1}, a _ {T - 1} \\right\\}, \\quad 0 \\leq i \\leq T - 1 \\tag {13}", + "image_path": "134e28044983248d000dc12bb33fcbad8da0c3f0f4a3edbcb16ded3ed2b58e40.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 508, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 533 + ], + "type": "text", + "content": "Using these sub-segments, the overall loss of the D3PO algorithm gives:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 536, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 536, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 536, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} (\\theta) = - E _ {\\left(s _ {i}, \\sigma_ {w}, \\sigma_ {l}\\right)} [ \\log \\rho (\\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {w} \\mid s _ {i} ^ {w}\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(a _ {i} ^ {l} \\mid s _ {i} ^ {l}\\right)}) ], \\tag {14}", + "image_path": "f10dc78cbaf5c2fcbded8b80a8e5ac71f7f2c6c6330cacac8559cac9b5a0fee2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "inline_equation", + "content": "i \\in [0, T - 1]" + }, + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "text", + "content": ". Compared to Eq. (12), Eq. (14) uses every state-action pair for training, effectively increasing the data utilization of the segment by a factor of " + }, + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 564, + 287, + 600 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 600, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 624 + ], + "type": "text", + "content": "The overview of our method is shown in Fig. 1. The pseudocode of D3PO can be seen in Appendix A." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "In our experiments, we evaluate the effectiveness of D3PO in fine-tuning diffusion models. Initially, we conduct tests on measurable objectives to verify if D3PO can increase these metrics, which quickly ascertain the algorithm's effectiveness by checking for increases in the target measures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "type": "text", + "content": "Next, we apply D3PO to experiments aimed at lowering the rate of deformities in hands and full-body images generated by diffusion models. Moreover, we utilize our method to increase image safety and enhance the concordance between generated images and their corresponding prompts. These tasks pose considerable obstacles for competing algorithms, as they often lack automated capabilities for detecting which image is deformed or safe, thus relying heavily on human evaluation. We use Stable Diffusion v1.5 [47] to generate images in most of the experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 372, + 533, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 533, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 533, + 384 + ], + "type": "text", + "content": "5.1. Pre-defined Quantifiable Objectives Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": "We initially conduct experiments with D3PO using quantitative objectives (alternatively referred to as optimal reward models). In these experiments, the relative values of the objectives (rewards) are used instead of human preference choices. Preferences are established based on these objectives, meaning " + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": " is preferred if its objective surpasses that of " + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": ". After training, we validate the effectiveness of our approach by measuring the growth of metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 487, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 617 + ], + "type": "text", + "content": "In our experimental setup, we benchmark against several popular methods: DDPO [5], DPOK [21], and Reward Weighted [34], all of which necessitate a reward model. Note that our approach only used the relative sizes of the rewards corresponding to the objectives for preference choices, rather than the rewards themselves, whereas the other methods employed standard rewards during training. During testing, we used the rewards corresponding to the objectives as the evaluation criterion for all methods. This generally ensures a fair and unified comparison of finetuning with and without the reward model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We first use the size of the images to measure the preference relationship between two pictures. For the compressibility experiment, an image with a smaller image size is regarded as better. Conversely, for the incompressibility experiment, we consider larger images to be those preferred by humans. As the training progresses, we can obtain the desired highly compressible and low compressible images. We then utilize the LAION aesthetics predictor [50] to pre" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8946" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "dict the aesthetic rating of images. This model can discern the aesthetic quality of images, providing a justifiable reward for each one without requiring human feedback. The model can generate more aesthetic images after fine-tuning. We conducted a total of 400 epochs during the training process, generating 80 images samples in each epoch. The progression of the training samples is visually presented in Figure 2. More quantitative samples are shown in Figure 8. The testing curves of D3PO and other methods are shown in Figure 3. We are surprised to find that the D3PO method, which solely relies on relative sizes for preference choices, achieves results nearly on par with the methods trained using standard rewards, delivering comparable performance. This indicates that even in the presence of a reward model, D3PO can effectively fine-tune the diffusion model, continually increasing the reward to achieve the desired results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 274, + 263, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 263, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 263, + 286 + ], + "type": "text", + "content": "5.2. Experiments without Any Reward Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 288, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 377 + ], + "type": "text", + "content": "We conduct experiments for some objectives without any reward model. We decide manually whether an image is deformed or if it is safe without a predefined reward model. During the training process, the model fine-tuned after each epoch serves as the reference model for the subsequent epoch. For each prompt, we generate 7 images with the same initial state " + }, + { + "bbox": [ + 46, + 293, + 288, + 377 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 293, + 288, + 377 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 396, + 187, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 187, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 187, + 407 + ], + "type": "text", + "content": "5.2.1 Reduce Image Distortion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 416, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 582 + ], + "type": "text", + "content": "We use the prompt \"1 hand\" to generate images, manually selecting those that are deformed. Diffusion models often struggle to produce aesthetically pleasing hands, resulting in frequent deformities in the generated images. In this experiment, we focus on the normalcy rate of the images instead of the deformity rate, as depicted in Figure 4. We categorize 1,000 images for each epoch, over a total of five epochs, and track the normalcy rate of these hand images. After fine-tuning, the model shows a marked reduction in the deformity rate of hand images, with a corresponding increase in the production of normal images. Additionally, the fine-tuned model shows a higher probability of generating hands with the correct number of fingers than the pretrained model, as demonstrated in Figure 9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 584, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 288, + 693 + ], + "type": "text", + "content": "To assess the generality of our method, we generated images with the Anything v5 model " + }, + { + "bbox": [ + 46, + 584, + 288, + 693 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 46, + 584, + 288, + 693 + ], + "type": "text", + "content": " , renowned for creating anime character images. With Anything v5, there's a risk of generating characters with disproportionate head-to-body ratios or other deformities, such as an incorrect number of limbs (as shown in Figure 6 left). We categorize such outputs as deformed. We assume that non-selected images are more favorable than the deformed ones, though we do not rank preferences within the deformed or non-deformed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "sets. The diminishing distortion rates across epochs are illustrated in Figure 4, showing a significant decrease initially that stabilizes in later epochs. The visual examples are provided in Figure 6." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 129, + 535, + 255 + ], + "blocks": [ + { + "bbox": [ + 318, + 129, + 535, + 255 + ], + "lines": [ + { + "bbox": [ + 318, + 129, + 535, + 255 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 535, + 255 + ], + "type": "image", + "image_path": "15638d63e4b751f1e53b0d42ab5d1f5d75d5436dce644bbb98cb668c2598ee5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 263, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 319 + ], + "type": "text", + "content": "Figure 4. Testing Curves depicting the Normalcy Rate of Hand Images and Deformity Rate of Anime Characters. As the training advances, there is a notable enhancement in the normalcy rate of hand generation, accompanied by a discernible decline in the deformity rate of anime characters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 337, + 434, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 434, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 434, + 350 + ], + "type": "text", + "content": "5.2.2 Enhance Image Safety" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 356, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 524 + ], + "type": "text", + "content": "In this experiment, we utilized unsafe prompts to generate images using a diffusion model. These prompts contained edgy terms that could lead to the creation of both normal and Not Safe for Work (NSFW) images, examples being 'ambiguous beauty' and 'provocative art'. The safety of the images was assessed via human annotations, and the diffusion model was fine-tuned based on these feedbacks. Across 10 epochs, we generated 1,000 images per epoch. Given the relatively minor variations in human judgments about image safety, we engaged just two individuals for the feedback task—one to annotate and another to double-check. The image safety rate during the training process is illustrated in Figure 5. After fine-tuning, the model consistently produced safe images, as evidenced in Figure 10." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 335, + 533, + 517, + 643 + ], + "blocks": [ + { + "bbox": [ + 335, + 533, + 517, + 643 + ], + "lines": [ + { + "bbox": [ + 335, + 533, + 517, + 643 + ], + "spans": [ + { + "bbox": [ + 335, + 533, + 517, + 643 + ], + "type": "image", + "image_path": "6ce5b477a552d5771aa5ead7de7b068fb32209b1d77b2032c0f48fe197f60f4c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 652, + 525, + 663 + ], + "lines": [ + { + "bbox": [ + 326, + 652, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 326, + 652, + 525, + 663 + ], + "type": "text", + "content": "Figure 5. Safety rate curves of the training procession." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "type": "text", + "content": "5.2.3 Prompt-Image Alignment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "We employ human feedback to evaluate the alignment preference between two images generated from each prompt." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 235, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 235, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 235, + 714 + ], + "type": "text", + "content": "1https://huggingface.co/stablediffusionapi/anything-v5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8947" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 254, + 250 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 254, + 250 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 254, + 250 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 254, + 250 + ], + "type": "image", + "image_path": "6c4009da913d0a9a7446270fa01b84a9fb70316a54549d285dcd10050220ac30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "type": "text", + "content": "Figure 6. Image samples of the pre-trained model and the fine-tuned model. The images on the left side of the arrows are distorted images (such as having 3 legs in the image) generated by the pre-trained model, while the images on the right side are normal images generated after fine-tuning the model. Both sets of images used the same initial Gaussian noise, prompts, and seeds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 522, + 249 + ], + "blocks": [ + { + "bbox": [ + 268, + 143, + 325, + 172 + ], + "lines": [ + { + "bbox": [ + 268, + 143, + 325, + 172 + ], + "spans": [ + { + "bbox": [ + 268, + 143, + 325, + 172 + ], + "type": "text", + "content": "finetuning" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 70, + 522, + 249 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 522, + 249 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 522, + 249 + ], + "type": "image", + "image_path": "b95f3a01561f2c5011769c87219c3eb4c077a1be33496906b4f9d8281381885d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 309, + 279, + 432 + ], + "blocks": [ + { + "bbox": [ + 56, + 309, + 279, + 432 + ], + "lines": [ + { + "bbox": [ + 56, + 309, + 279, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 279, + 432 + ], + "type": "image", + "image_path": "f0a8bc25c6dbdcb81331fe34180c03e75e1e399413d5f3703b07355dbf23631d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 439, + 289, + 538 + ], + "lines": [ + { + "bbox": [ + 46, + 439, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 289, + 538 + ], + "type": "text", + "content": "Figure 7. Comparative evaluation of 300 text prompts: Our study involved generating images from the pre-trained diffusion model, the fine-tuned models using other methods, and our fine-tuned model—using the same text prompts. For each prompt, human evaluators are tasked with determining which image is better aligned with the text. Each image was assessed by 5 human raters, and we report the percentage of images that received favorable evaluations. We also highlight the percentage with more than half vote in the black box." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 616, + 286, + 656 + ], + "blocks": [ + { + "bbox": [ + 46, + 552, + 288, + 608 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 608 + ], + "type": "text", + "content": "Table 1. Evaluations of prompt-image alignment. We compare D3PO with those that (a) do not fine-tune, (b) use the preferred images to fine-tune, and (c) use Reward Weighted method to fine-tune. D3PO achieve the best performance. D3PO demonstrated superior performance across all comparisons." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 616, + 286, + 656 + ], + "lines": [ + { + "bbox": [ + 48, + 616, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 616, + 286, + 656 + ], + "type": "table", + "html": "
MethodsCLIP score (↑)BLIP score (↑)ImageReward score (↑)Human preference (↑)
No fine-tune30.71.950.0414.7%
Preferred images31.01.970.0811%
Reward Weighted31.52.010.1718.7%
D3PO31.92.060.2755.7%
", + "image_path": "1e1ec9b08c355219ec6880d2aae27646e3a77cd88ac4b6fceeb60532b88d7699.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "For each epoch, we use 4,000 prompts, generate two images per prompt, and assess preferences with feedback from 16 different evaluators. The training spans 10 epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 311, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 396 + ], + "type": "text", + "content": "The preference comparisons between images from the pretrained and fine-tuned models are conducted by an additional 5 evaluators, with the comparative preferences depicted in Figure 7. We also execute quantitative evaluations of models using metrics that measure the congruence between prompts and images, including CLIP [42], BLIP [35], and ImageReward [60], as presented in Table 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 407, + 379, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 379, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 379, + 420 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 428, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 547, + 657 + ], + "type": "text", + "content": "In this paper, we propose a direct preference denoising diffusion policy optimization method, named D3PO, to fine-tune diffusion models purely from human feedback without learning a separate reward model. D3PO views the denoising process as a multi-step MDP, making it possible to utilize the DPO-style optimization formula by formulating the action-value function " + }, + { + "bbox": [ + 304, + 428, + 547, + 657 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 428, + 547, + 657 + ], + "type": "text", + "content": " with the reference model and the fine-tuned model. D3PO updates parameters at each step of denoising and consumes much fewer GPU memory overheads than directly applying the DPO algorithm. The empirical experiments illustrate that our method achieves competitive or even better performance compared with a diffusion model fine-tuned with a reward model that is trained with a large amount of images and human preferences in terms of image compressibility, image compressibility and aesthetic quality. We further show that D3PO can also benefit challenging tasks such as reducing image distortion rates, enhancing the safety of the generated images, and aligning prompts and images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 669, + 409, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 409, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 409, + 682 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the STI 2030-Major Projects under Grant 2021ZD0201404." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8948" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 212, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 212, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 212, + 158 + ], + "type": "text", + "content": "[2] Anthropic. Introducing claude, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "type": "text", + "content": "[3] Philip Bachman and Doina Precup. Data generation as sequential decision making. Advances in Neural Information Processing Systems, 28, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "type": "text", + "content": "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "text", + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 328 + ], + "type": "text", + "content": "[6] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 352 + ], + "type": "text", + "content": "[7] Noam Brown and Tuomas Sandholm. Superhuman ai for multiplayer poker. Science, 365(6456):885-890, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 408 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 464 + ], + "type": "text", + "content": "[9] Robert Busa-Fekete, Balázs Szörényi, Paul Weng, Weiwei Cheng, and Eyke Hüllermeier. Preference-based reinforcement learning: evolutionary direct policy search using a preference-based racing algorithm. Machine learning, 97: 327-351, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 531 + ], + "type": "text", + "content": "[10] Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[11] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[12] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 1, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890-16902, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "text", + "content": "[16] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "type": "text", + "content": "[17] Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International Conference on Machine Learning, pages 8489-8510. PMLR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "text", + "content": "[18] Miroslav Dudík, Katja Hofmann, Robert E Schapire, Aleksandris Slivkins, and Masrour Zoghi. Contextual dueling bandits. In Conference on Learning Theory, pages 563-587. PMLR, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 336 + ], + "type": "text", + "content": "[20] Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[21] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. arXiv preprint arXiv:2305.16381, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[22] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 493, + 410, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 410, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 410, + 503 + ], + "type": "text", + "content": "[24] Google. Bard, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[25] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 570 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[28] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[29] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8949" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 215 + ], + "type": "text", + "content": "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[33] Kimin Lee, Laura Smith, and Pieter Abbeel. Pebble: Feedback-efficient interactive reinforcement learning via rel-labeling experience and unsupervised pre-training. arXiv preprint arXiv:2106.05091, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[34] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 369 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "text", + "content": "[36] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[37] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[38] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[39] Runze Liu, Yali Du, Fengshuo Bai, Jiafei Lyu, and Xiu Li. Zero-shot preference learning for offline rl via optimal transport. arXiv preprint arXiv:2306.03615, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 225, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 225, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 225, + 624 + ], + "type": "text", + "content": "[41] OpenAI. Gpt-4 technical report, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "preprint arXiv:2305.18290, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "type": "text", + "content": "[46] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 286 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 289, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 545, + 342 + ], + "type": "text", + "content": "[48] Aadirupa Saha, Aldo Pacchiano, and Jonathan Lee. Dueling rl: Reinforcement learning with trajectory preferences. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 6263-6289. PMLR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 410 + ], + "type": "text", + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 412, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 412, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 412, + 545, + 477 + ], + "type": "text", + "content": "[50] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 543 + ], + "type": "text", + "content": "[51] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "text", + "content": "[52] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550(7676): 354-359, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "text", + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8950" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 543 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[55] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 286, + 150 + ], + "type": "text", + "content": "[56] Richard S Sutton, Andrew G Barto, et al. Introduction to reinforcement learning. 1998. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "text", + "content": "[57] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 286, + 251 + ], + "type": "text", + "content": "[58] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 286, + 296 + ], + "type": "text", + "content": "[59] Aaron Wilson, Alan Fern, and Prasad Tadepalli. A bayesian approach for policy learning from trajectory preference queries. Advances in neural information processing systems, 25, 2012. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 286, + 351 + ], + "type": "text", + "content": "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.1,3,8,14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 286, + 397 + ], + "type": "text", + "content": "[61] Yisong Yue, Josef Broder, Robert Kleinberg, and Thorsten Joachims. The k-armed dueling bandits problem. Journal of Computer and System Sciences, 78(5):1538-1556, 2012. JCSS Special Issue: Cloud Computing 2011. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "text", + "content": "[62] Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. Huatuogpt, towards taming language model to be a doctor. arXiv preprint arXiv:2305.15075, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 286, + 497 + ], + "type": "text", + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 286, + 543 + ], + "type": "text", + "content": "[64] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8951" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_content_list.json b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6fa156627d11992a33e2aac9b5c2526d4c03afe --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "Utility-Fairness Trade-Offs and How to Find Them", + "text_level": 1, + "bbox": [ + 227, + 130, + 743, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sepehr Dehdashtian Bashir Sadeghi Vishnu Naresh Boddeti Michigan State University", + "bbox": [ + 228, + 180, + 743, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sepehr, sadeghib, vishnu}@msu.edu", + "bbox": [ + 331, + 219, + 633, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated and achievable fairness-utility trade-offs across multiple datasets and prediction tasks.", + "bbox": [ + 73, + 300, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As learning-based systems are increasingly being deployed in high-stakes applications, there is a dire need to ensure that they do not propagate or amplify any discriminative tendencies inherent in the training datasets. An ideal solution would impart fairness to prediction models while retaining the performance of the same model when learned without fairness considerations.", + "bbox": [ + 75, + 718, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realizing this goal necessitates optimizing two objectives: maximizing utility in predicting a label $Y$ for a target task (e.g., face identity) while minimizing the unfairness w.r.t. a demographic attribute $S$ (e.g., age or gender). However, when the statistical dependence between $Y$ and $S$", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e7bd30e645e7d173a7dc2a8924480d38c6247e42119da4cff837d352ef6f517.jpg", + "image_caption": [ + "(a)", + "Figure 1. The utility-fairness trade-offs. (a) Classification systems can be evaluated by their utility (e.g., accuracy) w.r.t. a target label $Y$ and their unfairness w.r.t. a demographic label $S$ . We introduce two trade-offs, Data Space Trade-Off (DST) and Label Space Trade-Off (LST). (b) We empirically estimate DST and LST on CelebA and evaluate the utility (high cheekbones) and fairness (gender & age) of over 100 zero-shot and 900 supervised models." + ], + "image_footnote": [], + "bbox": [ + 506, + 271, + 666, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/500b552bf89e130a4a75990ced3b0154c9f208c44ee0cca63507cce7732e3e48.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 674, + 271, + 875, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is not negligible, learning with fairness considerations will necessarily degrade the performance of the target predictor, i.e., a trade-off will exist between utility and fairness.", + "bbox": [ + 496, + 518, + 893, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existence of a utility-fairness trade-off has been well established, theoretically [11, 22, 29, 36, 37] and empirically [29], in multiple prior works. However, the focus of this body of work has been limited in multiple respects. First, prior work [29, 36] focused on just one type of trade-off, ignoring other possible trade-offs between utility and fairness. Second, prior work [36, 37] focused on establishing bounds or identifying the end-points of the trade-off of interest rather than attempting its precise characterization. Third, the majority of the prior work [11, 29, 36, 37] has investigated the utility-fairness trade-offs for one definition of group fairness, namely, demographic parity (DP). There are multiple fairness definitions [2], including those more practically relevant than DP, such as Equalized Opportunity (EO), for which the trade-offs have not been studied.", + "bbox": [ + 496, + 564, + 895, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these attempts, several questions related to the utility-fairness trade-offs remain outstanding.", + "bbox": [ + 498, + 790, + 893, + 821 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. What are the optimal utility-fairness trade-offs?", + "2. For a given prediction task and a demographic attribute, we wish to be fair w.r.t., how can we empirically estimate the trade-offs from data?" + ], + "bbox": [ + 506, + 832, + 885, + 892 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12037", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Addressing these questions by identifying and quantifying the trade-offs is the primary goal of this paper. The trade-offs are a function of the data triplet $(X,Y,S)$ , where $X$ is the data (e.g., images), $Y$ is the target label, and $S$ is the sensitive demographic label. Figure 1 illustrates the plausible trade-offs, their empirical estimation on CelebA [19], and their utility in empirically evaluating representations from pre-trained models.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Identifying Trade-Offs (§3). We identify two trade-offs: the Label-Space Trade-Off (LST) and Data-Space Trade-Off (DST). They can be defined for any group fairness definitions that can be expressed via independence and separation relations [2]. The LST corresponds to the trade-off obtained by an oracle fair classifier that depends only on the distributions of $Y$ and $S$ . Similarly, DST is the trade-off obtained by an optimally learned fair classifier and depends on $(X,Y,S)$ . By definition, LST necessarily dominates DST since it does not depend on the data $X$ .", + "bbox": [ + 75, + 217, + 472, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The trade-offs divide the utility-fairness plane into three regions shown in Fig. 1a. A possible region that can be attained by algorithms learned on the given data, a possible with extra data region that can be attained by learning on data beyond the given data, and an impossible region that cannot be attained by any algorithmic scheme due to the inherent dependence between the distributions of $Y$ and $S$ .", + "bbox": [ + 75, + 369, + 472, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantifying Trade-Offs (\\$4). Characterizing the exact trade-offs from data for a given task, demographic attribute, and fairness definition affords multiple benefits. It will allow researchers and practitioners to identify the achievable solution space for the task, evaluate how far a given predictor is from the optimal solution, and identify performance gaps and trends among existing solutions. To this end, we propose U-FaTE (Utility-Fairness Trade-Off Estimator), a method for quantifying the trade-offs from data triplets numerically. U-FaTE is an end-to-end model that adopts a statistical dependence measure as a proxy for utility and fairness and optimizes their weighted linear combination. U-FaTE can be flexibly adapted to estimate both the DST and LST from a finite labeled dataset.", + "bbox": [ + 75, + 481, + 472, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Usefulness of Trade-Offs (§5). The trade-offs illuminate the fundamental limits of learning algorithms in mitigating unfairness and present a new avenue to evaluate a given image representation in terms of its distance from the estimated trade-offs. We adopt this scheme to evaluate the representations of over 900 supervised and 100 zero-shot publicly available pre-trained models, derive insights, and identify trends and models that are close and far from the empirical trade-off estimates (§6).", + "bbox": [ + 75, + 698, + 472, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation: We denote scalars using lowercase letters, e.g., $d$ and $\\lambda$ . We denote deterministic vectors by boldface lowercase letters, e.g., $x$ , $y$ . Both scalar-valued and multidimensional random variables (RV)s are denoted by regular upper", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "case letters, e.g., $X$ , $Y$ . We denote deterministic matrices by boldface upper case letters, e.g., $\\pmb{K}$ , $\\Theta$ . Finite or infinite sets are denoted by calligraphic letters, e.g., $\\mathcal{A}$ , $\\mathcal{H}$ .", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 498, + 152, + 650, + 169 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A vast majority of prior work on designing fair classifiers focused primarily on uncovering disparities in practical tasks [5, 31] and learning a fair predictor [20, 32, 35] for a given fairness measure. An extended discussion of this body of work can be found in the supplementary material.", + "bbox": [ + 496, + 180, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utility-Fairness Trade-Offs: Many attempts on learning fair models [10, 20, 31, 32] ignored the existence of trade-offs. They sought to maximize accuracy on target tasks while minimizing unfairness, thus perhaps seeking an infeasible solution. Most studies on utility-fairness trade-offs are theoretical and under restricted settings in terms of the type of labels, notion of fairness, and bounds or extreme limits of trade-offs. For example, Zhao et al. [38] obtained a lower bound on DST when both $Y$ and $S$ are binary labels. McNamara et al. [21] provided both upper and lower bounds for binary labels. Only a couple of attempts [27, 29] have been made to numerically estimate utility-fairness trade-offs for independence related-based measures like demographic parity, both of them on features from pre-trained models, rather than raw data. Sadeghi et al. [27] obtained a simplified version of DST, but for linear models. Later on, in the context of invariant representation learning, this was extended to estimate a near-optimal DST-like trade-off called $\\mathcal{T}_{\\mathrm{Opt}}$ in [29]. But as we demonstrate in §6.2, the estimate of $\\mathcal{T}_{\\mathrm{Opt}}$ called K- $\\mathcal{T}_{\\mathrm{Opt}}$ does not span the entire trade-off.", + "bbox": [ + 496, + 260, + 892, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to this body of work, we identify two types of trade-offs, DST and LST, and propose a method to numerically quantify them from data. Our trade-offs and their empirical estimates apply to a wide range of prediction tasks for two different categories of fairness notions without any restrictions on the type of labels.", + "bbox": [ + 496, + 564, + 893, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning Fair Classifiers: Over the last decade, many methods have been developed for learning fair classifiers. These approaches follow the template of adopting a fairness constraint as a regularizer in addition to the objective for the target task. The approaches differ in the choice of measure as a proxy for quantifying the level of unfairness between the target label $Y$ and the prediction $\\hat{Y}$ , and the associated optimization technique. From an optimization perspective, they can be classified into three major categories-i.e., iterative adversarial methods (ARL[34], SARL[27], and MaxEnt-ARL[26]), non-iterative adversarial methods (FairHSIC [23], OptNet-ARL [28]), and closed-form solver methods (SARL [27], K- $\\mathcal{T}_{\\mathrm{Opt}}$ [29], LEACE [3], FairerCLIP [7]). Among these, ARL, SARL, MaxEnt-ARL, and OptNet-ARL measure mean dependence [1, 12], FairHSIC, FairerCLIP and K- $\\mathcal{T}_{\\mathrm{Opt}}$", + "bbox": [ + 496, + 659, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12038", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "measure full statistical dependence, i.e., all modes of dependence, and SARL and LEACE measures linear dependence.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "U-FaTE draws inspiration from $\\mathrm{K - T_{Opt}}$ [29]. By using a closed-form solver and a universal dependence measure that captures all non-linear dependencies, $\\mathrm{K - T_{Opt}}$ achieves a better utility-fairness trade-off and is more stable than the other fair learning methods discussed above. However, $\\mathrm{K - T_{Opt}}$ is limited in multiple respects and cannot be directly employed for estimating the trade-offs. 1) It operates on features and does not generalize to learning directly from high-dimensional raw data representations such as pixels for images. 2) $\\mathrm{K - T_{Opt}}$ optimizes an unconditional dependence measure, which limits its applicability to independence relation-based fairness definitions such as demographic parity. 3) As we demonstrate in §6.2, for demographic parity, $\\mathrm{K - T_{Opt}}$ 's trade-off is the closest to our DST estimate, but it does not span the entire utility-fairness trade-off front. Therefore, we adopt the positive aspects of $\\mathrm{K - T_{Opt}}$ , namely universal dependence measure and closed-form solver, into U-FaTE and overcome its drawbacks.", + "bbox": [ + 75, + 122, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Utility-Fairness Trade-Offs", + "text_level": 1, + "bbox": [ + 76, + 412, + 366, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fairness Notions: Group fairness notions are typically categorized into three classes [2], namely independence, separation, and sufficiency, each corresponding to different societal desiderata. We focus on the independence and separation relations, which can be expressed as independence $(\\hat{Y} \\perp S)$ and conditional independence $(\\hat{Y} \\perp S|Y = y)$ relations, respectively.", + "bbox": [ + 75, + 440, + 468, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider frequently used fairness criteria including Demographic Parity (DP) [17] which is an example of an independence relation, and Equalized Opportunity (EO) [14], and Equality of Odds (EOO) [14] which are both examples of separation relations. The corresponding unfairness metrics are Demographic Parity Violation $DPV \\coloneqq |P(\\hat{Y} = 1|S = 0) - P(\\hat{Y} = 1|S = 1)|$ , Equalized Opportunity Difference $EOD \\coloneqq |P(\\hat{Y} = 1|Y = 1,S = 0) - P(\\hat{Y} = 1|Y = 1,S = 1)|$ , and Equality of Odds Difference $EOOD \\coloneqq \\frac{1}{2}\\sum_{y\\in \\{0,1\\}}|P(\\hat{Y} = 1|Y = y,S = 0) - P(\\hat{Y} = 1|Y = y,S = 1)|$ , respectively.", + "bbox": [ + 75, + 547, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce the Data-Space Trade-Off and the Label-Space Trade-Off. In both, we employ a dependence measure $\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)$ to enforce the independence and separation based fairness constraints. The function $\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)\\geq 0$ is a parametric or non-parametric measure of statistical dependence. $\\mathrm{Dep}(P,Q|R = r) = 0$ implies that conditioned on $R = r$ , the random variables (RVs) $P$ and $Q$ are independent. $\\mathrm{Dep}(P,Q|R = r) > 0$ means that conditioned on $R = r$ , $P$ , and $Q$ are dependent, with larger values indicating larger degrees of dependence. When $r$ is the empty set $\\emptyset$ , we assume that $\\mathrm{Dep}(P,Q|R = \\emptyset)$ simply reduces to the unconditional dependence $\\mathrm{Dep}(P,Q)$ .", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. Data Space Trade-Off (DST)", + "text_level": 1, + "bbox": [ + 506, + 92, + 797, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {\\lambda} ^ {D S T} := \\underset {f \\in \\mathcal {H} _ {X}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {X, Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X)\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (X), S | Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 121, + 872, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $f$ is the encoder that maps data $X$ to a representation $Z$ , and $g$ is a classifier that predicts $\\hat{Y}$ from $Z$ . $\\mathcal{H}_X$ and $\\mathcal{H}_Y$ are the hypothesis classes of functions for $f$ and $g$ respectively. $\\mathcal{L}_Y(\\cdot, \\cdot)$ is the loss function corresponding to the utility, and $\\lambda$ controls the trade-off between utility and fairness, i.e., $\\lambda = 0$ corresponds to ignoring the fairness constraint and only optimizing the utility, while, $\\lambda \\rightarrow 1$ corresponds to the total fairness. The outcome $f_{\\lambda}^{DST}$ corresponds to the encoder for a given value of $\\lambda$ . This definition corresponds to the DST curve in Fig. 1a, where the utility-fairness plane below the DST corresponds to the region achievable by algorithms designed for this prediction task that learn from the data triplet $(X, Y, S) \\sim p(X, Y, S)$ .", + "bbox": [ + 498, + 183, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. Label Space Trade-Off (LST)", + "text_level": 1, + "bbox": [ + 506, + 392, + 799, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Z _ {\\lambda} ^ {L S T} := \\underset {Z \\in L ^ {2}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} (Z), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(Z, S \\mid Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 420, + 841, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $L^2$ is the space of all square-integrable RVs (i.e. $\\mathbb{E}_Z\\left[\\| Z\\|^2\\right] < \\infty$ ) in the probability space generated by the joint RV $(Y, S)$ . LST corresponds to the trade-off from an ideal representation space $Z_{\\lambda}^{LST}$ that is not constrained to be learned from the input data $X$ . It is the trade-off inherent to the task itself and is the best that any algorithm can hope to achieve for this task. Therefore, it necessarily dominates (or is equal to) DST in Definition 1. This definition corresponds to the LST curve in Fig. 1a, where the utility-fairness plane above the LST corresponds to the region that any algorithm cannot achieve.", + "bbox": [ + 498, + 484, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We stress that the above trade-offs are intrinsic to the underlying data, specifically the underlying distributions that generated that data. So, the trade-offs are a property of the data, not of any particular learning algorithm.", + "bbox": [ + 498, + 654, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LST and DST Divide: As illustrated by the yellow region in Fig. 1a, there is a potential gap between LST and DST. This gap at $\\lambda = 0$ stems from the irreducible error from the prediction $\\mathbb{E}[Y|X]$ of a Bayes Classifier, or when $Y$ is fully recoverable from $X$ . And, when $\\lambda > 0$ , the gap between LST and DST widens in two scenarios: 1) $Y \\not\\perp S$ : The model starts discarding $S$ from the representation $Z$ , which will lead to $Y$ being even less recoverable from $X$ compared to $\\lambda = 0$ ; and 2) $Y \\perp S$ : If $X$ entangles $Y$ and $S$ in such a way that $Y$ is not recoverable from $X$ when $S$ is discarded, it will lead to $Y$ being even less recoverable from $X$ compared to $\\lambda = 0$ .", + "bbox": [ + 498, + 719, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12039", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6158bb99a7fc1f4c25e651d7eb8bae7a3ff70d4d11460a0d8343ee273e8f7cd4.jpg", + "image_caption": [ + "Figure 2. Overview of U-FaTE: (Left) It comprises two components, a feature extractor and a fair encoder, that are trained end-to-end. Once U-FaTE is trained, the MLP classifier is trained to predict $Y$ from which fairness metrics can be computed. (Right) The fair encoder parameters are optimized through a closed-form solver operating on the features from the feature extractor. See text for more details." + ], + "image_footnote": [], + "bbox": [ + 150, + 89, + 823, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Numerically Quantifying the Trade-Offs", + "text_level": 1, + "bbox": [ + 75, + 275, + 441, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, we turn to the second goal of this paper, numerically quantifying the trade-offs from data. Fig. 2 shows a high-level overview of U-FaTE to learn a fair representation for a given trade-off parameter $\\lambda$ . U-FaTE comprises a feature extractor and a fair encoder. It receives raw data as input and uses a feature extractor to provide features for the fair encoder. The encoder uses the extracted features and employs a closed-form solver to find the optimum function that maps these features to a new feature space that minimizes the dependency on the sensitive attribute while maximizing the dependency on the target attribute. Following this, to predict the target $Y$ , a classifier is trained with the standard cross-entropy loss for classification problems. This process is repeated for multiple values of $\\lambda$ with $0 \\leq \\lambda < 1$ to obtain the full trade-off curves.", + "bbox": [ + 73, + 300, + 473, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Setup", + "text_level": 1, + "bbox": [ + 75, + 534, + 228, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start from Definition 1 and model the function $\\pmb{f}$ as a composition $f_{FE} \\circ f_{Enc}$ of the feature extractor and a fair encoder i.e., $\\pmb{f}(X; \\Theta) = f_{Enc}(f_{FE}(X; \\Theta_{FE}); \\Theta_{Enc})$ . We parameterize $\\pmb{f}$ with $\\Theta = [\\Theta_{FE}; \\Theta_{Enc}]$ where $\\Theta_{FE}$ are the parameters of $f_{FE}$ and $\\Theta_{Enc}$ are the parameters of $f_{Enc}$ . The objective function in Definition 1 is now", + "bbox": [ + 75, + 556, + 470, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {\\Theta}} \\left\\{\\left(1 - \\lambda\\right) \\inf _ {\\boldsymbol {\\Theta} _ {Y}} \\mathbb {E} _ {X, Y} \\left[ L _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X; \\boldsymbol {\\Theta}); \\boldsymbol {\\Theta} _ {Y}\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 656, + 472, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Dep}(\\pmb{f}(X_c; \\Theta), S_c)$ is equivalent to the term $\\operatorname{Dep}(\\pmb{f}(X), S|Y = y)$ in Definition 1 when $Y$ is not a continuous label. In this case, $X_c \\sim P(X|Y = y)$ and $S_c \\sim P(S|Y = y)$ are the random variables that represent the data and sensitive attribute conditioned on $Y = y$ , respectively. The fair representation is $Z = \\pmb{f}(X; \\Theta)$ .", + "bbox": [ + 75, + 718, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Optimization via Dependence Measures", + "text_level": 1, + "bbox": [ + 75, + 816, + 419, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The formulation in (1) can be directly optimized for an appropriate choice of dependence measure. Different choices of Dep lead to different fair representation learning methods. For instance, measuring Dep through an adversary", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "leads to the class of adversarial representation learning (ARL) methods [26-28, 34]. Similarly, employing the Hilbert Schmidt Independence Criterion (HSIC) [13] as Dep leads to FairHSIC [23]. However, due to challenges in optimization [26-28] and as we demonstrate in §6.2, these approaches are either very unstable, fail to span the trade-off or lead to sub-optimal trade-offs.", + "bbox": [ + 496, + 276, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, Sadeghi et al. [29] demonstrated that adopting an HSIC-like dependence measure for the fairness objective and the target loss leads to a closed-form solution that is both efficient and effective at finding a near-optimal trade-off. Therefore, we incorporate the HSIC-like dependence measure and the closed-form solver into U-FaTE. Thus (1) can be expressed as,", + "bbox": [ + 496, + 382, + 892, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sup _ {\\boldsymbol {f} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l l} (1 - \\lambda) \\operatorname {D e p} (\\boldsymbol {f} (X; \\boldsymbol {\\Theta}), Y) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 500, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_r$ is a function space that encourages the representations to be uncorrelated. It does not affect the optimality of the learned encoder [29] and improves the compactness of representation [4]. Note that while the first term involves all data $X$ , the second involves the conditional data $X_c$ .", + "bbox": [ + 496, + 569, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Choice of Dependence Measure: We adapt the dependence measure from [29] since it lends itself to a closed-form solution while capturing linear and non-linear dependencies under mild assumptions. While the dependence measure in [29] has been defined for absolute independence, our formulation in (2) also requires conditional independence to be compatible with separation based fairness definitions. Therefore, when $Y$ is not a continuous label, we define the conditional dependence measure as,", + "bbox": [ + 496, + 648, + 892, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\\\ \\sum_ {j = 1} ^ {r} \\sum_ {\\beta_ {S} \\in \\mathcal {U} _ {S}} \\mathbb {E} \\left[ \\left(f _ {j} \\left(X _ {c}\\right) - \\mathbb {E} f _ {j} \\left(X _ {c}\\right)\\right) \\left(\\beta_ {S} \\left(S _ {c}\\right) - \\mathbb {E} \\beta_ {S} \\left(S _ {c}\\right)\\right) \\right] \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 795, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{U}_S$ is a countable orthonormal basis set for the separable universal RKHS $\\mathcal{H}_S$ and $X_{c}\\sim P(X|Y = y)$ and", + "bbox": [ + 496, + 869, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12040", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S_{c}\\sim P(S|Y = y)$ are data and sensitive attributes, respectively. Empirically it can be estimated as,", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\frac {1}{n ^ {2}} \\| \\boldsymbol {\\Theta} \\boldsymbol {K} _ {X _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {S _ {c}} \\| _ {F} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 132, + 468, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ is the number of data samples, $\\pmb{K}_{X_c} \\in \\mathbb{R}^{n \\times n}$ is the Gram matrix corresponding to $\\mathcal{H}_X$ , $\\Theta$ is the encoder parameter in $\\pmb{f}(X) = \\Theta[k_{X_1}, k_{X_2}, \\dots, k_{X_n}]^T$ , $\\pmb{H} = \\pmb{I}_n - \\frac{1}{n} \\pmb{1}_n \\pmb{1}_n^T$ is the centering matrix, and $\\pmb{L}_{S_c}$ is a full column-rank matrix such that $\\pmb{L}_{S_c} \\pmb{L}_{S_c}^T = \\pmb{K}_{S_c}$ (Cholesky factorization).", + "bbox": [ + 76, + 171, + 468, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. A Solution to the Optimization Problem", + "text_level": 1, + "bbox": [ + 76, + 272, + 418, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Closed-Form Solver via Functions in RKHSs: Directly solving for all the parameters $\\Theta$ through (2) and (4) leads to abysmal performance in practice since the kernel $K_{X}$ has to be computed over the raw data space. Therefore, we instead define the fair encoder on the co-domain of the feature extractor $\\pmb{f}(\\cdot ;\\Theta_{FE})$ . So, in this case, (2) reduces to,", + "bbox": [ + 76, + 295, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sup _ {\\boldsymbol {f} _ {E n c} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l} (1 - \\lambda) \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} (\\tilde {X}; \\boldsymbol {\\Theta} _ {E n c}), Y\\right) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} \\left(\\tilde {X} _ {c}; \\boldsymbol {\\Theta} _ {E n c}\\right), S _ {c}\\right) \\right\\}, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 397, + 468, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{X} = f(X;\\Theta_{FE})$ , and the first and second terms are $\\frac{1}{n^2}\\| \\Theta_{Enc}K_{\\tilde{X}}HL_Y\\| _F^2$ and $\\frac{1}{n^2}\\left\\| \\Theta_{Enc}K_{\\tilde{X}_c}HL_{S_c}\\right\\| _F^2$ , respectively. The parameters $\\Theta_{Enc}$ can now be solved exactly via a closed-form solution:", + "bbox": [ + 76, + 467, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. A global optimizer of (5) is", + "bbox": [ + 76, + 542, + 349, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {f} _ {\\mathcal {H} _ {\\tilde {X}}} ^ {o p t} (\\tilde {X}; \\pmb {\\Theta} _ {E n c}) = \\pmb {\\Theta} _ {E n c} ^ {o p t} \\left[ k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {1}, \\tilde {X}), \\dots , k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {n}, \\tilde {X}) \\right] ^ {T}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 568, + 465, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta_{Enc}^{opt} = \\pmb{U}^T\\pmb{L}_{\\tilde{X}}^\\dagger \\in \\mathbb{R}^{r\\times n}$ and the columns of $\\pmb{U}$ are eigenvectors corresponding to the $r$ largest eigenvalues of the following generalized eigenvalue problem.", + "bbox": [ + 76, + 608, + 468, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left((1 - \\lambda) \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {Y} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} - \\lambda \\boldsymbol {L} _ {\\tilde {X} _ {c}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {S _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X} _ {c}}\\right) \\boldsymbol {u} \\\\ = \\lambda \\left(\\frac {1}{n} \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} + \\gamma \\boldsymbol {I}\\right) \\boldsymbol {u}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 666, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\mathbf{L}_{\\tilde{X}}\\mathbf{L}_{\\tilde{X}}^T = \\mathbf{K}_{\\tilde{X}},\\tilde{X}_c\\sim p(\\tilde{X} |Y = y)$ and $S_{c}\\sim$ $p(S|Y = y)$", + "bbox": [ + 76, + 738, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. The objective in (5) reduces to a generalized eigenvalue problem [18] by expressing it as a trace optimization problem. See supplementary for detailed proof. $\\square$", + "bbox": [ + 76, + 781, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While this is a general solution to (5), the solution for each group fairness case is detailed in the supplementary.", + "bbox": [ + 76, + 835, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alternating Optimization: Now we present our full algorithm to optimize (2). We adopt standard minibatch to learn", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the feature encoder's parameters $\\Theta_{FE}$ and the closed-form solver for the fair encoder parameters $\\Theta_{Enc}$ . We optimize them alternatively where in each iteration, we update $\\Theta_{Enc}$ while freezing $\\Theta_{FE}$ and vice-versa. Specifically, to optimize the fair encoder's parameters $\\Theta_{Enc}$ , we extract features from the data using the frozen feature extractor and use the closed-form solution in (5) to update $\\Theta_{Enc}$ . Then, we update the feature extractor's parameters $\\Theta_{FE}$ through minibatch SGD in (2) while freezing the encoder parameters. We repeat this process for every minibatch iteration. More details and an illustration of this alternating algorithm can be found in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Numerically Estimating the LST", + "text_level": 1, + "bbox": [ + 500, + 280, + 785, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Label Space Trade-off (LST) arises when the representation $Z$ is not restricted to be a function of $X$ . Following the discussion in the previous subsection, this trade-off can be formulated as,", + "bbox": [ + 496, + 303, + 890, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sup _ {Z \\in L _ {r} ^ {2}} \\left\\{(1 - \\lambda) \\operatorname {D e p} (Z, Y) - \\lambda \\operatorname {D e p} (Z, S | Y = y) \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 371, + 890, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_r^2$ is the space of all RVs of dimension $r$ with finite variance, i.e., $\\mathbb{E}_Z\\left[\\left\\| Z - \\mathbb{E}[Z]\\right\\|^2 < \\infty\\right]$ . From (7), observe that the optimal $Z$ is a function of $\\pmb{p}_{Y,S}$ only. Therefore, instead of directly optimizing $Z$ over $L_r^2$ , equivalently, we optimize for $\\Theta_{FE}$ and $\\Theta_{Enc}$ as", + "bbox": [ + 496, + 411, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}} \\left\\{\\left(1 - \\lambda\\right) \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(Y, S; \\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}\\right), Y\\right) \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (Y, S; \\Theta_ {F E}, \\Theta_ {E n c}), S | Y = y\\right) \\right\\}. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 498, + 890, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $f$ is a function of the labels $Y$ and $S$ , i.e., the model takes as input $Y$ and $S$ and seeks to remove the information corresponding to $S$ , including that present in $Y$ . In practice, to improve the stability of the optimization and facilitate learning, in addition to $Y$ and $S$ , we also use $X$ .", + "bbox": [ + 496, + 560, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Scheme for Evaluating Representations", + "text_level": 1, + "bbox": [ + 500, + 648, + 852, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The primary utility of the trade-offs is in illuminating the fundamental limits of learning algorithms in mitigating unfairness and evaluating the effectiveness of learned representations w.r.t. utility and fairness. This includes representations that provide a single solution in the utility-fairness plane and multiple solutions that span the trade-off between utility and fairness. Fairness evaluations in prior literature focused primarily on relative comparisons of the models to each other. Such a comparison, however, precludes an understanding of how far the solution is from the inherent limits of the task. Elucidating and numerically quantifying the inherent trade-offs facilitates such an understanding and drives further algorithmic development.", + "bbox": [ + 496, + 674, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A standard way to evaluate bi-objective solutions is to plot them on a 2-D plane, identify non-dominated solutions,", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4be1810a3edbf26f89d48fd89eaa3bdd2b4b3daa3b259ddbdf703b1ef54f28f.jpg", + "image_caption": [ + "Figure 3. Evaluating Fair Representation Learning Methods: Accuracy versus fairness trade-offs on CelebA (a)-(c) and FolkTable (d)-(f). (a) and (d) show the trade-off for Equalized Opportunity as the fairness constraint. (b) and (e) show the trade-off for Equality of Odds as the fairness constraint, and (c) and (f) show the trade-off for Demographic Parity as the fairness constraint. The solid lines represent the mean accuracy at a given fairness value, and the shaded region shows the uncertainty of the trade-off. Both DST and LST estimates from U-FaTE are stable. Among the FRL methods, $\\mathrm{K - T_{Opt}}$ is closest to the DST, while ARL has the most variance." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 908, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "or compare their dominance w.r.t. each other. More details can be found in the supplementary material.", + "bbox": [ + 75, + 306, + 468, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 351, + 312, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We designed experiments to answer the following:", + "bbox": [ + 76, + 377, + 413, + 392 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. How far are existing supervised fair representation learning methods from the two trade-offs? (§6.2)", + "2. How far are zero-shot representations from the two trade-offs? What is the effect of network architecture and pre-training dataset? (§6.3)", + "3. How far are pre-trained image representations trained in a supervised fashion from the two trade-offs? (§6.4)" + ], + "bbox": [ + 76, + 393, + 468, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 508, + 266, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We estimate the trade-offs through U-FaTE on an assortment of datasets. 1) CelebA [19] consists of more than 200K face images of celebrities in the wild annotated with 40 binary attributes. 2) FairFace [16] consists of face images from 7 different race groups labeled with race, sex, and age groups. 3) FolkTables [8] is a tabular dataset of individuals from fifty states derived from the US Census.", + "bbox": [ + 75, + 532, + 468, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For experiments on CelebA, the target attribute is high cheekbones, and the sensitive attribute is a combination of sex and age for a total of four classes (young woman, young man, old woman, and old man). For experiments on the FairFace dataset, sex (binary) is the target attribute with race (7 groups: East Asian, White, Black, Indian, Latino, South Asian, and Middle Eastern) being the sensitive attribute. For the FolkTables dataset, we consider data from Washington State and choose employment status (binary) and age (discrete value between 1 and 96) of the individuals as the target and sensitive attributes, respectively.", + "bbox": [ + 75, + 638, + 470, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: In all experiments, the utility is measured via classification accuracy. We consider three group fairness metrics (EOD, EOOD, and DPV). The zero-shot CLIP models are evaluated via cosine similarity. Moreover, on the Fair-Face dataset, we compared the models to the ideal solutions estimated by LST and DST using a weighted normalized", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Euclidean distance (see supplementary for details). We refer to the distance as $\\mathrm{Dist}_{LST}$ and $\\mathrm{Dist}_{DST}$ , respectively.", + "bbox": [ + 498, + 306, + 890, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: We use ResNet-18 as the feature extractor for U-FaTE and the FRL methods. The final classifier is a two-layer MLP. We evaluate the representations from pre-trained vision models by learning a logistic classifier. To obtain the trade-offs, we run U-FaTE and other FRL methods for multiple values of $\\lambda$ between zero and one, where zero corresponds to no fairness constraint, and one corresponds to only fairness.", + "bbox": [ + 496, + 340, + 892, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimizing U-FaTE: The trade-offs are defined through the dependence terms in (3) which involves an expectation over the joint distribution $p(X, Y, S)$ . However, due to practical considerations, we only have access to a finite set of samples to estimate the trade-offs. Therefore, we estimate the trade-offs using all the samples available in each dataset without splitting it into train, validation, and test sets. This choice ensures that the estimates account for any possible generalization gap between the train and test distributions and identify the best achievable utility-fairness trade-offs.", + "bbox": [ + 496, + 465, + 893, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Evaluating FRL Methods", + "text_level": 1, + "bbox": [ + 500, + 628, + 732, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FRL Baselines: We consider a wide range of FRL methods based on adversarial learning (ARL [34] and MaxEnt-ARL [26]), dependence measures (FairHSIC [23], OptNet-ARL [28]), and closed-form solvers $(\\mathrm{K - T_{Opt}}$ [29]).", + "bbox": [ + 498, + 652, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: We estimate the LST and DST through U-FaTE and the trade-offs from the other baselines across various settings and datasets. Figs. 3a to 3c show the trade-offs on the CelebA dataset for EOD, EOOD, and DPV, respectively. Similarly, Figs. 3d to 3f show the trade-offs on the Folktable dataset for EOD, EOOD, and DPV, respectively. In the plots, the solid lines represent the mean, and the light shadows represent the variance of the accuracy for a given fairness value. On FairFace, observe that trade-offs do not exist since, on this task, it is possible to mitigate unfairness without sacrificing accuracy. Hence, we present the results of FRL methods and the estimated LST and DST in Tab. 1.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12042", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2766db0713609a4154f374f1abaab0ee99aa64d034e5e51bbd36defbdf507d74.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (↑)Unfairness (↓)DistDST (↓)DistLST (↓)
EODARL [34]93.391.340.4480.559
FairHSIC [23]91.021.330.4450.557
OptNet-ARL [28]92.941.700.5980.709
U-FaTE-DST96.170.263-0.133
U-FaTE-LST100.00.0--
EOODARL [34]91.603.040.4470.71
FairHSIC [23]93.432.130.2360.498
OptNet-ARL [28]93.392.340.2840.546
U-FaTE-DST97.931.126-0.262
U-FaTE-LST100.00.0--
DPVARL [34]92.496.090.3500.351
FairHSIC [23]91.415.910.3290.332
OptNet-ARL [28]93.335.800.3160.317
U-FaTE-DST94.393.082-0.04
U-FaTE-LST100.03.10--
", + "bbox": [ + 78, + 88, + 470, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Evaluation of FRL methods on FairFace based on the distance to DST and LST estimated by U-FaTE. Color corresponds to the DST and LST trade-offs.", + "bbox": [ + 75, + 279, + 468, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observations: Although $\\mathrm{K - T_{Opt}}$ , OptNet-ARL and FairHSIC can achieve near-optimal accuracy in most cases, they are unable to span the whole range of fairness values. ARL is the most unstable but can span the whole range of fairness values. $\\mathrm{K - T_{Opt}}$ is the most stable method due to its closed-form solver, but is unable to span the whole range of fairness values (Figs. 3c and 3f).", + "bbox": [ + 75, + 337, + 468, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The gap between LST and DST demonstrates the information gap in $X$ for predicting $Y$ . From Fig. 3, we observe that the gap is $\\sim 20\\%$ of accuracy in low fairness regions and $\\sim 40\\%$ in high fairness regions for EOD and EOOD. For DPV, the trend reverses with a gap of $\\sim 20\\%$ for low fairness and gradually decreases with increasing unfairness.", + "bbox": [ + 75, + 445, + 468, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observe that the LST in Figs. 3a and 3b and Figs. 3d and 3e for EOD and EOOD is almost flat at $100\\%$ accuracy. This observation, however, is unsurprising since EOD and EOOD both condition on the label $Y$ , and thus an ideal classifier with $100\\%$ accuracy (i.e., $\\hat{Y} = Y$ ) will have zero EOD and EOOD. And, in LST, the Oracle classifier is $100\\%$ accurate since it has access to $Y$ and $S$ . So, the LST has sufficient information to minimize EOD and EOOD without sacrificing utility. The same, however, does not hold for DPV since it does not consider the target labels in its definition. Based on the above discussion, we deduce that EOD and EOOD are more pragmatic fairness metrics than DPV since they do not force the model to sacrifice predictive accuracy to ensure fairness. Thus, both offer a more balanced and practical approach to measuring fairness. Our empirical results provide independent confirmation of the same observations in [6, 14].", + "bbox": [ + 75, + 537, + 470, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparison of FRL methods in Tab. 1 based on $\\mathrm{Dist}_{LST}$ suggests that when models are optimized for EOD, ARL and FairHSIC find solutions closer to LST and DST than OptNet-ARL. When models are optimized for EOOD, FairHSIC finds the closest point to the LST and DST. OptNet-ARL performs slightly better than the other FRL methods when optimized for reducing DPV.", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Evaluating Zero-Shot CLIP Models", + "text_level": 1, + "bbox": [ + 498, + 90, + 808, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-Shot Models: To study the fairness of current zero-shot models, we consider more than 100 pre-trained models from OpenCLIP [15] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before.", + "bbox": [ + 496, + 118, + 890, + 179 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: Figs. 4a and 4b show results on CelebA and FairFace, respectively, for three group fairness definitions. Each point represents the result of one zero-shot CLIP model, with the color denoting the model's pre-training dataset. Plots also include DST and LST for comparison.", + "bbox": [ + 496, + 184, + 890, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observations: From Fig. 4a, we observe that zero-shot models perform poorly, in terms of accuracy, on the CelebA task. We hypothesize that it is so since the target task (predicting high cheekbones) is uncommon, even in the large datasets the models have been trained on. From a fairness perspective, we observe that models trained on CommonPool [9] (red dots) are more likely to be fair, while models trained on DataComp [9] have marginally better accuracy over the other models. Finally, the CLIP models are very far from the DST across the board.", + "bbox": [ + 496, + 263, + 890, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On FairFace (Fig. 4b), since the target task (sex) is abundantly represented in all large-scale pre-training datasets, we observe that the CLIP models exhibit high levels of accuracy. Two CLIP models pre-trained on OpenAI WIT [24] are the closest to DST and LST w.r.t. EOD. Similar to our observations on CelebA, models pre-trained on the CommonPool dataset are more likely to be fair but at the cost of accuracy. In contrast, models pre-trained on DataComp are more unfair but have greater accuracy. Finally, there is a clear positive correlation between accuracy and unfairness across all fairness metrics on FairFace (Fig. 4b).", + "bbox": [ + 496, + 416, + 890, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Evaluating Supervised Representations", + "text_level": 1, + "bbox": [ + 498, + 594, + 836, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Supervised Baselines: To study the fairness of image representations from models pre-trained in a supervised fashion, we consider more than 900 models from Pytorch Image Models [33] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before.", + "bbox": [ + 496, + 619, + 890, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: Figs. 4c and 4d show results on CelebA and Fair-Face datasets, respectively. Each point represents the result of one supervised model, with color denoting the model's pre-training dataset. Plots also include DST and LST for comparison, but some plots were magnified for better resolution. So, the LST may only be partially visible.", + "bbox": [ + 496, + 699, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observations: From the CelebA results in Fig. 4c, we observe a high positive correlation between the accuracy and EOD and EOOD. Thus, models with better accuracy are also more fair. And, in contrast to the zero-shot models, even though high cheekbones is a rare label, the representations have sufficient information to accurately detect it with a logistic regression classifier. Specifically, models pre", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a02c2de14caee88fc2cd783dc3499821d69be1240559551202411b3dc781772.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 89, + 480, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e26458bb9691ed5da278f0e48a7d3df3e94c7c47b2e5a68726c3e41cb6a39647.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 883, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95b0f331c6a56ff7065a57a305ad7e45312f6680cb82cdec8212acfda1bd0ee2.jpg", + "image_caption": [ + "(a)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 84, + 223, + 482, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78246bf3607b8d1fc7406e33a3730ded1a5cc1359b6ebfdb248d6ad000d641dd.jpg", + "image_caption": [ + "(d)", + "Figure 4. Evaluating Pre-Trained Zero-Shot and Supervised Models: Accuracy-fairness evaluation of more than 100 pre-trained zero-shot models on CelebA (a) and FairFace (b), and over 900 pre-trained image representations on CelebA (c), and FairFace (d)." + ], + "image_footnote": [], + "bbox": [ + 486, + 224, + 880, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained on ImageNet22K [25] (orange dots), and OpenAI WIT [24] have the best accuracy and reach the DST tradeoff in high unfairness regions. However, with more than $11\\%$ EOD and EOOD and more than $20\\%$ DPV, they have significant levels of bias between the two sexes.", + "bbox": [ + 75, + 391, + 468, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on FairFace in Fig. 4d also reiterate that models trained on OpenAI WIT and ImageNet22K are more fair and more accurate than other datasets. We also observe that the models are generally more fair on FairFace than on CelebA. A similar positive correlation exists between accuracy and fairness for EOD and EOOD. We also make an interesting observation from Fig. 4d. Some models surpass the DST and enter the \"Possible with Extra Data\" region of the utility-fairness plane illustrated in Fig. 1a. Recall that DST estimates the upper bound of the possible region for models trained with the same data as DST. The DST can be surpassed to enter the \"Possible with Extra Data\" region if the model is trained on additional data beyond what is used for estimating DST. From Fig. 4d (middle), we observe that a few models trained on OpenAI, LAION [30], ImageNet22K, ImageNet21K, and ImageNet12K can surpass the DST, plausibly since these datasets contain sufficient—both quality and quantity—samples from the distribution of the target attribute (sex).", + "bbox": [ + 75, + 467, + 470, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Concluding Remarks", + "text_level": 1, + "bbox": [ + 76, + 768, + 279, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As image classification systems are widely deployed in high-stakes applications, ensuring that their predictions do not exhibit demographic bias is paramount for gaining user trust. While it is desirable to mitigate bias without sacrificing accuracy, this is not always possible. This paper studied such inherent trade-offs between utility and fairness. First, we identified two types of trade-offs called Data-Space and", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Label-Space trade-offs corresponding to those achievable with and without data restrictions. But unlike prior theoretical studies on utility-fairness trade-offs, next, we focused on developing algorithmic tools for quantifying the trade-offs from data and proposed U-FaTE. As an illustration of its practical utility, we estimated the trade-offs on several image classification tasks, facilitating a large-scale evaluation of over 100 zero-shot and 900 supervised pre-trained models. The results revealed that, out of the box, pre-trained models are far from the best achievable limits of accuracy and fairness. Furthermore, we identified that, in some cases, larger datasets can improve accuracy and fairness and surpass the solutions represented by the DST.", + "bbox": [ + 496, + 391, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "U-FaTE was designed as a composition of a neural network with the last layer optimized to global optimality through a closed-form solver for a given feature representation. This design allowed it to estimate the two trade-offs reliably. However, U-FaTE does not provide convergence or optimality guarantees. Therefore, the estimated DST and LST are likely to be suboptimal. Nonetheless, they can serve as a valuable tool for understanding the nature of the problem (e.g., Does there exist a trade-off? or What are the possible, impossible and possible with extra data regions in the utility-fairness plane?) and how far a given fair learning algorithm is from the achievable limits. Furthermore, depending on which region of the utility-fairness plane a solution is and how far it is from the DST and LST reveals whether to focus on better optimization or better data.", + "bbox": [ + 496, + 614, + 892, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by the National Science Foundation (award #2147116).", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12044", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 2", + "[2] Solon Barocas, Moritz Hardt, and Arvind Narayanan. *Fairness and Machine Learning: Limitations and Opportunities*. MIT Press, 2023. 1, 2, 3", + "[3] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. LEACE: Perfect linear concept erasure in closed form. arXiv preprint arXiv:2306.03819, 2023. 2", + "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 4", + "[5] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, 2018. 2", + "[6] Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. big data 5, 2 (2017), 153-163. arXiv preprint arXiv:1610.07524, 2017. 7", + "[7] Sepehr Dehdashtian, Lan Wang, and Vishnu Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In International Conference on Learning Representations, 2024. 2", + "[8] Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. Retiring adult: New datasets for fair machine learning. In Advances in Neural Information Processing Systems, 2021. 6", + "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 7", + "[10] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating face recognition bias via group adaptive classifier. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[11] Thibaut Le Gouic, Jean-Michel Loubes, and Philippe Rigollet. Projection to fairness in statistical learning. arXiv preprint arXiv:2005.11720, 2020. 1", + "[12] Vincent Grari, Oualid El Hajouji, Sylvain Lamprier, and Marcin Detyniecki. Learning unbiased representations via Rényi minimization. arXiv preprint arXiv:2009.03183, 2020. 2", + "[13] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, and Bernhard Scholkopf. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12):2075-2129, 2005. 4", + "[14] Moritz Hardt, Eric Price, Nati Srebro, et al. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems, 2016. 3, 7" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7", + "[16] Kimmo Karkkainen and Jungseock Joo. FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 6", + "[17] Niki Kilbertus, Mateo Rojas Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Scholkopf. Avoiding discrimination through causal reasoning. Advances in Neural Information Processing Systems, 2017. 3", + "[18] Effrosini Kokiopoulou, Jie Chen, and Yousef Saad. Trace optimization and eigenproblems in dimension reduction methods. Numerical Linear Algebra with Applications, 18(3): 565-602, 2011. 5", + "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In IEEE International Conference on Computer Vision, 2015. 2, 6", + "[20] David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Learning adversarially fair and transferable representations. arXiv preprint arXiv:1802.06309, 2018. 2", + "[21] Daniel McNamara, Cheng Soon Ong, and Robert C Williamson. Costs and benefits of fair representation learning. AAAI/ACM Conference on AI, Ethics, and Society, 2019. 2", + "[22] Aditya Krishna Menon and Robert C Williamson. The cost of fairness in binary classification. Conference on Fairness, Accountability and Transparency, 2018. 1", + "[23] Novi Quadrianto, Viktoriia Sharmanska, and Oliver Thomas. Discovering fair representations in the data domain. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6, 7", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 7, 8", + "[25] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 8", + "[26] Proteek Roy and Vishnu Naresh Boddeti. Mitigating information leakage in image representations: A maximum entropy approach. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6", + "[27] Bashir Sadeghi, Runyi Yu, and Vishnu Boddeti. On the global optima of kernelized adversarial representation learning. IEEE International Conference on Computer Vision, 2019. 2", + "[28] Bashir Sadeghi, Lan Wang, and Vishnu Naresh Boddeti. Adversarial representation learning with closed-form solvers. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2021. 2, 4, 6, 7", + "[29] Bashir Sadeghi, Sepehr Dehdashtian, and Vishnu Boddeti. On characterizing the trade-off in invariant representation" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 3, 4, 6", + "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 8", + "[31] Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations. In IEEE/CVF International Conference on Computer Vision, 2019. 2", + "[32] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[33] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.7", + "[34] Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. Controllable invariance through adversarial feature learning. Advances in Neural Information Processing Systems, 2017. 2, 4, 6, 7", + "[35] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair representations. International Conference on Machine Learning, 2013. 2", + "[36] Han Zhao. Costs and benefits of wasserstein fair regression. arXiv preprint arXiv:2106.08812, 2021. 1", + "[37] Han Zhao and Geoffrey J Gordon. Inherent trade-offs in learning fair representations. arXiv preprint arXiv:1906.08386, 2019. 1", + "[38] Han Zhao, Jianfeng Chi, Yuan Tian, and Geoffrey J Gordon. Trade-offs and guarantees of adversarial representation learning for information obfuscation. arXiv preprint arXiv:1906.07902, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12046", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_model.json b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4407a9b1f8e7e2a239aa65d0eab455fcedaeba11 --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.131, + 0.745, + 0.154 + ], + "angle": 0, + "content": "Utility-Fairness Trade-Offs and How to Find Them" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.181, + 0.744, + 0.218 + ], + "angle": 0, + "content": "Sepehr Dehdashtian Bashir Sadeghi Vishnu Naresh Boddeti Michigan State University" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.22, + 0.635, + 0.235 + ], + "angle": 0, + "content": "{sepehr, sadeghib, vishnu}@msu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.301, + 0.474, + 0.665 + ], + "angle": 0, + "content": "When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated and achievable fairness-utility trade-offs across multiple datasets and prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.825 + ], + "angle": 0, + "content": "As learning-based systems are increasingly being deployed in high-stakes applications, there is a dire need to ensure that they do not propagate or amplify any discriminative tendencies inherent in the training datasets. An ideal solution would impart fairness to prediction models while retaining the performance of the same model when learned without fairness considerations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Realizing this goal necessitates optimizing two objectives: maximizing utility in predicting a label \\( Y \\) for a target task (e.g., face identity) while minimizing the unfairness w.r.t. a demographic attribute \\( S \\) (e.g., age or gender). However, when the statistical dependence between \\( Y \\) and \\( S \\)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.272, + 0.668, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.403, + 0.602, + 0.414 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.272, + 0.877, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.399, + 0.795, + 0.411 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.418, + 0.895, + 0.516 + ], + "angle": 0, + "content": "Figure 1. The utility-fairness trade-offs. (a) Classification systems can be evaluated by their utility (e.g., accuracy) w.r.t. a target label \\( Y \\) and their unfairness w.r.t. a demographic label \\( S \\). We introduce two trade-offs, Data Space Trade-Off (DST) and Label Space Trade-Off (LST). (b) We empirically estimate DST and LST on CelebA and evaluate the utility (high cheekbones) and fairness (gender & age) of over 100 zero-shot and 900 supervised models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.565 + ], + "angle": 0, + "content": "is not negligible, learning with fairness considerations will necessarily degrade the performance of the target predictor, i.e., a trade-off will exist between utility and fairness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.896, + 0.791 + ], + "angle": 0, + "content": "The existence of a utility-fairness trade-off has been well established, theoretically [11, 22, 29, 36, 37] and empirically [29], in multiple prior works. However, the focus of this body of work has been limited in multiple respects. First, prior work [29, 36] focused on just one type of trade-off, ignoring other possible trade-offs between utility and fairness. Second, prior work [36, 37] focused on establishing bounds or identifying the end-points of the trade-off of interest rather than attempting its precise characterization. Third, the majority of the prior work [11, 29, 36, 37] has investigated the utility-fairness trade-offs for one definition of group fairness, namely, demographic parity (DP). There are multiple fairness definitions [2], including those more practically relevant than DP, such as Equalized Opportunity (EO), for which the trade-offs have not been studied." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.791, + 0.894, + 0.822 + ], + "angle": 0, + "content": "Despite these attempts, several questions related to the utility-fairness trade-offs remain outstanding." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.833, + 0.845, + 0.848 + ], + "angle": 0, + "content": "1. What are the optimal utility-fairness trade-offs?" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.848, + 0.886, + 0.893 + ], + "angle": 0, + "content": "2. For a given prediction task and a demographic attribute, we wish to be fair w.r.t., how can we empirically estimate the trade-offs from data?" + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.833, + 0.886, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12037" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "Addressing these questions by identifying and quantifying the trade-offs is the primary goal of this paper. The trade-offs are a function of the data triplet \\((X,Y,S)\\), where \\(X\\) is the data (e.g., images), \\(Y\\) is the target label, and \\(S\\) is the sensitive demographic label. Figure 1 illustrates the plausible trade-offs, their empirical estimation on CelebA [19], and their utility in empirically evaluating representations from pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.473, + 0.369 + ], + "angle": 0, + "content": "Identifying Trade-Offs (§3). We identify two trade-offs: the Label-Space Trade-Off (LST) and Data-Space Trade-Off (DST). They can be defined for any group fairness definitions that can be expressed via independence and separation relations [2]. The LST corresponds to the trade-off obtained by an oracle fair classifier that depends only on the distributions of \\( Y \\) and \\( S \\). Similarly, DST is the trade-off obtained by an optimally learned fair classifier and depends on \\( (X,Y,S) \\). By definition, LST necessarily dominates DST since it does not depend on the data \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.477 + ], + "angle": 0, + "content": "The trade-offs divide the utility-fairness plane into three regions shown in Fig. 1a. A possible region that can be attained by algorithms learned on the given data, a possible with extra data region that can be attained by learning on data beyond the given data, and an impossible region that cannot be attained by any algorithmic scheme due to the inherent dependence between the distributions of \\( Y \\) and \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.473, + 0.693 + ], + "angle": 0, + "content": "Quantifying Trade-Offs (\\$4). Characterizing the exact trade-offs from data for a given task, demographic attribute, and fairness definition affords multiple benefits. It will allow researchers and practitioners to identify the achievable solution space for the task, evaluate how far a given predictor is from the optimal solution, and identify performance gaps and trends among existing solutions. To this end, we propose U-FaTE (Utility-Fairness Trade-Off Estimator), a method for quantifying the trade-offs from data triplets numerically. U-FaTE is an end-to-end model that adopts a statistical dependence measure as a proxy for utility and fairness and optimizes their weighted linear combination. U-FaTE can be flexibly adapted to estimate both the DST and LST from a finite labeled dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.473, + 0.837 + ], + "angle": 0, + "content": "Usefulness of Trade-Offs (§5). The trade-offs illuminate the fundamental limits of learning algorithms in mitigating unfairness and present a new avenue to evaluate a given image representation in terms of its distance from the estimated trade-offs. We adopt this scheme to evaluate the representations of over 900 supervised and 100 zero-shot publicly available pre-trained models, derive insights, and identify trends and models that are close and far from the empirical trade-off estimates (§6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Notation: We denote scalars using lowercase letters, e.g., \\(d\\) and \\(\\lambda\\). We denote deterministic vectors by boldface lowercase letters, e.g., \\(x\\), \\(y\\). Both scalar-valued and multidimensional random variables (RV)s are denoted by regular upper" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "case letters, e.g., \\(X\\), \\(Y\\). We denote deterministic matrices by boldface upper case letters, e.g., \\(\\pmb{K}\\), \\(\\Theta\\). Finite or infinite sets are denoted by calligraphic letters, e.g., \\(\\mathcal{A}\\), \\(\\mathcal{H}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.154, + 0.651, + 0.17 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.181, + 0.892, + 0.257 + ], + "angle": 0, + "content": "A vast majority of prior work on designing fair classifiers focused primarily on uncovering disparities in practical tasks [5, 31] and learning a fair predictor [20, 32, 35] for a given fairness measure. An extended discussion of this body of work can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Utility-Fairness Trade-Offs: Many attempts on learning fair models [10, 20, 31, 32] ignored the existence of trade-offs. They sought to maximize accuracy on target tasks while minimizing unfairness, thus perhaps seeking an infeasible solution. Most studies on utility-fairness trade-offs are theoretical and under restricted settings in terms of the type of labels, notion of fairness, and bounds or extreme limits of trade-offs. For example, Zhao et al. [38] obtained a lower bound on DST when both \\( Y \\) and \\( S \\) are binary labels. McNamara et al. [21] provided both upper and lower bounds for binary labels. Only a couple of attempts [27, 29] have been made to numerically estimate utility-fairness trade-offs for independence related-based measures like demographic parity, both of them on features from pre-trained models, rather than raw data. Sadeghi et al. [27] obtained a simplified version of DST, but for linear models. Later on, in the context of invariant representation learning, this was extended to estimate a near-optimal DST-like trade-off called \\( \\mathcal{T}_{\\mathrm{Opt}} \\) in [29]. But as we demonstrate in §6.2, the estimate of \\( \\mathcal{T}_{\\mathrm{Opt}} \\) called K- \\( \\mathcal{T}_{\\mathrm{Opt}} \\) does not span the entire trade-off." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.894, + 0.656 + ], + "angle": 0, + "content": "In contrast to this body of work, we identify two types of trade-offs, DST and LST, and propose a method to numerically quantify them from data. Our trade-offs and their empirical estimates apply to a wide range of prediction tasks for two different categories of fairness notions without any restrictions on the type of labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Learning Fair Classifiers: Over the last decade, many methods have been developed for learning fair classifiers. These approaches follow the template of adopting a fairness constraint as a regularizer in addition to the objective for the target task. The approaches differ in the choice of measure as a proxy for quantifying the level of unfairness between the target label \\( Y \\) and the prediction \\( \\hat{Y} \\), and the associated optimization technique. From an optimization perspective, they can be classified into three major categories-i.e., iterative adversarial methods (ARL[34], SARL[27], and MaxEnt-ARL[26]), non-iterative adversarial methods (FairHSIC [23], OptNet-ARL [28]), and closed-form solver methods (SARL [27], K- \\( \\mathcal{T}_{\\mathrm{Opt}} \\) [29], LEACE [3], FairerCLIP [7]). Among these, ARL, SARL, MaxEnt-ARL, and OptNet-ARL measure mean dependence [1, 12], FairHSIC, FairerCLIP and K- \\( \\mathcal{T}_{\\mathrm{Opt}} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12038" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "measure full statistical dependence, i.e., all modes of dependence, and SARL and LEACE measures linear dependence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.471, + 0.395 + ], + "angle": 0, + "content": "U-FaTE draws inspiration from \\(\\mathrm{K - T_{Opt}}\\) [29]. By using a closed-form solver and a universal dependence measure that captures all non-linear dependencies, \\(\\mathrm{K - T_{Opt}}\\) achieves a better utility-fairness trade-off and is more stable than the other fair learning methods discussed above. However, \\(\\mathrm{K - T_{Opt}}\\) is limited in multiple respects and cannot be directly employed for estimating the trade-offs. 1) It operates on features and does not generalize to learning directly from high-dimensional raw data representations such as pixels for images. 2) \\(\\mathrm{K - T_{Opt}}\\) optimizes an unconditional dependence measure, which limits its applicability to independence relation-based fairness definitions such as demographic parity. 3) As we demonstrate in §6.2, for demographic parity, \\(\\mathrm{K - T_{Opt}}\\) 's trade-off is the closest to our DST estimate, but it does not span the entire utility-fairness trade-off front. Therefore, we adopt the positive aspects of \\(\\mathrm{K - T_{Opt}}\\), namely universal dependence measure and closed-form solver, into U-FaTE and overcome its drawbacks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.367, + 0.431 + ], + "angle": 0, + "content": "3. The Utility-Fairness Trade-Offs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.469, + 0.546 + ], + "angle": 0, + "content": "Fairness Notions: Group fairness notions are typically categorized into three classes [2], namely independence, separation, and sufficiency, each corresponding to different societal desiderata. We focus on the independence and separation relations, which can be expressed as independence \\((\\hat{Y} \\perp S)\\) and conditional independence \\((\\hat{Y} \\perp S|Y = y)\\) relations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.718 + ], + "angle": 0, + "content": "We consider frequently used fairness criteria including Demographic Parity (DP) [17] which is an example of an independence relation, and Equalized Opportunity (EO) [14], and Equality of Odds (EOO) [14] which are both examples of separation relations. The corresponding unfairness metrics are Demographic Parity Violation \\( DPV \\coloneqq |P(\\hat{Y} = 1|S = 0) - P(\\hat{Y} = 1|S = 1)| \\), Equalized Opportunity Difference \\( EOD \\coloneqq |P(\\hat{Y} = 1|Y = 1,S = 0) - P(\\hat{Y} = 1|Y = 1,S = 1)| \\), and Equality of Odds Difference \\( EOOD \\coloneqq \\frac{1}{2}\\sum_{y\\in \\{0,1\\}}|P(\\hat{Y} = 1|Y = y,S = 0) - P(\\hat{Y} = 1|Y = y,S = 1)| \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We now introduce the Data-Space Trade-Off and the Label-Space Trade-Off. In both, we employ a dependence measure \\(\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)\\) to enforce the independence and separation based fairness constraints. The function \\(\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)\\geq 0\\) is a parametric or non-parametric measure of statistical dependence. \\(\\mathrm{Dep}(P,Q|R = r) = 0\\) implies that conditioned on \\(R = r\\), the random variables (RVs) \\(P\\) and \\(Q\\) are independent. \\(\\mathrm{Dep}(P,Q|R = r) > 0\\) means that conditioned on \\(R = r\\), \\(P\\), and \\(Q\\) are dependent, with larger values indicating larger degrees of dependence. When \\(r\\) is the empty set \\(\\emptyset\\), we assume that \\(\\mathrm{Dep}(P,Q|R = \\emptyset)\\) simply reduces to the unconditional dependence \\(\\mathrm{Dep}(P,Q)\\)." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.093, + 0.798, + 0.108 + ], + "angle": 0, + "content": "Definition 1. Data Space Trade-Off (DST)" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.122, + 0.873, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {\\lambda} ^ {D S T} := \\underset {f \\in \\mathcal {H} _ {X}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {X, Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X)\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (X), S | Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Here \\( f \\) is the encoder that maps data \\( X \\) to a representation \\( Z \\), and \\( g \\) is a classifier that predicts \\( \\hat{Y} \\) from \\( Z \\). \\( \\mathcal{H}_X \\) and \\( \\mathcal{H}_Y \\) are the hypothesis classes of functions for \\( f \\) and \\( g \\) respectively. \\( \\mathcal{L}_Y(\\cdot, \\cdot) \\) is the loss function corresponding to the utility, and \\( \\lambda \\) controls the trade-off between utility and fairness, i.e., \\( \\lambda = 0 \\) corresponds to ignoring the fairness constraint and only optimizing the utility, while, \\( \\lambda \\rightarrow 1 \\) corresponds to the total fairness. The outcome \\( f_{\\lambda}^{DST} \\) corresponds to the encoder for a given value of \\( \\lambda \\). This definition corresponds to the DST curve in Fig. 1a, where the utility-fairness plane below the DST corresponds to the region achievable by algorithms designed for this prediction task that learn from the data triplet \\( (X, Y, S) \\sim p(X, Y, S) \\)." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.393, + 0.8, + 0.409 + ], + "angle": 0, + "content": "Definition 2. Label Space Trade-Off (LST)" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.421, + 0.842, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Z _ {\\lambda} ^ {L S T} := \\underset {Z \\in L ^ {2}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} (Z), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(Z, S \\mid Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Here \\( L^2 \\) is the space of all square-integrable RVs (i.e. \\( \\mathbb{E}_Z\\left[\\| Z\\|^2\\right] < \\infty \\)) in the probability space generated by the joint RV \\( (Y, S) \\). LST corresponds to the trade-off from an ideal representation space \\( Z_{\\lambda}^{LST} \\) that is not constrained to be learned from the input data \\( X \\). It is the trade-off inherent to the task itself and is the best that any algorithm can hope to achieve for this task. Therefore, it necessarily dominates (or is equal to) DST in Definition 1. This definition corresponds to the LST curve in Fig. 1a, where the utility-fairness plane above the LST corresponds to the region that any algorithm cannot achieve." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We stress that the above trade-offs are intrinsic to the underlying data, specifically the underlying distributions that generated that data. So, the trade-offs are a property of the data, not of any particular learning algorithm." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The LST and DST Divide: As illustrated by the yellow region in Fig. 1a, there is a potential gap between LST and DST. This gap at \\(\\lambda = 0\\) stems from the irreducible error from the prediction \\(\\mathbb{E}[Y|X]\\) of a Bayes Classifier, or when \\(Y\\) is fully recoverable from \\(X\\). And, when \\(\\lambda > 0\\), the gap between LST and DST widens in two scenarios: 1) \\(Y \\not\\perp S\\): The model starts discarding \\(S\\) from the representation \\(Z\\), which will lead to \\(Y\\) being even less recoverable from \\(X\\) compared to \\(\\lambda = 0\\); and 2) \\(Y \\perp S\\): If \\(X\\) entangles \\(Y\\) and \\(S\\) in such a way that \\(Y\\) is not recoverable from \\(X\\) when \\(S\\) is discarded, it will lead to \\(Y\\) being even less recoverable from \\(X\\) compared to \\(\\lambda = 0\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12039" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.09, + 0.825, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.208, + 0.895, + 0.252 + ], + "angle": 0, + "content": "Figure 2. Overview of U-FaTE: (Left) It comprises two components, a feature extractor and a fair encoder, that are trained end-to-end. Once U-FaTE is trained, the MLP classifier is trained to predict \\( Y \\) from which fairness metrics can be computed. (Right) The fair encoder parameters are optimized through a closed-form solver operating on the features from the feature extractor. See text for more details." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.276, + 0.442, + 0.294 + ], + "angle": 0, + "content": "4. Numerically Quantifying the Trade-Offs" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.474, + 0.527 + ], + "angle": 0, + "content": "Now, we turn to the second goal of this paper, numerically quantifying the trade-offs from data. Fig. 2 shows a high-level overview of U-FaTE to learn a fair representation for a given trade-off parameter \\(\\lambda\\). U-FaTE comprises a feature extractor and a fair encoder. It receives raw data as input and uses a feature extractor to provide features for the fair encoder. The encoder uses the extracted features and employs a closed-form solver to find the optimum function that maps these features to a new feature space that minimizes the dependency on the sensitive attribute while maximizing the dependency on the target attribute. Following this, to predict the target \\(Y\\), a classifier is trained with the standard cross-entropy loss for classification problems. This process is repeated for multiple values of \\(\\lambda\\) with \\(0 \\leq \\lambda < 1\\) to obtain the full trade-off curves." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.535, + 0.23, + 0.552 + ], + "angle": 0, + "content": "4.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.472, + 0.651 + ], + "angle": 0, + "content": "We start from Definition 1 and model the function \\( \\pmb{f} \\) as a composition \\( f_{FE} \\circ f_{Enc} \\) of the feature extractor and a fair encoder i.e., \\( \\pmb{f}(X; \\Theta) = f_{Enc}(f_{FE}(X; \\Theta_{FE}); \\Theta_{Enc}) \\). We parameterize \\( \\pmb{f} \\) with \\( \\Theta = [\\Theta_{FE}; \\Theta_{Enc}] \\) where \\( \\Theta_{FE} \\) are the parameters of \\( f_{FE} \\) and \\( \\Theta_{Enc} \\) are the parameters of \\( f_{Enc} \\). The objective function in Definition 1 is now" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.657, + 0.473, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {\\Theta}} \\left\\{\\left(1 - \\lambda\\right) \\inf _ {\\boldsymbol {\\Theta} _ {Y}} \\mathbb {E} _ {X, Y} \\left[ L _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X; \\boldsymbol {\\Theta}); \\boldsymbol {\\Theta} _ {Y}\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.812 + ], + "angle": 0, + "content": "where \\(\\operatorname{Dep}(\\pmb{f}(X_c; \\Theta), S_c)\\) is equivalent to the term \\(\\operatorname{Dep}(\\pmb{f}(X), S|Y = y)\\) in Definition 1 when \\(Y\\) is not a continuous label. In this case, \\(X_c \\sim P(X|Y = y)\\) and \\(S_c \\sim P(S|Y = y)\\) are the random variables that represent the data and sensitive attribute conditioned on \\(Y = y\\), respectively. The fair representation is \\(Z = \\pmb{f}(X; \\Theta)\\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.818, + 0.421, + 0.835 + ], + "angle": 0, + "content": "4.2. Optimization via Dependence Measures" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The formulation in (1) can be directly optimized for an appropriate choice of dependence measure. Different choices of Dep lead to different fair representation learning methods. For instance, measuring Dep through an adversary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.383 + ], + "angle": 0, + "content": "leads to the class of adversarial representation learning (ARL) methods [26-28, 34]. Similarly, employing the Hilbert Schmidt Independence Criterion (HSIC) [13] as Dep leads to FairHSIC [23]. However, due to challenges in optimization [26-28] and as we demonstrate in §6.2, these approaches are either very unstable, fail to span the trade-off or lead to sub-optimal trade-offs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.489 + ], + "angle": 0, + "content": "Recently, Sadeghi et al. [29] demonstrated that adopting an HSIC-like dependence measure for the fairness objective and the target loss leads to a closed-form solution that is both efficient and effective at finding a near-optimal trade-off. Therefore, we incorporate the HSIC-like dependence measure and the closed-form solver into U-FaTE. Thus (1) can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.5, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sup _ {\\boldsymbol {f} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l l} (1 - \\lambda) \\operatorname {D e p} (\\boldsymbol {f} (X; \\boldsymbol {\\Theta}), Y) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.646 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_r\\) is a function space that encourages the representations to be uncorrelated. It does not affect the optimality of the learned encoder [29] and improves the compactness of representation [4]. Note that while the first term involves all data \\(X\\), the second involves the conditional data \\(X_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.893, + 0.786 + ], + "angle": 0, + "content": "Choice of Dependence Measure: We adapt the dependence measure from [29] since it lends itself to a closed-form solution while capturing linear and non-linear dependencies under mild assumptions. While the dependence measure in [29] has been defined for absolute independence, our formulation in (2) also requires conditional independence to be compatible with separation based fairness definitions. Therefore, when \\( Y \\) is not a continuous label, we define the conditional dependence measure as," + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.796, + 0.892, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\\\ \\sum_ {j = 1} ^ {r} \\sum_ {\\beta_ {S} \\in \\mathcal {U} _ {S}} \\mathbb {E} \\left[ \\left(f _ {j} \\left(X _ {c}\\right) - \\mathbb {E} f _ {j} \\left(X _ {c}\\right)\\right) \\left(\\beta_ {S} \\left(S _ {c}\\right) - \\mathbb {E} \\beta_ {S} \\left(S _ {c}\\right)\\right) \\right] \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{U}_S\\) is a countable orthonormal basis set for the separable universal RKHS \\(\\mathcal{H}_S\\) and \\(X_{c}\\sim P(X|Y = y)\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "\\(S_{c}\\sim P(S|Y = y)\\) are data and sensitive attributes, respectively. Empirically it can be estimated as," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.133, + 0.47, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\frac {1}{n ^ {2}} \\| \\boldsymbol {\\Theta} \\boldsymbol {K} _ {X _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {S _ {c}} \\| _ {F} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.172, + 0.469, + 0.263 + ], + "angle": 0, + "content": "where \\(n\\) is the number of data samples, \\(\\pmb{K}_{X_c} \\in \\mathbb{R}^{n \\times n}\\) is the Gram matrix corresponding to \\(\\mathcal{H}_X\\), \\(\\Theta\\) is the encoder parameter in \\(\\pmb{f}(X) = \\Theta[k_{X_1}, k_{X_2}, \\dots, k_{X_n}]^T\\), \\(\\pmb{H} = \\pmb{I}_n - \\frac{1}{n} \\pmb{1}_n \\pmb{1}_n^T\\) is the centering matrix, and \\(\\pmb{L}_{S_c}\\) is a full column-rank matrix such that \\(\\pmb{L}_{S_c} \\pmb{L}_{S_c}^T = \\pmb{K}_{S_c}\\) (Cholesky factorization)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.273, + 0.419, + 0.289 + ], + "angle": 0, + "content": "4.3. A Solution to the Optimization Problem" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.296, + 0.469, + 0.388 + ], + "angle": 0, + "content": "Closed-Form Solver via Functions in RKHSs: Directly solving for all the parameters \\(\\Theta\\) through (2) and (4) leads to abysmal performance in practice since the kernel \\(K_{X}\\) has to be computed over the raw data space. Therefore, we instead define the fair encoder on the co-domain of the feature extractor \\(\\pmb{f}(\\cdot ;\\Theta_{FE})\\). So, in this case, (2) reduces to," + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.398, + 0.469, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sup _ {\\boldsymbol {f} _ {E n c} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l} (1 - \\lambda) \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} (\\tilde {X}; \\boldsymbol {\\Theta} _ {E n c}), Y\\right) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} \\left(\\tilde {X} _ {c}; \\boldsymbol {\\Theta} _ {E n c}\\right), S _ {c}\\right) \\right\\}, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.533 + ], + "angle": 0, + "content": "where \\(\\tilde{X} = f(X;\\Theta_{FE})\\), and the first and second terms are \\(\\frac{1}{n^2}\\| \\Theta_{Enc}K_{\\tilde{X}}HL_Y\\| _F^2\\) and \\(\\frac{1}{n^2}\\left\\| \\Theta_{Enc}K_{\\tilde{X}_c}HL_{S_c}\\right\\| _F^2\\), respectively. The parameters \\(\\Theta_{Enc}\\) can now be solved exactly via a closed-form solution:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.543, + 0.35, + 0.558 + ], + "angle": 0, + "content": "Theorem 1. A global optimizer of (5) is" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.569, + 0.466, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\pmb {f} _ {\\mathcal {H} _ {\\tilde {X}}} ^ {o p t} (\\tilde {X}; \\pmb {\\Theta} _ {E n c}) = \\pmb {\\Theta} _ {E n c} ^ {o p t} \\left[ k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {1}, \\tilde {X}), \\dots , k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {n}, \\tilde {X}) \\right] ^ {T}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.609, + 0.47, + 0.657 + ], + "angle": 0, + "content": "where \\(\\Theta_{Enc}^{opt} = \\pmb{U}^T\\pmb{L}_{\\tilde{X}}^\\dagger \\in \\mathbb{R}^{r\\times n}\\) and the columns of \\(\\pmb{U}\\) are eigenvectors corresponding to the \\(r\\) largest eigenvalues of the following generalized eigenvalue problem." + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.667, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left((1 - \\lambda) \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {Y} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} - \\lambda \\boldsymbol {L} _ {\\tilde {X} _ {c}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {S _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X} _ {c}}\\right) \\boldsymbol {u} \\\\ = \\lambda \\left(\\frac {1}{n} \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} + \\gamma \\boldsymbol {I}\\right) \\boldsymbol {u}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.469, + 0.772 + ], + "angle": 0, + "content": "Here \\(\\mathbf{L}_{\\tilde{X}}\\mathbf{L}_{\\tilde{X}}^T = \\mathbf{K}_{\\tilde{X}},\\tilde{X}_c\\sim p(\\tilde{X} |Y = y)\\) and \\(S_{c}\\sim\\) \\(p(S|Y = y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.782, + 0.469, + 0.827 + ], + "angle": 0, + "content": "Proof. The objective in (5) reduces to a generalized eigenvalue problem [18] by expressing it as a trace optimization problem. See supplementary for detailed proof. \\(\\square\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.837, + 0.469, + 0.868 + ], + "angle": 0, + "content": "While this is a general solution to (5), the solution for each group fairness case is detailed in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Alternating Optimization: Now we present our full algorithm to optimize (2). We adopt standard minibatch to learn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "the feature encoder's parameters \\(\\Theta_{FE}\\) and the closed-form solver for the fair encoder parameters \\(\\Theta_{Enc}\\). We optimize them alternatively where in each iteration, we update \\(\\Theta_{Enc}\\) while freezing \\(\\Theta_{FE}\\) and vice-versa. Specifically, to optimize the fair encoder's parameters \\(\\Theta_{Enc}\\), we extract features from the data using the frozen feature extractor and use the closed-form solution in (5) to update \\(\\Theta_{Enc}\\). Then, we update the feature extractor's parameters \\(\\Theta_{FE}\\) through minibatch SGD in (2) while freezing the encoder parameters. We repeat this process for every minibatch iteration. More details and an illustration of this alternating algorithm can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.281, + 0.787, + 0.297 + ], + "angle": 0, + "content": "4.4. Numerically Estimating the LST" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.364 + ], + "angle": 0, + "content": "The Label Space Trade-off (LST) arises when the representation \\( Z \\) is not restricted to be a function of \\( X \\). Following the discussion in the previous subsection, this trade-off can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.372, + 0.892, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\sup _ {Z \\in L _ {r} ^ {2}} \\left\\{(1 - \\lambda) \\operatorname {D e p} (Z, Y) - \\lambda \\operatorname {D e p} (Z, S | Y = y) \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.491 + ], + "angle": 0, + "content": "where \\(L_r^2\\) is the space of all RVs of dimension \\(r\\) with finite variance, i.e., \\(\\mathbb{E}_Z\\left[\\left\\| Z - \\mathbb{E}[Z]\\right\\|^2 < \\infty\\right]\\). From (7), observe that the optimal \\(Z\\) is a function of \\(\\pmb{p}_{Y,S}\\) only. Therefore, instead of directly optimizing \\(Z\\) over \\(L_r^2\\), equivalently, we optimize for \\(\\Theta_{FE}\\) and \\(\\Theta_{Enc}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.499, + 0.891, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}} \\left\\{\\left(1 - \\lambda\\right) \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(Y, S; \\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}\\right), Y\\right) \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (Y, S; \\Theta_ {F E}, \\Theta_ {E n c}), S | Y = y\\right) \\right\\}. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Here \\( f \\) is a function of the labels \\( Y \\) and \\( S \\), i.e., the model takes as input \\( Y \\) and \\( S \\) and seeks to remove the information corresponding to \\( S \\), including that present in \\( Y \\). In practice, to improve the stability of the optimization and facilitate learning, in addition to \\( Y \\) and \\( S \\), we also use \\( X \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.854, + 0.667 + ], + "angle": 0, + "content": "5. Scheme for Evaluating Representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.87 + ], + "angle": 0, + "content": "The primary utility of the trade-offs is in illuminating the fundamental limits of learning algorithms in mitigating unfairness and evaluating the effectiveness of learned representations w.r.t. utility and fairness. This includes representations that provide a single solution in the utility-fairness plane and multiple solutions that span the trade-off between utility and fairness. Fairness evaluations in prior literature focused primarily on relative comparisons of the models to each other. Such a comparison, however, precludes an understanding of how far the solution is from the inherent limits of the task. Elucidating and numerically quantifying the inherent trade-offs facilitates such an understanding and drives further algorithmic development." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A standard way to evaluate bi-objective solutions is to plot them on a 2-D plane, identify non-dominated solutions," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.91, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Evaluating Fair Representation Learning Methods: Accuracy versus fairness trade-offs on CelebA (a)-(c) and FolkTable (d)-(f). (a) and (d) show the trade-off for Equalized Opportunity as the fairness constraint. (b) and (e) show the trade-off for Equality of Odds as the fairness constraint, and (c) and (f) show the trade-off for Demographic Parity as the fairness constraint. The solid lines represent the mean accuracy at a given fairness value, and the shaded region shows the uncertainty of the trade-off. Both DST and LST estimates from U-FaTE are stable. Among the FRL methods, \\(\\mathrm{K - T_{Opt}}\\) is closest to the DST, while ARL has the most variance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.339 + ], + "angle": 0, + "content": "or compare their dominance w.r.t. each other. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.313, + 0.369 + ], + "angle": 0, + "content": "6. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.378, + 0.414, + 0.393 + ], + "angle": 0, + "content": "We designed experiments to answer the following:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.394, + 0.469, + 0.424 + ], + "angle": 0, + "content": "1. How far are existing supervised fair representation learning methods from the two trade-offs? (§6.2)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.424, + 0.469, + 0.469 + ], + "angle": 0, + "content": "2. How far are zero-shot representations from the two trade-offs? What is the effect of network architecture and pre-training dataset? (§6.3)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.469, + 0.469, + 0.5 + ], + "angle": 0, + "content": "3. How far are pre-trained image representations trained in a supervised fashion from the two trade-offs? (§6.4)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.394, + 0.469, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.267, + 0.527 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Datasets: We estimate the trade-offs through U-FaTE on an assortment of datasets. 1) CelebA [19] consists of more than 200K face images of celebrities in the wild annotated with 40 binary attributes. 2) FairFace [16] consists of face images from 7 different race groups labeled with race, sex, and age groups. 3) FolkTables [8] is a tabular dataset of individuals from fifty states derived from the US Census." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.807 + ], + "angle": 0, + "content": "For experiments on CelebA, the target attribute is high cheekbones, and the sensitive attribute is a combination of sex and age for a total of four classes (young woman, young man, old woman, and old man). For experiments on the FairFace dataset, sex (binary) is the target attribute with race (7 groups: East Asian, White, Black, Indian, Latino, South Asian, and Middle Eastern) being the sensitive attribute. For the FolkTables dataset, we consider data from Washington State and choose employment status (binary) and age (discrete value between 1 and 96) of the individuals as the target and sensitive attributes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Metrics: In all experiments, the utility is measured via classification accuracy. We consider three group fairness metrics (EOD, EOOD, and DPV). The zero-shot CLIP models are evaluated via cosine similarity. Moreover, on the Fair-Face dataset, we compared the models to the ideal solutions estimated by LST and DST using a weighted normalized" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Euclidean distance (see supplementary for details). We refer to the distance as \\(\\mathrm{Dist}_{LST}\\) and \\(\\mathrm{Dist}_{DST}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.893, + 0.463 + ], + "angle": 0, + "content": "Implementation Details: We use ResNet-18 as the feature extractor for U-FaTE and the FRL methods. The final classifier is a two-layer MLP. We evaluate the representations from pre-trained vision models by learning a logistic classifier. To obtain the trade-offs, we run U-FaTE and other FRL methods for multiple values of \\(\\lambda\\) between zero and one, where zero corresponds to no fairness constraint, and one corresponds to only fairness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.894, + 0.619 + ], + "angle": 0, + "content": "Optimizing U-FaTE: The trade-offs are defined through the dependence terms in (3) which involves an expectation over the joint distribution \\( p(X, Y, S) \\). However, due to practical considerations, we only have access to a finite set of samples to estimate the trade-offs. Therefore, we estimate the trade-offs using all the samples available in each dataset without splitting it into train, validation, and test sets. This choice ensures that the estimates account for any possible generalization gap between the train and test distributions and identify the best achievable utility-fairness trade-offs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.733, + 0.646 + ], + "angle": 0, + "content": "6.2. Evaluating FRL Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.892, + 0.716 + ], + "angle": 0, + "content": "FRL Baselines: We consider a wide range of FRL methods based on adversarial learning (ARL [34] and MaxEnt-ARL [26]), dependence measures (FairHSIC [23], OptNet-ARL [28]), and closed-form solvers \\((\\mathrm{K - T_{Opt}}\\) [29])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Results: We estimate the LST and DST through U-FaTE and the trade-offs from the other baselines across various settings and datasets. Figs. 3a to 3c show the trade-offs on the CelebA dataset for EOD, EOOD, and DPV, respectively. Similarly, Figs. 3d to 3f show the trade-offs on the Folktable dataset for EOD, EOOD, and DPV, respectively. In the plots, the solid lines represent the mean, and the light shadows represent the variance of the accuracy for a given fairness value. On FairFace, observe that trade-offs do not exist since, on this task, it is possible to mitigate unfairness without sacrificing accuracy. Hence, we present the results of FRL methods and the estimated LST and DST in Tab. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12042" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.27 + ], + "angle": 0, + "content": "
MethodAccuracy (↑)Unfairness (↓)DistDST (↓)DistLST (↓)
EODARL [34]93.391.340.4480.559
FairHSIC [23]91.021.330.4450.557
OptNet-ARL [28]92.941.700.5980.709
U-FaTE-DST96.170.263-0.133
U-FaTE-LST100.00.0--
EOODARL [34]91.603.040.4470.71
FairHSIC [23]93.432.130.2360.498
OptNet-ARL [28]93.392.340.2840.546
U-FaTE-DST97.931.126-0.262
U-FaTE-LST100.00.0--
DPVARL [34]92.496.090.3500.351
FairHSIC [23]91.415.910.3290.332
OptNet-ARL [28]93.335.800.3160.317
U-FaTE-DST94.393.082-0.04
U-FaTE-LST100.03.10--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.322 + ], + "angle": 0, + "content": "Table 1. Evaluation of FRL methods on FairFace based on the distance to DST and LST estimated by U-FaTE. Color corresponds to the DST and LST trade-offs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.444 + ], + "angle": 0, + "content": "Observations: Although \\(\\mathrm{K - T_{Opt}}\\), OptNet-ARL and FairHSIC can achieve near-optimal accuracy in most cases, they are unable to span the whole range of fairness values. ARL is the most unstable but can span the whole range of fairness values. \\(\\mathrm{K - T_{Opt}}\\) is the most stable method due to its closed-form solver, but is unable to span the whole range of fairness values (Figs. 3c and 3f)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.537 + ], + "angle": 0, + "content": "The gap between LST and DST demonstrates the information gap in \\( X \\) for predicting \\( Y \\). From Fig. 3, we observe that the gap is \\( \\sim 20\\% \\) of accuracy in low fairness regions and \\( \\sim 40\\% \\) in high fairness regions for EOD and EOOD. For DPV, the trend reverses with a gap of \\( \\sim 20\\% \\) for low fairness and gradually decreases with increasing unfairness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.794 + ], + "angle": 0, + "content": "Observe that the LST in Figs. 3a and 3b and Figs. 3d and 3e for EOD and EOOD is almost flat at \\(100\\%\\) accuracy. This observation, however, is unsurprising since EOD and EOOD both condition on the label \\(Y\\), and thus an ideal classifier with \\(100\\%\\) accuracy (i.e., \\(\\hat{Y} = Y\\)) will have zero EOD and EOOD. And, in LST, the Oracle classifier is \\(100\\%\\) accurate since it has access to \\(Y\\) and \\(S\\). So, the LST has sufficient information to minimize EOD and EOOD without sacrificing utility. The same, however, does not hold for DPV since it does not consider the target labels in its definition. Based on the above discussion, we deduce that EOD and EOOD are more pragmatic fairness metrics than DPV since they do not force the model to sacrifice predictive accuracy to ensure fairness. Thus, both offer a more balanced and practical approach to measuring fairness. Our empirical results provide independent confirmation of the same observations in [6, 14]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The comparison of FRL methods in Tab. 1 based on \\(\\mathrm{Dist}_{LST}\\) suggests that when models are optimized for EOD, ARL and FairHSIC find solutions closer to LST and DST than OptNet-ARL. When models are optimized for EOOD, FairHSIC finds the closest point to the LST and DST. OptNet-ARL performs slightly better than the other FRL methods when optimized for reducing DPV." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.81, + 0.108 + ], + "angle": 0, + "content": "6.3. Evaluating Zero-Shot CLIP Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.119, + 0.892, + 0.18 + ], + "angle": 0, + "content": "Zero-Shot Models: To study the fairness of current zero-shot models, we consider more than 100 pre-trained models from OpenCLIP [15] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Results: Figs. 4a and 4b show results on CelebA and FairFace, respectively, for three group fairness definitions. Each point represents the result of one zero-shot CLIP model, with the color denoting the model's pre-training dataset. Plots also include DST and LST for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Observations: From Fig. 4a, we observe that zero-shot models perform poorly, in terms of accuracy, on the CelebA task. We hypothesize that it is so since the target task (predicting high cheekbones) is uncommon, even in the large datasets the models have been trained on. From a fairness perspective, we observe that models trained on CommonPool [9] (red dots) are more likely to be fair, while models trained on DataComp [9] have marginally better accuracy over the other models. Finally, the CLIP models are very far from the DST across the board." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.584 + ], + "angle": 0, + "content": "On FairFace (Fig. 4b), since the target task (sex) is abundantly represented in all large-scale pre-training datasets, we observe that the CLIP models exhibit high levels of accuracy. Two CLIP models pre-trained on OpenAI WIT [24] are the closest to DST and LST w.r.t. EOD. Similar to our observations on CelebA, models pre-trained on the CommonPool dataset are more likely to be fair but at the cost of accuracy. In contrast, models pre-trained on DataComp are more unfair but have greater accuracy. Finally, there is a clear positive correlation between accuracy and unfairness across all fairness metrics on FairFace (Fig. 4b)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.595, + 0.837, + 0.612 + ], + "angle": 0, + "content": "6.4. Evaluating Supervised Representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Supervised Baselines: To study the fairness of image representations from models pre-trained in a supervised fashion, we consider more than 900 models from Pytorch Image Models [33] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Results: Figs. 4c and 4d show results on CelebA and Fair-Face datasets, respectively. Each point represents the result of one supervised model, with color denoting the model's pre-training dataset. Plots also include DST and LST for comparison, but some plots were magnified for better resolution. So, the LST may only be partially visible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Observations: From the CelebA results in Fig. 4c, we observe a high positive correlation between the accuracy and EOD and EOOD. Thus, models with better accuracy are also more fair. And, in contrast to the zero-shot models, even though high cheekbones is a rare label, the representations have sufficient information to accurately detect it with a logistic regression classifier. Specifically, models pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12043" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.482, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.212, + 0.294, + 0.223 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.885, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.212, + 0.695, + 0.223 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.224, + 0.483, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.336, + 0.292, + 0.347 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.225, + 0.882, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.335, + 0.693, + 0.346 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.35, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 4. Evaluating Pre-Trained Zero-Shot and Supervised Models: Accuracy-fairness evaluation of more than 100 pre-trained zero-shot models on CelebA (a) and FairFace (b), and over 900 pre-trained image representations on CelebA (c), and FairFace (d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.467 + ], + "angle": 0, + "content": "trained on ImageNet22K [25] (orange dots), and OpenAI WIT [24] have the best accuracy and reach the DST tradeoff in high unfairness regions. However, with more than \\(11\\%\\) EOD and EOOD and more than \\(20\\%\\) DPV, they have significant levels of bias between the two sexes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.471, + 0.756 + ], + "angle": 0, + "content": "Results on FairFace in Fig. 4d also reiterate that models trained on OpenAI WIT and ImageNet22K are more fair and more accurate than other datasets. We also observe that the models are generally more fair on FairFace than on CelebA. A similar positive correlation exists between accuracy and fairness for EOD and EOOD. We also make an interesting observation from Fig. 4d. Some models surpass the DST and enter the \"Possible with Extra Data\" region of the utility-fairness plane illustrated in Fig. 1a. Recall that DST estimates the upper bound of the possible region for models trained with the same data as DST. The DST can be surpassed to enter the \"Possible with Extra Data\" region if the model is trained on additional data beyond what is used for estimating DST. From Fig. 4d (middle), we observe that a few models trained on OpenAI, LAION [30], ImageNet22K, ImageNet21K, and ImageNet12K can surpass the DST, plausibly since these datasets contain sufficient—both quality and quantity—samples from the distribution of the target attribute (sex)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.281, + 0.787 + ], + "angle": 0, + "content": "7. Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As image classification systems are widely deployed in high-stakes applications, ensuring that their predictions do not exhibit demographic bias is paramount for gaining user trust. While it is desirable to mitigate bias without sacrificing accuracy, this is not always possible. This paper studied such inherent trade-offs between utility and fairness. First, we identified two types of trade-offs called Data-Space and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Label-Space trade-offs corresponding to those achievable with and without data restrictions. But unlike prior theoretical studies on utility-fairness trade-offs, next, we focused on developing algorithmic tools for quantifying the trade-offs from data and proposed U-FaTE. As an illustration of its practical utility, we estimated the trade-offs on several image classification tasks, facilitating a large-scale evaluation of over 100 zero-shot and 900 supervised pre-trained models. The results revealed that, out of the box, pre-trained models are far from the best achievable limits of accuracy and fairness. Furthermore, we identified that, in some cases, larger datasets can improve accuracy and fairness and surpass the solutions represented by the DST." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.893, + 0.844 + ], + "angle": 0, + "content": "U-FaTE was designed as a composition of a neural network with the last layer optimized to global optimality through a closed-form solver for a given feature representation. This design allowed it to estimate the two trade-offs reliably. However, U-FaTE does not provide convergence or optimality guarantees. Therefore, the estimated DST and LST are likely to be suboptimal. Nonetheless, they can serve as a valuable tool for understanding the nature of the problem (e.g., Does there exist a trade-off? or What are the possible, impossible and possible with extra data regions in the utility-fairness plane?) and how far a given fair learning algorithm is from the achievable limits. Furthermore, depending on which region of the utility-fairness plane a solution is and how far it is from the DST and LST reveals whether to focus on better optimization or better data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by the National Science Foundation (award #2147116)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12044" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Solon Barocas, Moritz Hardt, and Arvind Narayanan. *Fairness and Machine Learning: Limitations and Opportunities*. MIT Press, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. LEACE: Perfect linear concept erasure in closed form. arXiv preprint arXiv:2306.03819, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. big data 5, 2 (2017), 153-163. arXiv preprint arXiv:1610.07524, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[7] Sepehr Dehdashtian, Lan Wang, and Vishnu Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[8] Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. Retiring adult: New datasets for fair machine learning. In Advances in Neural Information Processing Systems, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating face recognition bias via group adaptive classifier. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[11] Thibaut Le Gouic, Jean-Michel Loubes, and Philippe Rigollet. Projection to fairness in statistical learning. arXiv preprint arXiv:2005.11720, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[12] Vincent Grari, Oualid El Hajouji, Sylvain Lamprier, and Marcin Detyniecki. Learning unbiased representations via Rényi minimization. arXiv preprint arXiv:2009.03183, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[13] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, and Bernhard Scholkopf. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12):2075-2129, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Moritz Hardt, Eric Price, Nati Srebro, et al. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems, 2016. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[16] Kimmo Karkkainen and Jungseock Joo. FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[17] Niki Kilbertus, Mateo Rojas Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Scholkopf. Avoiding discrimination through causal reasoning. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[18] Effrosini Kokiopoulou, Jie Chen, and Yousef Saad. Trace optimization and eigenproblems in dimension reduction methods. Numerical Linear Algebra with Applications, 18(3): 565-602, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In IEEE International Conference on Computer Vision, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[20] David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Learning adversarially fair and transferable representations. arXiv preprint arXiv:1802.06309, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[21] Daniel McNamara, Cheng Soon Ong, and Robert C Williamson. Costs and benefits of fair representation learning. AAAI/ACM Conference on AI, Ethics, and Society, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[22] Aditya Krishna Menon and Robert C Williamson. The cost of fairness in binary classification. Conference on Fairness, Accountability and Transparency, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Novi Quadrianto, Viktoriia Sharmanska, and Oliver Thomas. Discovering fair representations in the data domain. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[25] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Proteek Roy and Vishnu Naresh Boddeti. Mitigating information leakage in image representations: A maximum entropy approach. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[27] Bashir Sadeghi, Runyi Yu, and Vishnu Boddeti. On the global optima of kernelized adversarial representation learning. IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Bashir Sadeghi, Lan Wang, and Vishnu Naresh Boddeti. Adversarial representation learning with closed-form solvers. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2021. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Bashir Sadeghi, Sepehr Dehdashtian, and Vishnu Boddeti. On characterizing the trade-off in invariant representation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12045" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "learning. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[31] Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations. In IEEE/CVF International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[32] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[33] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[34] Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. Controllable invariance through adversarial feature learning. Advances in Neural Information Processing Systems, 2017. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[35] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair representations. International Conference on Machine Learning, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.468, + 0.516 + ], + "angle": 0, + "content": "[36] Han Zhao. Costs and benefits of wasserstein fair regression. arXiv preprint arXiv:2106.08812, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[37] Han Zhao and Geoffrey J Gordon. Inherent trade-offs in learning fair representations. arXiv preprint arXiv:1906.08386, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[38] Han Zhao, Jianfeng Chi, Yuan Tian, and Geoffrey J Gordon. Trade-offs and guarantees of adversarial representation learning for information obfuscation. arXiv preprint arXiv:1906.07902, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12046" + } + ] +] \ No newline at end of file diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_origin.pdf b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6629239a0ff0800be7df6d050226aa6a63cf484 --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/dffbbb65-0327-47d2-9043-f34845ca6d7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b7dbd7266deddcb0b5bce993f27eaacff5eb8c7705ce2364b9119b20e373555 +size 1976677 diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/full.md b/2024/Utility-Fairness Trade-Offs and How to Find Them/full.md new file mode 100644 index 0000000000000000000000000000000000000000..337447189a7c766739f790629f724ab2dc26da66 --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/full.md @@ -0,0 +1,322 @@ +# Utility-Fairness Trade-Offs and How to Find Them + +Sepehr Dehdashtian Bashir Sadeghi Vishnu Naresh Boddeti Michigan State University + +{sepehr, sadeghib, vishnu}@msu.edu + +# Abstract + +When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated and achievable fairness-utility trade-offs across multiple datasets and prediction tasks. + +# 1. Introduction + +As learning-based systems are increasingly being deployed in high-stakes applications, there is a dire need to ensure that they do not propagate or amplify any discriminative tendencies inherent in the training datasets. An ideal solution would impart fairness to prediction models while retaining the performance of the same model when learned without fairness considerations. + +Realizing this goal necessitates optimizing two objectives: maximizing utility in predicting a label $Y$ for a target task (e.g., face identity) while minimizing the unfairness w.r.t. a demographic attribute $S$ (e.g., age or gender). However, when the statistical dependence between $Y$ and $S$ + +![](images/7e7bd30e645e7d173a7dc2a8924480d38c6247e42119da4cff837d352ef6f517.jpg) +(a) +Figure 1. The utility-fairness trade-offs. (a) Classification systems can be evaluated by their utility (e.g., accuracy) w.r.t. a target label $Y$ and their unfairness w.r.t. a demographic label $S$ . We introduce two trade-offs, Data Space Trade-Off (DST) and Label Space Trade-Off (LST). (b) We empirically estimate DST and LST on CelebA and evaluate the utility (high cheekbones) and fairness (gender & age) of over 100 zero-shot and 900 supervised models. + +![](images/500b552bf89e130a4a75990ced3b0154c9f208c44ee0cca63507cce7732e3e48.jpg) +(b) + +is not negligible, learning with fairness considerations will necessarily degrade the performance of the target predictor, i.e., a trade-off will exist between utility and fairness. + +The existence of a utility-fairness trade-off has been well established, theoretically [11, 22, 29, 36, 37] and empirically [29], in multiple prior works. However, the focus of this body of work has been limited in multiple respects. First, prior work [29, 36] focused on just one type of trade-off, ignoring other possible trade-offs between utility and fairness. Second, prior work [36, 37] focused on establishing bounds or identifying the end-points of the trade-off of interest rather than attempting its precise characterization. Third, the majority of the prior work [11, 29, 36, 37] has investigated the utility-fairness trade-offs for one definition of group fairness, namely, demographic parity (DP). There are multiple fairness definitions [2], including those more practically relevant than DP, such as Equalized Opportunity (EO), for which the trade-offs have not been studied. + +Despite these attempts, several questions related to the utility-fairness trade-offs remain outstanding. + +1. What are the optimal utility-fairness trade-offs? +2. For a given prediction task and a demographic attribute, we wish to be fair w.r.t., how can we empirically estimate the trade-offs from data? + +Addressing these questions by identifying and quantifying the trade-offs is the primary goal of this paper. The trade-offs are a function of the data triplet $(X,Y,S)$ , where $X$ is the data (e.g., images), $Y$ is the target label, and $S$ is the sensitive demographic label. Figure 1 illustrates the plausible trade-offs, their empirical estimation on CelebA [19], and their utility in empirically evaluating representations from pre-trained models. + +Identifying Trade-Offs (§3). We identify two trade-offs: the Label-Space Trade-Off (LST) and Data-Space Trade-Off (DST). They can be defined for any group fairness definitions that can be expressed via independence and separation relations [2]. The LST corresponds to the trade-off obtained by an oracle fair classifier that depends only on the distributions of $Y$ and $S$ . Similarly, DST is the trade-off obtained by an optimally learned fair classifier and depends on $(X,Y,S)$ . By definition, LST necessarily dominates DST since it does not depend on the data $X$ . + +The trade-offs divide the utility-fairness plane into three regions shown in Fig. 1a. A possible region that can be attained by algorithms learned on the given data, a possible with extra data region that can be attained by learning on data beyond the given data, and an impossible region that cannot be attained by any algorithmic scheme due to the inherent dependence between the distributions of $Y$ and $S$ . + +Quantifying Trade-Offs (\$4). Characterizing the exact trade-offs from data for a given task, demographic attribute, and fairness definition affords multiple benefits. It will allow researchers and practitioners to identify the achievable solution space for the task, evaluate how far a given predictor is from the optimal solution, and identify performance gaps and trends among existing solutions. To this end, we propose U-FaTE (Utility-Fairness Trade-Off Estimator), a method for quantifying the trade-offs from data triplets numerically. U-FaTE is an end-to-end model that adopts a statistical dependence measure as a proxy for utility and fairness and optimizes their weighted linear combination. U-FaTE can be flexibly adapted to estimate both the DST and LST from a finite labeled dataset. + +Usefulness of Trade-Offs (§5). The trade-offs illuminate the fundamental limits of learning algorithms in mitigating unfairness and present a new avenue to evaluate a given image representation in terms of its distance from the estimated trade-offs. We adopt this scheme to evaluate the representations of over 900 supervised and 100 zero-shot publicly available pre-trained models, derive insights, and identify trends and models that are close and far from the empirical trade-off estimates (§6). + +Notation: We denote scalars using lowercase letters, e.g., $d$ and $\lambda$ . We denote deterministic vectors by boldface lowercase letters, e.g., $x$ , $y$ . Both scalar-valued and multidimensional random variables (RV)s are denoted by regular upper + +case letters, e.g., $X$ , $Y$ . We denote deterministic matrices by boldface upper case letters, e.g., $\pmb{K}$ , $\Theta$ . Finite or infinite sets are denoted by calligraphic letters, e.g., $\mathcal{A}$ , $\mathcal{H}$ . + +# 2. Related Works + +A vast majority of prior work on designing fair classifiers focused primarily on uncovering disparities in practical tasks [5, 31] and learning a fair predictor [20, 32, 35] for a given fairness measure. An extended discussion of this body of work can be found in the supplementary material. + +Utility-Fairness Trade-Offs: Many attempts on learning fair models [10, 20, 31, 32] ignored the existence of trade-offs. They sought to maximize accuracy on target tasks while minimizing unfairness, thus perhaps seeking an infeasible solution. Most studies on utility-fairness trade-offs are theoretical and under restricted settings in terms of the type of labels, notion of fairness, and bounds or extreme limits of trade-offs. For example, Zhao et al. [38] obtained a lower bound on DST when both $Y$ and $S$ are binary labels. McNamara et al. [21] provided both upper and lower bounds for binary labels. Only a couple of attempts [27, 29] have been made to numerically estimate utility-fairness trade-offs for independence related-based measures like demographic parity, both of them on features from pre-trained models, rather than raw data. Sadeghi et al. [27] obtained a simplified version of DST, but for linear models. Later on, in the context of invariant representation learning, this was extended to estimate a near-optimal DST-like trade-off called $\mathcal{T}_{\mathrm{Opt}}$ in [29]. But as we demonstrate in §6.2, the estimate of $\mathcal{T}_{\mathrm{Opt}}$ called K- $\mathcal{T}_{\mathrm{Opt}}$ does not span the entire trade-off. + +In contrast to this body of work, we identify two types of trade-offs, DST and LST, and propose a method to numerically quantify them from data. Our trade-offs and their empirical estimates apply to a wide range of prediction tasks for two different categories of fairness notions without any restrictions on the type of labels. + +Learning Fair Classifiers: Over the last decade, many methods have been developed for learning fair classifiers. These approaches follow the template of adopting a fairness constraint as a regularizer in addition to the objective for the target task. The approaches differ in the choice of measure as a proxy for quantifying the level of unfairness between the target label $Y$ and the prediction $\hat{Y}$ , and the associated optimization technique. From an optimization perspective, they can be classified into three major categories-i.e., iterative adversarial methods (ARL[34], SARL[27], and MaxEnt-ARL[26]), non-iterative adversarial methods (FairHSIC [23], OptNet-ARL [28]), and closed-form solver methods (SARL [27], K- $\mathcal{T}_{\mathrm{Opt}}$ [29], LEACE [3], FairerCLIP [7]). Among these, ARL, SARL, MaxEnt-ARL, and OptNet-ARL measure mean dependence [1, 12], FairHSIC, FairerCLIP and K- $\mathcal{T}_{\mathrm{Opt}}$ + +measure full statistical dependence, i.e., all modes of dependence, and SARL and LEACE measures linear dependence. + +U-FaTE draws inspiration from $\mathrm{K - T_{Opt}}$ [29]. By using a closed-form solver and a universal dependence measure that captures all non-linear dependencies, $\mathrm{K - T_{Opt}}$ achieves a better utility-fairness trade-off and is more stable than the other fair learning methods discussed above. However, $\mathrm{K - T_{Opt}}$ is limited in multiple respects and cannot be directly employed for estimating the trade-offs. 1) It operates on features and does not generalize to learning directly from high-dimensional raw data representations such as pixels for images. 2) $\mathrm{K - T_{Opt}}$ optimizes an unconditional dependence measure, which limits its applicability to independence relation-based fairness definitions such as demographic parity. 3) As we demonstrate in §6.2, for demographic parity, $\mathrm{K - T_{Opt}}$ 's trade-off is the closest to our DST estimate, but it does not span the entire utility-fairness trade-off front. Therefore, we adopt the positive aspects of $\mathrm{K - T_{Opt}}$ , namely universal dependence measure and closed-form solver, into U-FaTE and overcome its drawbacks. + +# 3. The Utility-Fairness Trade-Offs + +Fairness Notions: Group fairness notions are typically categorized into three classes [2], namely independence, separation, and sufficiency, each corresponding to different societal desiderata. We focus on the independence and separation relations, which can be expressed as independence $(\hat{Y} \perp S)$ and conditional independence $(\hat{Y} \perp S|Y = y)$ relations, respectively. + +We consider frequently used fairness criteria including Demographic Parity (DP) [17] which is an example of an independence relation, and Equalized Opportunity (EO) [14], and Equality of Odds (EOO) [14] which are both examples of separation relations. The corresponding unfairness metrics are Demographic Parity Violation $DPV \coloneqq |P(\hat{Y} = 1|S = 0) - P(\hat{Y} = 1|S = 1)|$ , Equalized Opportunity Difference $EOD \coloneqq |P(\hat{Y} = 1|Y = 1,S = 0) - P(\hat{Y} = 1|Y = 1,S = 1)|$ , and Equality of Odds Difference $EOOD \coloneqq \frac{1}{2}\sum_{y\in \{0,1\}}|P(\hat{Y} = 1|Y = y,S = 0) - P(\hat{Y} = 1|Y = y,S = 1)|$ , respectively. + +We now introduce the Data-Space Trade-Off and the Label-Space Trade-Off. In both, we employ a dependence measure $\mathrm{Dep}(\cdot ,\cdot |\cdot)$ to enforce the independence and separation based fairness constraints. The function $\mathrm{Dep}(\cdot ,\cdot |\cdot)\geq 0$ is a parametric or non-parametric measure of statistical dependence. $\mathrm{Dep}(P,Q|R = r) = 0$ implies that conditioned on $R = r$ , the random variables (RVs) $P$ and $Q$ are independent. $\mathrm{Dep}(P,Q|R = r) > 0$ means that conditioned on $R = r$ , $P$ , and $Q$ are dependent, with larger values indicating larger degrees of dependence. When $r$ is the empty set $\emptyset$ , we assume that $\mathrm{Dep}(P,Q|R = \emptyset)$ simply reduces to the unconditional dependence $\mathrm{Dep}(P,Q)$ . + +# Definition 1. Data Space Trade-Off (DST) + +$$ +\begin{array}{l} f _ {\lambda} ^ {D S T} := \underset {f \in \mathcal {H} _ {X}} {\arg \inf } \left\{\left(1 - \lambda\right) \underset {g _ {Y} \in \mathcal {H} _ {Y}} {\inf } \mathbb {E} _ {X, Y} \left[ \mathcal {L} _ {Y} \left(g _ {Y} \left(\boldsymbol {f} (X)\right), Y\right) \right] \right. \\ \left. + \lambda \operatorname {D e p} \left(\boldsymbol {f} (X), S | Y = y\right) \right\}, \quad 0 \leq \lambda < 1 \\ \end{array} +$$ + +Here $f$ is the encoder that maps data $X$ to a representation $Z$ , and $g$ is a classifier that predicts $\hat{Y}$ from $Z$ . $\mathcal{H}_X$ and $\mathcal{H}_Y$ are the hypothesis classes of functions for $f$ and $g$ respectively. $\mathcal{L}_Y(\cdot, \cdot)$ is the loss function corresponding to the utility, and $\lambda$ controls the trade-off between utility and fairness, i.e., $\lambda = 0$ corresponds to ignoring the fairness constraint and only optimizing the utility, while, $\lambda \rightarrow 1$ corresponds to the total fairness. The outcome $f_{\lambda}^{DST}$ corresponds to the encoder for a given value of $\lambda$ . This definition corresponds to the DST curve in Fig. 1a, where the utility-fairness plane below the DST corresponds to the region achievable by algorithms designed for this prediction task that learn from the data triplet $(X, Y, S) \sim p(X, Y, S)$ . + +# Definition 2. Label Space Trade-Off (LST) + +$$ +\begin{array}{l} Z _ {\lambda} ^ {L S T} := \underset {Z \in L ^ {2}} {\arg \inf } \left\{\left(1 - \lambda\right) \underset {g _ {Y} \in \mathcal {H} _ {Y}} {\inf } \mathbb {E} _ {Y} \left[ \mathcal {L} _ {Y} \left(g _ {Y} (Z), Y\right) \right] \right. \\ \left. + \lambda \operatorname {D e p} \left(Z, S \mid Y = y\right) \right\}, \quad 0 \leq \lambda < 1 \\ \end{array} +$$ + +Here $L^2$ is the space of all square-integrable RVs (i.e. $\mathbb{E}_Z\left[\| Z\|^2\right] < \infty$ ) in the probability space generated by the joint RV $(Y, S)$ . LST corresponds to the trade-off from an ideal representation space $Z_{\lambda}^{LST}$ that is not constrained to be learned from the input data $X$ . It is the trade-off inherent to the task itself and is the best that any algorithm can hope to achieve for this task. Therefore, it necessarily dominates (or is equal to) DST in Definition 1. This definition corresponds to the LST curve in Fig. 1a, where the utility-fairness plane above the LST corresponds to the region that any algorithm cannot achieve. + +We stress that the above trade-offs are intrinsic to the underlying data, specifically the underlying distributions that generated that data. So, the trade-offs are a property of the data, not of any particular learning algorithm. + +The LST and DST Divide: As illustrated by the yellow region in Fig. 1a, there is a potential gap between LST and DST. This gap at $\lambda = 0$ stems from the irreducible error from the prediction $\mathbb{E}[Y|X]$ of a Bayes Classifier, or when $Y$ is fully recoverable from $X$ . And, when $\lambda > 0$ , the gap between LST and DST widens in two scenarios: 1) $Y \not\perp S$ : The model starts discarding $S$ from the representation $Z$ , which will lead to $Y$ being even less recoverable from $X$ compared to $\lambda = 0$ ; and 2) $Y \perp S$ : If $X$ entangles $Y$ and $S$ in such a way that $Y$ is not recoverable from $X$ when $S$ is discarded, it will lead to $Y$ being even less recoverable from $X$ compared to $\lambda = 0$ . + +![](images/6158bb99a7fc1f4c25e651d7eb8bae7a3ff70d4d11460a0d8343ee273e8f7cd4.jpg) +Figure 2. Overview of U-FaTE: (Left) It comprises two components, a feature extractor and a fair encoder, that are trained end-to-end. Once U-FaTE is trained, the MLP classifier is trained to predict $Y$ from which fairness metrics can be computed. (Right) The fair encoder parameters are optimized through a closed-form solver operating on the features from the feature extractor. See text for more details. + +# 4. Numerically Quantifying the Trade-Offs + +Now, we turn to the second goal of this paper, numerically quantifying the trade-offs from data. Fig. 2 shows a high-level overview of U-FaTE to learn a fair representation for a given trade-off parameter $\lambda$ . U-FaTE comprises a feature extractor and a fair encoder. It receives raw data as input and uses a feature extractor to provide features for the fair encoder. The encoder uses the extracted features and employs a closed-form solver to find the optimum function that maps these features to a new feature space that minimizes the dependency on the sensitive attribute while maximizing the dependency on the target attribute. Following this, to predict the target $Y$ , a classifier is trained with the standard cross-entropy loss for classification problems. This process is repeated for multiple values of $\lambda$ with $0 \leq \lambda < 1$ to obtain the full trade-off curves. + +# 4.1. Problem Setup + +We start from Definition 1 and model the function $\pmb{f}$ as a composition $f_{FE} \circ f_{Enc}$ of the feature extractor and a fair encoder i.e., $\pmb{f}(X; \Theta) = f_{Enc}(f_{FE}(X; \Theta_{FE}); \Theta_{Enc})$ . We parameterize $\pmb{f}$ with $\Theta = [\Theta_{FE}; \Theta_{Enc}]$ where $\Theta_{FE}$ are the parameters of $f_{FE}$ and $\Theta_{Enc}$ are the parameters of $f_{Enc}$ . The objective function in Definition 1 is now + +$$ +\begin{array}{l} \min _ {\boldsymbol {\Theta}} \left\{\left(1 - \lambda\right) \inf _ {\boldsymbol {\Theta} _ {Y}} \mathbb {E} _ {X, Y} \left[ L _ {Y} \left(g _ {Y} \left(\boldsymbol {f} (X; \boldsymbol {\Theta}); \boldsymbol {\Theta} _ {Y}\right), Y\right) \right] \right. \\ \left. + \lambda \operatorname {D e p} \left(\boldsymbol {f} \left(X _ {c}; \boldsymbol {\Theta}\right), S _ {c}\right) \right\}, \quad 0 \leq \lambda < 1. \tag {1} \\ \end{array} +$$ + +where $\operatorname{Dep}(\pmb{f}(X_c; \Theta), S_c)$ is equivalent to the term $\operatorname{Dep}(\pmb{f}(X), S|Y = y)$ in Definition 1 when $Y$ is not a continuous label. In this case, $X_c \sim P(X|Y = y)$ and $S_c \sim P(S|Y = y)$ are the random variables that represent the data and sensitive attribute conditioned on $Y = y$ , respectively. The fair representation is $Z = \pmb{f}(X; \Theta)$ . + +# 4.2. Optimization via Dependence Measures + +The formulation in (1) can be directly optimized for an appropriate choice of dependence measure. Different choices of Dep lead to different fair representation learning methods. For instance, measuring Dep through an adversary + +leads to the class of adversarial representation learning (ARL) methods [26-28, 34]. Similarly, employing the Hilbert Schmidt Independence Criterion (HSIC) [13] as Dep leads to FairHSIC [23]. However, due to challenges in optimization [26-28] and as we demonstrate in §6.2, these approaches are either very unstable, fail to span the trade-off or lead to sub-optimal trade-offs. + +Recently, Sadeghi et al. [29] demonstrated that adopting an HSIC-like dependence measure for the fairness objective and the target loss leads to a closed-form solution that is both efficient and effective at finding a near-optimal trade-off. Therefore, we incorporate the HSIC-like dependence measure and the closed-form solver into U-FaTE. Thus (1) can be expressed as, + +$$ +\begin{array}{l} \sup _ {\boldsymbol {f} \in \mathcal {A} _ {r}} \left\{ \begin{array}{l l} (1 - \lambda) \operatorname {D e p} (\boldsymbol {f} (X; \boldsymbol {\Theta}), Y) \end{array} \right. \\ \left. - \lambda \operatorname {D e p} \left(\boldsymbol {f} \left(X _ {c}; \boldsymbol {\Theta}\right), S _ {c}\right) \right\}, \tag {2} \\ \end{array} +$$ + +where $\mathcal{A}_r$ is a function space that encourages the representations to be uncorrelated. It does not affect the optimality of the learned encoder [29] and improves the compactness of representation [4]. Note that while the first term involves all data $X$ , the second involves the conditional data $X_c$ . + +Choice of Dependence Measure: We adapt the dependence measure from [29] since it lends itself to a closed-form solution while capturing linear and non-linear dependencies under mild assumptions. While the dependence measure in [29] has been defined for absolute independence, our formulation in (2) also requires conditional independence to be compatible with separation based fairness definitions. Therefore, when $Y$ is not a continuous label, we define the conditional dependence measure as, + +$$ +\begin{array}{l} \operatorname {D e p} (\boldsymbol {f} (X), S | Y = y) := \\ \sum_ {j = 1} ^ {r} \sum_ {\beta_ {S} \in \mathcal {U} _ {S}} \mathbb {E} \left[ \left(f _ {j} \left(X _ {c}\right) - \mathbb {E} f _ {j} \left(X _ {c}\right)\right) \left(\beta_ {S} \left(S _ {c}\right) - \mathbb {E} \beta_ {S} \left(S _ {c}\right)\right) \right] \tag {3} \\ \end{array} +$$ + +where $\mathcal{U}_S$ is a countable orthonormal basis set for the separable universal RKHS $\mathcal{H}_S$ and $X_{c}\sim P(X|Y = y)$ and + +$S_{c}\sim P(S|Y = y)$ are data and sensitive attributes, respectively. Empirically it can be estimated as, + +$$ +\operatorname {D e p} (\boldsymbol {f} (X), S | Y = y) := \frac {1}{n ^ {2}} \| \boldsymbol {\Theta} \boldsymbol {K} _ {X _ {c}} \boldsymbol {H} \boldsymbol {L} _ {S _ {c}} \| _ {F} ^ {2}, \tag {4} +$$ + +where $n$ is the number of data samples, $\pmb{K}_{X_c} \in \mathbb{R}^{n \times n}$ is the Gram matrix corresponding to $\mathcal{H}_X$ , $\Theta$ is the encoder parameter in $\pmb{f}(X) = \Theta[k_{X_1}, k_{X_2}, \dots, k_{X_n}]^T$ , $\pmb{H} = \pmb{I}_n - \frac{1}{n} \pmb{1}_n \pmb{1}_n^T$ is the centering matrix, and $\pmb{L}_{S_c}$ is a full column-rank matrix such that $\pmb{L}_{S_c} \pmb{L}_{S_c}^T = \pmb{K}_{S_c}$ (Cholesky factorization). + +# 4.3. A Solution to the Optimization Problem + +Closed-Form Solver via Functions in RKHSs: Directly solving for all the parameters $\Theta$ through (2) and (4) leads to abysmal performance in practice since the kernel $K_{X}$ has to be computed over the raw data space. Therefore, we instead define the fair encoder on the co-domain of the feature extractor $\pmb{f}(\cdot ;\Theta_{FE})$ . So, in this case, (2) reduces to, + +$$ +\begin{array}{l} \sup _ {\boldsymbol {f} _ {E n c} \in \mathcal {A} _ {r}} \left\{ \begin{array}{l} (1 - \lambda) \operatorname {D e p} \left(\boldsymbol {f} _ {E n c} (\tilde {X}; \boldsymbol {\Theta} _ {E n c}), Y\right) \end{array} \right. \\ \left. - \lambda \operatorname {D e p} \left(\boldsymbol {f} _ {E n c} \left(\tilde {X} _ {c}; \boldsymbol {\Theta} _ {E n c}\right), S _ {c}\right) \right\}, \tag {5} \\ \end{array} +$$ + +where $\tilde{X} = f(X;\Theta_{FE})$ , and the first and second terms are $\frac{1}{n^2}\| \Theta_{Enc}K_{\tilde{X}}HL_Y\| _F^2$ and $\frac{1}{n^2}\left\| \Theta_{Enc}K_{\tilde{X}_c}HL_{S_c}\right\| _F^2$ , respectively. The parameters $\Theta_{Enc}$ can now be solved exactly via a closed-form solution: + +Theorem 1. A global optimizer of (5) is + +$$ +\pmb {f} _ {\mathcal {H} _ {\tilde {X}}} ^ {o p t} (\tilde {X}; \pmb {\Theta} _ {E n c}) = \pmb {\Theta} _ {E n c} ^ {o p t} \left[ k _ {\tilde {X}} (\tilde {\pmb {x}} _ {1}, \tilde {X}), \dots , k _ {\tilde {X}} (\tilde {\pmb {x}} _ {n}, \tilde {X}) \right] ^ {T} +$$ + +where $\Theta_{Enc}^{opt} = \pmb{U}^T\pmb{L}_{\tilde{X}}^\dagger \in \mathbb{R}^{r\times n}$ and the columns of $\pmb{U}$ are eigenvectors corresponding to the $r$ largest eigenvalues of the following generalized eigenvalue problem. + +$$ +\begin{array}{l} \left((1 - \lambda) \boldsymbol {L} _ {\tilde {X}} ^ {T} \boldsymbol {H} \boldsymbol {K} _ {Y} \boldsymbol {H} \boldsymbol {L} _ {\tilde {X}} - \lambda \boldsymbol {L} _ {\tilde {X} _ {c}} ^ {T} \boldsymbol {H} \boldsymbol {K} _ {S _ {c}} \boldsymbol {H} \boldsymbol {L} _ {\tilde {X} _ {c}}\right) \boldsymbol {u} \\ = \lambda \left(\frac {1}{n} \boldsymbol {L} _ {\tilde {X}} ^ {T} \boldsymbol {H} \boldsymbol {L} _ {\tilde {X}} + \gamma \boldsymbol {I}\right) \boldsymbol {u}. \tag {6} \\ \end{array} +$$ + +Here $\mathbf{L}_{\tilde{X}}\mathbf{L}_{\tilde{X}}^T = \mathbf{K}_{\tilde{X}},\tilde{X}_c\sim p(\tilde{X} |Y = y)$ and $S_{c}\sim$ $p(S|Y = y)$ + +Proof. The objective in (5) reduces to a generalized eigenvalue problem [18] by expressing it as a trace optimization problem. See supplementary for detailed proof. $\square$ + +While this is a general solution to (5), the solution for each group fairness case is detailed in the supplementary. + +Alternating Optimization: Now we present our full algorithm to optimize (2). We adopt standard minibatch to learn + +the feature encoder's parameters $\Theta_{FE}$ and the closed-form solver for the fair encoder parameters $\Theta_{Enc}$ . We optimize them alternatively where in each iteration, we update $\Theta_{Enc}$ while freezing $\Theta_{FE}$ and vice-versa. Specifically, to optimize the fair encoder's parameters $\Theta_{Enc}$ , we extract features from the data using the frozen feature extractor and use the closed-form solution in (5) to update $\Theta_{Enc}$ . Then, we update the feature extractor's parameters $\Theta_{FE}$ through minibatch SGD in (2) while freezing the encoder parameters. We repeat this process for every minibatch iteration. More details and an illustration of this alternating algorithm can be found in the supplementary material. + +# 4.4. Numerically Estimating the LST + +The Label Space Trade-off (LST) arises when the representation $Z$ is not restricted to be a function of $X$ . Following the discussion in the previous subsection, this trade-off can be formulated as, + +$$ +\sup _ {Z \in L _ {r} ^ {2}} \left\{(1 - \lambda) \operatorname {D e p} (Z, Y) - \lambda \operatorname {D e p} (Z, S | Y = y) \right\}, \tag {7} +$$ + +where $L_r^2$ is the space of all RVs of dimension $r$ with finite variance, i.e., $\mathbb{E}_Z\left[\left\| Z - \mathbb{E}[Z]\right\|^2 < \infty\right]$ . From (7), observe that the optimal $Z$ is a function of $\pmb{p}_{Y,S}$ only. Therefore, instead of directly optimizing $Z$ over $L_r^2$ , equivalently, we optimize for $\Theta_{FE}$ and $\Theta_{Enc}$ as + +$$ +\begin{array}{l} \max _ {\boldsymbol {\Theta} _ {F E}, \boldsymbol {\Theta} _ {E n c}} \left\{\left(1 - \lambda\right) \operatorname {D e p} \left(\boldsymbol {f} \left(Y, S; \boldsymbol {\Theta} _ {F E}, \boldsymbol {\Theta} _ {E n c}\right), Y\right) \right. \\ \left. + \lambda \operatorname {D e p} \left(\boldsymbol {f} (Y, S; \Theta_ {F E}, \Theta_ {E n c}), S | Y = y\right) \right\}. \tag {8} \\ \end{array} +$$ + +Here $f$ is a function of the labels $Y$ and $S$ , i.e., the model takes as input $Y$ and $S$ and seeks to remove the information corresponding to $S$ , including that present in $Y$ . In practice, to improve the stability of the optimization and facilitate learning, in addition to $Y$ and $S$ , we also use $X$ . + +# 5. Scheme for Evaluating Representations + +The primary utility of the trade-offs is in illuminating the fundamental limits of learning algorithms in mitigating unfairness and evaluating the effectiveness of learned representations w.r.t. utility and fairness. This includes representations that provide a single solution in the utility-fairness plane and multiple solutions that span the trade-off between utility and fairness. Fairness evaluations in prior literature focused primarily on relative comparisons of the models to each other. Such a comparison, however, precludes an understanding of how far the solution is from the inherent limits of the task. Elucidating and numerically quantifying the inherent trade-offs facilitates such an understanding and drives further algorithmic development. + +A standard way to evaluate bi-objective solutions is to plot them on a 2-D plane, identify non-dominated solutions, + +![](images/f4be1810a3edbf26f89d48fd89eaa3bdd2b4b3daa3b259ddbdf703b1ef54f28f.jpg) +Figure 3. Evaluating Fair Representation Learning Methods: Accuracy versus fairness trade-offs on CelebA (a)-(c) and FolkTable (d)-(f). (a) and (d) show the trade-off for Equalized Opportunity as the fairness constraint. (b) and (e) show the trade-off for Equality of Odds as the fairness constraint, and (c) and (f) show the trade-off for Demographic Parity as the fairness constraint. The solid lines represent the mean accuracy at a given fairness value, and the shaded region shows the uncertainty of the trade-off. Both DST and LST estimates from U-FaTE are stable. Among the FRL methods, $\mathrm{K - T_{Opt}}$ is closest to the DST, while ARL has the most variance. + +or compare their dominance w.r.t. each other. More details can be found in the supplementary material. + +# 6. Experimental Evaluation + +We designed experiments to answer the following: + +1. How far are existing supervised fair representation learning methods from the two trade-offs? (§6.2) +2. How far are zero-shot representations from the two trade-offs? What is the effect of network architecture and pre-training dataset? (§6.3) +3. How far are pre-trained image representations trained in a supervised fashion from the two trade-offs? (§6.4) + +# 6.1. Experimental Setup + +Datasets: We estimate the trade-offs through U-FaTE on an assortment of datasets. 1) CelebA [19] consists of more than 200K face images of celebrities in the wild annotated with 40 binary attributes. 2) FairFace [16] consists of face images from 7 different race groups labeled with race, sex, and age groups. 3) FolkTables [8] is a tabular dataset of individuals from fifty states derived from the US Census. + +For experiments on CelebA, the target attribute is high cheekbones, and the sensitive attribute is a combination of sex and age for a total of four classes (young woman, young man, old woman, and old man). For experiments on the FairFace dataset, sex (binary) is the target attribute with race (7 groups: East Asian, White, Black, Indian, Latino, South Asian, and Middle Eastern) being the sensitive attribute. For the FolkTables dataset, we consider data from Washington State and choose employment status (binary) and age (discrete value between 1 and 96) of the individuals as the target and sensitive attributes, respectively. + +Metrics: In all experiments, the utility is measured via classification accuracy. We consider three group fairness metrics (EOD, EOOD, and DPV). The zero-shot CLIP models are evaluated via cosine similarity. Moreover, on the Fair-Face dataset, we compared the models to the ideal solutions estimated by LST and DST using a weighted normalized + +Euclidean distance (see supplementary for details). We refer to the distance as $\mathrm{Dist}_{LST}$ and $\mathrm{Dist}_{DST}$ , respectively. + +Implementation Details: We use ResNet-18 as the feature extractor for U-FaTE and the FRL methods. The final classifier is a two-layer MLP. We evaluate the representations from pre-trained vision models by learning a logistic classifier. To obtain the trade-offs, we run U-FaTE and other FRL methods for multiple values of $\lambda$ between zero and one, where zero corresponds to no fairness constraint, and one corresponds to only fairness. + +Optimizing U-FaTE: The trade-offs are defined through the dependence terms in (3) which involves an expectation over the joint distribution $p(X, Y, S)$ . However, due to practical considerations, we only have access to a finite set of samples to estimate the trade-offs. Therefore, we estimate the trade-offs using all the samples available in each dataset without splitting it into train, validation, and test sets. This choice ensures that the estimates account for any possible generalization gap between the train and test distributions and identify the best achievable utility-fairness trade-offs. + +# 6.2. Evaluating FRL Methods + +FRL Baselines: We consider a wide range of FRL methods based on adversarial learning (ARL [34] and MaxEnt-ARL [26]), dependence measures (FairHSIC [23], OptNet-ARL [28]), and closed-form solvers $(\mathrm{K - T_{Opt}}$ [29]). + +Results: We estimate the LST and DST through U-FaTE and the trade-offs from the other baselines across various settings and datasets. Figs. 3a to 3c show the trade-offs on the CelebA dataset for EOD, EOOD, and DPV, respectively. Similarly, Figs. 3d to 3f show the trade-offs on the Folktable dataset for EOD, EOOD, and DPV, respectively. In the plots, the solid lines represent the mean, and the light shadows represent the variance of the accuracy for a given fairness value. On FairFace, observe that trade-offs do not exist since, on this task, it is possible to mitigate unfairness without sacrificing accuracy. Hence, we present the results of FRL methods and the estimated LST and DST in Tab. 1. + +
MethodAccuracy (↑)Unfairness (↓)DistDST (↓)DistLST (↓)
EODARL [34]93.391.340.4480.559
FairHSIC [23]91.021.330.4450.557
OptNet-ARL [28]92.941.700.5980.709
U-FaTE-DST96.170.263-0.133
U-FaTE-LST100.00.0--
EOODARL [34]91.603.040.4470.71
FairHSIC [23]93.432.130.2360.498
OptNet-ARL [28]93.392.340.2840.546
U-FaTE-DST97.931.126-0.262
U-FaTE-LST100.00.0--
DPVARL [34]92.496.090.3500.351
FairHSIC [23]91.415.910.3290.332
OptNet-ARL [28]93.335.800.3160.317
U-FaTE-DST94.393.082-0.04
U-FaTE-LST100.03.10--
+ +Table 1. Evaluation of FRL methods on FairFace based on the distance to DST and LST estimated by U-FaTE. Color corresponds to the DST and LST trade-offs. + +Observations: Although $\mathrm{K - T_{Opt}}$ , OptNet-ARL and FairHSIC can achieve near-optimal accuracy in most cases, they are unable to span the whole range of fairness values. ARL is the most unstable but can span the whole range of fairness values. $\mathrm{K - T_{Opt}}$ is the most stable method due to its closed-form solver, but is unable to span the whole range of fairness values (Figs. 3c and 3f). + +The gap between LST and DST demonstrates the information gap in $X$ for predicting $Y$ . From Fig. 3, we observe that the gap is $\sim 20\%$ of accuracy in low fairness regions and $\sim 40\%$ in high fairness regions for EOD and EOOD. For DPV, the trend reverses with a gap of $\sim 20\%$ for low fairness and gradually decreases with increasing unfairness. + +Observe that the LST in Figs. 3a and 3b and Figs. 3d and 3e for EOD and EOOD is almost flat at $100\%$ accuracy. This observation, however, is unsurprising since EOD and EOOD both condition on the label $Y$ , and thus an ideal classifier with $100\%$ accuracy (i.e., $\hat{Y} = Y$ ) will have zero EOD and EOOD. And, in LST, the Oracle classifier is $100\%$ accurate since it has access to $Y$ and $S$ . So, the LST has sufficient information to minimize EOD and EOOD without sacrificing utility. The same, however, does not hold for DPV since it does not consider the target labels in its definition. Based on the above discussion, we deduce that EOD and EOOD are more pragmatic fairness metrics than DPV since they do not force the model to sacrifice predictive accuracy to ensure fairness. Thus, both offer a more balanced and practical approach to measuring fairness. Our empirical results provide independent confirmation of the same observations in [6, 14]. + +The comparison of FRL methods in Tab. 1 based on $\mathrm{Dist}_{LST}$ suggests that when models are optimized for EOD, ARL and FairHSIC find solutions closer to LST and DST than OptNet-ARL. When models are optimized for EOOD, FairHSIC finds the closest point to the LST and DST. OptNet-ARL performs slightly better than the other FRL methods when optimized for reducing DPV. + +# 6.3. Evaluating Zero-Shot CLIP Models + +Zero-Shot Models: To study the fairness of current zero-shot models, we consider more than 100 pre-trained models from OpenCLIP [15] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before. + +Results: Figs. 4a and 4b show results on CelebA and FairFace, respectively, for three group fairness definitions. Each point represents the result of one zero-shot CLIP model, with the color denoting the model's pre-training dataset. Plots also include DST and LST for comparison. + +Observations: From Fig. 4a, we observe that zero-shot models perform poorly, in terms of accuracy, on the CelebA task. We hypothesize that it is so since the target task (predicting high cheekbones) is uncommon, even in the large datasets the models have been trained on. From a fairness perspective, we observe that models trained on CommonPool [9] (red dots) are more likely to be fair, while models trained on DataComp [9] have marginally better accuracy over the other models. Finally, the CLIP models are very far from the DST across the board. + +On FairFace (Fig. 4b), since the target task (sex) is abundantly represented in all large-scale pre-training datasets, we observe that the CLIP models exhibit high levels of accuracy. Two CLIP models pre-trained on OpenAI WIT [24] are the closest to DST and LST w.r.t. EOD. Similar to our observations on CelebA, models pre-trained on the CommonPool dataset are more likely to be fair but at the cost of accuracy. In contrast, models pre-trained on DataComp are more unfair but have greater accuracy. Finally, there is a clear positive correlation between accuracy and unfairness across all fairness metrics on FairFace (Fig. 4b). + +# 6.4. Evaluating Supervised Representations + +Supervised Baselines: To study the fairness of image representations from models pre-trained in a supervised fashion, we consider more than 900 models from Pytorch Image Models [33] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before. + +Results: Figs. 4c and 4d show results on CelebA and Fair-Face datasets, respectively. Each point represents the result of one supervised model, with color denoting the model's pre-training dataset. Plots also include DST and LST for comparison, but some plots were magnified for better resolution. So, the LST may only be partially visible. + +Observations: From the CelebA results in Fig. 4c, we observe a high positive correlation between the accuracy and EOD and EOOD. Thus, models with better accuracy are also more fair. And, in contrast to the zero-shot models, even though high cheekbones is a rare label, the representations have sufficient information to accurately detect it with a logistic regression classifier. Specifically, models pre + +![](images/0a02c2de14caee88fc2cd783dc3499821d69be1240559551202411b3dc781772.jpg) + +![](images/e26458bb9691ed5da278f0e48a7d3df3e94c7c47b2e5a68726c3e41cb6a39647.jpg) +(b) + +![](images/95b0f331c6a56ff7065a57a305ad7e45312f6680cb82cdec8212acfda1bd0ee2.jpg) +(a) +(c) + +![](images/78246bf3607b8d1fc7406e33a3730ded1a5cc1359b6ebfdb248d6ad000d641dd.jpg) +(d) +Figure 4. Evaluating Pre-Trained Zero-Shot and Supervised Models: Accuracy-fairness evaluation of more than 100 pre-trained zero-shot models on CelebA (a) and FairFace (b), and over 900 pre-trained image representations on CelebA (c), and FairFace (d). + +trained on ImageNet22K [25] (orange dots), and OpenAI WIT [24] have the best accuracy and reach the DST tradeoff in high unfairness regions. However, with more than $11\%$ EOD and EOOD and more than $20\%$ DPV, they have significant levels of bias between the two sexes. + +Results on FairFace in Fig. 4d also reiterate that models trained on OpenAI WIT and ImageNet22K are more fair and more accurate than other datasets. We also observe that the models are generally more fair on FairFace than on CelebA. A similar positive correlation exists between accuracy and fairness for EOD and EOOD. We also make an interesting observation from Fig. 4d. Some models surpass the DST and enter the "Possible with Extra Data" region of the utility-fairness plane illustrated in Fig. 1a. Recall that DST estimates the upper bound of the possible region for models trained with the same data as DST. The DST can be surpassed to enter the "Possible with Extra Data" region if the model is trained on additional data beyond what is used for estimating DST. From Fig. 4d (middle), we observe that a few models trained on OpenAI, LAION [30], ImageNet22K, ImageNet21K, and ImageNet12K can surpass the DST, plausibly since these datasets contain sufficient—both quality and quantity—samples from the distribution of the target attribute (sex). + +# 7. Concluding Remarks + +As image classification systems are widely deployed in high-stakes applications, ensuring that their predictions do not exhibit demographic bias is paramount for gaining user trust. While it is desirable to mitigate bias without sacrificing accuracy, this is not always possible. This paper studied such inherent trade-offs between utility and fairness. First, we identified two types of trade-offs called Data-Space and + +Label-Space trade-offs corresponding to those achievable with and without data restrictions. But unlike prior theoretical studies on utility-fairness trade-offs, next, we focused on developing algorithmic tools for quantifying the trade-offs from data and proposed U-FaTE. As an illustration of its practical utility, we estimated the trade-offs on several image classification tasks, facilitating a large-scale evaluation of over 100 zero-shot and 900 supervised pre-trained models. The results revealed that, out of the box, pre-trained models are far from the best achievable limits of accuracy and fairness. Furthermore, we identified that, in some cases, larger datasets can improve accuracy and fairness and surpass the solutions represented by the DST. + +U-FaTE was designed as a composition of a neural network with the last layer optimized to global optimality through a closed-form solver for a given feature representation. This design allowed it to estimate the two trade-offs reliably. However, U-FaTE does not provide convergence or optimality guarantees. Therefore, the estimated DST and LST are likely to be suboptimal. Nonetheless, they can serve as a valuable tool for understanding the nature of the problem (e.g., Does there exist a trade-off? or What are the possible, impossible and possible with extra data regions in the utility-fairness plane?) and how far a given fair learning algorithm is from the achievable limits. Furthermore, depending on which region of the utility-fairness plane a solution is and how far it is from the DST and LST reveals whether to focus on better optimization or better data. + +Acknowledgements: This work was supported by the National Science Foundation (award #2147116). + +# References + +[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 2 +[2] Solon Barocas, Moritz Hardt, and Arvind Narayanan. *Fairness and Machine Learning: Limitations and Opportunities*. MIT Press, 2023. 1, 2, 3 +[3] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. LEACE: Perfect linear concept erasure in closed form. arXiv preprint arXiv:2306.03819, 2023. 2 +[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 4 +[5] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, 2018. 2 +[6] Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. big data 5, 2 (2017), 153-163. arXiv preprint arXiv:1610.07524, 2017. 7 +[7] Sepehr Dehdashtian, Lan Wang, and Vishnu Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In International Conference on Learning Representations, 2024. 2 +[8] Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. Retiring adult: New datasets for fair machine learning. In Advances in Neural Information Processing Systems, 2021. 6 +[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 7 +[10] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating face recognition bias via group adaptive classifier. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[11] Thibaut Le Gouic, Jean-Michel Loubes, and Philippe Rigollet. Projection to fairness in statistical learning. arXiv preprint arXiv:2005.11720, 2020. 1 +[12] Vincent Grari, Oualid El Hajouji, Sylvain Lamprier, and Marcin Detyniecki. Learning unbiased representations via Rényi minimization. arXiv preprint arXiv:2009.03183, 2020. 2 +[13] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, and Bernhard Scholkopf. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12):2075-2129, 2005. 4 +[14] Moritz Hardt, Eric Price, Nati Srebro, et al. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems, 2016. 3, 7 + +[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7 +[16] Kimmo Karkkainen and Jungseock Joo. FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 6 +[17] Niki Kilbertus, Mateo Rojas Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Scholkopf. Avoiding discrimination through causal reasoning. Advances in Neural Information Processing Systems, 2017. 3 +[18] Effrosini Kokiopoulou, Jie Chen, and Yousef Saad. Trace optimization and eigenproblems in dimension reduction methods. Numerical Linear Algebra with Applications, 18(3): 565-602, 2011. 5 +[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In IEEE International Conference on Computer Vision, 2015. 2, 6 +[20] David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Learning adversarially fair and transferable representations. arXiv preprint arXiv:1802.06309, 2018. 2 +[21] Daniel McNamara, Cheng Soon Ong, and Robert C Williamson. Costs and benefits of fair representation learning. AAAI/ACM Conference on AI, Ethics, and Society, 2019. 2 +[22] Aditya Krishna Menon and Robert C Williamson. The cost of fairness in binary classification. Conference on Fairness, Accountability and Transparency, 2018. 1 +[23] Novi Quadrianto, Viktoriia Sharmanska, and Oliver Thomas. Discovering fair representations in the data domain. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6, 7 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 7, 8 +[25] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 8 +[26] Proteek Roy and Vishnu Naresh Boddeti. Mitigating information leakage in image representations: A maximum entropy approach. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6 +[27] Bashir Sadeghi, Runyi Yu, and Vishnu Boddeti. On the global optima of kernelized adversarial representation learning. IEEE International Conference on Computer Vision, 2019. 2 +[28] Bashir Sadeghi, Lan Wang, and Vishnu Naresh Boddeti. Adversarial representation learning with closed-form solvers. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2021. 2, 4, 6, 7 +[29] Bashir Sadeghi, Sepehr Dehdashtian, and Vishnu Boddeti. On characterizing the trade-off in invariant representation + +learning. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 3, 4, 6 +[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 8 +[31] Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations. In IEEE/CVF International Conference on Computer Vision, 2019. 2 +[32] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[33] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.7 +[34] Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. Controllable invariance through adversarial feature learning. Advances in Neural Information Processing Systems, 2017. 2, 4, 6, 7 +[35] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair representations. International Conference on Machine Learning, 2013. 2 +[36] Han Zhao. Costs and benefits of wasserstein fair regression. arXiv preprint arXiv:2106.08812, 2021. 1 +[37] Han Zhao and Geoffrey J Gordon. Inherent trade-offs in learning fair representations. arXiv preprint arXiv:1906.08386, 2019. 1 +[38] Han Zhao, Jianfeng Chi, Yuan Tian, and Geoffrey J Gordon. Trade-offs and guarantees of adversarial representation learning for information obfuscation. arXiv preprint arXiv:1906.07902, 2019. 2 \ No newline at end of file diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/images.zip b/2024/Utility-Fairness Trade-Offs and How to Find Them/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..271151e66cbdbfaa143d7bf8405f69a6308a2460 --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82938714c4d6077b43193fd21ade08280b3608c0ac2725e825aaacd7632bc106 +size 348631 diff --git a/2024/Utility-Fairness Trade-Offs and How to Find Them/layout.json b/2024/Utility-Fairness Trade-Offs and How to Find Them/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..deb5767c3a9d27dd59516730d4ecddb6c9a77e9d --- /dev/null +++ b/2024/Utility-Fairness Trade-Offs and How to Find Them/layout.json @@ -0,0 +1,10517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "type": "text", + "content": "Utility-Fairness Trade-Offs and How to Find Them" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 143, + 455, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 143, + 455, + 172 + ], + "spans": [ + { + "bbox": [ + 140, + 143, + 455, + 172 + ], + "type": "text", + "content": "Sepehr Dehdashtian Bashir Sadeghi Vishnu Naresh Boddeti Michigan State University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 174, + 388, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 174, + 388, + 186 + ], + "spans": [ + { + "bbox": [ + 203, + 174, + 388, + 186 + ], + "type": "text", + "content": "{sepehr, sadeghib, vishnu}@msu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 238, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 526 + ], + "type": "text", + "content": "When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated and achievable fairness-utility trade-offs across multiple datasets and prediction tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "content": "As learning-based systems are increasingly being deployed in high-stakes applications, there is a dire need to ensure that they do not propagate or amplify any discriminative tendencies inherent in the training datasets. An ideal solution would impart fairness to prediction models while retaining the performance of the same model when learned without fairness considerations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "Realizing this goal necessitates optimizing two objectives: maximizing utility in predicting a label " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " for a target task (e.g., face identity) while minimizing the unfairness w.r.t. a demographic attribute " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " (e.g., age or gender). However, when the statistical dependence between " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 408, + 315 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 408, + 315 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 408, + 315 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 408, + 315 + ], + "type": "image", + "image_path": "7e7bd30e645e7d173a7dc2a8924480d38c6247e42119da4cff837d352ef6f517.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 319, + 368, + 327 + ], + "lines": [ + { + "bbox": [ + 357, + 319, + 368, + 327 + ], + "spans": [ + { + "bbox": [ + 357, + 319, + 368, + 327 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "type": "text", + "content": "Figure 1. The utility-fairness trade-offs. (a) Classification systems can be evaluated by their utility (e.g., accuracy) w.r.t. a target label " + }, + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "type": "text", + "content": " and their unfairness w.r.t. a demographic label " + }, + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 331, + 547, + 408 + ], + "type": "text", + "content": ". We introduce two trade-offs, Data Space Trade-Off (DST) and Label Space Trade-Off (LST). (b) We empirically estimate DST and LST on CelebA and evaluate the utility (high cheekbones) and fairness (gender & age) of over 100 zero-shot and 900 supervised models." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 413, + 215, + 536, + 312 + ], + "blocks": [ + { + "bbox": [ + 413, + 215, + 536, + 312 + ], + "lines": [ + { + "bbox": [ + 413, + 215, + 536, + 312 + ], + "spans": [ + { + "bbox": [ + 413, + 215, + 536, + 312 + ], + "type": "image", + "image_path": "500b552bf89e130a4a75990ced3b0154c9f208c44ee0cca63507cce7732e3e48.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 316, + 486, + 325 + ], + "lines": [ + { + "bbox": [ + 474, + 316, + 486, + 325 + ], + "spans": [ + { + "bbox": [ + 474, + 316, + 486, + 325 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 447 + ], + "type": "text", + "content": "is not negligible, learning with fairness considerations will necessarily degrade the performance of the target predictor, i.e., a trade-off will exist between utility and fairness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 447, + 548, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 548, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 548, + 626 + ], + "type": "text", + "content": "The existence of a utility-fairness trade-off has been well established, theoretically [11, 22, 29, 36, 37] and empirically [29], in multiple prior works. However, the focus of this body of work has been limited in multiple respects. First, prior work [29, 36] focused on just one type of trade-off, ignoring other possible trade-offs between utility and fairness. Second, prior work [36, 37] focused on establishing bounds or identifying the end-points of the trade-off of interest rather than attempting its precise characterization. Third, the majority of the prior work [11, 29, 36, 37] has investigated the utility-fairness trade-offs for one definition of group fairness, namely, demographic parity (DP). There are multiple fairness definitions [2], including those more practically relevant than DP, such as Equalized Opportunity (EO), for which the trade-offs have not been studied." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 626, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 626, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 626, + 547, + 651 + ], + "type": "text", + "content": "Despite these attempts, several questions related to the utility-fairness trade-offs remain outstanding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 659, + 542, + 707 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 310, + 659, + 517, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 659, + 517, + 671 + ], + "spans": [ + { + "bbox": [ + 310, + 659, + 517, + 671 + ], + "type": "text", + "content": "1. What are the optimal utility-fairness trade-offs?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 671, + 542, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 671, + 542, + 707 + ], + "spans": [ + { + "bbox": [ + 310, + 671, + 542, + 707 + ], + "type": "text", + "content": "2. For a given prediction task and a demographic attribute, we wish to be fair w.r.t., how can we empirically estimate the trade-offs from data?" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12037" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "Addressing these questions by identifying and quantifying the trade-offs is the primary goal of this paper. The trade-offs are a function of the data triplet " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "(X,Y,S)" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " is the data (e.g., images), " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " is the target label, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " is the sensitive demographic label. Figure 1 illustrates the plausible trade-offs, their empirical estimation on CelebA [19], and their utility in empirically evaluating representations from pre-trained models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "content": "Identifying Trade-Offs (§3). We identify two trade-offs: the Label-Space Trade-Off (LST) and Data-Space Trade-Off (DST). They can be defined for any group fairness definitions that can be expressed via independence and separation relations [2]. The LST corresponds to the trade-off obtained by an oracle fair classifier that depends only on the distributions of " + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "content": ". Similarly, DST is the trade-off obtained by an optimally learned fair classifier and depends on " + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "inline_equation", + "content": "(X,Y,S)" + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "content": ". By definition, LST necessarily dominates DST since it does not depend on the data " + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 172, + 289, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "text", + "content": "The trade-offs divide the utility-fairness plane into three regions shown in Fig. 1a. A possible region that can be attained by algorithms learned on the given data, a possible with extra data region that can be attained by learning on data beyond the given data, and an impossible region that cannot be attained by any algorithmic scheme due to the inherent dependence between the distributions of " + }, + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 293, + 289, + 377 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 381, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 289, + 548 + ], + "type": "text", + "content": "Quantifying Trade-Offs (\\$4). Characterizing the exact trade-offs from data for a given task, demographic attribute, and fairness definition affords multiple benefits. It will allow researchers and practitioners to identify the achievable solution space for the task, evaluate how far a given predictor is from the optimal solution, and identify performance gaps and trends among existing solutions. To this end, we propose U-FaTE (Utility-Fairness Trade-Off Estimator), a method for quantifying the trade-offs from data triplets numerically. U-FaTE is an end-to-end model that adopts a statistical dependence measure as a proxy for utility and fairness and optimizes their weighted linear combination. U-FaTE can be flexibly adapted to estimate both the DST and LST from a finite labeled dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 553, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 289, + 662 + ], + "type": "text", + "content": "Usefulness of Trade-Offs (§5). The trade-offs illuminate the fundamental limits of learning algorithms in mitigating unfairness and present a new avenue to evaluate a given image representation in terms of its distance from the estimated trade-offs. We adopt this scheme to evaluate the representations of over 900 supervised and 100 zero-shot publicly available pre-trained models, derive insights, and identify trends and models that are close and far from the empirical trade-off estimates (§6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Notation: We denote scalars using lowercase letters, e.g., " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": ". We denote deterministic vectors by boldface lowercase letters, e.g., " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": ". Both scalar-valued and multidimensional random variables (RV)s are denoted by regular upper" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "case letters, e.g., " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ". We denote deterministic matrices by boldface upper case letters, e.g., " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ". Finite or infinite sets are denoted by calligraphic letters, e.g., " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 121, + 398, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 398, + 134 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 398, + 134 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "type": "text", + "content": "A vast majority of prior work on designing fair classifiers focused primarily on uncovering disparities in practical tasks [5, 31] and learning a fair predictor [20, 32, 35] for a given fairness measure. An extended discussion of this body of work can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": "Utility-Fairness Trade-Offs: Many attempts on learning fair models [10, 20, 31, 32] ignored the existence of trade-offs. They sought to maximize accuracy on target tasks while minimizing unfairness, thus perhaps seeking an infeasible solution. Most studies on utility-fairness trade-offs are theoretical and under restricted settings in terms of the type of labels, notion of fairness, and bounds or extreme limits of trade-offs. For example, Zhao et al. [38] obtained a lower bound on DST when both " + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": " are binary labels. McNamara et al. [21] provided both upper and lower bounds for binary labels. Only a couple of attempts [27, 29] have been made to numerically estimate utility-fairness trade-offs for independence related-based measures like demographic parity, both of them on features from pre-trained models, rather than raw data. Sadeghi et al. [27] obtained a simplified version of DST, but for linear models. Later on, in the context of invariant representation learning, this was extended to estimate a near-optimal DST-like trade-off called " + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Opt}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": " in [29]. But as we demonstrate in §6.2, the estimate of " + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Opt}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": " called K- " + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Opt}}" + }, + { + "bbox": [ + 304, + 206, + 546, + 445 + ], + "type": "text", + "content": " does not span the entire trade-off." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 519 + ], + "type": "text", + "content": "In contrast to this body of work, we identify two types of trade-offs, DST and LST, and propose a method to numerically quantify them from data. Our trade-offs and their empirical estimates apply to a wide range of prediction tasks for two different categories of fairness notions without any restrictions on the type of labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": "Learning Fair Classifiers: Over the last decade, many methods have been developed for learning fair classifiers. These approaches follow the template of adopting a fairness constraint as a regularizer in addition to the objective for the target task. The approaches differ in the choice of measure as a proxy for quantifying the level of unfairness between the target label " + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": " and the prediction " + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": ", and the associated optimization technique. From an optimization perspective, they can be classified into three major categories-i.e., iterative adversarial methods (ARL[34], SARL[27], and MaxEnt-ARL[26]), non-iterative adversarial methods (FairHSIC [23], OptNet-ARL [28]), and closed-form solver methods (SARL [27], K- " + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Opt}}" + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": " [29], LEACE [3], FairerCLIP [7]). Among these, ARL, SARL, MaxEnt-ARL, and OptNet-ARL measure mean dependence [1, 12], FairHSIC, FairerCLIP and K- " + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Opt}}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12038" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "measure full statistical dependence, i.e., all modes of dependence, and SARL and LEACE measures linear dependence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": "U-FaTE draws inspiration from " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": " [29]. By using a closed-form solver and a universal dependence measure that captures all non-linear dependencies, " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": " achieves a better utility-fairness trade-off and is more stable than the other fair learning methods discussed above. However, " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": " is limited in multiple respects and cannot be directly employed for estimating the trade-offs. 1) It operates on features and does not generalize to learning directly from high-dimensional raw data representations such as pixels for images. 2) " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": " optimizes an unconditional dependence measure, which limits its applicability to independence relation-based fairness definitions such as demographic parity. 3) As we demonstrate in §6.2, for demographic parity, " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": " 's trade-off is the closest to our DST estimate, but it does not span the entire utility-fairness trade-off front. Therefore, we adopt the positive aspects of " + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 97, + 288, + 312 + ], + "type": "text", + "content": ", namely universal dependence measure and closed-form solver, into U-FaTE and overcome its drawbacks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 327, + 224, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 224, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 224, + 341 + ], + "type": "text", + "content": "3. The Utility-Fairness Trade-Offs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "content": "Fairness Notions: Group fairness notions are typically categorized into three classes [2], namely independence, separation, and sufficiency, each corresponding to different societal desiderata. We focus on the independence and separation relations, which can be expressed as independence " + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "inline_equation", + "content": "(\\hat{Y} \\perp S)" + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "content": " and conditional independence " + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "inline_equation", + "content": "(\\hat{Y} \\perp S|Y = y)" + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "content": " relations, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "text", + "content": "We consider frequently used fairness criteria including Demographic Parity (DP) [17] which is an example of an independence relation, and Equalized Opportunity (EO) [14], and Equality of Odds (EOO) [14] which are both examples of separation relations. The corresponding unfairness metrics are Demographic Parity Violation " + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "inline_equation", + "content": "DPV \\coloneqq |P(\\hat{Y} = 1|S = 0) - P(\\hat{Y} = 1|S = 1)|" + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "text", + "content": ", Equalized Opportunity Difference " + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "inline_equation", + "content": "EOD \\coloneqq |P(\\hat{Y} = 1|Y = 1,S = 0) - P(\\hat{Y} = 1|Y = 1,S = 1)|" + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "text", + "content": ", and Equality of Odds Difference " + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "inline_equation", + "content": "EOOD \\coloneqq \\frac{1}{2}\\sum_{y\\in \\{0,1\\}}|P(\\hat{Y} = 1|Y = y,S = 0) - P(\\hat{Y} = 1|Y = y,S = 1)|" + }, + { + "bbox": [ + 46, + 434, + 287, + 568 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "We now introduce the Data-Space Trade-Off and the Label-Space Trade-Off. In both, we employ a dependence measure " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " to enforce the independence and separation based fairness constraints. The function " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(\\cdot ,\\cdot |\\cdot)\\geq 0" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " is a parametric or non-parametric measure of statistical dependence. " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(P,Q|R = r) = 0" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " implies that conditioned on " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "R = r" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", the random variables (RVs) " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " are independent. " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(P,Q|R = r) > 0" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " means that conditioned on " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "R = r" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " are dependent, with larger values indicating larger degrees of dependence. When " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " is the empty set " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": ", we assume that " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(P,Q|R = \\emptyset)" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " simply reduces to the unconditional dependence " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Dep}(P,Q)" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 73, + 488, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 73, + 488, + 85 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 488, + 85 + ], + "type": "text", + "content": "Definition 1. Data Space Trade-Off (DST)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 96, + 534, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 96, + 534, + 136 + ], + "spans": [ + { + "bbox": [ + 310, + 96, + 534, + 136 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {\\lambda} ^ {D S T} := \\underset {f \\in \\mathcal {H} _ {X}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {X, Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X)\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (X), S | Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}", + "image_path": "876a6c768c64acfdbd51344e9585242452c7ef7593482e600ded3bf86ff042a3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " is the encoder that maps data " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " to a representation " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " is a classifier that predicts " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_X" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_Y" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " are the hypothesis classes of functions for " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_Y(\\cdot, \\cdot)" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " is the loss function corresponding to the utility, and " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " controls the trade-off between utility and fairness, i.e., " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " corresponds to ignoring the fairness constraint and only optimizing the utility, while, " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\lambda \\rightarrow 1" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " corresponds to the total fairness. The outcome " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "f_{\\lambda}^{DST}" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": " corresponds to the encoder for a given value of " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": ". This definition corresponds to the DST curve in Fig. 1a, where the utility-fairness plane below the DST corresponds to the region achievable by algorithms designed for this prediction task that learn from the data triplet " + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "inline_equation", + "content": "(X, Y, S) \\sim p(X, Y, S)" + }, + { + "bbox": [ + 305, + 145, + 545, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 311, + 489, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 311, + 489, + 323 + ], + "spans": [ + { + "bbox": [ + 310, + 311, + 489, + 323 + ], + "type": "text", + "content": "Definition 2. Label Space Trade-Off (LST)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 333, + 515, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 515, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 515, + 373 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Z _ {\\lambda} ^ {L S T} := \\underset {Z \\in L ^ {2}} {\\arg \\inf } \\left\\{\\left(1 - \\lambda\\right) \\underset {g _ {Y} \\in \\mathcal {H} _ {Y}} {\\inf } \\mathbb {E} _ {Y} \\left[ \\mathcal {L} _ {Y} \\left(g _ {Y} (Z), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(Z, S \\mid Y = y\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1 \\\\ \\end{array}", + "image_path": "13331f5f817f70df4d237f90d7b65936d0b020a42ce4419e7015b8687cd672eb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": " is the space of all square-integrable RVs (i.e. " + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_Z\\left[\\| Z\\|^2\\right] < \\infty" + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": ") in the probability space generated by the joint RV " + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "(Y, S)" + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": ". LST corresponds to the trade-off from an ideal representation space " + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "Z_{\\lambda}^{LST}" + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": " that is not constrained to be learned from the input data " + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 384, + 545, + 517 + ], + "type": "text", + "content": ". It is the trade-off inherent to the task itself and is the best that any algorithm can hope to achieve for this task. Therefore, it necessarily dominates (or is equal to) DST in Definition 1. This definition corresponds to the LST curve in Fig. 1a, where the utility-fairness plane above the LST corresponds to the region that any algorithm cannot achieve." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 518, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 545, + 567 + ], + "type": "text", + "content": "We stress that the above trade-offs are intrinsic to the underlying data, specifically the underlying distributions that generated that data. So, the trade-offs are a property of the data, not of any particular learning algorithm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": "The LST and DST Divide: As illustrated by the yellow region in Fig. 1a, there is a potential gap between LST and DST. This gap at " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " stems from the irreducible error from the prediction " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[Y|X]" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " of a Bayes Classifier, or when " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " is fully recoverable from " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": ". And, when " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": ", the gap between LST and DST widens in two scenarios: 1) " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y \\not\\perp S" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": ": The model starts discarding " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " from the representation " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": ", which will lead to " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " being even less recoverable from " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": "; and 2) " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y \\perp S" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": ": If " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " entangles " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " in such a way that " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " is not recoverable from " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " is discarded, it will lead to " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " being even less recoverable from " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12039" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 71, + 504, + 159 + ], + "blocks": [ + { + "bbox": [ + 92, + 71, + 504, + 159 + ], + "lines": [ + { + "bbox": [ + 92, + 71, + 504, + 159 + ], + "spans": [ + { + "bbox": [ + 92, + 71, + 504, + 159 + ], + "type": "image", + "image_path": "6158bb99a7fc1f4c25e651d7eb8bae7a3ff70d4d11460a0d8343ee273e8f7cd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 164, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 45, + 164, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 45, + 164, + 547, + 199 + ], + "type": "text", + "content": "Figure 2. Overview of U-FaTE: (Left) It comprises two components, a feature extractor and a fair encoder, that are trained end-to-end. Once U-FaTE is trained, the MLP classifier is trained to predict " + }, + { + "bbox": [ + 45, + 164, + 547, + 199 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 45, + 164, + 547, + 199 + ], + "type": "text", + "content": " from which fairness metrics can be computed. (Right) The fair encoder parameters are optimized through a closed-form solver operating on the features from the feature extractor. See text for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 270, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 270, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 270, + 232 + ], + "type": "text", + "content": "4. Numerically Quantifying the Trade-Offs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "content": "Now, we turn to the second goal of this paper, numerically quantifying the trade-offs from data. Fig. 2 shows a high-level overview of U-FaTE to learn a fair representation for a given trade-off parameter " + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "content": ". U-FaTE comprises a feature extractor and a fair encoder. It receives raw data as input and uses a feature extractor to provide features for the fair encoder. The encoder uses the extracted features and employs a closed-form solver to find the optimum function that maps these features to a new feature space that minimizes the dependency on the sensitive attribute while maximizing the dependency on the target attribute. Following this, to predict the target " + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "content": ", a classifier is trained with the standard cross-entropy loss for classification problems. This process is repeated for multiple values of " + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda < 1" + }, + { + "bbox": [ + 45, + 238, + 290, + 417 + ], + "type": "text", + "content": " to obtain the full trade-off curves." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 140, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 140, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 140, + 437 + ], + "type": "text", + "content": "4.1. Problem Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": "We start from Definition 1 and model the function " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " as a composition " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "f_{FE} \\circ f_{Enc}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " of the feature extractor and a fair encoder i.e., " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\pmb{f}(X; \\Theta) = f_{Enc}(f_{FE}(X; \\Theta_{FE}); \\Theta_{Enc})" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": ". We parameterize " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\Theta = [\\Theta_{FE}; \\Theta_{Enc}]" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\Theta_{FE}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "f_{FE}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "inline_equation", + "content": "f_{Enc}" + }, + { + "bbox": [ + 46, + 441, + 288, + 515 + ], + "type": "text", + "content": ". The objective function in Definition 1 is now" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 520, + 289, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 289, + 564 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {\\Theta}} \\left\\{\\left(1 - \\lambda\\right) \\inf _ {\\boldsymbol {\\Theta} _ {Y}} \\mathbb {E} _ {X, Y} \\left[ L _ {Y} \\left(g _ {Y} \\left(\\boldsymbol {f} (X; \\boldsymbol {\\Theta}); \\boldsymbol {\\Theta} _ {Y}\\right), Y\\right) \\right] \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\quad 0 \\leq \\lambda < 1. \\tag {1} \\\\ \\end{array}", + "image_path": "9553e6db07f5a1a3ec5d6b48752e5bbc4af76c18404b478b915a2a128538643f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\operatorname{Dep}(\\pmb{f}(X_c; \\Theta), S_c)" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": " is equivalent to the term " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\operatorname{Dep}(\\pmb{f}(X), S|Y = y)" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": " in Definition 1 when " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": " is not a continuous label. In this case, " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "X_c \\sim P(X|Y = y)" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "S_c \\sim P(S|Y = y)" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": " are the random variables that represent the data and sensitive attribute conditioned on " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "Y = y" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": ", respectively. The fair representation is " + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "inline_equation", + "content": "Z = \\pmb{f}(X; \\Theta)" + }, + { + "bbox": [ + 46, + 569, + 287, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 647, + 257, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 257, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 257, + 661 + ], + "type": "text", + "content": "4.2. Optimization via Dependence Measures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "The formulation in (1) can be directly optimized for an appropriate choice of dependence measure. Different choices of Dep lead to different fair representation learning methods. For instance, measuring Dep through an adversary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 303 + ], + "type": "text", + "content": "leads to the class of adversarial representation learning (ARL) methods [26-28, 34]. Similarly, employing the Hilbert Schmidt Independence Criterion (HSIC) [13] as Dep leads to FairHSIC [23]. However, due to challenges in optimization [26-28] and as we demonstrate in §6.2, these approaches are either very unstable, fail to span the trade-off or lead to sub-optimal trade-offs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "text", + "content": "Recently, Sadeghi et al. [29] demonstrated that adopting an HSIC-like dependence measure for the fairness objective and the target loss leads to a closed-form solution that is both efficient and effective at finding a near-optimal trade-off. Therefore, we incorporate the HSIC-like dependence measure and the closed-form solver into U-FaTE. Thus (1) can be expressed as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 396, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 396, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 344, + 396, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sup _ {\\boldsymbol {f} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l l} (1 - \\lambda) \\operatorname {D e p} (\\boldsymbol {f} (X; \\boldsymbol {\\Theta}), Y) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(X _ {c}; \\boldsymbol {\\Theta}\\right), S _ {c}\\right) \\right\\}, \\tag {2} \\\\ \\end{array}", + "image_path": "1e1e37eff04daed4b6e2cb3f8c79fd0330ecfb8bfcad92674c52d22961c2f0dc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_r" + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "text", + "content": " is a function space that encourages the representations to be uncorrelated. It does not affect the optimality of the learned encoder [29] and improves the compactness of representation [4]. Note that while the first term involves all data " + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "text", + "content": ", the second involves the conditional data " + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "inline_equation", + "content": "X_c" + }, + { + "bbox": [ + 304, + 451, + 545, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 514, + 546, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 546, + 622 + ], + "type": "text", + "content": "Choice of Dependence Measure: We adapt the dependence measure from [29] since it lends itself to a closed-form solution while capturing linear and non-linear dependencies under mild assumptions. While the dependence measure in [29] has been defined for absolute independence, our formulation in (2) also requires conditional independence to be compatible with separation based fairness definitions. Therefore, when " + }, + { + "bbox": [ + 304, + 514, + 546, + 622 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 514, + 546, + 622 + ], + "type": "text", + "content": " is not a continuous label, we define the conditional dependence measure as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 309, + 630, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 630, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 309, + 630, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\\\ \\sum_ {j = 1} ^ {r} \\sum_ {\\beta_ {S} \\in \\mathcal {U} _ {S}} \\mathbb {E} \\left[ \\left(f _ {j} \\left(X _ {c}\\right) - \\mathbb {E} f _ {j} \\left(X _ {c}\\right)\\right) \\left(\\beta_ {S} \\left(S _ {c}\\right) - \\mathbb {E} \\beta_ {S} \\left(S _ {c}\\right)\\right) \\right] \\tag {3} \\\\ \\end{array}", + "image_path": "8ffb337aff2f02a7f56280eb0c0ad213d847477e187b7aef4e5e95b176cedb97.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_S" + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": " is a countable orthonormal basis set for the separable universal RKHS " + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_S" + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "X_{c}\\sim P(X|Y = y)" + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12040" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "S_{c}\\sim P(S|Y = y)" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " are data and sensitive attributes, respectively. Empirically it can be estimated as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 105, + 287, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 105, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 61, + 105, + 287, + 128 + ], + "type": "interline_equation", + "content": "\\operatorname {D e p} (\\boldsymbol {f} (X), S | Y = y) := \\frac {1}{n ^ {2}} \\| \\boldsymbol {\\Theta} \\boldsymbol {K} _ {X _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {S _ {c}} \\| _ {F} ^ {2}, \\tag {4}", + "image_path": "d41429281ff9ede59240772aa6547b3185c019243f6e60809bfef11c20d7bfbc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " is the number of data samples, " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\pmb{K}_{X_c} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " is the Gram matrix corresponding to " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_X" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " is the encoder parameter in " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\pmb{f}(X) = \\Theta[k_{X_1}, k_{X_2}, \\dots, k_{X_n}]^T" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\pmb{H} = \\pmb{I}_n - \\frac{1}{n} \\pmb{1}_n \\pmb{1}_n^T" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " is the centering matrix, and " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\pmb{L}_{S_c}" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " is a full column-rank matrix such that " + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\pmb{L}_{S_c} \\pmb{L}_{S_c}^T = \\pmb{K}_{S_c}" + }, + { + "bbox": [ + 47, + 136, + 287, + 208 + ], + "type": "text", + "content": " (Cholesky factorization)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 216, + 256, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 256, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 256, + 228 + ], + "type": "text", + "content": "4.3. A Solution to the Optimization Problem" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "text", + "content": "Closed-Form Solver via Functions in RKHSs: Directly solving for all the parameters " + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "text", + "content": " through (2) and (4) leads to abysmal performance in practice since the kernel " + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "inline_equation", + "content": "K_{X}" + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "text", + "content": " has to be computed over the raw data space. Therefore, we instead define the fair encoder on the co-domain of the feature extractor " + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\pmb{f}(\\cdot ;\\Theta_{FE})" + }, + { + "bbox": [ + 47, + 234, + 287, + 307 + ], + "type": "text", + "content": ". So, in this case, (2) reduces to," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 315, + 287, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 287, + 362 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sup _ {\\boldsymbol {f} _ {E n c} \\in \\mathcal {A} _ {r}} \\left\\{ \\begin{array}{l} (1 - \\lambda) \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} (\\tilde {X}; \\boldsymbol {\\Theta} _ {E n c}), Y\\right) \\end{array} \\right. \\\\ \\left. - \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} _ {E n c} \\left(\\tilde {X} _ {c}; \\boldsymbol {\\Theta} _ {E n c}\\right), S _ {c}\\right) \\right\\}, \\tag {5} \\\\ \\end{array}", + "image_path": "d189cf315f6eb1a1005202c7edf8ae008be5167a39a8b0c923e46a15fe37a23e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{X} = f(X;\\Theta_{FE})" + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "content": ", and the first and second terms are " + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\frac{1}{n^2}\\| \\Theta_{Enc}K_{\\tilde{X}}HL_Y\\| _F^2" + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\frac{1}{n^2}\\left\\| \\Theta_{Enc}K_{\\tilde{X}_c}HL_{S_c}\\right\\| _F^2" + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "content": ", respectively. The parameters " + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 47, + 370, + 287, + 422 + ], + "type": "text", + "content": " can now be solved exactly via a closed-form solution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 430, + 214, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 214, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 214, + 441 + ], + "type": "text", + "content": "Theorem 1. A global optimizer of (5) is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 450, + 285, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 285, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 285, + 473 + ], + "type": "interline_equation", + "content": "\\pmb {f} _ {\\mathcal {H} _ {\\tilde {X}}} ^ {o p t} (\\tilde {X}; \\pmb {\\Theta} _ {E n c}) = \\pmb {\\Theta} _ {E n c} ^ {o p t} \\left[ k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {1}, \\tilde {X}), \\dots , k _ {\\tilde {X}} (\\tilde {\\pmb {x}} _ {n}, \\tilde {X}) \\right] ^ {T}", + "image_path": "96088a7357267d07a0dfd1cda71e7517ca72e0792fd976815253ad859d2ac557.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}^{opt} = \\pmb{U}^T\\pmb{L}_{\\tilde{X}}^\\dagger \\in \\mathbb{R}^{r\\times n}" + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "text", + "content": " and the columns of " + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "text", + "content": " are eigenvectors corresponding to the " + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 482, + 287, + 520 + ], + "type": "text", + "content": " largest eigenvalues of the following generalized eigenvalue problem." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 528, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 528, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 528, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left((1 - \\lambda) \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {Y} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} - \\lambda \\boldsymbol {L} _ {\\tilde {X} _ {c}} ^ {T} \\boldsymbol {H} \\boldsymbol {K} _ {S _ {c}} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X} _ {c}}\\right) \\boldsymbol {u} \\\\ = \\lambda \\left(\\frac {1}{n} \\boldsymbol {L} _ {\\tilde {X}} ^ {T} \\boldsymbol {H} \\boldsymbol {L} _ {\\tilde {X}} + \\gamma \\boldsymbol {I}\\right) \\boldsymbol {u}. \\tag {6} \\\\ \\end{array}", + "image_path": "3953cf9021557f4ee4789779742ee7033cc3746519a79d3c38e52ccc01225ad5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\tilde{X}}\\mathbf{L}_{\\tilde{X}}^T = \\mathbf{K}_{\\tilde{X}},\\tilde{X}_c\\sim p(\\tilde{X} |Y = y)" + }, + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "inline_equation", + "content": "S_{c}\\sim" + }, + { + "bbox": [ + 47, + 585, + 287, + 611 + ], + "type": "inline_equation", + "content": "p(S|Y = y)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 619, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 287, + 654 + ], + "type": "text", + "content": "Proof. The objective in (5) reduces to a generalized eigenvalue problem [18] by expressing it as a trace optimization problem. See supplementary for detailed proof. " + }, + { + "bbox": [ + 47, + 619, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\square" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "type": "text", + "content": "While this is a general solution to (5), the solution for each group fairness case is detailed in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Alternating Optimization: Now we present our full algorithm to optimize (2). We adopt standard minibatch to learn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "the feature encoder's parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{FE}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and the closed-form solver for the fair encoder parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". We optimize them alternatively where in each iteration, we update " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " while freezing " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{FE}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and vice-versa. Specifically, to optimize the fair encoder's parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ", we extract features from the data using the frozen feature extractor and use the closed-form solution in (5) to update " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". Then, we update the feature extractor's parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Theta_{FE}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " through minibatch SGD in (2) while freezing the encoder parameters. We repeat this process for every minibatch iteration. More details and an illustration of this alternating algorithm can be found in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 222, + 481, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 481, + 235 + ], + "type": "text", + "content": "4.4. Numerically Estimating the LST" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "text", + "content": "The Label Space Trade-off (LST) arises when the representation " + }, + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "text", + "content": " is not restricted to be a function of " + }, + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 240, + 545, + 288 + ], + "type": "text", + "content": ". Following the discussion in the previous subsection, this trade-off can be formulated as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 294, + 545, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 294, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 309, + 294, + 545, + 318 + ], + "type": "interline_equation", + "content": "\\sup _ {Z \\in L _ {r} ^ {2}} \\left\\{(1 - \\lambda) \\operatorname {D e p} (Z, Y) - \\lambda \\operatorname {D e p} (Z, S | Y = y) \\right\\}, \\tag {7}", + "image_path": "531dd665f072eb056d0b82171a50ddbb6c185cd45fa990ea580101c26f749dff.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "L_r^2" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " is the space of all RVs of dimension " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " with finite variance, i.e., " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_Z\\left[\\left\\| Z - \\mathbb{E}[Z]\\right\\|^2 < \\infty\\right]" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": ". From (7), observe that the optimal " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " is a function of " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{Y,S}" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " only. Therefore, instead of directly optimizing " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "L_r^2" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": ", equivalently, we optimize for " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\Theta_{FE}" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\Theta_{Enc}" + }, + { + "bbox": [ + 304, + 326, + 545, + 388 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 395, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 439 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}} \\left\\{\\left(1 - \\lambda\\right) \\operatorname {D e p} \\left(\\boldsymbol {f} \\left(Y, S; \\boldsymbol {\\Theta} _ {F E}, \\boldsymbol {\\Theta} _ {E n c}\\right), Y\\right) \\right. \\\\ \\left. + \\lambda \\operatorname {D e p} \\left(\\boldsymbol {f} (Y, S; \\Theta_ {F E}, \\Theta_ {E n c}), S | Y = y\\right) \\right\\}. \\tag {8} \\\\ \\end{array}", + "image_path": "0bc2b5c489b277239efd3b5d9b15e17f5ba4d5696c6b0db25f47d34e4b8b656f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": " is a function of the labels " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": ", i.e., the model takes as input " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": " and seeks to remove the information corresponding to " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": ", including that present in " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": ". In practice, to improve the stability of the optimization and facilitate learning, in addition to " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": ", we also use " + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 444, + 545, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 514, + 522, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 522, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 522, + 528 + ], + "type": "text", + "content": "5. Scheme for Evaluating Representations" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": "The primary utility of the trade-offs is in illuminating the fundamental limits of learning algorithms in mitigating unfairness and evaluating the effectiveness of learned representations w.r.t. utility and fairness. This includes representations that provide a single solution in the utility-fairness plane and multiple solutions that span the trade-off between utility and fairness. Fairness evaluations in prior literature focused primarily on relative comparisons of the models to each other. Such a comparison, however, precludes an understanding of how far the solution is from the inherent limits of the task. Elucidating and numerically quantifying the inherent trade-offs facilitates such an understanding and drives further algorithmic development." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "A standard way to evaluate bi-objective solutions is to plot them on a 2-D plane, identify non-dominated solutions," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12041" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 556, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 556, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 556, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 556, + 156 + ], + "type": "image", + "image_path": "f4be1810a3edbf26f89d48fd89eaa3bdd2b4b3daa3b259ddbdf703b1ef54f28f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 222 + ], + "type": "text", + "content": "Figure 3. Evaluating Fair Representation Learning Methods: Accuracy versus fairness trade-offs on CelebA (a)-(c) and FolkTable (d)-(f). (a) and (d) show the trade-off for Equalized Opportunity as the fairness constraint. (b) and (e) show the trade-off for Equality of Odds as the fairness constraint, and (c) and (f) show the trade-off for Demographic Parity as the fairness constraint. The solid lines represent the mean accuracy at a given fairness value, and the shaded region shows the uncertainty of the trade-off. Both DST and LST estimates from U-FaTE are stable. Among the FRL methods, " + }, + { + "bbox": [ + 46, + 166, + 547, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 166, + 547, + 222 + ], + "type": "text", + "content": " is closest to the DST, while ARL has the most variance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 268 + ], + "type": "text", + "content": "or compare their dominance w.r.t. each other. More details can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 278, + 191, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 191, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 191, + 292 + ], + "type": "text", + "content": "6. Experimental Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 299, + 253, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 253, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 253, + 311 + ], + "type": "text", + "content": "We designed experiments to answer the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 287, + 396 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 312, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 287, + 335 + ], + "type": "text", + "content": "1. How far are existing supervised fair representation learning methods from the two trade-offs? (§6.2)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "type": "text", + "content": "2. How far are zero-shot representations from the two trade-offs? What is the effect of network architecture and pre-training dataset? (§6.3)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "type": "text", + "content": "3. How far are pre-trained image representations trained in a supervised fashion from the two trade-offs? (§6.4)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 403, + 163, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 163, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 163, + 417 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "type": "text", + "content": "Datasets: We estimate the trade-offs through U-FaTE on an assortment of datasets. 1) CelebA [19] consists of more than 200K face images of celebrities in the wild annotated with 40 binary attributes. 2) FairFace [16] consists of face images from 7 different race groups labeled with race, sex, and age groups. 3) FolkTables [8] is a tabular dataset of individuals from fifty states derived from the US Census." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 506, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 639 + ], + "type": "text", + "content": "For experiments on CelebA, the target attribute is high cheekbones, and the sensitive attribute is a combination of sex and age for a total of four classes (young woman, young man, old woman, and old man). For experiments on the FairFace dataset, sex (binary) is the target attribute with race (7 groups: East Asian, White, Black, Indian, Latino, South Asian, and Middle Eastern) being the sensitive attribute. For the FolkTables dataset, we consider data from Washington State and choose employment status (binary) and age (discrete value between 1 and 96) of the individuals as the target and sensitive attributes, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Metrics: In all experiments, the utility is measured via classification accuracy. We consider three group fairness metrics (EOD, EOOD, and DPV). The zero-shot CLIP models are evaluated via cosine similarity. Moreover, on the Fair-Face dataset, we compared the models to the ideal solutions estimated by LST and DST using a weighted normalized" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "content": "Euclidean distance (see supplementary for details). We refer to the distance as " + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}_{LST}" + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}_{DST}" + }, + { + "bbox": [ + 305, + 243, + 545, + 267 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 546, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 366 + ], + "type": "text", + "content": "Implementation Details: We use ResNet-18 as the feature extractor for U-FaTE and the FRL methods. The final classifier is a two-layer MLP. We evaluate the representations from pre-trained vision models by learning a logistic classifier. To obtain the trade-offs, we run U-FaTE and other FRL methods for multiple values of " + }, + { + "bbox": [ + 304, + 270, + 546, + 366 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 270, + 546, + 366 + ], + "type": "text", + "content": " between zero and one, where zero corresponds to no fairness constraint, and one corresponds to only fairness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": "Optimizing U-FaTE: The trade-offs are defined through the dependence terms in (3) which involves an expectation over the joint distribution " + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "inline_equation", + "content": "p(X, Y, S)" + }, + { + "bbox": [ + 304, + 369, + 547, + 490 + ], + "type": "text", + "content": ". However, due to practical considerations, we only have access to a finite set of samples to estimate the trade-offs. Therefore, we estimate the trade-offs using all the samples available in each dataset without splitting it into train, validation, and test sets. This choice ensures that the estimates account for any possible generalization gap between the train and test distributions and identify the best achievable utility-fairness trade-offs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 498, + 448, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 448, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 448, + 511 + ], + "type": "text", + "content": "6.2. Evaluating FRL Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 517, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 567 + ], + "type": "text", + "content": "FRL Baselines: We consider a wide range of FRL methods based on adversarial learning (ARL [34] and MaxEnt-ARL [26]), dependence measures (FairHSIC [23], OptNet-ARL [28]), and closed-form solvers " + }, + { + "bbox": [ + 305, + 517, + 545, + 567 + ], + "type": "inline_equation", + "content": "(\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 305, + 517, + 545, + 567 + ], + "type": "text", + "content": " [29])." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Results: We estimate the LST and DST through U-FaTE and the trade-offs from the other baselines across various settings and datasets. Figs. 3a to 3c show the trade-offs on the CelebA dataset for EOD, EOOD, and DPV, respectively. Similarly, Figs. 3d to 3f show the trade-offs on the Folktable dataset for EOD, EOOD, and DPV, respectively. In the plots, the solid lines represent the mean, and the light shadows represent the variance of the accuracy for a given fairness value. On FairFace, observe that trade-offs do not exist since, on this task, it is possible to mitigate unfairness without sacrificing accuracy. Hence, we present the results of FRL methods and the estimated LST and DST in Tab. 1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12042" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 213 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 213 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 213 + ], + "type": "table", + "html": "
MethodAccuracy (↑)Unfairness (↓)DistDST (↓)DistLST (↓)
EODARL [34]93.391.340.4480.559
FairHSIC [23]91.021.330.4450.557
OptNet-ARL [28]92.941.700.5980.709
U-FaTE-DST96.170.263-0.133
U-FaTE-LST100.00.0--
EOODARL [34]91.603.040.4470.71
FairHSIC [23]93.432.130.2360.498
OptNet-ARL [28]93.392.340.2840.546
U-FaTE-DST97.931.126-0.262
U-FaTE-LST100.00.0--
DPVARL [34]92.496.090.3500.351
FairHSIC [23]91.415.910.3290.332
OptNet-ARL [28]93.335.800.3160.317
U-FaTE-DST94.393.082-0.04
U-FaTE-LST100.03.10--
", + "image_path": "2766db0713609a4154f374f1abaab0ee99aa64d034e5e51bbd36defbdf507d74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "text", + "content": "Table 1. Evaluation of FRL methods on FairFace based on the distance to DST and LST estimated by U-FaTE. Color corresponds to the DST and LST trade-offs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "text", + "content": "Observations: Although " + }, + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "text", + "content": ", OptNet-ARL and FairHSIC can achieve near-optimal accuracy in most cases, they are unable to span the whole range of fairness values. ARL is the most unstable but can span the whole range of fairness values. " + }, + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{K - T_{Opt}}" + }, + { + "bbox": [ + 46, + 267, + 287, + 351 + ], + "type": "text", + "content": " is the most stable method due to its closed-form solver, but is unable to span the whole range of fairness values (Figs. 3c and 3f)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": "The gap between LST and DST demonstrates the information gap in " + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": " for predicting " + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": ". From Fig. 3, we observe that the gap is " + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": " of accuracy in low fairness regions and " + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\sim 40\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": " in high fairness regions for EOD and EOOD. For DPV, the trend reverses with a gap of " + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": " for low fairness and gradually decreases with increasing unfairness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": "Observe that the LST in Figs. 3a and 3b and Figs. 3d and 3e for EOD and EOOD is almost flat at " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": " accuracy. This observation, however, is unsurprising since EOD and EOOD both condition on the label " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": ", and thus an ideal classifier with " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": " accuracy (i.e., " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Y} = Y" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": ") will have zero EOD and EOOD. And, in LST, the Oracle classifier is " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": " accurate since it has access to " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 426, + 288, + 628 + ], + "type": "text", + "content": ". So, the LST has sufficient information to minimize EOD and EOOD without sacrificing utility. The same, however, does not hold for DPV since it does not consider the target labels in its definition. Based on the above discussion, we deduce that EOD and EOOD are more pragmatic fairness metrics than DPV since they do not force the model to sacrifice predictive accuracy to ensure fairness. Thus, both offer a more balanced and practical approach to measuring fairness. Our empirical results provide independent confirmation of the same observations in [6, 14]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "The comparison of FRL methods in Tab. 1 based on " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}_{LST}" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " suggests that when models are optimized for EOD, ARL and FairHSIC find solutions closer to LST and DST than OptNet-ARL. When models are optimized for EOOD, FairHSIC finds the closest point to the LST and DST. OptNet-ARL performs slightly better than the other FRL methods when optimized for reducing DPV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "text", + "content": "6.3. Evaluating Zero-Shot CLIP Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 94, + 545, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 94, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 94, + 545, + 142 + ], + "type": "text", + "content": "Zero-Shot Models: To study the fairness of current zero-shot models, we consider more than 100 pre-trained models from OpenCLIP [15] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 146, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 545, + 206 + ], + "type": "text", + "content": "Results: Figs. 4a and 4b show results on CelebA and FairFace, respectively, for three group fairness definitions. Each point represents the result of one zero-shot CLIP model, with the color denoting the model's pre-training dataset. Plots also include DST and LST for comparison." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 328 + ], + "type": "text", + "content": "Observations: From Fig. 4a, we observe that zero-shot models perform poorly, in terms of accuracy, on the CelebA task. We hypothesize that it is so since the target task (predicting high cheekbones) is uncommon, even in the large datasets the models have been trained on. From a fairness perspective, we observe that models trained on CommonPool [9] (red dots) are more likely to be fair, while models trained on DataComp [9] have marginally better accuracy over the other models. Finally, the CLIP models are very far from the DST across the board." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 462 + ], + "type": "text", + "content": "On FairFace (Fig. 4b), since the target task (sex) is abundantly represented in all large-scale pre-training datasets, we observe that the CLIP models exhibit high levels of accuracy. Two CLIP models pre-trained on OpenAI WIT [24] are the closest to DST and LST w.r.t. EOD. Similar to our observations on CelebA, models pre-trained on the CommonPool dataset are more likely to be fair but at the cost of accuracy. In contrast, models pre-trained on DataComp are more unfair but have greater accuracy. Finally, there is a clear positive correlation between accuracy and unfairness across all fairness metrics on FairFace (Fig. 4b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 471, + 512, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 512, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 512, + 484 + ], + "type": "text", + "content": "6.4. Evaluating Supervised Representations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 491, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 550 + ], + "type": "text", + "content": "Supervised Baselines: To study the fairness of image representations from models pre-trained in a supervised fashion, we consider more than 900 models from Pytorch Image Models [33] and evaluate them on CelebA and FairFace for the same target and sensitive labels as before." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 554, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 626 + ], + "type": "text", + "content": "Results: Figs. 4c and 4d show results on CelebA and Fair-Face datasets, respectively. Each point represents the result of one supervised model, with color denoting the model's pre-training dataset. Plots also include DST and LST for comparison, but some plots were magnified for better resolution. So, the LST may only be partially visible." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Observations: From the CelebA results in Fig. 4c, we observe a high positive correlation between the accuracy and EOD and EOOD. Thus, models with better accuracy are also more fair. And, in contrast to the zero-shot models, even though high cheekbones is a rare label, the representations have sufficient information to accurately detect it with a logistic regression classifier. Specifically, models pre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12043" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 294, + 163 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 294, + 163 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 294, + 163 + ], + "type": "image", + "image_path": "0a02c2de14caee88fc2cd783dc3499821d69be1240559551202411b3dc781772.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 541, + 163 + ], + "type": "image", + "image_path": "e26458bb9691ed5da278f0e48a7d3df3e94c7c47b2e5a68726c3e41cb6a39647.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 167, + 425, + 176 + ], + "lines": [ + { + "bbox": [ + 414, + 167, + 425, + 176 + ], + "spans": [ + { + "bbox": [ + 414, + 167, + 425, + 176 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 177, + 295, + 262 + ], + "blocks": [ + { + "bbox": [ + 169, + 167, + 179, + 176 + ], + "lines": [ + { + "bbox": [ + 169, + 167, + 179, + 176 + ], + "spans": [ + { + "bbox": [ + 169, + 167, + 179, + 176 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 177, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 52, + 177, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 52, + 177, + 295, + 262 + ], + "type": "image", + "image_path": "95b0f331c6a56ff7065a57a305ad7e45312f6680cb82cdec8212acfda1bd0ee2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 266, + 178, + 274 + ], + "lines": [ + { + "bbox": [ + 168, + 266, + 178, + 274 + ], + "spans": [ + { + "bbox": [ + 168, + 266, + 178, + 274 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 178, + 539, + 262 + ], + "blocks": [ + { + "bbox": [ + 298, + 178, + 539, + 262 + ], + "lines": [ + { + "bbox": [ + 298, + 178, + 539, + 262 + ], + "spans": [ + { + "bbox": [ + 298, + 178, + 539, + 262 + ], + "type": "image", + "image_path": "78246bf3607b8d1fc7406e33a3730ded1a5cc1359b6ebfdb248d6ad000d641dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 265, + 424, + 274 + ], + "lines": [ + { + "bbox": [ + 413, + 265, + 424, + 274 + ], + "spans": [ + { + "bbox": [ + 413, + 265, + 424, + 274 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 277, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 545, + 300 + ], + "type": "text", + "content": "Figure 4. Evaluating Pre-Trained Zero-Shot and Supervised Models: Accuracy-fairness evaluation of more than 100 pre-trained zero-shot models on CelebA (a) and FairFace (b), and over 900 pre-trained image representations on CelebA (c), and FairFace (d)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": "trained on ImageNet22K [25] (orange dots), and OpenAI WIT [24] have the best accuracy and reach the DST tradeoff in high unfairness regions. However, with more than " + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": " EOD and EOOD and more than " + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 310, + 287, + 369 + ], + "type": "text", + "content": " DPV, they have significant levels of bias between the two sexes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 370, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 598 + ], + "type": "text", + "content": "Results on FairFace in Fig. 4d also reiterate that models trained on OpenAI WIT and ImageNet22K are more fair and more accurate than other datasets. We also observe that the models are generally more fair on FairFace than on CelebA. A similar positive correlation exists between accuracy and fairness for EOD and EOOD. We also make an interesting observation from Fig. 4d. Some models surpass the DST and enter the \"Possible with Extra Data\" region of the utility-fairness plane illustrated in Fig. 1a. Recall that DST estimates the upper bound of the possible region for models trained with the same data as DST. The DST can be surpassed to enter the \"Possible with Extra Data\" region if the model is trained on additional data beyond what is used for estimating DST. From Fig. 4d (middle), we observe that a few models trained on OpenAI, LAION [30], ImageNet22K, ImageNet21K, and ImageNet12K can surpass the DST, plausibly since these datasets contain sufficient—both quality and quantity—samples from the distribution of the target attribute (sex)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 609, + 171, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 171, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 171, + 623 + ], + "type": "text", + "content": "7. Concluding Remarks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "As image classification systems are widely deployed in high-stakes applications, ensuring that their predictions do not exhibit demographic bias is paramount for gaining user trust. While it is desirable to mitigate bias without sacrificing accuracy, this is not always possible. This paper studied such inherent trade-offs between utility and fairness. First, we identified two types of trade-offs called Data-Space and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 310, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 467 + ], + "type": "text", + "content": "Label-Space trade-offs corresponding to those achievable with and without data restrictions. But unlike prior theoretical studies on utility-fairness trade-offs, next, we focused on developing algorithmic tools for quantifying the trade-offs from data and proposed U-FaTE. As an illustration of its practical utility, we estimated the trade-offs on several image classification tasks, facilitating a large-scale evaluation of over 100 zero-shot and 900 supervised pre-trained models. The results revealed that, out of the box, pre-trained models are far from the best achievable limits of accuracy and fairness. Furthermore, we identified that, in some cases, larger datasets can improve accuracy and fairness and surpass the solutions represented by the DST." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 487, + 546, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 546, + 668 + ], + "type": "text", + "content": "U-FaTE was designed as a composition of a neural network with the last layer optimized to global optimality through a closed-form solver for a given feature representation. This design allowed it to estimate the two trade-offs reliably. However, U-FaTE does not provide convergence or optimality guarantees. Therefore, the estimated DST and LST are likely to be suboptimal. Nonetheless, they can serve as a valuable tool for understanding the nature of the problem (e.g., Does there exist a trade-off? or What are the possible, impossible and possible with extra data regions in the utility-fairness plane?) and how far a given fair learning algorithm is from the achievable limits. Furthermore, depending on which region of the utility-fairness plane a solution is and how far it is from the DST and LST reveals whether to focus on better optimization or better data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by the National Science Foundation (award #2147116)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12044" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "text", + "content": "[2] Solon Barocas, Moritz Hardt, and Arvind Narayanan. *Fairness and Machine Learning: Limitations and Opportunities*. MIT Press, 2023. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "text", + "content": "[3] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. LEACE: Perfect linear concept erasure in closed form. arXiv preprint arXiv:2306.03819, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[5] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[6] Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. big data 5, 2 (2017), 153-163. arXiv preprint arXiv:1610.07524, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "text", + "content": "[7] Sepehr Dehdashtian, Lan Wang, and Vishnu Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "type": "text", + "content": "[8] Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. Retiring adult: New datasets for fair machine learning. In Advances in Neural Information Processing Systems, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "type": "text", + "content": "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating face recognition bias via group adaptive classifier. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "text", + "content": "[11] Thibaut Le Gouic, Jean-Michel Loubes, and Philippe Rigollet. Projection to fairness in statistical learning. arXiv preprint arXiv:2005.11720, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[12] Vincent Grari, Oualid El Hajouji, Sylvain Lamprier, and Marcin Detyniecki. Learning unbiased representations via Rényi minimization. arXiv preprint arXiv:2009.03183, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, and Bernhard Scholkopf. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12):2075-2129, 2005. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Moritz Hardt, Eric Price, Nati Srebro, et al. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems, 2016. 3, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[16] Kimmo Karkkainen and Jungseock Joo. FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "type": "text", + "content": "[17] Niki Kilbertus, Mateo Rojas Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Scholkopf. Avoiding discrimination through causal reasoning. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "text", + "content": "[18] Effrosini Kokiopoulou, Jie Chen, and Yousef Saad. Trace optimization and eigenproblems in dimension reduction methods. Numerical Linear Algebra with Applications, 18(3): 565-602, 2011. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In IEEE International Conference on Computer Vision, 2015. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "text", + "content": "[20] David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Learning adversarially fair and transferable representations. arXiv preprint arXiv:1802.06309, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[21] Daniel McNamara, Cheng Soon Ong, and Robert C Williamson. Costs and benefits of fair representation learning. AAAI/ACM Conference on AI, Ethics, and Society, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[22] Aditya Krishna Menon and Robert C Williamson. The cost of fairness in binary classification. Conference on Fairness, Accountability and Transparency, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "text", + "content": "[23] Novi Quadrianto, Viktoriia Sharmanska, and Oliver Thomas. Discovering fair representations in the data domain. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 525 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 526, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 559 + ], + "type": "text", + "content": "[25] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[26] Proteek Roy and Vishnu Naresh Boddeti. Mitigating information leakage in image representations: A maximum entropy approach. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "text", + "content": "[27] Bashir Sadeghi, Runyi Yu, and Vishnu Boddeti. On the global optima of kernelized adversarial representation learning. IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[28] Bashir Sadeghi, Lan Wang, and Vishnu Naresh Boddeti. Adversarial representation learning with closed-form solvers. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2021. 2, 4, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Bashir Sadeghi, Sepehr Dehdashtian, and Vishnu Boddeti. On characterizing the trade-off in invariant representation" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12045" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 487 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "learning. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[31] Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations. In IEEE/CVF International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "text", + "content": "[32] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 286, + 306 + ], + "type": "text", + "content": "[33] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "type": "text", + "content": "[34] Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. Controllable invariance through adversarial feature learning. Advances in Neural Information Processing Systems, 2017. 2, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 385 + ], + "type": "text", + "content": "[35] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair representations. International Conference on Machine Learning, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 286, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 286, + 408 + ], + "type": "text", + "content": "[36] Han Zhao. Costs and benefits of wasserstein fair regression. arXiv preprint arXiv:2106.08812, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "text", + "content": "[37] Han Zhao and Geoffrey J Gordon. Inherent trade-offs in learning fair representations. arXiv preprint arXiv:1906.08386, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[38] Han Zhao, Jianfeng Chi, Yuan Tian, and Geoffrey J Gordon. Trade-offs and guarantees of adversarial representation learning for information obfuscation. arXiv preprint arXiv:1906.07902, 2019. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12046" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_content_list.json b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d97b24c31be0418c5f145409d43c6089a83b032 --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_content_list.json @@ -0,0 +1,1998 @@ +[ + { + "type": "text", + "text": "VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models", + "text_level": 1, + "bbox": [ + 120, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Li* Qianli Shen* Kenji Kawaguchi National University of Singapore {xiangli, qianli, kenji} @ comp.nus.edu.sg", + "bbox": [ + 287, + 203, + 679, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0631ba4c845e782141a22326e2ad2e181b579a6c8d0cea4a869e2ce2abdc0a45.jpg", + "image_caption": [ + "Figure 1. Given a copyrighted image $y_{C}$ and a Text-to-Image (T2I) generative model with probabilistic copyright protection, our proposed virtually assured amplification attack (VA3) significantly amplifies the probability of producing infringing generations with persistent interactions of online adversarial prompt selection." + ], + "image_footnote": [], + "bbox": [ + 81, + 290, + 888, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 631, + 313, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The booming use of text-to-image generative models has raised concerns about their high risk of producing copyright-infringing content. While probabilistic copyright protection methods provide a probabilistic guarantee against such infringement, in this paper, we introduce Virtually Assured Amplification Attack (VA3), a novel online attack framework that exposes the vulnerabilities of these protection mechanisms. The proposed framework significantly amplifies the probability of generating infringing content on the sustained interactions with generative models and a non-trivial lower-bound on the success probability of each engagement. Our theoretical and experimental results demonstrate the effectiveness of our approach under various scenarios. These findings highlight the potential risk of implementing probabilistic copyright protection in", + "bbox": [ + 75, + 652, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "practical applications of text-to-image generative models. Code is available at https://github.com/South7X/VA3.", + "bbox": [ + 500, + 633, + 888, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 678, + 630, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the advancement of large generative models [17, 47, 50] has revolutionized high-quality image synthesis [34, 37, 40], paving the way for commercial applications that enable the public to effortlessly craft their own artworks and designs [2, 13, 20, 27, 38, 39]. Nevertheless, these models exhibit notable memorization capabilities to produce generations highly similar to the training data [3]. This resemblance raises growing concerns about copyright infringement, especially when copyrighted data is used for training [12, 16, 49, 52].", + "bbox": [ + 496, + 703, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these concerns, there has been a surge in research focused on protecting copyrighted data from potential infringement by outputs of generative models [14, 21,", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 887, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/058a8d52955a1282d97e39bfa285658827fcbef3f22eafc817222f3ccbc682b0.jpg", + "image_caption": [ + "(a) Output w/o CP- $k$" + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 277, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aded96dea7d52576e4526e27ede13d16ef9395664899488443e80336beed8921.jpg", + "image_caption": [ + "(b) Output w/ CP-k." + ], + "image_footnote": [], + "bbox": [ + 282, + 87, + 482, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/db87137dff15d6aafeff4e0ec47530872e9e84a098d99398dfbfac2c13495388.jpg", + "image_caption": [ + "(c) Amp. output w/ CP- $k$", + "Figure 2. Example outputs given the copyright image in Fig. 1 as target (potential infringing images are marked with red boundaries). In (a), using a benign prompt, we observe a high incidence of infringing content from models without copyright protection (\"w/o CP- $k$ )\". In contrast, (b) shows that after applying the copyright protection mechanism (\"w/ CP- $k$ )\", all samples are safe as CP- $k$ rejects all infringing content. In (c), we find that amplification (Amp.) attack with a benign prompt results in limited success. Notably, by amplification attack with an adversarial prompt obtained from our proposed Anti-NAF algorithm, almost all output in (d) are copyright-infringed." + ], + "image_footnote": [], + "bbox": [ + 486, + 87, + 687, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/589a8d6610525929211e58c24d80cd8dfef503320baadfbf37b019d813da4f53.jpg", + "image_caption": [ + "(d) Anti-NAF Amp. output w/ CP- $k$" + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 890, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "24, 43, 45, 52, 58]. Among these studies, a pivotal concept involves establishing a probabilistic upper-bound against the generation of infringing content by generative models. We refer to this suite of approaches as probabilistic copyright protection. Most notably, Vyas et al. [52] introduce a mathematical definition of copyright known as near-access freeness (NAF). Their method enforces generative diffusion models to exhibit akin behaviors as safe models, which has no access to the copyrighted image. By leveraging the improbability of safe models generating infringing content, the probability of generative models doing the same is thereby substantially reduced. The copyright protection algorithm of NAF, CP- $k$ , can filter out infringing content generated by the models with high probability, even when the input prompts are adversarially designed.", + "bbox": [ + 73, + 364, + 472, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Virtually Assured Amplification Attack (VA3), a novel online attack framework, to show the vulnerability of probabilistic copyright protection. This framework induces text-to-image generative models with probabilistic copyright protection to generate infringing content. Our approach is grounded in the realization that in real-world scenarios, a malicious attacker intending to induce copyright infringement could engage in multiple interactions with the generative model via prompts. This persistent engagement poses a significant challenge to probabilistic protection methods, as it amplified the probability of producing infringing content of each generation.", + "bbox": [ + 75, + 595, + 468, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our proposed framework, the attacker functions as a conditional prompt generator, creating adversarial prompts iteratively based on previous interactions with the generative model. Our primary theoretical result, Theorem 1, suggests that the amplification attack is guaranteed to succeed with high probability, given a sufficient number of interactions with the generative model and a strictly positive lower-bound on success probability of each single engagement.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding practical algorithms, our work encompasses two technical innovations. Firstly, we present effective strategies to manage the exploration-exploitation dilemma in online prompt selection, thereby enhancing the stability of the attack. Secondly, we propose Anti-NAF, a theoretically motivated adversarial prompt optimization algorithm tailored for NAF copyright protection, to generate prompts fulfilling the conditions of Theorem 1.", + "bbox": [ + 496, + 364, + 890, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experimental results validate the efficacy of our proposed online attack approach under diverse scenarios. These findings underline the potential copyright infringement risk of applying probabilistic protection in practical applications of text-to-image generative models, for both providers and users.", + "bbox": [ + 496, + 486, + 892, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 593, + 640, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Copyright Issues in Generative Models", + "text_level": 1, + "bbox": [ + 500, + 619, + 834, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image generative models, trained on large-scale datasets like LAION [44], have been equipped with enhanced memorization ability to generate outputs of high semantic similarity to their training data [3, 48]. Given the prevalence of copyrighted works in these datasets, the significant risk of copyright infringement for these generations has raised great concerns from the public [6, 18] and researchers [1, 4, 16, 42, 48, 49, 52]. Many efforts have been made to safeguard copyrighted materials from being infringed by generative diffusion models. Some researchers [23, 24, 41, 45] introduced data perturbation, where input data is modified to hinder the model to imitate copyrighted features. Another separate line of works [14, 21, 26, 43, 58] exploited concept removal that erases unsafe concepts from existing pre-trained diffusion models to mitigate the risk of undesirable generations. In an alternative approach, researchers studied watermark protection for copyrighted data", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12364", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[9, 29, 36, 54, 56, 59] to encode ownership information into potentially infringed outputs.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A notable contribution, Vyas et al. [52] first provided a mathematical probabilistic upper-bound against copyright-infringed generation. They asserted that the proposed near access-freeness (NAF) offers robust guarantees for copyright protection. However, Elkin-Koren et al. [12] argued the limitation of this method for reducing copyright to a matter of privacy from a legal perspective. In this paper, we build upon these discussions to present a significant challenge to these probabilistic copyright protection methods through the amplification attack.", + "bbox": [ + 75, + 121, + 470, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vulnerability of Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 281, + 369, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A rising number of works focus on the vulnerability of diffusion models to various types of attacks. Poisoning attack [5, 8, 46, 55, 57] studies the problem of manipulating training data to induce unsafe behaviors in diffusion models during the training phase. In the context of the real-world inference phase for existing text-to-image diffusion models, other works also investigate the robustness and safety against different prompt inputs. Researchers [15, 25, 28, 60] have shown that injecting a slight perturbation to the input prompts will mislead the unprotected model to generate semantically unrelated images. Furthermore, with carefully crafted prompts, unsafe images can easily evade detection-based safety filters in text-to-image diffusion models [32, 35]. Other studies [7, 51] red-teaming concept removal copyright protection methods by finding problematic prompts that can recover erased unsafe concepts to yield undesirable generations. Diverging from these existing attack approaches on heuristic protection methods, in this paper, our focus is on challenging the vulnerability of probabilistic copyright protection methods.", + "bbox": [ + 75, + 303, + 470, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 619, + 215, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 643, + 356, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Take DDPM [17] for example, a typical diffusion process consists of a predefined forward process and a reverse process. Specifically, the forward process corrupts the original data $\\mathbf{x}_0 \\sim q(\\mathbf{x})$ into a standard Gaussian noise $\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ through $T$ timesteps. In the reverse process, the denoising network $\\epsilon_{\\theta}$ is learned to denoise the corrupted $\\mathbf{x}_t$ by predicting the sampled noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ that added to $\\mathbf{x}_0$ . The objective of the diffusion model can be simplified into the following form:", + "bbox": [ + 75, + 666, + 468, + 803 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 811, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image diffusion models further use prompts to guide the sampling process for generating desired images. Take Stable Diffusion [37] for example, they incorporate a pre-trained CLIP [33] as text encoder $\\pmb{\\tau}$ to encode an input", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text $\\mathbf{y}$ , where $\\mathbf{c} = \\boldsymbol{\\tau}(\\mathbf{y})$ . An image encoder $\\mathcal{E}$ is employed to map an input image $\\mathbf{x}_0$ into its latent representation $\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)$ . The training objective is formulated as:", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\boldsymbol {\\epsilon}, \\mathbf {c}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| _ {2} ^ {2} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 145, + 890, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Near Access-Freeness Copyright Protection", + "text_level": 1, + "bbox": [ + 498, + 171, + 867, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a pioneering work of probabilistic copyright protection method, Vyas et al. [52] formally defines Near Access-Freeness (NAF) to provide a probabilistic guarantee against the generation of infringing content. They provide practical algorithms (CP- $k$ ) to protect an arbitrary generative model $p$ from copyright infringement. Suppose we have a cover of safe models $\\mathcal{S}$ satisfying that for any piece of copyrighted data $y_{C}$ , there exists some $q = \\text{safe}_{C} \\in \\mathcal{S}$ trained without access to $y_{C}$ . The copyright protection at a pre-given threshold $k$ can be achieved by sampling $y \\sim p(\\cdot | x)$ with a prompt $x$ and accepting $y$ if:", + "bbox": [ + 496, + 194, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (y | x) := \\max _ {q \\in S} \\log \\frac {p (y | x)}{q (y | x)} \\leq k \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 369, + 890, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, $\\rho(y|x)$ serves as a criterion to distinguish non-infringing content from infringing content, while $k$ determines the threshold for judgment. Let $p_k$ be the protected model with threshold $k$ and $\\mathcal{Y}_C$ be the set of infringing contents, CP- $k$ provides the following probabilistic upper-bound for copyright infringement:", + "bbox": [ + 496, + 410, + 890, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} (y \\in \\mathcal {V} _ {C} | x) \\leq \\frac {2 ^ {k}}{\\nu_ {k} (x)} \\cdot \\operatorname {s a f e} _ {C} (y \\in \\mathcal {V} _ {C} | x) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 510, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\nu_{k}(x)$ denotes the acceptance rate of $p_k$ . For different prompts $x$ , the optimal choice of $k$ changes with varying distribution of $\\rho(y|x)$ . In the subsequent sections of our paper, we use $k_x$ to emphasize the dependency.", + "bbox": [ + 496, + 551, + 890, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 626, + 589, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 650, + 700, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we consider a text-to-image generative model $\\tilde{p}$ equipped with probabilistic copyright protection. The probability that $\\tilde{p}$ generates infringing content with prompt $x$ is upper-bounded. Following the typical adversarial prompt attack setting, we consider a malicious attacker who seeks to manipulate $\\tilde{p}$ to produce content that violates the copyright of a specific piece of target copyrighted data $y_{C}$ , of which the infringing contents form $\\mathcal{Y}_C \\subseteq \\mathcal{Y}$ .", + "bbox": [ + 496, + 674, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the standard paradigm, we consider an online attack scenario where the attacker is allowed to interact with $\\tilde{p}$ for $T$ times within each attack trial. During the $t = 1,\\dots ,T$ -th interaction, the attacker inputs a prompt $x_{t}$ and receives a generated sample $y_{t}\\sim \\tilde{p} (\\cdot |x_{t})$ . Besides furnishing a prompt, the attacker is prohibited from intervening directly in the generation process. This constraint", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is consistent with real-world scenarios where users interact with generative models as black boxes through APIs. At the end of each attack trial, the attacker will select $y^{*} \\in \\{y_{t}\\}_{t=1}^{T}$ as the final output, and the attack is regarded successful if $y^{*} \\in \\mathcal{V}_{C}$ . The formal objective of the attacker is to maximize the success rate of attack, i.e.,", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} \\left(y ^ {*} \\in \\mathcal {Y} _ {C}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 199, + 468, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Virtually Assured Amplification Attack on Probabilistic Copyright Protection", + "text_level": 1, + "bbox": [ + 76, + 231, + 468, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Copyright protection mechanisms in generative models, e.g., $NAF$ , are designed to yield a low upper-bound on the probability of generating infringing samples. Such protection is confined to single or infrequent generations. However, in scenarios where malicious attackers engage in multiple targeted generation requests, the probability of a successful attack can be amplified.", + "bbox": [ + 75, + 270, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a universal online amplification attack framework, detailed in Algorithm 1, for a virtually assured success against probabilistic copyright protection. In this framework, the attacker is modeled as an adversarial prompt generator $\\mathcal{A}(x_t|\\langle (x_s,y_s)^{t - 1}_s\\rangle)$ , which iteratively generates the prompt $x_{t}$ for the current step $t$ based on the interaction history with the generative model. For selecting the optimal sample, we employ a scoring function $\\mathcal{S}:\\mathcal{V}\\to \\mathbb{R}$ to evaluate all the samples and return the highest-scoring sample as the final result. An attack is deemed successful if the score of the returned result is higher than the target score $S_{tar}$ . For example, an essential choice of the scoring function could be the indicator function $\\mathbb{I}(y\\in \\mathcal{V}_C)$ , with the target score $S_{tar} = 0$ , which ensures that any infringing sample generated will be returned as the final result, making the attack as successful. In practice, to reduce the manual effort in identifying copyright infringement, a computable surrogate scoring function with a target score may be utilized to establish a standard for copyright infringement.", + "bbox": [ + 75, + 377, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the following theorem, which suggests the virtually assured success of the amplification attack with an assumption on the lower-bound of success probability for each single-shot attack attempt.", + "bbox": [ + 75, + 665, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Following the notations in Algorithm 1, for any $\\varepsilon \\in (0,1)$ , the attack is successful with probability at least $1 - \\varepsilon$ if $T > \\log_{1 - \\sigma} \\varepsilon$ , where $\\sigma > 0$ is a strictly positive lower-bound on the success probability shared by every single attack.", + "bbox": [ + 75, + 736, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intuition behind the amplification attack and Theorem 1 is straightforward. If the attacker is granted sufficient sampling opportunities with a set of prompts that yield even a modest lower-bound on the single success probability, as the number of sampling attempts increases, the probability", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Amplification Attack on Probabilistic Copyright Protection", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Generative model $\\tilde{p}$ with probabilistic copyright protection, target copyrighted data $\\mathcal{C}$ , adversarial prompt generator $\\mathcal{A}$ , maximum number of steps $T$ , score function $\\mathcal{S}$ .", + "bbox": [ + 501, + 125, + 890, + 184 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $t = 1, \\dots, T$ do", + "2: Sample prompt $x_{t} \\sim \\mathcal{A}(\\cdot |\\langle (x_{s},y_{s})\\rangle_{s = 1}^{t - 1})$", + "3: Feed $x_{t}$ to $\\tilde{p}$ and receive $y_{t} \\sim \\tilde{p}(\\cdot | x_{t})$", + "4: end for", + "5: return $y^{*} = \\arg \\max_{y\\in \\{y_{1:T}\\}}S(y)$" + ], + "bbox": [ + 509, + 186, + 821, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of generating infringing samples accumulates at an exponentially fast rate, eventually making an attack almost guaranteed to succeed. For instance, suppose our attack possesses a relatively low single-shot success probability, say $1\\%$ . Provided that we are permitted to repeat the attack approximately $\\log_{0.99} 0.01 \\approx 459$ times, the probability of at least one successful instance is then amplified to $99\\%$ .", + "bbox": [ + 496, + 294, + 890, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Online Prompt Selection", + "text_level": 1, + "bbox": [ + 498, + 411, + 725, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we consider a specific scenario within the framework described in Sec. 4.2, where the conditional prompt generator (the attacker) is restricted to select a prompt among $K$ candidate prompts $\\{x^{1},\\dots ,x^{K}\\}$ , based on previous choices and the scores of received samples. The generation of candidate prompts can either be independent of or tailored specifically for certain models and copyright protection mechanisms, as we will discuss in Sec. 4.4, designed for diffusion model and NAF copyright protection.", + "bbox": [ + 496, + 434, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $a_{t} \\in \\{1, \\dots, K\\}$ denote the decision at step $t$ and $\\pi$ denote the policy of prompt selection, the prompt generation procedure can be formally described as", + "bbox": [ + 498, + 571, + 890, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = x ^ {a _ {t}}, a _ {t} \\sim \\pi (\\cdot | \\langle (x _ {s}, r _ {s} = \\mathcal {S} (y _ {s})) \\rangle_ {s = 1} ^ {t - 1}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 628, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this context, online prompt generation is reformulated as a variant of the multi-armed bandit problem, where each candidate prompt is conceptualized as an arm, and the score of the corresponding sample is the reward for pulling the arm. Unlike the classic multi-armed bandit framework, our goal is to maximize the probability of obtaining a reward exceeding a specified threshold $S_{tar}$ at least once within $T$ trials. This involves the trade-off between exploration and exploitation. Herein, we present two variants of the $\\varepsilon$ -greedy algorithm. Initially, akin to the conventional algorithm, we conduct $m \\leq \\lfloor T / K \\rfloor$ trials for each arm. For the rest $T - mK$ steps, a random action will be taken with probability $\\varepsilon > 0$ for exploration. Otherwise, the best action $\\hat{a}_t^* = \\arg \\max_a \\hat{Q}_t(a)$ according to the evaluation $\\hat{Q}_t(a)$ will be taken for exploitation. The choice of $\\hat{Q}_t$ diverges into two distinct variants.", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12366", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\varepsilon$ greedy-max Inspired by the fact that only the max reward matters, the maximum reward received so far is employed for evaluation, i.e., $\\hat{Q}_t(a) = \\max \\{r_s : a_s = a\\}_{s=1}^t$ . $\\varepsilon$ greedy-cdf Essentially, we aim to maximize the probability that the reward for the next step exceeds a threshold $S_{tar}$ . If we assume that the reward distribution for each arm adheres to a normal distribution, we have $\\hat{Q}_t(a) = 1 - \\Phi\\left(\\frac{S_{tar} - \\hat{\\mu}_t(a)}{\\hat{\\sigma}_t(a)}\\right)$ , where $\\hat{\\mu}_t(a) = \\frac{1}{N_t(a)} \\sum_{s=1}^t r_s \\mathbb{I}(a_s = a)$ , $\\hat{\\sigma}_t^2(a) = \\frac{1}{N_t(a) - 1} \\sum_{s=1}^t (r_s \\mathbb{I}(a_s = a) - \\hat{\\mu}_t(a))^2$ , and $\\Phi$ denotes the cumulative distribution function (cdf) of standard normal distribution.", + "bbox": [ + 76, + 90, + 472, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our empirical findings suggest that a well-calibrated balance between exploration and exploitation can enhance the stability of the attack.", + "bbox": [ + 76, + 265, + 470, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection", + "text_level": 1, + "bbox": [ + 76, + 316, + 470, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we narrow the scope of discussion to adversarial prompt discovery against CP- $k$ , a NAF-based copyright protection algorithm. CP- $k$ modifies the probability density function of an unprotected generative model $p$ as", + "bbox": [ + 76, + 356, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} (y | x) \\propto p (y | x) \\mathbb {I} (\\rho (y | x) \\leq k _ {x}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 425, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ is defined in Eq. (3) and $k_{x}$ is a prompt-dependent threshold determined by the generative system, which is typically inaccessible to the user. CP- $k$ provides a probabilistic upper-bound for copyright infringement in Eq. (4).", + "bbox": [ + 76, + 449, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building upon our initial assumptions, we further assume that the attacker is granted full access, i.e., knowing the structures and parameters of these models, to the unprotected generative model $p$ and safe models $q \\in S$ . While this white-box setting renders our attack narrowed to open-source models and those at risk of backend leakage, the potential threat of our attack remains substantial, as it does not require intervention in the generation process. This means that if a malicious attacker gains access to a leaked model and discovers adversarial prompts, any other users, lacking access to the underlying models, can readily replicate these adversarial prompts to reproduce the attack.", + "bbox": [ + 75, + 510, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that our objective is to find adversarial prompts capable of inducing models protected by CP- $k$ to generate infringing content with a strictly positive probability, i.e.,", + "bbox": [ + 76, + 691, + 468, + 737 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 744, + 468, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We deduce a lower-bound for the infringement probability with optimal adversarial prompt in Theorem 2, subsequently guiding the development of a practical algorithm.", + "bbox": [ + 76, + 772, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1 (Local Continuity). Given a distance measure $\\mathcal{D}$ defined in $\\mathcal{V}$ , a model $p$ is called $(\\epsilon, \\alpha)$ -local-continuous around $y_0 \\in \\mathcal{V}$ if for any prompt $x \\in \\mathcal{X}$ , there exists $\\epsilon, \\alpha > 0$ such that for any $y \\in \\mathcal{B}_{\\mathcal{D}}(y_0, \\epsilon) := \\{y \\in \\mathcal{Y} : \\mathcal{D}(y_0, y) < \\epsilon\\}$ , $|p(y_0|x) - p(y|x)| < \\alpha \\mathcal{D}(y_0, y)$ .", + "bbox": [ + 76, + 825, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection" + ], + "code_body": "Require: Denoising network $\\epsilon_{\\theta_p},\\epsilon_{\\theta_q}$ , text encoder $\\pmb{\\tau}$ , target $y_{C}$ , denoising steps $T$ , optimization step $S$ , loss clip bound $\\varphi$ , loss weight $\\lambda$ , learning rate $\\gamma$ \n1: $\\mathbf{P} = \\{\\mathbf{e}_1,\\dots ,\\mathbf{e}_n\\} \\sim E^{|\\mathcal{V}|\\times d}$ \n2: for 1,.., S do \n3: $x = \\mathrm{Proj}(\\mathbf{P})$ \n4: Diffusion reconstruction task \n5: $t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})$ \n6: $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n7: $y_{t} = \\sqrt{\\bar{\\alpha}_{t}} y_{C} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon$ \n8: $\\mathcal{L}_p = \\| \\epsilon -\\epsilon_{\\theta_p}(y_t,t,\\tau (x))\\| _2^2$ \n9: $\\mathcal{L}_q = \\| \\epsilon -\\epsilon_{\\theta_q}(y_t,t,\\tau (x))\\| _2^2$ for all $q\\in S$ \n10: Calculate the gradient w.r.t projected embedding \n11: $g = \\nabla_{x}(\\lambda \\cdot \\max (\\mathcal{L}_{p},\\varphi) + (1 - \\lambda)\\cdot \\max_{q\\in S}(\\mathcal{L}_{q}))$ \n12: Apply the gradient on continuous embedding \n13: $\\mathbf{P} = \\mathbf{P} - \\gamma g$ \n14: end for \n15: return Proj(P)", + "bbox": [ + 501, + 125, + 890, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Assume there is a distance measure $\\mathcal{D}$ defined in $\\mathcal{V}$ such that (i) $p$ is $(\\epsilon_p, \\alpha)$ -local-continuous around $y_C$ , (ii) every $q \\in S$ is local-continuous around $y_C$ , and (iii) there exists $\\epsilon_c > 0$ such that $\\mathcal{B}_{\\mathcal{D}}(y_C, \\epsilon_c) \\subseteq \\mathcal{V}_C$ . The objective defined in Eq. (8) has the following lower-bound for any $\\eta, \\delta > 0$ ,", + "bbox": [ + 496, + 429, + 893, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\geq \\max _ {x \\in \\tilde {\\chi} _ {\\eta , \\delta}} \\eta C _ {1} - \\alpha C _ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 532, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathcal{X}}_{\\eta ,\\delta} = \\{x\\in \\mathcal{X}:p(y_C|x)\\geq \\eta ,\\rho (y_C|x) < k_x - \\delta \\}$ and $C_1,C_2$ are constants independent on $x$ given as", + "bbox": [ + 498, + 571, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {1} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} d y, C _ {2} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} \\mathcal {D} (y _ {C}, y) d y,\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 614, + 879, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon = \\min (\\epsilon_p,\\epsilon_c,\\epsilon_\\rho)$ with $\\epsilon_{\\rho}\\coloneqq \\inf_{x\\in \\tilde{\\mathcal{X}}_{\\eta ,\\delta}}\\sup \\{\\epsilon :$ $\\rho (y|x) < k_x,\\forall y\\in \\mathcal{B}_D(y_C,\\epsilon)\\} .$", + "bbox": [ + 498, + 660, + 890, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the lower-bound in the Theorem 2 is nontrivial, we need to set $\\eta > \\alpha C_2 / C_1$ and search along prompts satisfying (a) $\\rho(y_C|x) < k_x$ , (b) $p(y_C|x) \\geq \\eta$ . Unfortunately, the mechanism to determine the threshold $k_x$ is assumed inaccessible to the attacker. To obtain a feasible optimization objective, we instead minimize $\\rho(y_C|x)$ subject to $p(y_C|x) \\geq \\eta$ , as an alternative. The optimization objective then becomes", + "bbox": [ + 496, + 704, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x} \\rho \\left(y _ {C} | x\\right) \\quad \\text {s . t .} p \\left(y _ {C} | x\\right) \\geq \\eta . \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 842, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the reconstruction task of the original text-to-image diffusion models for direct optimization. In aligning", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12367", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the attack objective of Eq. (10), we develop the optimization objective from two aspects:", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon}, x} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta_ {p}} (y _ {t}, t, \\boldsymbol {\\tau} (x)) \\| _ {2} ^ {2} ] \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 132, + 468, + 152 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {q} = \\mathbb {E} _ {t, \\epsilon , x} [ \\| \\epsilon - \\epsilon_ {\\theta_ {q}} (y _ {t}, t, \\tau (x)) \\| _ {2} ^ {2} ] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 162, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $y_{t}$ is the noisy version of target copyrighted data $y_{C}$ at denoising step $t$ ; $\\epsilon_{\\theta_p}, \\epsilon_{\\theta_q}$ is the denoising network of generative model $p$ and safe model $q \\in S$ respectively.", + "bbox": [ + 76, + 186, + 468, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While $\\mathcal{L}_p$ is designed to increase the possibility of model $p$ in generating the desired infringed content $y_{C}$ , the minimization of $\\mathcal{L}_q$ corresponds to the maximization of $q(y_C|x)$ consequently leading to the reduction of $\\rho (y_C|x)$ . Hence, the overall optimization objective can be formulated as a weighted sum of $\\mathcal{L}_p$ and $\\mathcal{L}_q$ :", + "bbox": [ + 75, + 233, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\cdot \\max \\left(\\mathcal {L} _ {p}, \\varphi\\right) + (1 - \\lambda) \\cdot \\max _ {q \\in \\mathcal {S}} \\left(\\mathcal {L} _ {q}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 338, + 468, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\varphi$ is a loss clip to mitigate the conflict between $\\mathcal{L}_p$ and $\\mathcal{L}_q$ in minimizing $\\rho$ .", + "bbox": [ + 76, + 368, + 468, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The complete Anti-NAF prompt optimization process is detailed in Algorithm 2. We conduct optimization in the continuous embedding space with a sequence of learnable embeddings $\\mathbf{P} = \\{\\mathbf{e}_1, \\ldots, \\mathbf{e}_n | \\mathbf{e}_i \\in \\mathbb{R}^d\\}$ , where $n$ is the sequence length and $d$ is the embedding dimension. Given that each word token $w$ in a vocabulary $\\mathcal{V}$ can be represented as a corresponding embedding $\\mathrm{EMB}(w)$ using an embedding matrix $E^{|\\mathcal{V}| \\times d}$ , each embedding $\\mathbf{e}$ in the sequence can be mapped to its nearest token embedding in $\\mathcal{V}$ under some similarity metric, e.g., cosine similarity. The projection function can be defined as $\\mathrm{Proj}(\\mathbf{e}) = \\arg \\min_{w \\in \\mathcal{V}} \\cos(\\mathrm{EMB}(w), \\mathbf{e})$ . The resultant text prompt is a sequence of these projected tokens: $x = \\mathrm{Proj}(\\mathbf{P}) = \\{\\mathrm{Proj}(\\mathbf{e}_1), \\ldots, \\mathrm{Proj}(\\mathbf{e}_n)\\}$ . In addition, following the approach in [53], the continuous embeddings $\\mathbf{P}$ are projected into discrete tokens $x$ for each forward pass in every optimization step. The gradient of the projected prompt $x$ is then applied to update the continuous embeddings $\\mathbf{P}$ .", + "bbox": [ + 75, + 398, + 468, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 684, + 209, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 709, + 282, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1 Evaluated Datasets and Models", + "text_level": 1, + "bbox": [ + 76, + 732, + 349, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since training large text-to-image diffusion models from scratch is impractical, we fine-tune the pre-trained StableDiffusion-v1-4 model provided by Huggingface with two datasets: POKEMON [30] and LAION-mi [10]. Note that both datasets do not overlap with the pre-training dataset of Stable Diffusion, ensuring that the safe models fine-tuned on them are inaccessible to copyrighted data. Following [52], each dataset is split into two disjoint shards", + "bbox": [ + 75, + 756, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{D}_1$ and $\\mathcal{D}_2$ to train generative model $q_{1}$ and $q_{2}$ respectively. For copyrighted data $y_{C}$ , $q_{2}$ is served as safe model safe $C$ if $y_{C}\\in \\mathcal{D}_{1}$ but $y_{C}\\notin \\mathcal{D}_{2}$ . For better experimental illustrations, each copyrighted data is repeated to make up $1\\%$ of the dataset shard.", + "bbox": [ + 500, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "POKEMON [30]. This dataset consists of 833 image-caption pairs, where captions are obtained from the BLIP model [22]. We evaluate our attack on 5 copyrighted images added to one of the shards separately and fine-tune models $q_{1}$ and $q_{2}$ on each shard for 5000 steps.", + "bbox": [ + 496, + 167, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LAION-mi [10]. The dataset is originally constructed for membership inference attacks on diffusion models. We only use the nonmembers part of this dataset, which holds a similar data distribution but is disjoint with the pre-training dataset of Stable Diffusion. We evaluate our attack on 5 selected copyrighted images, which are separately combined with one of non-copyrighted data shards. Each dataset shard is of size 5000. We fine-tune model $q_{1}$ and $q_{2}$ for 25,000 steps. Details for fine-tuning can be found in the appendix.", + "bbox": [ + 496, + 243, + 892, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2 Implementation details", + "text_level": 1, + "bbox": [ + 500, + 402, + 714, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the sampling setting of text-to-image diffusion models, images of size $512 \\times 512$ are generated using a classifier-free guidance scale of 7.5 and 50 sampling steps with the default scheduler. The amplification step is set to 100 and 500 for POKEMON and LAION-mi respectively. For the prompt optimization process, the learning rate is set to 0.01 using Adagrad [11] optimizer with 25,000 optimization steps, the gradient accumulation step is set to 5, the loss clip $\\varphi$ for model $p$ is 0.01, and the loss weight $\\lambda$ is 0.95. The length of optimized prompts is set to 8 tokens. All experiments are conducted on A100 GPUs.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Settings", + "text_level": 1, + "bbox": [ + 500, + 604, + 686, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate with a model $p$ behaves with an equal chance of sampling from either model $q_{1}$ or $q_{2}$ . We consider such model $p$ as a strict protection model that even with the original caption as prompt, there is a $50\\%$ chance of generating from the safe model. In contrast, a model $p$ trained on the entire dataset could frequently produce highly similar generations to the target copyrighted image, making it extremely inefficient for the selection of a valid threshold $k$ . Next, we will separately discuss the data, the metrics, and the threat prompts for evaluation.", + "bbox": [ + 496, + 628, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Data. We use samples generated by $p$ for evaluation. To determine whether they infringe the copyright of the target image, we rely on the similarity between samples and the target image as there is no widely acknowledged computable standard for infringement to our knowledge. We employ the current SOTA similarity metric SSCD [31] for copy detection. As SSCD scores of infringed samples for different target images differ significantly by obser", + "bbox": [ + 496, + 780, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/CompVis/stable-diffusion-v1-4", + "bbox": [ + 94, + 886, + 387, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12368", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95b53acae85c213b8c4795de3ee7bd052202387f749c08848dee09eb70dc65cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 135, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5ef27f72007279251d2fbdfdbdbc79de60cc263bdaeb1dee2e59dc9766468e3.jpg", + "image_caption": [ + "a black and white zebrastanding next to a star" + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 241, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1eb20425d9017640ab071fa2756f7c90852d367381700449458a1aa6b144dea8.jpg", + "image_caption": [ + "zebra weet unex mee" + ], + "image_footnote": [], + "bbox": [ + 243, + 89, + 348, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1451d9ee728b3151c69393fd0353a405c1bea4c177058ca4038e21d7111971b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 89, + 403, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab818ab2deb108b12b205fdd2705ac17705ee3d706f79e5f00080cc3b9ce7e22.jpg", + "image_caption": [ + "a picture of a white horsewith orange and yellow slings" + ], + "image_footnote": [], + "bbox": [ + 418, + 89, + 511, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c28ca20a4883a1fdbc3e654731085d20c83dc23d151f2aec655e47369092c03d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 89, + 614, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4738cbedbf6175654f511a0a3d75b518c6b02a54500f3438a1d31228f813bbb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 89, + 673, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68d118a425ea0f8ba18455c62ab779d2c9843cf3e3d7e5d68fac65589c0038e4.jpg", + "image_caption": [ + "a cartoon rabbit with asocce around its neck" + ], + "image_footnote": [], + "bbox": [ + 679, + 89, + 779, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62eec3ba1ce2c6fcab3b98dd02c157513021b7365b6ddfea6007c861a948c35b.jpg", + "image_caption": [ + "rabbit, lamewridayday catunmibrated the room" + ], + "image_footnote": [], + "bbox": [ + 792, + 89, + 883, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fc6e76e20515ddc3275a1cde8f58f1ed400bf6904dda158406af806e2215961.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 152, + 135, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cb4dbaa395da339682983e353f04fdef965ec257ea73baf3ba19e0dce2d0328.jpg", + "image_caption": [ + "Giraffes Canvas Art Print" + ], + "image_footnote": [], + "bbox": [ + 138, + 152, + 241, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d3d505ba0e79062a8f745735c89fdcd32a36dc48c978b7a758e633cb9ccbab5.jpg", + "image_caption": [ + "giraffe guintspitnoche", + "tallent big etary", + "Figure 3. Visualization of generated images on different copyright targets. The examples in the first and second rows are selected from POKEMON and LAION-mi respectively. The prompts used to generate output are given below each group of images. Remarkably, the copyright-infringed content generated with Anti-NAF amplification reveals the vulnerability of probabilistic copyright protection CP- $k$ ." + ], + "image_footnote": [], + "bbox": [ + 245, + 152, + 346, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1643aec6851c347c356b1280fd57891a61eb3ebff289cc4df0af831e31d43289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 152, + 403, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9e5962edd28115ac91b9514d0e16c1695e27b78a744294ded66906e036fa139.jpg", + "image_caption": [ + "Product photo wall tattoo children's rostradberries with the inksticker set" + ], + "image_footnote": [], + "bbox": [ + 408, + 152, + 511, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9004b5819071094d27dbf3d52bb4ee071e43f32f3557f60df2ed2499c9694c82.jpg", + "image_caption": [ + "horse millennials symbolism", + "kano jacoblissseasoned" + ], + "image_footnote": [], + "bbox": [ + 514, + 152, + 616, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3f7f659a6df2a56c22629c0845bac9a343e2d45d6db9524d597e7a4b75e999f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 152, + 673, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fa8a38bb0628737b71ffb02cfe8ac30a11e3b36a401a42d34ec388733954484.jpg", + "image_caption": [ + "Penguins pictures of the New", + "Year, March with their own hands" + ], + "image_footnote": [], + "bbox": [ + 684, + 152, + 779, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2896a3a56ee6d52cd313290683163f9aaa112c52044fc0e8c6792352039a4da.jpg", + "image_caption": [ + "penguins knobs swimming", + "orbcsport orbfunday puppies have" + ], + "image_footnote": [], + "bbox": [ + 785, + 152, + 834, + 191 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/912d5157b8075d9cefc8afe69492873eea10ec2d78a33fc09a49c18a4d6a5d7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPOKEMONLAION-mi
CIRFAR@5%AR↑FAR@15%AR↑CIRFAR@10%AR↑FAR@30%AR↑FAR@50%AR↑
Caption (w/o Amp.)47.00%0.40%3.28%49.64%0.00%0.00%0.04%
CLIP-Int. (w/o Amp.)26.92%0.84%2.20%48.12%0.84%1.52%2.96%
PEZ (w/o Amp.)7.80%1.32%2.76%15.80%0.08%0.24%0.40%
Anti-NAF (w/o Amp.)12.88%8.52%10.00%33.84%2.64%4.16%7.00%
Caption (w/ Amp.)100.00%13.64%40.16%100.00%0.00%0.00%14.64%
CLIP-Int. (w/ Amp.)99.72%22.84%47.32%100.00%34.84%37.12%75.60%
PEZ (w/ Amp.)66.92%22.80%37.48%98.16%38.04%50.68%59.64%
Anti-NAF (w/ Amp.)99.84%77.36%91.36%100.00%56.12%73.32%95.92%
", + "bbox": [ + 86, + 272, + 883, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results. The performance with the amplification attack (\"w/ Amp.\") is significantly superior to scenarios without amplification (\"w/o Amp.\") Additionally, our proposed Anti-NAF demonstrates notably promising outcomes for providing a substantial probability of copyright infringement when probabilistic protection is applied. (CLIP-Int. is the abbreviation for CLIP-Interrogator).", + "bbox": [ + 75, + 419, + 892, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vation (shown in Fig. 5), we do not indicate a fixed score threshold for all target images. Instead, we use relative thresholds determined as percentiles of the similarity scores among samples generated by $p$ with the original caption of the target image as the prompt, e.g. SSCD-50%. Recognizing the variable criteria of infringement, we report results with other choices of thresholds in Sec. 11.3.", + "bbox": [ + 75, + 488, + 468, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. The CP- $k$ method achieves copyright protection by selectively accepting generated samples using the threshold $k_{x}$ , which can be indicated by the Acceptance Rate (AR). For a given AR, a good protection system is expected to rarely accept infringing content, i.e., have a low False Accept Rate (FAR), as defined in Eq. (8). It is worth noting that the choice of $k_{x}$ dictates the trade-off between model safety and efficiency. Furthermore, to our knowledge, there is no principled way to determine $k_{x}$ . As a result, we evaluate the success of attack by reporting the FAR at different AR, e.g., FAR@5%AR. Additionally, the copyright infringement rate (CIR) is also presented for scenarios without copyright protection, i.e., $\\mathrm{AR} = 100\\%$ .", + "bbox": [ + 75, + 603, + 468, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Threat Prompts. Given that generating images similar to the target image is a necessary condition for a successful attack, we focus on the following threat prompts: (i) Target caption: the original caption of the target image; (ii) PEZ [53]: a gradient-based discrete optimization approach to discover prompts semantically similar to the target im", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "age; (iii) $CLIP$ -Interrogator2: a prompt consists of the BLIP caption of the target image and top- $k$ keywords greedily sampled from a keywords collection; (iv) Anti-NAF: an adversarial prompt against NAF copyright protection as described in Sec. 4.4. For each threat prompt, we sample 6400 images from model $p$ for evaluation. For the online prompt selection setting, we take the aforementioned three prompts (exclude the target caption) as a prompt candidate set and consider $\\varepsilon$ -greedy-max-/cdf Bandit Amplification as bandit strategies as described in Sec. 4.3.", + "bbox": [ + 496, + 488, + 892, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 500, + 652, + 591, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 2, we show example outputs under four scenarios. It is evident that with amplification, the CP- $k$ protected generative model does indeed output copyright-infringed content with high probability, especially when using the Anti-NAF prompt. These examples suggest that probabilistic safeguards against copyright infringement are vulnerable to the amplification attack. Additional visualizations of outputs on various copyrighted images are demonstrated in Fig. 3. In Tab. 1, we report CIRs and FARs at various ARs on two datasets. Notably, there is a significant growth in both metrics when amplification attack is employed, highlighting its effectiveness in amplifying the probability of infringing generations. The superior performance of Anti-NAF under", + "bbox": [ + 496, + 676, + 892, + 873 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/pharmapsychotic/clip-interrogator", + "bbox": [ + 514, + 886, + 805, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12369", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fb7f79a69147f808b80b5828cc17ae37d7f803124c0c812bd1a489313060ef8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 272, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4e9ae3771e25aa7a7b15c35e579417aed5bc7338d969fefb74fc22f4acbb04e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 89, + 467, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f47bfac79e07c02ea7a3fef834b477560fbe3ea215da2b9a198281e1b757fa85.jpg", + "image_caption": [ + "(c) FAR-AR of Anti-NAF with different Amp. steps on POKEMON.", + "Figure 4. The overall FAR-AR curves. The results show that amplification is strongly effective in amplifying the possibility of infringed output, even with small amplification steps." + ], + "image_footnote": [], + "bbox": [ + 80, + 270, + 271, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/645819ffcfa46b03dabb3c216801f96a8dece9ad2c35481fe494a92243d65264.jpg", + "image_caption": [ + "(d) FAR-AR of Anti-NAF with different Amp. steps on LAION-mi." + ], + "image_footnote": [], + "bbox": [ + 276, + 270, + 467, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scores its efficacy in rendering infringing generations with a substantial probability. The overall FAR-AR curves are illustrated in Fig. 4. We can observe that bandit variants of amplification lead to a smaller variance across different target copyrighted images, especially at lower acceptance rates, indicating that our bandit strategies achieve a more steady attack. In Figs. 4c and 4d, we plot the overall FAR-AR curves over different amplification steps. There is a clear trend of rapidly improved performance with increased amplification steps, due to the cumulative probability of generating infringing samples. This finding underlines the potential risk in practical applications of probabilistic copyright protections, given the high frequency of daily interactions with text-to-image generative models. We provide additional results and human evaluations in Sec. 11.", + "bbox": [ + 75, + 535, + 472, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 771, + 230, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 2, we study how different components of our optimization objective in Eq. (13), affect the performance. We discover that an exclusive focus on minimizing $\\mathcal{L}_p$ (\" $\\mathcal{L}_p$ only\") leads to a notable drop in performance. This indicates that prompts designed merely to reconstruct the target image are insufficient for a successful attack, as the infringing output of such a prompt can be easily identified and", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/92b18b739f03a27935fd18f1c21163bbe9007a976b84ef4995b4d81c5bbfdcb5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCIRFAR@5%AR↑FAR@15%AR↑
Anti-NAF12.88%8.52%10.00%
\\( \\mathcal{L}_p \\) only10.60%1.84%2.04%
w/o \\( \\varphi \\)12.12%1.36%1.48%
w/o \\( \\mathcal{L}_q \\)13.72%2.00%2.76%
", + "bbox": [ + 526, + 89, + 867, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study for Anti-NAF algorithm on POKEMON.", + "bbox": [ + 501, + 175, + 888, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "blocked by the copyright protection system. Furthermore, a direct combination of $\\mathcal{L}_q$ with $\\mathcal{L}_p$ (\"w/o $\\varphi$ ) results in additional performance degradation because of the conflicting objectives of $\\mathcal{L}_p$ and $\\mathcal{L}_q$ in minimizing $\\rho(y_C|x)$ . On the other hand, we can observe a slight performance improvement when implementing a loss clip $\\varphi$ on $\\mathcal{L}_p$ (\"w/o $\\mathcal{L}_q$ ) to constrain the learning of $\\mathcal{L}_p$ . Overall, these results validate the effectiveness of our well-balanced optimization objective. Additional ablation studies are provided in Sec. 12.", + "bbox": [ + 496, + 217, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 367, + 617, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we shed light on the vulnerability of probabilistic copyright protection methods for text-to-image generative models, especially in real-world scenarios involving persistent and targeted interactions. Our proposed Virtually Assured Amplification Attack (VA3) framework presents the feasibility of inducing the protected model to generate copyright-infringed content with an amplified probability. Despite our focus on a narrow scenario of online prompt selection within this framework, the experimental results highlight its effectiveness in challenging even the most advanced existing probabilistic copyright protection methods. However, a broader scope of potential strategies remains unexplored within the VA3 framework, such as online prompt optimization, which may provide more powerful attacks against copyright protection. Furthermore, our Anti-NAF algorithm relies on access to generative models for adversarial prompt optimization, an assumption that may not be satisfied in completely black-box scenarios. We leave these more complicated and general attack methods for future investigation. In conclusion, our findings emphasize the significant risk of copyright infringement when applying probabilistic copyright protection methods in practice. Therefore, we hope that this work can inspire the development of more robust copyright protection approaches.", + "bbox": [ + 496, + 392, + 892, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 660, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The authors are grateful to Nikhil Vyas for providing the source code of CP- $k$ algorithm. This research is partially supported by the National Research Foundation Singapore under the AI Singapore Programme (AISG Award No: AISG2-TC-2023-010-SGIL) and the Singapore Ministry of Education Academic Research Fund Tier 1 (Award No: T1 251RES2207).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12370", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hossein Aboutalebi, Daniel Mao, Carol Xu, and Alexander Wong. Deepfakeart challenge: A benchmark dataset for generative ai art forgery and data poisoning detection. arXiv preprint arXiv:2306.01272, 2023. 2", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1", + "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2", + "[4] Stephen Casper, Zifan Guo, Shreya Mogulothu, Zachary Marinov, Chinmay Deshpande, Rui-Jie Yew, Zheng Dai, and Dylan Hadfield-Menell. Measuring the success of diffusion models at imitating human artists. arXiv preprint arXiv:2307.04028, 2023. 2", + "[5] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3", + "[6] David Chess. Some light infringement? https://ceoln.wordpress.com/2022/12/16/some-light-infringement/, 2022.2", + "[7] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3", + "[8] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 3", + "[9] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3", + "[10] Jan Dubiński, Antoni Kowalczuk, Stanisław Pawlak, Przemysław Rokita, Tomasz Trzeciński, and Pawel Morawiecki. Towards more realistic membership inference attacks on large diffusion models. arXiv preprint arXiv:2306.12983, 2023. 6", + "[11] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 6", + "[12] Niva Elkin-Koren, Uri Hacohen, Roi Livni, and Shay Moran. Can copyright be reduced to privacy? arXiv preprint arXiv:2305.14822, 2023. 1, 3", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023. 1, 2", + "[15] Hongcheng Gao, Hao Zhang, Yinpeng Dong, and Zhi-jie Deng. Evaluating the robustness of text-to-image diffusion models against real-world attacks. arXiv preprint arXiv:2306.13103, 2023. 3", + "[16] Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. arXiv preprint arXiv:2303.15715, 2023. 1, 2", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[18] Matthew Butterick Joseph Saveri. Stable diffusion litigation. https://stablediffusionlitigation.com/, 2023. 2", + "[19] Kenji Kawaguchi, Zhun Deng, Xu Ji, and Jiaoyang Huang. How does information bottleneck help deep learning? In International Conference on Machine Learning (ICML), 2023. 2", + "[20] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1", + "[21] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516, 2023. 1, 2", + "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 12888-12900. PMLR, 2022. 6", + "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2", + "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 20763-20786. PMLR, 2023. 2", + "[25] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974, 2023. 3", + "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 21548-21566. PMLR, 2023. 2", + "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1", + "[28] Natalie Maus, Patrick Chao, Eric Wong, and Jacob Gardner. Adversarial prompting for black box foundation models. arXiv preprint arXiv:2302.04237, 2023. 3", + "[29] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 3", + "[30] Justin N. M. Pinkney. Pokemon blip captions. https://huggingface.co/datasets/lambdaabs/pokemon-blip-captions/, 2022.6", + "[31] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14512-14522. IEEE, 2022. 6, 2", + "[32] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. arXiv preprint arXiv:2305.13873, 2023. 3", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1", + "[35] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 3", + "[36] Arkadip Ray and Somaditya Roy. Recent trends in image watermarking techniques for copyright protection: a survey. International Journal of Multimedia Information Retrieval, 9(4):249-270, 2020. 3", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1", + "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[41] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 29894-29918. PMLR, 2023. 2", + "[42] Sarah Scheffler, Eran Tromer, and Mayank Varia. Formalizing human ingenuity: A quantitative framework for copyright law's substantial similarity. In Proceedings of the 2022 Symposium on Computer Science and Law, pages 37-49, 2022. 2", + "[43] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2", + "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[45] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2", + "[46] Shawn Shan, Wenxin Ding, Josephine Passananti, Haitao Zheng, and Ben Y Zhao. Prompt-specific poisoning attacks on text-to-image generative models. arXiv preprint arXiv:2310.13828, 2023. 3", + "[47] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1", + "[48] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 6048-6058. IEEE, 2023. 2", + "[49] Growthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 2", + "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1", + "[51] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12372", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 1, 2, 3, 6", + "[53] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 6, 7", + "[54] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 3", + "[55] Yixin Wu, Ning Yu, Michael Backes, Yun Shen, and Yang Zhang. On the proactive generation of unsafe images from text-to-image models using benign prompts. arXiv preprint arXiv:2310.16613, 2023. 3", + "[56] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023. 3", + "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. arXiv preprint arXiv:2305.04175, 2023. 3", + "[58] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2", + "[59] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3", + "[60] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023 - Workshops, Vancouver, BC, Canada, June 17-24, 2023, pages 2385-2392. IEEE, 2023. 3", + "[61] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 670 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_model.json b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf8290e55860ca05867fc93cd0240c2f88cd4a06 --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.85, + 0.175 + ], + "angle": 0, + "content": "VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.204, + 0.681, + 0.258 + ], + "angle": 0, + "content": "Xiang Li* Qianli Shen* Kenji Kawaguchi National University of Singapore {xiangli, qianli, kenji} @ comp.nus.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.291, + 0.889, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.577, + 0.893, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Given a copyrighted image \\( y_{C} \\) and a Text-to-Image (T2I) generative model with probabilistic copyright protection, our proposed virtually assured amplification attack (VA3) significantly amplifies the probability of producing infringing generations with persistent interactions of online adversarial prompt selection." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.632, + 0.314, + 0.646 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.471, + 0.88 + ], + "angle": 0, + "content": "The booming use of text-to-image generative models has raised concerns about their high risk of producing copyright-infringing content. While probabilistic copyright protection methods provide a probabilistic guarantee against such infringement, in this paper, we introduce Virtually Assured Amplification Attack (VA3), a novel online attack framework that exposes the vulnerabilities of these protection mechanisms. The proposed framework significantly amplifies the probability of generating infringing content on the sustained interactions with generative models and a non-trivial lower-bound on the success probability of each engagement. Our theoretical and experimental results demonstrate the effectiveness of our approach under various scenarios. These findings highlight the potential risk of implementing probabilistic copyright protection in" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.634, + 0.89, + 0.663 + ], + "angle": 0, + "content": "practical applications of text-to-image generative models. Code is available at https://github.com/South7X/VA3." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.679, + 0.631, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.854 + ], + "angle": 0, + "content": "In recent years, the advancement of large generative models [17, 47, 50] has revolutionized high-quality image synthesis [34, 37, 40], paving the way for commercial applications that enable the public to effortlessly craft their own artworks and designs [2, 13, 20, 27, 38, 39]. Nevertheless, these models exhibit notable memorization capabilities to produce generations highly similar to the training data [3]. This resemblance raises growing concerns about copyright infringement, especially when copyrighted data is used for training [12, 16, 49, 52]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To address these concerns, there has been a surge in research focused on protecting copyrighted data from potential infringement by outputs of generative models [14, 21," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.204, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.278, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.244, + 0.235, + 0.256 + ], + "angle": 0, + "content": "(a) Output w/o CP- \\(k\\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.088, + 0.483, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.244, + 0.436, + 0.256 + ], + "angle": 0, + "content": "(b) Output w/ CP-k." + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.088, + 0.688, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.244, + 0.656, + 0.256 + ], + "angle": 0, + "content": "(c) Amp. output w/ CP- \\(k\\)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.089, + 0.892, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.244, + 0.889, + 0.256 + ], + "angle": 0, + "content": "(d) Anti-NAF Amp. output w/ CP- \\(k\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.894, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Example outputs given the copyright image in Fig. 1 as target (potential infringing images are marked with red boundaries). In (a), using a benign prompt, we observe a high incidence of infringing content from models without copyright protection (\"w/o CP-\\(k\\))\". In contrast, (b) shows that after applying the copyright protection mechanism (\"w/ CP-\\(k\\))\", all samples are safe as CP-\\(k\\) rejects all infringing content. In (c), we find that amplification (Amp.) attack with a benign prompt results in limited success. Notably, by amplification attack with an adversarial prompt obtained from our proposed Anti-NAF algorithm, almost all output in (d) are copyright-infringed." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.365, + 0.473, + 0.592 + ], + "angle": 0, + "content": "24, 43, 45, 52, 58]. Among these studies, a pivotal concept involves establishing a probabilistic upper-bound against the generation of infringing content by generative models. We refer to this suite of approaches as probabilistic copyright protection. Most notably, Vyas et al. [52] introduce a mathematical definition of copyright known as near-access freeness (NAF). Their method enforces generative diffusion models to exhibit akin behaviors as safe models, which has no access to the copyrighted image. By leveraging the improbability of safe models generating infringing content, the probability of generative models doing the same is thereby substantially reduced. The copyright protection algorithm of NAF, CP-\\(k\\), can filter out infringing content generated by the models with high probability, even when the input prompts are adversarially designed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.777 + ], + "angle": 0, + "content": "In this paper, we propose Virtually Assured Amplification Attack (VA3), a novel online attack framework, to show the vulnerability of probabilistic copyright protection. This framework induces text-to-image generative models with probabilistic copyright protection to generate infringing content. Our approach is grounded in the realization that in real-world scenarios, a malicious attacker intending to induce copyright infringement could engage in multiple interactions with the generative model via prompts. This persistent engagement poses a significant challenge to probabilistic protection methods, as it amplified the probability of producing infringing content of each generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In our proposed framework, the attacker functions as a conditional prompt generator, creating adversarial prompts iteratively based on previous interactions with the generative model. Our primary theoretical result, Theorem 1, suggests that the amplification attack is guaranteed to succeed with high probability, given a sufficient number of interactions with the generative model and a strictly positive lower-bound on success probability of each single engagement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Regarding practical algorithms, our work encompasses two technical innovations. Firstly, we present effective strategies to manage the exploration-exploitation dilemma in online prompt selection, thereby enhancing the stability of the attack. Secondly, we propose Anti-NAF, a theoretically motivated adversarial prompt optimization algorithm tailored for NAF copyright protection, to generate prompts fulfilling the conditions of Theorem 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.893, + 0.577 + ], + "angle": 0, + "content": "Our experimental results validate the efficacy of our proposed online attack approach under diverse scenarios. These findings underline the potential copyright infringement risk of applying probabilistic protection in practical applications of text-to-image generative models, for both providers and users." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.642, + 0.61 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.836, + 0.636 + ], + "angle": 0, + "content": "2.1. Copyright Issues in Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Text-to-image generative models, trained on large-scale datasets like LAION [44], have been equipped with enhanced memorization ability to generate outputs of high semantic similarity to their training data [3, 48]. Given the prevalence of copyrighted works in these datasets, the significant risk of copyright infringement for these generations has raised great concerns from the public [6, 18] and researchers [1, 4, 16, 42, 48, 49, 52]. Many efforts have been made to safeguard copyrighted materials from being infringed by generative diffusion models. Some researchers [23, 24, 41, 45] introduced data perturbation, where input data is modified to hinder the model to imitate copyrighted features. Another separate line of works [14, 21, 26, 43, 58] exploited concept removal that erases unsafe concepts from existing pre-trained diffusion models to mitigate the risk of undesirable generations. In an alternative approach, researchers studied watermark protection for copyrighted data" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12364" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "[9, 29, 36, 54, 56, 59] to encode ownership information into potentially infringed outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.471, + 0.273 + ], + "angle": 0, + "content": "A notable contribution, Vyas et al. [52] first provided a mathematical probabilistic upper-bound against copyright-infringed generation. They asserted that the proposed near access-freeness (NAF) offers robust guarantees for copyright protection. However, Elkin-Koren et al. [12] argued the limitation of this method for reducing copyright to a matter of privacy from a legal perspective. In this paper, we build upon these discussions to present a significant challenge to these probabilistic copyright protection methods through the amplification attack." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.37, + 0.297 + ], + "angle": 0, + "content": "2.2. Vulnerability of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.607 + ], + "angle": 0, + "content": "A rising number of works focus on the vulnerability of diffusion models to various types of attacks. Poisoning attack [5, 8, 46, 55, 57] studies the problem of manipulating training data to induce unsafe behaviors in diffusion models during the training phase. In the context of the real-world inference phase for existing text-to-image diffusion models, other works also investigate the robustness and safety against different prompt inputs. Researchers [15, 25, 28, 60] have shown that injecting a slight perturbation to the input prompts will mislead the unprotected model to generate semantically unrelated images. Furthermore, with carefully crafted prompts, unsafe images can easily evade detection-based safety filters in text-to-image diffusion models [32, 35]. Other studies [7, 51] red-teaming concept removal copyright protection methods by finding problematic prompts that can recover erased unsafe concepts to yield undesirable generations. Diverging from these existing attack approaches on heuristic protection methods, in this paper, our focus is on challenging the vulnerability of probabilistic copyright protection methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.216, + 0.635 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.357, + 0.66 + ], + "angle": 0, + "content": "3.1. Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.469, + 0.804 + ], + "angle": 0, + "content": "Take DDPM [17] for example, a typical diffusion process consists of a predefined forward process and a reverse process. Specifically, the forward process corrupts the original data \\(\\mathbf{x}_0 \\sim q(\\mathbf{x})\\) into a standard Gaussian noise \\(\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})\\) through \\(T\\) timesteps. In the reverse process, the denoising network \\(\\epsilon_{\\theta}\\) is learned to denoise the corrupted \\(\\mathbf{x}_t\\) by predicting the sampled noise \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\) that added to \\(\\mathbf{x}_0\\). The objective of the diffusion model can be simplified into the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.813, + 0.469, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Text-to-image diffusion models further use prompts to guide the sampling process for generating desired images. Take Stable Diffusion [37] for example, they incorporate a pre-trained CLIP [33] as text encoder \\(\\pmb{\\tau}\\) to encode an input" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "text \\(\\mathbf{y}\\), where \\(\\mathbf{c} = \\boldsymbol{\\tau}(\\mathbf{y})\\). An image encoder \\(\\mathcal{E}\\) is employed to map an input image \\(\\mathbf{x}_0\\) into its latent representation \\(\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)\\). The training objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.146, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\boldsymbol {\\epsilon}, \\mathbf {c}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| _ {2} ^ {2} \\right] \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.172, + 0.869, + 0.188 + ], + "angle": 0, + "content": "3.2. Near Access-Freeness Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.362 + ], + "angle": 0, + "content": "As a pioneering work of probabilistic copyright protection method, Vyas et al. [52] formally defines Near Access-Freeness (NAF) to provide a probabilistic guarantee against the generation of infringing content. They provide practical algorithms (CP-\\(k\\)) to protect an arbitrary generative model \\(p\\) from copyright infringement. Suppose we have a cover of safe models \\(\\mathcal{S}\\) satisfying that for any piece of copyrighted data \\(y_{C}\\), there exists some \\(q = \\text{safe}_{C} \\in \\mathcal{S}\\) trained without access to \\(y_{C}\\). The copyright protection at a pre-given threshold \\(k\\) can be achieved by sampling \\(y \\sim p(\\cdot | x)\\) with a prompt \\(x\\) and accepting \\(y\\) if:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.37, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\rho (y | x) := \\max _ {q \\in S} \\log \\frac {p (y | x)}{q (y | x)} \\leq k \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Intuitively, \\(\\rho(y|x)\\) serves as a criterion to distinguish non-infringing content from infringing content, while \\(k\\) determines the threshold for judgment. Let \\(p_k\\) be the protected model with threshold \\(k\\) and \\(\\mathcal{Y}_C\\) be the set of infringing contents, CP-\\(k\\) provides the following probabilistic upper-bound for copyright infringement:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.511, + 0.892, + 0.545 + ], + "angle": 0, + "content": "\\[\np _ {k} (y \\in \\mathcal {V} _ {C} | x) \\leq \\frac {2 ^ {k}}{\\nu_ {k} (x)} \\cdot \\operatorname {s a f e} _ {C} (y \\in \\mathcal {V} _ {C} | x) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.614 + ], + "angle": 0, + "content": "where \\(\\nu_{k}(x)\\) denotes the acceptance rate of \\(p_k\\). For different prompts \\(x\\), the optimal choice of \\(k\\) changes with varying distribution of \\(\\rho(y|x)\\). In the subsequent sections of our paper, we use \\(k_x\\) to emphasize the dependency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.591, + 0.642 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.702, + 0.666 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this paper, we consider a text-to-image generative model \\(\\tilde{p}\\) equipped with probabilistic copyright protection. The probability that \\(\\tilde{p}\\) generates infringing content with prompt \\(x\\) is upper-bounded. Following the typical adversarial prompt attack setting, we consider a malicious attacker who seeks to manipulate \\(\\tilde{p}\\) to produce content that violates the copyright of a specific piece of target copyrighted data \\(y_{C}\\), of which the infringing contents form \\(\\mathcal{Y}_C \\subseteq \\mathcal{Y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Different from the standard paradigm, we consider an online attack scenario where the attacker is allowed to interact with \\(\\tilde{p}\\) for \\(T\\) times within each attack trial. During the \\(t = 1,\\dots ,T\\)-th interaction, the attacker inputs a prompt \\(x_{t}\\) and receives a generated sample \\(y_{t}\\sim \\tilde{p} (\\cdot |x_{t})\\). Besides furnishing a prompt, the attacker is prohibited from intervening directly in the generation process. This constraint" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "is consistent with real-world scenarios where users interact with generative models as black boxes through APIs. At the end of each attack trial, the attacker will select \\( y^{*} \\in \\{y_{t}\\}_{t=1}^{T} \\) as the final output, and the attack is regarded successful if \\( y^{*} \\in \\mathcal{V}_{C} \\). The formal objective of the attacker is to maximize the success rate of attack, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.2, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} \\left(y ^ {*} \\in \\mathcal {Y} _ {C}\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.232, + 0.47, + 0.263 + ], + "angle": 0, + "content": "4.2. Virtually Assured Amplification Attack on Probabilistic Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.376 + ], + "angle": 0, + "content": "Copyright protection mechanisms in generative models, e.g., \\(NAF\\), are designed to yield a low upper-bound on the probability of generating infringing samples. Such protection is confined to single or infrequent generations. However, in scenarios where malicious attackers engage in multiple targeted generation requests, the probability of a successful attack can be amplified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.665 + ], + "angle": 0, + "content": "We propose a universal online amplification attack framework, detailed in Algorithm 1, for a virtually assured success against probabilistic copyright protection. In this framework, the attacker is modeled as an adversarial prompt generator \\(\\mathcal{A}(x_t|\\langle (x_s,y_s)^{t - 1}_s\\rangle)\\), which iteratively generates the prompt \\(x_{t}\\) for the current step \\(t\\) based on the interaction history with the generative model. For selecting the optimal sample, we employ a scoring function \\(\\mathcal{S}:\\mathcal{V}\\to \\mathbb{R}\\) to evaluate all the samples and return the highest-scoring sample as the final result. An attack is deemed successful if the score of the returned result is higher than the target score \\(S_{tar}\\). For example, an essential choice of the scoring function could be the indicator function \\(\\mathbb{I}(y\\in \\mathcal{V}_C)\\), with the target score \\(S_{tar} = 0\\), which ensures that any infringing sample generated will be returned as the final result, making the attack as successful. In practice, to reduce the manual effort in identifying copyright infringement, a computable surrogate scoring function with a target score may be utilized to establish a standard for copyright infringement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.726 + ], + "angle": 0, + "content": "We introduce the following theorem, which suggests the virtually assured success of the amplification attack with an assumption on the lower-bound of success probability for each single-shot attack attempt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.47, + 0.814 + ], + "angle": 0, + "content": "Theorem 1. Following the notations in Algorithm 1, for any \\(\\varepsilon \\in (0,1)\\), the attack is successful with probability at least \\(1 - \\varepsilon\\) if \\(T > \\log_{1 - \\sigma} \\varepsilon\\), where \\(\\sigma > 0\\) is a strictly positive lower-bound on the success probability shared by every single attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The intuition behind the amplification attack and Theorem 1 is straightforward. If the attacker is granted sufficient sampling opportunities with a set of prompts that yield even a modest lower-bound on the single success probability, as the number of sampling attempts increases, the probability" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 Amplification Attack on Probabilistic Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.126, + 0.892, + 0.185 + ], + "angle": 0, + "content": "Require: Generative model \\(\\tilde{p}\\) with probabilistic copyright protection, target copyrighted data \\(\\mathcal{C}\\), adversarial prompt generator \\(\\mathcal{A}\\), maximum number of steps \\(T\\), score function \\(\\mathcal{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.187, + 0.664, + 0.2 + ], + "angle": 0, + "content": "1: for \\( t = 1, \\dots, T \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.201, + 0.822, + 0.217 + ], + "angle": 0, + "content": "2: Sample prompt \\( x_{t} \\sim \\mathcal{A}(\\cdot |\\langle (x_{s},y_{s})\\rangle_{s = 1}^{t - 1}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.217, + 0.799, + 0.232 + ], + "angle": 0, + "content": "3: Feed \\( x_{t} \\) to \\( \\tilde{p} \\) and receive \\( y_{t} \\sim \\tilde{p}(\\cdot | x_{t}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.233, + 0.584, + 0.245 + ], + "angle": 0, + "content": "4: end for" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.247, + 0.764, + 0.264 + ], + "angle": 0, + "content": "5: return \\( y^{*} = \\arg \\max_{y\\in \\{y_{1:T}\\}}S(y) \\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.187, + 0.822, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.402 + ], + "angle": 0, + "content": "of generating infringing samples accumulates at an exponentially fast rate, eventually making an attack almost guaranteed to succeed. For instance, suppose our attack possesses a relatively low single-shot success probability, say \\(1\\%\\). Provided that we are permitted to repeat the attack approximately \\(\\log_{0.99} 0.01 \\approx 459\\) times, the probability of at least one successful instance is then amplified to \\(99\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.412, + 0.726, + 0.429 + ], + "angle": 0, + "content": "4.3. Online Prompt Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.572 + ], + "angle": 0, + "content": "In this section, we consider a specific scenario within the framework described in Sec. 4.2, where the conditional prompt generator (the attacker) is restricted to select a prompt among \\( K \\) candidate prompts \\( \\{x^{1},\\dots ,x^{K}\\} \\), based on previous choices and the scores of received samples. The generation of candidate prompts can either be independent of or tailored specifically for certain models and copyright protection mechanisms, as we will discuss in Sec. 4.4, designed for diffusion model and NAF copyright protection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Let \\( a_{t} \\in \\{1, \\dots, K\\} \\) denote the decision at step \\( t \\) and \\( \\pi \\) denote the policy of prompt selection, the prompt generation procedure can be formally described as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.629, + 0.892, + 0.648 + ], + "angle": 0, + "content": "\\[\nx _ {t} = x ^ {a _ {t}}, a _ {t} \\sim \\pi (\\cdot | \\langle (x _ {s}, r _ {s} = \\mathcal {S} (y _ {s})) \\rangle_ {s = 1} ^ {t - 1}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this context, online prompt generation is reformulated as a variant of the multi-armed bandit problem, where each candidate prompt is conceptualized as an arm, and the score of the corresponding sample is the reward for pulling the arm. Unlike the classic multi-armed bandit framework, our goal is to maximize the probability of obtaining a reward exceeding a specified threshold \\( S_{tar} \\) at least once within \\( T \\) trials. This involves the trade-off between exploration and exploitation. Herein, we present two variants of the \\( \\varepsilon \\)-greedy algorithm. Initially, akin to the conventional algorithm, we conduct \\( m \\leq \\lfloor T / K \\rfloor \\) trials for each arm. For the rest \\( T - mK \\) steps, a random action will be taken with probability \\( \\varepsilon > 0 \\) for exploration. Otherwise, the best action \\( \\hat{a}_t^* = \\arg \\max_a \\hat{Q}_t(a) \\) according to the evaluation \\( \\hat{Q}_t(a) \\) will be taken for exploitation. The choice of \\( \\hat{Q}_t \\) diverges into two distinct variants." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.265 + ], + "angle": 0, + "content": "\\(\\varepsilon\\) greedy-max Inspired by the fact that only the max reward matters, the maximum reward received so far is employed for evaluation, i.e., \\(\\hat{Q}_t(a) = \\max \\{r_s : a_s = a\\}_{s=1}^t\\). \\(\\varepsilon\\) greedy-cdf Essentially, we aim to maximize the probability that the reward for the next step exceeds a threshold \\(S_{tar}\\). If we assume that the reward distribution for each arm adheres to a normal distribution, we have \\(\\hat{Q}_t(a) = 1 - \\Phi\\left(\\frac{S_{tar} - \\hat{\\mu}_t(a)}{\\hat{\\sigma}_t(a)}\\right)\\), where \\(\\hat{\\mu}_t(a) = \\frac{1}{N_t(a)} \\sum_{s=1}^t r_s \\mathbb{I}(a_s = a)\\), \\(\\hat{\\sigma}_t^2(a) = \\frac{1}{N_t(a) - 1} \\sum_{s=1}^t (r_s \\mathbb{I}(a_s = a) - \\hat{\\mu}_t(a))^2\\), and \\(\\Phi\\) denotes the cumulative distribution function (cdf) of standard normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.266, + 0.472, + 0.311 + ], + "angle": 0, + "content": "Our empirical findings suggest that a well-calibrated balance between exploration and exploitation can enhance the stability of the attack." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.318, + 0.471, + 0.351 + ], + "angle": 0, + "content": "4.4. Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.47, + 0.419 + ], + "angle": 0, + "content": "In this section, we narrow the scope of discussion to adversarial prompt discovery against CP-\\(k\\), a NAF-based copyright protection algorithm. CP-\\(k\\) modifies the probability density function of an unprotected generative model \\(p\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.426, + 0.47, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} (y | x) \\propto p (y | x) \\mathbb {I} (\\rho (y | x) \\leq k _ {x}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.47, + 0.51 + ], + "angle": 0, + "content": "where \\(\\rho\\) is defined in Eq. (3) and \\(k_{x}\\) is a prompt-dependent threshold determined by the generative system, which is typically inaccessible to the user. CP-\\(k\\) provides a probabilistic upper-bound for copyright infringement in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Building upon our initial assumptions, we further assume that the attacker is granted full access, i.e., knowing the structures and parameters of these models, to the unprotected generative model \\( p \\) and safe models \\( q \\in S \\). While this white-box setting renders our attack narrowed to open-source models and those at risk of backend leakage, the potential threat of our attack remains substantial, as it does not require intervention in the generation process. This means that if a malicious attacker gains access to a leaked model and discovers adversarial prompts, any other users, lacking access to the underlying models, can readily replicate these adversarial prompts to reproduce the attack." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.692, + 0.47, + 0.738 + ], + "angle": 0, + "content": "Recall that our objective is to find adversarial prompts capable of inducing models protected by CP-\\(k\\) to generate infringing content with a strictly positive probability, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.745, + 0.469, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.773, + 0.47, + 0.821 + ], + "angle": 0, + "content": "We deduce a lower-bound for the infringement probability with optimal adversarial prompt in Theorem 2, subsequently guiding the development of a practical algorithm." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Definition 1 (Local Continuity). Given a distance measure \\(\\mathcal{D}\\) defined in \\(\\mathcal{V}\\), a model \\(p\\) is called \\((\\epsilon, \\alpha)\\)-local-continuous around \\(y_0 \\in \\mathcal{V}\\) if for any prompt \\(x \\in \\mathcal{X}\\), there exists \\(\\epsilon, \\alpha > 0\\) such that for any \\(y \\in \\mathcal{B}_{\\mathcal{D}}(y_0, \\epsilon) := \\{y \\in \\mathcal{Y} : \\mathcal{D}(y_0, y) < \\epsilon\\}\\), \\(|p(y_0|x) - p(y|x)| < \\alpha \\mathcal{D}(y_0, y)\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Algorithm 2 Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection" + }, + { + "type": "algorithm", + "bbox": [ + 0.502, + 0.126, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Require: Denoising network \\(\\epsilon_{\\theta_p},\\epsilon_{\\theta_q}\\) , text encoder \\(\\pmb{\\tau}\\) , target \\(y_{C}\\) , denoising steps \\(T\\) , optimization step \\(S\\) , loss clip bound \\(\\varphi\\) , loss weight \\(\\lambda\\) , learning rate \\(\\gamma\\) \n1: \\(\\mathbf{P} = \\{\\mathbf{e}_1,\\dots ,\\mathbf{e}_n\\} \\sim E^{|\\mathcal{V}|\\times d}\\) \n2: for 1,.., S do \n3: \\(x = \\mathrm{Proj}(\\mathbf{P})\\) \n4: Diffusion reconstruction task \n5: \\(t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})\\) \n6: \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n7: \\(y_{t} = \\sqrt{\\bar{\\alpha}_{t}} y_{C} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon\\) \n8: \\(\\mathcal{L}_p = \\| \\epsilon -\\epsilon_{\\theta_p}(y_t,t,\\tau (x))\\| _2^2\\) \n9: \\(\\mathcal{L}_q = \\| \\epsilon -\\epsilon_{\\theta_q}(y_t,t,\\tau (x))\\| _2^2\\) for all \\(q\\in S\\) \n10: Calculate the gradient w.r.t projected embedding \n11: \\(g = \\nabla_{x}(\\lambda \\cdot \\max (\\mathcal{L}_{p},\\varphi) + (1 - \\lambda)\\cdot \\max_{q\\in S}(\\mathcal{L}_{q}))\\) \n12: Apply the gradient on continuous embedding \n13: \\(\\mathbf{P} = \\mathbf{P} - \\gamma g\\) \n14: end for \n15: return Proj(P)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.894, + 0.52 + ], + "angle": 0, + "content": "Theorem 2. Assume there is a distance measure \\(\\mathcal{D}\\) defined in \\(\\mathcal{V}\\) such that (i) \\(p\\) is \\((\\epsilon_p, \\alpha)\\)-local-continuous around \\(y_C\\), (ii) every \\(q \\in S\\) is local-continuous around \\(y_C\\), and (iii) there exists \\(\\epsilon_c > 0\\) such that \\(\\mathcal{B}_{\\mathcal{D}}(y_C, \\epsilon_c) \\subseteq \\mathcal{V}_C\\). The objective defined in Eq. (8) has the following lower-bound for any \\(\\eta, \\delta > 0\\)," + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.533, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\geq \\max _ {x \\in \\tilde {\\chi} _ {\\eta , \\delta}} \\eta C _ {1} - \\alpha C _ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.892, + 0.606 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathcal{X}}_{\\eta ,\\delta} = \\{x\\in \\mathcal{X}:p(y_C|x)\\geq \\eta ,\\rho (y_C|x) < k_x - \\delta \\}\\) and \\(C_1,C_2\\) are constants independent on \\(x\\) given as" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.616, + 0.88, + 0.651 + ], + "angle": 0, + "content": "\\[\nC _ {1} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} d y, C _ {2} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} \\mathcal {D} (y _ {C}, y) d y,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.661, + 0.892, + 0.696 + ], + "angle": 0, + "content": "where \\(\\epsilon = \\min (\\epsilon_p,\\epsilon_c,\\epsilon_\\rho)\\) with \\(\\epsilon_{\\rho}\\coloneqq \\inf_{x\\in \\tilde{\\mathcal{X}}_{\\eta ,\\delta}}\\sup \\{\\epsilon :\\) \\(\\rho (y|x) < k_x,\\forall y\\in \\mathcal{B}_D(y_C,\\epsilon)\\} .\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.827 + ], + "angle": 0, + "content": "To ensure that the lower-bound in the Theorem 2 is nontrivial, we need to set \\(\\eta > \\alpha C_2 / C_1\\) and search along prompts satisfying (a) \\(\\rho(y_C|x) < k_x\\), (b) \\(p(y_C|x) \\geq \\eta\\). Unfortunately, the mechanism to determine the threshold \\(k_x\\) is assumed inaccessible to the attacker. To obtain a feasible optimization objective, we instead minimize \\(\\rho(y_C|x)\\) subject to \\(p(y_C|x) \\geq \\eta\\), as an alternative. The optimization objective then becomes" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.843, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\min _ {x} \\rho \\left(y _ {C} | x\\right) \\quad \\text {s . t .} p \\left(y _ {C} | x\\right) \\geq \\eta . \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We follow the reconstruction task of the original text-to-image diffusion models for direct optimization. In aligning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12367" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "with the attack objective of Eq. (10), we develop the optimization objective from two aspects:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.133, + 0.47, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon}, x} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta_ {p}} (y _ {t}, t, \\boldsymbol {\\tau} (x)) \\| _ {2} ^ {2} ] \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.163, + 0.469, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {q} = \\mathbb {E} _ {t, \\epsilon , x} [ \\| \\epsilon - \\epsilon_ {\\theta_ {q}} (y _ {t}, t, \\tau (x)) \\| _ {2} ^ {2} ] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.188, + 0.469, + 0.233 + ], + "angle": 0, + "content": "Where \\(y_{t}\\) is the noisy version of target copyrighted data \\(y_{C}\\) at denoising step \\(t\\); \\(\\epsilon_{\\theta_p}, \\epsilon_{\\theta_q}\\) is the denoising network of generative model \\(p\\) and safe model \\(q \\in S\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.469, + 0.326 + ], + "angle": 0, + "content": "While \\(\\mathcal{L}_p\\) is designed to increase the possibility of model \\(p\\) in generating the desired infringed content \\(y_{C}\\), the minimization of \\(\\mathcal{L}_q\\) corresponds to the maximization of \\(q(y_C|x)\\) consequently leading to the reduction of \\(\\rho (y_C|x)\\). Hence, the overall optimization objective can be formulated as a weighted sum of \\(\\mathcal{L}_p\\) and \\(\\mathcal{L}_q\\):" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.339, + 0.469, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\cdot \\max \\left(\\mathcal {L} _ {p}, \\varphi\\right) + (1 - \\lambda) \\cdot \\max _ {q \\in \\mathcal {S}} \\left(\\mathcal {L} _ {q}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.469, + 0.399 + ], + "angle": 0, + "content": "Where \\(\\varphi\\) is a loss clip to mitigate the conflict between \\(\\mathcal{L}_p\\) and \\(\\mathcal{L}_q\\) in minimizing \\(\\rho\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.671 + ], + "angle": 0, + "content": "The complete Anti-NAF prompt optimization process is detailed in Algorithm 2. We conduct optimization in the continuous embedding space with a sequence of learnable embeddings \\(\\mathbf{P} = \\{\\mathbf{e}_1, \\ldots, \\mathbf{e}_n | \\mathbf{e}_i \\in \\mathbb{R}^d\\}\\), where \\(n\\) is the sequence length and \\(d\\) is the embedding dimension. Given that each word token \\(w\\) in a vocabulary \\(\\mathcal{V}\\) can be represented as a corresponding embedding \\(\\mathrm{EMB}(w)\\) using an embedding matrix \\(E^{|\\mathcal{V}| \\times d}\\), each embedding \\(\\mathbf{e}\\) in the sequence can be mapped to its nearest token embedding in \\(\\mathcal{V}\\) under some similarity metric, e.g., cosine similarity. The projection function can be defined as \\(\\mathrm{Proj}(\\mathbf{e}) = \\arg \\min_{w \\in \\mathcal{V}} \\cos(\\mathrm{EMB}(w), \\mathbf{e})\\). The resultant text prompt is a sequence of these projected tokens: \\(x = \\mathrm{Proj}(\\mathbf{P}) = \\{\\mathrm{Proj}(\\mathbf{e}_1), \\ldots, \\mathrm{Proj}(\\mathbf{e}_n)\\}\\). In addition, following the approach in [53], the continuous embeddings \\(\\mathbf{P}\\) are projected into discrete tokens \\(x\\) for each forward pass in every optimization step. The gradient of the projected prompt \\(x\\) is then applied to update the continuous embeddings \\(\\mathbf{P}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.21, + 0.701 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.284, + 0.726 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.35, + 0.747 + ], + "angle": 0, + "content": "5.1.1 Evaluated Datasets and Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Since training large text-to-image diffusion models from scratch is impractical, we fine-tune the pre-trained StableDiffusion-v1-4 model provided by Huggingface with two datasets: POKEMON [30] and LAION-mi [10]. Note that both datasets do not overlap with the pre-training dataset of Stable Diffusion, ensuring that the safe models fine-tuned on them are inaccessible to copyrighted data. Following [52], each dataset is split into two disjoint shards" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\) to train generative model \\(q_{1}\\) and \\(q_{2}\\) respectively. For copyrighted data \\(y_{C}\\), \\(q_{2}\\) is served as safe model safe \\(C\\) if \\(y_{C}\\in \\mathcal{D}_{1}\\) but \\(y_{C}\\notin \\mathcal{D}_{2}\\). For better experimental illustrations, each copyrighted data is repeated to make up \\(1\\%\\) of the dataset shard." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.243 + ], + "angle": 0, + "content": "POKEMON [30]. This dataset consists of 833 image-caption pairs, where captions are obtained from the BLIP model [22]. We evaluate our attack on 5 copyrighted images added to one of the shards separately and fine-tune models \\( q_{1} \\) and \\( q_{2} \\) on each shard for 5000 steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.893, + 0.381 + ], + "angle": 0, + "content": "LAION-mi [10]. The dataset is originally constructed for membership inference attacks on diffusion models. We only use the nonmembers part of this dataset, which holds a similar data distribution but is disjoint with the pre-training dataset of Stable Diffusion. We evaluate our attack on 5 selected copyrighted images, which are separately combined with one of non-copyrighted data shards. Each dataset shard is of size 5000. We fine-tune model \\( q_{1} \\) and \\( q_{2} \\) for 25,000 steps. Details for fine-tuning can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.715, + 0.417 + ], + "angle": 0, + "content": "5.1.2 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.593 + ], + "angle": 0, + "content": "For the sampling setting of text-to-image diffusion models, images of size \\(512 \\times 512\\) are generated using a classifier-free guidance scale of 7.5 and 50 sampling steps with the default scheduler. The amplification step is set to 100 and 500 for POKEMON and LAION-mi respectively. For the prompt optimization process, the learning rate is set to 0.01 using Adagrad [11] optimizer with 25,000 optimization steps, the gradient accumulation step is set to 5, the loss clip \\(\\varphi\\) for model \\(p\\) is 0.01, and the loss weight \\(\\lambda\\) is 0.95. The length of optimized prompts is set to 8 tokens. All experiments are conducted on A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.687, + 0.621 + ], + "angle": 0, + "content": "5.2. Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We evaluate with a model \\( p \\) behaves with an equal chance of sampling from either model \\( q_{1} \\) or \\( q_{2} \\). We consider such model \\( p \\) as a strict protection model that even with the original caption as prompt, there is a \\( 50\\% \\) chance of generating from the safe model. In contrast, a model \\( p \\) trained on the entire dataset could frequently produce highly similar generations to the target copyrighted image, making it extremely inefficient for the selection of a valid threshold \\( k \\). Next, we will separately discuss the data, the metrics, and the threat prompts for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation Data. We use samples generated by \\( p \\) for evaluation. To determine whether they infringe the copyright of the target image, we rely on the similarity between samples and the target image as there is no widely acknowledged computable standard for infringement to our knowledge. We employ the current SOTA similarity metric SSCD [31] for copy detection. As SSCD scores of infringed samples for different target images differ significantly by obser" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.388, + 0.901 + ], + "angle": 0, + "content": "1https://huggingface.co/CompVis/stable-diffusion-v1-4" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.136, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.242, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.14, + 0.223, + 0.148 + ], + "angle": 0, + "content": "a black and white zebrastanding next to a star" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.09, + 0.349, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.14, + 0.325, + 0.145 + ], + "angle": 0, + "content": "zebra weet unex mee" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.09, + 0.404, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.09, + 0.512, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.14, + 0.498, + 0.147 + ], + "angle": 0, + "content": "a picture of a white horsewith orange and yellow slings" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.615, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.09, + 0.674, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.09, + 0.78, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.14, + 0.762, + 0.147 + ], + "angle": 0, + "content": "a cartoon rabbit with asocce around its neck" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.09, + 0.885, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.14, + 0.866, + 0.147 + ], + "angle": 0, + "content": "rabbit, lamewridayday catunmibrated the room" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.153, + 0.136, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.153, + 0.242, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.196, + 0.226, + 0.201 + ], + "angle": 0, + "content": "Giraffes Canvas Art Print" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.153, + 0.348, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.193, + 0.326, + 0.198 + ], + "angle": 0, + "content": "giraffe guintspitnoche" + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.198, + 0.325, + 0.203 + ], + "angle": 0, + "content": "tallent big etary" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.153, + 0.405, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.153, + 0.512, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.193, + 0.511, + 0.205 + ], + "angle": 0, + "content": "Product photo wall tattoo children's rostradberries with the inksticker set" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.153, + 0.617, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.193, + 0.604, + 0.198 + ], + "angle": 0, + "content": "horse millennials symbolism" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.198, + 0.604, + 0.203 + ], + "angle": 0, + "content": "kano jacoblissseasoned" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.153, + 0.674, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.153, + 0.78, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.193, + 0.772, + 0.2 + ], + "angle": 0, + "content": "Penguins pictures of the New" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.2, + 0.772, + 0.205 + ], + "angle": 0, + "content": "Year, March with their own hands" + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.153, + 0.836, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.193, + 0.875, + 0.2 + ], + "angle": 0, + "content": "penguins knobs swimming" + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.2, + 0.875, + 0.205 + ], + "angle": 0, + "content": "orbcsport orbfunday puppies have" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.893, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Visualization of generated images on different copyright targets. The examples in the first and second rows are selected from POKEMON and LAION-mi respectively. The prompts used to generate output are given below each group of images. Remarkably, the copyright-infringed content generated with Anti-NAF amplification reveals the vulnerability of probabilistic copyright protection CP-\\(k\\)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.273, + 0.885, + 0.41 + ], + "angle": 0, + "content": "
MethodsPOKEMONLAION-mi
CIRFAR@5%AR↑FAR@15%AR↑CIRFAR@10%AR↑FAR@30%AR↑FAR@50%AR↑
Caption (w/o Amp.)47.00%0.40%3.28%49.64%0.00%0.00%0.04%
CLIP-Int. (w/o Amp.)26.92%0.84%2.20%48.12%0.84%1.52%2.96%
PEZ (w/o Amp.)7.80%1.32%2.76%15.80%0.08%0.24%0.40%
Anti-NAF (w/o Amp.)12.88%8.52%10.00%33.84%2.64%4.16%7.00%
Caption (w/ Amp.)100.00%13.64%40.16%100.00%0.00%0.00%14.64%
CLIP-Int. (w/ Amp.)99.72%22.84%47.32%100.00%34.84%37.12%75.60%
PEZ (w/ Amp.)66.92%22.80%37.48%98.16%38.04%50.68%59.64%
Anti-NAF (w/ Amp.)99.84%77.36%91.36%100.00%56.12%73.32%95.92%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.464 + ], + "angle": 0, + "content": "Table 1. Quantitative results. The performance with the amplification attack (\"w/ Amp.\") is significantly superior to scenarios without amplification (\"w/o Amp.\") Additionally, our proposed Anti-NAF demonstrates notably promising outcomes for providing a substantial probability of copyright infringement when probabilistic protection is applied. (CLIP-Int. is the abbreviation for CLIP-Interrogator)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.593 + ], + "angle": 0, + "content": "vation (shown in Fig. 5), we do not indicate a fixed score threshold for all target images. Instead, we use relative thresholds determined as percentiles of the similarity scores among samples generated by \\( p \\) with the original caption of the target image as the prompt, e.g. SSCD-50%. Recognizing the variable criteria of infringement, we report results with other choices of thresholds in Sec. 11.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.801 + ], + "angle": 0, + "content": "Evaluation Metrics. The CP-\\(k\\) method achieves copyright protection by selectively accepting generated samples using the threshold \\(k_{x}\\), which can be indicated by the Acceptance Rate (AR). For a given AR, a good protection system is expected to rarely accept infringing content, i.e., have a low False Accept Rate (FAR), as defined in Eq. (8). It is worth noting that the choice of \\(k_{x}\\) dictates the trade-off between model safety and efficiency. Furthermore, to our knowledge, there is no principled way to determine \\(k_{x}\\). As a result, we evaluate the success of attack by reporting the FAR at different AR, e.g., FAR@5%AR. Additionally, the copyright infringement rate (CIR) is also presented for scenarios without copyright protection, i.e., \\(\\mathrm{AR} = 100\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Threat Prompts. Given that generating images similar to the target image is a necessary condition for a successful attack, we focus on the following threat prompts: (i) Target caption: the original caption of the target image; (ii) PEZ [53]: a gradient-based discrete optimization approach to discover prompts semantically similar to the target im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.893, + 0.64 + ], + "angle": 0, + "content": "age; (iii) \\(CLIP\\)-Interrogator2: a prompt consists of the BLIP caption of the target image and top-\\(k\\) keywords greedily sampled from a keywords collection; (iv) Anti-NAF: an adversarial prompt against NAF copyright protection as described in Sec. 4.4. For each threat prompt, we sample 6400 images from model \\(p\\) for evaluation. For the online prompt selection setting, we take the aforementioned three prompts (exclude the target caption) as a prompt candidate set and consider \\(\\varepsilon\\)-greedy-max-/cdf Bandit Amplification as bandit strategies as described in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.653, + 0.593, + 0.667 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.893, + 0.874 + ], + "angle": 0, + "content": "In Fig. 2, we show example outputs under four scenarios. It is evident that with amplification, the CP-\\(k\\) protected generative model does indeed output copyright-infringed content with high probability, especially when using the Anti-NAF prompt. These examples suggest that probabilistic safeguards against copyright infringement are vulnerable to the amplification attack. Additional visualizations of outputs on various copyrighted images are demonstrated in Fig. 3. In Tab. 1, we report CIRs and FARs at various ARs on two datasets. Notably, there is a significant growth in both metrics when amplification attack is employed, highlighting its effectiveness in amplifying the probability of infringing generations. The superior performance of Anti-NAF under" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.806, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/pharmapsychotic/clip-interrogator" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12369" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.273, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.09, + 0.468, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.271, + 0.272, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.271, + 0.468, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.43, + 0.27, + 0.455 + ], + "angle": 0, + "content": "(c) FAR-AR of Anti-NAF with different Amp. steps on POKEMON." + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.43, + 0.468, + 0.455 + ], + "angle": 0, + "content": "(d) FAR-AR of Anti-NAF with different Amp. steps on LAION-mi." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Figure 4. The overall FAR-AR curves. The results show that amplification is strongly effective in amplifying the possibility of infringed output, even with small amplification steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.765 + ], + "angle": 0, + "content": "scores its efficacy in rendering infringing generations with a substantial probability. The overall FAR-AR curves are illustrated in Fig. 4. We can observe that bandit variants of amplification lead to a smaller variance across different target copyrighted images, especially at lower acceptance rates, indicating that our bandit strategies achieve a more steady attack. In Figs. 4c and 4d, we plot the overall FAR-AR curves over different amplification steps. There is a clear trend of rapidly improved performance with increased amplification steps, due to the cumulative probability of generating infringing samples. This finding underlines the potential risk in practical applications of probabilistic copyright protections, given the high frequency of daily interactions with text-to-image generative models. We provide additional results and human evaluations in Sec. 11." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.232, + 0.789 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In Tab. 2, we study how different components of our optimization objective in Eq. (13), affect the performance. We discover that an exclusive focus on minimizing \\(\\mathcal{L}_p\\) (\"\\(\\mathcal{L}_p\\) only\") leads to a notable drop in performance. This indicates that prompts designed merely to reconstruct the target image are insufficient for a successful attack, as the infringing output of such a prompt can be easily identified and" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.09, + 0.868, + 0.166 + ], + "angle": 0, + "content": "
MethodsCIRFAR@5%AR↑FAR@15%AR↑
Anti-NAF12.88%8.52%10.00%
\\( \\mathcal{L}_p \\) only10.60%1.84%2.04%
w/o \\( \\varphi \\)12.12%1.36%1.48%
w/o \\( \\mathcal{L}_q \\)13.72%2.00%2.76%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.176, + 0.889, + 0.19 + ], + "angle": 0, + "content": "Table 2. Ablation study for Anti-NAF algorithm on POKEMON." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.354 + ], + "angle": 0, + "content": "blocked by the copyright protection system. Furthermore, a direct combination of \\(\\mathcal{L}_q\\) with \\(\\mathcal{L}_p\\) (\"w/o \\(\\varphi\\)) results in additional performance degradation because of the conflicting objectives of \\(\\mathcal{L}_p\\) and \\(\\mathcal{L}_q\\) in minimizing \\(\\rho(y_C|x)\\). On the other hand, we can observe a slight performance improvement when implementing a loss clip \\(\\varphi\\) on \\(\\mathcal{L}_p\\) (\"w/o \\(\\mathcal{L}_q\\)) to constrain the learning of \\(\\mathcal{L}_p\\). Overall, these results validate the effectiveness of our well-balanced optimization objective. Additional ablation studies are provided in Sec. 12." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.368, + 0.619, + 0.383 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.756 + ], + "angle": 0, + "content": "In this paper, we shed light on the vulnerability of probabilistic copyright protection methods for text-to-image generative models, especially in real-world scenarios involving persistent and targeted interactions. Our proposed Virtually Assured Amplification Attack (VA3) framework presents the feasibility of inducing the protected model to generate copyright-infringed content with an amplified probability. Despite our focus on a narrow scenario of online prompt selection within this framework, the experimental results highlight its effectiveness in challenging even the most advanced existing probabilistic copyright protection methods. However, a broader scope of potential strategies remains unexplored within the VA3 framework, such as online prompt optimization, which may provide more powerful attacks against copyright protection. Furthermore, our Anti-NAF algorithm relies on access to generative models for adversarial prompt optimization, an assumption that may not be satisfied in completely black-box scenarios. We leave these more complicated and general attack methods for future investigation. In conclusion, our findings emphasize the significant risk of copyright infringement when applying probabilistic copyright protection methods in practice. Therefore, we hope that this work can inspire the development of more robust copyright protection approaches." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.77, + 0.661, + 0.788 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The authors are grateful to Nikhil Vyas for providing the source code of CP-\\(k\\) algorithm. This research is partially supported by the National Research Foundation Singapore under the AI Singapore Programme (AISG Award No: AISG2-TC-2023-010-SGIL) and the Singapore Ministry of Education Academic Research Fund Tier 1 (Award No: T1 251RES2207)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12370" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Hossein Aboutalebi, Daniel Mao, Carol Xu, and Alexander Wong. Deepfakeart challenge: A benchmark dataset for generative ai art forgery and data poisoning detection. arXiv preprint arXiv:2306.01272, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Stephen Casper, Zifan Guo, Shreya Mogulothu, Zachary Marinov, Chinmay Deshpande, Rui-Jie Yew, Zheng Dai, and Dylan Hadfield-Menell. Measuring the success of diffusion models at imitating human artists. arXiv preprint arXiv:2307.04028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[6] David Chess. Some light infringement? https://ceoln.wordpress.com/2022/12/16/some-light-infringement/, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[7] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[8] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[9] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[10] Jan Dubiński, Antoni Kowalczuk, Stanisław Pawlak, Przemysław Rokita, Tomasz Trzeciński, and Pawel Morawiecki. Towards more realistic membership inference attacks on large diffusion models. arXiv preprint arXiv:2306.12983, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[11] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[12] Niva Elkin-Koren, Uri Hacohen, Roi Livni, and Shay Moran. Can copyright be reduced to privacy? arXiv preprint arXiv:2305.14822, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Hongcheng Gao, Hao Zhang, Yinpeng Dong, and Zhi-jie Deng. Evaluating the robustness of text-to-image diffusion models against real-world attacks. arXiv preprint arXiv:2306.13103, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. arXiv preprint arXiv:2303.15715, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[18] Matthew Butterick Joseph Saveri. Stable diffusion litigation. https://stablediffusionlitigation.com/, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Kenji Kawaguchi, Zhun Deng, Xu Ji, and Jiaoyang Huang. How does information bottleneck help deep learning? In International Conference on Machine Learning (ICML), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 12888-12900. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 20763-20786. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 21548-21566. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12371" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Natalie Maus, Patrick Chao, Eric Wong, and Jacob Gardner. Adversarial prompting for black box foundation models. arXiv preprint arXiv:2302.04237, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[29] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[30] Justin N. M. Pinkney. Pokemon blip captions. https://huggingface.co/datasets/lambdaabs/pokemon-blip-captions/, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[31] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14512-14522. IEEE, 2022. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[32] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. arXiv preprint arXiv:2305.13873, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[35] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Arkadip Ray and Somaditya Roy. Recent trends in image watermarking techniques for copyright protection: a survey. International Journal of Multimedia Information Retrieval, 9(4):249-270, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[41] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 29894-29918. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[42] Sarah Scheffler, Eran Tromer, and Mayank Varia. Formalizing human ingenuity: A quantitative framework for copyright law's substantial similarity. In Proceedings of the 2022 Symposium on Computer Science and Law, pages 37-49, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[43] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[45] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[46] Shawn Shan, Wenxin Ding, Josephine Passananti, Haitao Zheng, and Ben Y Zhao. Prompt-specific poisoning attacks on text-to-image generative models. arXiv preprint arXiv:2310.13828, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[47] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[48] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 6048-6058. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[49] Growthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[51] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12372" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[53] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[54] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[55] Yixin Wu, Ning Yu, Michael Backes, Yun Shen, and Yang Zhang. On the proactive generation of unsafe images from text-to-image models using benign prompts. arXiv preprint arXiv:2310.16613, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[56] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. arXiv preprint arXiv:2305.04175, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[58] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.503, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[59] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[60] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023 - Workshops, Vancouver, BC, Canada, June 17-24, 2023, pages 2385-2392. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[61] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12373" + } + ] +] \ No newline at end of file diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_origin.pdf b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c637d654896063092c0ea02213608cdc495c91a3 --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/294458f2-3d11-4022-9469-5c90ec662dfc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f52217422f54afd6de3e87a7873aaf4e37213cf97ee8813c5d636565813c6a41 +size 5867992 diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/full.md b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6bddb917d96afce866d4ec8afd8f31129e74e960 --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/full.md @@ -0,0 +1,420 @@ +# VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models + +Xiang Li* Qianli Shen* Kenji Kawaguchi National University of Singapore {xiangli, qianli, kenji} @ comp.nus.edu.sg + +![](images/0631ba4c845e782141a22326e2ad2e181b579a6c8d0cea4a869e2ce2abdc0a45.jpg) +Figure 1. Given a copyrighted image $y_{C}$ and a Text-to-Image (T2I) generative model with probabilistic copyright protection, our proposed virtually assured amplification attack (VA3) significantly amplifies the probability of producing infringing generations with persistent interactions of online adversarial prompt selection. + +# Abstract + +The booming use of text-to-image generative models has raised concerns about their high risk of producing copyright-infringing content. While probabilistic copyright protection methods provide a probabilistic guarantee against such infringement, in this paper, we introduce Virtually Assured Amplification Attack (VA3), a novel online attack framework that exposes the vulnerabilities of these protection mechanisms. The proposed framework significantly amplifies the probability of generating infringing content on the sustained interactions with generative models and a non-trivial lower-bound on the success probability of each engagement. Our theoretical and experimental results demonstrate the effectiveness of our approach under various scenarios. These findings highlight the potential risk of implementing probabilistic copyright protection in + +practical applications of text-to-image generative models. Code is available at https://github.com/South7X/VA3. + +# 1. Introduction + +In recent years, the advancement of large generative models [17, 47, 50] has revolutionized high-quality image synthesis [34, 37, 40], paving the way for commercial applications that enable the public to effortlessly craft their own artworks and designs [2, 13, 20, 27, 38, 39]. Nevertheless, these models exhibit notable memorization capabilities to produce generations highly similar to the training data [3]. This resemblance raises growing concerns about copyright infringement, especially when copyrighted data is used for training [12, 16, 49, 52]. + +To address these concerns, there has been a surge in research focused on protecting copyrighted data from potential infringement by outputs of generative models [14, 21, + +![](images/058a8d52955a1282d97e39bfa285658827fcbef3f22eafc817222f3ccbc682b0.jpg) +(a) Output w/o CP- $k$ + +![](images/aded96dea7d52576e4526e27ede13d16ef9395664899488443e80336beed8921.jpg) +(b) Output w/ CP-k. + +![](images/db87137dff15d6aafeff4e0ec47530872e9e84a098d99398dfbfac2c13495388.jpg) +(c) Amp. output w/ CP- $k$ +Figure 2. Example outputs given the copyright image in Fig. 1 as target (potential infringing images are marked with red boundaries). In (a), using a benign prompt, we observe a high incidence of infringing content from models without copyright protection ("w/o CP- $k$ )". In contrast, (b) shows that after applying the copyright protection mechanism ("w/ CP- $k$ )", all samples are safe as CP- $k$ rejects all infringing content. In (c), we find that amplification (Amp.) attack with a benign prompt results in limited success. Notably, by amplification attack with an adversarial prompt obtained from our proposed Anti-NAF algorithm, almost all output in (d) are copyright-infringed. + +![](images/589a8d6610525929211e58c24d80cd8dfef503320baadfbf37b019d813da4f53.jpg) +(d) Anti-NAF Amp. output w/ CP- $k$ + +24, 43, 45, 52, 58]. Among these studies, a pivotal concept involves establishing a probabilistic upper-bound against the generation of infringing content by generative models. We refer to this suite of approaches as probabilistic copyright protection. Most notably, Vyas et al. [52] introduce a mathematical definition of copyright known as near-access freeness (NAF). Their method enforces generative diffusion models to exhibit akin behaviors as safe models, which has no access to the copyrighted image. By leveraging the improbability of safe models generating infringing content, the probability of generative models doing the same is thereby substantially reduced. The copyright protection algorithm of NAF, CP- $k$ , can filter out infringing content generated by the models with high probability, even when the input prompts are adversarially designed. + +In this paper, we propose Virtually Assured Amplification Attack (VA3), a novel online attack framework, to show the vulnerability of probabilistic copyright protection. This framework induces text-to-image generative models with probabilistic copyright protection to generate infringing content. Our approach is grounded in the realization that in real-world scenarios, a malicious attacker intending to induce copyright infringement could engage in multiple interactions with the generative model via prompts. This persistent engagement poses a significant challenge to probabilistic protection methods, as it amplified the probability of producing infringing content of each generation. + +In our proposed framework, the attacker functions as a conditional prompt generator, creating adversarial prompts iteratively based on previous interactions with the generative model. Our primary theoretical result, Theorem 1, suggests that the amplification attack is guaranteed to succeed with high probability, given a sufficient number of interactions with the generative model and a strictly positive lower-bound on success probability of each single engagement. + +Regarding practical algorithms, our work encompasses two technical innovations. Firstly, we present effective strategies to manage the exploration-exploitation dilemma in online prompt selection, thereby enhancing the stability of the attack. Secondly, we propose Anti-NAF, a theoretically motivated adversarial prompt optimization algorithm tailored for NAF copyright protection, to generate prompts fulfilling the conditions of Theorem 1. + +Our experimental results validate the efficacy of our proposed online attack approach under diverse scenarios. These findings underline the potential copyright infringement risk of applying probabilistic protection in practical applications of text-to-image generative models, for both providers and users. + +# 2. Related Work + +# 2.1. Copyright Issues in Generative Models + +Text-to-image generative models, trained on large-scale datasets like LAION [44], have been equipped with enhanced memorization ability to generate outputs of high semantic similarity to their training data [3, 48]. Given the prevalence of copyrighted works in these datasets, the significant risk of copyright infringement for these generations has raised great concerns from the public [6, 18] and researchers [1, 4, 16, 42, 48, 49, 52]. Many efforts have been made to safeguard copyrighted materials from being infringed by generative diffusion models. Some researchers [23, 24, 41, 45] introduced data perturbation, where input data is modified to hinder the model to imitate copyrighted features. Another separate line of works [14, 21, 26, 43, 58] exploited concept removal that erases unsafe concepts from existing pre-trained diffusion models to mitigate the risk of undesirable generations. In an alternative approach, researchers studied watermark protection for copyrighted data + +[9, 29, 36, 54, 56, 59] to encode ownership information into potentially infringed outputs. + +A notable contribution, Vyas et al. [52] first provided a mathematical probabilistic upper-bound against copyright-infringed generation. They asserted that the proposed near access-freeness (NAF) offers robust guarantees for copyright protection. However, Elkin-Koren et al. [12] argued the limitation of this method for reducing copyright to a matter of privacy from a legal perspective. In this paper, we build upon these discussions to present a significant challenge to these probabilistic copyright protection methods through the amplification attack. + +# 2.2. Vulnerability of Diffusion Models + +A rising number of works focus on the vulnerability of diffusion models to various types of attacks. Poisoning attack [5, 8, 46, 55, 57] studies the problem of manipulating training data to induce unsafe behaviors in diffusion models during the training phase. In the context of the real-world inference phase for existing text-to-image diffusion models, other works also investigate the robustness and safety against different prompt inputs. Researchers [15, 25, 28, 60] have shown that injecting a slight perturbation to the input prompts will mislead the unprotected model to generate semantically unrelated images. Furthermore, with carefully crafted prompts, unsafe images can easily evade detection-based safety filters in text-to-image diffusion models [32, 35]. Other studies [7, 51] red-teaming concept removal copyright protection methods by finding problematic prompts that can recover erased unsafe concepts to yield undesirable generations. Diverging from these existing attack approaches on heuristic protection methods, in this paper, our focus is on challenging the vulnerability of probabilistic copyright protection methods. + +# 3. Preliminaries + +# 3.1. Text-to-Image Diffusion Models + +Take DDPM [17] for example, a typical diffusion process consists of a predefined forward process and a reverse process. Specifically, the forward process corrupts the original data $\mathbf{x}_0 \sim q(\mathbf{x})$ into a standard Gaussian noise $\mathbf{x}_T \sim \mathcal{N}(0, \mathbf{I})$ through $T$ timesteps. In the reverse process, the denoising network $\epsilon_{\theta}$ is learned to denoise the corrupted $\mathbf{x}_t$ by predicting the sampled noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ that added to $\mathbf{x}_0$ . The objective of the diffusion model can be simplified into the following form: + +$$ +\mathcal {L} _ {\mathrm {D D P M}} = \mathbb {E} _ {t, \mathbf {x} _ {0}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t) \| ^ {2} \right] \tag {1} +$$ + +Text-to-image diffusion models further use prompts to guide the sampling process for generating desired images. Take Stable Diffusion [37] for example, they incorporate a pre-trained CLIP [33] as text encoder $\pmb{\tau}$ to encode an input + +text $\mathbf{y}$ , where $\mathbf{c} = \boldsymbol{\tau}(\mathbf{y})$ . An image encoder $\mathcal{E}$ is employed to map an input image $\mathbf{x}_0$ into its latent representation $\mathbf{z}_0 = \mathcal{E}(\mathbf{x}_0)$ . The training objective is formulated as: + +$$ +\mathcal {L} _ {\mathrm {S D}} = \mathbb {E} _ {t, \mathbf {z} _ {0}, \boldsymbol {\epsilon}, \mathbf {c}} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\mathbf {z} _ {t}, t, \mathbf {c}) \| _ {2} ^ {2} \right] \tag {2} +$$ + +# 3.2. Near Access-Freeness Copyright Protection + +As a pioneering work of probabilistic copyright protection method, Vyas et al. [52] formally defines Near Access-Freeness (NAF) to provide a probabilistic guarantee against the generation of infringing content. They provide practical algorithms (CP- $k$ ) to protect an arbitrary generative model $p$ from copyright infringement. Suppose we have a cover of safe models $\mathcal{S}$ satisfying that for any piece of copyrighted data $y_{C}$ , there exists some $q = \text{safe}_{C} \in \mathcal{S}$ trained without access to $y_{C}$ . The copyright protection at a pre-given threshold $k$ can be achieved by sampling $y \sim p(\cdot | x)$ with a prompt $x$ and accepting $y$ if: + +$$ +\rho (y | x) := \max _ {q \in S} \log \frac {p (y | x)}{q (y | x)} \leq k \tag {3} +$$ + +Intuitively, $\rho(y|x)$ serves as a criterion to distinguish non-infringing content from infringing content, while $k$ determines the threshold for judgment. Let $p_k$ be the protected model with threshold $k$ and $\mathcal{Y}_C$ be the set of infringing contents, CP- $k$ provides the following probabilistic upper-bound for copyright infringement: + +$$ +p _ {k} (y \in \mathcal {V} _ {C} | x) \leq \frac {2 ^ {k}}{\nu_ {k} (x)} \cdot \operatorname {s a f e} _ {C} (y \in \mathcal {V} _ {C} | x) \tag {4} +$$ + +where $\nu_{k}(x)$ denotes the acceptance rate of $p_k$ . For different prompts $x$ , the optimal choice of $k$ changes with varying distribution of $\rho(y|x)$ . In the subsequent sections of our paper, we use $k_x$ to emphasize the dependency. + +# 4. Method + +# 4.1. Problem Formulation + +In this paper, we consider a text-to-image generative model $\tilde{p}$ equipped with probabilistic copyright protection. The probability that $\tilde{p}$ generates infringing content with prompt $x$ is upper-bounded. Following the typical adversarial prompt attack setting, we consider a malicious attacker who seeks to manipulate $\tilde{p}$ to produce content that violates the copyright of a specific piece of target copyrighted data $y_{C}$ , of which the infringing contents form $\mathcal{Y}_C \subseteq \mathcal{Y}$ . + +Different from the standard paradigm, we consider an online attack scenario where the attacker is allowed to interact with $\tilde{p}$ for $T$ times within each attack trial. During the $t = 1,\dots ,T$ -th interaction, the attacker inputs a prompt $x_{t}$ and receives a generated sample $y_{t}\sim \tilde{p} (\cdot |x_{t})$ . Besides furnishing a prompt, the attacker is prohibited from intervening directly in the generation process. This constraint + +is consistent with real-world scenarios where users interact with generative models as black boxes through APIs. At the end of each attack trial, the attacker will select $y^{*} \in \{y_{t}\}_{t=1}^{T}$ as the final output, and the attack is regarded successful if $y^{*} \in \mathcal{V}_{C}$ . The formal objective of the attacker is to maximize the success rate of attack, i.e., + +$$ +\max _ {x \in \mathcal {X}} P _ {y \sim \tilde {p} (\cdot | x)} \left(y ^ {*} \in \mathcal {Y} _ {C}\right) \tag {5} +$$ + +# 4.2. Virtually Assured Amplification Attack on Probabilistic Copyright Protection + +Copyright protection mechanisms in generative models, e.g., $NAF$ , are designed to yield a low upper-bound on the probability of generating infringing samples. Such protection is confined to single or infrequent generations. However, in scenarios where malicious attackers engage in multiple targeted generation requests, the probability of a successful attack can be amplified. + +We propose a universal online amplification attack framework, detailed in Algorithm 1, for a virtually assured success against probabilistic copyright protection. In this framework, the attacker is modeled as an adversarial prompt generator $\mathcal{A}(x_t|\langle (x_s,y_s)^{t - 1}_s\rangle)$ , which iteratively generates the prompt $x_{t}$ for the current step $t$ based on the interaction history with the generative model. For selecting the optimal sample, we employ a scoring function $\mathcal{S}:\mathcal{V}\to \mathbb{R}$ to evaluate all the samples and return the highest-scoring sample as the final result. An attack is deemed successful if the score of the returned result is higher than the target score $S_{tar}$ . For example, an essential choice of the scoring function could be the indicator function $\mathbb{I}(y\in \mathcal{V}_C)$ , with the target score $S_{tar} = 0$ , which ensures that any infringing sample generated will be returned as the final result, making the attack as successful. In practice, to reduce the manual effort in identifying copyright infringement, a computable surrogate scoring function with a target score may be utilized to establish a standard for copyright infringement. + +We introduce the following theorem, which suggests the virtually assured success of the amplification attack with an assumption on the lower-bound of success probability for each single-shot attack attempt. + +Theorem 1. Following the notations in Algorithm 1, for any $\varepsilon \in (0,1)$ , the attack is successful with probability at least $1 - \varepsilon$ if $T > \log_{1 - \sigma} \varepsilon$ , where $\sigma > 0$ is a strictly positive lower-bound on the success probability shared by every single attack. + +The intuition behind the amplification attack and Theorem 1 is straightforward. If the attacker is granted sufficient sampling opportunities with a set of prompts that yield even a modest lower-bound on the single success probability, as the number of sampling attempts increases, the probability + +Algorithm 1 Amplification Attack on Probabilistic Copyright Protection + +Require: Generative model $\tilde{p}$ with probabilistic copyright protection, target copyrighted data $\mathcal{C}$ , adversarial prompt generator $\mathcal{A}$ , maximum number of steps $T$ , score function $\mathcal{S}$ . + +1: for $t = 1, \dots, T$ do +2: Sample prompt $x_{t} \sim \mathcal{A}(\cdot |\langle (x_{s},y_{s})\rangle_{s = 1}^{t - 1})$ +3: Feed $x_{t}$ to $\tilde{p}$ and receive $y_{t} \sim \tilde{p}(\cdot | x_{t})$ +4: end for +5: return $y^{*} = \arg \max_{y\in \{y_{1:T}\}}S(y)$ + +of generating infringing samples accumulates at an exponentially fast rate, eventually making an attack almost guaranteed to succeed. For instance, suppose our attack possesses a relatively low single-shot success probability, say $1\%$ . Provided that we are permitted to repeat the attack approximately $\log_{0.99} 0.01 \approx 459$ times, the probability of at least one successful instance is then amplified to $99\%$ . + +# 4.3. Online Prompt Selection + +In this section, we consider a specific scenario within the framework described in Sec. 4.2, where the conditional prompt generator (the attacker) is restricted to select a prompt among $K$ candidate prompts $\{x^{1},\dots ,x^{K}\}$ , based on previous choices and the scores of received samples. The generation of candidate prompts can either be independent of or tailored specifically for certain models and copyright protection mechanisms, as we will discuss in Sec. 4.4, designed for diffusion model and NAF copyright protection. + +Let $a_{t} \in \{1, \dots, K\}$ denote the decision at step $t$ and $\pi$ denote the policy of prompt selection, the prompt generation procedure can be formally described as + +$$ +x _ {t} = x ^ {a _ {t}}, a _ {t} \sim \pi (\cdot | \langle (x _ {s}, r _ {s} = \mathcal {S} (y _ {s})) \rangle_ {s = 1} ^ {t - 1}). \tag {6} +$$ + +In this context, online prompt generation is reformulated as a variant of the multi-armed bandit problem, where each candidate prompt is conceptualized as an arm, and the score of the corresponding sample is the reward for pulling the arm. Unlike the classic multi-armed bandit framework, our goal is to maximize the probability of obtaining a reward exceeding a specified threshold $S_{tar}$ at least once within $T$ trials. This involves the trade-off between exploration and exploitation. Herein, we present two variants of the $\varepsilon$ -greedy algorithm. Initially, akin to the conventional algorithm, we conduct $m \leq \lfloor T / K \rfloor$ trials for each arm. For the rest $T - mK$ steps, a random action will be taken with probability $\varepsilon > 0$ for exploration. Otherwise, the best action $\hat{a}_t^* = \arg \max_a \hat{Q}_t(a)$ according to the evaluation $\hat{Q}_t(a)$ will be taken for exploitation. The choice of $\hat{Q}_t$ diverges into two distinct variants. + +$\varepsilon$ greedy-max Inspired by the fact that only the max reward matters, the maximum reward received so far is employed for evaluation, i.e., $\hat{Q}_t(a) = \max \{r_s : a_s = a\}_{s=1}^t$ . $\varepsilon$ greedy-cdf Essentially, we aim to maximize the probability that the reward for the next step exceeds a threshold $S_{tar}$ . If we assume that the reward distribution for each arm adheres to a normal distribution, we have $\hat{Q}_t(a) = 1 - \Phi\left(\frac{S_{tar} - \hat{\mu}_t(a)}{\hat{\sigma}_t(a)}\right)$ , where $\hat{\mu}_t(a) = \frac{1}{N_t(a)} \sum_{s=1}^t r_s \mathbb{I}(a_s = a)$ , $\hat{\sigma}_t^2(a) = \frac{1}{N_t(a) - 1} \sum_{s=1}^t (r_s \mathbb{I}(a_s = a) - \hat{\mu}_t(a))^2$ , and $\Phi$ denotes the cumulative distribution function (cdf) of standard normal distribution. + +Our empirical findings suggest that a well-calibrated balance between exploration and exploitation can enhance the stability of the attack. + +# 4.4. Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection + +In this section, we narrow the scope of discussion to adversarial prompt discovery against CP- $k$ , a NAF-based copyright protection algorithm. CP- $k$ modifies the probability density function of an unprotected generative model $p$ as + +$$ +\tilde {p} (y | x) \propto p (y | x) \mathbb {I} (\rho (y | x) \leq k _ {x}) \tag {7} +$$ + +where $\rho$ is defined in Eq. (3) and $k_{x}$ is a prompt-dependent threshold determined by the generative system, which is typically inaccessible to the user. CP- $k$ provides a probabilistic upper-bound for copyright infringement in Eq. (4). + +Building upon our initial assumptions, we further assume that the attacker is granted full access, i.e., knowing the structures and parameters of these models, to the unprotected generative model $p$ and safe models $q \in S$ . While this white-box setting renders our attack narrowed to open-source models and those at risk of backend leakage, the potential threat of our attack remains substantial, as it does not require intervention in the generation process. This means that if a malicious attacker gains access to a leaked model and discovers adversarial prompts, any other users, lacking access to the underlying models, can readily replicate these adversarial prompts to reproduce the attack. + +Recall that our objective is to find adversarial prompts capable of inducing models protected by CP- $k$ to generate infringing content with a strictly positive probability, i.e., + +$$ +\max _ {x \in \mathcal {X}} P _ {y \sim \tilde {p} (\cdot | x)} (y \in \mathcal {Y} _ {C}) \tag {8} +$$ + +We deduce a lower-bound for the infringement probability with optimal adversarial prompt in Theorem 2, subsequently guiding the development of a practical algorithm. + +Definition 1 (Local Continuity). Given a distance measure $\mathcal{D}$ defined in $\mathcal{V}$ , a model $p$ is called $(\epsilon, \alpha)$ -local-continuous around $y_0 \in \mathcal{V}$ if for any prompt $x \in \mathcal{X}$ , there exists $\epsilon, \alpha > 0$ such that for any $y \in \mathcal{B}_{\mathcal{D}}(y_0, \epsilon) := \{y \in \mathcal{Y} : \mathcal{D}(y_0, y) < \epsilon\}$ , $|p(y_0|x) - p(y|x)| < \alpha \mathcal{D}(y_0, y)$ . + +Algorithm 2 Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection +Require: Denoising network $\epsilon_{\theta_p},\epsilon_{\theta_q}$ , text encoder $\pmb{\tau}$ , target $y_{C}$ , denoising steps $T$ , optimization step $S$ , loss clip bound $\varphi$ , loss weight $\lambda$ , learning rate $\gamma$ +1: $\mathbf{P} = \{\mathbf{e}_1,\dots ,\mathbf{e}_n\} \sim E^{|\mathcal{V}|\times d}$ +2: for 1,.., S do +3: $x = \mathrm{Proj}(\mathbf{P})$ +4: Diffusion reconstruction task +5: $t\sim \mathrm{Uniform}(\{1,\ldots ,T\})$ +6: $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +7: $y_{t} = \sqrt{\bar{\alpha}_{t}} y_{C} + \sqrt{1 - \bar{\alpha}_{t}}\epsilon$ +8: $\mathcal{L}_p = \| \epsilon -\epsilon_{\theta_p}(y_t,t,\tau (x))\| _2^2$ +9: $\mathcal{L}_q = \| \epsilon -\epsilon_{\theta_q}(y_t,t,\tau (x))\| _2^2$ for all $q\in S$ +10: Calculate the gradient w.r.t projected embedding +11: $g = \nabla_{x}(\lambda \cdot \max (\mathcal{L}_{p},\varphi) + (1 - \lambda)\cdot \max_{q\in S}(\mathcal{L}_{q}))$ +12: Apply the gradient on continuous embedding +13: $\mathbf{P} = \mathbf{P} - \gamma g$ +14: end for +15: return Proj(P) + +Theorem 2. Assume there is a distance measure $\mathcal{D}$ defined in $\mathcal{V}$ such that (i) $p$ is $(\epsilon_p, \alpha)$ -local-continuous around $y_C$ , (ii) every $q \in S$ is local-continuous around $y_C$ , and (iii) there exists $\epsilon_c > 0$ such that $\mathcal{B}_{\mathcal{D}}(y_C, \epsilon_c) \subseteq \mathcal{V}_C$ . The objective defined in Eq. (8) has the following lower-bound for any $\eta, \delta > 0$ , + +$$ +\max _ {x \in \mathcal {X}} P _ {y \sim \tilde {p} (\cdot | x)} (y \in \mathcal {Y} _ {C}) \geq \max _ {x \in \tilde {\chi} _ {\eta , \delta}} \eta C _ {1} - \alpha C _ {2} \tag {9} +$$ + +where $\tilde{\mathcal{X}}_{\eta ,\delta} = \{x\in \mathcal{X}:p(y_C|x)\geq \eta ,\rho (y_C|x) < k_x - \delta \}$ and $C_1,C_2$ are constants independent on $x$ given as + +$$ +C _ {1} = \int_ {y \in \mathcal {B} _ {\mathcal {D}} (y _ {C}, \epsilon)} d y, C _ {2} = \int_ {y \in \mathcal {B} _ {\mathcal {D}} (y _ {C}, \epsilon)} \mathcal {D} (y _ {C}, y) d y, +$$ + +where $\epsilon = \min (\epsilon_p,\epsilon_c,\epsilon_\rho)$ with $\epsilon_{\rho}\coloneqq \inf_{x\in \tilde{\mathcal{X}}_{\eta ,\delta}}\sup \{\epsilon :$ $\rho (y|x) < k_x,\forall y\in \mathcal{B}_D(y_C,\epsilon)\} .$ + +To ensure that the lower-bound in the Theorem 2 is nontrivial, we need to set $\eta > \alpha C_2 / C_1$ and search along prompts satisfying (a) $\rho(y_C|x) < k_x$ , (b) $p(y_C|x) \geq \eta$ . Unfortunately, the mechanism to determine the threshold $k_x$ is assumed inaccessible to the attacker. To obtain a feasible optimization objective, we instead minimize $\rho(y_C|x)$ subject to $p(y_C|x) \geq \eta$ , as an alternative. The optimization objective then becomes + +$$ +\min _ {x} \rho \left(y _ {C} | x\right) \quad \text {s . t .} p \left(y _ {C} | x\right) \geq \eta . \tag {10} +$$ + +We follow the reconstruction task of the original text-to-image diffusion models for direct optimization. In aligning + +with the attack objective of Eq. (10), we develop the optimization objective from two aspects: + +$$ +\mathcal {L} _ {p} = \mathbb {E} _ {t, \boldsymbol {\epsilon}, x} [ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta_ {p}} (y _ {t}, t, \boldsymbol {\tau} (x)) \| _ {2} ^ {2} ] \tag {11} +$$ + +$$ +\mathcal {L} _ {q} = \mathbb {E} _ {t, \epsilon , x} [ \| \epsilon - \epsilon_ {\theta_ {q}} (y _ {t}, t, \tau (x)) \| _ {2} ^ {2} ] \tag {12} +$$ + +Where $y_{t}$ is the noisy version of target copyrighted data $y_{C}$ at denoising step $t$ ; $\epsilon_{\theta_p}, \epsilon_{\theta_q}$ is the denoising network of generative model $p$ and safe model $q \in S$ respectively. + +While $\mathcal{L}_p$ is designed to increase the possibility of model $p$ in generating the desired infringed content $y_{C}$ , the minimization of $\mathcal{L}_q$ corresponds to the maximization of $q(y_C|x)$ consequently leading to the reduction of $\rho (y_C|x)$ . Hence, the overall optimization objective can be formulated as a weighted sum of $\mathcal{L}_p$ and $\mathcal{L}_q$ : + +$$ +\mathcal {L} = \lambda \cdot \max \left(\mathcal {L} _ {p}, \varphi\right) + (1 - \lambda) \cdot \max _ {q \in \mathcal {S}} \left(\mathcal {L} _ {q}\right) \tag {13} +$$ + +Where $\varphi$ is a loss clip to mitigate the conflict between $\mathcal{L}_p$ and $\mathcal{L}_q$ in minimizing $\rho$ . + +The complete Anti-NAF prompt optimization process is detailed in Algorithm 2. We conduct optimization in the continuous embedding space with a sequence of learnable embeddings $\mathbf{P} = \{\mathbf{e}_1, \ldots, \mathbf{e}_n | \mathbf{e}_i \in \mathbb{R}^d\}$ , where $n$ is the sequence length and $d$ is the embedding dimension. Given that each word token $w$ in a vocabulary $\mathcal{V}$ can be represented as a corresponding embedding $\mathrm{EMB}(w)$ using an embedding matrix $E^{|\mathcal{V}| \times d}$ , each embedding $\mathbf{e}$ in the sequence can be mapped to its nearest token embedding in $\mathcal{V}$ under some similarity metric, e.g., cosine similarity. The projection function can be defined as $\mathrm{Proj}(\mathbf{e}) = \arg \min_{w \in \mathcal{V}} \cos(\mathrm{EMB}(w), \mathbf{e})$ . The resultant text prompt is a sequence of these projected tokens: $x = \mathrm{Proj}(\mathbf{P}) = \{\mathrm{Proj}(\mathbf{e}_1), \ldots, \mathrm{Proj}(\mathbf{e}_n)\}$ . In addition, following the approach in [53], the continuous embeddings $\mathbf{P}$ are projected into discrete tokens $x$ for each forward pass in every optimization step. The gradient of the projected prompt $x$ is then applied to update the continuous embeddings $\mathbf{P}$ . + +# 5. Experiments + +# 5.1. Experimental Settings + +# 5.1.1 Evaluated Datasets and Models + +Since training large text-to-image diffusion models from scratch is impractical, we fine-tune the pre-trained StableDiffusion-v1-4 model provided by Huggingface with two datasets: POKEMON [30] and LAION-mi [10]. Note that both datasets do not overlap with the pre-training dataset of Stable Diffusion, ensuring that the safe models fine-tuned on them are inaccessible to copyrighted data. Following [52], each dataset is split into two disjoint shards + +$\mathcal{D}_1$ and $\mathcal{D}_2$ to train generative model $q_{1}$ and $q_{2}$ respectively. For copyrighted data $y_{C}$ , $q_{2}$ is served as safe model safe $C$ if $y_{C}\in \mathcal{D}_{1}$ but $y_{C}\notin \mathcal{D}_{2}$ . For better experimental illustrations, each copyrighted data is repeated to make up $1\%$ of the dataset shard. + +POKEMON [30]. This dataset consists of 833 image-caption pairs, where captions are obtained from the BLIP model [22]. We evaluate our attack on 5 copyrighted images added to one of the shards separately and fine-tune models $q_{1}$ and $q_{2}$ on each shard for 5000 steps. + +LAION-mi [10]. The dataset is originally constructed for membership inference attacks on diffusion models. We only use the nonmembers part of this dataset, which holds a similar data distribution but is disjoint with the pre-training dataset of Stable Diffusion. We evaluate our attack on 5 selected copyrighted images, which are separately combined with one of non-copyrighted data shards. Each dataset shard is of size 5000. We fine-tune model $q_{1}$ and $q_{2}$ for 25,000 steps. Details for fine-tuning can be found in the appendix. + +# 5.1.2 Implementation details + +For the sampling setting of text-to-image diffusion models, images of size $512 \times 512$ are generated using a classifier-free guidance scale of 7.5 and 50 sampling steps with the default scheduler. The amplification step is set to 100 and 500 for POKEMON and LAION-mi respectively. For the prompt optimization process, the learning rate is set to 0.01 using Adagrad [11] optimizer with 25,000 optimization steps, the gradient accumulation step is set to 5, the loss clip $\varphi$ for model $p$ is 0.01, and the loss weight $\lambda$ is 0.95. The length of optimized prompts is set to 8 tokens. All experiments are conducted on A100 GPUs. + +# 5.2. Evaluation Settings + +We evaluate with a model $p$ behaves with an equal chance of sampling from either model $q_{1}$ or $q_{2}$ . We consider such model $p$ as a strict protection model that even with the original caption as prompt, there is a $50\%$ chance of generating from the safe model. In contrast, a model $p$ trained on the entire dataset could frequently produce highly similar generations to the target copyrighted image, making it extremely inefficient for the selection of a valid threshold $k$ . Next, we will separately discuss the data, the metrics, and the threat prompts for evaluation. + +Evaluation Data. We use samples generated by $p$ for evaluation. To determine whether they infringe the copyright of the target image, we rely on the similarity between samples and the target image as there is no widely acknowledged computable standard for infringement to our knowledge. We employ the current SOTA similarity metric SSCD [31] for copy detection. As SSCD scores of infringed samples for different target images differ significantly by obser + +![](images/95b53acae85c213b8c4795de3ee7bd052202387f749c08848dee09eb70dc65cb.jpg) + +![](images/b5ef27f72007279251d2fbdfdbdbc79de60cc263bdaeb1dee2e59dc9766468e3.jpg) +a black and white zebrastanding next to a star + +![](images/1eb20425d9017640ab071fa2756f7c90852d367381700449458a1aa6b144dea8.jpg) +zebra weet unex mee + +![](images/1451d9ee728b3151c69393fd0353a405c1bea4c177058ca4038e21d7111971b4.jpg) + +![](images/ab818ab2deb108b12b205fdd2705ac17705ee3d706f79e5f00080cc3b9ce7e22.jpg) +a picture of a white horsewith orange and yellow slings + +![](images/c28ca20a4883a1fdbc3e654731085d20c83dc23d151f2aec655e47369092c03d.jpg) + +![](images/4738cbedbf6175654f511a0a3d75b518c6b02a54500f3438a1d31228f813bbb4.jpg) + +![](images/68d118a425ea0f8ba18455c62ab779d2c9843cf3e3d7e5d68fac65589c0038e4.jpg) +a cartoon rabbit with asocce around its neck + +![](images/62eec3ba1ce2c6fcab3b98dd02c157513021b7365b6ddfea6007c861a948c35b.jpg) +rabbit, lamewridayday catunmibrated the room + +![](images/7fc6e76e20515ddc3275a1cde8f58f1ed400bf6904dda158406af806e2215961.jpg) + +![](images/3cb4dbaa395da339682983e353f04fdef965ec257ea73baf3ba19e0dce2d0328.jpg) +Giraffes Canvas Art Print + +![](images/6d3d505ba0e79062a8f745735c89fdcd32a36dc48c978b7a758e633cb9ccbab5.jpg) +giraffe guintspitnoche +tallent big etary +Figure 3. Visualization of generated images on different copyright targets. The examples in the first and second rows are selected from POKEMON and LAION-mi respectively. The prompts used to generate output are given below each group of images. Remarkably, the copyright-infringed content generated with Anti-NAF amplification reveals the vulnerability of probabilistic copyright protection CP- $k$ . + +![](images/1643aec6851c347c356b1280fd57891a61eb3ebff289cc4df0af831e31d43289.jpg) + +![](images/b9e5962edd28115ac91b9514d0e16c1695e27b78a744294ded66906e036fa139.jpg) +Product photo wall tattoo children's rostradberries with the inksticker set + +![](images/9004b5819071094d27dbf3d52bb4ee071e43f32f3557f60df2ed2499c9694c82.jpg) +horse millennials symbolism +kano jacoblissseasoned + +![](images/a3f7f659a6df2a56c22629c0845bac9a343e2d45d6db9524d597e7a4b75e999f.jpg) + +![](images/2fa8a38bb0628737b71ffb02cfe8ac30a11e3b36a401a42d34ec388733954484.jpg) +Penguins pictures of the New +Year, March with their own hands + +![](images/a2896a3a56ee6d52cd313290683163f9aaa112c52044fc0e8c6792352039a4da.jpg) +penguins knobs swimming +orbcsport orbfunday puppies have + +
MethodsPOKEMONLAION-mi
CIRFAR@5%AR↑FAR@15%AR↑CIRFAR@10%AR↑FAR@30%AR↑FAR@50%AR↑
Caption (w/o Amp.)47.00%0.40%3.28%49.64%0.00%0.00%0.04%
CLIP-Int. (w/o Amp.)26.92%0.84%2.20%48.12%0.84%1.52%2.96%
PEZ (w/o Amp.)7.80%1.32%2.76%15.80%0.08%0.24%0.40%
Anti-NAF (w/o Amp.)12.88%8.52%10.00%33.84%2.64%4.16%7.00%
Caption (w/ Amp.)100.00%13.64%40.16%100.00%0.00%0.00%14.64%
CLIP-Int. (w/ Amp.)99.72%22.84%47.32%100.00%34.84%37.12%75.60%
PEZ (w/ Amp.)66.92%22.80%37.48%98.16%38.04%50.68%59.64%
Anti-NAF (w/ Amp.)99.84%77.36%91.36%100.00%56.12%73.32%95.92%
+ +Table 1. Quantitative results. The performance with the amplification attack ("w/ Amp.") is significantly superior to scenarios without amplification ("w/o Amp.") Additionally, our proposed Anti-NAF demonstrates notably promising outcomes for providing a substantial probability of copyright infringement when probabilistic protection is applied. (CLIP-Int. is the abbreviation for CLIP-Interrogator). + +vation (shown in Fig. 5), we do not indicate a fixed score threshold for all target images. Instead, we use relative thresholds determined as percentiles of the similarity scores among samples generated by $p$ with the original caption of the target image as the prompt, e.g. SSCD-50%. Recognizing the variable criteria of infringement, we report results with other choices of thresholds in Sec. 11.3. + +Evaluation Metrics. The CP- $k$ method achieves copyright protection by selectively accepting generated samples using the threshold $k_{x}$ , which can be indicated by the Acceptance Rate (AR). For a given AR, a good protection system is expected to rarely accept infringing content, i.e., have a low False Accept Rate (FAR), as defined in Eq. (8). It is worth noting that the choice of $k_{x}$ dictates the trade-off between model safety and efficiency. Furthermore, to our knowledge, there is no principled way to determine $k_{x}$ . As a result, we evaluate the success of attack by reporting the FAR at different AR, e.g., FAR@5%AR. Additionally, the copyright infringement rate (CIR) is also presented for scenarios without copyright protection, i.e., $\mathrm{AR} = 100\%$ . + +Threat Prompts. Given that generating images similar to the target image is a necessary condition for a successful attack, we focus on the following threat prompts: (i) Target caption: the original caption of the target image; (ii) PEZ [53]: a gradient-based discrete optimization approach to discover prompts semantically similar to the target im + +age; (iii) $CLIP$ -Interrogator2: a prompt consists of the BLIP caption of the target image and top- $k$ keywords greedily sampled from a keywords collection; (iv) Anti-NAF: an adversarial prompt against NAF copyright protection as described in Sec. 4.4. For each threat prompt, we sample 6400 images from model $p$ for evaluation. For the online prompt selection setting, we take the aforementioned three prompts (exclude the target caption) as a prompt candidate set and consider $\varepsilon$ -greedy-max-/cdf Bandit Amplification as bandit strategies as described in Sec. 4.3. + +# 5.3. Results + +In Fig. 2, we show example outputs under four scenarios. It is evident that with amplification, the CP- $k$ protected generative model does indeed output copyright-infringed content with high probability, especially when using the Anti-NAF prompt. These examples suggest that probabilistic safeguards against copyright infringement are vulnerable to the amplification attack. Additional visualizations of outputs on various copyrighted images are demonstrated in Fig. 3. In Tab. 1, we report CIRs and FARs at various ARs on two datasets. Notably, there is a significant growth in both metrics when amplification attack is employed, highlighting its effectiveness in amplifying the probability of infringing generations. The superior performance of Anti-NAF under + +![](images/5fb7f79a69147f808b80b5828cc17ae37d7f803124c0c812bd1a489313060ef8.jpg) + +![](images/a4e9ae3771e25aa7a7b15c35e579417aed5bc7338d969fefb74fc22f4acbb04e.jpg) + +![](images/f47bfac79e07c02ea7a3fef834b477560fbe3ea215da2b9a198281e1b757fa85.jpg) +(c) FAR-AR of Anti-NAF with different Amp. steps on POKEMON. +Figure 4. The overall FAR-AR curves. The results show that amplification is strongly effective in amplifying the possibility of infringed output, even with small amplification steps. + +![](images/645819ffcfa46b03dabb3c216801f96a8dece9ad2c35481fe494a92243d65264.jpg) +(d) FAR-AR of Anti-NAF with different Amp. steps on LAION-mi. + +scores its efficacy in rendering infringing generations with a substantial probability. The overall FAR-AR curves are illustrated in Fig. 4. We can observe that bandit variants of amplification lead to a smaller variance across different target copyrighted images, especially at lower acceptance rates, indicating that our bandit strategies achieve a more steady attack. In Figs. 4c and 4d, we plot the overall FAR-AR curves over different amplification steps. There is a clear trend of rapidly improved performance with increased amplification steps, due to the cumulative probability of generating infringing samples. This finding underlines the potential risk in practical applications of probabilistic copyright protections, given the high frequency of daily interactions with text-to-image generative models. We provide additional results and human evaluations in Sec. 11. + +# 5.4. Ablation Study + +In Tab. 2, we study how different components of our optimization objective in Eq. (13), affect the performance. We discover that an exclusive focus on minimizing $\mathcal{L}_p$ (" $\mathcal{L}_p$ only") leads to a notable drop in performance. This indicates that prompts designed merely to reconstruct the target image are insufficient for a successful attack, as the infringing output of such a prompt can be easily identified and + +
MethodsCIRFAR@5%AR↑FAR@15%AR↑
Anti-NAF12.88%8.52%10.00%
\( \mathcal{L}_p \) only10.60%1.84%2.04%
w/o \( \varphi \)12.12%1.36%1.48%
w/o \( \mathcal{L}_q \)13.72%2.00%2.76%
+ +Table 2. Ablation study for Anti-NAF algorithm on POKEMON. + +blocked by the copyright protection system. Furthermore, a direct combination of $\mathcal{L}_q$ with $\mathcal{L}_p$ ("w/o $\varphi$ ) results in additional performance degradation because of the conflicting objectives of $\mathcal{L}_p$ and $\mathcal{L}_q$ in minimizing $\rho(y_C|x)$ . On the other hand, we can observe a slight performance improvement when implementing a loss clip $\varphi$ on $\mathcal{L}_p$ ("w/o $\mathcal{L}_q$ ) to constrain the learning of $\mathcal{L}_p$ . Overall, these results validate the effectiveness of our well-balanced optimization objective. Additional ablation studies are provided in Sec. 12. + +# 6. Conclusion + +In this paper, we shed light on the vulnerability of probabilistic copyright protection methods for text-to-image generative models, especially in real-world scenarios involving persistent and targeted interactions. Our proposed Virtually Assured Amplification Attack (VA3) framework presents the feasibility of inducing the protected model to generate copyright-infringed content with an amplified probability. Despite our focus on a narrow scenario of online prompt selection within this framework, the experimental results highlight its effectiveness in challenging even the most advanced existing probabilistic copyright protection methods. However, a broader scope of potential strategies remains unexplored within the VA3 framework, such as online prompt optimization, which may provide more powerful attacks against copyright protection. Furthermore, our Anti-NAF algorithm relies on access to generative models for adversarial prompt optimization, an assumption that may not be satisfied in completely black-box scenarios. We leave these more complicated and general attack methods for future investigation. In conclusion, our findings emphasize the significant risk of copyright infringement when applying probabilistic copyright protection methods in practice. Therefore, we hope that this work can inspire the development of more robust copyright protection approaches. + +# Acknowledgement + +The authors are grateful to Nikhil Vyas for providing the source code of CP- $k$ algorithm. This research is partially supported by the National Research Foundation Singapore under the AI Singapore Programme (AISG Award No: AISG2-TC-2023-010-SGIL) and the Singapore Ministry of Education Academic Research Fund Tier 1 (Award No: T1 251RES2207). + +# References + +[1] Hossein Aboutalebi, Daniel Mao, Carol Xu, and Alexander Wong. Deepfakeart challenge: A benchmark dataset for generative ai art forgery and data poisoning detection. arXiv preprint arXiv:2306.01272, 2023. 2 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1 +[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2 +[4] Stephen Casper, Zifan Guo, Shreya Mogulothu, Zachary Marinov, Chinmay Deshpande, Rui-Jie Yew, Zheng Dai, and Dylan Hadfield-Menell. Measuring the success of diffusion models at imitating human artists. arXiv preprint arXiv:2307.04028, 2023. 2 +[5] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3 +[6] David Chess. Some light infringement? https://ceoln.wordpress.com/2022/12/16/some-light-infringement/, 2022.2 +[7] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3 +[8] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 3 +[9] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3 +[10] Jan Dubiński, Antoni Kowalczuk, Stanisław Pawlak, Przemysław Rokita, Tomasz Trzeciński, and Pawel Morawiecki. Towards more realistic membership inference attacks on large diffusion models. arXiv preprint arXiv:2306.12983, 2023. 6 +[11] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 6 +[12] Niva Elkin-Koren, Uri Hacohen, Roi Livni, and Shay Moran. Can copyright be reduced to privacy? arXiv preprint arXiv:2305.14822, 2023. 1, 3 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1 + +[14] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023. 1, 2 +[15] Hongcheng Gao, Hao Zhang, Yinpeng Dong, and Zhi-jie Deng. Evaluating the robustness of text-to-image diffusion models against real-world attacks. arXiv preprint arXiv:2306.13103, 2023. 3 +[16] Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. arXiv preprint arXiv:2303.15715, 2023. 1, 2 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[18] Matthew Butterick Joseph Saveri. Stable diffusion litigation. https://stablediffusionlitigation.com/, 2023. 2 +[19] Kenji Kawaguchi, Zhun Deng, Xu Ji, and Jiaoyang Huang. How does information bottleneck help deep learning? In International Conference on Machine Learning (ICML), 2023. 2 +[20] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1 +[21] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516, 2023. 1, 2 +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 12888-12900. PMLR, 2022. 6 +[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2 +[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 20763-20786. PMLR, 2023. 2 +[25] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974, 2023. 3 +[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 21548-21566. PMLR, 2023. 2 +[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting + +using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1 +[28] Natalie Maus, Patrick Chao, Eric Wong, and Jacob Gardner. Adversarial prompting for black box foundation models. arXiv preprint arXiv:2302.04237, 2023. 3 +[29] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 3 +[30] Justin N. M. Pinkney. Pokemon blip captions. https://huggingface.co/datasets/lambdaabs/pokemon-blip-captions/, 2022.6 +[31] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14512-14522. IEEE, 2022. 6, 2 +[32] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. arXiv preprint arXiv:2305.13873, 2023. 3 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1 +[35] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 3 +[36] Arkadip Ray and Somaditya Roy. Recent trends in image watermarking techniques for copyright protection: a survey. International Journal of Multimedia Information Retrieval, 9(4):249-270, 2020. 3 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1 +[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[41] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 29894-29918. PMLR, 2023. 2 +[42] Sarah Scheffler, Eran Tromer, and Mayank Varia. Formalizing human ingenuity: A quantitative framework for copyright law's substantial similarity. In Proceedings of the 2022 Symposium on Computer Science and Law, pages 37-49, 2022. 2 +[43] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2 +[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[45] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2 +[46] Shawn Shan, Wenxin Ding, Josephine Passananti, Haitao Zheng, and Ben Y Zhao. Prompt-specific poisoning attacks on text-to-image generative models. arXiv preprint arXiv:2310.13828, 2023. 3 +[47] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1 +[48] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 6048-6058. IEEE, 2023. 2 +[49] Growthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 2 +[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1 +[51] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 3 + +[52] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 1, 2, 3, 6 +[53] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 6, 7 +[54] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 3 +[55] Yixin Wu, Ning Yu, Michael Backes, Yun Shen, and Yang Zhang. On the proactive generation of unsafe images from text-to-image models using benign prompts. arXiv preprint arXiv:2310.16613, 2023. 3 +[56] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023. 3 +[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. arXiv preprint arXiv:2305.04175, 2023. 3 +[58] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2 +[59] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3 +[60] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023 - Workshops, Vancouver, BC, Canada, June 17-24, 2023, pages 2385-2392. IEEE, 2023. 3 +[61] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 2 \ No newline at end of file diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/images.zip b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0cb3988d2cc0987d9cffa866144d85d59a0852b --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5372ad958ba8ea6e43553cb07713b72abb7ec5d884f024597c693b6b266f498 +size 648376 diff --git a/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/layout.json b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cff71c2caacb007ca616c9371288de49343616b7 --- /dev/null +++ b/2024/VA3_ Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models/layout.json @@ -0,0 +1,13635 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "text", + "content": "VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 161, + 416, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 161, + 416, + 204 + ], + "spans": [ + { + "bbox": [ + 176, + 161, + 416, + 204 + ], + "type": "text", + "content": "Xiang Li* Qianli Shen* Kenji Kawaguchi National University of Singapore {xiangli, qianli, kenji} @ comp.nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 230, + 544, + 449 + ], + "blocks": [ + { + "bbox": [ + 50, + 230, + 544, + 449 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 544, + 449 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 544, + 449 + ], + "type": "image", + "image_path": "0631ba4c845e782141a22326e2ad2e181b579a6c8d0cea4a869e2ce2abdc0a45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 456, + 546, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 546, + 491 + ], + "type": "text", + "content": "Figure 1. Given a copyrighted image " + }, + { + "bbox": [ + 46, + 456, + 546, + 491 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 46, + 456, + 546, + 491 + ], + "type": "text", + "content": " and a Text-to-Image (T2I) generative model with probabilistic copyright protection, our proposed virtually assured amplification attack (VA3) significantly amplifies the probability of producing infringing generations with persistent interactions of online adversarial prompt selection." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 500, + 192, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 500, + 192, + 511 + ], + "spans": [ + { + "bbox": [ + 143, + 500, + 192, + 511 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "type": "text", + "content": "The booming use of text-to-image generative models has raised concerns about their high risk of producing copyright-infringing content. While probabilistic copyright protection methods provide a probabilistic guarantee against such infringement, in this paper, we introduce Virtually Assured Amplification Attack (VA3), a novel online attack framework that exposes the vulnerabilities of these protection mechanisms. The proposed framework significantly amplifies the probability of generating infringing content on the sustained interactions with generative models and a non-trivial lower-bound on the success probability of each engagement. Our theoretical and experimental results demonstrate the effectiveness of our approach under various scenarios. These findings highlight the potential risk of implementing probabilistic copyright protection in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 502, + 544, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 544, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 544, + 525 + ], + "type": "text", + "content": "practical applications of text-to-image generative models. Code is available at https://github.com/South7X/VA3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 537, + 386, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 386, + 549 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 386, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "content": "In recent years, the advancement of large generative models [17, 47, 50] has revolutionized high-quality image synthesis [34, 37, 40], paving the way for commercial applications that enable the public to effortlessly craft their own artworks and designs [2, 13, 20, 27, 38, 39]. Nevertheless, these models exhibit notable memorization capabilities to produce generations highly similar to the training data [3]. This resemblance raises growing concerns about copyright infringement, especially when copyrighted data is used for training [12, 16, 49, 52]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "To address these concerns, there has been a surge in research focused on protecting copyrighted data from potential infringement by outputs of generative models [14, 21," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12363" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 170, + 191 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 170, + 191 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 170, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 170, + 191 + ], + "type": "image", + "image_path": "058a8d52955a1282d97e39bfa285658827fcbef3f22eafc817222f3ccbc682b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 193, + 143, + 202 + ], + "lines": [ + { + "bbox": [ + 74, + 193, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 74, + 193, + 143, + 202 + ], + "type": "text", + "content": "(a) Output w/o CP- " + }, + { + "bbox": [ + 74, + 193, + 143, + 202 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 69, + 295, + 191 + ], + "blocks": [ + { + "bbox": [ + 173, + 69, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 173, + 69, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 173, + 69, + 295, + 191 + ], + "type": "image", + "image_path": "aded96dea7d52576e4526e27ede13d16ef9395664899488443e80336beed8921.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 193, + 266, + 202 + ], + "lines": [ + { + "bbox": [ + 201, + 193, + 266, + 202 + ], + "spans": [ + { + "bbox": [ + 201, + 193, + 266, + 202 + ], + "type": "text", + "content": "(b) Output w/ CP-k." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 69, + 421, + 191 + ], + "blocks": [ + { + "bbox": [ + 298, + 69, + 421, + 191 + ], + "lines": [ + { + "bbox": [ + 298, + 69, + 421, + 191 + ], + "spans": [ + { + "bbox": [ + 298, + 69, + 421, + 191 + ], + "type": "image", + "image_path": "db87137dff15d6aafeff4e0ec47530872e9e84a098d99398dfbfac2c13495388.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 193, + 401, + 202 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 401, + 202 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 401, + 202 + ], + "type": "text", + "content": "(c) Amp. output w/ CP- " + }, + { + "bbox": [ + 317, + 193, + 401, + 202 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "text", + "content": "Figure 2. Example outputs given the copyright image in Fig. 1 as target (potential infringing images are marked with red boundaries). In (a), using a benign prompt, we observe a high incidence of infringing content from models without copyright protection (\"w/o CP-" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "text", + "content": ")\". In contrast, (b) shows that after applying the copyright protection mechanism (\"w/ CP-" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "text", + "content": ")\", all samples are safe as CP-" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 212, + 547, + 268 + ], + "type": "text", + "content": " rejects all infringing content. In (c), we find that amplification (Amp.) attack with a benign prompt results in limited success. Notably, by amplification attack with an adversarial prompt obtained from our proposed Anti-NAF algorithm, almost all output in (d) are copyright-infringed." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 545, + 191 + ], + "type": "image", + "image_path": "589a8d6610525929211e58c24d80cd8dfef503320baadfbf37b019d813da4f53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 193, + 544, + 202 + ], + "lines": [ + { + "bbox": [ + 425, + 193, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 425, + 193, + 544, + 202 + ], + "type": "text", + "content": "(d) Anti-NAF Amp. output w/ CP- " + }, + { + "bbox": [ + 425, + 193, + 544, + 202 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 289, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 289, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 45, + 289, + 289, + 468 + ], + "type": "text", + "content": "24, 43, 45, 52, 58]. Among these studies, a pivotal concept involves establishing a probabilistic upper-bound against the generation of infringing content by generative models. We refer to this suite of approaches as probabilistic copyright protection. Most notably, Vyas et al. [52] introduce a mathematical definition of copyright known as near-access freeness (NAF). Their method enforces generative diffusion models to exhibit akin behaviors as safe models, which has no access to the copyrighted image. By leveraging the improbability of safe models generating infringing content, the probability of generative models doing the same is thereby substantially reduced. The copyright protection algorithm of NAF, CP-" + }, + { + "bbox": [ + 45, + 289, + 289, + 468 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 289, + 289, + 468 + ], + "type": "text", + "content": ", can filter out infringing content generated by the models with high probability, even when the input prompts are adversarially designed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 472, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 615 + ], + "type": "text", + "content": "In this paper, we propose Virtually Assured Amplification Attack (VA3), a novel online attack framework, to show the vulnerability of probabilistic copyright protection. This framework induces text-to-image generative models with probabilistic copyright protection to generate infringing content. Our approach is grounded in the realization that in real-world scenarios, a malicious attacker intending to induce copyright infringement could engage in multiple interactions with the generative model via prompts. This persistent engagement poses a significant challenge to probabilistic protection methods, as it amplified the probability of producing infringing content of each generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "In our proposed framework, the attacker functions as a conditional prompt generator, creating adversarial prompts iteratively based on previous interactions with the generative model. Our primary theoretical result, Theorem 1, suggests that the amplification attack is guaranteed to succeed with high probability, given a sufficient number of interactions with the generative model and a strictly positive lower-bound on success probability of each single engagement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 384 + ], + "type": "text", + "content": "Regarding practical algorithms, our work encompasses two technical innovations. Firstly, we present effective strategies to manage the exploration-exploitation dilemma in online prompt selection, thereby enhancing the stability of the attack. Secondly, we propose Anti-NAF, a theoretically motivated adversarial prompt optimization algorithm tailored for NAF copyright protection, to generate prompts fulfilling the conditions of Theorem 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 385, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 546, + 456 + ], + "type": "text", + "content": "Our experimental results validate the efficacy of our proposed online attack approach under diverse scenarios. These findings underline the potential copyright infringement risk of applying probabilistic protection in practical applications of text-to-image generative models, for both providers and users." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 491, + 511, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 511, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 511, + 503 + ], + "type": "text", + "content": "2.1. Copyright Issues in Generative Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Text-to-image generative models, trained on large-scale datasets like LAION [44], have been equipped with enhanced memorization ability to generate outputs of high semantic similarity to their training data [3, 48]. Given the prevalence of copyrighted works in these datasets, the significant risk of copyright infringement for these generations has raised great concerns from the public [6, 18] and researchers [1, 4, 16, 42, 48, 49, 52]. Many efforts have been made to safeguard copyrighted materials from being infringed by generative diffusion models. Some researchers [23, 24, 41, 45] introduced data perturbation, where input data is modified to hinder the model to imitate copyrighted features. Another separate line of works [14, 21, 26, 43, 58] exploited concept removal that erases unsafe concepts from existing pre-trained diffusion models to mitigate the risk of undesirable generations. In an alternative approach, researchers studied watermark protection for copyrighted data" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12364" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "[9, 29, 36, 54, 56, 59] to encode ownership information into potentially infringed outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 288, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 288, + 216 + ], + "type": "text", + "content": "A notable contribution, Vyas et al. [52] first provided a mathematical probabilistic upper-bound against copyright-infringed generation. They asserted that the proposed near access-freeness (NAF) offers robust guarantees for copyright protection. However, Elkin-Koren et al. [12] argued the limitation of this method for reducing copyright to a matter of privacy from a legal perspective. In this paper, we build upon these discussions to present a significant challenge to these probabilistic copyright protection methods through the amplification attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 226, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 226, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 226, + 235 + ], + "type": "text", + "content": "2.2. Vulnerability of Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 288, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 480 + ], + "type": "text", + "content": "A rising number of works focus on the vulnerability of diffusion models to various types of attacks. Poisoning attack [5, 8, 46, 55, 57] studies the problem of manipulating training data to induce unsafe behaviors in diffusion models during the training phase. In the context of the real-world inference phase for existing text-to-image diffusion models, other works also investigate the robustness and safety against different prompt inputs. Researchers [15, 25, 28, 60] have shown that injecting a slight perturbation to the input prompts will mislead the unprotected model to generate semantically unrelated images. Furthermore, with carefully crafted prompts, unsafe images can easily evade detection-based safety filters in text-to-image diffusion models [32, 35]. Other studies [7, 51] red-teaming concept removal copyright protection methods by finding problematic prompts that can recover erased unsafe concepts to yield undesirable generations. Diverging from these existing attack approaches on heuristic protection methods, in this paper, our focus is on challenging the vulnerability of probabilistic copyright protection methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 132, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 132, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 132, + 502 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 510, + 218, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 218, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 218, + 522 + ], + "type": "text", + "content": "3.1. Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": "Take DDPM [17] for example, a typical diffusion process consists of a predefined forward process and a reverse process. Specifically, the forward process corrupts the original data " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " into a standard Gaussian noise " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " timesteps. In the reverse process, the denoising network " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " is learned to denoise the corrupted " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " by predicting the sampled noise " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": " that added to " + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": ". The objective of the diffusion model can be simplified into the following form:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 643, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 643, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 91, + 643, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right] \\tag {1}", + "image_path": "a7572ac44c88f1104f5496fdb67151125dd2d879e722c6752139e119a4439915.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Text-to-image diffusion models further use prompts to guide the sampling process for generating desired images. Take Stable Diffusion [37] for example, they incorporate a pre-trained CLIP [33] as text encoder " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " to encode an input" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "text " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\boldsymbol{\\tau}(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ". An image encoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " is employed to map an input image " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " into its latent representation " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ". The training objective is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 115, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 115, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 349, + 115, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\boldsymbol {\\epsilon}, \\mathbf {c}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| _ {2} ^ {2} \\right] \\tag {2}", + "image_path": "3ede9c3cd177f8ce86b579f3e3208bd52e80628efcf434280bf86e671b0eed08.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 136, + 531, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 136, + 531, + 148 + ], + "spans": [ + { + "bbox": [ + 305, + 136, + 531, + 148 + ], + "type": "text", + "content": "3.2. Near Access-Freeness Copyright Protection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": "As a pioneering work of probabilistic copyright protection method, Vyas et al. [52] formally defines Near Access-Freeness (NAF) to provide a probabilistic guarantee against the generation of infringing content. They provide practical algorithms (CP-" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": ") to protect an arbitrary generative model " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " from copyright infringement. Suppose we have a cover of safe models " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " satisfying that for any piece of copyrighted data " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": ", there exists some " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "q = \\text{safe}_{C} \\in \\mathcal{S}" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " trained without access to " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": ". The copyright protection at a pre-given threshold " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " can be achieved by sampling " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "y \\sim p(\\cdot | x)" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " with a prompt " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " and accepting " + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 154, + 545, + 286 + ], + "type": "text", + "content": " if:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 293, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 293, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 361, + 293, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\rho (y | x) := \\max _ {q \\in S} \\log \\frac {p (y | x)}{q (y | x)} \\leq k \\tag {3}", + "image_path": "b076f812b7224d941adc0154abb609e8320c1d59efeb83989bf41d46dabf1372.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\rho(y|x)" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " serves as a criterion to distinguish non-infringing content from infringing content, while " + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " determines the threshold for judgment. Let " + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " be the protected model with threshold " + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_C" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " be the set of infringing contents, CP-" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 325, + 545, + 397 + ], + "type": "text", + "content": " provides the following probabilistic upper-bound for copyright infringement:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 404, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 404, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 336, + 404, + 545, + 431 + ], + "type": "interline_equation", + "content": "p _ {k} (y \\in \\mathcal {V} _ {C} | x) \\leq \\frac {2 ^ {k}}{\\nu_ {k} (x)} \\cdot \\operatorname {s a f e} _ {C} (y \\in \\mathcal {V} _ {C} | x) \\tag {4}", + "image_path": "d2695511794e7b817c27e644c3dd879d62530a85a02210655e257d793419a0a6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\nu_{k}(x)" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": " denotes the acceptance rate of " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": ". For different prompts " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": ", the optimal choice of " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": " changes with varying distribution of " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\rho(y|x)" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": ". In the subsequent sections of our paper, we use " + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "inline_equation", + "content": "k_x" + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": " to emphasize the dependency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 496, + 361, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 361, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 361, + 508 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 515, + 429, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 429, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 429, + 527 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": "In this paper, we consider a text-to-image generative model " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " equipped with probabilistic copyright protection. The probability that " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " generates infringing content with prompt " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " is upper-bounded. Following the typical adversarial prompt attack setting, we consider a malicious attacker who seeks to manipulate " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " to produce content that violates the copyright of a specific piece of target copyrighted data " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": ", of which the infringing contents form " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_C \\subseteq \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Different from the standard paradigm, we consider an online attack scenario where the attacker is allowed to interact with " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " times within each attack trial. During the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "-th interaction, the attacker inputs a prompt " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and receives a generated sample " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{t}\\sim \\tilde{p} (\\cdot |x_{t})" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Besides furnishing a prompt, the attacker is prohibited from intervening directly in the generation process. This constraint" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12365" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "is consistent with real-world scenarios where users interact with generative models as black boxes through APIs. At the end of each attack trial, the attacker will select " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "y^{*} \\in \\{y_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " as the final output, and the attack is regarded successful if " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "y^{*} \\in \\mathcal{V}_{C}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": ". The formal objective of the attacker is to maximize the success rate of attack, i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 158, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 158, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 113, + 158, + 287, + 175 + ], + "type": "interline_equation", + "content": "\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} \\left(y ^ {*} \\in \\mathcal {Y} _ {C}\\right) \\tag {5}", + "image_path": "137d61819db8825e38be89d10cb0415977eae969d24abfd0d4fb9294132b2306.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 287, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 287, + 208 + ], + "type": "text", + "content": "4.2. Virtually Assured Amplification Attack on Probabilistic Copyright Protection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 297 + ], + "type": "text", + "content": "Copyright protection mechanisms in generative models, e.g., " + }, + { + "bbox": [ + 46, + 214, + 287, + 297 + ], + "type": "inline_equation", + "content": "NAF" + }, + { + "bbox": [ + 46, + 214, + 287, + 297 + ], + "type": "text", + "content": ", are designed to yield a low upper-bound on the probability of generating infringing samples. Such protection is confined to single or infrequent generations. However, in scenarios where malicious attackers engage in multiple targeted generation requests, the probability of a successful attack can be amplified." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": "We propose a universal online amplification attack framework, detailed in Algorithm 1, for a virtually assured success against probabilistic copyright protection. In this framework, the attacker is modeled as an adversarial prompt generator " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(x_t|\\langle (x_s,y_s)^{t - 1}_s\\rangle)" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": ", which iteratively generates the prompt " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": " for the current step " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": " based on the interaction history with the generative model. For selecting the optimal sample, we employ a scoring function " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{S}:\\mathcal{V}\\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": " to evaluate all the samples and return the highest-scoring sample as the final result. An attack is deemed successful if the score of the returned result is higher than the target score " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "S_{tar}" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": ". For example, an essential choice of the scoring function could be the indicator function " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(y\\in \\mathcal{V}_C)" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": ", with the target score " + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "inline_equation", + "content": "S_{tar} = 0" + }, + { + "bbox": [ + 46, + 299, + 287, + 526 + ], + "type": "text", + "content": ", which ensures that any infringing sample generated will be returned as the final result, making the attack as successful. In practice, to reduce the manual effort in identifying copyright infringement, a computable surrogate scoring function with a target score may be utilized to establish a standard for copyright infringement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 527, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 574 + ], + "type": "text", + "content": "We introduce the following theorem, which suggests the virtually assured success of the amplification attack with an assumption on the lower-bound of success probability for each single-shot attack attempt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "content": "Theorem 1. Following the notations in Algorithm 1, for any " + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "inline_equation", + "content": "\\varepsilon \\in (0,1)" + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "content": ", the attack is successful with probability at least " + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "inline_equation", + "content": "1 - \\varepsilon" + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "inline_equation", + "content": "T > \\log_{1 - \\sigma} \\varepsilon" + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "inline_equation", + "content": "\\sigma > 0" + }, + { + "bbox": [ + 46, + 583, + 287, + 644 + ], + "type": "text", + "content": " is a strictly positive lower-bound on the success probability shared by every single attack." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "The intuition behind the amplification attack and Theorem 1 is straightforward. If the attacker is granted sufficient sampling opportunities with a set of prompts that yield even a modest lower-bound on the single success probability, as the number of sampling attempts increases, the probability" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "Algorithm 1 Amplification Attack on Probabilistic Copyright Protection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": "Require: Generative model " + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": " with probabilistic copyright protection, target copyrighted data " + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": ", adversarial prompt generator " + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": ", maximum number of steps " + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": ", score function " + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 307, + 99, + 545, + 146 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 148, + 503, + 209 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 312, + 148, + 406, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 406, + 158 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 406, + 158 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 312, + 148, + 406, + 158 + ], + "type": "inline_equation", + "content": "t = 1, \\dots, T" + }, + { + "bbox": [ + 312, + 148, + 406, + 158 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 159, + 503, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 159, + 503, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 159, + 503, + 171 + ], + "type": "text", + "content": "2: Sample prompt " + }, + { + "bbox": [ + 312, + 159, + 503, + 171 + ], + "type": "inline_equation", + "content": "x_{t} \\sim \\mathcal{A}(\\cdot |\\langle (x_{s},y_{s})\\rangle_{s = 1}^{t - 1})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "text", + "content": "3: Feed " + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "text", + "content": " and receive " + }, + { + "bbox": [ + 312, + 171, + 488, + 183 + ], + "type": "inline_equation", + "content": "y_{t} \\sim \\tilde{p}(\\cdot | x_{t})" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 184, + 357, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 357, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 357, + 194 + ], + "type": "text", + "content": "4: end for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 195, + 467, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 467, + 209 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 467, + 209 + ], + "type": "text", + "content": "5: return " + }, + { + "bbox": [ + 312, + 195, + 467, + 209 + ], + "type": "inline_equation", + "content": "y^{*} = \\arg \\max_{y\\in \\{y_{1:T}\\}}S(y)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "text", + "content": "of generating infringing samples accumulates at an exponentially fast rate, eventually making an attack almost guaranteed to succeed. For instance, suppose our attack possesses a relatively low single-shot success probability, say " + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "text", + "content": ". Provided that we are permitted to repeat the attack approximately " + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\log_{0.99} 0.01 \\approx 459" + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "text", + "content": " times, the probability of at least one successful instance is then amplified to " + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 233, + 545, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 326, + 444, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 444, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 444, + 339 + ], + "type": "text", + "content": "4.3. Online Prompt Selection" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "text", + "content": "In this section, we consider a specific scenario within the framework described in Sec. 4.2, where the conditional prompt generator (the attacker) is restricted to select a prompt among " + }, + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "text", + "content": " candidate prompts " + }, + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\{x^{1},\\dots ,x^{K}\\}" + }, + { + "bbox": [ + 304, + 344, + 545, + 453 + ], + "type": "text", + "content": ", based on previous choices and the scores of received samples. The generation of candidate prompts can either be independent of or tailored specifically for certain models and copyright protection mechanisms, as we will discuss in Sec. 4.4, designed for diffusion model and NAF copyright protection." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "inline_equation", + "content": "a_{t} \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "text", + "content": " denote the decision at step " + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 453, + 545, + 488 + ], + "type": "text", + "content": " denote the policy of prompt selection, the prompt generation procedure can be formally described as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 335, + 498, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 498, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 335, + 498, + 545, + 513 + ], + "type": "interline_equation", + "content": "x _ {t} = x ^ {a _ {t}}, a _ {t} \\sim \\pi (\\cdot | \\langle (x _ {s}, r _ {s} = \\mathcal {S} (y _ {s})) \\rangle_ {s = 1} ^ {t - 1}). \\tag {6}", + "image_path": "a00b68f4c26c86d1f131a235d9e373ffb1d2aa96158b4b17a3633dc0e32d59a3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "In this context, online prompt generation is reformulated as a variant of the multi-armed bandit problem, where each candidate prompt is conceptualized as an arm, and the score of the corresponding sample is the reward for pulling the arm. Unlike the classic multi-armed bandit framework, our goal is to maximize the probability of obtaining a reward exceeding a specified threshold " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{tar}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " at least once within " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " trials. This involves the trade-off between exploration and exploitation. Herein, we present two variants of the " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "-greedy algorithm. Initially, akin to the conventional algorithm, we conduct " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "m \\leq \\lfloor T / K \\rfloor" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " trials for each arm. For the rest " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "T - mK" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " steps, a random action will be taken with probability " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon > 0" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " for exploration. Otherwise, the best action " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{a}_t^* = \\arg \\max_a \\hat{Q}_t(a)" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " according to the evaluation " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Q}_t(a)" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " will be taken for exploitation. The choice of " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Q}_t" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " diverges into two distinct variants." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12366" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": " greedy-max Inspired by the fact that only the max reward matters, the maximum reward received so far is employed for evaluation, i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\hat{Q}_t(a) = \\max \\{r_s : a_s = a\\}_{s=1}^t" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": " greedy-cdf Essentially, we aim to maximize the probability that the reward for the next step exceeds a threshold " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "S_{tar}" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": ". If we assume that the reward distribution for each arm adheres to a normal distribution, we have " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\hat{Q}_t(a) = 1 - \\Phi\\left(\\frac{S_{tar} - \\hat{\\mu}_t(a)}{\\hat{\\sigma}_t(a)}\\right)" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t(a) = \\frac{1}{N_t(a)} \\sum_{s=1}^t r_s \\mathbb{I}(a_s = a)" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}_t^2(a) = \\frac{1}{N_t(a) - 1} \\sum_{s=1}^t (r_s \\mathbb{I}(a_s = a) - \\hat{\\mu}_t(a))^2" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 72, + 289, + 209 + ], + "type": "text", + "content": " denotes the cumulative distribution function (cdf) of standard normal distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 210, + 288, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 288, + 246 + ], + "type": "text", + "content": "Our empirical findings suggest that a well-calibrated balance between exploration and exploitation can enhance the stability of the attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 251, + 288, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 288, + 277 + ], + "type": "text", + "content": "4.4. Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "text", + "content": "In this section, we narrow the scope of discussion to adversarial prompt discovery against CP-" + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "text", + "content": ", a NAF-based copyright protection algorithm. CP-" + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "text", + "content": " modifies the probability density function of an unprotected generative model " + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 282, + 287, + 331 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 337, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 337, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 100, + 337, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\tilde {p} (y | x) \\propto p (y | x) \\mathbb {I} (\\rho (y | x) \\leq k _ {x}) \\tag {7}", + "image_path": "77b311d96a95ad47f63d93a4bf1965468328773b3cd44c127d7317db7500b5d9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "text", + "content": " is defined in Eq. (3) and " + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "inline_equation", + "content": "k_{x}" + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "text", + "content": " is a prompt-dependent threshold determined by the generative system, which is typically inaccessible to the user. CP-" + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 356, + 287, + 403 + ], + "type": "text", + "content": " provides a probabilistic upper-bound for copyright infringement in Eq. (4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "text", + "content": "Building upon our initial assumptions, we further assume that the attacker is granted full access, i.e., knowing the structures and parameters of these models, to the unprotected generative model " + }, + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "text", + "content": " and safe models " + }, + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "inline_equation", + "content": "q \\in S" + }, + { + "bbox": [ + 46, + 404, + 287, + 547 + ], + "type": "text", + "content": ". While this white-box setting renders our attack narrowed to open-source models and those at risk of backend leakage, the potential threat of our attack remains substantial, as it does not require intervention in the generation process. This means that if a malicious attacker gains access to a leaked model and discovers adversarial prompts, any other users, lacking access to the underlying models, can readily replicate these adversarial prompts to reproduce the attack." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 548, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 584 + ], + "type": "text", + "content": "Recall that our objective is to find adversarial prompts capable of inducing models protected by CP-" + }, + { + "bbox": [ + 47, + 548, + 287, + 584 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 548, + 287, + 584 + ], + "type": "text", + "content": " to generate infringing content with a strictly positive probability, i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 590, + 287, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 590, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 287, + 607 + ], + "type": "interline_equation", + "content": "\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\tag {8}", + "image_path": "80c77a72bc8776f7776d281af8870f93385e3aaf41aea23bd95d56ba45c2549e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 612, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 287, + 650 + ], + "type": "text", + "content": "We deduce a lower-bound for the infringement probability with optimal adversarial prompt in Theorem 2, subsequently guiding the development of a practical algorithm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": "Definition 1 (Local Continuity). Given a distance measure " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": " defined in " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": ", a model " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": " is called " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\epsilon, \\alpha)" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": "-local-continuous around " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "y_0 \\in \\mathcal{V}" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": " if for any prompt " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": ", there exists " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon, \\alpha > 0" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": " such that for any " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{B}_{\\mathcal{D}}(y_0, \\epsilon) := \\{y \\in \\mathcal{Y} : \\mathcal{D}(y_0, y) < \\epsilon\\}" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "|p(y_0|x) - p(y|x)| < \\alpha \\mathcal{D}(y_0, y)" + }, + { + "bbox": [ + 47, + 654, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 307, + 99, + 545, + 318 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "Algorithm 2 Anti-NAF: Adversarial Prompt Optimization Against NAF Copyright Protection" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": "Require: Denoising network " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_p},\\epsilon_{\\theta_q}" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , text encoder " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , target " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , denoising steps " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , optimization step " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , loss clip bound " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , loss weight " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " , learning rate " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{e}_1,\\dots ,\\mathbf{e}_n\\} \\sim E^{|\\mathcal{V}|\\times d}" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n2: for 1,.., S do \n3: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "x = \\mathrm{Proj}(\\mathbf{P})" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n4: Diffusion reconstruction task \n5: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "y_{t} = \\sqrt{\\bar{\\alpha}_{t}} y_{C} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p = \\| \\epsilon -\\epsilon_{\\theta_p}(y_t,t,\\tau (x))\\| _2^2" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q = \\| \\epsilon -\\epsilon_{\\theta_q}(y_t,t,\\tau (x))\\| _2^2" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "q\\in S" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n10: Calculate the gradient w.r.t projected embedding \n11: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "g = \\nabla_{x}(\\lambda \\cdot \\max (\\mathcal{L}_{p},\\varphi) + (1 - \\lambda)\\cdot \\max_{q\\in S}(\\mathcal{L}_{q}))" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n12: Apply the gradient on continuous embedding \n13: " + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\mathbf{P} - \\gamma g" + }, + { + "bbox": [ + 307, + 99, + 545, + 318 + ], + "type": "text", + "content": " \n14: end for \n15: return Proj(P)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": "Theorem 2. Assume there is a distance measure " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": " defined in " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": " such that (i) " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "(\\epsilon_p, \\alpha)" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": "-local-continuous around " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "y_C" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": ", (ii) every " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "q \\in S" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": " is local-continuous around " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "y_C" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": ", and (iii) there exists " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon_c > 0" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{D}}(y_C, \\epsilon_c) \\subseteq \\mathcal{V}_C" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": ". The objective defined in Eq. (8) has the following lower-bound for any " + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\eta, \\delta > 0" + }, + { + "bbox": [ + 304, + 340, + 547, + 411 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 422, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 422, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 331, + 422, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\max _ {x \\in \\mathcal {X}} P _ {y \\sim \\tilde {p} (\\cdot | x)} (y \\in \\mathcal {Y} _ {C}) \\geq \\max _ {x \\in \\tilde {\\chi} _ {\\eta , \\delta}} \\eta C _ {1} - \\alpha C _ {2} \\tag {9}", + "image_path": "fdc7a3dc672ebb280ca4b53bf494d8217f7ad929e2ff99395da8d7ab819e59b8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{X}}_{\\eta ,\\delta} = \\{x\\in \\mathcal{X}:p(y_C|x)\\geq \\eta ,\\rho (y_C|x) < k_x - \\delta \\}" + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "inline_equation", + "content": "C_1,C_2" + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "text", + "content": " are constants independent on " + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 453, + 545, + 479 + ], + "type": "text", + "content": " given as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 538, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 538, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 538, + 515 + ], + "type": "interline_equation", + "content": "C _ {1} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} d y, C _ {2} = \\int_ {y \\in \\mathcal {B} _ {\\mathcal {D}} (y _ {C}, \\epsilon)} \\mathcal {D} (y _ {C}, y) d y,", + "image_path": "a4055bb9317a0d5c5538a1ae7074452000292b7ee8000648b5a45b7a823fffe5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\epsilon = \\min (\\epsilon_p,\\epsilon_c,\\epsilon_\\rho)" + }, + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\rho}\\coloneqq \\inf_{x\\in \\tilde{\\mathcal{X}}_{\\eta ,\\delta}}\\sup \\{\\epsilon :" + }, + { + "bbox": [ + 305, + 523, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\rho (y|x) < k_x,\\forall y\\in \\mathcal{B}_D(y_C,\\epsilon)\\} ." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": "To ensure that the lower-bound in the Theorem 2 is nontrivial, we need to set " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\eta > \\alpha C_2 / C_1" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": " and search along prompts satisfying (a) " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\rho(y_C|x) < k_x" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": ", (b) " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "p(y_C|x) \\geq \\eta" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": ". Unfortunately, the mechanism to determine the threshold " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "k_x" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": " is assumed inaccessible to the attacker. To obtain a feasible optimization objective, we instead minimize " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\rho(y_C|x)" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": " subject to " + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "inline_equation", + "content": "p(y_C|x) \\geq \\eta" + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": ", as an alternative. The optimization objective then becomes" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 667, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 667, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 358, + 667, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\min _ {x} \\rho \\left(y _ {C} | x\\right) \\quad \\text {s . t .} p \\left(y _ {C} | x\\right) \\geq \\eta . \\tag {10}", + "image_path": "e738ea8f119a68e347d5b5a28629d8ef7195cfaead43039159a731c8fbf66a13.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We follow the reconstruction task of the original text-to-image diffusion models for direct optimization. In aligning" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12367" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "with the attack objective of Eq. (10), we develop the optimization objective from two aspects:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 105, + 287, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 105, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 105, + 287, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\mathbb {E} _ {t, \\boldsymbol {\\epsilon}, x} [ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta_ {p}} (y _ {t}, t, \\boldsymbol {\\tau} (x)) \\| _ {2} ^ {2} ] \\tag {11}", + "image_path": "578881cab979363af0554dee541187447a9a0352ac7c78f1b132333dc80d1192.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 129, + 287, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 129, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 93, + 129, + 287, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {q} = \\mathbb {E} _ {t, \\epsilon , x} [ \\| \\epsilon - \\epsilon_ {\\theta_ {q}} (y _ {t}, t, \\tau (x)) \\| _ {2} ^ {2} ] \\tag {12}", + "image_path": "48207f5985815a3197b96fb590fc3016161480e9d7b0a86238064ba14ae70074.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": " is the noisy version of target copyrighted data " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": " at denoising step " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_p}, \\epsilon_{\\theta_q}" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": " is the denoising network of generative model " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": " and safe model " + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "inline_equation", + "content": "q \\in S" + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": " is designed to increase the possibility of model " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": " in generating the desired infringed content " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": ", the minimization of " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": " corresponds to the maximization of " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "q(y_C|x)" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": " consequently leading to the reduction of " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\rho (y_C|x)" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": ". Hence, the overall optimization objective can be formulated as a weighted sum of " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 46, + 185, + 287, + 258 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 268, + 287, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 268, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 80, + 268, + 287, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\cdot \\max \\left(\\mathcal {L} _ {p}, \\varphi\\right) + (1 - \\lambda) \\cdot \\max _ {q \\in \\mathcal {S}} \\left(\\mathcal {L} _ {q}\\right) \\tag {13}", + "image_path": "0fe3e3911ab760f0c25bf3e14bd675a61cfb77d97b4be7d9399e1e49da49d278.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": " is a loss clip to mitigate the conflict between " + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": " in minimizing " + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": "The complete Anti-NAF prompt optimization process is detailed in Algorithm 2. We conduct optimization in the continuous embedding space with a sequence of learnable embeddings " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{e}_1, \\ldots, \\mathbf{e}_n | \\mathbf{e}_i \\in \\mathbb{R}^d\\}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " is the embedding dimension. Given that each word token " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " in a vocabulary " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " can be represented as a corresponding embedding " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{EMB}(w)" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " using an embedding matrix " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "E^{|\\mathcal{V}| \\times d}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": ", each embedding " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " in the sequence can be mapped to its nearest token embedding in " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " under some similarity metric, e.g., cosine similarity. The projection function can be defined as " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{Proj}(\\mathbf{e}) = \\arg \\min_{w \\in \\mathcal{V}} \\cos(\\mathrm{EMB}(w), \\mathbf{e})" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": ". The resultant text prompt is a sequence of these projected tokens: " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "x = \\mathrm{Proj}(\\mathbf{P}) = \\{\\mathrm{Proj}(\\mathbf{e}_1), \\ldots, \\mathrm{Proj}(\\mathbf{e}_n)\\}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": ". In addition, following the approach in [53], the continuous embeddings " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " are projected into discrete tokens " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " for each forward pass in every optimization step. The gradient of the projected prompt " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": " is then applied to update the continuous embeddings " + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 562, + 173, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 173, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 173, + 574 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 580, + 214, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 214, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 214, + 591 + ], + "type": "text", + "content": "5.1.1 Evaluated Datasets and Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "content": "Since training large text-to-image diffusion models from scratch is impractical, we fine-tune the pre-trained StableDiffusion-v1-4 model provided by Huggingface with two datasets: POKEMON [30] and LAION-mi [10]. Note that both datasets do not overlap with the pre-training dataset of Stable Diffusion, ensuring that the safe models fine-tuned on them are inaccessible to copyrighted data. Following [52], each dataset is split into two disjoint shards" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " to train generative model " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " respectively. For copyrighted data " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "y_{C}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " is served as safe model safe " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "y_{C}\\in \\mathcal{D}_{1}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "y_{C}\\notin \\mathcal{D}_{2}" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": ". For better experimental illustrations, each copyrighted data is repeated to make up " + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": " of the dataset shard." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "content": "POKEMON [30]. This dataset consists of 833 image-caption pairs, where captions are obtained from the BLIP model [22]. We evaluate our attack on 5 copyrighted images added to one of the shards separately and fine-tune models " + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 304, + 133, + 545, + 192 + ], + "type": "text", + "content": " on each shard for 5000 steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "text", + "content": "LAION-mi [10]. The dataset is originally constructed for membership inference attacks on diffusion models. We only use the nonmembers part of this dataset, which holds a similar data distribution but is disjoint with the pre-training dataset of Stable Diffusion. We evaluate our attack on 5 selected copyrighted images, which are separately combined with one of non-copyrighted data shards. Each dataset shard is of size 5000. We fine-tune model " + }, + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 304, + 193, + 546, + 301 + ], + "type": "text", + "content": " for 25,000 steps. Details for fine-tuning can be found in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 319, + 437, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 437, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 437, + 330 + ], + "type": "text", + "content": "5.1.2 Implementation details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "For the sampling setting of text-to-image diffusion models, images of size " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " are generated using a classifier-free guidance scale of 7.5 and 50 sampling steps with the default scheduler. The amplification step is set to 100 and 500 for POKEMON and LAION-mi respectively. For the prompt optimization process, the learning rate is set to 0.01 using Adagrad [11] optimizer with 25,000 optimization steps, the gradient accumulation step is set to 5, the loss clip " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " for model " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " is 0.01, and the loss weight " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " is 0.95. The length of optimized prompts is set to 8 tokens. All experiments are conducted on A100 GPUs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 479, + 420, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 420, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 420, + 491 + ], + "type": "text", + "content": "5.2. Evaluation Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": "We evaluate with a model " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " behaves with an equal chance of sampling from either model " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "q_{2}" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": ". We consider such model " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " as a strict protection model that even with the original caption as prompt, there is a " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " chance of generating from the safe model. In contrast, a model " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " trained on the entire dataset could frequently produce highly similar generations to the target copyrighted image, making it extremely inefficient for the selection of a valid threshold " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": ". Next, we will separately discuss the data, the metrics, and the threat prompts for evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": "Evaluation Data. We use samples generated by " + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": " for evaluation. To determine whether they infringe the copyright of the target image, we rely on the similarity between samples and the target image as there is no widely acknowledged computable standard for infringement to our knowledge. We employ the current SOTA similarity metric SSCD [31] for copy detection. As SSCD scores of infringed samples for different target images differ significantly by obser" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "text", + "content": "1https://huggingface.co/CompVis/stable-diffusion-v1-4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12368" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 83, + 109 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 83, + 109 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 83, + 109 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 83, + 109 + ], + "type": "image", + "image_path": "95b53acae85c213b8c4795de3ee7bd052202387f749c08848dee09eb70dc65cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 71, + 148, + 109 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 148, + 109 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 148, + 109 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 148, + 109 + ], + "type": "image", + "image_path": "b5ef27f72007279251d2fbdfdbdbc79de60cc263bdaeb1dee2e59dc9766468e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 110, + 136, + 117 + ], + "lines": [ + { + "bbox": [ + 96, + 110, + 136, + 117 + ], + "spans": [ + { + "bbox": [ + 96, + 110, + 136, + 117 + ], + "type": "text", + "content": "a black and white zebrastanding next to a star" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 71, + 213, + 110 + ], + "blocks": [ + { + "bbox": [ + 149, + 71, + 213, + 110 + ], + "lines": [ + { + "bbox": [ + 149, + 71, + 213, + 110 + ], + "spans": [ + { + "bbox": [ + 149, + 71, + 213, + 110 + ], + "type": "image", + "image_path": "1eb20425d9017640ab071fa2756f7c90852d367381700449458a1aa6b144dea8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 110, + 198, + 114 + ], + "lines": [ + { + "bbox": [ + 162, + 110, + 198, + 114 + ], + "spans": [ + { + "bbox": [ + 162, + 110, + 198, + 114 + ], + "type": "text", + "content": "zebra weet unex mee" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 216, + 71, + 247, + 110 + ], + "blocks": [ + { + "bbox": [ + 216, + 71, + 247, + 110 + ], + "lines": [ + { + "bbox": [ + 216, + 71, + 247, + 110 + ], + "spans": [ + { + "bbox": [ + 216, + 71, + 247, + 110 + ], + "type": "image", + "image_path": "1451d9ee728b3151c69393fd0353a405c1bea4c177058ca4038e21d7111971b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 256, + 71, + 313, + 110 + ], + "blocks": [ + { + "bbox": [ + 256, + 71, + 313, + 110 + ], + "lines": [ + { + "bbox": [ + 256, + 71, + 313, + 110 + ], + "spans": [ + { + "bbox": [ + 256, + 71, + 313, + 110 + ], + "type": "image", + "image_path": "ab818ab2deb108b12b205fdd2705ac17705ee3d706f79e5f00080cc3b9ce7e22.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 110, + 304, + 116 + ], + "lines": [ + { + "bbox": [ + 256, + 110, + 304, + 116 + ], + "spans": [ + { + "bbox": [ + 256, + 110, + 304, + 116 + ], + "type": "text", + "content": "a picture of a white horsewith orange and yellow slings" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 376, + 110 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 376, + 110 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 376, + 110 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 376, + 110 + ], + "type": "image", + "image_path": "c28ca20a4883a1fdbc3e654731085d20c83dc23d151f2aec655e47369092c03d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 381, + 71, + 412, + 110 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 412, + 110 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 412, + 110 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 412, + 110 + ], + "type": "image", + "image_path": "4738cbedbf6175654f511a0a3d75b518c6b02a54500f3438a1d31228f813bbb4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 416, + 71, + 477, + 110 + ], + "blocks": [ + { + "bbox": [ + 416, + 71, + 477, + 110 + ], + "lines": [ + { + "bbox": [ + 416, + 71, + 477, + 110 + ], + "spans": [ + { + "bbox": [ + 416, + 71, + 477, + 110 + ], + "type": "image", + "image_path": "68d118a425ea0f8ba18455c62ab779d2c9843cf3e3d7e5d68fac65589c0038e4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 110, + 466, + 116 + ], + "lines": [ + { + "bbox": [ + 427, + 110, + 466, + 116 + ], + "spans": [ + { + "bbox": [ + 427, + 110, + 466, + 116 + ], + "type": "text", + "content": "a cartoon rabbit with asocce around its neck" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 485, + 71, + 541, + 110 + ], + "blocks": [ + { + "bbox": [ + 485, + 71, + 541, + 110 + ], + "lines": [ + { + "bbox": [ + 485, + 71, + 541, + 110 + ], + "spans": [ + { + "bbox": [ + 485, + 71, + 541, + 110 + ], + "type": "image", + "image_path": "62eec3ba1ce2c6fcab3b98dd02c157513021b7365b6ddfea6007c861a948c35b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 110, + 529, + 116 + ], + "lines": [ + { + "bbox": [ + 493, + 110, + 529, + 116 + ], + "spans": [ + { + "bbox": [ + 493, + 110, + 529, + 116 + ], + "type": "text", + "content": "rabbit, lamewridayday catunmibrated the room" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 121, + 83, + 152 + ], + "blocks": [ + { + "bbox": [ + 51, + 121, + 83, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 121, + 83, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 121, + 83, + 152 + ], + "type": "image", + "image_path": "7fc6e76e20515ddc3275a1cde8f58f1ed400bf6904dda158406af806e2215961.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 85, + 121, + 148, + 152 + ], + "blocks": [ + { + "bbox": [ + 85, + 121, + 148, + 152 + ], + "lines": [ + { + "bbox": [ + 85, + 121, + 148, + 152 + ], + "spans": [ + { + "bbox": [ + 85, + 121, + 148, + 152 + ], + "type": "image", + "image_path": "3cb4dbaa395da339682983e353f04fdef965ec257ea73baf3ba19e0dce2d0328.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 155, + 138, + 159 + ], + "lines": [ + { + "bbox": [ + 94, + 155, + 138, + 159 + ], + "spans": [ + { + "bbox": [ + 94, + 155, + 138, + 159 + ], + "type": "text", + "content": "Giraffes Canvas Art Print" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 150, + 121, + 212, + 152 + ], + "blocks": [ + { + "bbox": [ + 150, + 121, + 212, + 152 + ], + "lines": [ + { + "bbox": [ + 150, + 121, + 212, + 152 + ], + "spans": [ + { + "bbox": [ + 150, + 121, + 212, + 152 + ], + "type": "image", + "image_path": "6d3d505ba0e79062a8f745735c89fdcd32a36dc48c978b7a758e633cb9ccbab5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 152, + 199, + 156 + ], + "lines": [ + { + "bbox": [ + 164, + 152, + 199, + 156 + ], + "spans": [ + { + "bbox": [ + 164, + 152, + 199, + 156 + ], + "type": "text", + "content": "giraffe guintspitnoche" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 156, + 198, + 160 + ], + "lines": [ + { + "bbox": [ + 164, + 156, + 198, + 160 + ], + "spans": [ + { + "bbox": [ + 164, + 156, + 198, + 160 + ], + "type": "text", + "content": "tallent big etary" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 171, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 546, + 205 + ], + "type": "text", + "content": "Figure 3. Visualization of generated images on different copyright targets. The examples in the first and second rows are selected from POKEMON and LAION-mi respectively. The prompts used to generate output are given below each group of images. Remarkably, the copyright-infringed content generated with Anti-NAF amplification reveals the vulnerability of probabilistic copyright protection CP-" + }, + { + "bbox": [ + 46, + 171, + 546, + 205 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 171, + 546, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 216, + 121, + 247, + 152 + ], + "blocks": [ + { + "bbox": [ + 216, + 121, + 247, + 152 + ], + "lines": [ + { + "bbox": [ + 216, + 121, + 247, + 152 + ], + "spans": [ + { + "bbox": [ + 216, + 121, + 247, + 152 + ], + "type": "image", + "image_path": "1643aec6851c347c356b1280fd57891a61eb3ebff289cc4df0af831e31d43289.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 250, + 121, + 313, + 152 + ], + "blocks": [ + { + "bbox": [ + 250, + 121, + 313, + 152 + ], + "lines": [ + { + "bbox": [ + 250, + 121, + 313, + 152 + ], + "spans": [ + { + "bbox": [ + 250, + 121, + 313, + 152 + ], + "type": "image", + "image_path": "b9e5962edd28115ac91b9514d0e16c1695e27b78a744294ded66906e036fa139.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 152, + 312, + 162 + ], + "lines": [ + { + "bbox": [ + 250, + 152, + 312, + 162 + ], + "spans": [ + { + "bbox": [ + 250, + 152, + 312, + 162 + ], + "type": "text", + "content": "Product photo wall tattoo children's rostradberries with the inksticker set" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 121, + 377, + 152 + ], + "blocks": [ + { + "bbox": [ + 315, + 121, + 377, + 152 + ], + "lines": [ + { + "bbox": [ + 315, + 121, + 377, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 121, + 377, + 152 + ], + "type": "image", + "image_path": "9004b5819071094d27dbf3d52bb4ee071e43f32f3557f60df2ed2499c9694c82.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 152, + 369, + 156 + ], + "lines": [ + { + "bbox": [ + 323, + 152, + 369, + 156 + ], + "spans": [ + { + "bbox": [ + 323, + 152, + 369, + 156 + ], + "type": "text", + "content": "horse millennials symbolism" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 156, + 369, + 160 + ], + "lines": [ + { + "bbox": [ + 323, + 156, + 369, + 160 + ], + "spans": [ + { + "bbox": [ + 323, + 156, + 369, + 160 + ], + "type": "text", + "content": "kano jacoblissseasoned" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 381, + 121, + 412, + 152 + ], + "blocks": [ + { + "bbox": [ + 381, + 121, + 412, + 152 + ], + "lines": [ + { + "bbox": [ + 381, + 121, + 412, + 152 + ], + "spans": [ + { + "bbox": [ + 381, + 121, + 412, + 152 + ], + "type": "image", + "image_path": "a3f7f659a6df2a56c22629c0845bac9a343e2d45d6db9524d597e7a4b75e999f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 419, + 121, + 477, + 152 + ], + "blocks": [ + { + "bbox": [ + 419, + 121, + 477, + 152 + ], + "lines": [ + { + "bbox": [ + 419, + 121, + 477, + 152 + ], + "spans": [ + { + "bbox": [ + 419, + 121, + 477, + 152 + ], + "type": "image", + "image_path": "2fa8a38bb0628737b71ffb02cfe8ac30a11e3b36a401a42d34ec388733954484.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 152, + 472, + 158 + ], + "lines": [ + { + "bbox": [ + 420, + 152, + 472, + 158 + ], + "spans": [ + { + "bbox": [ + 420, + 152, + 472, + 158 + ], + "type": "text", + "content": "Penguins pictures of the New" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 158, + 472, + 162 + ], + "lines": [ + { + "bbox": [ + 420, + 158, + 472, + 162 + ], + "spans": [ + { + "bbox": [ + 420, + 158, + 472, + 162 + ], + "type": "text", + "content": "Year, March with their own hands" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 481, + 121, + 511, + 152 + ], + "blocks": [ + { + "bbox": [ + 481, + 121, + 511, + 152 + ], + "lines": [ + { + "bbox": [ + 481, + 121, + 511, + 152 + ], + "spans": [ + { + "bbox": [ + 481, + 121, + 511, + 152 + ], + "type": "image", + "image_path": "a2896a3a56ee6d52cd313290683163f9aaa112c52044fc0e8c6792352039a4da.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 152, + 535, + 158 + ], + "lines": [ + { + "bbox": [ + 489, + 152, + 535, + 158 + ], + "spans": [ + { + "bbox": [ + 489, + 152, + 535, + 158 + ], + "type": "text", + "content": "penguins knobs swimming" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 158, + 535, + 162 + ], + "lines": [ + { + "bbox": [ + 488, + 158, + 535, + 162 + ], + "spans": [ + { + "bbox": [ + 488, + 158, + 535, + 162 + ], + "type": "text", + "content": "orbcsport orbfunday puppies have" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 53, + 216, + 541, + 324 + ], + "blocks": [ + { + "bbox": [ + 53, + 216, + 541, + 324 + ], + "lines": [ + { + "bbox": [ + 53, + 216, + 541, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 541, + 324 + ], + "type": "table", + "html": "
MethodsPOKEMONLAION-mi
CIRFAR@5%AR↑FAR@15%AR↑CIRFAR@10%AR↑FAR@30%AR↑FAR@50%AR↑
Caption (w/o Amp.)47.00%0.40%3.28%49.64%0.00%0.00%0.04%
CLIP-Int. (w/o Amp.)26.92%0.84%2.20%48.12%0.84%1.52%2.96%
PEZ (w/o Amp.)7.80%1.32%2.76%15.80%0.08%0.24%0.40%
Anti-NAF (w/o Amp.)12.88%8.52%10.00%33.84%2.64%4.16%7.00%
Caption (w/ Amp.)100.00%13.64%40.16%100.00%0.00%0.00%14.64%
CLIP-Int. (w/ Amp.)99.72%22.84%47.32%100.00%34.84%37.12%75.60%
PEZ (w/ Amp.)66.92%22.80%37.48%98.16%38.04%50.68%59.64%
Anti-NAF (w/ Amp.)99.84%77.36%91.36%100.00%56.12%73.32%95.92%
", + "image_path": "912d5157b8075d9cefc8afe69492873eea10ec2d78a33fc09a49c18a4d6a5d7b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "type": "text", + "content": "Table 1. Quantitative results. The performance with the amplification attack (\"w/ Amp.\") is significantly superior to scenarios without amplification (\"w/o Amp.\") Additionally, our proposed Anti-NAF demonstrates notably promising outcomes for providing a substantial probability of copyright infringement when probabilistic protection is applied. (CLIP-Int. is the abbreviation for CLIP-Interrogator)." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 387, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 469 + ], + "type": "text", + "content": "vation (shown in Fig. 5), we do not indicate a fixed score threshold for all target images. Instead, we use relative thresholds determined as percentiles of the similarity scores among samples generated by " + }, + { + "bbox": [ + 46, + 387, + 287, + 469 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 387, + 287, + 469 + ], + "type": "text", + "content": " with the original caption of the target image as the prompt, e.g. SSCD-50%. Recognizing the variable criteria of infringement, we report results with other choices of thresholds in Sec. 11.3." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": "Evaluation Metrics. The CP-" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": " method achieves copyright protection by selectively accepting generated samples using the threshold " + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "inline_equation", + "content": "k_{x}" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": ", which can be indicated by the Acceptance Rate (AR). For a given AR, a good protection system is expected to rarely accept infringing content, i.e., have a low False Accept Rate (FAR), as defined in Eq. (8). It is worth noting that the choice of " + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "inline_equation", + "content": "k_{x}" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": " dictates the trade-off between model safety and efficiency. Furthermore, to our knowledge, there is no principled way to determine " + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "inline_equation", + "content": "k_{x}" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": ". As a result, we evaluate the success of attack by reporting the FAR at different AR, e.g., FAR@5%AR. Additionally, the copyright infringement rate (CIR) is also presented for scenarios without copyright protection, i.e., " + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{AR} = 100\\%" + }, + { + "bbox": [ + 46, + 478, + 287, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Threat Prompts. Given that generating images similar to the target image is a necessary condition for a successful attack, we focus on the following threat prompts: (i) Target caption: the original caption of the target image; (ii) PEZ [53]: a gradient-based discrete optimization approach to discover prompts semantically similar to the target im" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": "age; (iii) " + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "inline_equation", + "content": "CLIP" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": "-Interrogator2: a prompt consists of the BLIP caption of the target image and top-" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": " keywords greedily sampled from a keywords collection; (iv) Anti-NAF: an adversarial prompt against NAF copyright protection as described in Sec. 4.4. For each threat prompt, we sample 6400 images from model " + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": " for evaluation. For the online prompt selection setting, we take the aforementioned three prompts (exclude the target caption) as a prompt candidate set and consider " + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 304, + 387, + 546, + 506 + ], + "type": "text", + "content": "-greedy-max-/cdf Bandit Amplification as bandit strategies as described in Sec. 4.3." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 517, + 362, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 362, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 362, + 528 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 536, + 546, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 546, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 546, + 692 + ], + "type": "text", + "content": "In Fig. 2, we show example outputs under four scenarios. It is evident that with amplification, the CP-" + }, + { + "bbox": [ + 304, + 536, + 546, + 692 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 536, + 546, + 692 + ], + "type": "text", + "content": " protected generative model does indeed output copyright-infringed content with high probability, especially when using the Anti-NAF prompt. These examples suggest that probabilistic safeguards against copyright infringement are vulnerable to the amplification attack. Additional visualizations of outputs on various copyrighted images are demonstrated in Fig. 3. In Tab. 1, we report CIRs and FARs at various ARs on two datasets. Notably, there is a significant growth in both metrics when amplification attack is employed, highlighting its effectiveness in amplifying the probability of infringing generations. The superior performance of Anti-NAF under" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 493, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 493, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 493, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 315, + 702, + 493, + 713 + ], + "type": "text", + "content": "https://github.com/pharmapsychotic/clip-interrogator" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12369" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 167, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 167, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 167, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 167, + 210 + ], + "type": "image", + "image_path": "5fb7f79a69147f808b80b5828cc17ae37d7f803124c0c812bd1a489313060ef8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 71, + 286, + 211 + ], + "blocks": [ + { + "bbox": [ + 169, + 71, + 286, + 211 + ], + "lines": [ + { + "bbox": [ + 169, + 71, + 286, + 211 + ], + "spans": [ + { + "bbox": [ + 169, + 71, + 286, + 211 + ], + "type": "image", + "image_path": "a4e9ae3771e25aa7a7b15c35e579417aed5bc7338d969fefb74fc22f4acbb04e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 214, + 166, + 337 + ], + "blocks": [ + { + "bbox": [ + 49, + 214, + 166, + 337 + ], + "lines": [ + { + "bbox": [ + 49, + 214, + 166, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 214, + 166, + 337 + ], + "type": "image", + "image_path": "f47bfac79e07c02ea7a3fef834b477560fbe3ea215da2b9a198281e1b757fa85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 340, + 165, + 360 + ], + "lines": [ + { + "bbox": [ + 48, + 340, + 165, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 340, + 165, + 360 + ], + "type": "text", + "content": "(c) FAR-AR of Anti-NAF with different Amp. steps on POKEMON." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 369, + 287, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 403 + ], + "type": "text", + "content": "Figure 4. The overall FAR-AR curves. The results show that amplification is strongly effective in amplifying the possibility of infringed output, even with small amplification steps." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 214, + 286, + 337 + ], + "blocks": [ + { + "bbox": [ + 169, + 214, + 286, + 337 + ], + "lines": [ + { + "bbox": [ + 169, + 214, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 169, + 214, + 286, + 337 + ], + "type": "image", + "image_path": "645819ffcfa46b03dabb3c216801f96a8dece9ad2c35481fe494a92243d65264.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 340, + 286, + 360 + ], + "lines": [ + { + "bbox": [ + 169, + 340, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 169, + 340, + 286, + 360 + ], + "type": "text", + "content": "(d) FAR-AR of Anti-NAF with different Amp. steps on LAION-mi." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 605 + ], + "type": "text", + "content": "scores its efficacy in rendering infringing generations with a substantial probability. The overall FAR-AR curves are illustrated in Fig. 4. We can observe that bandit variants of amplification lead to a smaller variance across different target copyrighted images, especially at lower acceptance rates, indicating that our bandit strategies achieve a more steady attack. In Figs. 4c and 4d, we plot the overall FAR-AR curves over different amplification steps. There is a clear trend of rapidly improved performance with increased amplification steps, due to the cumulative probability of generating infringing samples. This finding underlines the potential risk in practical applications of probabilistic copyright protections, given the high frequency of daily interactions with text-to-image generative models. We provide additional results and human evaluations in Sec. 11." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 141, + 624 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "In Tab. 2, we study how different components of our optimization objective in Eq. (13), affect the performance. We discover that an exclusive focus on minimizing " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " (\"" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " only\") leads to a notable drop in performance. This indicates that prompts designed merely to reconstruct the target image are insufficient for a successful attack, as the infringing output of such a prompt can be easily identified and" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 322, + 71, + 531, + 131 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 531, + 131 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 531, + 131 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 531, + 131 + ], + "type": "table", + "html": "
MethodsCIRFAR@5%AR↑FAR@15%AR↑
Anti-NAF12.88%8.52%10.00%
\\( \\mathcal{L}_p \\) only10.60%1.84%2.04%
w/o \\( \\varphi \\)12.12%1.36%1.48%
w/o \\( \\mathcal{L}_q \\)13.72%2.00%2.76%
", + "image_path": "92b18b739f03a27935fd18f1c21163bbe9007a976b84ef4995b4d81c5bbfdcb5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 139, + 544, + 150 + ], + "lines": [ + { + "bbox": [ + 307, + 139, + 544, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 544, + 150 + ], + "type": "text", + "content": "Table 2. Ablation study for Anti-NAF algorithm on POKEMON." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": "blocked by the copyright protection system. Furthermore, a direct combination of " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " (\"w/o " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": ") results in additional performance degradation because of the conflicting objectives of " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " in minimizing " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\rho(y_C|x)" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": ". On the other hand, we can observe a slight performance improvement when implementing a loss clip " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": " (\"w/o " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": ") to constrain the learning of " + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 172, + 545, + 280 + ], + "type": "text", + "content": ". Overall, these results validate the effectiveness of our well-balanced optimization objective. Additional ablation studies are provided in Sec. 12." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 291, + 378, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 291, + 378, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 291, + 378, + 303 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 311, + 546, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 546, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 546, + 598 + ], + "type": "text", + "content": "In this paper, we shed light on the vulnerability of probabilistic copyright protection methods for text-to-image generative models, especially in real-world scenarios involving persistent and targeted interactions. Our proposed Virtually Assured Amplification Attack (VA3) framework presents the feasibility of inducing the protected model to generate copyright-infringed content with an amplified probability. Despite our focus on a narrow scenario of online prompt selection within this framework, the experimental results highlight its effectiveness in challenging even the most advanced existing probabilistic copyright protection methods. However, a broader scope of potential strategies remains unexplored within the VA3 framework, such as online prompt optimization, which may provide more powerful attacks against copyright protection. Furthermore, our Anti-NAF algorithm relies on access to generative models for adversarial prompt optimization, an assumption that may not be satisfied in completely black-box scenarios. We leave these more complicated and general attack methods for future investigation. In conclusion, our findings emphasize the significant risk of copyright infringement when applying probabilistic copyright protection methods in practice. Therefore, we hope that this work can inspire the development of more robust copyright protection approaches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The authors are grateful to Nikhil Vyas for providing the source code of CP-" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " algorithm. This research is partially supported by the National Research Foundation Singapore under the AI Singapore Programme (AISG Award No: AISG2-TC-2023-010-SGIL) and the Singapore Ministry of Education Academic Research Fund Tier 1 (Award No: T1 251RES2207)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12370" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Hossein Aboutalebi, Daniel Mao, Carol Xu, and Alexander Wong. Deepfakeart challenge: A benchmark dataset for generative ai art forgery and data poisoning detection. arXiv preprint arXiv:2306.01272, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Stephen Casper, Zifan Guo, Shreya Mogulothu, Zachary Marinov, Chinmay Deshpande, Rui-Jie Yew, Zheng Dai, and Dylan Hadfield-Menell. Measuring the success of diffusion models at imitating human artists. arXiv preprint arXiv:2307.04028, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[5] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "type": "text", + "content": "[6] David Chess. Some light infringement? https://ceoln.wordpress.com/2022/12/16/some-light-infringement/, 2022.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "text", + "content": "[7] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[8] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 513 + ], + "type": "text", + "content": "[9] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[10] Jan Dubiński, Antoni Kowalczuk, Stanisław Pawlak, Przemysław Rokita, Tomasz Trzeciński, and Pawel Morawiecki. Towards more realistic membership inference attacks on large diffusion models. arXiv preprint arXiv:2306.12983, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[11] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12(7), 2011. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[12] Niva Elkin-Koren, Uri Hacohen, Roi Livni, and Shay Moran. Can copyright be reduced to privacy? arXiv preprint arXiv:2305.14822, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[15] Hongcheng Gao, Hao Zhang, Yinpeng Dong, and Zhi-jie Deng. Evaluating the robustness of text-to-image diffusion models against real-world attacks. arXiv preprint arXiv:2306.13103, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. arXiv preprint arXiv:2303.15715, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "text", + "content": "[18] Matthew Butterick Joseph Saveri. Stable diffusion litigation. https://stablediffusionlitigation.com/, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[19] Kenji Kawaguchi, Zhun Deng, Xu Ji, and Jiaoyang Huang. How does information bottleneck help deep learning? In International Conference on Machine Learning (ICML), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "text", + "content": "[20] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "type": "text", + "content": "[21] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 12888-12900. PMLR, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "type": "text", + "content": "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 20763-20786. PMLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "text", + "content": "[25] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 21548-21566. PMLR, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12371" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[28] Natalie Maus, Patrick Chao, Eric Wong, and Jacob Gardner. Adversarial prompting for black box foundation models. arXiv preprint arXiv:2302.04237, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[29] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Protecting the intellectual property of diffusion models by the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[30] Justin N. M. Pinkney. Pokemon blip captions. https://huggingface.co/datasets/lambdaabs/pokemon-blip-captions/, 2022.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "type": "text", + "content": "[31] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14512-14522. IEEE, 2022. 6, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 337 + ], + "type": "text", + "content": "[32] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. arXiv preprint arXiv:2305.13873, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[35] Javier Rando, Daniel Paleka, David Lindner, Lennard Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[36] Arkadip Ray and Somaditya Roy. Recent trends in image watermarking techniques for copyright protection: a survey. International Journal of Multimedia Information Retrieval, 9(4):249-270, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 635 + ], + "type": "text", + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[41] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 29894-29918. PMLR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "type": "text", + "content": "[42] Sarah Scheffler, Eran Tromer, and Mayank Varia. Formalizing human ingenuity: A quantitative framework for copyright law's substantial similarity. In Proceedings of the 2022 Symposium on Computer Science and Law, pages 37-49, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "text", + "content": "[43] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[45] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "text", + "content": "[46] Shawn Shan, Wenxin Ding, Josephine Passananti, Haitao Zheng, and Ben Y Zhao. Prompt-specific poisoning attacks on text-to-image generative models. arXiv preprint arXiv:2310.13828, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[47] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 558 + ], + "type": "text", + "content": "[48] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 6048-6058. IEEE, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[49] Growthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "text", + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[51] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12372" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 531 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[52] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "type": "text", + "content": "[53] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 287, + 217 + ], + "type": "text", + "content": "[54] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 261 + ], + "type": "text", + "content": "[55] Yixin Wu, Ning Yu, Michael Backes, Yun Shen, and Yang Zhang. On the proactive generation of unsafe images from text-to-image models using benign prompts. arXiv preprint arXiv:2310.16613, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 306 + ], + "type": "text", + "content": "[56] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "type": "text", + "content": "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. arXiv preprint arXiv:2305.04175, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "type": "text", + "content": "[58] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 398, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 398, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 398, + 287, + 430 + ], + "type": "text", + "content": "[59] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 287, + 495 + ], + "type": "text", + "content": "[60] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023 - Workshops, Vancouver, BC, Canada, June 17-24, 2023, pages 2385-2392. IEEE, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 287, + 531 + ], + "type": "text", + "content": "[61] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12373" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_content_list.json b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b03ba81534cf9c19ef7b449864b45f26b37d4e8c --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_content_list.json @@ -0,0 +1,1143 @@ +[ + { + "type": "text", + "text": "VAREN: Very Accurate and Realistic Equine Network", + "text_level": 1, + "bbox": [ + 210, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Silvia Zuffi1, Ylva Mellbin2, Ci Li3, Markus Hoeschle4, Hedvig Kjellström3,2, Senya Polikovsky4, Elin Hernlund2, and Michael J. Black4", + "bbox": [ + 99, + 191, + 870, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1IMATI-CNR, Milan, Italy, 2SLU, Uppsala, Sweden, 3KTH, Stockholm, Sweden, 4Max Planck Institute for Intelligent Systems, Tübingen, Germany", + "bbox": [ + 163, + 239, + 807, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93023e9dfad391bddb5a2e6a18b8517fe6f887e16d1d0a0f789370d6b8d993ee.jpg", + "image_caption": [ + "Figure 1. VAREN is learned from real horses and models pose-dependent deformations at the muscle level, enabling the generation of 3D horses with high realism. Inspired by Muybridge [24], we show a sequence of frames of horses in motion, generated by VAREN, where the first frame illustrates, for each row, the horse in the rest pose. We show three different shapes, obtained by sampling the VAREN model." + ], + "image_footnote": [], + "bbox": [ + 161, + 297, + 815, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 558, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data-driven three-dimensional parametric shape models of the human body have gained enormous popularity both for the analysis of visual data and for the generation of synthetic humans. Following a similar approach for animals does not scale to the multitude of existing animal species, not to mention the difficulty of accessing subjects to scan in 3D. However, we argue that for domestic species of great importance, like the horse, it is a highly valuable investment to put effort into gathering a large dataset of real 3D scans, and learn a realistic 3D articulated shape model. We introduce VAREN, a novel 3D articulated parametric shape model learned from 3D scans of many real horses. VAREN bridges synthesis and analysis tasks, as the generated model instances have unprecedented realism, while being able to represent horses of different sizes and shapes. Differently from previous body models, VAREN has two resolutions, an anatomical skeleton, and interpretable, learned pose-dependent deformations, which are related to the body muscles. We show with experiments that this formulation has superior performance with respect to previous strategies for modeling pose-dependent deformations.", + "bbox": [ + 73, + 583, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mations in the human body case, while also being more compact and allowing an analysis of the relationship between articulation and muscle deformation during articulated motion. The VAREN model and data are available at https://varen.is.tue.mpg.de.", + "bbox": [ + 498, + 559, + 893, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 648, + 633, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Horses are arguably the most valuable domestic animal and there is a large industry focused on their breeding, care, training, and use in sports. Buying a horse is a large investment, and keeping a horse requires resources and time. Although they are large animals, horses are delicate. The accumulated loads from training and competition frequently result in unrecoverable injuries in their skeletal or tendinous limb structures, ultimately leading to euthanasia. Consequently, among domestic animals, horses are widely studied from a behavioral and biomechanical perspective, with the aim of evaluating their performance, interpreting their state of pain, and preventing injuries. To facilitate such analysis using computer vision, our goal is to develop a highly accurate and detailed 3D model of horses that can be articulated, fit to data, and animated. Here we learn such a model", + "bbox": [ + 496, + 672, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5374", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8827cc573e621c2dd373afb5a98d9b50392b6407c5e3f30426e2d11e2a558fb2.jpg", + "image_caption": [ + "Figure 2. Dynamic 3D horse scanner. The horse is asked to lift one hind limb." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 450, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(see Fig. 1) using a novel formulation that captures pose-dependent muscle deformations. Such a model may facilitate 3D motion analysis in-the-wild, body shape estimation, diagnosis of illness and performance, horse-human interaction and behavior analysis, and generation of 3D synthetic horses for VFX, VR, AR, and gaming. In particular, accurate markerless motion capture of horses would provide a valuable new tool for the study of horse motion during challenging activities in natural environments.", + "bbox": [ + 75, + 295, + 467, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For humans, there has been great progress on capturing their 3D shape and pose from images and video. This work often exploits the SMPL body model [22], a 3D mesh-based representation of the human body, controlled by 3D shape and pose parameters. SMPL provides strong priors over human shape and pose, enabling 3D pose reconstruction from ambiguous 2D data. Creating such a 3D model of horses is more challenging. While 3D articulated shape models of horses have been created in the past, they lack realism and expressiveness. In contrast, models like SMPL are learned from thousands of scans, and include an expressive shape space and a pose-dependent deformation model to represent how the body deforms under articulation. Unfortunately, obtaining high-quality 3D scans of horses in a variety of poses is challenging. To fill the gap, we need to (1) capture 3D scans of horses in motion and (2) formulate a new parametric 3D model of horses that represents important aspects of their physiology. In both cases, our solutions deviate from prior work on humans to accommodate the unique aspects of horses.", + "bbox": [ + 75, + 431, + 467, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset. We use a novel setup (3dMD Ltd.) to capture dynamic 3D scans of horses over time (see Fig. 2). Scanning a large number of horses presents technical and practical challenges, requiring a large cooperative effort between computer scientists and animal care experts. Our dataset includes 50 horses of different breeds and sizes, ranging from small ponies to large horses, with a height difference of more than 1.5 meters. This shape variation is significantly larger than with humans. To model pose-dependent shape changes, we need to capture the horses in a range of poses. Additionally, the animal is always managed by a care per", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/344b873cea3f8996ca2cf958a65eb4e93c967df1fc3c98dd5db450b30bfabd13.jpg", + "image_caption": [ + "Figure 3. Samples from the VAREN model result in a diversity of horse shapes." + ], + "image_footnote": [], + "bbox": [ + 555, + 88, + 841, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "son who may occlude the horse during scanning, resulting in missing data. All these issues present novel challenges that we overcome. As a result, our dataset includes in total approximately 4000 cleaned up raw 3D scans, for which we provide accurately registered 3D meshes – training and testing split. Written consent was provided by all animal owners before the animals entered the study. The data collection procedure did not include any invasive technique, therefore no animal ethical approval was needed.", + "bbox": [ + 496, + 292, + 890, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model. Using this unique dataset, we train a new model named VAREN (pronounced Varenne1 in French). VAREN is the first 3D articulated model of horses that is learned from real data and that models pose-dependent muscle deformations. Samples from the VAREN model are shown in a neutral pose in Fig. 3 and animated using motion capture data in Fig. 1. To model horses in motion, we must also capture the non-rigid deformations of the body that occur during movement. Previous work on humans models pose-dependent deformations that are learned from 3D scans and conditioned on the body articulation. In particular, STAR [25] models such deformations locally based on the distance from the nearby joints. Such methods do not explicitly model the deformation of the muscles. Scans of humans are typically acquired with some form of clothing (not fully naked) and most humans have a layer of subcutaneous adipose tissue that hides the musculature. Horses, however, are different. They generally carry less body fat than humans and their hair is short, making their musculature more directly observable. Modeling the visible muscle deformation is important because it relates directly to the health and performance of the animal. To that end, we propose a novel learned muscle-based deformation model. Specifically, we group the body surface points into regions that correspond to the superficial muscles, rather than based on their distance to the joints. Note that a single muscle can span more than one body part (see for example the back muscle in Fig. 4) and it is important to capture these long-range cor", + "bbox": [ + 496, + 429, + 890, + 852 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Varenne is considered to be the best trotter of all time. No other trotter has won so many of the most important races in the world and set as many records as Varenne. Source: Wikipedia.", + "bbox": [ + 500, + 862, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5375", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ac79c7ce7328841311ad7779ed1f9a38170aa7ee5d9b7b4ba80673d28ed2e2d.jpg", + "image_caption": [ + "Figure 4. Skeleton and muscles. From left: anatomical skeleton manually aligned to the model template, selected joints, muscle structure and skin labels on the template. We consider 76 muscles. Head, tail and hooves have no muscle definitions." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 464, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relations. In VAREN, we define deformations per-muscle and learn the influence of body part articulation on the observable muscle deformation from the scan data. We find that, in comparison with previous formulations like those used in SMPL and STAR, our proposed muscle-inspired pose-dependent deformation model has superior accuracy, while also being more compact and interpretable. This is relevant for our high-resolution horse model. While we do not measure muscle activation directly, our formulation can be relevant for research on body pose using measurements of muscle activation [11]. A further novelty of our model, compared with previous work, is that we define the model skeleton, that is the location of the joints in the kinematic tree, based on real horse anatomy. Exploiting a 3D model of the horse skeleton, and with the help of an expert on horse biomechanics, we define the articulation points for the model. Previous human body models define the joint locations as a result of an optimization procedure, resulting in a per-model definition with poor correspondence to the true anatomical skeleton, in particular at the shoulders and hip joints [17]. These models are motivated by applications in computer graphics and vision, where, in many cases, a precise anatomical skeleton may not be needed. For a model like VAREN to be valuable to horse breeders, trainers, and veterinarians, it should provide information that is anatomically relevant. Moreover, an anatomical definition of the model joints, being model-independent, could facilitate comparison and motion data transfer between different models, and, eventually, species.", + "bbox": [ + 76, + 239, + 467, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make two key contributions. First, we provide a new dataset for the modeling of horse shape and pose. The quantity, resolution, variation, and quality of the VAREN data dwarfs any previous dataset of animal shape and pose, and opens up the possibility for the community to explore new questions in 3D shape representation. Second, we develop a new shape model that drives the observable deformation of the muscles based on the animal pose using an anatomical skeleton.", + "bbox": [ + 75, + 679, + 467, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 218, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We review prior work on learning human body models from data, as our problem and setting share similarities with methods proposed in this case. Then, we present the few", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D models for animals previously proposed. In addition, we provide an overview of methods for 3D animal reconstruction from visual data, in order to set the context in which we see our novel model applied.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Body Models. There is a long history of learning 3D mesh-based models of the human body from 3D scans [4-6, 10, 13, 22, 26, 35]. There is also recent work on implicit representations that we do not consider here, e.g. [3, 21, 23]. Most recent work is based on SMPL [22], a 3D body model that is based on vertex deformations defined by linear spaces. SMPL is fully data-driven: pose dependent deformations are defined as global representations and learned from data. One issue with SMPL is that the pose-dependent deformations are not local and capture spurious long-range correlations in the data. This is addressed by STAR [25], which adds locality to the joints to the pose-dependent deformations, weighting the outcome of global linear spaces with the distance from skeleton joints. While STAR conditions the pose deformation on the rough BMI of the body, none of the models above explicitly model the observable muscle deformation. VAREN addresses this for the first time. Also, unlike VAREN, the prior methods model articulation using a kinematic tree only loosely based on the human skeleton. A first direction into modeling the human skeleton is OSSO [17], who models the human body skeleton, followed by SKEL [18], which drives the pose of SMPL with a biomechanical skeletal model. SKEL, however, uses the default SMPL model deformations and, unlike VAREN does not train deformations that are driven by the skeletal pose.", + "bbox": [ + 496, + 155, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Animal Body Models. The modeling of animal shape for computer vision applications has received less attention. Since 3D scanning of animals is challenging, research has mainly focused on learning from images; e.g. Ramanan et al. learn 2D articulated models [27], while Cashman and Fitzgibbon [9] adapt a rough mean shape using 2D image cues to reconstruct dolphins and bears in 3D. Zuffi et al. learn SMAL, a parametric 3D animal shape model, using 3D scans of toys [44]. Wang et al. learn the 3D shape variation of birds from images starting from a synthetic model [32]. Li et al. learn a horse-specific model similar to SMAL, using 3D scans of horse figurines [20]. Still, others learn to represent category-specific shapes as either meshes [39] or implicit surfaces [14], but do not decouple shape and pose. None of these methods learn 3D articulated models from real 3D scans, which contain noise. The multi-species SMAL model [44] contains an Equine class, but SMAL is learned from a limited number of toys, thus horses generated with the SMAL model do not exhibit a wide range of diverse shapes. The more recent hSMAL model is horse-specific [20]. While learned from a larger number of horses of different shapes, hSMAL is still learned from toys, and has a limited resolution. While useful for reconstructing", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5376", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "horses from images and video, we find in our experiments that hSMAL is not expressive enough to accurately represent real horses with high detail. Being learned from rigid toys, both models do not address body deformations.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Animal 3D Reconstruction. The 3D reconstruction of animals from images and video follows a model-based or a model free approach. In the first case, an existing 3D shape model of the species of interest is available, and the method outputs the model parameters of 3D shape and pose. This approach is useful when dealing with challenging input modalities, like in-the-wild monocular images, or the goal is to estimate animal shape and pose for downstream analysis tasks. For example, Zuffi et al. address multi-species 3D reconstruction [45], Kanazawa et al. learn 3D deformations of animals [15], Badger et al. reconstruct 3D birds [7], while several methods estimate the shape and pose of dogs [8, 29, 30]. Our work fits in this category, but provides a 3D model with a higher level of realism than any previous methods. Model-free methods, in contrast, do not assume that a 3D model of the animal class is available, and the output of these methods is typically a 3D surface, in case of static input, or, eventually, a 3D animatable object in case of a video input. Works in this class include CMR [16] and DOVE [33], which reconstruct birds from images and ViSER [37], LASR [36], BANMo [38] and PPR [40], which reconstruct articulated 3D shapes from video, the latter incorporating physical constraints. Kokkinos and Kokkinos learn 3D reconstruction from video without using a shape model by relying only on a template [19]. MagicPony [34], LASSIE [41], Hi-LASSIE [42], ARTIC3D [43] learn 3D reconstruction from image collections.", + "bbox": [ + 75, + 162, + 472, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 587, + 169, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn the VAREN model in two stages. First, we learn an articulated parametric shape model from a set of scans, that we call prototypes, which are of different horses in a neutral pose, that is, not undergoing pose-dependent deformations. This gives us a model with a shape space and generic articulation. We then align, in an unsupervised way, this static model to the dynamic 3D scans in which the prototype horses perform different movements. From this data, namely the scans and the alignment parameters of 3D shape, pose and translation, the VAREN network learns a pose-dependent deformation model that improves the matching between the learned model and the scans. The novel deformation model is a function of the shape and pose parameters, such that, at test time, the network generates a shaped and posed horse, with realistic pose-dependent deformations, given the parameters. The pose-dependent deformation model is defined on the basis of anatomical structures that we incorporate into VAREN from a purchased realistic graphics (CG) model [1], which includes the body skeleton", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/05f75e39a519dbb3f5470eccf022a5fbd28bdad8d6975d40cd53a653a7742272.jpg", + "image_caption": [ + "Figure 5. Comparison between the average equine model in SMAL (left), the hSMAL+ template (middle) and the VAREN template (right). Learned from real horses, the VAREN template has a more proportioned neck with respect to hSMAL+." + ], + "image_footnote": [], + "bbox": [ + 542, + 90, + 849, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f7aa45cbca78a02af6b02fee14aac131b2780ea002af4a38d4b2aae467ae515.jpg", + "image_caption": [ + "Figure 6. Visualization of the VAREN shape space. We capture small ponies, large breeds, and many in between, obtaining a shape space with variation in size, morphology and face features. The tail and mane are tied, and not considered as part of body shape change." + ], + "image_footnote": [], + "bbox": [ + 534, + 214, + 857, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the muscles. In the following, we detail the procedure to learn the horse model from the prototypes, the creation of the training dataset, and the definition of the VAREN network.", + "bbox": [ + 496, + 431, + 893, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. VAREN Horse Model", + "text_level": 1, + "bbox": [ + 500, + 501, + 702, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment of the Prototype Scans. We follow a model-based approach to express the prototypes in a common topology, specifically, as the alignment model, we use hSMAL [20], a recently proposed articulated horse model (Fig. 5). The hSMAL model follows the formulation of SMPL, and is defined by a triangular mesh template $\\mathbf{v}_t$ , with $n_V$ vertices, a matrix $B$ of shape $3n_V \\times n_B$ containing the $n_B$ basis vectors of a linear shape deformation space, a joint regressor $J_r$ that maps model vertices to a set of $n_J$ joint locations, and a skinning weight matrix $W$ . A horse is generated, given shape parameters $\\beta$ and pose parameters $\\theta$ , by first deforming the template into an intrinsic shape $\\mathbf{v}_s$ , then applying Linear Blend Skinning (LBS) to rotate the body parts according to the given pose:", + "bbox": [ + 496, + 523, + 893, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {v} _ {s} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} \\\\ \\mathbf {v} = L B S \\left(\\mathbf {v} _ {s}, \\theta ; W, J _ {r}\\right). \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 746, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While able to represent horses of different shapes, we found that the hSMAL model has, in our case, two limitations: a low resolution (the model has about $1.5K$ vertices) and a shape space that is not sufficiently expressive to accurately capture the shape of real horses. To overcome these limitations, we create an hSMAL model with an increased resolution (3647 vertices) and additional shape deformations.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5377", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We call this model hSMAL+. Furthermore, we observed that the purchased CG model [1], in addition to containing anatomical structures, has a more natural posture than hSMAL+. Therefore, we modify the rest pose of hSMAL+ to match the pose of the CG model. This is done by fitting hSMAL+ to the CG model and using the obtained mesh as the new template. In hSMAL+ we retain skinning weights and part segmentations from the original hSMAL model.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additional Shape Deformations. We augment the hSML shape space by defining additional dimensions in the shape deformation matrix $B$ (Eq. 1) to obtain a linear scaling of a set of body parts. Note that limb scaling has been applied before to animal models, but our formulation is novel, as previous work [8] applies scaling during the LBS process, while here we define scaling in the model shape space. Let $j$ be a body part for which we want to add scaling as an additional dimension in the shape space. We define a new shape matrix $B_{1}$ of dimension $3n_V \\times (n_B + 1)$ by adding a new column to the shape matrix $B$ :", + "bbox": [ + 75, + 220, + 472, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB _ {1} = B | B _ {j}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 395, + 323, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB _ {j} = s _ {j} \\mathbf {v} _ {j, c}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 414, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{j,c}$ is a column vector corresponding to the template, but with non-zero values only for the coordinate $c$ of the vertices of the part $j$ we want to scale, and $s_j$ is a scaling factor. Here $j$ can also indicate a set of parts, for example, all the segments of the tail. Expanding $B$ is not sufficient, as the deformation we obtain is expressed for the model template $\\mathbf{v}_t$ , while we need it to be applied to the part vertices of the intrinsic shape $\\mathbf{v}_s$ (Eq. 1). Therefore, we add an iterative procedure that estimates a vertex shift $d_v$ for the part vertices. Let $J$ be the joints of the extended model with shape space $B_1$ , $J_0$ be the joints of the original model, and $J_s$ a vector of cumulative joint shifts initialized with zero values. In an iterative process that considers all the body joints from the root to the leaves of the kinematic tree, for each body part $j$ , we compute the vertex shift as $d_v(j) = J_s(j) + J_0(j) - J(j)$ , add it to the part vertices, and then add $d_v(j)$ to the children of the part $j$ in $J_s$ , such that the joint shift is propagated through the kinematic tree. The set of body parts we consider for limb scaling are the ears, the legs and the tail, resulting in an additional set of 6 deformation vectors, as we group the left and right ears in a single part. The scaling factors in Equation 2 for these parts are 0.05 for the ears, where we scale the whole part, and 0.1 for the legs and tail, where we scale only the vertical axis. Note that the scaling factors in Equation 2 are applied to define the new shape space dimensions, but the scaling is then weighted with corresponding shape variables as for the original PCA dimensions.", + "bbox": [ + 75, + 439, + 472, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model-based Alignment. In order to align the hSMAL+ model to the prototypes, first, we manually annotate land-", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4145ee81a8f22b5a86340cd10f9c31055f3a9ff1f452070e9fba4bd054c4a09a.jpg", + "image_caption": [ + "Figure 7. Model alignment to the scans. Examples of noisy (top) and clean (bottom) registrations. In white is the scan and in purple is the model registration." + ], + "image_footnote": [], + "bbox": [ + 547, + 92, + 836, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "marks on the selected scans: for most of the prototypes we annotate 7 landmarks: the tip of the four hooves, the tip of the ears, and the tip of the tail. Note that this process, while manual, is applied to only one frame per horse, resulting in a very small fraction of the time invested to record a specific subject. The alignment procedure minimizes an energy with data terms that consist of the mesh-to-scan and scan-to-mesh distances between the scan and the model, and the distance between landmarks defined on the model and annotated on the scans. We add regularization terms for the $\\beta$ and the $\\theta$ variables, set as the square of the elements of the $\\beta$ vector, and the $L1$ norm of the $\\theta$ values, respectively. The latter formulation is as a consequence of the fact that in this stage we are aligning scans that are close to the model neutral pose by selection. The pose prior is stronger for the tail, which is often partially observed, and weaker for the ears, which might not be in a neutral pose. We optimize the energy with Chumpy [2], with an annealing schedule that decreases the weight of the pose prior and increases the weight of the data terms. After the model-based alignment, we increase the resolution of the hSMAL+ model (from 3647 to 13873 vertices) and perform a model-free alignment, where, to capture the fine-level details of the scans, we optimize over the vertices. In this case, we use two regularization terms: as-rigid-as-possible (ARAP) [31] regularization, and a coupling term between the 3D vertices and the model solution. Note that when increasing the model resolution, we do not alter the vertices that are inside the head (tongue and palate). Moreover, we keep the low-resolution vertices as the first 3647 vertices of the model, such that VAREN can be easily used at two resolutions.", + "bbox": [ + 496, + 287, + 893, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VAREN Shape Space. The VAREN horse model shares the same formulation of SMPL and hSMAL expressed by Equation 1. The shape space of the model, indicated by the matrix $B$ , is learned in the following way. Once the prototype scans have been aligned, they are all brought into the neutral pose of the alignment model (they are already close to this pose). The mean template shape is computed, giving an averaged prototype horse (Fig. 5), which is subtracted from the prototypes. Then Principal Component Analysis", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5378", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(PCA) is applied, giving a linear shape space of skin deformations (Fig. 6). On the new template, we define the skeleton and muscle skin labels.", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1.1 Anatomical Structures", + "text_level": 1, + "bbox": [ + 76, + 147, + 290, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we describe how we incorporate anatomical structures, namely the skeleton and the muscles, into the VAREN model. Our goal is to define the model skeleton joints in correspondence with real joints, and characterize the model vertices according to the muscle they are closest to, as this association will be exploited in the pose-dependent deformation model. To this end, we exploit a realistic CG model of a horse [1] that includes a skeleton and muscles. We have already exploited the CG model to define the neutral pose for VAREN (Fig. 5), with the further advantage that now the VAREN template has the same pose of the skeleton of the CG model. However, given the body part proportions differ between the GC model and the VAREN template, we manually deform the individual skeleton bones to better match the template. We then select a set of joint locations on the skeleton that are anatomically relevant and closest to the hSMAL+ joints (Fig. 4). Finally, we re-compute the joint regressor of the VAREN model such that the anatomical joints are used. In addition to the existing hSMAL joints, we define two additional joints, corresponding to the left and right scapula. We found this necessary to fit the model to complex poses. A joint in correspondence to the top of the scapula models biomechanical principles for horses and dogs [12]. The muscles in the purchased CG model are grouped by large sections. We separate them into individual muscles, label them, and find the closest muscle for each vertex of the VAREN registration to the CG model. These labels are then transferred to the VAREN template. In this way, we obtain muscle labels for each vertex of the VAREN horse model (Fig. 4). Now that we have a model with anatomical joints and pervertex muscle indices, first we re-register the prototypes to obtain the shape and pose parameters for the new model, then we learn pose-dependent deformations from the dynamic scans. The deformations are learned from an optimization network on a dataset of 3D scans aligned to the horse model. In the following, we describe the generation of the training set and the VAREN network.", + "bbox": [ + 75, + 171, + 472, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Training Dataset", + "text_level": 1, + "bbox": [ + 76, + 756, + 243, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training dataset is composed by a set of scans and corresponding alignment parameters. The scans are frames of a set of clips containing the same prototype horses used to learn the shape model performing different activities: standing still, moving the neck forward and toward a side, and moving the legs (see Fig. 2). As a pre-processing step, the scans are cleaned-up to remove the floor and the horse owner using a simple procedure based on point cloudCLUS", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47c8f21d8b23b3a21e431cca26294012d39977074c7e86d5115b0d74f935d481.jpg", + "image_caption": [ + "Figure 8. Post-alignment cleaning. We exploit the model registration to clean up the 3D scans. Left column: input scan, middle: cleaning with a single bounding box around the registration. Right: cleaning with a set of bounding boxes. The horse head is sometimes not captured due to the presence of a person." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 854, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fering. We register the VAREN model to the dynamic sequences with an alignment procedure similar to the one used for the prototypes. The difference here is that we do not optimize for the horse shape, but keep the shape parameter fixed to the one computed with the VAREN prototype registration. For the first frame, we also initialize with the prototype pose. After the first, each frame is initialized with the previous solution, and the pose prior takes the form of the $L1$ distance from the previous pose. The alignment procedure does not require manual annotations. However, due to the presence of noisy frames, where the pre-processing clean-up failed, sometimes the alignment process gives a wrong result, and we need to annotate a few landmarks to correct such cases. We annotate about the $1\\%$ of the frames with about 4 landmarks. Moreover, we need to remove from the training set those frames where complete parts are missing. While we can deal with large holes in the scans, the alignment may return an incorrect solution if the entire head or a whole limb is missing. On the other hand, we allow ears to be missing in the scans, as the pose-dependent deformation model does not include them (see Fig. 4). Examples of registrations are shown in Fig. 7. The model registration can be used to further clean up and filter the data. We do this with a procedure that creates a set of bounding boxes around the registration to filter out the scan regions that are not pertinent to the horse (Fig. 8). We then compute the maximum scan-to-mesh distance to filter out scans with residual outliers. We retain, in total, about $3.6K$ training scans. Note that not all the prototype scans have been used to create the training set, as we leave out a few for testing.", + "bbox": [ + 496, + 297, + 892, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. VAREN Network", + "text_level": 1, + "bbox": [ + 500, + 762, + 668, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pose-dependent deformations model is an additive term augmenting the intrinsic body shape (see Eq. 1):", + "bbox": [ + 498, + 786, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {s + d} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} + \\mathbf {d v} _ {m} (\\theta , \\beta), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 827, + 890, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{d}\\mathbf{v}_m$ are the muscle deformations, obtained as:", + "bbox": [ + 500, + 857, + 852, + 872 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d v} _ {m, I _ {i}} (\\theta , \\beta) = D _ {i} (\\beta_ {m}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5379", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i$ is the muscle index, $D_{i}$ is a decoder composed by a linear layer of dimension $(4(n_{J} - 1) + 2)\\times 3n_{I_{i}}$ , with $I_{i}$ the set of indices of the skin vertices associated with muscle $i$ . The number of joints $n_J = 38$ , excluding the root, is multiplied by 4 as we use quaternions to represent pose. Equation 4 is applied for each muscle. The muscle deformation variable $\\beta_{m}$ is defined as:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {m} = A \\circ W _ {m} \\left(\\theta_ {2,.. n, j}, \\beta_ {1, 2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 210, + 468, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A$ is a selection matrix of dimension $n_M \\times (4(n_J - 1) + 2)$ , with $n_M = 76$ being the number of muscles. We initialize $A(i, 4k: 4k + 4) = 1$ if muscle $i$ belongs to part $j$ , with $k \\in N(j)$ , otherwise 0. Here $N(j)$ indicates the set composed by the part $j$ and its neighbors. An advantage of our formulation is that each row of the product $A \\circ W_m$ is now a weighting vector for the pose parameters indicating the influence between muscles and body parts, regardless of the pose, while the muscle deformation variable $\\beta_m$ indicates the strength of the muscle deformation. Both $A$ and $W_m$ are optimized during training. The last column of Fig. 9 shows the association between muscle deformation and body parts.", + "bbox": [ + 75, + 239, + 472, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3.1 Network Training", + "text_level": 1, + "bbox": [ + 76, + 448, + 256, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, the network reads the training set, and stores the 3D pose, shape and translation estimated by the alignment procedure. Input scans are resampled at a fixed size of 20000 points to have a uniform size within a batch, with size 4. We store the alignment parameters in memory such that, during network optimization, pose and translation variables can be fine-tuned. However, we found in our experiments that fine-tuning the alignment variables does not significantly change the results, indicating that the training registrations are already of good quality. Here we present results for which we did not fine-tune the training data. We train using different losses: the Chamfer distance (from the PyTorch3D library [28]) between the generated horse mesh and the input scan, a regularization term, implemented as an edge length minimization loss, also from PyTorch3D, applied to the deformations $\\mathbf{d}\\mathbf{v}_m$ and only to the mesh triangles on the muscle boundaries, and a regularization loss on the matrix $A$ , to favor small values for body parts that are likely not to influence the muscles, namely the tail, feet, mouth and ears. This is implemented as the $L_{1}$ norm of the matrix entries for these parts. Weights for the losses are: $\\alpha_{dist} = 1e^{3}$ , $\\alpha_{reg} = 10$ , $\\alpha_{bound} = 100$ , $\\alpha_{A} = 1e^{3}$ .", + "bbox": [ + 75, + 473, + 472, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 820, + 209, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply the VAREN network to a set of held-out alignments. As a baseline, we consider the model without pose-dependent deformations. We also compare with defining pose-dependent deformations as in SMPL and STAR.", + "bbox": [ + 75, + 845, + 468, + 907 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SMPL Pose-dependent Deformations. The SMPL model defines the pose-dependent deformations as a linear model over the pose vector of relative rotation matrices (see Eq. 9 in [22]). The pose feature vector is therefore of size $9(n_{J} - 1)$ , and the pose blend shapes is a matrix of shape $3n_{V} \\times 9(n_{J} - 1)$ . When training to learn the pose blend shapes, we use the Chamfer distance loss, the regularization loss applied to the deformations, and a loss on the $L_{1}$ norm of the pose blend shapes values. Weights for these losses are: $\\alpha_{dist} = 1e^{3}$ , $\\alpha_{reg} = 100$ , $\\alpha_{reg,B_{p}} = 1e^{3}$ .", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "STAR Pose-dependent Deformations. The STAR model defines the pose-dependent deformations as a set of perpart linear models over the part and neighbors rotations, expressed with quaternions. The second value $\\beta_{2}$ of the shape vector is added to the set of pose features, as it correlates with the human BMI index. Here, lacking this information for horses, we condition the deformation with the first two elements of the $\\beta$ vector, as in the VAREN shape space all the components capture morphological information (see Fig. 6). Each of the $n_J - 1$ linear models has dimension $(4(n_{Jj} + 1) + 2)\\times 3n_V$ , where $n_{Jj}$ is the number of neighbors of the part $j$ . STAR defines the rotations relative to the human neutral pose, as the template pose in the STAR model is not neutral. This is not necessary for VAREN, where the template is in a natural rest pose. STAR also defines a set of activation weights, $A$ , for each vertex, to weight the output of each linear model. Here, $A$ is a matrix of shape $(n_J - 1)\\times n_V$ that we initialize with the distance between vertices and joints. Following STAR, $A$ is passed through a nonlinear rectifying function. During training, we use as losses the Chamfer distance and the regularization loss on the deformations. Weights for these losses are: $\\alpha_{dist} = 1e^{3}$ , $\\alpha_{reg} = 100$ . All networks are trained for 100 epochs, with a learning rate of $1e^{-5}$ .", + "bbox": [ + 496, + 250, + 892, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Test Datasets and Results", + "text_level": 1, + "bbox": [ + 498, + 619, + 732, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider two test sets. The In-shape test set, with 275 frames from 5 horses, includes the prototype horses used to train the shape deformation model, but not used for training the VAREN network. The Out-shape test set, 154 frames for 6 horses, includes subjects that have not been used at all. On the first set, we expect to obtain lower errors, and see only the effect of the pose-dependent deformation term, while the second set will provide a general performance assessment of the whole VAREN model. Results are reported in Table 1 and Table 2. We report errors in terms of average Chamfer distance, and of mesh-to-scan distance over a subset of model vertices obtained by removing the head, tail, ears, ankles and hooves. We do this to reduce the influence of the outliers and missing parts on the error scores. VAREN provides the best accuracy in comparison with the baseline and previous work. Figure 9 shows results for the Out-shape dataset. Notice how VAREN generates realistic", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5380", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45f5ef74d5cdacb66c56a893e2bd88c813eac258c53d73cee8d7fa6580c01c6e.jpg", + "image_caption": [ + "Figure 9. Results. From left: scan, baseline (purple), baseline mesh-to-scan distance (two views), VAREN (blue), VAREN mesh-to-scan distance (two views), VAREN with part colors, VAREN with muscle colored with the color of the part that gives the main contribution to the deformation, computed as the argmax over body parts of the absolute value of the muscle deformation variable $\\beta_{m}$ (Eq. 5). For the mesh-to-scan visualization, distances are clipped at $4\\mathrm{cm}$ , then normalized; darker regions indicate higher distance." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 872, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed166a5cbbfdcaf6f6d35c34968b64531520c2957515b9ad62c4169999846891.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
Baseline10.321.038.200.89
SMPL9.720.987.060.72
STAR9.510.676.990.63
VAREN (Our)9.370.746.350.59
", + "bbox": [ + 88, + 530, + 455, + 628 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c99fd5bd236b2ddd343aaaf7ead08c27ae9c80151b60e330e0909c861a87cf56.jpg", + "table_caption": [ + "Table 1. Results on the In-shape testset. Baseline is the model without pose-dependent deformations. Errors in mm." + ], + "table_footnote": [], + "table_body": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
hSMAL21.923.0217.012.51
Baseline12.001.9610.072.55
SMPL11.381.668.842.01
STAR11.031.578.361.77
VAREN (Our)10.881.597.781.71
", + "bbox": [ + 86, + 657, + 454, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Results on the Out-shape testset. hSMAL uses the hS- MAL model for alignment. Baseline is the model without posedependent deformations. Errors in mm. Wilcoxon significance test for STAR and VAREN comparison: p-values are 0.04 and 3.3e-06 for Chamfer and mesh-to-scan distance, respectively.", + "bbox": [ + 75, + 773, + 468, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "deformations, in particular for the neck region. We also illustrate, for each muscle, the body part that most contributes to its deformation.", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 531, + 619, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced VAREN, the first parametric 3D model of horses learned from real data. In contrast to previous work, VAREN exploits a novel formulation that captures pose-dependent muscle deformations, resulting in better accuracy compared to state-of-the-art approaches, while also being more compact and connecting 3D pose to muscle deformations. By focusing on quality and anatomical realism, VAREN can support a wide set of AI applications in the equestrian world.", + "bbox": [ + 496, + 551, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Silvia Zuffi is supported by PNRR project FAIR - Future AI Research (PE00000013), Spoke 8 - Pervasive AI (CUP H97G22000210007) under the NRRP MUR program by NextGenerationEU. We thank Minsoek Kim, Peter Kulits, Gökce Ergünfor, and Emil Breustedtfor for helping with pre-processing the scans. We also thank Dr. Zala Zgank, for assisting with data collection, and the horse owners for their time and for bringing their horses. Disclosure. MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon. MJB has financial interests in Amazon and Meshcapade GmbH. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5381", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.3dscanstore.com/3d-body-models/ecorche-3d-models/horse-ecorche-3d-model.4,5,6", + "[2] http://chumpy.org.5", + "[3] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5461-5470, 2021. 3", + "[4] Brett Allen, Brian Curless, and Zoran Popovic. The space of human body shapes: Reconstruction and parameterization from range scans. ACM Trans. Graph., 22(3):587-594, 2003. 3", + "[5] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a correlated model of identity and pose-dependent body shape variation for real-time synthesis. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 147-156, Airela-Ville, Switzerland, Switzerland, 2006. Eurographics Association.", + "[6] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE: Shape Completion and Animation of PEople. ACM Trans. Graph. (Proc. SIGGRAPH, 24(3):408-416, 2005. 3", + "[7] Marc Badger, Yufu Wang, Adarsh Modh, Ammon Perkes, Nikos Kolotouros, Bernd G Pfrommer, Marc F Schmidt, and Kostas Daniilidis. 3d bird reconstruction: a dataset, model, and shape recovery from a single view. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-17, 2020. 4", + "[8] Benjamin Biggs, Oliver Boyne, James Charles, Andrew Fitzgibbon, and Roberto Cipolla. Who left the dogs out? 3D animal reconstruction with expectation maximization in the loop. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 4, 5", + "[9] T. J. Cashman and A. W. Fitzgibbon. What shape are dolphins? building 3D morphable models from 2d images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):232-244, 2013. 3", + "[10] Yinpeng Chen, Zicheng Liu, and Zhengyou Zhang. Tensor-based human body modeling. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-112, 2013. 3", + "[11] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22091-22101, 2023. 3", + "[12] M.S. Fischer, K.E. Lilje, J. Laustroer, and A. Andikfar. Dogs in Motion. Kosmos, 2011. 6", + "[13] N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.P. Seidel. A statistical model of human pose and body shape. Computer Graphics Forum, 28(2):337-346, 2009. 3", + "[14] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3D: Learning articulated 3D animals by distilling 2d diffusion. In International Conference on 3D Vision (3DV), 2024. 3" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Angjoo Kanazawa, Shahar Kovalsky, Ronen Basri, and David Jacobs. Learning 3D deformation of animals from 2d images. Comput. Graph. Forum, 35(2):365-374, 2016. 4", + "[16] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 4", + "[17] Marilyn Keller, Silvia Zuffi, Michael J. Black, and Sergi Pujades. OSSO: Obtaining skeletal shape from outside. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 20492-20501, 2022. 3", + "[18] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, Liu C. Karen, and Michael J. Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM ToG, Proc. SIGGRAPH Asia, 2023. 3", + "[19] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3D reconstruction of articulated categories from motion. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1737-1746, 2021. 4", + "[20] Ci Li, Nima Ghorbani, Sofia Broome, Maheen Rashid, Michael J. Black, Elin Hernlund, Hedvig Kjellström, and Silvia Zuffi. hsmal: Detailed horse shape and pose reconstruction for motion pattern recognition. In First CV4Animals Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4", + "[21] Sandro Lombardi, Bangbang Yang, Tianxing Fan, Hujun Bao, Guofeng Zhang, Marc Pollefeys, and Zhaopeng Cui. Latenthuman: Shape-and-pose disentangled latent representation for human bodies. In International Conference on 3D Vision (3DV), 2021. 3", + "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 7", + "[23] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[24] Eadweard Muybridge. Animal Locomotion. An Electro-Photographic Investigation of Consecutive Phases of Animal Movements. Philadelphia: University of Pennsylvania Press, 1887. 1", + "[25] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Proceedings of the European Conference on Computer Vision (ECCV), pages 598-613, 2020. 2, 3", + "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[27] Deva Ramanan, David A Forsyth, and Kobus Barnard. Building models of animals from video. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(8):1319-1334, 2006. 3", + "[28] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5382", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 7", + "[29] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876-3884, 2022. 4", + "[30] Nadine Ruegg, Shashank Tripathi, Konrad Schindler, Michael J. Black, and Silvia Zuffi. BITE: Beyond priors for improved three-D dog pose estimation. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8867-8876, 2023. 4", + "[31] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Proceedings of the Fifth Eurographics Symposium on Geometry Processing, Barcelona, Spain, July 4-6, 2007, pages 109-116, 2007. 5", + "[32] Yufu Wang, Nikos Kolotouros, Kostas Daniilidis, and Marc Badger. Birds of a feather: Capturing avian shape models from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 14739-14749, 2021. 3", + "[33] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3D objects by watching videos. International Journal of Computer Vision (IJCV), 2023. 4", + "[34] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3D animals in the wild. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4", + "[35] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[36] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 4", + "[37] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3D shape reconstruction. In Neural Information Processing Systems (NeurIPS), 2021. 4", + "[38] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. BANMo: Building animatable 3D neural models from many casual videos. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2853-2863, 2022. 4", + "[39] Gengshan Yang, Chaoyang Wang, N Dinesh Reddy, and Deva Ramanan. Reconstructing animatable categories from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16995-17005, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3914-3924, 2023. 4", + "[41] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shape from sparse image ensemble via 3D part discovery. In Neural Information Processing Systems (NeurIPS), 2022. 4", + "[42] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4", + "[43] Chun-Han Yao, Amit Raj, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Artic3d: Learning robust articulated 3D shapes from noisy web image collections. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4", + "[44] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5524-5532, 2017. 3", + "[45] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3955-3963, 2018. 4" + ], + "bbox": [ + 501, + 92, + 892, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5383", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_model.json b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0f480e7e9af803c991fa67d6820816eb3c295466 --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_model.json @@ -0,0 +1,1661 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.211, + 0.131, + 0.761, + 0.154 + ], + "angle": 0, + "content": "VAREN: Very Accurate and Realistic Equine Network" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.193, + 0.871, + 0.229 + ], + "angle": 0, + "content": "Silvia Zuffi1, Ylva Mellbin2, Ci Li3, Markus Hoeschle4, Hedvig Kjellström3,2, Senya Polikovsky4, Elin Hernlund2, and Michael J. Black4" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.241, + 0.808, + 0.278 + ], + "angle": 0, + "content": "1IMATI-CNR, Milan, Italy, 2SLU, Uppsala, Sweden, 3KTH, Stockholm, Sweden, 4Max Planck Institute for Intelligent Systems, Tübingen, Germany" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.299, + 0.816, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.481, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Figure 1. VAREN is learned from real horses and models pose-dependent deformations at the muscle level, enabling the generation of 3D horses with high realism. Inspired by Muybridge [24], we show a sequence of frames of horses in motion, generated by VAREN, where the first frame illustrates, for each row, the horse in the rest pose. We show three different shapes, obtained by sampling the VAREN model." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.559, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Data-driven three-dimensional parametric shape models of the human body have gained enormous popularity both for the analysis of visual data and for the generation of synthetic humans. Following a similar approach for animals does not scale to the multitude of existing animal species, not to mention the difficulty of accessing subjects to scan in 3D. However, we argue that for domestic species of great importance, like the horse, it is a highly valuable investment to put effort into gathering a large dataset of real 3D scans, and learn a realistic 3D articulated shape model. We introduce VAREN, a novel 3D articulated parametric shape model learned from 3D scans of many real horses. VAREN bridges synthesis and analysis tasks, as the generated model instances have unprecedented realism, while being able to represent horses of different sizes and shapes. Differently from previous body models, VAREN has two resolutions, an anatomical skeleton, and interpretable, learned pose-dependent deformations, which are related to the body muscles. We show with experiments that this formulation has superior performance with respect to previous strategies for modeling pose-dependent deformations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.56, + 0.895, + 0.637 + ], + "angle": 0, + "content": "mations in the human body case, while also being more compact and allowing an analysis of the relationship between articulation and muscle deformation during articulated motion. The VAREN model and data are available at https://varen.is.tue.mpg.de." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Horses are arguably the most valuable domestic animal and there is a large industry focused on their breeding, care, training, and use in sports. Buying a horse is a large investment, and keeping a horse requires resources and time. Although they are large animals, horses are delicate. The accumulated loads from training and competition frequently result in unrecoverable injuries in their skeletal or tendinous limb structures, ultimately leading to euthanasia. Consequently, among domestic animals, horses are widely studied from a behavioral and biomechanical perspective, with the aim of evaluating their performance, interpreting their state of pain, and preventing injuries. To facilitate such analysis using computer vision, our goal is to develop a highly accurate and detailed 3D model of horses that can be articulated, fit to data, and animated. Here we learn such a model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.451, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Dynamic 3D horse scanner. The horse is asked to lift one hind limb." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.468, + 0.431 + ], + "angle": 0, + "content": "(see Fig. 1) using a novel formulation that captures pose-dependent muscle deformations. Such a model may facilitate 3D motion analysis in-the-wild, body shape estimation, diagnosis of illness and performance, horse-human interaction and behavior analysis, and generation of 3D synthetic horses for VFX, VR, AR, and gaming. In particular, accurate markerless motion capture of horses would provide a valuable new tool for the study of horse motion during challenging activities in natural environments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.734 + ], + "angle": 0, + "content": "For humans, there has been great progress on capturing their 3D shape and pose from images and video. This work often exploits the SMPL body model [22], a 3D mesh-based representation of the human body, controlled by 3D shape and pose parameters. SMPL provides strong priors over human shape and pose, enabling 3D pose reconstruction from ambiguous 2D data. Creating such a 3D model of horses is more challenging. While 3D articulated shape models of horses have been created in the past, they lack realism and expressiveness. In contrast, models like SMPL are learned from thousands of scans, and include an expressive shape space and a pose-dependent deformation model to represent how the body deforms under articulation. Unfortunately, obtaining high-quality 3D scans of horses in a variety of poses is challenging. To fill the gap, we need to (1) capture 3D scans of horses in motion and (2) formulate a new parametric 3D model of horses that represents important aspects of their physiology. In both cases, our solutions deviate from prior work on humans to accommodate the unique aspects of horses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Dataset. We use a novel setup (3dMD Ltd.) to capture dynamic 3D scans of horses over time (see Fig. 2). Scanning a large number of horses presents technical and practical challenges, requiring a large cooperative effort between computer scientists and animal care experts. Our dataset includes 50 horses of different breeds and sizes, ranging from small ponies to large horses, with a height difference of more than 1.5 meters. This shape variation is significantly larger than with humans. To model pose-dependent shape changes, we need to capture the horses in a range of poses. Additionally, the animal is always managed by a care per" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.089, + 0.843, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.251, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Samples from the VAREN model result in a diversity of horse shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.429 + ], + "angle": 0, + "content": "son who may occlude the horse during scanning, resulting in missing data. All these issues present novel challenges that we overcome. As a result, our dataset includes in total approximately 4000 cleaned up raw 3D scans, for which we provide accurately registered 3D meshes – training and testing split. Written consent was provided by all animal owners before the animals entered the study. The data collection procedure did not include any invasive technique, therefore no animal ethical approval was needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Model. Using this unique dataset, we train a new model named VAREN (pronounced Varenne1 in French). VAREN is the first 3D articulated model of horses that is learned from real data and that models pose-dependent muscle deformations. Samples from the VAREN model are shown in a neutral pose in Fig. 3 and animated using motion capture data in Fig. 1. To model horses in motion, we must also capture the non-rigid deformations of the body that occur during movement. Previous work on humans models pose-dependent deformations that are learned from 3D scans and conditioned on the body articulation. In particular, STAR [25] models such deformations locally based on the distance from the nearby joints. Such methods do not explicitly model the deformation of the muscles. Scans of humans are typically acquired with some form of clothing (not fully naked) and most humans have a layer of subcutaneous adipose tissue that hides the musculature. Horses, however, are different. They generally carry less body fat than humans and their hair is short, making their musculature more directly observable. Modeling the visible muscle deformation is important because it relates directly to the health and performance of the animal. To that end, we propose a novel learned muscle-based deformation model. Specifically, we group the body surface points into regions that correspond to the superficial muscles, rather than based on their distance to the joints. Note that a single muscle can span more than one body part (see for example the back muscle in Fig. 4) and it is important to capture these long-range cor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1Varenne is considered to be the best trotter of all time. No other trotter has won so many of the most important races in the world and set as many records as Varenne. Source: Wikipedia." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5375" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.465, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.171, + 0.468, + 0.226 + ], + "angle": 0, + "content": "Figure 4. Skeleton and muscles. From left: anatomical skeleton manually aligned to the model template, selected joints, muscle structure and skin labels on the template. We consider 76 muscles. Head, tail and hooves have no muscle definitions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.241, + 0.468, + 0.679 + ], + "angle": 0, + "content": "relations. In VAREN, we define deformations per-muscle and learn the influence of body part articulation on the observable muscle deformation from the scan data. We find that, in comparison with previous formulations like those used in SMPL and STAR, our proposed muscle-inspired pose-dependent deformation model has superior accuracy, while also being more compact and interpretable. This is relevant for our high-resolution horse model. While we do not measure muscle activation directly, our formulation can be relevant for research on body pose using measurements of muscle activation [11]. A further novelty of our model, compared with previous work, is that we define the model skeleton, that is the location of the joints in the kinematic tree, based on real horse anatomy. Exploiting a 3D model of the horse skeleton, and with the help of an expert on horse biomechanics, we define the articulation points for the model. Previous human body models define the joint locations as a result of an optimization procedure, resulting in a per-model definition with poor correspondence to the true anatomical skeleton, in particular at the shoulders and hip joints [17]. These models are motivated by applications in computer graphics and vision, where, in many cases, a precise anatomical skeleton may not be needed. For a model like VAREN to be valuable to horse breeders, trainers, and veterinarians, it should provide information that is anatomically relevant. Moreover, an anatomical definition of the model joints, being model-independent, could facilitate comparison and motion data transfer between different models, and, eventually, species." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.468, + 0.814 + ], + "angle": 0, + "content": "In summary, we make two key contributions. First, we provide a new dataset for the modeling of horse shape and pose. The quantity, resolution, variation, and quality of the VAREN data dwarfs any previous dataset of animal shape and pose, and opens up the possibility for the community to explore new questions in 3D shape representation. Second, we develop a new shape model that drives the observable deformation of the muscles based on the animal pose using an anatomical skeleton." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.219, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We review prior work on learning human body models from data, as our problem and setting share similarities with methods proposed in this case. Then, we present the few" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "3D models for animals previously proposed. In addition, we provide an overview of methods for 3D animal reconstruction from visual data, in order to set the context in which we see our novel model applied." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Human Body Models. There is a long history of learning 3D mesh-based models of the human body from 3D scans [4-6, 10, 13, 22, 26, 35]. There is also recent work on implicit representations that we do not consider here, e.g. [3, 21, 23]. Most recent work is based on SMPL [22], a 3D body model that is based on vertex deformations defined by linear spaces. SMPL is fully data-driven: pose dependent deformations are defined as global representations and learned from data. One issue with SMPL is that the pose-dependent deformations are not local and capture spurious long-range correlations in the data. This is addressed by STAR [25], which adds locality to the joints to the pose-dependent deformations, weighting the outcome of global linear spaces with the distance from skeleton joints. While STAR conditions the pose deformation on the rough BMI of the body, none of the models above explicitly model the observable muscle deformation. VAREN addresses this for the first time. Also, unlike VAREN, the prior methods model articulation using a kinematic tree only loosely based on the human skeleton. A first direction into modeling the human skeleton is OSSO [17], who models the human body skeleton, followed by SKEL [18], which drives the pose of SMPL with a biomechanical skeletal model. SKEL, however, uses the default SMPL model deformations and, unlike VAREN does not train deformations that are driven by the skeletal pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Animal Body Models. The modeling of animal shape for computer vision applications has received less attention. Since 3D scanning of animals is challenging, research has mainly focused on learning from images; e.g. Ramanan et al. learn 2D articulated models [27], while Cashman and Fitzgibbon [9] adapt a rough mean shape using 2D image cues to reconstruct dolphins and bears in 3D. Zuffi et al. learn SMAL, a parametric 3D animal shape model, using 3D scans of toys [44]. Wang et al. learn the 3D shape variation of birds from images starting from a synthetic model [32]. Li et al. learn a horse-specific model similar to SMAL, using 3D scans of horse figurines [20]. Still, others learn to represent category-specific shapes as either meshes [39] or implicit surfaces [14], but do not decouple shape and pose. None of these methods learn 3D articulated models from real 3D scans, which contain noise. The multi-species SMAL model [44] contains an Equine class, but SMAL is learned from a limited number of toys, thus horses generated with the SMAL model do not exhibit a wide range of diverse shapes. The more recent hSMAL model is horse-specific [20]. While learned from a larger number of horses of different shapes, hSMAL is still learned from toys, and has a limited resolution. While useful for reconstructing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5376" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "horses from images and video, we find in our experiments that hSMAL is not expressive enough to accurately represent real horses with high detail. Being learned from rigid toys, both models do not address body deformations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.163, + 0.473, + 0.571 + ], + "angle": 0, + "content": "Animal 3D Reconstruction. The 3D reconstruction of animals from images and video follows a model-based or a model free approach. In the first case, an existing 3D shape model of the species of interest is available, and the method outputs the model parameters of 3D shape and pose. This approach is useful when dealing with challenging input modalities, like in-the-wild monocular images, or the goal is to estimate animal shape and pose for downstream analysis tasks. For example, Zuffi et al. address multi-species 3D reconstruction [45], Kanazawa et al. learn 3D deformations of animals [15], Badger et al. reconstruct 3D birds [7], while several methods estimate the shape and pose of dogs [8, 29, 30]. Our work fits in this category, but provides a 3D model with a higher level of realism than any previous methods. Model-free methods, in contrast, do not assume that a 3D model of the animal class is available, and the output of these methods is typically a 3D surface, in case of static input, or, eventually, a 3D animatable object in case of a video input. Works in this class include CMR [16] and DOVE [33], which reconstruct birds from images and ViSER [37], LASR [36], BANMo [38] and PPR [40], which reconstruct articulated 3D shapes from video, the latter incorporating physical constraints. Kokkinos and Kokkinos learn 3D reconstruction from video without using a shape model by relying only on a template [19]. MagicPony [34], LASSIE [41], Hi-LASSIE [42], ARTIC3D [43] learn 3D reconstruction from image collections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.17, + 0.603 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We learn the VAREN model in two stages. First, we learn an articulated parametric shape model from a set of scans, that we call prototypes, which are of different horses in a neutral pose, that is, not undergoing pose-dependent deformations. This gives us a model with a shape space and generic articulation. We then align, in an unsupervised way, this static model to the dynamic 3D scans in which the prototype horses perform different movements. From this data, namely the scans and the alignment parameters of 3D shape, pose and translation, the VAREN network learns a pose-dependent deformation model that improves the matching between the learned model and the scans. The novel deformation model is a function of the shape and pose parameters, such that, at test time, the network generates a shaped and posed horse, with realistic pose-dependent deformations, given the parameters. The pose-dependent deformation model is defined on the basis of anatomical structures that we incorporate into VAREN from a purchased realistic graphics (CG) model [1], which includes the body skeleton" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.092, + 0.85, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.157, + 0.894, + 0.213 + ], + "angle": 0, + "content": "Figure 5. Comparison between the average equine model in SMAL (left), the hSMAL+ template (middle) and the VAREN template (right). Learned from real horses, the VAREN template has a more proportioned neck with respect to hSMAL+." + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.215, + 0.858, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.35, + 0.893, + 0.42 + ], + "angle": 0, + "content": "Figure 6. Visualization of the VAREN shape space. We capture small ponies, large breeds, and many in between, obtaining a shape space with variation in size, morphology and face features. The tail and mane are tied, and not considered as part of body shape change." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.894, + 0.492 + ], + "angle": 0, + "content": "and the muscles. In the following, we detail the procedure to learn the horse model from the prototypes, the creation of the training dataset, and the definition of the VAREN network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.502, + 0.704, + 0.517 + ], + "angle": 0, + "content": "3.1. VAREN Horse Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.895, + 0.738 + ], + "angle": 0, + "content": "Alignment of the Prototype Scans. We follow a model-based approach to express the prototypes in a common topology, specifically, as the alignment model, we use hSMAL [20], a recently proposed articulated horse model (Fig. 5). The hSMAL model follows the formulation of SMPL, and is defined by a triangular mesh template \\(\\mathbf{v}_t\\), with \\(n_V\\) vertices, a matrix \\(B\\) of shape \\(3n_V \\times n_B\\) containing the \\(n_B\\) basis vectors of a linear shape deformation space, a joint regressor \\(J_r\\) that maps model vertices to a set of \\(n_J\\) joint locations, and a skinning weight matrix \\(W\\). A horse is generated, given shape parameters \\(\\beta\\) and pose parameters \\(\\theta\\), by first deforming the template into an intrinsic shape \\(\\mathbf{v}_s\\), then applying Linear Blend Skinning (LBS) to rotate the body parts according to the given pose:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.747, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {v} _ {s} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} \\\\ \\mathbf {v} = L B S \\left(\\mathbf {v} _ {s}, \\theta ; W, J _ {r}\\right). \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "While able to represent horses of different shapes, we found that the hSMAL model has, in our case, two limitations: a low resolution (the model has about \\(1.5K\\) vertices) and a shape space that is not sufficiently expressive to accurately capture the shape of real horses. To overcome these limitations, we create an hSMAL model with an increased resolution (3647 vertices) and additional shape deformations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "We call this model hSMAL+. Furthermore, we observed that the purchased CG model [1], in addition to containing anatomical structures, has a more natural posture than hSMAL+. Therefore, we modify the rest pose of hSMAL+ to match the pose of the CG model. This is done by fitting hSMAL+ to the CG model and using the obtained mesh as the new template. In hSMAL+ we retain skinning weights and part segmentations from the original hSMAL model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.473, + 0.388 + ], + "angle": 0, + "content": "Additional Shape Deformations. We augment the hSML shape space by defining additional dimensions in the shape deformation matrix \\( B \\) (Eq. 1) to obtain a linear scaling of a set of body parts. Note that limb scaling has been applied before to animal models, but our formulation is novel, as previous work [8] applies scaling during the LBS process, while here we define scaling in the model shape space. Let \\( j \\) be a body part for which we want to add scaling as an additional dimension in the shape space. We define a new shape matrix \\( B_{1} \\) of dimension \\( 3n_V \\times (n_B + 1) \\) by adding a new column to the shape matrix \\( B \\):" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.396, + 0.324, + 0.413 + ], + "angle": 0, + "content": "\\[\nB _ {1} = B | B _ {j}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.415, + 0.47, + 0.433 + ], + "angle": 0, + "content": "\\[\nB _ {j} = s _ {j} \\mathbf {v} _ {j, c}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.473, + 0.862 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{j,c}\\) is a column vector corresponding to the template, but with non-zero values only for the coordinate \\(c\\) of the vertices of the part \\(j\\) we want to scale, and \\(s_j\\) is a scaling factor. Here \\(j\\) can also indicate a set of parts, for example, all the segments of the tail. Expanding \\(B\\) is not sufficient, as the deformation we obtain is expressed for the model template \\(\\mathbf{v}_t\\), while we need it to be applied to the part vertices of the intrinsic shape \\(\\mathbf{v}_s\\) (Eq. 1). Therefore, we add an iterative procedure that estimates a vertex shift \\(d_v\\) for the part vertices. Let \\(J\\) be the joints of the extended model with shape space \\(B_1\\), \\(J_0\\) be the joints of the original model, and \\(J_s\\) a vector of cumulative joint shifts initialized with zero values. In an iterative process that considers all the body joints from the root to the leaves of the kinematic tree, for each body part \\(j\\), we compute the vertex shift as \\(d_v(j) = J_s(j) + J_0(j) - J(j)\\), add it to the part vertices, and then add \\(d_v(j)\\) to the children of the part \\(j\\) in \\(J_s\\), such that the joint shift is propagated through the kinematic tree. The set of body parts we consider for limb scaling are the ears, the legs and the tail, resulting in an additional set of 6 deformation vectors, as we group the left and right ears in a single part. The scaling factors in Equation 2 for these parts are 0.05 for the ears, where we scale the whole part, and 0.1 for the legs and tail, where we scale only the vertical axis. Note that the scaling factors in Equation 2 are applied to define the new shape space dimensions, but the scaling is then weighted with corresponding shape variables as for the original PCA dimensions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Model-based Alignment. In order to align the hSMAL+ model to the prototypes, first, we manually annotate land-" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.093, + 0.837, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.234, + 0.895, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Model alignment to the scans. Examples of noisy (top) and clean (bottom) registrations. In white is the scan and in purple is the model registration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.756 + ], + "angle": 0, + "content": "marks on the selected scans: for most of the prototypes we annotate 7 landmarks: the tip of the four hooves, the tip of the ears, and the tip of the tail. Note that this process, while manual, is applied to only one frame per horse, resulting in a very small fraction of the time invested to record a specific subject. The alignment procedure minimizes an energy with data terms that consist of the mesh-to-scan and scan-to-mesh distances between the scan and the model, and the distance between landmarks defined on the model and annotated on the scans. We add regularization terms for the \\(\\beta\\) and the \\(\\theta\\) variables, set as the square of the elements of the \\(\\beta\\) vector, and the \\(L1\\) norm of the \\(\\theta\\) values, respectively. The latter formulation is as a consequence of the fact that in this stage we are aligning scans that are close to the model neutral pose by selection. The pose prior is stronger for the tail, which is often partially observed, and weaker for the ears, which might not be in a neutral pose. We optimize the energy with Chumpy [2], with an annealing schedule that decreases the weight of the pose prior and increases the weight of the data terms. After the model-based alignment, we increase the resolution of the hSMAL+ model (from 3647 to 13873 vertices) and perform a model-free alignment, where, to capture the fine-level details of the scans, we optimize over the vertices. In this case, we use two regularization terms: as-rigid-as-possible (ARAP) [31] regularization, and a coupling term between the 3D vertices and the model solution. Note that when increasing the model resolution, we do not alter the vertices that are inside the head (tongue and palate). Moreover, we keep the low-resolution vertices as the first 3647 vertices of the model, such that VAREN can be easily used at two resolutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "VAREN Shape Space. The VAREN horse model shares the same formulation of SMPL and hSMAL expressed by Equation 1. The shape space of the model, indicated by the matrix \\( B \\), is learned in the following way. Once the prototype scans have been aligned, they are all brought into the neutral pose of the alignment model (they are already close to this pose). The mean template shape is computed, giving an averaged prototype horse (Fig. 5), which is subtracted from the prototypes. Then Principal Component Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5378" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "(PCA) is applied, giving a linear shape space of skin deformations (Fig. 6). On the new template, we define the skeleton and muscle skin labels." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.148, + 0.291, + 0.162 + ], + "angle": 0, + "content": "3.1.1 Anatomical Structures" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.172, + 0.473, + 0.747 + ], + "angle": 0, + "content": "In this section we describe how we incorporate anatomical structures, namely the skeleton and the muscles, into the VAREN model. Our goal is to define the model skeleton joints in correspondence with real joints, and characterize the model vertices according to the muscle they are closest to, as this association will be exploited in the pose-dependent deformation model. To this end, we exploit a realistic CG model of a horse [1] that includes a skeleton and muscles. We have already exploited the CG model to define the neutral pose for VAREN (Fig. 5), with the further advantage that now the VAREN template has the same pose of the skeleton of the CG model. However, given the body part proportions differ between the GC model and the VAREN template, we manually deform the individual skeleton bones to better match the template. We then select a set of joint locations on the skeleton that are anatomically relevant and closest to the hSMAL+ joints (Fig. 4). Finally, we re-compute the joint regressor of the VAREN model such that the anatomical joints are used. In addition to the existing hSMAL joints, we define two additional joints, corresponding to the left and right scapula. We found this necessary to fit the model to complex poses. A joint in correspondence to the top of the scapula models biomechanical principles for horses and dogs [12]. The muscles in the purchased CG model are grouped by large sections. We separate them into individual muscles, label them, and find the closest muscle for each vertex of the VAREN registration to the CG model. These labels are then transferred to the VAREN template. In this way, we obtain muscle labels for each vertex of the VAREN horse model (Fig. 4). Now that we have a model with anatomical joints and pervertex muscle indices, first we re-register the prototypes to obtain the shape and pose parameters for the new model, then we learn pose-dependent deformations from the dynamic scans. The deformations are learned from an optimization network on a dataset of 3D scans aligned to the horse model. In the following, we describe the generation of the training set and the VAREN network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.245, + 0.773 + ], + "angle": 0, + "content": "3.2. Training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The training dataset is composed by a set of scans and corresponding alignment parameters. The scans are frames of a set of clips containing the same prototype horses used to learn the shape model performing different activities: standing still, moving the neck forward and toward a side, and moving the legs (see Fig. 2). As a pre-processing step, the scans are cleaned-up to remove the floor and the horse owner using a simple procedure based on point cloudCLUS" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.855, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Figure 8. Post-alignment cleaning. We exploit the model registration to clean up the 3D scans. Left column: input scan, middle: cleaning with a single bounding box around the registration. Right: cleaning with a set of bounding boxes. The horse head is sometimes not captured due to the presence of a person." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.893, + 0.754 + ], + "angle": 0, + "content": "fering. We register the VAREN model to the dynamic sequences with an alignment procedure similar to the one used for the prototypes. The difference here is that we do not optimize for the horse shape, but keep the shape parameter fixed to the one computed with the VAREN prototype registration. For the first frame, we also initialize with the prototype pose. After the first, each frame is initialized with the previous solution, and the pose prior takes the form of the \\( L1 \\) distance from the previous pose. The alignment procedure does not require manual annotations. However, due to the presence of noisy frames, where the pre-processing clean-up failed, sometimes the alignment process gives a wrong result, and we need to annotate a few landmarks to correct such cases. We annotate about the \\( 1\\% \\) of the frames with about 4 landmarks. Moreover, we need to remove from the training set those frames where complete parts are missing. While we can deal with large holes in the scans, the alignment may return an incorrect solution if the entire head or a whole limb is missing. On the other hand, we allow ears to be missing in the scans, as the pose-dependent deformation model does not include them (see Fig. 4). Examples of registrations are shown in Fig. 7. The model registration can be used to further clean up and filter the data. We do this with a procedure that creates a set of bounding boxes around the registration to filter out the scan regions that are not pertinent to the horse (Fig. 8). We then compute the maximum scan-to-mesh distance to filter out scans with residual outliers. We retain, in total, about \\( 3.6K \\) training scans. Note that not all the prototype scans have been used to create the training set, as we leave out a few for testing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.669, + 0.778 + ], + "angle": 0, + "content": "3.3. VAREN Network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.787, + 0.892, + 0.817 + ], + "angle": 0, + "content": "The pose-dependent deformations model is an additive term augmenting the intrinsic body shape (see Eq. 1):" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.828, + 0.891, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {s + d} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} + \\mathbf {d v} _ {m} (\\theta , \\beta), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.858, + 0.854, + 0.873 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}\\mathbf{v}_m\\) are the muscle deformations, obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d v} _ {m, I _ {i}} (\\theta , \\beta) = D _ {i} (\\beta_ {m}), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5379" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "where \\(i\\) is the muscle index, \\(D_{i}\\) is a decoder composed by a linear layer of dimension \\((4(n_{J} - 1) + 2)\\times 3n_{I_{i}}\\), with \\(I_{i}\\) the set of indices of the skin vertices associated with muscle \\(i\\). The number of joints \\(n_J = 38\\), excluding the root, is multiplied by 4 as we use quaternions to represent pose. Equation 4 is applied for each muscle. The muscle deformation variable \\(\\beta_{m}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.211, + 0.47, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\beta_ {m} = A \\circ W _ {m} \\left(\\theta_ {2,.. n, j}, \\beta_ {1, 2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.473, + 0.438 + ], + "angle": 0, + "content": "where \\( A \\) is a selection matrix of dimension \\( n_M \\times (4(n_J - 1) + 2) \\), with \\( n_M = 76 \\) being the number of muscles. We initialize \\( A(i, 4k: 4k + 4) = 1 \\) if muscle \\( i \\) belongs to part \\( j \\), with \\( k \\in N(j) \\), otherwise 0. Here \\( N(j) \\) indicates the set composed by the part \\( j \\) and its neighbors. An advantage of our formulation is that each row of the product \\( A \\circ W_m \\) is now a weighting vector for the pose parameters indicating the influence between muscles and body parts, regardless of the pose, while the muscle deformation variable \\( \\beta_m \\) indicates the strength of the muscle deformation. Both \\( A \\) and \\( W_m \\) are optimized during training. The last column of Fig. 9 shows the association between muscle deformation and body parts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.258, + 0.464 + ], + "angle": 0, + "content": "3.3.1 Network Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.473, + 0.807 + ], + "angle": 0, + "content": "During training, the network reads the training set, and stores the 3D pose, shape and translation estimated by the alignment procedure. Input scans are resampled at a fixed size of 20000 points to have a uniform size within a batch, with size 4. We store the alignment parameters in memory such that, during network optimization, pose and translation variables can be fine-tuned. However, we found in our experiments that fine-tuning the alignment variables does not significantly change the results, indicating that the training registrations are already of good quality. Here we present results for which we did not fine-tune the training data. We train using different losses: the Chamfer distance (from the PyTorch3D library [28]) between the generated horse mesh and the input scan, a regularization term, implemented as an edge length minimization loss, also from PyTorch3D, applied to the deformations \\(\\mathbf{d}\\mathbf{v}_m\\) and only to the mesh triangles on the muscle boundaries, and a regularization loss on the matrix \\(A\\), to favor small values for body parts that are likely not to influence the muscles, namely the tail, feet, mouth and ears. This is implemented as the \\(L_{1}\\) norm of the matrix entries for these parts. Weights for the losses are: \\(\\alpha_{dist} = 1e^{3}\\), \\(\\alpha_{reg} = 10\\), \\(\\alpha_{bound} = 100\\), \\(\\alpha_{A} = 1e^{3}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.21, + 0.838 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.847, + 0.47, + 0.908 + ], + "angle": 0, + "content": "We apply the VAREN network to a set of held-out alignments. As a baseline, we consider the model without pose-dependent deformations. We also compare with defining pose-dependent deformations as in SMPL and STAR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "SMPL Pose-dependent Deformations. The SMPL model defines the pose-dependent deformations as a linear model over the pose vector of relative rotation matrices (see Eq. 9 in [22]). The pose feature vector is therefore of size \\( 9(n_{J} - 1) \\), and the pose blend shapes is a matrix of shape \\( 3n_{V} \\times 9(n_{J} - 1) \\). When training to learn the pose blend shapes, we use the Chamfer distance loss, the regularization loss applied to the deformations, and a loss on the \\( L_{1} \\) norm of the pose blend shapes values. Weights for these losses are: \\( \\alpha_{dist} = 1e^{3} \\), \\( \\alpha_{reg} = 100 \\), \\( \\alpha_{reg,B_{p}} = 1e^{3} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.893, + 0.614 + ], + "angle": 0, + "content": "STAR Pose-dependent Deformations. The STAR model defines the pose-dependent deformations as a set of perpart linear models over the part and neighbors rotations, expressed with quaternions. The second value \\(\\beta_{2}\\) of the shape vector is added to the set of pose features, as it correlates with the human BMI index. Here, lacking this information for horses, we condition the deformation with the first two elements of the \\(\\beta\\) vector, as in the VAREN shape space all the components capture morphological information (see Fig. 6). Each of the \\(n_J - 1\\) linear models has dimension \\((4(n_{Jj} + 1) + 2)\\times 3n_V\\), where \\(n_{Jj}\\) is the number of neighbors of the part \\(j\\). STAR defines the rotations relative to the human neutral pose, as the template pose in the STAR model is not neutral. This is not necessary for VAREN, where the template is in a natural rest pose. STAR also defines a set of activation weights, \\(A\\), for each vertex, to weight the output of each linear model. Here, \\(A\\) is a matrix of shape \\((n_J - 1)\\times n_V\\) that we initialize with the distance between vertices and joints. Following STAR, \\(A\\) is passed through a nonlinear rectifying function. During training, we use as losses the Chamfer distance and the regularization loss on the deformations. Weights for these losses are: \\(\\alpha_{dist} = 1e^{3}\\), \\(\\alpha_{reg} = 100\\). All networks are trained for 100 epochs, with a learning rate of \\(1e^{-5}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.621, + 0.733, + 0.636 + ], + "angle": 0, + "content": "4.1. Test Datasets and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We consider two test sets. The In-shape test set, with 275 frames from 5 horses, includes the prototype horses used to train the shape deformation model, but not used for training the VAREN network. The Out-shape test set, 154 frames for 6 horses, includes subjects that have not been used at all. On the first set, we expect to obtain lower errors, and see only the effect of the pose-dependent deformation term, while the second set will provide a general performance assessment of the whole VAREN model. Results are reported in Table 1 and Table 2. We report errors in terms of average Chamfer distance, and of mesh-to-scan distance over a subset of model vertices obtained by removing the head, tail, ears, ankles and hooves. We do this to reduce the influence of the outliers and missing parts on the error scores. VAREN provides the best accuracy in comparison with the baseline and previous work. Figure 9 shows results for the Out-shape dataset. Notice how VAREN generates realistic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5380" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.874, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 9. Results. From left: scan, baseline (purple), baseline mesh-to-scan distance (two views), VAREN (blue), VAREN mesh-to-scan distance (two views), VAREN with part colors, VAREN with muscle colored with the color of the part that gives the main contribution to the deformation, computed as the argmax over body parts of the absolute value of the muscle deformation variable \\(\\beta_{m}\\) (Eq. 5). For the mesh-to-scan visualization, distances are clipped at \\(4\\mathrm{cm}\\), then normalized; darker regions indicate higher distance." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.531, + 0.456, + 0.63 + ], + "angle": 0, + "content": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
Baseline10.321.038.200.89
SMPL9.720.987.060.72
STAR9.510.676.990.63
VAREN (Our)9.370.746.350.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.632, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Table 1. Results on the In-shape testset. Baseline is the model without pose-dependent deformations. Errors in mm." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.659, + 0.455, + 0.771 + ], + "angle": 0, + "content": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
hSMAL21.923.0217.012.51
Baseline12.001.9610.072.55
SMPL11.381.668.842.01
STAR11.031.578.361.77
VAREN (Our)10.881.597.781.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.847 + ], + "angle": 0, + "content": "Table 2. Results on the Out-shape testset. hSMAL uses the hS- MAL model for alignment. Baseline is the model without posedependent deformations. Errors in mm. Wilcoxon significance test for STAR and VAREN comparison: p-values are 0.04 and 3.3e-06 for Chamfer and mesh-to-scan distance, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "deformations, in particular for the neck region. We also illustrate, for each muscle, the body part that most contributes to its deformation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.62, + 0.547 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.687 + ], + "angle": 0, + "content": "We introduced VAREN, the first parametric 3D model of horses learned from real data. In contrast to previous work, VAREN exploits a novel formulation that captures pose-dependent muscle deformations, resulting in better accuracy compared to state-of-the-art approaches, while also being more compact and connecting 3D pose to muscle deformations. By focusing on quality and anatomical realism, VAREN can support a wide set of AI applications in the equestrian world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. Silvia Zuffi is supported by PNRR project FAIR - Future AI Research (PE00000013), Spoke 8 - Pervasive AI (CUP H97G22000210007) under the NRRP MUR program by NextGenerationEU. We thank Minsoek Kim, Peter Kulits, Gökce Ergünfor, and Emil Breustedtfor for helping with pre-processing the scans. We also thank Dr. Zala Zgank, for assisting with data collection, and the horse owners for their time and for bringing their horses. Disclosure. MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon. MJB has financial interests in Amazon and Meshcapade GmbH. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "5381" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] https://www.3dscanstore.com/3d-body-models/ecorche-3d-models/horse-ecorche-3d-model.4,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.281, + 0.173 + ], + "angle": 0, + "content": "[2] http://chumpy.org.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.468, + 0.243 + ], + "angle": 0, + "content": "[3] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5461-5470, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[4] Brett Allen, Brian Curless, and Zoran Popovic. The space of human body shapes: Reconstruction and parameterization from range scans. ACM Trans. Graph., 22(3):587-594, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[5] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a correlated model of identity and pose-dependent body shape variation for real-time synthesis. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 147-156, Airela-Ville, Switzerland, Switzerland, 2006. Eurographics Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.468, + 0.456 + ], + "angle": 0, + "content": "[6] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE: Shape Completion and Animation of PEople. ACM Trans. Graph. (Proc. SIGGRAPH, 24(3):408-416, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.468, + 0.541 + ], + "angle": 0, + "content": "[7] Marc Badger, Yufu Wang, Adarsh Modh, Ammon Perkes, Nikos Kolotouros, Bernd G Pfrommer, Marc F Schmidt, and Kostas Daniilidis. 3d bird reconstruction: a dataset, model, and shape recovery from a single view. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-17, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[8] Benjamin Biggs, Oliver Boyne, James Charles, Andrew Fitzgibbon, and Roberto Cipolla. Who left the dogs out? 3D animal reconstruction with expectation maximization in the loop. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.468, + 0.669 + ], + "angle": 0, + "content": "[9] T. J. Cashman and A. W. Fitzgibbon. What shape are dolphins? building 3D morphable models from 2d images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):232-244, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.468, + 0.727 + ], + "angle": 0, + "content": "[10] Yinpeng Chen, Zicheng Liu, and Zhengyou Zhang. Tensor-based human body modeling. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-112, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[11] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22091-22101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[12] M.S. Fischer, K.E. Lilje, J. Laustroer, and A. Andikfar. Dogs in Motion. Kosmos, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[13] N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.P. Seidel. A statistical model of human pose and body shape. Computer Graphics Forum, 28(2):337-346, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3D: Learning articulated 3D animals by distilling 2d diffusion. In International Conference on 3D Vision (3DV), 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Angjoo Kanazawa, Shahar Kovalsky, Ronen Basri, and David Jacobs. Learning 3D deformation of animals from 2d images. Comput. Graph. Forum, 35(2):365-374, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] Marilyn Keller, Silvia Zuffi, Michael J. Black, and Sergi Pujades. OSSO: Obtaining skeletal shape from outside. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 20492-20501, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[18] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, Liu C. Karen, and Michael J. Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM ToG, Proc. SIGGRAPH Asia, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[19] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3D reconstruction of articulated categories from motion. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1737-1746, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[20] Ci Li, Nima Ghorbani, Sofia Broome, Maheen Rashid, Michael J. Black, Elin Hernlund, Hedvig Kjellström, and Silvia Zuffi. hsmal: Detailed horse shape and pose reconstruction for motion pattern recognition. In First CV4Animals Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Sandro Lombardi, Bangbang Yang, Tianxing Fan, Hujun Bao, Guofeng Zhang, Marc Pollefeys, and Zhaopeng Cui. Latenthuman: Shape-and-pose disentangled latent representation for human bodies. In International Conference on 3D Vision (3DV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[23] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[24] Eadweard Muybridge. Animal Locomotion. An Electro-Photographic Investigation of Consecutive Phases of Animal Movements. Philadelphia: University of Pennsylvania Press, 1887. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Proceedings of the European Conference on Computer Vision (ECCV), pages 598-613, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Deva Ramanan, David A Forsyth, and Kobus Barnard. Building models of animals from video. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(8):1319-1334, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[29] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876-3884, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[30] Nadine Ruegg, Shashank Tripathi, Konrad Schindler, Michael J. Black, and Silvia Zuffi. BITE: Beyond priors for improved three-D dog pose estimation. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8867-8876, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[31] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Proceedings of the Fifth Eurographics Symposium on Geometry Processing, Barcelona, Spain, July 4-6, 2007, pages 109-116, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[32] Yufu Wang, Nikos Kolotouros, Kostas Daniilidis, and Marc Badger. Birds of a feather: Capturing avian shape models from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 14739-14749, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[33] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3D objects by watching videos. International Journal of Computer Vision (IJCV), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[34] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3D animals in the wild. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[35] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[36] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[37] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3D shape reconstruction. In Neural Information Processing Systems (NeurIPS), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[38] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. BANMo: Building animatable 3D neural models from many casual videos. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2853-2863, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Gengshan Yang, Chaoyang Wang, N Dinesh Reddy, and Deva Ramanan. Reconstructing animatable categories from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16995-17005, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[40] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3914-3924, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shape from sparse image ensemble via 3D part discovery. In Neural Information Processing Systems (NeurIPS), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[42] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[43] Chun-Han Yao, Amit Raj, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Artic3d: Learning robust articulated 3D shapes from noisy web image collections. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[44] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5524-5532, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3955-3963, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5383" + } + ] +] \ No newline at end of file diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_origin.pdf b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0029b3cbcbc5cc0b8b7ec3338122a5fbd1ba9308 --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/0f31a4f3-cf8b-45be-b24e-a0b28f1cf54f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91392743fa9fac31d4bfdd9e5b26bdd4783b081c32b268c62af8ebc85bfe6ae6 +size 7676621 diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/full.md b/2024/VAREN_ Very Accurate and Realistic Equine Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b0f9b8aff46f8b0ac68592b5b0b52f4b6dc82f7 --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/full.md @@ -0,0 +1,224 @@ +# VAREN: Very Accurate and Realistic Equine Network + +Silvia Zuffi1, Ylva Mellbin2, Ci Li3, Markus Hoeschle4, Hedvig Kjellström3,2, Senya Polikovsky4, Elin Hernlund2, and Michael J. Black4 + +1IMATI-CNR, Milan, Italy, 2SLU, Uppsala, Sweden, 3KTH, Stockholm, Sweden, 4Max Planck Institute for Intelligent Systems, Tübingen, Germany + +![](images/93023e9dfad391bddb5a2e6a18b8517fe6f887e16d1d0a0f789370d6b8d993ee.jpg) +Figure 1. VAREN is learned from real horses and models pose-dependent deformations at the muscle level, enabling the generation of 3D horses with high realism. Inspired by Muybridge [24], we show a sequence of frames of horses in motion, generated by VAREN, where the first frame illustrates, for each row, the horse in the rest pose. We show three different shapes, obtained by sampling the VAREN model. + +# Abstract + +Data-driven three-dimensional parametric shape models of the human body have gained enormous popularity both for the analysis of visual data and for the generation of synthetic humans. Following a similar approach for animals does not scale to the multitude of existing animal species, not to mention the difficulty of accessing subjects to scan in 3D. However, we argue that for domestic species of great importance, like the horse, it is a highly valuable investment to put effort into gathering a large dataset of real 3D scans, and learn a realistic 3D articulated shape model. We introduce VAREN, a novel 3D articulated parametric shape model learned from 3D scans of many real horses. VAREN bridges synthesis and analysis tasks, as the generated model instances have unprecedented realism, while being able to represent horses of different sizes and shapes. Differently from previous body models, VAREN has two resolutions, an anatomical skeleton, and interpretable, learned pose-dependent deformations, which are related to the body muscles. We show with experiments that this formulation has superior performance with respect to previous strategies for modeling pose-dependent deformations. + +mations in the human body case, while also being more compact and allowing an analysis of the relationship between articulation and muscle deformation during articulated motion. The VAREN model and data are available at https://varen.is.tue.mpg.de. + +# 1. Introduction + +Horses are arguably the most valuable domestic animal and there is a large industry focused on their breeding, care, training, and use in sports. Buying a horse is a large investment, and keeping a horse requires resources and time. Although they are large animals, horses are delicate. The accumulated loads from training and competition frequently result in unrecoverable injuries in their skeletal or tendinous limb structures, ultimately leading to euthanasia. Consequently, among domestic animals, horses are widely studied from a behavioral and biomechanical perspective, with the aim of evaluating their performance, interpreting their state of pain, and preventing injuries. To facilitate such analysis using computer vision, our goal is to develop a highly accurate and detailed 3D model of horses that can be articulated, fit to data, and animated. Here we learn such a model + +![](images/8827cc573e621c2dd373afb5a98d9b50392b6407c5e3f30426e2d11e2a558fb2.jpg) +Figure 2. Dynamic 3D horse scanner. The horse is asked to lift one hind limb. + +(see Fig. 1) using a novel formulation that captures pose-dependent muscle deformations. Such a model may facilitate 3D motion analysis in-the-wild, body shape estimation, diagnosis of illness and performance, horse-human interaction and behavior analysis, and generation of 3D synthetic horses for VFX, VR, AR, and gaming. In particular, accurate markerless motion capture of horses would provide a valuable new tool for the study of horse motion during challenging activities in natural environments. + +For humans, there has been great progress on capturing their 3D shape and pose from images and video. This work often exploits the SMPL body model [22], a 3D mesh-based representation of the human body, controlled by 3D shape and pose parameters. SMPL provides strong priors over human shape and pose, enabling 3D pose reconstruction from ambiguous 2D data. Creating such a 3D model of horses is more challenging. While 3D articulated shape models of horses have been created in the past, they lack realism and expressiveness. In contrast, models like SMPL are learned from thousands of scans, and include an expressive shape space and a pose-dependent deformation model to represent how the body deforms under articulation. Unfortunately, obtaining high-quality 3D scans of horses in a variety of poses is challenging. To fill the gap, we need to (1) capture 3D scans of horses in motion and (2) formulate a new parametric 3D model of horses that represents important aspects of their physiology. In both cases, our solutions deviate from prior work on humans to accommodate the unique aspects of horses. + +Dataset. We use a novel setup (3dMD Ltd.) to capture dynamic 3D scans of horses over time (see Fig. 2). Scanning a large number of horses presents technical and practical challenges, requiring a large cooperative effort between computer scientists and animal care experts. Our dataset includes 50 horses of different breeds and sizes, ranging from small ponies to large horses, with a height difference of more than 1.5 meters. This shape variation is significantly larger than with humans. To model pose-dependent shape changes, we need to capture the horses in a range of poses. Additionally, the animal is always managed by a care per + +![](images/344b873cea3f8996ca2cf958a65eb4e93c967df1fc3c98dd5db450b30bfabd13.jpg) +Figure 3. Samples from the VAREN model result in a diversity of horse shapes. + +son who may occlude the horse during scanning, resulting in missing data. All these issues present novel challenges that we overcome. As a result, our dataset includes in total approximately 4000 cleaned up raw 3D scans, for which we provide accurately registered 3D meshes – training and testing split. Written consent was provided by all animal owners before the animals entered the study. The data collection procedure did not include any invasive technique, therefore no animal ethical approval was needed. + +Model. Using this unique dataset, we train a new model named VAREN (pronounced Varenne1 in French). VAREN is the first 3D articulated model of horses that is learned from real data and that models pose-dependent muscle deformations. Samples from the VAREN model are shown in a neutral pose in Fig. 3 and animated using motion capture data in Fig. 1. To model horses in motion, we must also capture the non-rigid deformations of the body that occur during movement. Previous work on humans models pose-dependent deformations that are learned from 3D scans and conditioned on the body articulation. In particular, STAR [25] models such deformations locally based on the distance from the nearby joints. Such methods do not explicitly model the deformation of the muscles. Scans of humans are typically acquired with some form of clothing (not fully naked) and most humans have a layer of subcutaneous adipose tissue that hides the musculature. Horses, however, are different. They generally carry less body fat than humans and their hair is short, making their musculature more directly observable. Modeling the visible muscle deformation is important because it relates directly to the health and performance of the animal. To that end, we propose a novel learned muscle-based deformation model. Specifically, we group the body surface points into regions that correspond to the superficial muscles, rather than based on their distance to the joints. Note that a single muscle can span more than one body part (see for example the back muscle in Fig. 4) and it is important to capture these long-range cor + +![](images/9ac79c7ce7328841311ad7779ed1f9a38170aa7ee5d9b7b4ba80673d28ed2e2d.jpg) +Figure 4. Skeleton and muscles. From left: anatomical skeleton manually aligned to the model template, selected joints, muscle structure and skin labels on the template. We consider 76 muscles. Head, tail and hooves have no muscle definitions. + +relations. In VAREN, we define deformations per-muscle and learn the influence of body part articulation on the observable muscle deformation from the scan data. We find that, in comparison with previous formulations like those used in SMPL and STAR, our proposed muscle-inspired pose-dependent deformation model has superior accuracy, while also being more compact and interpretable. This is relevant for our high-resolution horse model. While we do not measure muscle activation directly, our formulation can be relevant for research on body pose using measurements of muscle activation [11]. A further novelty of our model, compared with previous work, is that we define the model skeleton, that is the location of the joints in the kinematic tree, based on real horse anatomy. Exploiting a 3D model of the horse skeleton, and with the help of an expert on horse biomechanics, we define the articulation points for the model. Previous human body models define the joint locations as a result of an optimization procedure, resulting in a per-model definition with poor correspondence to the true anatomical skeleton, in particular at the shoulders and hip joints [17]. These models are motivated by applications in computer graphics and vision, where, in many cases, a precise anatomical skeleton may not be needed. For a model like VAREN to be valuable to horse breeders, trainers, and veterinarians, it should provide information that is anatomically relevant. Moreover, an anatomical definition of the model joints, being model-independent, could facilitate comparison and motion data transfer between different models, and, eventually, species. + +In summary, we make two key contributions. First, we provide a new dataset for the modeling of horse shape and pose. The quantity, resolution, variation, and quality of the VAREN data dwarfs any previous dataset of animal shape and pose, and opens up the possibility for the community to explore new questions in 3D shape representation. Second, we develop a new shape model that drives the observable deformation of the muscles based on the animal pose using an anatomical skeleton. + +# 2. Related Work + +We review prior work on learning human body models from data, as our problem and setting share similarities with methods proposed in this case. Then, we present the few + +3D models for animals previously proposed. In addition, we provide an overview of methods for 3D animal reconstruction from visual data, in order to set the context in which we see our novel model applied. + +Human Body Models. There is a long history of learning 3D mesh-based models of the human body from 3D scans [4-6, 10, 13, 22, 26, 35]. There is also recent work on implicit representations that we do not consider here, e.g. [3, 21, 23]. Most recent work is based on SMPL [22], a 3D body model that is based on vertex deformations defined by linear spaces. SMPL is fully data-driven: pose dependent deformations are defined as global representations and learned from data. One issue with SMPL is that the pose-dependent deformations are not local and capture spurious long-range correlations in the data. This is addressed by STAR [25], which adds locality to the joints to the pose-dependent deformations, weighting the outcome of global linear spaces with the distance from skeleton joints. While STAR conditions the pose deformation on the rough BMI of the body, none of the models above explicitly model the observable muscle deformation. VAREN addresses this for the first time. Also, unlike VAREN, the prior methods model articulation using a kinematic tree only loosely based on the human skeleton. A first direction into modeling the human skeleton is OSSO [17], who models the human body skeleton, followed by SKEL [18], which drives the pose of SMPL with a biomechanical skeletal model. SKEL, however, uses the default SMPL model deformations and, unlike VAREN does not train deformations that are driven by the skeletal pose. + +Animal Body Models. The modeling of animal shape for computer vision applications has received less attention. Since 3D scanning of animals is challenging, research has mainly focused on learning from images; e.g. Ramanan et al. learn 2D articulated models [27], while Cashman and Fitzgibbon [9] adapt a rough mean shape using 2D image cues to reconstruct dolphins and bears in 3D. Zuffi et al. learn SMAL, a parametric 3D animal shape model, using 3D scans of toys [44]. Wang et al. learn the 3D shape variation of birds from images starting from a synthetic model [32]. Li et al. learn a horse-specific model similar to SMAL, using 3D scans of horse figurines [20]. Still, others learn to represent category-specific shapes as either meshes [39] or implicit surfaces [14], but do not decouple shape and pose. None of these methods learn 3D articulated models from real 3D scans, which contain noise. The multi-species SMAL model [44] contains an Equine class, but SMAL is learned from a limited number of toys, thus horses generated with the SMAL model do not exhibit a wide range of diverse shapes. The more recent hSMAL model is horse-specific [20]. While learned from a larger number of horses of different shapes, hSMAL is still learned from toys, and has a limited resolution. While useful for reconstructing + +horses from images and video, we find in our experiments that hSMAL is not expressive enough to accurately represent real horses with high detail. Being learned from rigid toys, both models do not address body deformations. + +Animal 3D Reconstruction. The 3D reconstruction of animals from images and video follows a model-based or a model free approach. In the first case, an existing 3D shape model of the species of interest is available, and the method outputs the model parameters of 3D shape and pose. This approach is useful when dealing with challenging input modalities, like in-the-wild monocular images, or the goal is to estimate animal shape and pose for downstream analysis tasks. For example, Zuffi et al. address multi-species 3D reconstruction [45], Kanazawa et al. learn 3D deformations of animals [15], Badger et al. reconstruct 3D birds [7], while several methods estimate the shape and pose of dogs [8, 29, 30]. Our work fits in this category, but provides a 3D model with a higher level of realism than any previous methods. Model-free methods, in contrast, do not assume that a 3D model of the animal class is available, and the output of these methods is typically a 3D surface, in case of static input, or, eventually, a 3D animatable object in case of a video input. Works in this class include CMR [16] and DOVE [33], which reconstruct birds from images and ViSER [37], LASR [36], BANMo [38] and PPR [40], which reconstruct articulated 3D shapes from video, the latter incorporating physical constraints. Kokkinos and Kokkinos learn 3D reconstruction from video without using a shape model by relying only on a template [19]. MagicPony [34], LASSIE [41], Hi-LASSIE [42], ARTIC3D [43] learn 3D reconstruction from image collections. + +# 3. Method + +We learn the VAREN model in two stages. First, we learn an articulated parametric shape model from a set of scans, that we call prototypes, which are of different horses in a neutral pose, that is, not undergoing pose-dependent deformations. This gives us a model with a shape space and generic articulation. We then align, in an unsupervised way, this static model to the dynamic 3D scans in which the prototype horses perform different movements. From this data, namely the scans and the alignment parameters of 3D shape, pose and translation, the VAREN network learns a pose-dependent deformation model that improves the matching between the learned model and the scans. The novel deformation model is a function of the shape and pose parameters, such that, at test time, the network generates a shaped and posed horse, with realistic pose-dependent deformations, given the parameters. The pose-dependent deformation model is defined on the basis of anatomical structures that we incorporate into VAREN from a purchased realistic graphics (CG) model [1], which includes the body skeleton + +![](images/05f75e39a519dbb3f5470eccf022a5fbd28bdad8d6975d40cd53a653a7742272.jpg) +Figure 5. Comparison between the average equine model in SMAL (left), the hSMAL+ template (middle) and the VAREN template (right). Learned from real horses, the VAREN template has a more proportioned neck with respect to hSMAL+. + +![](images/8f7aa45cbca78a02af6b02fee14aac131b2780ea002af4a38d4b2aae467ae515.jpg) +Figure 6. Visualization of the VAREN shape space. We capture small ponies, large breeds, and many in between, obtaining a shape space with variation in size, morphology and face features. The tail and mane are tied, and not considered as part of body shape change. + +and the muscles. In the following, we detail the procedure to learn the horse model from the prototypes, the creation of the training dataset, and the definition of the VAREN network. + +# 3.1. VAREN Horse Model + +Alignment of the Prototype Scans. We follow a model-based approach to express the prototypes in a common topology, specifically, as the alignment model, we use hSMAL [20], a recently proposed articulated horse model (Fig. 5). The hSMAL model follows the formulation of SMPL, and is defined by a triangular mesh template $\mathbf{v}_t$ , with $n_V$ vertices, a matrix $B$ of shape $3n_V \times n_B$ containing the $n_B$ basis vectors of a linear shape deformation space, a joint regressor $J_r$ that maps model vertices to a set of $n_J$ joint locations, and a skinning weight matrix $W$ . A horse is generated, given shape parameters $\beta$ and pose parameters $\theta$ , by first deforming the template into an intrinsic shape $\mathbf{v}_s$ , then applying Linear Blend Skinning (LBS) to rotate the body parts according to the given pose: + +$$ +\begin{array}{l} \mathbf {v} _ {s} = \mathbf {v} _ {t} + B \boldsymbol {\beta} ^ {T} \\ \mathbf {v} = L B S \left(\mathbf {v} _ {s}, \theta ; W, J _ {r}\right). \tag {1} \\ \end{array} +$$ + +While able to represent horses of different shapes, we found that the hSMAL model has, in our case, two limitations: a low resolution (the model has about $1.5K$ vertices) and a shape space that is not sufficiently expressive to accurately capture the shape of real horses. To overcome these limitations, we create an hSMAL model with an increased resolution (3647 vertices) and additional shape deformations. + +We call this model hSMAL+. Furthermore, we observed that the purchased CG model [1], in addition to containing anatomical structures, has a more natural posture than hSMAL+. Therefore, we modify the rest pose of hSMAL+ to match the pose of the CG model. This is done by fitting hSMAL+ to the CG model and using the obtained mesh as the new template. In hSMAL+ we retain skinning weights and part segmentations from the original hSMAL model. + +Additional Shape Deformations. We augment the hSML shape space by defining additional dimensions in the shape deformation matrix $B$ (Eq. 1) to obtain a linear scaling of a set of body parts. Note that limb scaling has been applied before to animal models, but our formulation is novel, as previous work [8] applies scaling during the LBS process, while here we define scaling in the model shape space. Let $j$ be a body part for which we want to add scaling as an additional dimension in the shape space. We define a new shape matrix $B_{1}$ of dimension $3n_V \times (n_B + 1)$ by adding a new column to the shape matrix $B$ : + +$$ +B _ {1} = B | B _ {j} +$$ + +$$ +B _ {j} = s _ {j} \mathbf {v} _ {j, c}, \tag {2} +$$ + +where $\mathbf{v}_{j,c}$ is a column vector corresponding to the template, but with non-zero values only for the coordinate $c$ of the vertices of the part $j$ we want to scale, and $s_j$ is a scaling factor. Here $j$ can also indicate a set of parts, for example, all the segments of the tail. Expanding $B$ is not sufficient, as the deformation we obtain is expressed for the model template $\mathbf{v}_t$ , while we need it to be applied to the part vertices of the intrinsic shape $\mathbf{v}_s$ (Eq. 1). Therefore, we add an iterative procedure that estimates a vertex shift $d_v$ for the part vertices. Let $J$ be the joints of the extended model with shape space $B_1$ , $J_0$ be the joints of the original model, and $J_s$ a vector of cumulative joint shifts initialized with zero values. In an iterative process that considers all the body joints from the root to the leaves of the kinematic tree, for each body part $j$ , we compute the vertex shift as $d_v(j) = J_s(j) + J_0(j) - J(j)$ , add it to the part vertices, and then add $d_v(j)$ to the children of the part $j$ in $J_s$ , such that the joint shift is propagated through the kinematic tree. The set of body parts we consider for limb scaling are the ears, the legs and the tail, resulting in an additional set of 6 deformation vectors, as we group the left and right ears in a single part. The scaling factors in Equation 2 for these parts are 0.05 for the ears, where we scale the whole part, and 0.1 for the legs and tail, where we scale only the vertical axis. Note that the scaling factors in Equation 2 are applied to define the new shape space dimensions, but the scaling is then weighted with corresponding shape variables as for the original PCA dimensions. + +Model-based Alignment. In order to align the hSMAL+ model to the prototypes, first, we manually annotate land- + +![](images/4145ee81a8f22b5a86340cd10f9c31055f3a9ff1f452070e9fba4bd054c4a09a.jpg) +Figure 7. Model alignment to the scans. Examples of noisy (top) and clean (bottom) registrations. In white is the scan and in purple is the model registration. + +marks on the selected scans: for most of the prototypes we annotate 7 landmarks: the tip of the four hooves, the tip of the ears, and the tip of the tail. Note that this process, while manual, is applied to only one frame per horse, resulting in a very small fraction of the time invested to record a specific subject. The alignment procedure minimizes an energy with data terms that consist of the mesh-to-scan and scan-to-mesh distances between the scan and the model, and the distance between landmarks defined on the model and annotated on the scans. We add regularization terms for the $\beta$ and the $\theta$ variables, set as the square of the elements of the $\beta$ vector, and the $L1$ norm of the $\theta$ values, respectively. The latter formulation is as a consequence of the fact that in this stage we are aligning scans that are close to the model neutral pose by selection. The pose prior is stronger for the tail, which is often partially observed, and weaker for the ears, which might not be in a neutral pose. We optimize the energy with Chumpy [2], with an annealing schedule that decreases the weight of the pose prior and increases the weight of the data terms. After the model-based alignment, we increase the resolution of the hSMAL+ model (from 3647 to 13873 vertices) and perform a model-free alignment, where, to capture the fine-level details of the scans, we optimize over the vertices. In this case, we use two regularization terms: as-rigid-as-possible (ARAP) [31] regularization, and a coupling term between the 3D vertices and the model solution. Note that when increasing the model resolution, we do not alter the vertices that are inside the head (tongue and palate). Moreover, we keep the low-resolution vertices as the first 3647 vertices of the model, such that VAREN can be easily used at two resolutions. + +VAREN Shape Space. The VAREN horse model shares the same formulation of SMPL and hSMAL expressed by Equation 1. The shape space of the model, indicated by the matrix $B$ , is learned in the following way. Once the prototype scans have been aligned, they are all brought into the neutral pose of the alignment model (they are already close to this pose). The mean template shape is computed, giving an averaged prototype horse (Fig. 5), which is subtracted from the prototypes. Then Principal Component Analysis + +(PCA) is applied, giving a linear shape space of skin deformations (Fig. 6). On the new template, we define the skeleton and muscle skin labels. + +# 3.1.1 Anatomical Structures + +In this section we describe how we incorporate anatomical structures, namely the skeleton and the muscles, into the VAREN model. Our goal is to define the model skeleton joints in correspondence with real joints, and characterize the model vertices according to the muscle they are closest to, as this association will be exploited in the pose-dependent deformation model. To this end, we exploit a realistic CG model of a horse [1] that includes a skeleton and muscles. We have already exploited the CG model to define the neutral pose for VAREN (Fig. 5), with the further advantage that now the VAREN template has the same pose of the skeleton of the CG model. However, given the body part proportions differ between the GC model and the VAREN template, we manually deform the individual skeleton bones to better match the template. We then select a set of joint locations on the skeleton that are anatomically relevant and closest to the hSMAL+ joints (Fig. 4). Finally, we re-compute the joint regressor of the VAREN model such that the anatomical joints are used. In addition to the existing hSMAL joints, we define two additional joints, corresponding to the left and right scapula. We found this necessary to fit the model to complex poses. A joint in correspondence to the top of the scapula models biomechanical principles for horses and dogs [12]. The muscles in the purchased CG model are grouped by large sections. We separate them into individual muscles, label them, and find the closest muscle for each vertex of the VAREN registration to the CG model. These labels are then transferred to the VAREN template. In this way, we obtain muscle labels for each vertex of the VAREN horse model (Fig. 4). Now that we have a model with anatomical joints and pervertex muscle indices, first we re-register the prototypes to obtain the shape and pose parameters for the new model, then we learn pose-dependent deformations from the dynamic scans. The deformations are learned from an optimization network on a dataset of 3D scans aligned to the horse model. In the following, we describe the generation of the training set and the VAREN network. + +# 3.2. Training Dataset + +The training dataset is composed by a set of scans and corresponding alignment parameters. The scans are frames of a set of clips containing the same prototype horses used to learn the shape model performing different activities: standing still, moving the neck forward and toward a side, and moving the legs (see Fig. 2). As a pre-processing step, the scans are cleaned-up to remove the floor and the horse owner using a simple procedure based on point cloudCLUS + +![](images/47c8f21d8b23b3a21e431cca26294012d39977074c7e86d5115b0d74f935d481.jpg) +Figure 8. Post-alignment cleaning. We exploit the model registration to clean up the 3D scans. Left column: input scan, middle: cleaning with a single bounding box around the registration. Right: cleaning with a set of bounding boxes. The horse head is sometimes not captured due to the presence of a person. + +fering. We register the VAREN model to the dynamic sequences with an alignment procedure similar to the one used for the prototypes. The difference here is that we do not optimize for the horse shape, but keep the shape parameter fixed to the one computed with the VAREN prototype registration. For the first frame, we also initialize with the prototype pose. After the first, each frame is initialized with the previous solution, and the pose prior takes the form of the $L1$ distance from the previous pose. The alignment procedure does not require manual annotations. However, due to the presence of noisy frames, where the pre-processing clean-up failed, sometimes the alignment process gives a wrong result, and we need to annotate a few landmarks to correct such cases. We annotate about the $1\%$ of the frames with about 4 landmarks. Moreover, we need to remove from the training set those frames where complete parts are missing. While we can deal with large holes in the scans, the alignment may return an incorrect solution if the entire head or a whole limb is missing. On the other hand, we allow ears to be missing in the scans, as the pose-dependent deformation model does not include them (see Fig. 4). Examples of registrations are shown in Fig. 7. The model registration can be used to further clean up and filter the data. We do this with a procedure that creates a set of bounding boxes around the registration to filter out the scan regions that are not pertinent to the horse (Fig. 8). We then compute the maximum scan-to-mesh distance to filter out scans with residual outliers. We retain, in total, about $3.6K$ training scans. Note that not all the prototype scans have been used to create the training set, as we leave out a few for testing. + +# 3.3. VAREN Network + +The pose-dependent deformations model is an additive term augmenting the intrinsic body shape (see Eq. 1): + +$$ +\mathbf {v} _ {s + d} = \mathbf {v} _ {t} + B \boldsymbol {\beta} ^ {T} + \mathbf {d v} _ {m} (\theta , \beta), \tag {3} +$$ + +where $\mathbf{d}\mathbf{v}_m$ are the muscle deformations, obtained as: + +$$ +\mathbf {d v} _ {m, I _ {i}} (\theta , \beta) = D _ {i} (\beta_ {m}), \tag {4} +$$ + +where $i$ is the muscle index, $D_{i}$ is a decoder composed by a linear layer of dimension $(4(n_{J} - 1) + 2)\times 3n_{I_{i}}$ , with $I_{i}$ the set of indices of the skin vertices associated with muscle $i$ . The number of joints $n_J = 38$ , excluding the root, is multiplied by 4 as we use quaternions to represent pose. Equation 4 is applied for each muscle. The muscle deformation variable $\beta_{m}$ is defined as: + +$$ +\beta_ {m} = A \circ W _ {m} \left(\theta_ {2,.. n, j}, \beta_ {1, 2}\right), \tag {5} +$$ + +where $A$ is a selection matrix of dimension $n_M \times (4(n_J - 1) + 2)$ , with $n_M = 76$ being the number of muscles. We initialize $A(i, 4k: 4k + 4) = 1$ if muscle $i$ belongs to part $j$ , with $k \in N(j)$ , otherwise 0. Here $N(j)$ indicates the set composed by the part $j$ and its neighbors. An advantage of our formulation is that each row of the product $A \circ W_m$ is now a weighting vector for the pose parameters indicating the influence between muscles and body parts, regardless of the pose, while the muscle deformation variable $\beta_m$ indicates the strength of the muscle deformation. Both $A$ and $W_m$ are optimized during training. The last column of Fig. 9 shows the association between muscle deformation and body parts. + +# 3.3.1 Network Training + +During training, the network reads the training set, and stores the 3D pose, shape and translation estimated by the alignment procedure. Input scans are resampled at a fixed size of 20000 points to have a uniform size within a batch, with size 4. We store the alignment parameters in memory such that, during network optimization, pose and translation variables can be fine-tuned. However, we found in our experiments that fine-tuning the alignment variables does not significantly change the results, indicating that the training registrations are already of good quality. Here we present results for which we did not fine-tune the training data. We train using different losses: the Chamfer distance (from the PyTorch3D library [28]) between the generated horse mesh and the input scan, a regularization term, implemented as an edge length minimization loss, also from PyTorch3D, applied to the deformations $\mathbf{d}\mathbf{v}_m$ and only to the mesh triangles on the muscle boundaries, and a regularization loss on the matrix $A$ , to favor small values for body parts that are likely not to influence the muscles, namely the tail, feet, mouth and ears. This is implemented as the $L_{1}$ norm of the matrix entries for these parts. Weights for the losses are: $\alpha_{dist} = 1e^{3}$ , $\alpha_{reg} = 10$ , $\alpha_{bound} = 100$ , $\alpha_{A} = 1e^{3}$ . + +# 4. Experiments + +We apply the VAREN network to a set of held-out alignments. As a baseline, we consider the model without pose-dependent deformations. We also compare with defining pose-dependent deformations as in SMPL and STAR. + +SMPL Pose-dependent Deformations. The SMPL model defines the pose-dependent deformations as a linear model over the pose vector of relative rotation matrices (see Eq. 9 in [22]). The pose feature vector is therefore of size $9(n_{J} - 1)$ , and the pose blend shapes is a matrix of shape $3n_{V} \times 9(n_{J} - 1)$ . When training to learn the pose blend shapes, we use the Chamfer distance loss, the regularization loss applied to the deformations, and a loss on the $L_{1}$ norm of the pose blend shapes values. Weights for these losses are: $\alpha_{dist} = 1e^{3}$ , $\alpha_{reg} = 100$ , $\alpha_{reg,B_{p}} = 1e^{3}$ . + +STAR Pose-dependent Deformations. The STAR model defines the pose-dependent deformations as a set of perpart linear models over the part and neighbors rotations, expressed with quaternions. The second value $\beta_{2}$ of the shape vector is added to the set of pose features, as it correlates with the human BMI index. Here, lacking this information for horses, we condition the deformation with the first two elements of the $\beta$ vector, as in the VAREN shape space all the components capture morphological information (see Fig. 6). Each of the $n_J - 1$ linear models has dimension $(4(n_{Jj} + 1) + 2)\times 3n_V$ , where $n_{Jj}$ is the number of neighbors of the part $j$ . STAR defines the rotations relative to the human neutral pose, as the template pose in the STAR model is not neutral. This is not necessary for VAREN, where the template is in a natural rest pose. STAR also defines a set of activation weights, $A$ , for each vertex, to weight the output of each linear model. Here, $A$ is a matrix of shape $(n_J - 1)\times n_V$ that we initialize with the distance between vertices and joints. Following STAR, $A$ is passed through a nonlinear rectifying function. During training, we use as losses the Chamfer distance and the regularization loss on the deformations. Weights for these losses are: $\alpha_{dist} = 1e^{3}$ , $\alpha_{reg} = 100$ . All networks are trained for 100 epochs, with a learning rate of $1e^{-5}$ . + +# 4.1. Test Datasets and Results + +We consider two test sets. The In-shape test set, with 275 frames from 5 horses, includes the prototype horses used to train the shape deformation model, but not used for training the VAREN network. The Out-shape test set, 154 frames for 6 horses, includes subjects that have not been used at all. On the first set, we expect to obtain lower errors, and see only the effect of the pose-dependent deformation term, while the second set will provide a general performance assessment of the whole VAREN model. Results are reported in Table 1 and Table 2. We report errors in terms of average Chamfer distance, and of mesh-to-scan distance over a subset of model vertices obtained by removing the head, tail, ears, ankles and hooves. We do this to reduce the influence of the outliers and missing parts on the error scores. VAREN provides the best accuracy in comparison with the baseline and previous work. Figure 9 shows results for the Out-shape dataset. Notice how VAREN generates realistic + +![](images/45f5ef74d5cdacb66c56a893e2bd88c813eac258c53d73cee8d7fa6580c01c6e.jpg) +Figure 9. Results. From left: scan, baseline (purple), baseline mesh-to-scan distance (two views), VAREN (blue), VAREN mesh-to-scan distance (two views), VAREN with part colors, VAREN with muscle colored with the color of the part that gives the main contribution to the deformation, computed as the argmax over body parts of the absolute value of the muscle deformation variable $\beta_{m}$ (Eq. 5). For the mesh-to-scan visualization, distances are clipped at $4\mathrm{cm}$ , then normalized; darker regions indicate higher distance. + +
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
Baseline10.321.038.200.89
SMPL9.720.987.060.72
STAR9.510.676.990.63
VAREN (Our)9.370.746.350.59
+ +Table 1. Results on the In-shape testset. Baseline is the model without pose-dependent deformations. Errors in mm. + +
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
hSMAL21.923.0217.012.51
Baseline12.001.9610.072.55
SMPL11.381.668.842.01
STAR11.031.578.361.77
VAREN (Our)10.881.597.781.71
+ +Table 2. Results on the Out-shape testset. hSMAL uses the hS- MAL model for alignment. Baseline is the model without posedependent deformations. Errors in mm. Wilcoxon significance test for STAR and VAREN comparison: p-values are 0.04 and 3.3e-06 for Chamfer and mesh-to-scan distance, respectively. + +deformations, in particular for the neck region. We also illustrate, for each muscle, the body part that most contributes to its deformation. + +# 5. Conclusion + +We introduced VAREN, the first parametric 3D model of horses learned from real data. In contrast to previous work, VAREN exploits a novel formulation that captures pose-dependent muscle deformations, resulting in better accuracy compared to state-of-the-art approaches, while also being more compact and connecting 3D pose to muscle deformations. By focusing on quality and anatomical realism, VAREN can support a wide set of AI applications in the equestrian world. + +Acknowledgements. Silvia Zuffi is supported by PNRR project FAIR - Future AI Research (PE00000013), Spoke 8 - Pervasive AI (CUP H97G22000210007) under the NRRP MUR program by NextGenerationEU. We thank Minsoek Kim, Peter Kulits, Gökce Ergünfor, and Emil Breustedtfor for helping with pre-processing the scans. We also thank Dr. Zala Zgank, for assisting with data collection, and the horse owners for their time and for bringing their horses. Disclosure. MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon. MJB has financial interests in Amazon and Meshcapade GmbH. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society. + +# References + +[1] https://www.3dscanstore.com/3d-body-models/ecorche-3d-models/horse-ecorche-3d-model.4,5,6 +[2] http://chumpy.org.5 +[3] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5461-5470, 2021. 3 +[4] Brett Allen, Brian Curless, and Zoran Popovic. The space of human body shapes: Reconstruction and parameterization from range scans. ACM Trans. Graph., 22(3):587-594, 2003. 3 +[5] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a correlated model of identity and pose-dependent body shape variation for real-time synthesis. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 147-156, Airela-Ville, Switzerland, Switzerland, 2006. Eurographics Association. +[6] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE: Shape Completion and Animation of PEople. ACM Trans. Graph. (Proc. SIGGRAPH, 24(3):408-416, 2005. 3 +[7] Marc Badger, Yufu Wang, Adarsh Modh, Ammon Perkes, Nikos Kolotouros, Bernd G Pfrommer, Marc F Schmidt, and Kostas Daniilidis. 3d bird reconstruction: a dataset, model, and shape recovery from a single view. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-17, 2020. 4 +[8] Benjamin Biggs, Oliver Boyne, James Charles, Andrew Fitzgibbon, and Roberto Cipolla. Who left the dogs out? 3D animal reconstruction with expectation maximization in the loop. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 4, 5 +[9] T. J. Cashman and A. W. Fitzgibbon. What shape are dolphins? building 3D morphable models from 2d images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):232-244, 2013. 3 +[10] Yinpeng Chen, Zicheng Liu, and Zhengyou Zhang. Tensor-based human body modeling. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-112, 2013. 3 +[11] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22091-22101, 2023. 3 +[12] M.S. Fischer, K.E. Lilje, J. Laustroer, and A. Andikfar. Dogs in Motion. Kosmos, 2011. 6 +[13] N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.P. Seidel. A statistical model of human pose and body shape. Computer Graphics Forum, 28(2):337-346, 2009. 3 +[14] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3D: Learning articulated 3D animals by distilling 2d diffusion. In International Conference on 3D Vision (3DV), 2024. 3 + +[15] Angjoo Kanazawa, Shahar Kovalsky, Ronen Basri, and David Jacobs. Learning 3D deformation of animals from 2d images. Comput. Graph. Forum, 35(2):365-374, 2016. 4 +[16] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 4 +[17] Marilyn Keller, Silvia Zuffi, Michael J. Black, and Sergi Pujades. OSSO: Obtaining skeletal shape from outside. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 20492-20501, 2022. 3 +[18] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, Liu C. Karen, and Michael J. Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM ToG, Proc. SIGGRAPH Asia, 2023. 3 +[19] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3D reconstruction of articulated categories from motion. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1737-1746, 2021. 4 +[20] Ci Li, Nima Ghorbani, Sofia Broome, Maheen Rashid, Michael J. Black, Elin Hernlund, Hedvig Kjellström, and Silvia Zuffi. hsmal: Detailed horse shape and pose reconstruction for motion pattern recognition. In First CV4Animals Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4 +[21] Sandro Lombardi, Bangbang Yang, Tianxing Fan, Hujun Bao, Guofeng Zhang, Marc Pollefeys, and Zhaopeng Cui. Latenthuman: Shape-and-pose disentangled latent representation for human bodies. In International Conference on 3D Vision (3DV), 2021. 3 +[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 7 +[23] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[24] Eadweard Muybridge. Animal Locomotion. An Electro-Photographic Investigation of Consecutive Phases of Animal Movements. Philadelphia: University of Pennsylvania Press, 1887. 1 +[25] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Proceedings of the European Conference on Computer Vision (ECCV), pages 598-613, 2020. 2, 3 +[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[27] Deva Ramanan, David A Forsyth, and Kobus Barnard. Building models of animals from video. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(8):1319-1334, 2006. 3 +[28] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia + +Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 7 +[29] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876-3884, 2022. 4 +[30] Nadine Ruegg, Shashank Tripathi, Konrad Schindler, Michael J. Black, and Silvia Zuffi. BITE: Beyond priors for improved three-D dog pose estimation. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8867-8876, 2023. 4 +[31] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Proceedings of the Fifth Eurographics Symposium on Geometry Processing, Barcelona, Spain, July 4-6, 2007, pages 109-116, 2007. 5 +[32] Yufu Wang, Nikos Kolotouros, Kostas Daniilidis, and Marc Badger. Birds of a feather: Capturing avian shape models from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 14739-14749, 2021. 3 +[33] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3D objects by watching videos. International Journal of Computer Vision (IJCV), 2023. 4 +[34] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3D animals in the wild. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4 +[35] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[36] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 4 +[37] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3D shape reconstruction. In Neural Information Processing Systems (NeurIPS), 2021. 4 +[38] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. BANMo: Building animatable 3D neural models from many casual videos. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2853-2863, 2022. 4 +[39] Gengshan Yang, Chaoyang Wang, N Dinesh Reddy, and Deva Ramanan. Reconstructing animatable categories from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16995-17005, 2023. 3 + +[40] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3914-3924, 2023. 4 +[41] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shape from sparse image ensemble via 3D part discovery. In Neural Information Processing Systems (NeurIPS), 2022. 4 +[42] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4 +[43] Chun-Han Yao, Amit Raj, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Artic3d: Learning robust articulated 3D shapes from noisy web image collections. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4 +[44] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5524-5532, 2017. 3 +[45] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3955-3963, 2018. 4 \ No newline at end of file diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/images.zip b/2024/VAREN_ Very Accurate and Realistic Equine Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed5fcbb118e421b280a64f39f2ce535069a5dad0 --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98c640b11ba702b47d3ae704ade925968f45793a314c908a96829e1d9e174b4c +size 415487 diff --git a/2024/VAREN_ Very Accurate and Realistic Equine Network/layout.json b/2024/VAREN_ Very Accurate and Realistic Equine Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1206ad7958965db0525510d9a240e645d08f8e14 --- /dev/null +++ b/2024/VAREN_ Very Accurate and Realistic Equine Network/layout.json @@ -0,0 +1,6957 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "type": "text", + "content": "VAREN: Very Accurate and Realistic Equine Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 152, + 533, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 533, + 181 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 533, + 181 + ], + "type": "text", + "content": "Silvia Zuffi1, Ylva Mellbin2, Ci Li3, Markus Hoeschle4, Hedvig Kjellström3,2, Senya Polikovsky4, Elin Hernlund2, and Michael J. Black4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 190, + 494, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 190, + 494, + 220 + ], + "spans": [ + { + "bbox": [ + 100, + 190, + 494, + 220 + ], + "type": "text", + "content": "1IMATI-CNR, Milan, Italy, 2SLU, Uppsala, Sweden, 3KTH, Stockholm, Sweden, 4Max Planck Institute for Intelligent Systems, Tübingen, Germany" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 99, + 236, + 499, + 376 + ], + "blocks": [ + { + "bbox": [ + 99, + 236, + 499, + 376 + ], + "lines": [ + { + "bbox": [ + 99, + 236, + 499, + 376 + ], + "spans": [ + { + "bbox": [ + 99, + 236, + 499, + 376 + ], + "type": "image", + "image_path": "93023e9dfad391bddb5a2e6a18b8517fe6f887e16d1d0a0f789370d6b8d993ee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 547, + 415 + ], + "type": "text", + "content": "Figure 1. VAREN is learned from real horses and models pose-dependent deformations at the muscle level, enabling the generation of 3D horses with high realism. Inspired by Muybridge [24], we show a sequence of frames of horses in motion, generated by VAREN, where the first frame illustrates, for each row, the horse in the rest pose. We show three different shapes, obtained by sampling the VAREN model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 462, + 289, + 715 + ], + "type": "text", + "content": "Data-driven three-dimensional parametric shape models of the human body have gained enormous popularity both for the analysis of visual data and for the generation of synthetic humans. Following a similar approach for animals does not scale to the multitude of existing animal species, not to mention the difficulty of accessing subjects to scan in 3D. However, we argue that for domestic species of great importance, like the horse, it is a highly valuable investment to put effort into gathering a large dataset of real 3D scans, and learn a realistic 3D articulated shape model. We introduce VAREN, a novel 3D articulated parametric shape model learned from 3D scans of many real horses. VAREN bridges synthesis and analysis tasks, as the generated model instances have unprecedented realism, while being able to represent horses of different sizes and shapes. Differently from previous body models, VAREN has two resolutions, an anatomical skeleton, and interpretable, learned pose-dependent deformations, which are related to the body muscles. We show with experiments that this formulation has superior performance with respect to previous strategies for modeling pose-dependent deformations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 443, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 443, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 547, + 504 + ], + "type": "text", + "content": "mations in the human body case, while also being more compact and allowing an analysis of the relationship between articulation and muscle deformation during articulated motion. The VAREN model and data are available at https://varen.is.tue.mpg.de." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 388, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "content": "Horses are arguably the most valuable domestic animal and there is a large industry focused on their breeding, care, training, and use in sports. Buying a horse is a large investment, and keeping a horse requires resources and time. Although they are large animals, horses are delicate. The accumulated loads from training and competition frequently result in unrecoverable injuries in their skeletal or tendinous limb structures, ultimately leading to euthanasia. Consequently, among domestic animals, horses are widely studied from a behavioral and biomechanical perspective, with the aim of evaluating their performance, interpreting their state of pain, and preventing injuries. To facilitate such analysis using computer vision, our goal is to develop a highly accurate and detailed 3D model of horses that can be articulated, fit to data, and animated. Here we learn such a model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5374" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 276, + 192 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 192 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 192 + ], + "type": "image", + "image_path": "8827cc573e621c2dd373afb5a98d9b50392b6407c5e3f30426e2d11e2a558fb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 222 + ], + "type": "text", + "content": "Figure 2. Dynamic 3D horse scanner. The horse is asked to lift one hind limb." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 286, + 341 + ], + "type": "text", + "content": "(see Fig. 1) using a novel formulation that captures pose-dependent muscle deformations. Such a model may facilitate 3D motion analysis in-the-wild, body shape estimation, diagnosis of illness and performance, horse-human interaction and behavior analysis, and generation of 3D synthetic horses for VFX, VR, AR, and gaming. In particular, accurate markerless motion capture of horses would provide a valuable new tool for the study of horse motion during challenging activities in natural environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 581 + ], + "type": "text", + "content": "For humans, there has been great progress on capturing their 3D shape and pose from images and video. This work often exploits the SMPL body model [22], a 3D mesh-based representation of the human body, controlled by 3D shape and pose parameters. SMPL provides strong priors over human shape and pose, enabling 3D pose reconstruction from ambiguous 2D data. Creating such a 3D model of horses is more challenging. While 3D articulated shape models of horses have been created in the past, they lack realism and expressiveness. In contrast, models like SMPL are learned from thousands of scans, and include an expressive shape space and a pose-dependent deformation model to represent how the body deforms under articulation. Unfortunately, obtaining high-quality 3D scans of horses in a variety of poses is challenging. To fill the gap, we need to (1) capture 3D scans of horses in motion and (2) formulate a new parametric 3D model of horses that represents important aspects of their physiology. In both cases, our solutions deviate from prior work on humans to accommodate the unique aspects of horses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Dataset. We use a novel setup (3dMD Ltd.) to capture dynamic 3D scans of horses over time (see Fig. 2). Scanning a large number of horses presents technical and practical challenges, requiring a large cooperative effort between computer scientists and animal care experts. Our dataset includes 50 horses of different breeds and sizes, ranging from small ponies to large horses, with a height difference of more than 1.5 meters. This shape variation is significantly larger than with humans. To model pose-dependent shape changes, we need to capture the horses in a range of poses. Additionally, the animal is always managed by a care per" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 515, + 190 + ], + "blocks": [ + { + "bbox": [ + 340, + 70, + 515, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 515, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 515, + 190 + ], + "type": "image", + "image_path": "344b873cea3f8996ca2cf958a65eb4e93c967df1fc3c98dd5db450b30bfabd13.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 198, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 198, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 545, + 221 + ], + "type": "text", + "content": "Figure 3. Samples from the VAREN model result in a diversity of horse shapes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 232, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 339 + ], + "type": "text", + "content": "son who may occlude the horse during scanning, resulting in missing data. All these issues present novel challenges that we overcome. As a result, our dataset includes in total approximately 4000 cleaned up raw 3D scans, for which we provide accurately registered 3D meshes – training and testing split. Written consent was provided by all animal owners before the animals entered the study. The data collection procedure did not include any invasive technique, therefore no animal ethical approval was needed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 340, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 675 + ], + "type": "text", + "content": "Model. Using this unique dataset, we train a new model named VAREN (pronounced Varenne1 in French). VAREN is the first 3D articulated model of horses that is learned from real data and that models pose-dependent muscle deformations. Samples from the VAREN model are shown in a neutral pose in Fig. 3 and animated using motion capture data in Fig. 1. To model horses in motion, we must also capture the non-rigid deformations of the body that occur during movement. Previous work on humans models pose-dependent deformations that are learned from 3D scans and conditioned on the body articulation. In particular, STAR [25] models such deformations locally based on the distance from the nearby joints. Such methods do not explicitly model the deformation of the muscles. Scans of humans are typically acquired with some form of clothing (not fully naked) and most humans have a layer of subcutaneous adipose tissue that hides the musculature. Horses, however, are different. They generally carry less body fat than humans and their hair is short, making their musculature more directly observable. Modeling the visible muscle deformation is important because it relates directly to the health and performance of the animal. To that end, we propose a novel learned muscle-based deformation model. Specifically, we group the body surface points into regions that correspond to the superficial muscles, rather than based on their distance to the joints. Note that a single muscle can span more than one body part (see for example the back muscle in Fig. 4) and it is important to capture these long-range cor" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 712 + ], + "type": "text", + "content": "1Varenne is considered to be the best trotter of all time. No other trotter has won so many of the most important races in the world and set as many records as Varenne. Source: Wikipedia." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5375" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 284, + 125 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 284, + 125 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 284, + 125 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 284, + 125 + ], + "type": "image", + "image_path": "9ac79c7ce7328841311ad7779ed1f9a38170aa7ee5d9b7b4ba80673d28ed2e2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 135, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 46, + 135, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 286, + 178 + ], + "type": "text", + "content": "Figure 4. Skeleton and muscles. From left: anatomical skeleton manually aligned to the model template, selected joints, muscle structure and skin labels on the template. We consider 76 muscles. Head, tail and hooves have no muscle definitions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 286, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 286, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 286, + 537 + ], + "type": "text", + "content": "relations. In VAREN, we define deformations per-muscle and learn the influence of body part articulation on the observable muscle deformation from the scan data. We find that, in comparison with previous formulations like those used in SMPL and STAR, our proposed muscle-inspired pose-dependent deformation model has superior accuracy, while also being more compact and interpretable. This is relevant for our high-resolution horse model. While we do not measure muscle activation directly, our formulation can be relevant for research on body pose using measurements of muscle activation [11]. A further novelty of our model, compared with previous work, is that we define the model skeleton, that is the location of the joints in the kinematic tree, based on real horse anatomy. Exploiting a 3D model of the horse skeleton, and with the help of an expert on horse biomechanics, we define the articulation points for the model. Previous human body models define the joint locations as a result of an optimization procedure, resulting in a per-model definition with poor correspondence to the true anatomical skeleton, in particular at the shoulders and hip joints [17]. These models are motivated by applications in computer graphics and vision, where, in many cases, a precise anatomical skeleton may not be needed. For a model like VAREN to be valuable to horse breeders, trainers, and veterinarians, it should provide information that is anatomically relevant. Moreover, an anatomical definition of the model joints, being model-independent, could facilitate comparison and motion data transfer between different models, and, eventually, species." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 538, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 286, + 644 + ], + "type": "text", + "content": "In summary, we make two key contributions. First, we provide a new dataset for the modeling of horse shape and pose. The quantity, resolution, variation, and quality of the VAREN data dwarfs any previous dataset of animal shape and pose, and opens up the possibility for the community to explore new questions in 3D shape representation. Second, we develop a new shape model that drives the observable deformation of the muscles based on the animal pose using an anatomical skeleton." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "We review prior work on learning human body models from data, as our problem and setting share similarities with methods proposed in this case. Then, we present the few" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "3D models for animals previously proposed. In addition, we provide an overview of methods for 3D animal reconstruction from visual data, in order to set the context in which we see our novel model applied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 123, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 434 + ], + "type": "text", + "content": "Human Body Models. There is a long history of learning 3D mesh-based models of the human body from 3D scans [4-6, 10, 13, 22, 26, 35]. There is also recent work on implicit representations that we do not consider here, e.g. [3, 21, 23]. Most recent work is based on SMPL [22], a 3D body model that is based on vertex deformations defined by linear spaces. SMPL is fully data-driven: pose dependent deformations are defined as global representations and learned from data. One issue with SMPL is that the pose-dependent deformations are not local and capture spurious long-range correlations in the data. This is addressed by STAR [25], which adds locality to the joints to the pose-dependent deformations, weighting the outcome of global linear spaces with the distance from skeleton joints. While STAR conditions the pose deformation on the rough BMI of the body, none of the models above explicitly model the observable muscle deformation. VAREN addresses this for the first time. Also, unlike VAREN, the prior methods model articulation using a kinematic tree only loosely based on the human skeleton. A first direction into modeling the human skeleton is OSSO [17], who models the human body skeleton, followed by SKEL [18], which drives the pose of SMPL with a biomechanical skeletal model. SKEL, however, uses the default SMPL model deformations and, unlike VAREN does not train deformations that are driven by the skeletal pose." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "Animal Body Models. The modeling of animal shape for computer vision applications has received less attention. Since 3D scanning of animals is challenging, research has mainly focused on learning from images; e.g. Ramanan et al. learn 2D articulated models [27], while Cashman and Fitzgibbon [9] adapt a rough mean shape using 2D image cues to reconstruct dolphins and bears in 3D. Zuffi et al. learn SMAL, a parametric 3D animal shape model, using 3D scans of toys [44]. Wang et al. learn the 3D shape variation of birds from images starting from a synthetic model [32]. Li et al. learn a horse-specific model similar to SMAL, using 3D scans of horse figurines [20]. Still, others learn to represent category-specific shapes as either meshes [39] or implicit surfaces [14], but do not decouple shape and pose. None of these methods learn 3D articulated models from real 3D scans, which contain noise. The multi-species SMAL model [44] contains an Equine class, but SMAL is learned from a limited number of toys, thus horses generated with the SMAL model do not exhibit a wide range of diverse shapes. The more recent hSMAL model is horse-specific [20]. While learned from a larger number of horses of different shapes, hSMAL is still learned from toys, and has a limited resolution. While useful for reconstructing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5376" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "horses from images and video, we find in our experiments that hSMAL is not expressive enough to accurately represent real horses with high detail. Being learned from rigid toys, both models do not address body deformations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 129, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 129, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 289, + 452 + ], + "type": "text", + "content": "Animal 3D Reconstruction. The 3D reconstruction of animals from images and video follows a model-based or a model free approach. In the first case, an existing 3D shape model of the species of interest is available, and the method outputs the model parameters of 3D shape and pose. This approach is useful when dealing with challenging input modalities, like in-the-wild monocular images, or the goal is to estimate animal shape and pose for downstream analysis tasks. For example, Zuffi et al. address multi-species 3D reconstruction [45], Kanazawa et al. learn 3D deformations of animals [15], Badger et al. reconstruct 3D birds [7], while several methods estimate the shape and pose of dogs [8, 29, 30]. Our work fits in this category, but provides a 3D model with a higher level of realism than any previous methods. Model-free methods, in contrast, do not assume that a 3D model of the animal class is available, and the output of these methods is typically a 3D surface, in case of static input, or, eventually, a 3D animatable object in case of a video input. Works in this class include CMR [16] and DOVE [33], which reconstruct birds from images and ViSER [37], LASR [36], BANMo [38] and PPR [40], which reconstruct articulated 3D shapes from video, the latter incorporating physical constraints. Kokkinos and Kokkinos learn 3D reconstruction from video without using a shape model by relying only on a template [19]. MagicPony [34], LASSIE [41], Hi-LASSIE [42], ARTIC3D [43] learn 3D reconstruction from image collections." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "We learn the VAREN model in two stages. First, we learn an articulated parametric shape model from a set of scans, that we call prototypes, which are of different horses in a neutral pose, that is, not undergoing pose-dependent deformations. This gives us a model with a shape space and generic articulation. We then align, in an unsupervised way, this static model to the dynamic 3D scans in which the prototype horses perform different movements. From this data, namely the scans and the alignment parameters of 3D shape, pose and translation, the VAREN network learns a pose-dependent deformation model that improves the matching between the learned model and the scans. The novel deformation model is a function of the shape and pose parameters, such that, at test time, the network generates a shaped and posed horse, with realistic pose-dependent deformations, given the parameters. The pose-dependent deformation model is defined on the basis of anatomical structures that we incorporate into VAREN from a purchased realistic graphics (CG) model [1], which includes the body skeleton" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 332, + 72, + 520, + 121 + ], + "blocks": [ + { + "bbox": [ + 332, + 72, + 520, + 121 + ], + "lines": [ + { + "bbox": [ + 332, + 72, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 332, + 72, + 520, + 121 + ], + "type": "image", + "image_path": "05f75e39a519dbb3f5470eccf022a5fbd28bdad8d6975d40cd53a653a7742272.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 124, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 304, + 124, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 547, + 168 + ], + "type": "text", + "content": "Figure 5. Comparison between the average equine model in SMAL (left), the hSMAL+ template (middle) and the VAREN template (right). Learned from real horses, the VAREN template has a more proportioned neck with respect to hSMAL+." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 327, + 170, + 525, + 273 + ], + "blocks": [ + { + "bbox": [ + 327, + 170, + 525, + 273 + ], + "lines": [ + { + "bbox": [ + 327, + 170, + 525, + 273 + ], + "spans": [ + { + "bbox": [ + 327, + 170, + 525, + 273 + ], + "type": "image", + "image_path": "8f7aa45cbca78a02af6b02fee14aac131b2780ea002af4a38d4b2aae467ae515.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 277, + 546, + 332 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 332 + ], + "type": "text", + "content": "Figure 6. Visualization of the VAREN shape space. We capture small ponies, large breeds, and many in between, obtaining a shape space with variation in size, morphology and face features. The tail and mane are tied, and not considered as part of body shape change." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 389 + ], + "type": "text", + "content": "and the muscles. In the following, we detail the procedure to learn the horse model from the prototypes, the creation of the training dataset, and the definition of the VAREN network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 397, + 430, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 430, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 430, + 409 + ], + "type": "text", + "content": "3.1. VAREN Horse Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": "Alignment of the Prototype Scans. We follow a model-based approach to express the prototypes in a common topology, specifically, as the alignment model, we use hSMAL [20], a recently proposed articulated horse model (Fig. 5). The hSMAL model follows the formulation of SMPL, and is defined by a triangular mesh template " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "n_V" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " vertices, a matrix " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "3n_V \\times n_B" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " containing the " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "n_B" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " basis vectors of a linear shape deformation space, a joint regressor " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "J_r" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " that maps model vertices to a set of " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "n_J" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " joint locations, and a skinning weight matrix " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": ". A horse is generated, given shape parameters " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": " and pose parameters " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": ", by first deforming the template into an intrinsic shape " + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 415, + 547, + 584 + ], + "type": "text", + "content": ", then applying Linear Blend Skinning (LBS) to rotate the body parts according to the given pose:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 591, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 591, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 365, + 591, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {v} _ {s} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} \\\\ \\mathbf {v} = L B S \\left(\\mathbf {v} _ {s}, \\theta ; W, J _ {r}\\right). \\tag {1} \\\\ \\end{array}", + "image_path": "a1e22742b4a716fac6af285125076d998c60a37e2f9766e0810fca88a95aeeb1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "While able to represent horses of different shapes, we found that the hSMAL model has, in our case, two limitations: a low resolution (the model has about " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.5K" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " vertices) and a shape space that is not sufficiently expressive to accurately capture the shape of real horses. To overcome these limitations, we create an hSMAL model with an increased resolution (3647 vertices) and additional shape deformations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5377" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "We call this model hSMAL+. Furthermore, we observed that the purchased CG model [1], in addition to containing anatomical structures, has a more natural posture than hSMAL+. Therefore, we modify the rest pose of hSMAL+ to match the pose of the CG model. This is done by fitting hSMAL+ to the CG model and using the obtained mesh as the new template. In hSMAL+ we retain skinning weights and part segmentations from the original hSMAL model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": "Additional Shape Deformations. We augment the hSML shape space by defining additional dimensions in the shape deformation matrix " + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": " (Eq. 1) to obtain a linear scaling of a set of body parts. Note that limb scaling has been applied before to animal models, but our formulation is novel, as previous work [8] applies scaling during the LBS process, while here we define scaling in the model shape space. Let " + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": " be a body part for which we want to add scaling as an additional dimension in the shape space. We define a new shape matrix " + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "inline_equation", + "content": "B_{1}" + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "inline_equation", + "content": "3n_V \\times (n_B + 1)" + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": " by adding a new column to the shape matrix " + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 175, + 289, + 307 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 313, + 198, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 198, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 198, + 327 + ], + "type": "interline_equation", + "content": "B _ {1} = B | B _ {j}", + "image_path": "cfb5e296849f6b7a82b4e145a281dc3352921cf6267ae858df012b71c5618f1c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 328, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 328, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 328, + 287, + 342 + ], + "type": "interline_equation", + "content": "B _ {j} = s _ {j} \\mathbf {v} _ {j, c}, \\tag {2}", + "image_path": "bc23f34182b6cf03f72dcfd7dd031b5d1a297cfe57ee2e36c5a539a8c0938869.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{j,c}" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " is a column vector corresponding to the template, but with non-zero values only for the coordinate " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " of the vertices of the part " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " we want to scale, and " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " is a scaling factor. Here " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " can also indicate a set of parts, for example, all the segments of the tail. Expanding " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " is not sufficient, as the deformation we obtain is expressed for the model template " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": ", while we need it to be applied to the part vertices of the intrinsic shape " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " (Eq. 1). Therefore, we add an iterative procedure that estimates a vertex shift " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " for the part vertices. Let " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " be the joints of the extended model with shape space " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "B_1" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "J_0" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " be the joints of the original model, and " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "J_s" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " a vector of cumulative joint shifts initialized with zero values. In an iterative process that considers all the body joints from the root to the leaves of the kinematic tree, for each body part " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": ", we compute the vertex shift as " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "d_v(j) = J_s(j) + J_0(j) - J(j)" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": ", add it to the part vertices, and then add " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "d_v(j)" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " to the children of the part " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "inline_equation", + "content": "J_s" + }, + { + "bbox": [ + 46, + 348, + 289, + 682 + ], + "type": "text", + "content": ", such that the joint shift is propagated through the kinematic tree. The set of body parts we consider for limb scaling are the ears, the legs and the tail, resulting in an additional set of 6 deformation vectors, as we group the left and right ears in a single part. The scaling factors in Equation 2 for these parts are 0.05 for the ears, where we scale the whole part, and 0.1 for the legs and tail, where we scale only the vertical axis. Note that the scaling factors in Equation 2 are applied to define the new shape space dimensions, but the scaling is then weighted with corresponding shape variables as for the original PCA dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Model-based Alignment. In order to align the hSMAL+ model to the prototypes, first, we manually annotate land-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 335, + 73, + 512, + 182 + ], + "blocks": [ + { + "bbox": [ + 335, + 73, + 512, + 182 + ], + "lines": [ + { + "bbox": [ + 335, + 73, + 512, + 182 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 512, + 182 + ], + "type": "image", + "image_path": "4145ee81a8f22b5a86340cd10f9c31055f3a9ff1f452070e9fba4bd054c4a09a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 185, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 305, + 185, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 547, + 217 + ], + "type": "text", + "content": "Figure 7. Model alignment to the scans. Examples of noisy (top) and clean (bottom) registrations. In white is the scan and in purple is the model registration." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": "marks on the selected scans: for most of the prototypes we annotate 7 landmarks: the tip of the four hooves, the tip of the ears, and the tip of the tail. Note that this process, while manual, is applied to only one frame per horse, resulting in a very small fraction of the time invested to record a specific subject. The alignment procedure minimizes an energy with data terms that consist of the mesh-to-scan and scan-to-mesh distances between the scan and the model, and the distance between landmarks defined on the model and annotated on the scans. We add regularization terms for the " + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": " variables, set as the square of the elements of the " + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": " vector, and the " + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": " norm of the " + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 228, + 547, + 598 + ], + "type": "text", + "content": " values, respectively. The latter formulation is as a consequence of the fact that in this stage we are aligning scans that are close to the model neutral pose by selection. The pose prior is stronger for the tail, which is often partially observed, and weaker for the ears, which might not be in a neutral pose. We optimize the energy with Chumpy [2], with an annealing schedule that decreases the weight of the pose prior and increases the weight of the data terms. After the model-based alignment, we increase the resolution of the hSMAL+ model (from 3647 to 13873 vertices) and perform a model-free alignment, where, to capture the fine-level details of the scans, we optimize over the vertices. In this case, we use two regularization terms: as-rigid-as-possible (ARAP) [31] regularization, and a coupling term between the 3D vertices and the model solution. Note that when increasing the model resolution, we do not alter the vertices that are inside the head (tongue and palate). Moreover, we keep the low-resolution vertices as the first 3647 vertices of the model, such that VAREN can be easily used at two resolutions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "VAREN Shape Space. The VAREN horse model shares the same formulation of SMPL and hSMAL expressed by Equation 1. The shape space of the model, indicated by the matrix " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", is learned in the following way. Once the prototype scans have been aligned, they are all brought into the neutral pose of the alignment model (they are already close to this pose). The mean template shape is computed, giving an averaged prototype horse (Fig. 5), which is subtracted from the prototypes. Then Principal Component Analysis" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5378" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "(PCA) is applied, giving a linear shape space of skin deformations (Fig. 6). On the new template, we define the skeleton and muscle skin labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 178, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 178, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 178, + 128 + ], + "type": "text", + "content": "3.1.1 Anatomical Structures" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 136, + 289, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 136, + 289, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 136, + 289, + 591 + ], + "type": "text", + "content": "In this section we describe how we incorporate anatomical structures, namely the skeleton and the muscles, into the VAREN model. Our goal is to define the model skeleton joints in correspondence with real joints, and characterize the model vertices according to the muscle they are closest to, as this association will be exploited in the pose-dependent deformation model. To this end, we exploit a realistic CG model of a horse [1] that includes a skeleton and muscles. We have already exploited the CG model to define the neutral pose for VAREN (Fig. 5), with the further advantage that now the VAREN template has the same pose of the skeleton of the CG model. However, given the body part proportions differ between the GC model and the VAREN template, we manually deform the individual skeleton bones to better match the template. We then select a set of joint locations on the skeleton that are anatomically relevant and closest to the hSMAL+ joints (Fig. 4). Finally, we re-compute the joint regressor of the VAREN model such that the anatomical joints are used. In addition to the existing hSMAL joints, we define two additional joints, corresponding to the left and right scapula. We found this necessary to fit the model to complex poses. A joint in correspondence to the top of the scapula models biomechanical principles for horses and dogs [12]. The muscles in the purchased CG model are grouped by large sections. We separate them into individual muscles, label them, and find the closest muscle for each vertex of the VAREN registration to the CG model. These labels are then transferred to the VAREN template. In this way, we obtain muscle labels for each vertex of the VAREN horse model (Fig. 4). Now that we have a model with anatomical joints and pervertex muscle indices, first we re-register the prototypes to obtain the shape and pose parameters for the new model, then we learn pose-dependent deformations from the dynamic scans. The deformations are learned from an optimization network on a dataset of 3D scans aligned to the horse model. In the following, we describe the generation of the training set and the VAREN network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "type": "text", + "content": "3.2. Training Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "The training dataset is composed by a set of scans and corresponding alignment parameters. The scans are frames of a set of clips containing the same prototype horses used to learn the shape model performing different activities: standing still, moving the neck forward and toward a side, and moving the legs (see Fig. 2). As a pre-processing step, the scans are cleaned-up to remove the floor and the horse owner using a simple procedure based on point cloudCLUS" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 523, + 167 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 523, + 167 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 523, + 167 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 523, + 167 + ], + "type": "image", + "image_path": "47c8f21d8b23b3a21e431cca26294012d39977074c7e86d5115b0d74f935d481.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 227 + ], + "type": "text", + "content": "Figure 8. Post-alignment cleaning. We exploit the model registration to clean up the 3D scans. Left column: input scan, middle: cleaning with a single bounding box around the registration. Right: cleaning with a set of bounding boxes. The horse head is sometimes not captured due to the presence of a person." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "text", + "content": "fering. We register the VAREN model to the dynamic sequences with an alignment procedure similar to the one used for the prototypes. The difference here is that we do not optimize for the horse shape, but keep the shape parameter fixed to the one computed with the VAREN prototype registration. For the first frame, we also initialize with the prototype pose. After the first, each frame is initialized with the previous solution, and the pose prior takes the form of the " + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "text", + "content": " distance from the previous pose. The alignment procedure does not require manual annotations. However, due to the presence of noisy frames, where the pre-processing clean-up failed, sometimes the alignment process gives a wrong result, and we need to annotate a few landmarks to correct such cases. We annotate about the " + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "text", + "content": " of the frames with about 4 landmarks. Moreover, we need to remove from the training set those frames where complete parts are missing. While we can deal with large holes in the scans, the alignment may return an incorrect solution if the entire head or a whole limb is missing. On the other hand, we allow ears to be missing in the scans, as the pose-dependent deformation model does not include them (see Fig. 4). Examples of registrations are shown in Fig. 7. The model registration can be used to further clean up and filter the data. We do this with a procedure that creates a set of bounding boxes around the registration to filter out the scan regions that are not pertinent to the horse (Fig. 8). We then compute the maximum scan-to-mesh distance to filter out scans with residual outliers. We retain, in total, about " + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "inline_equation", + "content": "3.6K" + }, + { + "bbox": [ + 304, + 236, + 546, + 597 + ], + "type": "text", + "content": " training scans. Note that not all the prototype scans have been used to create the training set, as we leave out a few for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 604, + 409, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 409, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 409, + 616 + ], + "type": "text", + "content": "3.3. VAREN Network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "type": "text", + "content": "The pose-dependent deformations model is an additive term augmenting the intrinsic body shape (see Eq. 1):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 655, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 655, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 358, + 655, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {s + d} = \\mathbf {v} _ {t} + B \\boldsymbol {\\beta} ^ {T} + \\mathbf {d v} _ {m} (\\theta , \\beta), \\tag {3}", + "image_path": "f280c1864ae0a2f7aab61d912f88971da4cc0f56bb17ed1744cbd1068b749450.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 679, + 522, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 679, + 522, + 691 + ], + "spans": [ + { + "bbox": [ + 306, + 679, + 522, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 679, + 522, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{d}\\mathbf{v}_m" + }, + { + "bbox": [ + 306, + 679, + 522, + 691 + ], + "type": "text", + "content": " are the muscle deformations, obtained as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 374, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {d v} _ {m, I _ {i}} (\\theta , \\beta) = D _ {i} (\\beta_ {m}), \\tag {4}", + "image_path": "b8aafeea7f7c90ba8a8f0ca1722c69fc464b03a2601e8403107b5566a30d71cf.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5379" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is the muscle index, " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is a decoder composed by a linear layer of dimension " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "(4(n_{J} - 1) + 2)\\times 3n_{I_{i}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " the set of indices of the skin vertices associated with muscle " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". The number of joints " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "n_J = 38" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", excluding the root, is multiplied by 4 as we use quaternions to represent pose. Equation 4 is applied for each muscle. The muscle deformation variable " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\beta_{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 167, + 287, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 167, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 287, + 180 + ], + "type": "interline_equation", + "content": "\\beta_ {m} = A \\circ W _ {m} \\left(\\theta_ {2,.. n, j}, \\beta_ {1, 2}\\right), \\tag {5}", + "image_path": "d9c77c8b933aa1f38fa87468292098e792efa5a8a86a156821346a8d9f767860.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " is a selection matrix of dimension " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "n_M \\times (4(n_J - 1) + 2)" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "n_M = 76" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " being the number of muscles. We initialize " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "A(i, 4k: 4k + 4) = 1" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " if muscle " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " belongs to part " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "k \\in N(j)" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": ", otherwise 0. Here " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "N(j)" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " indicates the set composed by the part " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " and its neighbors. An advantage of our formulation is that each row of the product " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "A \\circ W_m" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " is now a weighting vector for the pose parameters indicating the influence between muscles and body parts, regardless of the pose, while the muscle deformation variable " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\beta_m" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " indicates the strength of the muscle deformation. Both " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "inline_equation", + "content": "W_m" + }, + { + "bbox": [ + 46, + 190, + 289, + 346 + ], + "type": "text", + "content": " are optimized during training. The last column of Fig. 9 shows the association between muscle deformation and body parts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 355, + 157, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 157, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 157, + 367 + ], + "type": "text", + "content": "3.3.1 Network Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": "During training, the network reads the training set, and stores the 3D pose, shape and translation estimated by the alignment procedure. Input scans are resampled at a fixed size of 20000 points to have a uniform size within a batch, with size 4. We store the alignment parameters in memory such that, during network optimization, pose and translation variables can be fine-tuned. However, we found in our experiments that fine-tuning the alignment variables does not significantly change the results, indicating that the training registrations are already of good quality. Here we present results for which we did not fine-tune the training data. We train using different losses: the Chamfer distance (from the PyTorch3D library [28]) between the generated horse mesh and the input scan, a regularization term, implemented as an edge length minimization loss, also from PyTorch3D, applied to the deformations " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{d}\\mathbf{v}_m" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": " and only to the mesh triangles on the muscle boundaries, and a regularization loss on the matrix " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": ", to favor small values for body parts that are likely not to influence the muscles, namely the tail, feet, mouth and ears. This is implemented as the " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": " norm of the matrix entries for these parts. Weights for the losses are: " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{dist} = 1e^{3}" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{reg} = 10" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{bound} = 100" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{A} = 1e^{3}" + }, + { + "bbox": [ + 46, + 375, + 289, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 650, + 128, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 128, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 128, + 663 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 670, + 287, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 670, + 287, + 719 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 287, + 719 + ], + "type": "text", + "content": "We apply the VAREN network to a set of held-out alignments. As a baseline, we consider the model without pose-dependent deformations. We also compare with defining pose-dependent deformations as in SMPL and STAR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "SMPL Pose-dependent Deformations. The SMPL model defines the pose-dependent deformations as a linear model over the pose vector of relative rotation matrices (see Eq. 9 in [22]). The pose feature vector is therefore of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "9(n_{J} - 1)" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", and the pose blend shapes is a matrix of shape " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "3n_{V} \\times 9(n_{J} - 1)" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ". When training to learn the pose blend shapes, we use the Chamfer distance loss, the regularization loss applied to the deformations, and a loss on the " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " norm of the pose blend shapes values. Weights for these losses are: " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\alpha_{dist} = 1e^{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\alpha_{reg} = 100" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\alpha_{reg,B_{p}} = 1e^{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": "STAR Pose-dependent Deformations. The STAR model defines the pose-dependent deformations as a set of perpart linear models over the part and neighbors rotations, expressed with quaternions. The second value " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " of the shape vector is added to the set of pose features, as it correlates with the human BMI index. Here, lacking this information for horses, we condition the deformation with the first two elements of the " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " vector, as in the VAREN shape space all the components capture morphological information (see Fig. 6). Each of the " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "n_J - 1" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " linear models has dimension " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "(4(n_{Jj} + 1) + 2)\\times 3n_V" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "n_{Jj}" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " is the number of neighbors of the part " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": ". STAR defines the rotations relative to the human neutral pose, as the template pose in the STAR model is not neutral. This is not necessary for VAREN, where the template is in a natural rest pose. STAR also defines a set of activation weights, " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": ", for each vertex, to weight the output of each linear model. Here, " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " is a matrix of shape " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "(n_J - 1)\\times n_V" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " that we initialize with the distance between vertices and joints. Following STAR, " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": " is passed through a nonlinear rectifying function. During training, we use as losses the Chamfer distance and the regularization loss on the deformations. Weights for these losses are: " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\alpha_{dist} = 1e^{3}" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\alpha_{reg} = 100" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": ". All networks are trained for 100 epochs, with a learning rate of " + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 304, + 198, + 546, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 491, + 448, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 448, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 448, + 503 + ], + "type": "text", + "content": "4.1. Test Datasets and Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We consider two test sets. The In-shape test set, with 275 frames from 5 horses, includes the prototype horses used to train the shape deformation model, but not used for training the VAREN network. The Out-shape test set, 154 frames for 6 horses, includes subjects that have not been used at all. On the first set, we expect to obtain lower errors, and see only the effect of the pose-dependent deformation term, while the second set will provide a general performance assessment of the whole VAREN model. Results are reported in Table 1 and Table 2. We report errors in terms of average Chamfer distance, and of mesh-to-scan distance over a subset of model vertices obtained by removing the head, tail, ears, ankles and hooves. We do this to reduce the influence of the outliers and missing parts on the error scores. VAREN provides the best accuracy in comparison with the baseline and previous work. Figure 9 shows results for the Out-shape dataset. Notice how VAREN generates realistic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5380" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 534, + 357 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 534, + 357 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 534, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 534, + 357 + ], + "type": "image", + "image_path": "45f5ef74d5cdacb66c56a893e2bd88c813eac258c53d73cee8d7fa6580c01c6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "type": "text", + "content": "Figure 9. Results. From left: scan, baseline (purple), baseline mesh-to-scan distance (two views), VAREN (blue), VAREN mesh-to-scan distance (two views), VAREN with part colors, VAREN with muscle colored with the color of the part that gives the main contribution to the deformation, computed as the argmax over body parts of the absolute value of the muscle deformation variable " + }, + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "type": "inline_equation", + "content": "\\beta_{m}" + }, + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "type": "text", + "content": " (Eq. 5). For the mesh-to-scan visualization, distances are clipped at " + }, + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "type": "inline_equation", + "content": "4\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 360, + 547, + 407 + ], + "type": "text", + "content": ", then normalized; darker regions indicate higher distance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 420, + 279, + 498 + ], + "blocks": [ + { + "bbox": [ + 54, + 420, + 279, + 498 + ], + "lines": [ + { + "bbox": [ + 54, + 420, + 279, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 279, + 498 + ], + "type": "table", + "html": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
Baseline10.321.038.200.89
SMPL9.720.987.060.72
STAR9.510.676.990.63
VAREN (Our)9.370.746.350.59
", + "image_path": "ed166a5cbbfdcaf6f6d35c34968b64531520c2957515b9ad62c4169999846891.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 521, + 278, + 610 + ], + "blocks": [ + { + "bbox": [ + 47, + 500, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 47, + 500, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 287, + 521 + ], + "type": "text", + "content": "Table 1. Results on the In-shape testset. Baseline is the model without pose-dependent deformations. Errors in mm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 521, + 278, + 610 + ], + "lines": [ + { + "bbox": [ + 53, + 521, + 278, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 278, + 610 + ], + "type": "table", + "html": "
Chamfer DistanceMesh-to-Scan
meanstdmeanstd
hSMAL21.923.0217.012.51
Baseline12.001.9610.072.55
SMPL11.381.668.842.01
STAR11.031.578.361.77
VAREN (Our)10.881.597.781.71
", + "image_path": "c99fd5bd236b2ddd343aaaf7ead08c27ae9c80151b60e330e0909c861a87cf56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 613, + 287, + 670 + ], + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 670 + ], + "type": "text", + "content": "Table 2. Results on the Out-shape testset. hSMAL uses the hS- MAL model for alignment. Baseline is the model without posedependent deformations. Errors in mm. Wilcoxon significance test for STAR and VAREN comparison: p-values are 0.04 and 3.3e-06 for Chamfer and mesh-to-scan distance, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "deformations, in particular for the neck region. We also illustrate, for each muscle, the body part that most contributes to its deformation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 544 + ], + "type": "text", + "content": "We introduced VAREN, the first parametric 3D model of horses learned from real data. In contrast to previous work, VAREN exploits a novel formulation that captures pose-dependent muscle deformations, resulting in better accuracy compared to state-of-the-art approaches, while also being more compact and connecting 3D pose to muscle deformations. By focusing on quality and anatomical realism, VAREN can support a wide set of AI applications in the equestrian world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgements. Silvia Zuffi is supported by PNRR project FAIR - Future AI Research (PE00000013), Spoke 8 - Pervasive AI (CUP H97G22000210007) under the NRRP MUR program by NextGenerationEU. We thank Minsoek Kim, Peter Kulits, Gökce Ergünfor, and Emil Breustedtfor for helping with pre-processing the scans. We also thank Dr. Zala Zgank, for assisting with data collection, and the horse owners for their time and for bringing their horses. Disclosure. MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon. MJB has financial interests in Amazon and Meshcapade GmbH. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "5381" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] https://www.3dscanstore.com/3d-body-models/ecorche-3d-models/horse-ecorche-3d-model.4,5,6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 171, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 171, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 171, + 137 + ], + "type": "text", + "content": "[2] http://chumpy.org.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 138, + 286, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 138, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 286, + 192 + ], + "type": "text", + "content": "[3] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5461-5470, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 236 + ], + "type": "text", + "content": "[4] Brett Allen, Brian Curless, and Zoran Popovic. The space of human body shapes: Reconstruction and parameterization from range scans. ACM Trans. Graph., 22(3):587-594, 2003. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 316 + ], + "type": "text", + "content": "[5] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a correlated model of identity and pose-dependent body shape variation for real-time synthesis. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 147-156, Airela-Ville, Switzerland, Switzerland, 2006. Eurographics Association." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "text", + "content": "[6] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE: Shape Completion and Animation of PEople. ACM Trans. Graph. (Proc. SIGGRAPH, 24(3):408-416, 2005. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 286, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 286, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 286, + 428 + ], + "type": "text", + "content": "[7] Marc Badger, Yufu Wang, Adarsh Modh, Ammon Perkes, Nikos Kolotouros, Bernd G Pfrommer, Marc F Schmidt, and Kostas Daniilidis. 3d bird reconstruction: a dataset, model, and shape recovery from a single view. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-17, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 286, + 485 + ], + "type": "text", + "content": "[8] Benjamin Biggs, Oliver Boyne, James Charles, Andrew Fitzgibbon, and Roberto Cipolla. Who left the dogs out? 3D animal reconstruction with expectation maximization in the loop. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2020. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 286, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 286, + 529 + ], + "type": "text", + "content": "[9] T. J. Cashman and A. W. Fitzgibbon. What shape are dolphins? building 3D morphable models from 2d images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):232-244, 2013. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "text", + "content": "[10] Yinpeng Chen, Zicheng Liu, and Zhengyou Zhang. Tensor-based human body modeling. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-112, 2013. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 286, + 610 + ], + "type": "text", + "content": "[11] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22091-22101, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 633 + ], + "type": "text", + "content": "[12] M.S. Fischer, K.E. Lilje, J. Laustroer, and A. Andikfar. Dogs in Motion. Kosmos, 2011. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 668 + ], + "type": "text", + "content": "[13] N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.P. Seidel. A statistical model of human pose and body shape. Computer Graphics Forum, 28(2):337-346, 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[14] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3D: Learning articulated 3D animals by distilling 2d diffusion. In International Conference on 3D Vision (3DV), 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Angjoo Kanazawa, Shahar Kovalsky, Ronen Basri, and David Jacobs. Learning 3D deformation of animals from 2d images. Comput. Graph. Forum, 35(2):365-374, 2016. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[16] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[17] Marilyn Keller, Silvia Zuffi, Michael J. Black, and Sergi Pujades. OSSO: Obtaining skeletal shape from outside. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 20492-20501, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "text", + "content": "[18] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, Liu C. Karen, and Michael J. Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM ToG, Proc. SIGGRAPH Asia, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[19] Filippos Kokkinos and Iasonas Kokkinos. Learning monocular 3D reconstruction of articulated categories from motion. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1737-1746, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 348 + ], + "type": "text", + "content": "[20] Ci Li, Nima Ghorbani, Sofia Broome, Maheen Rashid, Michael J. Black, Elin Hernlund, Hedvig Kjellström, and Silvia Zuffi. hsmal: Detailed horse shape and pose reconstruction for motion pattern recognition. In First CV4Animals Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[21] Sandro Lombardi, Bangbang Yang, Tianxing Fan, Hujun Bao, Guofeng Zhang, Marc Pollefeys, and Zhaopeng Cui. Latenthuman: Shape-and-pose disentangled latent representation for human bodies. In International Conference on 3D Vision (3DV), 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[23] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "text", + "content": "[24] Eadweard Muybridge. Animal Locomotion. An Electro-Photographic Investigation of Consecutive Phases of Animal Movements. Philadelphia: University of Pennsylvania Press, 1887. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[25] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse trained articulated human body regressor. In Proceedings of the European Conference on Computer Vision (ECCV), pages 598-613, 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "type": "text", + "content": "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Deva Ramanan, David A Forsyth, and Kobus Barnard. Building models of animals from video. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(8):1319-1334, 2006. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5382" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[29] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876-3884, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[30] Nadine Ruegg, Shashank Tripathi, Konrad Schindler, Michael J. Black, and Silvia Zuffi. BITE: Beyond priors for improved three-D dog pose estimation. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8867-8876, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[31] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Proceedings of the Fifth Eurographics Symposium on Geometry Processing, Barcelona, Spain, July 4-6, 2007, pages 109-116, 2007. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[32] Yufu Wang, Nikos Kolotouros, Kostas Daniilidis, and Marc Badger. Birds of a feather: Capturing avian shape models from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 14739-14749, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[33] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. DOVE: Learning deformable 3D objects by watching videos. International Journal of Computer Vision (IJCV), 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 410 + ], + "type": "text", + "content": "[34] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3D animals in the wild. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "text", + "content": "[35] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[36] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[37] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. ViSER: Video-specific surface embeddings for articulated 3D shape reconstruction. In Neural Information Processing Systems (NeurIPS), 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[38] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. BANMo: Building animatable 3D neural models from many casual videos. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2853-2863, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Gengshan Yang, Chaoyang Wang, N Dinesh Reddy, and Deva Ramanan. Reconstructing animatable categories from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16995-17005, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 418 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[40] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3914-3924, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[41] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: Learning articulated shape from sparse image ensemble via 3D part discovery. In Neural Information Processing Systems (NeurIPS), 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 249 + ], + "type": "text", + "content": "[42] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 251, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 306 + ], + "type": "text", + "content": "[43] Chun-Han Yao, Amit Raj, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Artic3d: Learning robust articulated 3D shapes from noisy web image collections. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "text", + "content": "[44] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5524-5532, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "type": "text", + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3955-3963, 2018. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5383" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_content_list.json b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..368c2e17493d7a8f5cd58ea28414049492264659 --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_content_list.json @@ -0,0 +1,1522 @@ +[ + { + "type": "text", + "text": "VBenchmark: Comprehensive Benchmark Suite for Video Generative Models", + "text_level": 1, + "bbox": [ + 119, + 117, + 849, + 138 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqi Huang $^{1*}$ Yinan He $^{2*}$ Jiashuo Yu $^{2*}$ Fan Zhang $^{2*}$ Chenyang Si $^{1}$ Yuming Jiang $^{1}$ \nYuanhan Zhang $^{1}$ Tianxing Wu $^{1}$ Qingyang Jin $^{1}$ Nattapol Chanpaisit $^{1}$ \nYaohui Wang $^{2}$ Xinyuan Chen $^{2}$ Limin Wang $^{4,2}$ Dahua Lin $^{2,3\\boxtimes}$ Yu Qiao $^{2\\boxtimes}$ Ziwei Liu $^{1\\boxtimes}$", + "bbox": [ + 109, + 166, + 880, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ The Chinese University of Hong Kong $^{4}$ Nanjing University", + "bbox": [ + 127, + 238, + 841, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://vchitect.github.io/VBench-project/", + "bbox": [ + 274, + 289, + 692, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53e38a1b6b63f2c36f3d80e8036ca07f24f19673565536d356f21e1c424766cb.jpg", + "image_caption": [ + "Figure 1. Overview of VBench. We propose VBench, a comprehensive benchmark suite for video generative models. We design a comprehensive and hierarchical Evaluation Dimension Suite to decompose \"video generation quality\" into multiple well-defined dimensions to facilitate fine-grained and objective evaluation. For each dimension and each content category, we carefully design a Prompt Suite as test cases, and sample Generated Videos from a set of video generation models. For each evaluation dimension, we specifically design an Evaluation Method Suite, which uses a carefully crafted method or designated pipeline for automatic objective evaluation. We also conduct Human Preference Annotation for the generated videos for each dimension and show that VBench evaluation results are well aligned with human perceptions. VBench can provide valuable insights from multiple perspectives." + ], + "image_footnote": [], + "bbox": [ + 81, + 333, + 439, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6d763e4a79961e303efc773cb5e13ede49ac8e6e10f18e9aabe77fa894557da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 332, + 888, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 667, + 313, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBenchmark, a comprehensive benchmark suite that dissects \"video generation quality\" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBenchmark has three appealing properties: 1) Comprehensive Dimensions: VBenchmark comprises", + "bbox": [ + 75, + 693, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "16 dimensions in video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc.). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VBench to drive forward the field of video generation.", + "bbox": [ + 500, + 669, + 890, + 896 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contributions. corresponding authors. Code is available", + "bbox": [ + 99, + 886, + 450, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21807", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3d0229a0445214c16208b34564756131840011b11ee92cb4c8960824875f840.jpg", + "image_caption": [ + "Figure 2. VBench Evaluation Results of Video Generative Models. We visualize the evaluation results of four video generation models in 16 VBench dimensions. We normalize the results per dimension for clearer comparisons. For comprehensive numerical results, please refer to Table 1." + ], + "image_footnote": [], + "bbox": [ + 133, + 56, + 410, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 356, + 209, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image generation models have made rapid progress in the past few years, such as Variational Autoencoders (VAEs) [51], Generative Adversarial Networks (GANs) [6, 19, 20, 24, 39, 41, 44-47, 67], vector quantized (VQ) based approaches [16, 40, 85], and diffusion models [31, 78, 79]. This fuels recent explorations in video generation [3, 5, 10, 21, 23, 27, 29, 34, 52, 62, 66, 77, 86, 87, 89, 91, 101, 107-109], which goes beyond static imagery and models the dynamics and kinematics of real-world scenes. With the growth of video generation models, there arises a critical need for effective evaluation methods. The evaluation should be able to accurately reflect human perception of generated videos, providing reliable measures of a model's performance. Additionally, it should reflect each model's specific strengths and weaknesses, offering insights that inform the data, training, and architectural choices of future video generation models.", + "bbox": [ + 75, + 382, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing metrics for video generation such as Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83, 84], and CLIPSIM [73] are inconsistent with human judgement [15, 69]. Meanwhile, the Video Quality Assessment (VQA) methods [55, 82, 93-99] are primarily designed for real videos, thereby neglecting the unique challenges posed by generative models, such as artifacts in synthesized videos. Hence, there is a pressing need for an evaluation framework that aligns closely with human perception, and specifically designed for the characteristics of video generation models.", + "bbox": [ + 75, + 641, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce VBench, a comprehensive benchmark suite for evaluating video generation model performance. VBench has three appealing properties: 1) comprehensive evaluation dimensions, 2) human alignment, and 3) valuable insights.", + "bbox": [ + 75, + 808, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, our framework includes an evaluation dimension", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "suite that employs a hierarchical and disentangled approach to the decomposition of \"video generation quality\". This suite systematically breaks down the evaluation into two primary dimensions at a coarse level: Video Quality and Video-Condition Consistency. Each of these dimensions is further subdivided into more granular criteria. This hierarchical separation ensures that each dimension isolates and evaluates a single aspect of video quality, without interference from other variables, as illustrated in Figure 1. Recognizing video generation's unique challenges, we have tailored evaluation dimensions to its specific characteristics. For example, in terms of Video Quality, maintaining consistent subject identity (e.g., a teddy bear) in generated videos is crucial, and is a problem rarely encountered in real-world videos. Additionally, Video-Condition Consistency is vital for conditional video generation tasks, requiring its dedicated evaluation criteria. For each evaluation dimension, we carefully prepared around 100 text prompts as test cases for text-to-video (T2V) generation, and devised specialized evaluation methods tailored to each dimension. In addition to multi-dimensional evaluations, we also assess T2V models across diverse content categories. We organized prompt suites for eight distinct types, such as animal, architecture, human, and scenery, allowing for a separate evaluation within each category. This exploration reveals variable competencies in T2V generation across different content types, highlighting areas of proficiency and those requiring further enhancement.", + "bbox": [ + 496, + 90, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we systematically demonstrate that our evaluation method suite is closely aligned with human perception in every fine-grained evaluation dimension. We collected human preference annotations for each dimension. Specifically, we use various T2V models to sample videos from our prompt suites. Then given two videos sampled from the same prompt, we ask human annotators to indicate preferences according to each VBenchmark dimension respectively. We show that VBenchmark evaluations highly correlate with human preferences. Additionally, the human preference annotations can be utilized for multiple purposes, such as fine-tuning generation or evaluation models to enhance alignment with human perceptions. For instance, we utilize the annotations to implement Instruction Tuning within a Visual-Language Model (VLM), enhancing its T2V evaluation alignment with human preferences.", + "bbox": [ + 496, + 518, + 890, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Third, VBench's multi-dimensional and multicategorical approach can provide valuable insights to the video generation community. Our multi-dimensional system enables detailed feedback on the strengths and weaknesses of video generation models across various ability aspects. This approach not only ensures a comprehensive evaluation of existing models but also provides valuable insights into the training of advanced video generation models, guiding architectural and data choices", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21808", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for improved video generation outcomes. Additionally, VBench can be readily applied to evaluate image generation models, and thus we investigate the disparities between video and image generation models. In Section 5, we discuss in detail on various observations and insights drawn from VBench evaluations.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are open-sourcing VBench, including its evaluation dimension suite, evaluation method suite, prompt suite, generated videos, and the dataset of human preference annotations. We also encourage more video generation models to participate in the VBench challenge.", + "bbox": [ + 75, + 186, + 470, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 291, + 225, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Generative Models. Recently, diffusion models [14, 31, 78, 79] have achieved significant progress in image synthesis [25, 37, 38, 68, 71, 74, 75], and enabled a line of works towards video generation [5, 9, 22, 26, 28, 29, 32, 33, 42, 50, 62, 63, 77, 87, 91, 102, 103, 109, 113]. Many recent diffusion-based works [29, 62, 87, 91] are text-to-video (T2V) models. Other guidance modalities are also available, including image-to-video [11, 13, 17, 106], video-to-video [8, 59, 70, 72, 105], and a variety of control maps [12, 43, 50, 64, 90, 110, 111] such as pose, depth, and sketch. The boom of video generation models requires a comprehensive evaluation system to inform their current capabilities and guide future developments, and VBench takes the initiative in providing a comprehensive benchmark suite for fine-grained and human-aligned evaluation.", + "bbox": [ + 75, + 321, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation of Visual Generative Models. Existing video generation models typically use metrics like Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83], and CLIPSIM [73] for evaluation. The UCF-101 [80] dataset's class labels often serve as text prompts for IS, FID, and FVD, whereas MSR-VTT [104]'s human-labeled video captions are used for CLIPSIM. Despite covering various real-world scenarios, these prompts lack diversity and specificity, limiting accurate and fine-grained evaluation of video generation. For text-to-image (T2I) models, several benchmarks [2, 4, 35, 54, 65, 75, 88] are proposed to assess various capabilities like compositionality [35] and editing ability [4, 88]. However, video generative models still lack comprehensive evaluation benchmarks for detailed and human-aligned feedback. Our work differs from concurrent research [60, 61] in three key ways: 1) We have created 16 distinct evaluation dimensions, each with specialized prompts for precise assessment; 2) We have empirically validated that every dimension aligns closely with human perception; 3) Our multi-dimensional and multi-categorical evaluation offers valuable and comprehensive insights into video generation.", + "bbox": [ + 75, + 551, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VBench Suite", + "text_level": 1, + "bbox": [ + 498, + 89, + 638, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the main components of VBench. In Section 3.1, we present our rationale for designing the 16 evaluation dimensions, as well as each dimension's definition and evaluation method. We then elaborate on the prompt suites we use in Section 3.2. To validate VBench's alignment with human perception, we conduct human preference annotation for each dimension (see Section 3.3). The experiments and the insights drawn from VBench will be detailed in Section 4 and Section 5.", + "bbox": [ + 496, + 114, + 890, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Evaluation Dimension Suite", + "text_level": 1, + "bbox": [ + 498, + 258, + 750, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce our evaluation dimensions and their corresponding evaluation methods.", + "bbox": [ + 496, + 281, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing evaluation metrics like FVD [83] often conclude video generation model performance to a single number. This oversimplifies the evaluation and has several risks. First, a single number can obscure an individual model's strengths and weaknesses, and it fails to provide insights into specific areas where a model excels or underperforms. This makes it challenging to derive insights for future architectural and training designs based on single-valued metrics. Second, the notion of \"high-quality video generation\" is complex and multifaceted, with individuals prioritizing different video attributes based on the intended application. For instance, some may prioritize the absence of temporal flickering, while others may consider fidelity to the text prompt as the most significant, with less emphasis on flickering. Therefore, in contrast with performing single-valued evaluations of video generation quality, we propose a disaggregated approach by decomposing the brand notion of \"video generation performance\" into multiple discrete dimensions for fine-grained evaluation.", + "bbox": [ + 496, + 311, + 892, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we break \"video generation quality\" down into 16 disentangled dimensions in a top-down manner, with each evaluation dimension assessing one aspect of video generation quality. On the top level, we evaluate T2V performance from two broad perspectives: 1) Video Quality — \"Without considering alignment with the text prompt, does the video alone look good?\", which focuses on the perceptual quality of the synthesized video, and does not consider the input condition (e.g., text prompt), and 2) Video-Condition Consistency — \"Is the video consistent with what the user wants to generate?\", which focuses on whether the synthesized video is consistent with the guiding condition that the user provides (e.g., the text prompt for T2V generation). Under both \"Video Quality\" and \"Video-Condition Consistency\", we further break the coarse-grained dimensions into more fine-grained dimensions, as shown in Figure 1. We briefly introduce each dimension here. Please refer to the Supplementary File for the detailed definition and evaluation method of each dimension.", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21809", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Video Quality", + "text_level": 1, + "bbox": [ + 76, + 90, + 227, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We split \"Video Quality\" into two disentangled aspects, \"Temporal Quality\" and \"Frame-Wise Quality\", where the former only considers the cross-frame consistency and dynamics, and the latter only considers the quality of each individual frame without taking temporal quality into concern. For \"Temporal Quality\", we further devise five evaluation dimensions, where each focusing on a different aspect of temporal quality.", + "bbox": [ + 76, + 116, + 467, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Quality - Subject Consistency. For a subject (e.g., a person, a car, or a cat) in the video, we assess whether its appearance remains consistent throughout the whole video. To this end, we calculate the DINO [7] feature similarity across frames.", + "bbox": [ + 76, + 239, + 467, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Quality - Background Consistency. We evaluate the temporal consistency of the background scenes by calculating CLIP [73] feature similarity across frames.", + "bbox": [ + 76, + 315, + 467, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Quality - Temporal Flickering. Generated videos can exhibit imperfect temporal consistency at local and high-frequency details. We take static frames and compute the mean absolute difference across frames.", + "bbox": [ + 76, + 362, + 467, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Quality - Motion Smoothness. Subject Consistency and Background Consistency focus on temporal consistency of the \"look\" instead of the smoothness of \"movement and motion\". We believe it is important to evaluate whether the motion in the generated video is smooth, and follows the physical law of the real world. We utilize the motion priors in the video frame interpolation model [58] to evaluate the smoothness of generated motions.", + "bbox": [ + 76, + 422, + 467, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Quality - Dynamic Degree. Since a completely static video can score well in the aforementioned temporal quality dimensions, it is important to also evaluate the degree of dynamics (i.e., whether it contains large motions) generated by each model. We use RAFT [81] to estimate the degree of dynamics in synthesized videos.", + "bbox": [ + 76, + 545, + 467, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame-Wise Quality - Aesthetic Quality. We evaluate the artistic and beauty value perceived by humans towards each video frame using the LAION aesthetic predictor [53]. It can reflect aesthetic aspects such as the layout, the richness and harmony of colors, the photo-realism, naturalness, and artistic quality of the video frames.", + "bbox": [ + 76, + 637, + 467, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame-Wise Quality - Imaging Quality. Imaging quality refers to the distortion (e.g., over-exposure, noise, blur) presented in the generated frames, and we evaluate it using the MUSIQ [49] image quality predictor trained on the SPAQ [18] dataset.", + "bbox": [ + 76, + 729, + 467, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Video-Condition Consistency", + "text_level": 1, + "bbox": [ + 76, + 829, + 333, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We mainly dissect \"Video-Condition Consistency\" into \"Semantics\" (i.e., the type of the entities and their attributes) and \"Style\" (i.e., whether the generated video", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is consistent with user-requested style), with each decomposed into more fine-grained dimensions.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Object Class. We use GRiT [100] to detect the success rate of generating the specific class of objects depicted in the text prompt.", + "bbox": [ + 498, + 121, + 888, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Multiple Objects. Other than generating a single object of a particular class, the ability to compose multiple objects from different classes in the same frame is also an essential ability in video generation. We detect the success rate of generating all the objects specified in the text prompt within each video frame.", + "bbox": [ + 498, + 167, + 888, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Human Action. Human action is an important aspect in human-centric video generation. We apply UMT [57] to evaluate whether human subjects in generated videos can accurately execute the specific actions mentioned in the text prompts.", + "bbox": [ + 498, + 258, + 888, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Color. To evaluate whether synthesized object colors align with the text prompt, we use GRiT [100] for color captioning and comparison with expected colors.", + "bbox": [ + 498, + 334, + 888, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Spatial Relationship. Other than classes and attributes of synthesized objects, we also evaluate whether their spatial relationship follows what is specified by the text prompt. We focus on four primary types of spatial relationships, and perform rule-based evaluation similar to [35].", + "bbox": [ + 498, + 380, + 888, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics - Scene. We need to evaluate whether the synthesized video is consistent with the intended scene described by the text prompt. For example, when prompted \"ocean\", the generated video should be \"ocean\" instead of \"river\". We use Tag2Text [36] to caption the generated scenes, and then check its correspondence with scene descriptions in the text prompt.", + "bbox": [ + 498, + 455, + 888, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style - Appearance Style. Apart from semantics consistency with the text prompt, another important pillar in video-condition consistency is style. There are many styles that alter the look, color, and texture of synthesized video frames, such as \"oil painting style\", \"black and white style\", \"watercolor painting style\", \"cyberpunk style\", \"black and white\" etc. We calculate the CLIP [73] feature similarity between synthesized frames and these style descriptions.", + "bbox": [ + 498, + 563, + 888, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style - Temporal Style. Apart from appearance styles, videos also have temporal styles like various camera motions. We use ViCLIP [92] to calculate the video feature and the temporal style description feature similarity to reflect temporal style consistency.", + "bbox": [ + 498, + 684, + 888, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall Consistency. We further use overall video-text consistency computed by ViCLIP [92] on general text prompts as an aiding metric to reflect both semantics and style consistency.", + "bbox": [ + 498, + 758, + 888, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Prompt Suite", + "text_level": 1, + "bbox": [ + 500, + 830, + 638, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sampling procedure of current diffusion-based video generation models [29, 87, 91] is time-consuming (e.g., 90 seconds per video for LaVie [91], and more than 2 minutes", + "bbox": [ + 498, + 854, + 888, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21810", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2061c7ebb5579f224dfebbfe68f8a8fb80e68f609e1dc4b211d1fc03181a7bea.jpg", + "image_caption": [ + "Figure 3. Prompt Suite Statistics. The two graphs provide an overview of our prompt suites. Left: the word cloud to visualize word distribution of our prompt suites. Right: the number of prompts across different evaluation dimensions and different content categories.", + "Figure 4. Interface for Human Preference Annotation. Top: prompt and question. Right: choices that annotators can make. Bottom left: control for stop and playback." + ], + "image_footnote": [], + "bbox": [ + 78, + 61, + 604, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/736f108c33ad8d4ce70b2d30beae340d6d8ebff7191df1d985ba75fa14571dc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 61, + 885, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "per video for CogVideo [34]). Therefore, we need to control the amount of test cases for efficient evaluation. Meanwhile, we need to maintain the diversity and comprehensiveness of our prompt suite, so we design compact yet representative prompts in terms of both the evaluation dimensions and the content categories. We visualize our prompt suite distributions in Figure 3.", + "bbox": [ + 75, + 267, + 472, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt Suite per Dimension. For each VBenchmark evaluation dimension, we carefully designed a suite of around 100 prompts as test cases. The prompt suite is carefully curated to probe the specific ability relevant to the dimension tested. For example, for the \"Subject Consistency\" dimension which aims to evaluate the consistency of subjects' appearances throughout the video, we ensure every prompt has a movable subject (e.g., animals or vehicles) performing non-static actions, where their consistency might be compromised due to inconsistency introduced by their movements or changing locations. In \"Object Class\" dimension, we ensure the existence of a specific class of object in every prompt. For \"Human Action\", each test prompt contains a human subject performing a well-defined action from the Kinetics-400 dataset [48], where 100 representative actions are selected with minimal semantic overlaps among themselves. Please refer to the Supplementary File for the design rationale of the prompt suite for all 16 dimensions.", + "bbox": [ + 75, + 380, + 472, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt Suite per Category. When designing prompts for each dimension, the focus was to showcase models' ability in that specific dimension. We further incorporate prompt suites for eight content categories to provide insights into the performance across varied content types. To this end, we prepare a collection of human-curated prompts from the Internet and divide them into 8 distinctive categories following YouTube's categorization. Subsequently, we feed both the category labels and prompts into a Large Language Model (LLM) [112] (see more implementation details in Supplementary File), obtaining multi-label outputs for each caption. We select 800 prompts and manually clean their labels to serve as per-category prompt suites. Finally, we obtain 100 prompts for each of these eight categories: Animal, Architecture, Food, Human, Lifestyle, Plant, Scenery, and Vehicles.", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Human Preference Annotation", + "text_level": 1, + "bbox": [ + 500, + 267, + 772, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform human preference labeling on massive generated videos. The primary goal is to validate VBench evaluation's alignment with human perception in each of the 16 evaluation dimensions, and the verification results will be detailed in Section 4.2. We also show that our human preference annotations can be useful in future tasks of finetuning generation and evaluation models to enhance alignment with human perceptions.", + "bbox": [ + 496, + 291, + 893, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Preparation. Given a text prompt $p_i$ , and four video generation models to be evaluated $\\{A, B, C, D\\}$ , we use each model to generate a video, forming a \"group\" of videos $G_{i,j} = \\{V_{i,A,j}, V_{i,B,j}, V_{i,C,j}, V_{i,D,j}\\}$ . For each prompt $p_i$ , we sample five such groups of videos $\\{G_{i,0}, G_{i,1}, G_{i,2}, G_{i,3}, G_{i,4}\\}$ . For each group, we pair the videos up in pair-wise combinations, yielding six pairs: $(V_A, V_B)$ , $(V_A, V_C)$ , $(V_A, V_D)$ , $(V_B, V_C)$ , $(V_B, V_D)$ , $(V_C, V_D)$ , and ask human annotators to indicate their preferred video for each pair. Within the VBench evaluation framework, a prompt suite of $N$ prompts produces $N \\times 5 \\times 6$ pairwise video comparisons. The video order within each pair is randomized to ensure unbiased annotation.", + "bbox": [ + 496, + 414, + 893, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Labeling Rules. Specifically, the human annotators are asked to only consider the specific evaluation dimension of interest and select the preferred video. For example, in Figure 4, for the Appearance Style dimension, the question is \"Is the style of the video in the Van Gogh style?\", and human annotators are instructed to only focus on whether the generated video's style belongs to the Van Gogh style and should not consider other quality aspects of the generated video, such as potential issues like the degree of temporal flickering. In the example in this figure, video A resembles the Van Gogh better than video B, and the annotator is expected to select \"A is better\". For every dimension, we carefully prepare instructions and train the human annotators to understand the definition of the dimension, and perform multiple quality assurance protocols via a pre-labeling trial, and two rounds of post-labeling checks", + "bbox": [ + 496, + 611, + 895, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotations for VLM Tuning. We map VBench evaluation scores from various dimensions to the scale of 0-10 and combine them with human preference annotations to", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21811", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b437b1a8fb969d649bfa861638817df4cce4344633d911b4c86da6bfe2afb0e2.jpg", + "table_caption": [ + "Table 1. VBenchmark Evaluation Results per Dimension. This table compares the performance of four video generation models across each of the 16 VBenchmark dimensions. A higher score indicates relatively better performance for a particular dimension. We also provide two specially built baselines, i.e., Empirical Min and Max (the approximated achievable min and max scores for each dimension), as references." + ], + "table_footnote": [], + "table_body": "
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessDynamic DegreeAesthetic QualityImaging QualityObject Class
LaVie [91]91.41%97.47%98.30%96.38%49.72%54.94%61.90%91.82%
ModelScope [62, 87]89.87%95.29%98.28%95.79%66.39%52.06%58.57%82.25%
VideoCrafter [29]86.24%92.88%97.60%91.79%89.72%44.41%57.22%87.34%
CogVideo [34]92.19%96.20%97.64%96.47%42.22%38.18%41.03%73.40%
Empirical Min14.62%26.15%62.93%70.60%0.00%0.00%0.00%0.00%
Empirical Max100.00%100.00%100.00%99.75%100.00%100.00%100.00%100.00%
ModelsMultiple ObjectsHuman ActionColorSpatial RelationshipSceneAppearance StyleTemporal StyleOverall Consistency
LaVie [91]33.32%96.80%86.39%34.09%52.69%23.56%25.93%26.41%
ModelScope [62, 87]38.98%92.40%81.72%33.68%39.26%23.39%25.37%25.67%
VideoCrafter [29]25.93%93.00%78.84%36.74%43.36%21.57%25.42%25.21%
CogVideo [34]18.11%78.20%79.57%18.24%28.24%22.01%7.80%7.70%
Empirical Min0.00%0.00%0.00%0.00%0.00%0.09%0.00%0.00%
Empirical Max100.00%100.00%100.00%100.00%82.22%28.55%36.40%36.40%
", + "bbox": [ + 200, + 109, + 767, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "form the instruction data, which is then used to fine-tune the pre-trained VideoChat [56] model to demonstrate improved evaluation capabilities. For implementation details and tuning results, please refer to the Supplementary File.", + "bbox": [ + 75, + 291, + 470, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 364, + 209, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the video generation models LaVie [91], ModelScope [62, 87], VideoCrafter [29], and CogVideo [34] for VBench evaluation, and more will be added as they become open-sourced. Details of the models and sampling procedures are in the Supplementary File.", + "bbox": [ + 75, + 385, + 470, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Per-Dimension Evaluation", + "text_level": 1, + "bbox": [ + 76, + 470, + 316, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For every dimension, we calculate the VBench scores using the evaluation method suite described in Section 3.1, and show the results using Figure 2 and Table 1. We additionally designed three reference baselines, namely Empirical Max, Empirical Min, and WebVid-Avg. The first two approximate the maximum / minimum scores that videos might be able to achieve, and WebVid-Avg reflects the WebVid-10M [1] dataset quality in each VBench dimension.", + "bbox": [ + 75, + 488, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Empirical Max. For most dimensions, to approximate the maximum achievable values, we first retrieve WebVid-10M [1] videos according to our prompt suites. We use CLIP [73] to extract text features of both WebVid-10M's captions and our prompts. For each prompt, we retrieve the top-5 WebVid-10M videos according to text feature similarity with the given prompt. Given that the generated videos are usually 2 seconds in length, we randomly select a 2-second segment from each retrieved video and sample frames at 8 frames per second (FPS). For each dimension, we use the retrieved videos according to its prompt suite and report the highest-scoring video's result as Empirical Max.", + "bbox": [ + 75, + 611, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Empirical Min. To approximate the minimum achievable values, we use randomly generated 2-second Gaussian noise clips to calculate results for the \"Video-Condition Consistency\" dimensions. For most \"Video Quality\" dimensions, we select frames from real videos and design frame concatenation for each dimension, approximating the minimum score achievable for each VBench dimension.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WebVid-Avg. Similar to Empirical Max, we compute the average for each dimension on retrieved WebVid-10M [1] videos. This baseline could reflect the average per-dimension quality of the commonly used video generation training dataset WebVid-10M, and provide a reference for model performances. The comparison against WebVid-Avg and Empirical Max is visualized in Figure 6 (b).", + "bbox": [ + 496, + 291, + 890, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Validating Human Alignment of VBench", + "text_level": 1, + "bbox": [ + 498, + 402, + 846, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate that our evaluation method can faithfully reflect human perception, we performed a large-scale human annotation for each dimension, as mentioned in Section 3.3. We show the correlation between VBenchmark evaluation results and human preference annotations in Figure 5.", + "bbox": [ + 496, + 420, + 890, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Win Ratio. Given the human labels, we calculate the win ratio of each model. During pairwise comparisons, if a model's video is selected as better, then the model scores 1 and the other model scores 0. If there is a tie, then both models score 0.5. For each model, the win ratio is calculated as the total score divided by the total number of pairwise comparisons participated.", + "bbox": [ + 496, + 496, + 890, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Per-Dimension Evaluation. For each evaluation dimension, we calculate the model win ratio based on (1) VBenchmark results, and (2) human annotation results, respectively, and compute their correlations, as shown in Figure 5. We observe that VBenchmark's per-dimension evaluation results are highly correlated with human preference annotations.", + "bbox": [ + 496, + 603, + 890, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Per-Category Evaluation", + "text_level": 1, + "bbox": [ + 498, + 699, + 728, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the T2V models across eight different content categories, by generating videos based on Prompt Suite per Category described in Section 3.2, and then calculating their performance across different evaluation dimensions. Figure 7 visualizes the evaluation results of each model in terms of the eight content categories.", + "bbox": [ + 496, + 717, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Video Generation V.S. Image Generation", + "text_level": 1, + "bbox": [ + 498, + 819, + 849, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comparative analysis of the frame-wise generation capability exhibited by text-to-video (T2V) models and text-to-image (T2I) models with two primary objectives: first, to assess the extent to which T2V models", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21812", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18e1edf20a57807908f97ff5944543f3b40659999ea81933f208949b94d8926c.jpg", + "image_caption": [ + "Figure 5. Validate VBench's Human Alignment. Our experiments show that VBench evaluations across all dimensions closely match human perceptions. Each plot shows the alignment verification result of a specific VBench dimension. In each plot, a dot represents the human preference win ratio (horizontal axis) and VBench evaluation win ratio (vertical axis) for a particular video generation model. We linearly fit a straight line to visualize the correlation, and calculate the Spearman's correlation coefficient $(\\rho)$ for each dimension." + ], + "image_footnote": [], + "bbox": [ + 83, + 65, + 883, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/271dbb03f501abe078ec14e0f96007705aa2898b4577a30f243fea91fd1d8dd0.jpg", + "image_caption": [ + "Figure 6. More Comparisons of Video Generation Models with Other Models and Baselines. We use VBench to evaluate other models and baselines for further comparative analysis of T2V models. (a) Comparison with text-to-image (T2I) generation models. (b) Comparison with WebVid-Avg and Empirical Max baselines. See the Supplementary File for comprehensive numerical results and details on normalization methods." + ], + "image_footnote": [], + "bbox": [ + 78, + 303, + 473, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "have successfully inherited the frame-wise generative capability of the T2I models; and second, to investigate the frame-wise generation capability gap between existing T2I and T2V models. As an initial exploration into this problem, we compare video generation models with three image generation models, namely Stable Diffusion (SD) 1.4 [74], SD2.1 [74], and SDXL [71]. We choose 10 V-Bench dimensions that can encompass frame-wise generation capabilities, and sample frames from all the image and video generation models according to Prompt Suite per Evaluation Dimension described in Section 3.2. Figure 6 (a) visualizes the evaluation results of T2V versus T2I models.", + "bbox": [ + 75, + 570, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Insights and Discussions", + "text_level": 1, + "bbox": [ + 76, + 760, + 307, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss the observations and insights we draw from our comprehensive evaluation experiments.", + "bbox": [ + 75, + 779, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Trade-off across Ability Dimensions. We have noticed a trade-off in video generation models between 1) temporal consistency (Subject Consistency, Background Consistency, Temporal Flickering, Motion Smoothness) and 2) Dynamic Degree. Models strong in temporal consistency often have a lower Dynamic Degree, as these two aspects are some", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "what complementary (see Figure 2 and Table 1). For example, LaVie excels in Background Consistency and Temporal Flickering but has a low Dynamic Degree, probably because generating relatively static scenes can \"cheat\" to get high temporal consistency scores. Conversely, VideoCrafter shows a high Dynamic Degree but suffers from poor performance in all temporal consistency dimensions. This trend highlights the current challenge for models to achieve temporal consistency with dynamic content of large motions. Future research should focus on enhancing both aspects simultaneously, as improving only one might indicate compromising the other.", + "bbox": [ + 496, + 306, + 890, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Uncovering Hidden Potential of T2V Models in Specific Content Categories. Our analysis reveals that the capabilities of some models vary significantly across different content types. For instance, for Aesthetic Quality, CogVideo scores well for Food (see Figure 7 rightmost chart), whereas it underperforms in others like Animal and Vehicles. The average results across various prompts might suggest a lower overall \"Aesthetic Quality\" (as seen in Figure 2), but CogVideo demonstrates relatively strong aesthetics in at least the Food category. This suggests that with tailored training data and strategies, CogVideo could potentially match other models in aesthetics by improving such ability in other content types. Therefore, we recommend evaluating video generation models not just based on ability dimensions but also considering specific content categories to uncover their hidden potential.", + "bbox": [ + 496, + 489, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Bottleneck in Temporally Complex Categories Affecting Spatial and Temporal Performance. For spatially complex categories (e.g., Animal, LifeStyle, Human, Vehicles), models all perform relatively poorly mainly in Aesthetic Quality (shown in Figure 7). This is likely due to the challenges in synthesizing harmonious color schemes, articulated structures, and appealing layouts amidst complex elements. On the other hand, for categories involving complex and intense motions like Human and Vehicle (see their Dynamic Degree in Supplementary File), performance is relatively poor across all dimen", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21813", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72706807e6048bcc1277c117c1ec346373f4f827e80256234c39f52dface877d.jpg", + "image_caption": [ + "Figure 7. VBench Results across Eight Content Categories (best viewed in color). For each chart, we plot the VBench evaluation results across eight different content categories, benchmarked by our Prompt Suite per Category. The results are linearly normalized between 0 and 1 for better visibility across categories. See the Supplementary File for comprehensive numerical results, and normalization details." + ], + "image_footnote": [], + "bbox": [ + 78, + 60, + 890, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sions. This suggests that motion complexity and dynamic intensity significantly hinder synthesis, impacting both spatial and temporal dimensions, probably because poor temporal modeling results in distorted and blurred imagery. This highlights the need for improved handling of dynamic motions in video generation models.", + "bbox": [ + 75, + 277, + 468, + 368 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Challenges of Data Quantity in Handling Complex Categories like Human. The WebVid-10M dataset [1] allocates $26\\%$ of its content to the Human category, which is the largest share among the eight categories (see statistics in Supplementary File). However, the Human category exhibits one of the poorest results among eight categories (see Figure 7). This suggests that merely increasing data volume may not significantly enhance performance in complex categories like Human. A potential approach could involve integrating human-related priors or controls, such as skeletons, to better capture the articulated nature of human appearances and movements.", + "- Prioritizing Data Quality Over Quantity in Large-Scale Datasets. For Aesthetic Quality, Figure 7 shows that the Food category almost always tends to have the highest scores among all categories. This is corroborated by the WebVid-10M dataset [1], where Food ranks highest in Aesthetic Quality according to VBench evaluation (refer to Supplementary File for more details), despite comprising just $11\\%$ of the total data. This observation suggests that at million scales, data quality might hold greater importance than quantity. Furthermore, VBench's evaluation dimensions can be potentially useful for cleaning datasets in specified quality dimensions.", + "- Compositionality: T2I versus T2V. As shown in Figure 6 (a), T2V models significantly underperform in Multiple Objects and Spatial Relationship compared to T2I models (especially SDXL [71]), which highlights the need to enhance compositionality (i.e., correctly composing multiple objects in the same frame). We believe possible solutions might be: 1) curating training data incorporating multiple objects with corresponding captions explicitly depicting this compositionality, or 2) adding intermediate spatial control modules or modalities during video synthesis. Furthermore, the disparity of the text encoders might also account for the per" + ], + "bbox": [ + 75, + 369, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance gap. As T2I models leverage bigger (OpenCLIP ViT-H for SD2.1 [74]) or more sophisticated (CLIP ViT-L & OpenCLIP ViT-G for SDXL [71]) text encoders compared with T2V models (e.g., CLIP ViT-L alone for LaVie), more representative text embeddings could be featuring more accurate object composition comprehension.", + "bbox": [ + 496, + 277, + 893, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 383, + 619, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the growing focus on video generation, comprehensive evaluation of these models is essential to assess current advancements and guide future research. In this work, we take the first step forward and propose VBench, a comprehensive benchmark suite for evaluating video generation models. With its multi-dimensional, human-aligned, and insight-rich properties, VBench could play vital roles for evaluating future video generation models and inspiring further advancements in video generation. We believe that VBench is a significant contribution to the video generation and evaluation community.", + "bbox": [ + 496, + 402, + 890, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. We plan to extend VBenchmark to include more models that recently became available, and extend the evaluations aspects to additional video generation tasks, like image-to-video.", + "bbox": [ + 496, + 568, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Negative Societal Impacts. We also recognize the importance of considering ethical aspects in future iterations of VBench. While VBench currently does not assess safety and equality dimensions, we urge users to exercise caution with open-sourced video generation models.", + "bbox": [ + 496, + 628, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s); the National Key R&D Program of China (2022ZD0160201); National Natural Science Foundation of China (62076119, 61921006, 62102150); the Science and Technology Commission of Shanghai Municipality (23YF1461900, 23QD1400800); and Shanghai Artificial Intelligence Laboratory. We thank Shangchen Zhou, Jianyi Wang, and Ruicheng Feng for helpful suggestions.", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21814", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6, 8", + "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 3", + "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2", + "[4] Samyadeep Basu, Mehrdad Saberi, Shweta Bhardwaj, Atoosa Malemir Chegini, Daniela Massiceti, Maziar Sanjabi, Shell Xu Hu, and Soheil Feizi. Editval: Benchmarking diffusion based text-guided image editing methods. arXiv preprint arXiv:2310.02426, 2023. 3", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3", + "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 4", + "[8] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3", + "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3", + "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 2", + "[11] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 3", + "[12] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3", + "[13] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 3", + "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 3" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2", + "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[17] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3", + "[18] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 4", + "[19] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2", + "[20] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In ICCV, 2023. 2", + "[21] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2", + "[22] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3", + "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2", + "[24] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[25] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3", + "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 3", + "[27] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2", + "[28] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3", + "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21815", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 3, 4, 6", + "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 3", + "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3", + "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3", + "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3", + "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 5, 6", + "[35] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350, 2023. 3, 4", + "[36] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 4", + "[37] Ziqi Huang, Kelvin C.K. Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In CVPR, 2023. 3", + "[38] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C.K. Chan, and Ziwei Liu. ReVersion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3", + "[39] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 2", + "[40] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Textdriven controllable human image generation. ACM TOG, 2022. 2", + "[41] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE TPAMI, 2023. 2", + "[42] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2Performer: Textdriven human video generation. In ICCV, 2023. 3", + "[43] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 3", + "[44] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2" + ], + "bbox": [ + 86, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019.", + "[46] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020.", + "[47] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2", + "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5", + "[49] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: multi-scale image quality transformer. CoRR, abs/2108.05997, 2021. 4", + "[50] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3", + "[51] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[52] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2", + "[53] LAION-AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor, 2022.4", + "[54] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. arXiv preprint arXiv:2311.04287, 2023. 3", + "[55] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, 2019. 2", + "[56] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6", + "[57] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058, 2023. 4", + "[58] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 4", + "[59] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhong-cong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3", + "[60] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. In NeurIPS, 2023. 3" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21816", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In CVPR, 2024. 3", + "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 3, 6", + "[63] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3", + "[64] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3", + "[65] Kai Zhang Yujie Lu Xingyu Fu Wenwen Zhuang Wenhu Chen Max Ku, Tianle Li. Imagenhub: Standardizing the evaluation of conditional image generation models. In ICLR, 2024. 3", + "[66] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. arXiv preprint arXiv:2402.14797, 2024. 2", + "[67] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2", + "[68] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3", + "[69] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In CVPR, 2023. 2", + "[70] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 3", + "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8", + "[72] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3", + "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 6", + "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 86, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In CVPR, 2022. 3, 7, 8", + "[75] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3", + "[76] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2, 3", + "[77] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2.3", + "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3", + "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3", + "[80] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3", + "[81] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4", + "[82] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C. Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE TIP, 2021. 2", + "[83] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 3", + "[84] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In ICLRW, 2019. 2", + "[85] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2", + "[86] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2", + "[87] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3, 4, 6", + "[88] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and EditBench: Advancing and evaluating text-guided" + ], + "bbox": [ + 509, + 90, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21817", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image inpainting. arXiv preprint arXiv:2212.06909, 2022.3", + "[89] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, et al. Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2", + "[90] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3", + "[91] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3, 4, 6", + "[92] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 4", + "[93] Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fastvqa: Efficient end-to-end video quality assessment with fragment sampling. In ECCV, 2022. 2", + "[94] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin. Neighbourhood representative sampling for efficient end-to-end video quality assessment. arXiv preprint arXiv:2210.05357, 2022.", + "[95] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, and Weisi Lin. Discovqa: Temporal distortion-content transformers for video quality assessment. IEEE TCSVT, 2023.", + "[96] Haoning Wu, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring opinion-unaware video quality assessment with semantic affinity criterion. In ICME, 2023.", + "[97] Haoning Wu, Liang Liao, Annan Wang, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards robust text-prompted semantic criterion for in-the-wild video quality assessment. arXiv preprint arXiv:2304.14672, 2023.", + "[98] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023.", + "[99] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards explainable video quality assessment: A database and a language-prompted approach. In ACM MM, 2023. 2", + "[100] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[101] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2", + "[102] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 3", + "[103] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 3", + "[104] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3", + "[105] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3", + "[106] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3", + "[107] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2", + "[108] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023.", + "[109] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3", + "[110] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. Magicavatar: Multimodal avatar generation and animation. arXiv preprint arXiv:2308.14748, 2023. 3", + "[111] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3", + "[112] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5", + "[113] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 867 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21818", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_model.json b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb158b53e2411a91dfc4073b658276d58e411e8a --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_model.json @@ -0,0 +1,2754 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.118, + 0.85, + 0.14 + ], + "angle": 0, + "content": "VBenchmark: Comprehensive Benchmark Suite for Video Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.167, + 0.881, + 0.224 + ], + "angle": 0, + "content": "Ziqi Huang\\(^{1*}\\) Yinan He\\(^{2*}\\) Jiashuo Yu\\(^{2*}\\) Fan Zhang\\(^{2*}\\) Chenyang Si\\(^{1}\\) Yuming Jiang\\(^{1}\\) \nYuanhan Zhang\\(^{1}\\) Tianxing Wu\\(^{1}\\) Qingyang Jin\\(^{1}\\) Nattapol Chanpaisit\\(^{1}\\) \nYaohui Wang\\(^{2}\\) Xinyuan Chen\\(^{2}\\) Limin Wang\\(^{4,2}\\) Dahua Lin\\(^{2,3\\boxtimes}\\) Yu Qiao\\(^{2\\boxtimes}\\) Ziwei Liu\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.239, + 0.842, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\)The Chinese University of Hong Kong \\(^{4}\\)Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.29, + 0.694, + 0.307 + ], + "angle": 0, + "content": "https://vchitect.github.io/VBench-project/" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.334, + 0.44, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.333, + 0.89, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.559, + 0.895, + 0.657 + ], + "angle": 0, + "content": "Figure 1. Overview of VBench. We propose VBench, a comprehensive benchmark suite for video generative models. We design a comprehensive and hierarchical Evaluation Dimension Suite to decompose \"video generation quality\" into multiple well-defined dimensions to facilitate fine-grained and objective evaluation. For each dimension and each content category, we carefully design a Prompt Suite as test cases, and sample Generated Videos from a set of video generation models. For each evaluation dimension, we specifically design an Evaluation Method Suite, which uses a carefully crafted method or designated pipeline for automatic objective evaluation. We also conduct Human Preference Annotation for the generated videos for each dimension and show that VBench evaluation results are well aligned with human perceptions. VBench can provide valuable insights from multiple perspectives." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.669, + 0.314, + 0.685 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBenchmark, a comprehensive benchmark suite that dissects \"video generation quality\" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBenchmark has three appealing properties: 1) Comprehensive Dimensions: VBenchmark comprises" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.67, + 0.892, + 0.897 + ], + "angle": 0, + "content": "16 dimensions in video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc.). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VBench to drive forward the field of video generation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.451, + 0.901 + ], + "angle": 0, + "content": "*equal contributions. corresponding authors. Code is available" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.057, + 0.411, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.472, + 0.342 + ], + "angle": 0, + "content": "Figure 2. VBench Evaluation Results of Video Generative Models. We visualize the evaluation results of four video generation models in 16 VBench dimensions. We normalize the results per dimension for clearer comparisons. For comprehensive numerical results, please refer to Table 1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.357, + 0.21, + 0.373 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Image generation models have made rapid progress in the past few years, such as Variational Autoencoders (VAEs) [51], Generative Adversarial Networks (GANs) [6, 19, 20, 24, 39, 41, 44-47, 67], vector quantized (VQ) based approaches [16, 40, 85], and diffusion models [31, 78, 79]. This fuels recent explorations in video generation [3, 5, 10, 21, 23, 27, 29, 34, 52, 62, 66, 77, 86, 87, 89, 91, 101, 107-109], which goes beyond static imagery and models the dynamics and kinematics of real-world scenes. With the growth of video generation models, there arises a critical need for effective evaluation methods. The evaluation should be able to accurately reflect human perception of generated videos, providing reliable measures of a model's performance. Additionally, it should reflect each model's specific strengths and weaknesses, offering insights that inform the data, training, and architectural choices of future video generation models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.808 + ], + "angle": 0, + "content": "However, existing metrics for video generation such as Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83, 84], and CLIPSIM [73] are inconsistent with human judgement [15, 69]. Meanwhile, the Video Quality Assessment (VQA) methods [55, 82, 93-99] are primarily designed for real videos, thereby neglecting the unique challenges posed by generative models, such as artifacts in synthesized videos. Hence, there is a pressing need for an evaluation framework that aligns closely with human perception, and specifically designed for the characteristics of video generation models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.47, + 0.884 + ], + "angle": 0, + "content": "To this end, we introduce VBench, a comprehensive benchmark suite for evaluating video generation model performance. VBench has three appealing properties: 1) comprehensive evaluation dimensions, 2) human alignment, and 3) valuable insights." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.9 + ], + "angle": 0, + "content": "First, our framework includes an evaluation dimension" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.514 + ], + "angle": 0, + "content": "suite that employs a hierarchical and disentangled approach to the decomposition of \"video generation quality\". This suite systematically breaks down the evaluation into two primary dimensions at a coarse level: Video Quality and Video-Condition Consistency. Each of these dimensions is further subdivided into more granular criteria. This hierarchical separation ensures that each dimension isolates and evaluates a single aspect of video quality, without interference from other variables, as illustrated in Figure 1. Recognizing video generation's unique challenges, we have tailored evaluation dimensions to its specific characteristics. For example, in terms of Video Quality, maintaining consistent subject identity (e.g., a teddy bear) in generated videos is crucial, and is a problem rarely encountered in real-world videos. Additionally, Video-Condition Consistency is vital for conditional video generation tasks, requiring its dedicated evaluation criteria. For each evaluation dimension, we carefully prepared around 100 text prompts as test cases for text-to-video (T2V) generation, and devised specialized evaluation methods tailored to each dimension. In addition to multi-dimensional evaluations, we also assess T2V models across diverse content categories. We organized prompt suites for eight distinct types, such as animal, architecture, human, and scenery, allowing for a separate evaluation within each category. This exploration reveals variable competencies in T2V generation across different content types, highlighting areas of proficiency and those requiring further enhancement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Second, we systematically demonstrate that our evaluation method suite is closely aligned with human perception in every fine-grained evaluation dimension. We collected human preference annotations for each dimension. Specifically, we use various T2V models to sample videos from our prompt suites. Then given two videos sampled from the same prompt, we ask human annotators to indicate preferences according to each VBenchmark dimension respectively. We show that VBenchmark evaluations highly correlate with human preferences. Additionally, the human preference annotations can be utilized for multiple purposes, such as fine-tuning generation or evaluation models to enhance alignment with human perceptions. For instance, we utilize the annotations to implement Instruction Tuning within a Visual-Language Model (VLM), enhancing its T2V evaluation alignment with human preferences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Third, VBench's multi-dimensional and multicategorical approach can provide valuable insights to the video generation community. Our multi-dimensional system enables detailed feedback on the strengths and weaknesses of video generation models across various ability aspects. This approach not only ensures a comprehensive evaluation of existing models but also provides valuable insights into the training of advanced video generation models, guiding architectural and data choices" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "for improved video generation outcomes. Additionally, VBench can be readily applied to evaluate image generation models, and thus we investigate the disparities between video and image generation models. In Section 5, we discuss in detail on various observations and insights drawn from VBench evaluations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.472, + 0.264 + ], + "angle": 0, + "content": "We are open-sourcing VBench, including its evaluation dimension suite, evaluation method suite, prompt suite, generated videos, and the dataset of human preference annotations. We also encourage more video generation models to participate in the VBench challenge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.292, + 0.227, + 0.308 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Video Generative Models. Recently, diffusion models [14, 31, 78, 79] have achieved significant progress in image synthesis [25, 37, 38, 68, 71, 74, 75], and enabled a line of works towards video generation [5, 9, 22, 26, 28, 29, 32, 33, 42, 50, 62, 63, 77, 87, 91, 102, 103, 109, 113]. Many recent diffusion-based works [29, 62, 87, 91] are text-to-video (T2V) models. Other guidance modalities are also available, including image-to-video [11, 13, 17, 106], video-to-video [8, 59, 70, 72, 105], and a variety of control maps [12, 43, 50, 64, 90, 110, 111] such as pose, depth, and sketch. The boom of video generation models requires a comprehensive evaluation system to inform their current capabilities and guide future developments, and VBench takes the initiative in providing a comprehensive benchmark suite for fine-grained and human-aligned evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Evaluation of Visual Generative Models. Existing video generation models typically use metrics like Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83], and CLIPSIM [73] for evaluation. The UCF-101 [80] dataset's class labels often serve as text prompts for IS, FID, and FVD, whereas MSR-VTT [104]'s human-labeled video captions are used for CLIPSIM. Despite covering various real-world scenarios, these prompts lack diversity and specificity, limiting accurate and fine-grained evaluation of video generation. For text-to-image (T2I) models, several benchmarks [2, 4, 35, 54, 65, 75, 88] are proposed to assess various capabilities like compositionality [35] and editing ability [4, 88]. However, video generative models still lack comprehensive evaluation benchmarks for detailed and human-aligned feedback. Our work differs from concurrent research [60, 61] in three key ways: 1) We have created 16 distinct evaluation dimensions, each with specialized prompts for precise assessment; 2) We have empirically validated that every dimension aligns closely with human perception; 3) Our multi-dimensional and multi-categorical evaluation offers valuable and comprehensive insights into video generation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.64, + 0.106 + ], + "angle": 0, + "content": "3. VBench Suite" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.251 + ], + "angle": 0, + "content": "In this section, we introduce the main components of VBench. In Section 3.1, we present our rationale for designing the 16 evaluation dimensions, as well as each dimension's definition and evaluation method. We then elaborate on the prompt suites we use in Section 3.2. To validate VBench's alignment with human perception, we conduct human preference annotation for each dimension (see Section 3.3). The experiments and the insights drawn from VBench will be detailed in Section 4 and Section 5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.259, + 0.75, + 0.274 + ], + "angle": 0, + "content": "3.1. Evaluation Dimension Suite" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.891, + 0.312 + ], + "angle": 0, + "content": "We first introduce our evaluation dimensions and their corresponding evaluation methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.893, + 0.599 + ], + "angle": 0, + "content": "Existing evaluation metrics like FVD [83] often conclude video generation model performance to a single number. This oversimplifies the evaluation and has several risks. First, a single number can obscure an individual model's strengths and weaknesses, and it fails to provide insights into specific areas where a model excels or underperforms. This makes it challenging to derive insights for future architectural and training designs based on single-valued metrics. Second, the notion of \"high-quality video generation\" is complex and multifaceted, with individuals prioritizing different video attributes based on the intended application. For instance, some may prioritize the absence of temporal flickering, while others may consider fidelity to the text prompt as the most significant, with less emphasis on flickering. Therefore, in contrast with performing single-valued evaluations of video generation quality, we propose a disaggregated approach by decomposing the brand notion of \"video generation performance\" into multiple discrete dimensions for fine-grained evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Specifically, we break \"video generation quality\" down into 16 disentangled dimensions in a top-down manner, with each evaluation dimension assessing one aspect of video generation quality. On the top level, we evaluate T2V performance from two broad perspectives: 1) Video Quality — \"Without considering alignment with the text prompt, does the video alone look good?\", which focuses on the perceptual quality of the synthesized video, and does not consider the input condition (e.g., text prompt), and 2) Video-Condition Consistency — \"Is the video consistent with what the user wants to generate?\", which focuses on whether the synthesized video is consistent with the guiding condition that the user provides (e.g., the text prompt for T2V generation). Under both \"Video Quality\" and \"Video-Condition Consistency\", we further break the coarse-grained dimensions into more fine-grained dimensions, as shown in Figure 1. We briefly introduce each dimension here. Please refer to the Supplementary File for the detailed definition and evaluation method of each dimension." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21809" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.092, + 0.228, + 0.108 + ], + "angle": 0, + "content": "3.1.1 Video Quality" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.117, + 0.468, + 0.239 + ], + "angle": 0, + "content": "We split \"Video Quality\" into two disentangled aspects, \"Temporal Quality\" and \"Frame-Wise Quality\", where the former only considers the cross-frame consistency and dynamics, and the latter only considers the quality of each individual frame without taking temporal quality into concern. For \"Temporal Quality\", we further devise five evaluation dimensions, where each focusing on a different aspect of temporal quality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.24, + 0.468, + 0.315 + ], + "angle": 0, + "content": "Temporal Quality - Subject Consistency. For a subject (e.g., a person, a car, or a cat) in the video, we assess whether its appearance remains consistent throughout the whole video. To this end, we calculate the DINO [7] feature similarity across frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.316, + 0.468, + 0.362 + ], + "angle": 0, + "content": "Temporal Quality - Background Consistency. We evaluate the temporal consistency of the background scenes by calculating CLIP [73] feature similarity across frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.468, + 0.422 + ], + "angle": 0, + "content": "Temporal Quality - Temporal Flickering. Generated videos can exhibit imperfect temporal consistency at local and high-frequency details. We take static frames and compute the mean absolute difference across frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.468, + 0.544 + ], + "angle": 0, + "content": "Temporal Quality - Motion Smoothness. Subject Consistency and Background Consistency focus on temporal consistency of the \"look\" instead of the smoothness of \"movement and motion\". We believe it is important to evaluate whether the motion in the generated video is smooth, and follows the physical law of the real world. We utilize the motion priors in the video frame interpolation model [58] to evaluate the smoothness of generated motions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.468, + 0.636 + ], + "angle": 0, + "content": "Temporal Quality - Dynamic Degree. Since a completely static video can score well in the aforementioned temporal quality dimensions, it is important to also evaluate the degree of dynamics (i.e., whether it contains large motions) generated by each model. We use RAFT [81] to estimate the degree of dynamics in synthesized videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.468, + 0.728 + ], + "angle": 0, + "content": "Frame-Wise Quality - Aesthetic Quality. We evaluate the artistic and beauty value perceived by humans towards each video frame using the LAION aesthetic predictor [53]. It can reflect aesthetic aspects such as the layout, the richness and harmony of colors, the photo-realism, naturalness, and artistic quality of the video frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.468, + 0.805 + ], + "angle": 0, + "content": "Frame-Wise Quality - Imaging Quality. Imaging quality refers to the distortion (e.g., over-exposure, noise, blur) presented in the generated frames, and we evaluate it using the MUSIQ [49] image quality predictor trained on the SPAQ [18] dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.334, + 0.845 + ], + "angle": 0, + "content": "3.1.2 Video-Condition Consistency" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We mainly dissect \"Video-Condition Consistency\" into \"Semantics\" (i.e., the type of the entities and their attributes) and \"Style\" (i.e., whether the generated video" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "is consistent with user-requested style), with each decomposed into more fine-grained dimensions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.89, + 0.168 + ], + "angle": 0, + "content": "Semantics - Object Class. We use GRiT [100] to detect the success rate of generating the specific class of objects depicted in the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.169, + 0.89, + 0.258 + ], + "angle": 0, + "content": "Semantics - Multiple Objects. Other than generating a single object of a particular class, the ability to compose multiple objects from different classes in the same frame is also an essential ability in video generation. We detect the success rate of generating all the objects specified in the text prompt within each video frame." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.89, + 0.335 + ], + "angle": 0, + "content": "Semantics - Human Action. Human action is an important aspect in human-centric video generation. We apply UMT [57] to evaluate whether human subjects in generated videos can accurately execute the specific actions mentioned in the text prompts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.89, + 0.38 + ], + "angle": 0, + "content": "Semantics - Color. To evaluate whether synthesized object colors align with the text prompt, we use GRiT [100] for color captioning and comparison with expected colors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.89, + 0.457 + ], + "angle": 0, + "content": "Semantics - Spatial Relationship. Other than classes and attributes of synthesized objects, we also evaluate whether their spatial relationship follows what is specified by the text prompt. We focus on four primary types of spatial relationships, and perform rule-based evaluation similar to [35]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.89, + 0.563 + ], + "angle": 0, + "content": "Semantics - Scene. We need to evaluate whether the synthesized video is consistent with the intended scene described by the text prompt. For example, when prompted \"ocean\", the generated video should be \"ocean\" instead of \"river\". We use Tag2Text [36] to caption the generated scenes, and then check its correspondence with scene descriptions in the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.89, + 0.684 + ], + "angle": 0, + "content": "Style - Appearance Style. Apart from semantics consistency with the text prompt, another important pillar in video-condition consistency is style. There are many styles that alter the look, color, and texture of synthesized video frames, such as \"oil painting style\", \"black and white style\", \"watercolor painting style\", \"cyberpunk style\", \"black and white\" etc. We calculate the CLIP [73] feature similarity between synthesized frames and these style descriptions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.89, + 0.76 + ], + "angle": 0, + "content": "Style - Temporal Style. Apart from appearance styles, videos also have temporal styles like various camera motions. We use ViCLIP [92] to calculate the video feature and the temporal style description feature similarity to reflect temporal style consistency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.76, + 0.89, + 0.821 + ], + "angle": 0, + "content": "Overall Consistency. We further use overall video-text consistency computed by ViCLIP [92] on general text prompts as an aiding metric to reflect both semantics and style consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.64, + 0.848 + ], + "angle": 0, + "content": "3.2. Prompt Suite" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.89, + 0.901 + ], + "angle": 0, + "content": "The sampling procedure of current diffusion-based video generation models [29, 87, 91] is time-consuming (e.g., 90 seconds per video for LaVie [91], and more than 2 minutes" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.063, + 0.605, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.61, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Prompt Suite Statistics. The two graphs provide an overview of our prompt suites. Left: the word cloud to visualize word distribution of our prompt suites. Right: the number of prompts across different evaluation dimensions and different content categories." + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.063, + 0.887, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.198, + 0.89, + 0.254 + ], + "angle": 0, + "content": "Figure 4. Interface for Human Preference Annotation. Top: prompt and question. Right: choices that annotators can make. Bottom left: control for stop and playback." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.473, + 0.375 + ], + "angle": 0, + "content": "per video for CogVideo [34]). Therefore, we need to control the amount of test cases for efficient evaluation. Meanwhile, we need to maintain the diversity and comprehensiveness of our prompt suite, so we design compact yet representative prompts in terms of both the evaluation dimensions and the content categories. We visualize our prompt suite distributions in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.653 + ], + "angle": 0, + "content": "Prompt Suite per Dimension. For each VBenchmark evaluation dimension, we carefully designed a suite of around 100 prompts as test cases. The prompt suite is carefully curated to probe the specific ability relevant to the dimension tested. For example, for the \"Subject Consistency\" dimension which aims to evaluate the consistency of subjects' appearances throughout the video, we ensure every prompt has a movable subject (e.g., animals or vehicles) performing non-static actions, where their consistency might be compromised due to inconsistency introduced by their movements or changing locations. In \"Object Class\" dimension, we ensure the existence of a specific class of object in every prompt. For \"Human Action\", each test prompt contains a human subject performing a well-defined action from the Kinetics-400 dataset [48], where 100 representative actions are selected with minimal semantic overlaps among themselves. Please refer to the Supplementary File for the design rationale of the prompt suite for all 16 dimensions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Prompt Suite per Category. When designing prompts for each dimension, the focus was to showcase models' ability in that specific dimension. We further incorporate prompt suites for eight content categories to provide insights into the performance across varied content types. To this end, we prepare a collection of human-curated prompts from the Internet and divide them into 8 distinctive categories following YouTube's categorization. Subsequently, we feed both the category labels and prompts into a Large Language Model (LLM) [112] (see more implementation details in Supplementary File), obtaining multi-label outputs for each caption. We select 800 prompts and manually clean their labels to serve as per-category prompt suites. Finally, we obtain 100 prompts for each of these eight categories: Animal, Architecture, Food, Human, Lifestyle, Plant, Scenery, and Vehicles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.773, + 0.283 + ], + "angle": 0, + "content": "3.3. Human Preference Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.414 + ], + "angle": 0, + "content": "We perform human preference labeling on massive generated videos. The primary goal is to validate VBench evaluation's alignment with human perception in each of the 16 evaluation dimensions, and the verification results will be detailed in Section 4.2. We also show that our human preference annotations can be useful in future tasks of finetuning generation and evaluation models to enhance alignment with human perceptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Data Preparation. Given a text prompt \\( p_i \\), and four video generation models to be evaluated \\( \\{A, B, C, D\\} \\), we use each model to generate a video, forming a \"group\" of videos \\( G_{i,j} = \\{V_{i,A,j}, V_{i,B,j}, V_{i,C,j}, V_{i,D,j}\\} \\). For each prompt \\( p_i \\), we sample five such groups of videos \\( \\{G_{i,0}, G_{i,1}, G_{i,2}, G_{i,3}, G_{i,4}\\} \\). For each group, we pair the videos up in pair-wise combinations, yielding six pairs: \\( (V_A, V_B) \\), \\( (V_A, V_C) \\), \\( (V_A, V_D) \\), \\( (V_B, V_C) \\), \\( (V_B, V_D) \\), \\( (V_C, V_D) \\), and ask human annotators to indicate their preferred video for each pair. Within the VBench evaluation framework, a prompt suite of \\( N \\) prompts produces \\( N \\times 5 \\times 6 \\) pairwise video comparisons. The video order within each pair is randomized to ensure unbiased annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.896, + 0.854 + ], + "angle": 0, + "content": "Human Labeling Rules. Specifically, the human annotators are asked to only consider the specific evaluation dimension of interest and select the preferred video. For example, in Figure 4, for the Appearance Style dimension, the question is \"Is the style of the video in the Van Gogh style?\", and human annotators are instructed to only focus on whether the generated video's style belongs to the Van Gogh style and should not consider other quality aspects of the generated video, such as potential issues like the degree of temporal flickering. In the example in this figure, video A resembles the Van Gogh better than video B, and the annotator is expected to select \"A is better\". For every dimension, we carefully prepare instructions and train the human annotators to understand the definition of the dimension, and perform multiple quality assurance protocols via a pre-labeling trial, and two rounds of post-labeling checks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Annotations for VLM Tuning. We map VBench evaluation scores from various dimensions to the scale of 0-10 and combine them with human preference annotations to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "21811" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.065, + 0.894, + 0.108 + ], + "angle": 0, + "content": "Table 1. VBenchmark Evaluation Results per Dimension. This table compares the performance of four video generation models across each of the 16 VBenchmark dimensions. A higher score indicates relatively better performance for a particular dimension. We also provide two specially built baselines, i.e., Empirical Min and Max (the approximated achievable min and max scores for each dimension), as references." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.111, + 0.768, + 0.281 + ], + "angle": 0, + "content": "
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessDynamic DegreeAesthetic QualityImaging QualityObject Class
LaVie [91]91.41%97.47%98.30%96.38%49.72%54.94%61.90%91.82%
ModelScope [62, 87]89.87%95.29%98.28%95.79%66.39%52.06%58.57%82.25%
VideoCrafter [29]86.24%92.88%97.60%91.79%89.72%44.41%57.22%87.34%
CogVideo [34]92.19%96.20%97.64%96.47%42.22%38.18%41.03%73.40%
Empirical Min14.62%26.15%62.93%70.60%0.00%0.00%0.00%0.00%
Empirical Max100.00%100.00%100.00%99.75%100.00%100.00%100.00%100.00%
ModelsMultiple ObjectsHuman ActionColorSpatial RelationshipSceneAppearance StyleTemporal StyleOverall Consistency
LaVie [91]33.32%96.80%86.39%34.09%52.69%23.56%25.93%26.41%
ModelScope [62, 87]38.98%92.40%81.72%33.68%39.26%23.39%25.37%25.67%
VideoCrafter [29]25.93%93.00%78.84%36.74%43.36%21.57%25.42%25.21%
CogVideo [34]18.11%78.20%79.57%18.24%28.24%22.01%7.80%7.70%
Empirical Min0.00%0.00%0.00%0.00%0.00%0.09%0.00%0.00%
Empirical Max100.00%100.00%100.00%100.00%82.22%28.55%36.40%36.40%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.354 + ], + "angle": 0, + "content": "form the instruction data, which is then used to fine-tune the pre-trained VideoChat [56] model to demonstrate improved evaluation capabilities. For implementation details and tuning results, please refer to the Supplementary File." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.21, + 0.383 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.463 + ], + "angle": 0, + "content": "We adopt the video generation models LaVie [91], ModelScope [62, 87], VideoCrafter [29], and CogVideo [34] for VBench evaluation, and more will be added as they become open-sourced. Details of the models and sampling procedures are in the Supplementary File." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.471, + 0.317, + 0.485 + ], + "angle": 0, + "content": "4.1. Per-Dimension Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.61 + ], + "angle": 0, + "content": "For every dimension, we calculate the VBench scores using the evaluation method suite described in Section 3.1, and show the results using Figure 2 and Table 1. We additionally designed three reference baselines, namely Empirical Max, Empirical Min, and WebVid-Avg. The first two approximate the maximum / minimum scores that videos might be able to achieve, and WebVid-Avg reflects the WebVid-10M [1] dataset quality in each VBench dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Empirical Max. For most dimensions, to approximate the maximum achievable values, we first retrieve WebVid-10M [1] videos according to our prompt suites. We use CLIP [73] to extract text features of both WebVid-10M's captions and our prompts. For each prompt, we retrieve the top-5 WebVid-10M videos according to text feature similarity with the given prompt. Given that the generated videos are usually 2 seconds in length, we randomly select a 2-second segment from each retrieved video and sample frames at 8 frames per second (FPS). For each dimension, we use the retrieved videos according to its prompt suite and report the highest-scoring video's result as Empirical Max." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Empirical Min. To approximate the minimum achievable values, we use randomly generated 2-second Gaussian noise clips to calculate results for the \"Video-Condition Consistency\" dimensions. For most \"Video Quality\" dimensions, we select frames from real videos and design frame concatenation for each dimension, approximating the minimum score achievable for each VBench dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.398 + ], + "angle": 0, + "content": "WebVid-Avg. Similar to Empirical Max, we compute the average for each dimension on retrieved WebVid-10M [1] videos. This baseline could reflect the average per-dimension quality of the commonly used video generation training dataset WebVid-10M, and provide a reference for model performances. The comparison against WebVid-Avg and Empirical Max is visualized in Figure 6 (b)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.403, + 0.848, + 0.418 + ], + "angle": 0, + "content": "4.2. Validating Human Alignment of VBench" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.496 + ], + "angle": 0, + "content": "To validate that our evaluation method can faithfully reflect human perception, we performed a large-scale human annotation for each dimension, as mentioned in Section 3.3. We show the correlation between VBenchmark evaluation results and human preference annotations in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Win Ratio. Given the human labels, we calculate the win ratio of each model. During pairwise comparisons, if a model's video is selected as better, then the model scores 1 and the other model scores 0. If there is a tie, then both models score 0.5. For each model, the win ratio is calculated as the total score divided by the total number of pairwise comparisons participated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Per-Dimension Evaluation. For each evaluation dimension, we calculate the model win ratio based on (1) VBenchmark results, and (2) human annotation results, respectively, and compute their correlations, as shown in Figure 5. We observe that VBenchmark's per-dimension evaluation results are highly correlated with human preference annotations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.7, + 0.729, + 0.715 + ], + "angle": 0, + "content": "4.3. Per-Category Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We evaluate the T2V models across eight different content categories, by generating videos based on Prompt Suite per Category described in Section 3.2, and then calculating their performance across different evaluation dimensions. Figure 7 visualizes the evaluation results of each model in terms of the eight content categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.82, + 0.85, + 0.836 + ], + "angle": 0, + "content": "4.4. Video Generation V.S. Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct a comparative analysis of the frame-wise generation capability exhibited by text-to-video (T2V) models and text-to-image (T2I) models with two primary objectives: first, to assess the extent to which T2V models" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.066, + 0.885, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 5. Validate VBench's Human Alignment. Our experiments show that VBench evaluations across all dimensions closely match human perceptions. Each plot shows the alignment verification result of a specific VBench dimension. In each plot, a dot represents the human preference win ratio (horizontal axis) and VBench evaluation win ratio (vertical axis) for a particular video generation model. We linearly fit a straight line to visualize the correlation, and calculate the Spearman's correlation coefficient \\((\\rho)\\) for each dimension." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.304, + 0.475, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.46, + 0.471, + 0.556 + ], + "angle": 0, + "content": "Figure 6. More Comparisons of Video Generation Models with Other Models and Baselines. We use VBench to evaluate other models and baselines for further comparative analysis of T2V models. (a) Comparison with text-to-image (T2I) generation models. (b) Comparison with WebVid-Avg and Empirical Max baselines. See the Supplementary File for comprehensive numerical results and details on normalization methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.751 + ], + "angle": 0, + "content": "have successfully inherited the frame-wise generative capability of the T2I models; and second, to investigate the frame-wise generation capability gap between existing T2I and T2V models. As an initial exploration into this problem, we compare video generation models with three image generation models, namely Stable Diffusion (SD) 1.4 [74], SD2.1 [74], and SDXL [71]. We choose 10 V-Bench dimensions that can encompass frame-wise generation capabilities, and sample frames from all the image and video generation models according to Prompt Suite per Evaluation Dimension described in Section 3.2. Figure 6 (a) visualizes the evaluation results of T2V versus T2I models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.308, + 0.777 + ], + "angle": 0, + "content": "5. Insights and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.81 + ], + "angle": 0, + "content": "In this section, we discuss the observations and insights we draw from our comprehensive evaluation experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Trade-off across Ability Dimensions. We have noticed a trade-off in video generation models between 1) temporal consistency (Subject Consistency, Background Consistency, Temporal Flickering, Motion Smoothness) and 2) Dynamic Degree. Models strong in temporal consistency often have a lower Dynamic Degree, as these two aspects are some" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.489 + ], + "angle": 0, + "content": "what complementary (see Figure 2 and Table 1). For example, LaVie excels in Background Consistency and Temporal Flickering but has a low Dynamic Degree, probably because generating relatively static scenes can \"cheat\" to get high temporal consistency scores. Conversely, VideoCrafter shows a high Dynamic Degree but suffers from poor performance in all temporal consistency dimensions. This trend highlights the current challenge for models to achieve temporal consistency with dynamic content of large motions. Future research should focus on enhancing both aspects simultaneously, as improving only one might indicate compromising the other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.893, + 0.734 + ], + "angle": 0, + "content": "- Uncovering Hidden Potential of T2V Models in Specific Content Categories. Our analysis reveals that the capabilities of some models vary significantly across different content types. For instance, for Aesthetic Quality, CogVideo scores well for Food (see Figure 7 rightmost chart), whereas it underperforms in others like Animal and Vehicles. The average results across various prompts might suggest a lower overall \"Aesthetic Quality\" (as seen in Figure 2), but CogVideo demonstrates relatively strong aesthetics in at least the Food category. This suggests that with tailored training data and strategies, CogVideo could potentially match other models in aesthetics by improving such ability in other content types. Therefore, we recommend evaluating video generation models not just based on ability dimensions but also considering specific content categories to uncover their hidden potential." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- Bottleneck in Temporally Complex Categories Affecting Spatial and Temporal Performance. For spatially complex categories (e.g., Animal, LifeStyle, Human, Vehicles), models all perform relatively poorly mainly in Aesthetic Quality (shown in Figure 7). This is likely due to the challenges in synthesizing harmonious color schemes, articulated structures, and appealing layouts amidst complex elements. On the other hand, for categories involving complex and intense motions like Human and Vehicle (see their Dynamic Degree in Supplementary File), performance is relatively poor across all dimen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21813" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.061, + 0.891, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.897, + 0.267 + ], + "angle": 0, + "content": "Figure 7. VBench Results across Eight Content Categories (best viewed in color). For each chart, we plot the VBench evaluation results across eight different content categories, benchmarked by our Prompt Suite per Category. The results are linearly normalized between 0 and 1 for better visibility across categories. See the Supplementary File for comprehensive numerical results, and normalization details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.369 + ], + "angle": 0, + "content": "sions. This suggests that motion complexity and dynamic intensity significantly hinder synthesis, impacting both spatial and temporal dimensions, probably because poor temporal modeling results in distorted and blurred imagery. This highlights the need for improved handling of dynamic motions in video generation models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.55 + ], + "angle": 0, + "content": "- Challenges of Data Quantity in Handling Complex Categories like Human. The WebVid-10M dataset [1] allocates \\(26\\%\\) of its content to the Human category, which is the largest share among the eight categories (see statistics in Supplementary File). However, the Human category exhibits one of the poorest results among eight categories (see Figure 7). This suggests that merely increasing data volume may not significantly enhance performance in complex categories like Human. A potential approach could involve integrating human-related priors or controls, such as skeletons, to better capture the articulated nature of human appearances and movements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.734 + ], + "angle": 0, + "content": "- Prioritizing Data Quality Over Quantity in Large-Scale Datasets. For Aesthetic Quality, Figure 7 shows that the Food category almost always tends to have the highest scores among all categories. This is corroborated by the WebVid-10M dataset [1], where Food ranks highest in Aesthetic Quality according to VBench evaluation (refer to Supplementary File for more details), despite comprising just \\(11\\%\\) of the total data. This observation suggests that at million scales, data quality might hold greater importance than quantity. Furthermore, VBench's evaluation dimensions can be potentially useful for cleaning datasets in specified quality dimensions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- Compositionality: T2I versus T2V. As shown in Figure 6 (a), T2V models significantly underperform in Multiple Objects and Spatial Relationship compared to T2I models (especially SDXL [71]), which highlights the need to enhance compositionality (i.e., correctly composing multiple objects in the same frame). We believe possible solutions might be: 1) curating training data incorporating multiple objects with corresponding captions explicitly depicting this compositionality, or 2) adding intermediate spatial control modules or modalities during video synthesis. Furthermore, the disparity of the text encoders might also account for the per" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.371, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.895, + 0.371 + ], + "angle": 0, + "content": "formance gap. As T2I models leverage bigger (OpenCLIP ViT-H for SD2.1 [74]) or more sophisticated (CLIP ViT-L & OpenCLIP ViT-G for SDXL [71]) text encoders compared with T2V models (e.g., CLIP ViT-L alone for LaVie), more representative text embeddings could be featuring more accurate object composition comprehension." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.384, + 0.62, + 0.399 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.569 + ], + "angle": 0, + "content": "With the growing focus on video generation, comprehensive evaluation of these models is essential to assess current advancements and guide future research. In this work, we take the first step forward and propose VBench, a comprehensive benchmark suite for evaluating video generation models. With its multi-dimensional, human-aligned, and insight-rich properties, VBench could play vital roles for evaluating future video generation models and inspiring further advancements in video generation. We believe that VBench is a significant contribution to the video generation and evaluation community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Limitations and Future Work. We plan to extend VBenchmark to include more models that recently became available, and extend the evaluations aspects to additional video generation tasks, like image-to-video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Potential Negative Societal Impacts. We also recognize the importance of considering ethical aspects in future iterations of VBench. While VBench currently does not assess safety and equality dimensions, we urge users to exercise caution with open-sourced video generation models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s); the National Key R&D Program of China (2022ZD0160201); National Natural Science Foundation of China (62076119, 61921006, 62102150); the Science and Technology Commission of Shanghai Municipality (23YF1461900, 23QD1400800); and Shanghai Artificial Intelligence Laboratory. We thank Shangchen Zhou, Jianyi Wang, and Ruicheng Feng for helpful suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21814" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.159, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.215, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.285, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[4] Samyadeep Basu, Mehrdad Saberi, Shweta Bhardwaj, Atoosa Malemir Chegini, Daniela Massiceti, Maziar Sanjabi, Shell Xu Hu, and Soheil Feizi. Editval: Benchmarking diffusion based text-guided image editing methods. arXiv preprint arXiv:2310.02426, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.355, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.411, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.453, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.509, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[8] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.552, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.635, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.691, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.236, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[18] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.279, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[19] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[20] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[21] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[22] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.59, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[24] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.646, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[25] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.704, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[27] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[28] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "21815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.181, + 0.469, + 0.208 + ], + "angle": 0, + "content": "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.211, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.282, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.326, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.383, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[35] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.44, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[36] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.498, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[37] Ziqi Huang, Kelvin C.K. Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.542, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[38] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C.K. Chan, and Ziwei Liu. ReVersion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.6, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[39] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.643, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[40] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Textdriven controllable human image generation. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.701, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[41] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.758, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[42] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2Performer: Textdriven human video generation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[43] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[45] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[46] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[47] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.22, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[49] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: multi-scale image quality transformer. CoRR, abs/2108.05997, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.331, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[50] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.401, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[51] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.429, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[52] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.498, + 0.889, + 0.524 + ], + "angle": 0, + "content": "[53] LAION-AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[54] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. arXiv preprint arXiv:2311.04287, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[55] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.624, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[56] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.68, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[57] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[58] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[59] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhong-cong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[60] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. In NeurIPS, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21816" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[61] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[63] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[64] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[65] Kai Zhang Yujie Lu Xingyu Fu Wenwen Zhuang Wenhu Chen Max Ku, Tianle Li. Imagenhub: Standardizing the evaluation of conditional image generation models. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.391, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[66] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. arXiv preprint arXiv:2402.14797, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.462, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[67] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[68] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[69] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[70] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[72] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In CVPR, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[75] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[76] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[77] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[80] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[81] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[82] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C. Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE TIP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[83] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[84] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In ICLRW, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.689, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[85] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.719, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[86] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[87] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[88] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and EditBench: Advancing and evaluating text-guided" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "21817" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "image inpainting. arXiv preprint arXiv:2212.06909, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[89] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, et al. Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.191, + 0.468, + 0.259 + ], + "angle": 0, + "content": "[90] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.261, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[91] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.33, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[92] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.401, + 0.468, + 0.455 + ], + "angle": 0, + "content": "[93] Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fastvqa: Efficient end-to-end video quality assessment with fragment sampling. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.457, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[94] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin. Neighbourhood representative sampling for efficient end-to-end video quality assessment. arXiv preprint arXiv:2210.05357, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.525, + 0.468, + 0.579 + ], + "angle": 0, + "content": "[95] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, and Weisi Lin. Discovqa: Temporal distortion-content transformers for video quality assessment. IEEE TCSVT, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.581, + 0.468, + 0.635 + ], + "angle": 0, + "content": "[96] Haoning Wu, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring opinion-unaware video quality assessment with semantic affinity criterion. In ICME, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.637, + 0.468, + 0.705 + ], + "angle": 0, + "content": "[97] Haoning Wu, Liang Liao, Annan Wang, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards robust text-prompted semantic criterion for in-the-wild video quality assessment. arXiv preprint arXiv:2304.14672, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[98] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[99] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards explainable video quality assessment: A database and a language-prompted approach. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[100] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[101] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[102] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[103] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[104] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[105] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[106] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[107] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[108] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[109] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[110] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. Magicavatar: Multimodal avatar generation and animation. arXiv preprint arXiv:2308.14748, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[111] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.811 + ], + "angle": 0, + "content": "[112] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.814, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[113] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21818" + } + ] +] \ No newline at end of file diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_origin.pdf b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f181d92ab50e9d40f9b9e11d7c0d8b64e8478e7 --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/852c042a-18a2-4ced-8d57-cde6f2ed2650_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d789dec2ee994a4297a34c1447330269534d19479547334f6dae4e250f46319b +size 8711358 diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/full.md b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eedd5129ebbe833633b59762e98f19ea8f06370c --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/full.md @@ -0,0 +1,332 @@ +# VBenchmark: Comprehensive Benchmark Suite for Video Generative Models + +Ziqi Huang $^{1*}$ Yinan He $^{2*}$ Jiashuo Yu $^{2*}$ Fan Zhang $^{2*}$ Chenyang Si $^{1}$ Yuming Jiang $^{1}$ +Yuanhan Zhang $^{1}$ Tianxing Wu $^{1}$ Qingyang Jin $^{1}$ Nattapol Chanpaisit $^{1}$ +Yaohui Wang $^{2}$ Xinyuan Chen $^{2}$ Limin Wang $^{4,2}$ Dahua Lin $^{2,3\boxtimes}$ Yu Qiao $^{2\boxtimes}$ Ziwei Liu $^{1\boxtimes}$ + +$^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ The Chinese University of Hong Kong $^{4}$ Nanjing University + +https://vchitect.github.io/VBench-project/ + +![](images/53e38a1b6b63f2c36f3d80e8036ca07f24f19673565536d356f21e1c424766cb.jpg) +Figure 1. Overview of VBench. We propose VBench, a comprehensive benchmark suite for video generative models. We design a comprehensive and hierarchical Evaluation Dimension Suite to decompose "video generation quality" into multiple well-defined dimensions to facilitate fine-grained and objective evaluation. For each dimension and each content category, we carefully design a Prompt Suite as test cases, and sample Generated Videos from a set of video generation models. For each evaluation dimension, we specifically design an Evaluation Method Suite, which uses a carefully crafted method or designated pipeline for automatic objective evaluation. We also conduct Human Preference Annotation for the generated videos for each dimension and show that VBench evaluation results are well aligned with human perceptions. VBench can provide valuable insights from multiple perspectives. + +![](images/c6d763e4a79961e303efc773cb5e13ede49ac8e6e10f18e9aabe77fa894557da.jpg) + +# Abstract + +Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBenchmark, a comprehensive benchmark suite that dissects "video generation quality" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBenchmark has three appealing properties: 1) Comprehensive Dimensions: VBenchmark comprises + +16 dimensions in video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc.). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VBench to drive forward the field of video generation. + +![](images/a3d0229a0445214c16208b34564756131840011b11ee92cb4c8960824875f840.jpg) +Figure 2. VBench Evaluation Results of Video Generative Models. We visualize the evaluation results of four video generation models in 16 VBench dimensions. We normalize the results per dimension for clearer comparisons. For comprehensive numerical results, please refer to Table 1. + +# 1. Introduction + +Image generation models have made rapid progress in the past few years, such as Variational Autoencoders (VAEs) [51], Generative Adversarial Networks (GANs) [6, 19, 20, 24, 39, 41, 44-47, 67], vector quantized (VQ) based approaches [16, 40, 85], and diffusion models [31, 78, 79]. This fuels recent explorations in video generation [3, 5, 10, 21, 23, 27, 29, 34, 52, 62, 66, 77, 86, 87, 89, 91, 101, 107-109], which goes beyond static imagery and models the dynamics and kinematics of real-world scenes. With the growth of video generation models, there arises a critical need for effective evaluation methods. The evaluation should be able to accurately reflect human perception of generated videos, providing reliable measures of a model's performance. Additionally, it should reflect each model's specific strengths and weaknesses, offering insights that inform the data, training, and architectural choices of future video generation models. + +However, existing metrics for video generation such as Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83, 84], and CLIPSIM [73] are inconsistent with human judgement [15, 69]. Meanwhile, the Video Quality Assessment (VQA) methods [55, 82, 93-99] are primarily designed for real videos, thereby neglecting the unique challenges posed by generative models, such as artifacts in synthesized videos. Hence, there is a pressing need for an evaluation framework that aligns closely with human perception, and specifically designed for the characteristics of video generation models. + +To this end, we introduce VBench, a comprehensive benchmark suite for evaluating video generation model performance. VBench has three appealing properties: 1) comprehensive evaluation dimensions, 2) human alignment, and 3) valuable insights. + +First, our framework includes an evaluation dimension + +suite that employs a hierarchical and disentangled approach to the decomposition of "video generation quality". This suite systematically breaks down the evaluation into two primary dimensions at a coarse level: Video Quality and Video-Condition Consistency. Each of these dimensions is further subdivided into more granular criteria. This hierarchical separation ensures that each dimension isolates and evaluates a single aspect of video quality, without interference from other variables, as illustrated in Figure 1. Recognizing video generation's unique challenges, we have tailored evaluation dimensions to its specific characteristics. For example, in terms of Video Quality, maintaining consistent subject identity (e.g., a teddy bear) in generated videos is crucial, and is a problem rarely encountered in real-world videos. Additionally, Video-Condition Consistency is vital for conditional video generation tasks, requiring its dedicated evaluation criteria. For each evaluation dimension, we carefully prepared around 100 text prompts as test cases for text-to-video (T2V) generation, and devised specialized evaluation methods tailored to each dimension. In addition to multi-dimensional evaluations, we also assess T2V models across diverse content categories. We organized prompt suites for eight distinct types, such as animal, architecture, human, and scenery, allowing for a separate evaluation within each category. This exploration reveals variable competencies in T2V generation across different content types, highlighting areas of proficiency and those requiring further enhancement. + +Second, we systematically demonstrate that our evaluation method suite is closely aligned with human perception in every fine-grained evaluation dimension. We collected human preference annotations for each dimension. Specifically, we use various T2V models to sample videos from our prompt suites. Then given two videos sampled from the same prompt, we ask human annotators to indicate preferences according to each VBenchmark dimension respectively. We show that VBenchmark evaluations highly correlate with human preferences. Additionally, the human preference annotations can be utilized for multiple purposes, such as fine-tuning generation or evaluation models to enhance alignment with human perceptions. For instance, we utilize the annotations to implement Instruction Tuning within a Visual-Language Model (VLM), enhancing its T2V evaluation alignment with human preferences. + +Third, VBench's multi-dimensional and multicategorical approach can provide valuable insights to the video generation community. Our multi-dimensional system enables detailed feedback on the strengths and weaknesses of video generation models across various ability aspects. This approach not only ensures a comprehensive evaluation of existing models but also provides valuable insights into the training of advanced video generation models, guiding architectural and data choices + +for improved video generation outcomes. Additionally, VBench can be readily applied to evaluate image generation models, and thus we investigate the disparities between video and image generation models. In Section 5, we discuss in detail on various observations and insights drawn from VBench evaluations. + +We are open-sourcing VBench, including its evaluation dimension suite, evaluation method suite, prompt suite, generated videos, and the dataset of human preference annotations. We also encourage more video generation models to participate in the VBench challenge. + +# 2. Related Works + +Video Generative Models. Recently, diffusion models [14, 31, 78, 79] have achieved significant progress in image synthesis [25, 37, 38, 68, 71, 74, 75], and enabled a line of works towards video generation [5, 9, 22, 26, 28, 29, 32, 33, 42, 50, 62, 63, 77, 87, 91, 102, 103, 109, 113]. Many recent diffusion-based works [29, 62, 87, 91] are text-to-video (T2V) models. Other guidance modalities are also available, including image-to-video [11, 13, 17, 106], video-to-video [8, 59, 70, 72, 105], and a variety of control maps [12, 43, 50, 64, 90, 110, 111] such as pose, depth, and sketch. The boom of video generation models requires a comprehensive evaluation system to inform their current capabilities and guide future developments, and VBench takes the initiative in providing a comprehensive benchmark suite for fine-grained and human-aligned evaluation. + +Evaluation of Visual Generative Models. Existing video generation models typically use metrics like Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83], and CLIPSIM [73] for evaluation. The UCF-101 [80] dataset's class labels often serve as text prompts for IS, FID, and FVD, whereas MSR-VTT [104]'s human-labeled video captions are used for CLIPSIM. Despite covering various real-world scenarios, these prompts lack diversity and specificity, limiting accurate and fine-grained evaluation of video generation. For text-to-image (T2I) models, several benchmarks [2, 4, 35, 54, 65, 75, 88] are proposed to assess various capabilities like compositionality [35] and editing ability [4, 88]. However, video generative models still lack comprehensive evaluation benchmarks for detailed and human-aligned feedback. Our work differs from concurrent research [60, 61] in three key ways: 1) We have created 16 distinct evaluation dimensions, each with specialized prompts for precise assessment; 2) We have empirically validated that every dimension aligns closely with human perception; 3) Our multi-dimensional and multi-categorical evaluation offers valuable and comprehensive insights into video generation. + +# 3. VBench Suite + +In this section, we introduce the main components of VBench. In Section 3.1, we present our rationale for designing the 16 evaluation dimensions, as well as each dimension's definition and evaluation method. We then elaborate on the prompt suites we use in Section 3.2. To validate VBench's alignment with human perception, we conduct human preference annotation for each dimension (see Section 3.3). The experiments and the insights drawn from VBench will be detailed in Section 4 and Section 5. + +# 3.1. Evaluation Dimension Suite + +We first introduce our evaluation dimensions and their corresponding evaluation methods. + +Existing evaluation metrics like FVD [83] often conclude video generation model performance to a single number. This oversimplifies the evaluation and has several risks. First, a single number can obscure an individual model's strengths and weaknesses, and it fails to provide insights into specific areas where a model excels or underperforms. This makes it challenging to derive insights for future architectural and training designs based on single-valued metrics. Second, the notion of "high-quality video generation" is complex and multifaceted, with individuals prioritizing different video attributes based on the intended application. For instance, some may prioritize the absence of temporal flickering, while others may consider fidelity to the text prompt as the most significant, with less emphasis on flickering. Therefore, in contrast with performing single-valued evaluations of video generation quality, we propose a disaggregated approach by decomposing the brand notion of "video generation performance" into multiple discrete dimensions for fine-grained evaluation. + +Specifically, we break "video generation quality" down into 16 disentangled dimensions in a top-down manner, with each evaluation dimension assessing one aspect of video generation quality. On the top level, we evaluate T2V performance from two broad perspectives: 1) Video Quality — "Without considering alignment with the text prompt, does the video alone look good?", which focuses on the perceptual quality of the synthesized video, and does not consider the input condition (e.g., text prompt), and 2) Video-Condition Consistency — "Is the video consistent with what the user wants to generate?", which focuses on whether the synthesized video is consistent with the guiding condition that the user provides (e.g., the text prompt for T2V generation). Under both "Video Quality" and "Video-Condition Consistency", we further break the coarse-grained dimensions into more fine-grained dimensions, as shown in Figure 1. We briefly introduce each dimension here. Please refer to the Supplementary File for the detailed definition and evaluation method of each dimension. + +# 3.1.1 Video Quality + +We split "Video Quality" into two disentangled aspects, "Temporal Quality" and "Frame-Wise Quality", where the former only considers the cross-frame consistency and dynamics, and the latter only considers the quality of each individual frame without taking temporal quality into concern. For "Temporal Quality", we further devise five evaluation dimensions, where each focusing on a different aspect of temporal quality. + +Temporal Quality - Subject Consistency. For a subject (e.g., a person, a car, or a cat) in the video, we assess whether its appearance remains consistent throughout the whole video. To this end, we calculate the DINO [7] feature similarity across frames. + +Temporal Quality - Background Consistency. We evaluate the temporal consistency of the background scenes by calculating CLIP [73] feature similarity across frames. + +Temporal Quality - Temporal Flickering. Generated videos can exhibit imperfect temporal consistency at local and high-frequency details. We take static frames and compute the mean absolute difference across frames. + +Temporal Quality - Motion Smoothness. Subject Consistency and Background Consistency focus on temporal consistency of the "look" instead of the smoothness of "movement and motion". We believe it is important to evaluate whether the motion in the generated video is smooth, and follows the physical law of the real world. We utilize the motion priors in the video frame interpolation model [58] to evaluate the smoothness of generated motions. + +Temporal Quality - Dynamic Degree. Since a completely static video can score well in the aforementioned temporal quality dimensions, it is important to also evaluate the degree of dynamics (i.e., whether it contains large motions) generated by each model. We use RAFT [81] to estimate the degree of dynamics in synthesized videos. + +Frame-Wise Quality - Aesthetic Quality. We evaluate the artistic and beauty value perceived by humans towards each video frame using the LAION aesthetic predictor [53]. It can reflect aesthetic aspects such as the layout, the richness and harmony of colors, the photo-realism, naturalness, and artistic quality of the video frames. + +Frame-Wise Quality - Imaging Quality. Imaging quality refers to the distortion (e.g., over-exposure, noise, blur) presented in the generated frames, and we evaluate it using the MUSIQ [49] image quality predictor trained on the SPAQ [18] dataset. + +# 3.1.2 Video-Condition Consistency + +We mainly dissect "Video-Condition Consistency" into "Semantics" (i.e., the type of the entities and their attributes) and "Style" (i.e., whether the generated video + +is consistent with user-requested style), with each decomposed into more fine-grained dimensions. + +Semantics - Object Class. We use GRiT [100] to detect the success rate of generating the specific class of objects depicted in the text prompt. + +Semantics - Multiple Objects. Other than generating a single object of a particular class, the ability to compose multiple objects from different classes in the same frame is also an essential ability in video generation. We detect the success rate of generating all the objects specified in the text prompt within each video frame. + +Semantics - Human Action. Human action is an important aspect in human-centric video generation. We apply UMT [57] to evaluate whether human subjects in generated videos can accurately execute the specific actions mentioned in the text prompts. + +Semantics - Color. To evaluate whether synthesized object colors align with the text prompt, we use GRiT [100] for color captioning and comparison with expected colors. + +Semantics - Spatial Relationship. Other than classes and attributes of synthesized objects, we also evaluate whether their spatial relationship follows what is specified by the text prompt. We focus on four primary types of spatial relationships, and perform rule-based evaluation similar to [35]. + +Semantics - Scene. We need to evaluate whether the synthesized video is consistent with the intended scene described by the text prompt. For example, when prompted "ocean", the generated video should be "ocean" instead of "river". We use Tag2Text [36] to caption the generated scenes, and then check its correspondence with scene descriptions in the text prompt. + +Style - Appearance Style. Apart from semantics consistency with the text prompt, another important pillar in video-condition consistency is style. There are many styles that alter the look, color, and texture of synthesized video frames, such as "oil painting style", "black and white style", "watercolor painting style", "cyberpunk style", "black and white" etc. We calculate the CLIP [73] feature similarity between synthesized frames and these style descriptions. + +Style - Temporal Style. Apart from appearance styles, videos also have temporal styles like various camera motions. We use ViCLIP [92] to calculate the video feature and the temporal style description feature similarity to reflect temporal style consistency. + +Overall Consistency. We further use overall video-text consistency computed by ViCLIP [92] on general text prompts as an aiding metric to reflect both semantics and style consistency. + +# 3.2. Prompt Suite + +The sampling procedure of current diffusion-based video generation models [29, 87, 91] is time-consuming (e.g., 90 seconds per video for LaVie [91], and more than 2 minutes + +![](images/2061c7ebb5579f224dfebbfe68f8a8fb80e68f609e1dc4b211d1fc03181a7bea.jpg) +Figure 3. Prompt Suite Statistics. The two graphs provide an overview of our prompt suites. Left: the word cloud to visualize word distribution of our prompt suites. Right: the number of prompts across different evaluation dimensions and different content categories. +Figure 4. Interface for Human Preference Annotation. Top: prompt and question. Right: choices that annotators can make. Bottom left: control for stop and playback. + +![](images/736f108c33ad8d4ce70b2d30beae340d6d8ebff7191df1d985ba75fa14571dc5.jpg) + +per video for CogVideo [34]). Therefore, we need to control the amount of test cases for efficient evaluation. Meanwhile, we need to maintain the diversity and comprehensiveness of our prompt suite, so we design compact yet representative prompts in terms of both the evaluation dimensions and the content categories. We visualize our prompt suite distributions in Figure 3. + +Prompt Suite per Dimension. For each VBenchmark evaluation dimension, we carefully designed a suite of around 100 prompts as test cases. The prompt suite is carefully curated to probe the specific ability relevant to the dimension tested. For example, for the "Subject Consistency" dimension which aims to evaluate the consistency of subjects' appearances throughout the video, we ensure every prompt has a movable subject (e.g., animals or vehicles) performing non-static actions, where their consistency might be compromised due to inconsistency introduced by their movements or changing locations. In "Object Class" dimension, we ensure the existence of a specific class of object in every prompt. For "Human Action", each test prompt contains a human subject performing a well-defined action from the Kinetics-400 dataset [48], where 100 representative actions are selected with minimal semantic overlaps among themselves. Please refer to the Supplementary File for the design rationale of the prompt suite for all 16 dimensions. + +Prompt Suite per Category. When designing prompts for each dimension, the focus was to showcase models' ability in that specific dimension. We further incorporate prompt suites for eight content categories to provide insights into the performance across varied content types. To this end, we prepare a collection of human-curated prompts from the Internet and divide them into 8 distinctive categories following YouTube's categorization. Subsequently, we feed both the category labels and prompts into a Large Language Model (LLM) [112] (see more implementation details in Supplementary File), obtaining multi-label outputs for each caption. We select 800 prompts and manually clean their labels to serve as per-category prompt suites. Finally, we obtain 100 prompts for each of these eight categories: Animal, Architecture, Food, Human, Lifestyle, Plant, Scenery, and Vehicles. + +# 3.3. Human Preference Annotation + +We perform human preference labeling on massive generated videos. The primary goal is to validate VBench evaluation's alignment with human perception in each of the 16 evaluation dimensions, and the verification results will be detailed in Section 4.2. We also show that our human preference annotations can be useful in future tasks of finetuning generation and evaluation models to enhance alignment with human perceptions. + +Data Preparation. Given a text prompt $p_i$ , and four video generation models to be evaluated $\{A, B, C, D\}$ , we use each model to generate a video, forming a "group" of videos $G_{i,j} = \{V_{i,A,j}, V_{i,B,j}, V_{i,C,j}, V_{i,D,j}\}$ . For each prompt $p_i$ , we sample five such groups of videos $\{G_{i,0}, G_{i,1}, G_{i,2}, G_{i,3}, G_{i,4}\}$ . For each group, we pair the videos up in pair-wise combinations, yielding six pairs: $(V_A, V_B)$ , $(V_A, V_C)$ , $(V_A, V_D)$ , $(V_B, V_C)$ , $(V_B, V_D)$ , $(V_C, V_D)$ , and ask human annotators to indicate their preferred video for each pair. Within the VBench evaluation framework, a prompt suite of $N$ prompts produces $N \times 5 \times 6$ pairwise video comparisons. The video order within each pair is randomized to ensure unbiased annotation. + +Human Labeling Rules. Specifically, the human annotators are asked to only consider the specific evaluation dimension of interest and select the preferred video. For example, in Figure 4, for the Appearance Style dimension, the question is "Is the style of the video in the Van Gogh style?", and human annotators are instructed to only focus on whether the generated video's style belongs to the Van Gogh style and should not consider other quality aspects of the generated video, such as potential issues like the degree of temporal flickering. In the example in this figure, video A resembles the Van Gogh better than video B, and the annotator is expected to select "A is better". For every dimension, we carefully prepare instructions and train the human annotators to understand the definition of the dimension, and perform multiple quality assurance protocols via a pre-labeling trial, and two rounds of post-labeling checks + +Annotations for VLM Tuning. We map VBench evaluation scores from various dimensions to the scale of 0-10 and combine them with human preference annotations to + +Table 1. VBenchmark Evaluation Results per Dimension. This table compares the performance of four video generation models across each of the 16 VBenchmark dimensions. A higher score indicates relatively better performance for a particular dimension. We also provide two specially built baselines, i.e., Empirical Min and Max (the approximated achievable min and max scores for each dimension), as references. + +
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessDynamic DegreeAesthetic QualityImaging QualityObject Class
LaVie [91]91.41%97.47%98.30%96.38%49.72%54.94%61.90%91.82%
ModelScope [62, 87]89.87%95.29%98.28%95.79%66.39%52.06%58.57%82.25%
VideoCrafter [29]86.24%92.88%97.60%91.79%89.72%44.41%57.22%87.34%
CogVideo [34]92.19%96.20%97.64%96.47%42.22%38.18%41.03%73.40%
Empirical Min14.62%26.15%62.93%70.60%0.00%0.00%0.00%0.00%
Empirical Max100.00%100.00%100.00%99.75%100.00%100.00%100.00%100.00%
ModelsMultiple ObjectsHuman ActionColorSpatial RelationshipSceneAppearance StyleTemporal StyleOverall Consistency
LaVie [91]33.32%96.80%86.39%34.09%52.69%23.56%25.93%26.41%
ModelScope [62, 87]38.98%92.40%81.72%33.68%39.26%23.39%25.37%25.67%
VideoCrafter [29]25.93%93.00%78.84%36.74%43.36%21.57%25.42%25.21%
CogVideo [34]18.11%78.20%79.57%18.24%28.24%22.01%7.80%7.70%
Empirical Min0.00%0.00%0.00%0.00%0.00%0.09%0.00%0.00%
Empirical Max100.00%100.00%100.00%100.00%82.22%28.55%36.40%36.40%
+ +form the instruction data, which is then used to fine-tune the pre-trained VideoChat [56] model to demonstrate improved evaluation capabilities. For implementation details and tuning results, please refer to the Supplementary File. + +# 4. Experiments + +We adopt the video generation models LaVie [91], ModelScope [62, 87], VideoCrafter [29], and CogVideo [34] for VBench evaluation, and more will be added as they become open-sourced. Details of the models and sampling procedures are in the Supplementary File. + +# 4.1. Per-Dimension Evaluation + +For every dimension, we calculate the VBench scores using the evaluation method suite described in Section 3.1, and show the results using Figure 2 and Table 1. We additionally designed three reference baselines, namely Empirical Max, Empirical Min, and WebVid-Avg. The first two approximate the maximum / minimum scores that videos might be able to achieve, and WebVid-Avg reflects the WebVid-10M [1] dataset quality in each VBench dimension. + +Empirical Max. For most dimensions, to approximate the maximum achievable values, we first retrieve WebVid-10M [1] videos according to our prompt suites. We use CLIP [73] to extract text features of both WebVid-10M's captions and our prompts. For each prompt, we retrieve the top-5 WebVid-10M videos according to text feature similarity with the given prompt. Given that the generated videos are usually 2 seconds in length, we randomly select a 2-second segment from each retrieved video and sample frames at 8 frames per second (FPS). For each dimension, we use the retrieved videos according to its prompt suite and report the highest-scoring video's result as Empirical Max. + +Empirical Min. To approximate the minimum achievable values, we use randomly generated 2-second Gaussian noise clips to calculate results for the "Video-Condition Consistency" dimensions. For most "Video Quality" dimensions, we select frames from real videos and design frame concatenation for each dimension, approximating the minimum score achievable for each VBench dimension. + +WebVid-Avg. Similar to Empirical Max, we compute the average for each dimension on retrieved WebVid-10M [1] videos. This baseline could reflect the average per-dimension quality of the commonly used video generation training dataset WebVid-10M, and provide a reference for model performances. The comparison against WebVid-Avg and Empirical Max is visualized in Figure 6 (b). + +# 4.2. Validating Human Alignment of VBench + +To validate that our evaluation method can faithfully reflect human perception, we performed a large-scale human annotation for each dimension, as mentioned in Section 3.3. We show the correlation between VBenchmark evaluation results and human preference annotations in Figure 5. + +Win Ratio. Given the human labels, we calculate the win ratio of each model. During pairwise comparisons, if a model's video is selected as better, then the model scores 1 and the other model scores 0. If there is a tie, then both models score 0.5. For each model, the win ratio is calculated as the total score divided by the total number of pairwise comparisons participated. + +Per-Dimension Evaluation. For each evaluation dimension, we calculate the model win ratio based on (1) VBenchmark results, and (2) human annotation results, respectively, and compute their correlations, as shown in Figure 5. We observe that VBenchmark's per-dimension evaluation results are highly correlated with human preference annotations. + +# 4.3. Per-Category Evaluation + +We evaluate the T2V models across eight different content categories, by generating videos based on Prompt Suite per Category described in Section 3.2, and then calculating their performance across different evaluation dimensions. Figure 7 visualizes the evaluation results of each model in terms of the eight content categories. + +# 4.4. Video Generation V.S. Image Generation + +We conduct a comparative analysis of the frame-wise generation capability exhibited by text-to-video (T2V) models and text-to-image (T2I) models with two primary objectives: first, to assess the extent to which T2V models + +![](images/18e1edf20a57807908f97ff5944543f3b40659999ea81933f208949b94d8926c.jpg) +Figure 5. Validate VBench's Human Alignment. Our experiments show that VBench evaluations across all dimensions closely match human perceptions. Each plot shows the alignment verification result of a specific VBench dimension. In each plot, a dot represents the human preference win ratio (horizontal axis) and VBench evaluation win ratio (vertical axis) for a particular video generation model. We linearly fit a straight line to visualize the correlation, and calculate the Spearman's correlation coefficient $(\rho)$ for each dimension. + +![](images/271dbb03f501abe078ec14e0f96007705aa2898b4577a30f243fea91fd1d8dd0.jpg) +Figure 6. More Comparisons of Video Generation Models with Other Models and Baselines. We use VBench to evaluate other models and baselines for further comparative analysis of T2V models. (a) Comparison with text-to-image (T2I) generation models. (b) Comparison with WebVid-Avg and Empirical Max baselines. See the Supplementary File for comprehensive numerical results and details on normalization methods. + +have successfully inherited the frame-wise generative capability of the T2I models; and second, to investigate the frame-wise generation capability gap between existing T2I and T2V models. As an initial exploration into this problem, we compare video generation models with three image generation models, namely Stable Diffusion (SD) 1.4 [74], SD2.1 [74], and SDXL [71]. We choose 10 V-Bench dimensions that can encompass frame-wise generation capabilities, and sample frames from all the image and video generation models according to Prompt Suite per Evaluation Dimension described in Section 3.2. Figure 6 (a) visualizes the evaluation results of T2V versus T2I models. + +# 5. Insights and Discussions + +In this section, we discuss the observations and insights we draw from our comprehensive evaluation experiments. + +- Trade-off across Ability Dimensions. We have noticed a trade-off in video generation models between 1) temporal consistency (Subject Consistency, Background Consistency, Temporal Flickering, Motion Smoothness) and 2) Dynamic Degree. Models strong in temporal consistency often have a lower Dynamic Degree, as these two aspects are some + +what complementary (see Figure 2 and Table 1). For example, LaVie excels in Background Consistency and Temporal Flickering but has a low Dynamic Degree, probably because generating relatively static scenes can "cheat" to get high temporal consistency scores. Conversely, VideoCrafter shows a high Dynamic Degree but suffers from poor performance in all temporal consistency dimensions. This trend highlights the current challenge for models to achieve temporal consistency with dynamic content of large motions. Future research should focus on enhancing both aspects simultaneously, as improving only one might indicate compromising the other. + +- Uncovering Hidden Potential of T2V Models in Specific Content Categories. Our analysis reveals that the capabilities of some models vary significantly across different content types. For instance, for Aesthetic Quality, CogVideo scores well for Food (see Figure 7 rightmost chart), whereas it underperforms in others like Animal and Vehicles. The average results across various prompts might suggest a lower overall "Aesthetic Quality" (as seen in Figure 2), but CogVideo demonstrates relatively strong aesthetics in at least the Food category. This suggests that with tailored training data and strategies, CogVideo could potentially match other models in aesthetics by improving such ability in other content types. Therefore, we recommend evaluating video generation models not just based on ability dimensions but also considering specific content categories to uncover their hidden potential. + +- Bottleneck in Temporally Complex Categories Affecting Spatial and Temporal Performance. For spatially complex categories (e.g., Animal, LifeStyle, Human, Vehicles), models all perform relatively poorly mainly in Aesthetic Quality (shown in Figure 7). This is likely due to the challenges in synthesizing harmonious color schemes, articulated structures, and appealing layouts amidst complex elements. On the other hand, for categories involving complex and intense motions like Human and Vehicle (see their Dynamic Degree in Supplementary File), performance is relatively poor across all dimen + +![](images/72706807e6048bcc1277c117c1ec346373f4f827e80256234c39f52dface877d.jpg) +Figure 7. VBench Results across Eight Content Categories (best viewed in color). For each chart, we plot the VBench evaluation results across eight different content categories, benchmarked by our Prompt Suite per Category. The results are linearly normalized between 0 and 1 for better visibility across categories. See the Supplementary File for comprehensive numerical results, and normalization details. + +sions. This suggests that motion complexity and dynamic intensity significantly hinder synthesis, impacting both spatial and temporal dimensions, probably because poor temporal modeling results in distorted and blurred imagery. This highlights the need for improved handling of dynamic motions in video generation models. + +- Challenges of Data Quantity in Handling Complex Categories like Human. The WebVid-10M dataset [1] allocates $26\%$ of its content to the Human category, which is the largest share among the eight categories (see statistics in Supplementary File). However, the Human category exhibits one of the poorest results among eight categories (see Figure 7). This suggests that merely increasing data volume may not significantly enhance performance in complex categories like Human. A potential approach could involve integrating human-related priors or controls, such as skeletons, to better capture the articulated nature of human appearances and movements. +- Prioritizing Data Quality Over Quantity in Large-Scale Datasets. For Aesthetic Quality, Figure 7 shows that the Food category almost always tends to have the highest scores among all categories. This is corroborated by the WebVid-10M dataset [1], where Food ranks highest in Aesthetic Quality according to VBench evaluation (refer to Supplementary File for more details), despite comprising just $11\%$ of the total data. This observation suggests that at million scales, data quality might hold greater importance than quantity. Furthermore, VBench's evaluation dimensions can be potentially useful for cleaning datasets in specified quality dimensions. +- Compositionality: T2I versus T2V. As shown in Figure 6 (a), T2V models significantly underperform in Multiple Objects and Spatial Relationship compared to T2I models (especially SDXL [71]), which highlights the need to enhance compositionality (i.e., correctly composing multiple objects in the same frame). We believe possible solutions might be: 1) curating training data incorporating multiple objects with corresponding captions explicitly depicting this compositionality, or 2) adding intermediate spatial control modules or modalities during video synthesis. Furthermore, the disparity of the text encoders might also account for the per + +formance gap. As T2I models leverage bigger (OpenCLIP ViT-H for SD2.1 [74]) or more sophisticated (CLIP ViT-L & OpenCLIP ViT-G for SDXL [71]) text encoders compared with T2V models (e.g., CLIP ViT-L alone for LaVie), more representative text embeddings could be featuring more accurate object composition comprehension. + +# 6. Conclusion + +With the growing focus on video generation, comprehensive evaluation of these models is essential to assess current advancements and guide future research. In this work, we take the first step forward and propose VBench, a comprehensive benchmark suite for evaluating video generation models. With its multi-dimensional, human-aligned, and insight-rich properties, VBench could play vital roles for evaluating future video generation models and inspiring further advancements in video generation. We believe that VBench is a significant contribution to the video generation and evaluation community. + +Limitations and Future Work. We plan to extend VBenchmark to include more models that recently became available, and extend the evaluations aspects to additional video generation tasks, like image-to-video. + +Potential Negative Societal Impacts. We also recognize the importance of considering ethical aspects in future iterations of VBench. While VBench currently does not assess safety and equality dimensions, we urge users to exercise caution with open-sourced video generation models. + +Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s); the National Key R&D Program of China (2022ZD0160201); National Natural Science Foundation of China (62076119, 61921006, 62102150); the Science and Technology Commission of Shanghai Municipality (23YF1461900, 23QD1400800); and Shanghai Artificial Intelligence Laboratory. We thank Shangchen Zhou, Jianyi Wang, and Ruicheng Feng for helpful suggestions. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6, 8 +[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 3 +[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2 +[4] Samyadeep Basu, Mehrdad Saberi, Shweta Bhardwaj, Atoosa Malemir Chegini, Daniela Massiceti, Maziar Sanjabi, Shell Xu Hu, and Soheil Feizi. Editval: Benchmarking diffusion based text-guided image editing methods. arXiv preprint arXiv:2310.02426, 2023. 3 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3 +[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 4 +[8] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3 +[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3 +[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 2 +[11] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 3 +[12] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3 +[13] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 3 +[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 3 + +[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2 +[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[17] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3 +[18] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 4 +[19] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2 +[20] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In ICCV, 2023. 2 +[21] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2 +[22] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3 +[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2 +[24] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[25] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3 +[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 3 +[27] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2 +[28] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3 +[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high + +fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 3, 4, 6 +[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 3 +[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 +[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3 +[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3 +[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 5, 6 +[35] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350, 2023. 3, 4 +[36] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 4 +[37] Ziqi Huang, Kelvin C.K. Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In CVPR, 2023. 3 +[38] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C.K. Chan, and Ziwei Liu. ReVersion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3 +[39] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 2 +[40] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Textdriven controllable human image generation. ACM TOG, 2022. 2 +[41] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE TPAMI, 2023. 2 +[42] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2Performer: Textdriven human video generation. In ICCV, 2023. 3 +[43] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 3 +[44] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2 + +[45] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. +[46] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. +[47] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2 +[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5 +[49] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: multi-scale image quality transformer. CoRR, abs/2108.05997, 2021. 4 +[50] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3 +[51] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[52] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2 +[53] LAION-AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor, 2022.4 +[54] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. arXiv preprint arXiv:2311.04287, 2023. 3 +[55] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, 2019. 2 +[56] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6 +[57] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058, 2023. 4 +[58] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 4 +[59] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhong-cong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3 +[60] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. In NeurIPS, 2023. 3 + +[61] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In CVPR, 2024. 3 +[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 3, 6 +[63] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3 +[64] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3 +[65] Kai Zhang Yujie Lu Xingyu Fu Wenwen Zhuang Wenhu Chen Max Ku, Tianle Li. Imagenhub: Standardizing the evaluation of conditional image generation models. In ICLR, 2024. 3 +[66] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. arXiv preprint arXiv:2402.14797, 2024. 2 +[67] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[68] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3 +[69] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In CVPR, 2023. 2 +[70] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 3 +[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8 +[72] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3 +[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 6 +[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In CVPR, 2022. 3, 7, 8 +[75] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3 +[76] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2, 3 +[77] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2.3 +[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3 +[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3 +[80] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3 +[81] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4 +[82] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C. Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE TIP, 2021. 2 +[83] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 3 +[84] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In ICLRW, 2019. 2 +[85] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2 +[86] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2 +[87] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3, 4, 6 +[88] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and EditBench: Advancing and evaluating text-guided + +image inpainting. arXiv preprint arXiv:2212.06909, 2022.3 +[89] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, et al. Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2 +[90] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3 +[91] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3, 4, 6 +[92] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 4 +[93] Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fastvqa: Efficient end-to-end video quality assessment with fragment sampling. In ECCV, 2022. 2 +[94] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin. Neighbourhood representative sampling for efficient end-to-end video quality assessment. arXiv preprint arXiv:2210.05357, 2022. +[95] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, and Weisi Lin. Discovqa: Temporal distortion-content transformers for video quality assessment. IEEE TCSVT, 2023. +[96] Haoning Wu, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring opinion-unaware video quality assessment with semantic affinity criterion. In ICME, 2023. +[97] Haoning Wu, Liang Liao, Annan Wang, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards robust text-prompted semantic criterion for in-the-wild video quality assessment. arXiv preprint arXiv:2304.14672, 2023. +[98] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023. +[99] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards explainable video quality assessment: A database and a language-prompted approach. In ACM MM, 2023. 2 +[100] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 4 + +[101] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2 +[102] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 3 +[103] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 3 +[104] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3 +[105] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3 +[106] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3 +[107] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2 +[108] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. +[109] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3 +[110] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. Magicavatar: Multimodal avatar generation and animation. arXiv preprint arXiv:2308.14748, 2023. 3 +[111] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3 +[112] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5 +[113] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3 \ No newline at end of file diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/images.zip b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1126770698fe4b76b0aa712edeb9c7066f8edb02 --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4db4ce6755c6f6912dcb3e0dc53de987a6572d4388f4fa752b274a0c9ccea67 +size 454120 diff --git a/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/layout.json b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46426d10093221b307c3f89502f257ac162f9952 --- /dev/null +++ b/2024/VBench_ Comprehensive Benchmark Suite for Video Generative Models/layout.json @@ -0,0 +1,8960 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 93, + 520, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 93, + 520, + 110 + ], + "spans": [ + { + "bbox": [ + 73, + 93, + 520, + 110 + ], + "type": "text", + "content": "VBenchmark: Comprehensive Benchmark Suite for Video Generative Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "spans": [ + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": "Ziqi Huang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Yinan He" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Jiashuo Yu" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Fan Zhang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Chenyang Si" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Yuming Jiang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " \nYuanhan Zhang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Tianxing Wu" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Qingyang Jin" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Nattapol Chanpaisit" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " \nYaohui Wang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Xinyuan Chen" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Limin Wang" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{4,2}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Dahua Lin" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2,3\\boxtimes}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 67, + 132, + 539, + 177 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "spans": [ + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "text", + "content": "The Chinese University of Hong Kong " + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 189, + 515, + 219 + ], + "type": "text", + "content": "Nanjing University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 229, + 424, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 229, + 424, + 243 + ], + "spans": [ + { + "bbox": [ + 168, + 229, + 424, + 243 + ], + "type": "text", + "content": "https://vchitect.github.io/VBench-project/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 264, + 269, + 437 + ], + "blocks": [ + { + "bbox": [ + 50, + 264, + 269, + 437 + ], + "lines": [ + { + "bbox": [ + 50, + 264, + 269, + 437 + ], + "spans": [ + { + "bbox": [ + 50, + 264, + 269, + 437 + ], + "type": "image", + "image_path": "53e38a1b6b63f2c36f3d80e8036ca07f24f19673565536d356f21e1c424766cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 442, + 547, + 520 + ], + "lines": [ + { + "bbox": [ + 45, + 442, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 547, + 520 + ], + "type": "text", + "content": "Figure 1. Overview of VBench. We propose VBench, a comprehensive benchmark suite for video generative models. We design a comprehensive and hierarchical Evaluation Dimension Suite to decompose \"video generation quality\" into multiple well-defined dimensions to facilitate fine-grained and objective evaluation. For each dimension and each content category, we carefully design a Prompt Suite as test cases, and sample Generated Videos from a set of video generation models. For each evaluation dimension, we specifically design an Evaluation Method Suite, which uses a carefully crafted method or designated pipeline for automatic objective evaluation. We also conduct Human Preference Annotation for the generated videos for each dimension and show that VBench evaluation results are well aligned with human perceptions. VBench can provide valuable insights from multiple perspectives." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 259, + 263, + 544, + 437 + ], + "blocks": [ + { + "bbox": [ + 259, + 263, + 544, + 437 + ], + "lines": [ + { + "bbox": [ + 259, + 263, + 544, + 437 + ], + "spans": [ + { + "bbox": [ + 259, + 263, + 544, + 437 + ], + "type": "image", + "image_path": "c6d763e4a79961e303efc773cb5e13ede49ac8e6e10f18e9aabe77fa894557da.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 529, + 192, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 529, + 192, + 542 + ], + "spans": [ + { + "bbox": [ + 143, + 529, + 192, + 542 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "content": "Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBenchmark, a comprehensive benchmark suite that dissects \"video generation quality\" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBenchmark has three appealing properties: 1) Comprehensive Dimensions: VBenchmark comprises" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 530, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 545, + 710 + ], + "type": "text", + "content": "16 dimensions in video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc.). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VBench to drive forward the field of video generation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 276, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 276, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 276, + 713 + ], + "type": "text", + "content": "*equal contributions. corresponding authors. Code is available" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 45, + 251, + 210 + ], + "blocks": [ + { + "bbox": [ + 82, + 45, + 251, + 210 + ], + "lines": [ + { + "bbox": [ + 82, + 45, + 251, + 210 + ], + "spans": [ + { + "bbox": [ + 82, + 45, + 251, + 210 + ], + "type": "image", + "image_path": "a3d0229a0445214c16208b34564756131840011b11ee92cb4c8960824875f840.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 288, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 288, + 270 + ], + "type": "text", + "content": "Figure 2. VBench Evaluation Results of Video Generative Models. We visualize the evaluation results of four video generation models in 16 VBench dimensions. We normalize the results per dimension for clearer comparisons. For comprehensive numerical results, please refer to Table 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 282, + 128, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 128, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 128, + 295 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 506 + ], + "type": "text", + "content": "Image generation models have made rapid progress in the past few years, such as Variational Autoencoders (VAEs) [51], Generative Adversarial Networks (GANs) [6, 19, 20, 24, 39, 41, 44-47, 67], vector quantized (VQ) based approaches [16, 40, 85], and diffusion models [31, 78, 79]. This fuels recent explorations in video generation [3, 5, 10, 21, 23, 27, 29, 34, 52, 62, 66, 77, 86, 87, 89, 91, 101, 107-109], which goes beyond static imagery and models the dynamics and kinematics of real-world scenes. With the growth of video generation models, there arises a critical need for effective evaluation methods. The evaluation should be able to accurately reflect human perception of generated videos, providing reliable measures of a model's performance. Additionally, it should reflect each model's specific strengths and weaknesses, offering insights that inform the data, training, and architectural choices of future video generation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 639 + ], + "type": "text", + "content": "However, existing metrics for video generation such as Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83, 84], and CLIPSIM [73] are inconsistent with human judgement [15, 69]. Meanwhile, the Video Quality Assessment (VQA) methods [55, 82, 93-99] are primarily designed for real videos, thereby neglecting the unique challenges posed by generative models, such as artifacts in synthesized videos. Hence, there is a pressing need for an evaluation framework that aligns closely with human perception, and specifically designed for the characteristics of video generation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 640, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 700 + ], + "type": "text", + "content": "To this end, we introduce VBench, a comprehensive benchmark suite for evaluating video generation model performance. VBench has three appealing properties: 1) comprehensive evaluation dimensions, 2) human alignment, and 3) valuable insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "First, our framework includes an evaluation dimension" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "type": "text", + "content": "suite that employs a hierarchical and disentangled approach to the decomposition of \"video generation quality\". This suite systematically breaks down the evaluation into two primary dimensions at a coarse level: Video Quality and Video-Condition Consistency. Each of these dimensions is further subdivided into more granular criteria. This hierarchical separation ensures that each dimension isolates and evaluates a single aspect of video quality, without interference from other variables, as illustrated in Figure 1. Recognizing video generation's unique challenges, we have tailored evaluation dimensions to its specific characteristics. For example, in terms of Video Quality, maintaining consistent subject identity (e.g., a teddy bear) in generated videos is crucial, and is a problem rarely encountered in real-world videos. Additionally, Video-Condition Consistency is vital for conditional video generation tasks, requiring its dedicated evaluation criteria. For each evaluation dimension, we carefully prepared around 100 text prompts as test cases for text-to-video (T2V) generation, and devised specialized evaluation methods tailored to each dimension. In addition to multi-dimensional evaluations, we also assess T2V models across diverse content categories. We organized prompt suites for eight distinct types, such as animal, architecture, human, and scenery, allowing for a separate evaluation within each category. This exploration reveals variable competencies in T2V generation across different content types, highlighting areas of proficiency and those requiring further enhancement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 602 + ], + "type": "text", + "content": "Second, we systematically demonstrate that our evaluation method suite is closely aligned with human perception in every fine-grained evaluation dimension. We collected human preference annotations for each dimension. Specifically, we use various T2V models to sample videos from our prompt suites. Then given two videos sampled from the same prompt, we ask human annotators to indicate preferences according to each VBenchmark dimension respectively. We show that VBenchmark evaluations highly correlate with human preferences. Additionally, the human preference annotations can be utilized for multiple purposes, such as fine-tuning generation or evaluation models to enhance alignment with human perceptions. For instance, we utilize the annotations to implement Instruction Tuning within a Visual-Language Model (VLM), enhancing its T2V evaluation alignment with human preferences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Third, VBench's multi-dimensional and multicategorical approach can provide valuable insights to the video generation community. Our multi-dimensional system enables detailed feedback on the strengths and weaknesses of video generation models across various ability aspects. This approach not only ensures a comprehensive evaluation of existing models but also provides valuable insights into the training of advanced video generation models, guiding architectural and data choices" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21808" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "for improved video generation outcomes. Additionally, VBench can be readily applied to evaluate image generation models, and thus we investigate the disparities between video and image generation models. In Section 5, we discuss in detail on various observations and insights drawn from VBench evaluations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 148, + 288, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 288, + 209 + ], + "type": "text", + "content": "We are open-sourcing VBench, including its evaluation dimension suite, evaluation method suite, prompt suite, generated videos, and the dataset of human preference annotations. We also encourage more video generation models to participate in the VBench challenge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 231, + 138, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 138, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 138, + 243 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 434 + ], + "type": "text", + "content": "Video Generative Models. Recently, diffusion models [14, 31, 78, 79] have achieved significant progress in image synthesis [25, 37, 38, 68, 71, 74, 75], and enabled a line of works towards video generation [5, 9, 22, 26, 28, 29, 32, 33, 42, 50, 62, 63, 77, 87, 91, 102, 103, 109, 113]. Many recent diffusion-based works [29, 62, 87, 91] are text-to-video (T2V) models. Other guidance modalities are also available, including image-to-video [11, 13, 17, 106], video-to-video [8, 59, 70, 72, 105], and a variety of control maps [12, 43, 50, 64, 90, 110, 111] such as pose, depth, and sketch. The boom of video generation models requires a comprehensive evaluation system to inform their current capabilities and guide future developments, and VBench takes the initiative in providing a comprehensive benchmark suite for fine-grained and human-aligned evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "content": "Evaluation of Visual Generative Models. Existing video generation models typically use metrics like Inception Score (IS) [76], Fréchet inception distance (FID) [30], Fréchet Video Distance (FVD) [83], and CLIPSIM [73] for evaluation. The UCF-101 [80] dataset's class labels often serve as text prompts for IS, FID, and FVD, whereas MSR-VTT [104]'s human-labeled video captions are used for CLIPSIM. Despite covering various real-world scenarios, these prompts lack diversity and specificity, limiting accurate and fine-grained evaluation of video generation. For text-to-image (T2I) models, several benchmarks [2, 4, 35, 54, 65, 75, 88] are proposed to assess various capabilities like compositionality [35] and editing ability [4, 88]. However, video generative models still lack comprehensive evaluation benchmarks for detailed and human-aligned feedback. Our work differs from concurrent research [60, 61] in three key ways: 1) We have created 16 distinct evaluation dimensions, each with specialized prompts for precise assessment; 2) We have empirically validated that every dimension aligns closely with human perception; 3) Our multi-dimensional and multi-categorical evaluation offers valuable and comprehensive insights into video generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 71, + 391, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 391, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 391, + 83 + ], + "type": "text", + "content": "3. VBench Suite" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "type": "text", + "content": "In this section, we introduce the main components of VBench. In Section 3.1, we present our rationale for designing the 16 evaluation dimensions, as well as each dimension's definition and evaluation method. We then elaborate on the prompt suites we use in Section 3.2. To validate VBench's alignment with human perception, we conduct human preference annotation for each dimension (see Section 3.3). The experiments and the insights drawn from VBench will be detailed in Section 4 and Section 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 205, + 459, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 459, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 459, + 217 + ], + "type": "text", + "content": "3.1. Evaluation Dimension Suite" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 223, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 247 + ], + "type": "text", + "content": "We first introduce our evaluation dimensions and their corresponding evaluation methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 247, + 546, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 474 + ], + "type": "text", + "content": "Existing evaluation metrics like FVD [83] often conclude video generation model performance to a single number. This oversimplifies the evaluation and has several risks. First, a single number can obscure an individual model's strengths and weaknesses, and it fails to provide insights into specific areas where a model excels or underperforms. This makes it challenging to derive insights for future architectural and training designs based on single-valued metrics. Second, the notion of \"high-quality video generation\" is complex and multifaceted, with individuals prioritizing different video attributes based on the intended application. For instance, some may prioritize the absence of temporal flickering, while others may consider fidelity to the text prompt as the most significant, with less emphasis on flickering. Therefore, in contrast with performing single-valued evaluations of video generation quality, we propose a disaggregated approach by decomposing the brand notion of \"video generation performance\" into multiple discrete dimensions for fine-grained evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "Specifically, we break \"video generation quality\" down into 16 disentangled dimensions in a top-down manner, with each evaluation dimension assessing one aspect of video generation quality. On the top level, we evaluate T2V performance from two broad perspectives: 1) Video Quality — \"Without considering alignment with the text prompt, does the video alone look good?\", which focuses on the perceptual quality of the synthesized video, and does not consider the input condition (e.g., text prompt), and 2) Video-Condition Consistency — \"Is the video consistent with what the user wants to generate?\", which focuses on whether the synthesized video is consistent with the guiding condition that the user provides (e.g., the text prompt for T2V generation). Under both \"Video Quality\" and \"Video-Condition Consistency\", we further break the coarse-grained dimensions into more fine-grained dimensions, as shown in Figure 1. We briefly introduce each dimension here. Please refer to the Supplementary File for the detailed definition and evaluation method of each dimension." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21809" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 139, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 139, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 139, + 85 + ], + "type": "text", + "content": "3.1.1 Video Quality" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 286, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 286, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 286, + 189 + ], + "type": "text", + "content": "We split \"Video Quality\" into two disentangled aspects, \"Temporal Quality\" and \"Frame-Wise Quality\", where the former only considers the cross-frame consistency and dynamics, and the latter only considers the quality of each individual frame without taking temporal quality into concern. For \"Temporal Quality\", we further devise five evaluation dimensions, where each focusing on a different aspect of temporal quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 190, + 286, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 286, + 249 + ], + "type": "text", + "content": "Temporal Quality - Subject Consistency. For a subject (e.g., a person, a car, or a cat) in the video, we assess whether its appearance remains consistent throughout the whole video. To this end, we calculate the DINO [7] feature similarity across frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 250, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 286, + 286 + ], + "type": "text", + "content": "Temporal Quality - Background Consistency. We evaluate the temporal consistency of the background scenes by calculating CLIP [73] feature similarity across frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 287, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 286, + 334 + ], + "type": "text", + "content": "Temporal Quality - Temporal Flickering. Generated videos can exhibit imperfect temporal consistency at local and high-frequency details. We take static frames and compute the mean absolute difference across frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 335, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 286, + 430 + ], + "type": "text", + "content": "Temporal Quality - Motion Smoothness. Subject Consistency and Background Consistency focus on temporal consistency of the \"look\" instead of the smoothness of \"movement and motion\". We believe it is important to evaluate whether the motion in the generated video is smooth, and follows the physical law of the real world. We utilize the motion priors in the video frame interpolation model [58] to evaluate the smoothness of generated motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 432, + 286, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 286, + 503 + ], + "type": "text", + "content": "Temporal Quality - Dynamic Degree. Since a completely static video can score well in the aforementioned temporal quality dimensions, it is important to also evaluate the degree of dynamics (i.e., whether it contains large motions) generated by each model. We use RAFT [81] to estimate the degree of dynamics in synthesized videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 505, + 286, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 286, + 576 + ], + "type": "text", + "content": "Frame-Wise Quality - Aesthetic Quality. We evaluate the artistic and beauty value perceived by humans towards each video frame using the LAION aesthetic predictor [53]. It can reflect aesthetic aspects such as the layout, the richness and harmony of colors, the photo-realism, naturalness, and artistic quality of the video frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 578, + 286, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 286, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 286, + 637 + ], + "type": "text", + "content": "Frame-Wise Quality - Imaging Quality. Imaging quality refers to the distortion (e.g., over-exposure, noise, blur) presented in the generated frames, and we evaluate it using the MUSIQ [49] image quality predictor trained on the SPAQ [18] dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "type": "text", + "content": "3.1.2 Video-Condition Consistency" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "We mainly dissect \"Video-Condition Consistency\" into \"Semantics\" (i.e., the type of the entities and their attributes) and \"Style\" (i.e., whether the generated video" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "is consistent with user-requested style), with each decomposed into more fine-grained dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 96, + 544, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 544, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 544, + 133 + ], + "type": "text", + "content": "Semantics - Object Class. We use GRiT [100] to detect the success rate of generating the specific class of objects depicted in the text prompt." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 133, + 544, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 544, + 204 + ], + "type": "text", + "content": "Semantics - Multiple Objects. Other than generating a single object of a particular class, the ability to compose multiple objects from different classes in the same frame is also an essential ability in video generation. We detect the success rate of generating all the objects specified in the text prompt within each video frame." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 205, + 544, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 544, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 544, + 265 + ], + "type": "text", + "content": "Semantics - Human Action. Human action is an important aspect in human-centric video generation. We apply UMT [57] to evaluate whether human subjects in generated videos can accurately execute the specific actions mentioned in the text prompts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 265, + 544, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 544, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 544, + 300 + ], + "type": "text", + "content": "Semantics - Color. To evaluate whether synthesized object colors align with the text prompt, we use GRiT [100] for color captioning and comparison with expected colors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 301, + 544, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 544, + 361 + ], + "type": "text", + "content": "Semantics - Spatial Relationship. Other than classes and attributes of synthesized objects, we also evaluate whether their spatial relationship follows what is specified by the text prompt. We focus on four primary types of spatial relationships, and perform rule-based evaluation similar to [35]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 361, + 544, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 544, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 544, + 445 + ], + "type": "text", + "content": "Semantics - Scene. We need to evaluate whether the synthesized video is consistent with the intended scene described by the text prompt. For example, when prompted \"ocean\", the generated video should be \"ocean\" instead of \"river\". We use Tag2Text [36] to caption the generated scenes, and then check its correspondence with scene descriptions in the text prompt." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 446, + 544, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 544, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 544, + 541 + ], + "type": "text", + "content": "Style - Appearance Style. Apart from semantics consistency with the text prompt, another important pillar in video-condition consistency is style. There are many styles that alter the look, color, and texture of synthesized video frames, such as \"oil painting style\", \"black and white style\", \"watercolor painting style\", \"cyberpunk style\", \"black and white\" etc. We calculate the CLIP [73] feature similarity between synthesized frames and these style descriptions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 542, + 544, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 544, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 544, + 601 + ], + "type": "text", + "content": "Style - Temporal Style. Apart from appearance styles, videos also have temporal styles like various camera motions. We use ViCLIP [92] to calculate the video feature and the temporal style description feature similarity to reflect temporal style consistency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 601, + 544, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 544, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 544, + 650 + ], + "type": "text", + "content": "Overall Consistency. We further use overall video-text consistency computed by ViCLIP [92] on general text prompts as an aiding metric to reflect both semantics and style consistency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 658, + 391, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 391, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 391, + 671 + ], + "type": "text", + "content": "3.2. Prompt Suite" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 544, + 713 + ], + "type": "text", + "content": "The sampling procedure of current diffusion-based video generation models [29, 87, 91] is time-consuming (e.g., 90 seconds per video for LaVie [91], and more than 2 minutes" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21810" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 49, + 370, + 167 + ], + "blocks": [ + { + "bbox": [ + 48, + 49, + 370, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 49, + 370, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 49, + 370, + 167 + ], + "type": "image", + "image_path": "2061c7ebb5579f224dfebbfe68f8a8fb80e68f609e1dc4b211d1fc03181a7bea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 373, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 373, + 201 + ], + "type": "text", + "content": "Figure 3. Prompt Suite Statistics. The two graphs provide an overview of our prompt suites. Left: the word cloud to visualize word distribution of our prompt suites. Right: the number of prompts across different evaluation dimensions and different content categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 156, + 544, + 201 + ], + "lines": [ + { + "bbox": [ + 375, + 156, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 375, + 156, + 544, + 201 + ], + "type": "text", + "content": "Figure 4. Interface for Human Preference Annotation. Top: prompt and question. Right: choices that annotators can make. Bottom left: control for stop and playback." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 377, + 49, + 542, + 150 + ], + "blocks": [ + { + "bbox": [ + 377, + 49, + 542, + 150 + ], + "lines": [ + { + "bbox": [ + 377, + 49, + 542, + 150 + ], + "spans": [ + { + "bbox": [ + 377, + 49, + 542, + 150 + ], + "type": "image", + "image_path": "736f108c33ad8d4ce70b2d30beae340d6d8ebff7191df1d985ba75fa14571dc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 212, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 289, + 297 + ], + "type": "text", + "content": "per video for CogVideo [34]). Therefore, we need to control the amount of test cases for efficient evaluation. Meanwhile, we need to maintain the diversity and comprehensiveness of our prompt suite, so we design compact yet representative prompts in terms of both the evaluation dimensions and the content categories. We visualize our prompt suite distributions in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 301, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 517 + ], + "type": "text", + "content": "Prompt Suite per Dimension. For each VBenchmark evaluation dimension, we carefully designed a suite of around 100 prompts as test cases. The prompt suite is carefully curated to probe the specific ability relevant to the dimension tested. For example, for the \"Subject Consistency\" dimension which aims to evaluate the consistency of subjects' appearances throughout the video, we ensure every prompt has a movable subject (e.g., animals or vehicles) performing non-static actions, where their consistency might be compromised due to inconsistency introduced by their movements or changing locations. In \"Object Class\" dimension, we ensure the existence of a specific class of object in every prompt. For \"Human Action\", each test prompt contains a human subject performing a well-defined action from the Kinetics-400 dataset [48], where 100 representative actions are selected with minimal semantic overlaps among themselves. Please refer to the Supplementary File for the design rationale of the prompt suite for all 16 dimensions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Prompt Suite per Category. When designing prompts for each dimension, the focus was to showcase models' ability in that specific dimension. We further incorporate prompt suites for eight content categories to provide insights into the performance across varied content types. To this end, we prepare a collection of human-curated prompts from the Internet and divide them into 8 distinctive categories following YouTube's categorization. Subsequently, we feed both the category labels and prompts into a Large Language Model (LLM) [112] (see more implementation details in Supplementary File), obtaining multi-label outputs for each caption. We select 800 prompts and manually clean their labels to serve as per-category prompt suites. Finally, we obtain 100 prompts for each of these eight categories: Animal, Architecture, Food, Human, Lifestyle, Plant, Scenery, and Vehicles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 212, + 473, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 473, + 224 + ], + "type": "text", + "content": "3.3. Human Preference Annotation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 327 + ], + "type": "text", + "content": "We perform human preference labeling on massive generated videos. The primary goal is to validate VBench evaluation's alignment with human perception in each of the 16 evaluation dimensions, and the verification results will be detailed in Section 4.2. We also show that our human preference annotations can be useful in future tasks of finetuning generation and evaluation models to enhance alignment with human perceptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": "Data Preparation. Given a text prompt " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", and four video generation models to be evaluated " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\{A, B, C, D\\}" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", we use each model to generate a video, forming a \"group\" of videos " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "G_{i,j} = \\{V_{i,A,j}, V_{i,B,j}, V_{i,C,j}, V_{i,D,j}\\}" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ". For each prompt " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", we sample five such groups of videos " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\{G_{i,0}, G_{i,1}, G_{i,2}, G_{i,3}, G_{i,4}\\}" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ". For each group, we pair the videos up in pair-wise combinations, yielding six pairs: " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_A, V_B)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_A, V_C)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_A, V_D)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_B, V_C)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_B, V_D)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "(V_C, V_D)" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": ", and ask human annotators to indicate their preferred video for each pair. Within the VBench evaluation framework, a prompt suite of " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": " prompts produces " + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "inline_equation", + "content": "N \\times 5 \\times 6" + }, + { + "bbox": [ + 304, + 328, + 547, + 484 + ], + "type": "text", + "content": " pairwise video comparisons. The video order within each pair is randomized to ensure unbiased annotation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 548, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 548, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 548, + 676 + ], + "type": "text", + "content": "Human Labeling Rules. Specifically, the human annotators are asked to only consider the specific evaluation dimension of interest and select the preferred video. For example, in Figure 4, for the Appearance Style dimension, the question is \"Is the style of the video in the Van Gogh style?\", and human annotators are instructed to only focus on whether the generated video's style belongs to the Van Gogh style and should not consider other quality aspects of the generated video, such as potential issues like the degree of temporal flickering. In the example in this figure, video A resembles the Van Gogh better than video B, and the annotator is expected to select \"A is better\". For every dimension, we carefully prepare instructions and train the human annotators to understand the definition of the dimension, and perform multiple quality assurance protocols via a pre-labeling trial, and two rounds of post-labeling checks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Annotations for VLM Tuning. We map VBench evaluation scores from various dimensions to the scale of 0-10 and combine them with human preference annotations to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "21811" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 87, + 470, + 222 + ], + "blocks": [ + { + "bbox": [ + 46, + 51, + 547, + 85 + ], + "lines": [ + { + "bbox": [ + 46, + 51, + 547, + 85 + ], + "spans": [ + { + "bbox": [ + 46, + 51, + 547, + 85 + ], + "type": "text", + "content": "Table 1. VBenchmark Evaluation Results per Dimension. This table compares the performance of four video generation models across each of the 16 VBenchmark dimensions. A higher score indicates relatively better performance for a particular dimension. We also provide two specially built baselines, i.e., Empirical Min and Max (the approximated achievable min and max scores for each dimension), as references." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 123, + 87, + 470, + 222 + ], + "lines": [ + { + "bbox": [ + 123, + 87, + 470, + 222 + ], + "spans": [ + { + "bbox": [ + 123, + 87, + 470, + 222 + ], + "type": "table", + "html": "
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessDynamic DegreeAesthetic QualityImaging QualityObject Class
LaVie [91]91.41%97.47%98.30%96.38%49.72%54.94%61.90%91.82%
ModelScope [62, 87]89.87%95.29%98.28%95.79%66.39%52.06%58.57%82.25%
VideoCrafter [29]86.24%92.88%97.60%91.79%89.72%44.41%57.22%87.34%
CogVideo [34]92.19%96.20%97.64%96.47%42.22%38.18%41.03%73.40%
Empirical Min14.62%26.15%62.93%70.60%0.00%0.00%0.00%0.00%
Empirical Max100.00%100.00%100.00%99.75%100.00%100.00%100.00%100.00%
ModelsMultiple ObjectsHuman ActionColorSpatial RelationshipSceneAppearance StyleTemporal StyleOverall Consistency
LaVie [91]33.32%96.80%86.39%34.09%52.69%23.56%25.93%26.41%
ModelScope [62, 87]38.98%92.40%81.72%33.68%39.26%23.39%25.37%25.67%
VideoCrafter [29]25.93%93.00%78.84%36.74%43.36%21.57%25.42%25.21%
CogVideo [34]18.11%78.20%79.57%18.24%28.24%22.01%7.80%7.70%
Empirical Min0.00%0.00%0.00%0.00%0.00%0.09%0.00%0.00%
Empirical Max100.00%100.00%100.00%100.00%82.22%28.55%36.40%36.40%
", + "image_path": "b437b1a8fb969d649bfa861638817df4cce4344633d911b4c86da6bfe2afb0e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 280 + ], + "type": "text", + "content": "form the instruction data, which is then used to fine-tune the pre-trained VideoChat [56] model to demonstrate improved evaluation capabilities. For implementation details and tuning results, please refer to the Supplementary File." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 128, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 128, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 128, + 303 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 366 + ], + "type": "text", + "content": "We adopt the video generation models LaVie [91], ModelScope [62, 87], VideoCrafter [29], and CogVideo [34] for VBench evaluation, and more will be added as they become open-sourced. Details of the models and sampling procedures are in the Supplementary File." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 194, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 194, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 194, + 384 + ], + "type": "text", + "content": "4.1. Per-Dimension Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 387, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 483 + ], + "type": "text", + "content": "For every dimension, we calculate the VBench scores using the evaluation method suite described in Section 3.1, and show the results using Figure 2 and Table 1. We additionally designed three reference baselines, namely Empirical Max, Empirical Min, and WebVid-Avg. The first two approximate the maximum / minimum scores that videos might be able to achieve, and WebVid-Avg reflects the WebVid-10M [1] dataset quality in each VBench dimension." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": "Empirical Max. For most dimensions, to approximate the maximum achievable values, we first retrieve WebVid-10M [1] videos according to our prompt suites. We use CLIP [73] to extract text features of both WebVid-10M's captions and our prompts. For each prompt, we retrieve the top-5 WebVid-10M videos according to text feature similarity with the given prompt. Given that the generated videos are usually 2 seconds in length, we randomly select a 2-second segment from each retrieved video and sample frames at 8 frames per second (FPS). For each dimension, we use the retrieved videos according to its prompt suite and report the highest-scoring video's result as Empirical Max." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Empirical Min. To approximate the minimum achievable values, we use randomly generated 2-second Gaussian noise clips to calculate results for the \"Video-Condition Consistency\" dimensions. For most \"Video Quality\" dimensions, we select frames from real videos and design frame concatenation for each dimension, approximating the minimum score achievable for each VBench dimension." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 231, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 315 + ], + "type": "text", + "content": "WebVid-Avg. Similar to Empirical Max, we compute the average for each dimension on retrieved WebVid-10M [1] videos. This baseline could reflect the average per-dimension quality of the commonly used video generation training dataset WebVid-10M, and provide a reference for model performances. The comparison against WebVid-Avg and Empirical Max is visualized in Figure 6 (b)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 319, + 518, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 518, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 518, + 331 + ], + "type": "text", + "content": "4.2. Validating Human Alignment of VBench" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "content": "To validate that our evaluation method can faithfully reflect human perception, we performed a large-scale human annotation for each dimension, as mentioned in Section 3.3. We show the correlation between VBenchmark evaluation results and human preference annotations in Figure 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 393, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 477 + ], + "type": "text", + "content": "Win Ratio. Given the human labels, we calculate the win ratio of each model. During pairwise comparisons, if a model's video is selected as better, then the model scores 1 and the other model scores 0. If there is a tie, then both models score 0.5. For each model, the win ratio is calculated as the total score divided by the total number of pairwise comparisons participated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 478, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 550 + ], + "type": "text", + "content": "Per-Dimension Evaluation. For each evaluation dimension, we calculate the model win ratio based on (1) VBenchmark results, and (2) human annotation results, respectively, and compute their correlations, as shown in Figure 5. We observe that VBenchmark's per-dimension evaluation results are highly correlated with human preference annotations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 554, + 446, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 446, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 446, + 566 + ], + "type": "text", + "content": "4.3. Per-Category Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": "We evaluate the T2V models across eight different content categories, by generating videos based on Prompt Suite per Category described in Section 3.2, and then calculating their performance across different evaluation dimensions. Figure 7 visualizes the evaluation results of each model in terms of the eight content categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 649, + 520, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 649, + 520, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 649, + 520, + 662 + ], + "type": "text", + "content": "4.4. Video Generation V.S. Image Generation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "We conduct a comparative analysis of the frame-wise generation capability exhibited by text-to-video (T2V) models and text-to-image (T2I) models with two primary objectives: first, to assess the extent to which T2V models" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21812" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 52, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 51, + 52, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 51, + 52, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 51, + 52, + 541, + 186 + ], + "type": "image", + "image_path": "18e1edf20a57807908f97ff5944543f3b40659999ea81933f208949b94d8926c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "type": "text", + "content": "Figure 5. Validate VBench's Human Alignment. Our experiments show that VBench evaluations across all dimensions closely match human perceptions. Each plot shows the alignment verification result of a specific VBench dimension. In each plot, a dot represents the human preference win ratio (horizontal axis) and VBench evaluation win ratio (vertical axis) for a particular video generation model. We linearly fit a straight line to visualize the correlation, and calculate the Spearman's correlation coefficient " + }, + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "type": "inline_equation", + "content": "(\\rho)" + }, + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "type": "text", + "content": " for each dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 240, + 290, + 363 + ], + "blocks": [ + { + "bbox": [ + 48, + 240, + 290, + 363 + ], + "lines": [ + { + "bbox": [ + 48, + 240, + 290, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 290, + 363 + ], + "type": "image", + "image_path": "271dbb03f501abe078ec14e0f96007705aa2898b4577a30f243fea91fd1d8dd0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 364, + 288, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 364, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 288, + 440 + ], + "type": "text", + "content": "Figure 6. More Comparisons of Video Generation Models with Other Models and Baselines. We use VBench to evaluate other models and baselines for further comparative analysis of T2V models. (a) Comparison with text-to-image (T2I) generation models. (b) Comparison with WebVid-Avg and Empirical Max baselines. See the Supplementary File for comprehensive numerical results and details on normalization methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 452, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 594 + ], + "type": "text", + "content": "have successfully inherited the frame-wise generative capability of the T2I models; and second, to investigate the frame-wise generation capability gap between existing T2I and T2V models. As an initial exploration into this problem, we compare video generation models with three image generation models, namely Stable Diffusion (SD) 1.4 [74], SD2.1 [74], and SDXL [71]. We choose 10 V-Bench dimensions that can encompass frame-wise generation capabilities, and sample frames from all the image and video generation models according to Prompt Suite per Evaluation Dimension described in Section 3.2. Figure 6 (a) visualizes the evaluation results of T2V versus T2I models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 602, + 188, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 188, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 188, + 615 + ], + "type": "text", + "content": "5. Insights and Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "content": "In this section, we discuss the observations and insights we draw from our comprehensive evaluation experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "- Trade-off across Ability Dimensions. We have noticed a trade-off in video generation models between 1) temporal consistency (Subject Consistency, Background Consistency, Temporal Flickering, Motion Smoothness) and 2) Dynamic Degree. Models strong in temporal consistency often have a lower Dynamic Degree, as these two aspects are some" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 387 + ], + "type": "text", + "content": "what complementary (see Figure 2 and Table 1). For example, LaVie excels in Background Consistency and Temporal Flickering but has a low Dynamic Degree, probably because generating relatively static scenes can \"cheat\" to get high temporal consistency scores. Conversely, VideoCrafter shows a high Dynamic Degree but suffers from poor performance in all temporal consistency dimensions. This trend highlights the current challenge for models to achieve temporal consistency with dynamic content of large motions. Future research should focus on enhancing both aspects simultaneously, as improving only one might indicate compromising the other." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 581 + ], + "type": "text", + "content": "- Uncovering Hidden Potential of T2V Models in Specific Content Categories. Our analysis reveals that the capabilities of some models vary significantly across different content types. For instance, for Aesthetic Quality, CogVideo scores well for Food (see Figure 7 rightmost chart), whereas it underperforms in others like Animal and Vehicles. The average results across various prompts might suggest a lower overall \"Aesthetic Quality\" (as seen in Figure 2), but CogVideo demonstrates relatively strong aesthetics in at least the Food category. This suggests that with tailored training data and strategies, CogVideo could potentially match other models in aesthetics by improving such ability in other content types. Therefore, we recommend evaluating video generation models not just based on ability dimensions but also considering specific content categories to uncover their hidden potential." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "- Bottleneck in Temporally Complex Categories Affecting Spatial and Temporal Performance. For spatially complex categories (e.g., Animal, LifeStyle, Human, Vehicles), models all perform relatively poorly mainly in Aesthetic Quality (shown in Figure 7). This is likely due to the challenges in synthesizing harmonious color schemes, articulated structures, and appealing layouts amidst complex elements. On the other hand, for categories involving complex and intense motions like Human and Vehicle (see their Dynamic Degree in Supplementary File), performance is relatively poor across all dimen" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21813" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 48, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 48, + 48, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 48, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 48, + 545, + 173 + ], + "type": "image", + "image_path": "72706807e6048bcc1277c117c1ec346373f4f827e80256234c39f52dface877d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 548, + 211 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 548, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 548, + 211 + ], + "type": "text", + "content": "Figure 7. VBench Results across Eight Content Categories (best viewed in color). For each chart, we plot the VBench evaluation results across eight different content categories, benchmarked by our Prompt Suite per Category. The results are linearly normalized between 0 and 1 for better visibility across categories. See the Supplementary File for comprehensive numerical results, and normalization details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 292 + ], + "type": "text", + "content": "sions. This suggests that motion complexity and dynamic intensity significantly hinder synthesis, impacting both spatial and temporal dimensions, probably because poor temporal modeling results in distorted and blurred imagery. This highlights the need for improved handling of dynamic motions in video generation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "content": "- Challenges of Data Quantity in Handling Complex Categories like Human. The WebVid-10M dataset [1] allocates " + }, + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "content": " of its content to the Human category, which is the largest share among the eight categories (see statistics in Supplementary File). However, the Human category exhibits one of the poorest results among eight categories (see Figure 7). This suggests that merely increasing data volume may not significantly enhance performance in complex categories like Human. A potential approach could involve integrating human-related priors or controls, such as skeletons, to better capture the articulated nature of human appearances and movements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": "- Prioritizing Data Quality Over Quantity in Large-Scale Datasets. For Aesthetic Quality, Figure 7 shows that the Food category almost always tends to have the highest scores among all categories. This is corroborated by the WebVid-10M dataset [1], where Food ranks highest in Aesthetic Quality according to VBench evaluation (refer to Supplementary File for more details), despite comprising just " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " of the total data. This observation suggests that at million scales, data quality might hold greater importance than quantity. Furthermore, VBench's evaluation dimensions can be potentially useful for cleaning datasets in specified quality dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "- Compositionality: T2I versus T2V. As shown in Figure 6 (a), T2V models significantly underperform in Multiple Objects and Spatial Relationship compared to T2I models (especially SDXL [71]), which highlights the need to enhance compositionality (i.e., correctly composing multiple objects in the same frame). We believe possible solutions might be: 1) curating training data incorporating multiple objects with corresponding captions explicitly depicting this compositionality, or 2) adding intermediate spatial control modules or modalities during video synthesis. Furthermore, the disparity of the text encoders might also account for the per" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 220, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 293 + ], + "type": "text", + "content": "formance gap. As T2I models leverage bigger (OpenCLIP ViT-H for SD2.1 [74]) or more sophisticated (CLIP ViT-L & OpenCLIP ViT-G for SDXL [71]) text encoders compared with T2V models (e.g., CLIP ViT-L alone for LaVie), more representative text embeddings could be featuring more accurate object composition comprehension." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 379, + 316 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 450 + ], + "type": "text", + "content": "With the growing focus on video generation, comprehensive evaluation of these models is essential to assess current advancements and guide future research. In this work, we take the first step forward and propose VBench, a comprehensive benchmark suite for evaluating video generation models. With its multi-dimensional, human-aligned, and insight-rich properties, VBench could play vital roles for evaluating future video generation models and inspiring further advancements in video generation. We believe that VBench is a significant contribution to the video generation and evaluation community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 498 + ], + "type": "text", + "content": "Limitations and Future Work. We plan to extend VBenchmark to include more models that recently became available, and extend the evaluations aspects to additional video generation tasks, like image-to-video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 557 + ], + "type": "text", + "content": "Potential Negative Societal Impacts. We also recognize the importance of considering ethical aspects in future iterations of VBench. While VBench currently does not assess safety and equality dimensions, we urge users to exercise caution with open-sourced video generation models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s); the National Key R&D Program of China (2022ZD0160201); National Natural Science Foundation of China (62076119, 61921006, 62102150); the Science and Technology Commission of Shanghai Municipality (23YF1461900, 23QD1400800); and Shanghai Artificial Intelligence Laboratory. We thank Shangchen Zhou, Jianyi Wang, and Ruicheng Feng for helpful suggestions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21814" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 125, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 125, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 287, + 167 + ], + "type": "text", + "content": "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 225, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 225, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 287, + 278 + ], + "type": "text", + "content": "[4] Samyadeep Basu, Mehrdad Saberi, Shweta Bhardwaj, Atoosa Malemir Chegini, Daniela Massiceti, Maziar Sanjabi, Shell Xu Hu, and Soheil Feizi. Editval: Benchmarking diffusion based text-guided image editing methods. arXiv preprint arXiv:2310.02426, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 281, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 281, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 287, + 323 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 325, + 287, + 357 + ], + "type": "text", + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 287, + 401 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 287, + 435 + ], + "type": "text", + "content": "[8] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. arXiv preprint arXiv:2308.09592, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 437, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 437, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 58, + 437, + 287, + 500 + ], + "type": "text", + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "text", + "content": "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "text", + "content": "[11] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 287, + 635 + ], + "type": "text", + "content": "[12] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "type": "text", + "content": "[13] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "type": "text", + "content": "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 545, + 184 + ], + "type": "text", + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 218 + ], + "type": "text", + "content": "[18] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 545, + 262 + ], + "type": "text", + "content": "[19] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "type": "text", + "content": "[20] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Wayne Wu, and Ziwei Liu. Unitedhuman: Harnessing multi-source data for high-resolution human generation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 310, + 545, + 353 + ], + "type": "text", + "content": "[21] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 312, + 355, + 545, + 408 + ], + "type": "text", + "content": "[22] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 545, + 464 + ], + "type": "text", + "content": "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 467, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 545, + 509 + ], + "type": "text", + "content": "[24] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 511, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 511, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 312, + 511, + 545, + 554 + ], + "type": "text", + "content": "[25] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 557, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 599 + ], + "type": "text", + "content": "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "text", + "content": "[27] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "text", + "content": "[28] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "21815" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 97, + 287, + 140 + ], + "type": "text", + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 143, + 287, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 287, + 164 + ], + "type": "text", + "content": "[31] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 167, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 167, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 167, + 287, + 220 + ], + "type": "text", + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 223, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 223, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 223, + 287, + 255 + ], + "type": "text", + "content": "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[34] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "type": "text", + "content": "[35] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350, 2023. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 348, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 287, + 392 + ], + "type": "text", + "content": "[36] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 394, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 287, + 426 + ], + "type": "text", + "content": "[37] Ziqi Huang, Kelvin C.K. Chan, Yuming Jiang, and Ziwei Liu. Collaborative diffusion for multi-modal face generation and editing. In CVPR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 429, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 429, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 429, + 287, + 472 + ], + "type": "text", + "content": "[38] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C.K. Chan, and Ziwei Liu. ReVersion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 475, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 507 + ], + "type": "text", + "content": "[39] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 509, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 287, + 552 + ], + "type": "text", + "content": "[40] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Textdriven controllable human image generation. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 597 + ], + "type": "text", + "content": "[41] Yuming Jiang, Ziqi Huang, Tianxing Wu, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained 2d and 3d facial editing via dialog. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 600, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 600, + 287, + 633 + ], + "type": "text", + "content": "[42] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2Performer: Textdriven human video generation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 678 + ], + "type": "text", + "content": "[43] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[44] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "text", + "content": "[45] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "text", + "content": "[46] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "text", + "content": "[47] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 174, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 174, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 174, + 545, + 227 + ], + "type": "text", + "content": "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 259 + ], + "type": "text", + "content": "[49] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: multi-scale image quality transformer. CoRR, abs/2108.05997, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 262, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 545, + 316 + ], + "type": "text", + "content": "[50] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 317, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 317, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 317, + 545, + 338 + ], + "type": "text", + "content": "[51] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 339, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 392 + ], + "type": "text", + "content": "[52] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 394, + 544, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 394, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 394, + 544, + 415 + ], + "type": "text", + "content": "[53] LAION-AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor, 2022.4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 416, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 469 + ], + "type": "text", + "content": "[54] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. arXiv preprint arXiv:2311.04287, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 472, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 545, + 491 + ], + "type": "text", + "content": "[55] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 494, + 545, + 536 + ], + "type": "text", + "content": "[56] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 538, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 538, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 538, + 545, + 580 + ], + "type": "text", + "content": "[57] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058, 2023. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 624 + ], + "type": "text", + "content": "[58] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 625, + 545, + 668 + ], + "type": "text", + "content": "[59] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhong-cong Xu, and Jiashi Feng. Magicedit: High-fidelity and temporally coherent video editing. arXiv preprint arXiv:2308.14749, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "text", + "content": "[60] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "21816" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "text", + "content": "[61] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 287, + 173 + ], + "type": "text", + "content": "[62] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "text", + "content": "[63] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "text", + "content": "[64] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "text", + "content": "[65] Kai Zhang Yujie Lu Xingyu Fu Wenwen Zhuang Wenhu Chen Max Ku, Tianle Li. Imagenhub: Standardizing the evaluation of conditional image generation models. In ICLR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "type": "text", + "content": "[66] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. arXiv preprint arXiv:2402.14797, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 387 + ], + "type": "text", + "content": "[67] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 287, + 443 + ], + "type": "text", + "content": "[68] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "type": "text", + "content": "[69] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 287, + 533 + ], + "type": "text", + "content": "[70] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 535, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 287, + 589 + ], + "type": "text", + "content": "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "text", + "content": "[72] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "text", + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 331, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 545, + 95 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In CVPR, 2022. 3, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 162 + ], + "type": "text", + "content": "[75] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 196 + ], + "type": "text", + "content": "[76] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "type": "text", + "content": "[77] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.2.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 253, + 545, + 285 + ], + "type": "text", + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "type": "text", + "content": "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 545, + 374 + ], + "type": "text", + "content": "[80] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 376, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 545, + 399 + ], + "type": "text", + "content": "[81] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 545, + 443 + ], + "type": "text", + "content": "[82] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C. Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE TIP, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 498 + ], + "type": "text", + "content": "[83] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 500, + 545, + 544 + ], + "type": "text", + "content": "[84] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In ICLRW, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 545, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 545, + 567 + ], + "type": "text", + "content": "[85] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 569, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 569, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 569, + 545, + 623 + ], + "type": "text", + "content": "[86] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In ICLR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "text", + "content": "[87] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 714 + ], + "type": "text", + "content": "[88] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and EditBench: Advancing and evaluating text-guided" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21817" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "image inpainting. arXiv preprint arXiv:2212.06909, 2022.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 149 + ], + "type": "text", + "content": "[89] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, et al. Magicvideo-v2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 151, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 286, + 205 + ], + "type": "text", + "content": "[90] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 286, + 259 + ], + "type": "text", + "content": "[91] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "text", + "content": "[92] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 286, + 360 + ], + "type": "text", + "content": "[93] Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fastvqa: Efficient end-to-end video quality assessment with fragment sampling. In ECCV, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 361, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 286, + 414 + ], + "type": "text", + "content": "[94] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin. Neighbourhood representative sampling for efficient end-to-end video quality assessment. arXiv preprint arXiv:2210.05357, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 286, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 458 + ], + "type": "text", + "content": "[95] Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, and Weisi Lin. Discovqa: Temporal distortion-content transformers for video quality assessment. IEEE TCSVT, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 460, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 286, + 502 + ], + "type": "text", + "content": "[96] Haoning Wu, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring opinion-unaware video quality assessment with semantic affinity criterion. In ICME, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 504, + 286, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 286, + 558 + ], + "type": "text", + "content": "[97] Haoning Wu, Liang Liao, Annan Wang, Chaofeng Chen, Jingwen Hou Hou, Erli Zhang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards robust text-prompted semantic criterion for in-the-wild video quality assessment. arXiv preprint arXiv:2304.14672, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 286, + 612 + ], + "type": "text", + "content": "[98] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In ICCV, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 614, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 286, + 667 + ], + "type": "text", + "content": "[99] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Towards explainable video quality assessment: A database and a language-prompted approach. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[100] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280, 2022. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 687 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[101] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[102] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[103] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[104] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "type": "text", + "content": "[105] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[106] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[107] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[108] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 443, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 496 + ], + "type": "text", + "content": "[109] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "text", + "content": "[110] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. Magicavatar: Multimodal avatar generation and animation. arXiv preprint arXiv:2308.14748, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[111] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 588, + 545, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 642 + ], + "type": "text", + "content": "[112] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 644, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 644, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 644, + 545, + 687 + ], + "type": "text", + "content": "[113] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21818" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_content_list.json b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4624781b97b3d24047e538d5dccaa4bc286b60b9 --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_content_list.json @@ -0,0 +1,2383 @@ +[ + { + "type": "text", + "text": "VCoder: Versatile Vision Encoders for Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 102, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jitesh Jain $^{1}$ Jianwei Yang $^{2}$ Humphrey Shi $^{1,3}$", + "bbox": [ + 292, + 179, + 676, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SHI Labs @ Georgia Tech $^{2}$ Microsoft Research, Redmond $^{3}$ Picsart AI Research (PAIR)", + "bbox": [ + 205, + 200, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/SHI-Labs/VCoder", + "bbox": [ + 328, + 219, + 632, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans possess the remarkable skill of Visual Perception, the ability to see and understand the seen, helping them make sense of the visual world and, in turn, reason. Multimodal Large Language Models (MLLM) have recently achieved impressive performance on vision-language tasks ranging from visual question-answering and image captioning to visual reasoning and image generation. However, when prompted to identify or count (perceive) the entities in a given image, existing MLLM systems fail. Working towards developing an accurate MLLM system for perception and reasoning, we propose using Versatile vision en-Coders (VCoder) as perception eyes for Multimodal LLMs. We feed the VCoder with perception modalities such as segmentation or depth maps, improving the MLLM's perception abilities. Secondly, we leverage the images from COCO and outputs from off-the-shelf vision perception models to create our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object perception task. Thirdly, we introduce metrics to assess the object perception abilities in MLLMs on our COST dataset. Lastly, we provide extensive experimental evidence proving the VCoder's improved object-level perception skills over existing Multimodal LLMs, including GPT-4V. We open-source our dataset, code, and models to promote research.", + "bbox": [ + 76, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 680, + 207, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "'Perception is the soil; reasoning, the seed. Without fertile ground, the seed cannot flourish.'", + "bbox": [ + 94, + 707, + 449, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(GPT-4 [48], 2023)", + "bbox": [ + 318, + 737, + 447, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to think and reason is one of the most remarkable traits that help humans function daily. Generally, understanding the environment precedes the act of thinking and reasoning [28]. Following the success of ChatGPT-like instruction following AI agents [3, 11, 47, 48, 60] at language understanding and reasoning, researchers have leveraged LLMs to develop instruct frameworks [14, 40, 68, 74] that can understand vision and language inputs in an effort to imitate human perception and reasoning ability. We", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a33cf8d5cc97945d86c635dc43762b30a586590bd13807c0dc74ef2fda3a96b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 267, + 883, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e452c7f4f23ac429c598f27d17519c50e8933bca261b74aa876d83c01f1c714.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 435, + 549, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "at is happening in the image?", + "bbox": [ + 535, + 452, + 663, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Difficulty Level", + "bbox": [ + 741, + 441, + 818, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1575724ae16c41b02d01f25dee5ba0c162882520f5ba01dcbc765d7af2a9ec4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 470, + 532, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 535, + 476, + 578, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume.", + "bbox": [ + 514, + 494, + 844, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81344a50af2718e479e7c1e048a6000cb55be66574ef99b11bf39adc76755975.jpg", + "image_caption": [ + "Figure 1. GPT-4V [48] (access date: Dec 16, 2023) returns impressive responses when prompted to describe complex visual scenes. However, it fails at the simple task of counting in the same scene. Our VCoder returns the correct count of people." + ], + "image_footnote": [], + "bbox": [ + 509, + 537, + 532, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VCoder", + "text_level": 1, + "bbox": [ + 535, + 542, + 578, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party.", + "bbox": [ + 511, + 558, + 870, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7986ddffcc6976567fabd9c85ad0d55109c9d17abdc5a5bae8cbdbd9c18909f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 604, + 549, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can you count the number of people in the image?", + "bbox": [ + 511, + 623, + 624, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Difficulty Level", + "bbox": [ + 537, + 650, + 612, + 661 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a465be8f016c7c2fee0dac85a91c639e77ac5f6cc354fa3880e2db2a16bf65ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 607, + 665, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 666, + 609, + 710, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, there are ten people visible in the image.", + "bbox": [ + 643, + 630, + 736, + 665 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d8d0cadd883301e5c7d9b3144bd1721d42d9a3425657cc76202caf637b6ef2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 608, + 772, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VCoder", + "text_level": 1, + "bbox": [ + 774, + 609, + 818, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, there are a total of 14 people in the image.", + "bbox": [ + 754, + 632, + 866, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "refer to such systems as Multimodal LLMs (MLLM). Although MLLMs exhibit the ability to perform complex vision-language tasks like visual captioning [2, 4, 39], image generation [26, 27, 57], visual reasoning and grounding [22, 51, 72], they often display sub-par performance at simple tasks like counting objects (Fig. 1). As shown in Fig. 2, MLLMs output incorrect object counts (people, chairs) and hallucinate about the presence (handbag, couch) of certain objects when prompted to identify entities in a visual input. The perception performance is much worse", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27992", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "USER", + "text_level": 1, + "bbox": [ + 83, + 126, + 120, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What objects can be seen in the image?", + "bbox": [ + 88, + 143, + 261, + 155 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "USER", + "text_level": 1, + "bbox": [ + 272, + 126, + 307, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What objects can be seen in the image?", + "bbox": [ + 276, + 143, + 450, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "USER", + "text_level": 1, + "bbox": [ + 460, + 126, + 493, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What objects can be seen in the image?", + "bbox": [ + 464, + 143, + 637, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "USER", + "text_level": 1, + "bbox": [ + 653, + 126, + 687, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What is happening in the image?", + "bbox": [ + 656, + 143, + 797, + 154 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f64e7236a0404dc1fadf4fe07010d4e9e5a2108a61a75e673e718f9f6690caf3.jpg", + "image_caption": [ + "Figure 2. MLLMs counting and identifying objects. As shown in the first column, GPT-4V [48] (access date: Dec 16, 2023) and LLaVA-1.5 [39] both fail at counting people. Moreover, LLaVA-1.5 [39] misses background entities like window, wall, etc. and hallucinates about the presence of a handbag. VCoder can predict the people counts and other background entities accurately except chairs. Similarly, in the second column, GPT-4V and LLaVA-1.5 fail at counting chairs while the VCoder matches the Oracle's performance. Notably, all MLLMs can perceive objects accurately for a non-cluttered image in the third column, with LLaVA-1.5 failing at counting ties. Our VCoder can also accurately perform general question-answering tasks, as shown in the fourth column. We treat OneFormer [25] as the Oracle for object perception. Red text represents counting mistakes; pink text represents hallucination; blue text represents correct object perception." + ], + "image_footnote": [], + "bbox": [ + 83, + 159, + 267, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e239ad6519c95bb5f6688e29f8e8aa194e00a04c870ee0001134d8c7b3383520.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 159, + 454, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd6c711c84170ee68b6f4a19c7e1db48db20e232e1ea5c28cf980aadf48814c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 159, + 640, + 266 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e63190aef579ad86efb7a24543562ac404b65f177bc38939acb5afb9030d0ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 159, + 885, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c614f92a5cf193e12ec5f9dbbb292af9e0a2c0657bc7537ffcbad6306ab6e6de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 258, + 104, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 104, + 261, + 145, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: eleven people, one large potted plant, two ceiling lamps, one exit sign, and various pieces of furniture that appear to be office desks and chairs.", + "bbox": [ + 84, + 277, + 264, + 333 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b65306bffef51c66058ccede5b60e585c2c534a8d4ac2cc63e00ae3ad33b2016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 258, + 307, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "T-4V", + "text_level": 1, + "bbox": [ + 308, + 261, + 333, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: two Corgi dogs, two umbrellas, multiple chairs, two speakers, and one palm tree. The setting is an outdoor area with a beachfront view.", + "bbox": [ + 272, + 279, + 450, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 480, + 272, + 519, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: two people (1 male and 1 female), a table, windows, and buildings seen through the windows.", + "bbox": [ + 460, + 287, + 635, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GPT-4V", + "text_level": 1, + "bbox": [ + 671, + 267, + 712, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume.", + "bbox": [ + 653, + 282, + 877, + 343 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7552bb6475bf959d9117a786e069a22478c1f3ab3f8e31956a9a693e1676967d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 342, + 104, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLaVA-1.5", + "text_level": 1, + "bbox": [ + 104, + 343, + 156, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: 12 people, 1 potted plant, 1 TV, 2 chairs, 1 handbag.", + "bbox": [ + 84, + 359, + 256, + 396 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45973119da94a73ddf1a687610e895582eca4e3f4fce7bb4baaba606c1cdfc17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 349, + 308, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vA-1.5", + "text_level": 1, + "bbox": [ + 308, + 349, + 344, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: 2 dogs, 3 couches, 3 umbrellas, 1 table, and 2 chairs.", + "bbox": [ + 272, + 367, + 450, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LaVA-1.5", + "text_level": 1, + "bbox": [ + 482, + 345, + 529, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: man, woman, ties (2), suit (1), glass wall (1), building (1), desk (1).", + "bbox": [ + 460, + 364, + 638, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLaVA-1.5", + "text_level": 1, + "bbox": [ + 671, + 353, + 725, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the image, a group of people is gathered in a room, sitting around a table. They are dressed in costumes, possibly for a party or a themed event.", + "bbox": [ + 653, + 369, + 879, + 407 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e80386a4a303de7610ddb8bb5e3b51ee4b58f1973e284548425be081b994e518.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 404, + 102, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VCoder", + "text_level": 1, + "bbox": [ + 104, + 405, + 143, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, three chairs.", + "bbox": [ + 86, + 422, + 259, + 470 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e62166f332d6bf694d9ccb74d469d6c899c75c7f02b6fa208f5a556d777aeeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 411, + 307, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "oder", + "text_level": 1, + "bbox": [ + 308, + 414, + 330, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: sky, tree, pillow, mountain, water, grass, bed, two dogs, umbrella, potted plant, dining table, six chairs.", + "bbox": [ + 272, + 429, + 447, + 479 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d2131fbc61eef778ea71423e0776188fcb7f442143baa7784752142b9b6d90bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 409, + 517, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: window, wall, two people, tie.", + "bbox": [ + 460, + 426, + 633, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VCoder", + "text_level": 1, + "bbox": [ + 671, + 414, + 710, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party.", + "bbox": [ + 653, + 430, + 875, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OneFormer (Oracle)", + "text_level": 1, + "bbox": [ + 86, + 479, + 187, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, chair, TV.", + "bbox": [ + 86, + 494, + 259, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OneFormer (Oracle)", + "text_level": 1, + "bbox": [ + 274, + 488, + 375, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects present in the image are: sky, tree, sea, pillow, wall, bed, two dogs, dining table, six chairs, grass, umbrella, potted plant, mountain.", + "bbox": [ + 272, + 505, + 638, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text in red denotes wrong count.", + "bbox": [ + 656, + 503, + 812, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text in pink denotes hallucination.", + "bbox": [ + 656, + 517, + 816, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text in blue denotes correct object counts.", + "bbox": [ + 656, + 529, + 861, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "when the scenes are cluttered with many entities. Consequently, a natural question arises: \"How to develop MLLM systems that respond to perception questions accurately?\"", + "bbox": [ + 75, + 672, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work aims to improve Multimodal LLMs at the simple yet fundamental object-level perception skills, including counting. Our motivation stems from the intuition that one can only describe and reason about a visual scene with the correct understanding of the entities in the image. In our effort to develop an accurate Multimodal LLM perception system, we face three significant challenges: (i) the scarcity of a vision-language dataset focused on the object perception task; (ii) existing open-sourced Multimodal LLMs usually use the ViT from CLIP [52] with an RGB image as input as the visual component that majorly focuses only on salient objects, and (iii) the absence of evaluation metrics to", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quantitatively measure Multimodal LLMs' object perception and in particular, counting skills. We list our efforts to overcome the issues above in the following paragraphs.", + "bbox": [ + 496, + 672, + 890, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contemporary vision-language models [14, 34, 52] owe their success to the availability of large-scale image-text datasets [7, 50, 56]. However, these datasets are more focused on image captioning [33] and VQA [1] tasks, making them unfit for training Multimodal LLMs for basic perception skills like object identification and counting. To overcome the scarcity of fundamental perception-focused image-text data, we leverage images from the COCO [36] dataset and use predictions from off-the-shelf visual perception models [25, 49, 54] to prepare a COCO Segmentation Text (COST) dataset comprising of question-answer pairs about the objects (background and foreground) present in", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Object Perception", + "bbox": [ + 290, + 93, + 433, + 109 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "General Question-Answering", + "bbox": [ + 658, + 93, + 877, + 109 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27993", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "each image. We provide more details in Sec. 3.1.", + "bbox": [ + 76, + 90, + 400, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by diffusion models that add various perception \"control\" or \"context\" images [45, 66, 67, 70] as auxiliary inputs to aid image generation, we propose feeding extra perception modalities as control inputs through additional vision encoders, which we term as our Versatile vision encoders (VCoder). In this work, we focus on the task of object perception and leverage a segmentation map, depth map, or both as the control inputs; however, the same design can be extended to other modalities. Our VCoder projects the control inputs' information into the LLM's space as shown in Fig. 4. We hypothesize that this added control helps the MLLM improve its object perception ability.", + "bbox": [ + 75, + 107, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lastly, owing to the absence of metrics to quantify the counting ability in MLLMs, we propose computing a count score (CS) using one-to-one matching of object words in the ground truth and MLLM's answer. We also compute a hallucination score (HS) based on the extra objects in the MLLM's response that are absent from the ground truth. Similarly, we introduce a depth score (DS) to quantify the object order prediction performance in MLLMs.", + "bbox": [ + 75, + 289, + 467, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among the open-source MLLMs, we choose LLaVA-1.5 [39] as our base MLLM. Our extensive experimental analysis demonstrates the importance of our COST dataset and VCoder LLaVA-1.5's improved perception ability. To summarize, our contributions are as follows:", + "bbox": [ + 76, + 411, + 467, + 484 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose using extra (perception) control inputs and feeding those to a Versatile enCoder (VCoder) for improved object perception performance.", + "- We introduce a COCO Segmentation Text (COST) dataset to train and evaluate Multimodal LLM systems on the fundamental object-level perception tasks of object identification, counting, and order prediction.", + "- Furthermore, to quantify the object perception ability in MLLMs, we propose calculating a count score (CS), a hallucination score (HS) and a depth score (DS). Our experiments show that the VCoder-adapted LLaVA1.5 outperforms the baseline MLLMs on all metrics when validated on the COST dataset." + ], + "bbox": [ + 94, + 487, + 467, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 699, + 217, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Visual Perception", + "text_level": 1, + "bbox": [ + 76, + 724, + 246, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The fundamental nature of visual perception makes it a critical component in MLLM systems. Perception can be divided into sub-tasks, including dense prediction tasks like image segmentation [24, 41, 62] and depth estimation [16, 18, 54], and sparse prediction tasks like object detection [6, 63] and pose estimation [13, 58]. In the DL era, initial methods tackled the perception task using CNN based methods [8, 9, 21, 29, 30, 58] with recent methods shifting to the use of transformer-based architectures [10, 18, 25, 54, 65, 75]. In this work, we tackle the", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task of object-level perception, mainly focusing on predicting names, counts, and order of objects using MLLMs.", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Understanding with LLMs", + "text_level": 1, + "bbox": [ + 498, + 131, + 794, + 147 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using LLMs for vision applications is not a new concept. In a nutshell, developing Multimodal LLMs involves projecting [2, 4, 34, 61] the features from a vision encoder [15, 53] to the embedding space of a language model (LLM) [11, 59, 60], and, visual instruction-tuning.", + "bbox": [ + 496, + 155, + 890, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLaVA [40] proposed a pipeline to convert existing image-text data into dialog format and then finetuned a CLIP [52] and LLaMA [59] model end-to-end on their collected dataset showing one of the earliest evidence of visual-language instruction tuning. Concurrent to LLaVA, MiniGPT-4 [74] used the visual encoder from BLIP2 [34] and used a linear layer for projecting visual features into Vicuna's [11] feature space. InstructBLIP [14] open-sourced a collection of 16 different datasets covering various vision tasks like VQA, reasoning, captioning, classification, etc., and finetuned a BLIP2 model on their dataset. mPLUGOwl [68] proposed using a vision abstractor and finetuning the vision encoder. More recently, LLaVA-1.5 [39] proposed using an MLP as the projector and finetuned on academic instruction datasets to achieve state-of-the-art performance on various benchmarks [12, 17, 23, 35]. Among various open-source MLLMs [5, 22, 31, 32, 69], we chose LLaVA-1.5 as our baseline due to its superior performance.", + "bbox": [ + 496, + 231, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Perception Hallucination in MLLMs", + "text_level": 1, + "bbox": [ + 498, + 513, + 815, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the introduction of LLMs, there has been a comprehensive study about their ability to hallucinate [71] in the NLP community. However, the phenomenon of hallucination in Multimodal LLMs has received comparatively less attention. LRV-Instruction [38] introduced a new instruction-tuning dataset containing 400k visual instructions to prevent hallucination in MLLMs and measured performance treating responses from GPT-4 [48] as ground truths. More recently, HallusionBench [37] quantitatively benchmarked various failure modes in MLLMs that lead to hallucinations based primarily on logical consistency and reasoning. Unlike these works that tried to benchmark MLLMs mainly on VQA-type tasks, this paper focuses on the object-level hallucination in MLLMs.", + "bbox": [ + 496, + 537, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The two closest works to our objective are POPE [35] and CHAIR [55]. On the one hand, POPE [35] tried to measure hallucination in MLLMs using a binary “Yes”-“No” answer policy in response to questions based on the absence or presence of an object in the image. On the other hand, CHAIR [55] focused on measuring hallucination in image captioning based on only words and not counts for the objects. In our work, we consider not only object words but also the corresponding count to compute an object-level count score and hallucination score.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27994", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/96879fa2cf51322e094aa1136a5a31d22cf2f59569536a42fc019c52e93e991e.jpg", + "image_caption": [ + "Semantic Object Identification" + ], + "image_footnote": [ + "USER: What objects can be seen in the image? Perceive as done for semantic segmentation." + ], + "bbox": [ + 81, + 109, + 236, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/066a464a56fe331f09262e6ea63fed3bb02b376a606cff4c5810fc0201e25758.jpg", + "image_caption": [], + "image_footnote": [ + "USER: What objects can be seen in the image? Perceive as done for instance segmentation." + ], + "bbox": [ + 240, + 109, + 395, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a21f3b6a4be5b31abed17f36fa7580bb491c5dfd0575e73ff072c8194d0913f4.jpg", + "image_caption": [ + "Instance Object Identification" + ], + "image_footnote": [], + "bbox": [ + 411, + 109, + 568, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e5db06204c7bfcd8ec256e39f554b30237b895027b5b649bd81ab2606bbb964.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 109, + 727, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from DINov2 DPT", + "bbox": [ + 769, + 104, + 849, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from OneFormer", + "bbox": [ + 777, + 147, + 841, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": " from COCO dataset", + "bbox": [ + 769, + 190, + 849, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "question text from prompting GPT-4", + "bbox": [ + 754, + 232, + 864, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "answer text from OneFormer output", + "bbox": [ + 756, + 273, + 862, + 299 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a6f9272957ecdc8c391e59afc6476a399d10228ec0cf0f137297cc9322b9393.jpg", + "image_caption": [ + "Figure 3. Organization of the COST dataset. We incorporate the images from COCO [36], the questions from GPT-4 [48], and the segmentation outputs from OneFormer [25] in a question-answer format for training and evaluating MLLMs on the object identification task. We also extend COST to the object order perception task by incorporating depth map outputs from DINOv2 [49] DPT [54]. COST can be extended to more object-level tasks by similarly incorporating other modalities (for example, keypoint maps)." + ], + "image_footnote": [ + "ASSISTANT: The objects present in the image are: grass, person, tree, house, car, frisbee, fence, window" + ], + "bbox": [ + 81, + 313, + 236, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1d2bfb79a938f31da3ce43d62fc342a520b25002afa321309f128449e25a256.jpg", + "image_caption": [], + "image_footnote": [ + "USER: What objects can be seen in the image? Perceive as done for panoptic segmentation." + ], + "bbox": [ + 240, + 313, + 397, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dafc120373e7522f7fda861a262d307ffbe5e49e08a2f6645ef0921ebdb298e8.jpg", + "image_caption": [], + "image_footnote": [ + "ASSISTANT: The objects present in the image are: two people, car, frisbee.", + "ASSISTANT: The depth order for objects present in the image is: frisbee, person, person-2, car, tree, house, grass." + ], + "bbox": [ + 411, + 313, + 568, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/700d946e4e38e0865fdac7be190cf9e63df81028dbf61a2ee66c8a71f705564c.jpg", + "image_caption": [], + "image_footnote": [ + "ASSISTANT: The objects present in the image are: grass, tree, house, two people, frisbee, car." + ], + "bbox": [ + 571, + 313, + 727, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39a475e679b05d7d4bfeefdc95cd9e528e683c640ca2cf64889ab66cc65005db.jpg", + "image_caption": [], + "image_footnote": [ + "USER: Could you list the objects in the image in order of their perceived distance from the foreground to the background?" + ], + "bbox": [ + 730, + 313, + 887, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Panoptic Object Identification", + "text_level": 1, + "bbox": [ + 140, + 511, + 333, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Order Perception", + "text_level": 1, + "bbox": [ + 566, + 511, + 725, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Object Identification with MLLMs", + "text_level": 1, + "bbox": [ + 75, + 599, + 392, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose you are invited to a Halloween party and want to bring candies for every person at that party. You ask your friend to send you a picture (Fig. 1) of the party room so that you can estimate the number of people and the number of candies you need to buy. In a hurry, you ask GPT-4V [48]: \"Can you count the number of people in the image?\", and it responds: \"Yes, there are ten people visible in the image.\" Excited, you arrive at the party with ten candies but wait, you see fourteen people! Confused, you look at the image your friend sent you, and you can count fourteen people in that image, realizing that GPT-4V fails at the simple task of counting the people in the picture. At the same time, it can accurately describe the happening of a Halloween party in the image (Fig. 1). We refer to the phenomenon of Multimodal LLMs failing at simple visual perception tasks while succeeding at complex visual reasoning tasks as Moravec's Paradox [44] in perception.", + "bbox": [ + 73, + 626, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hypothesize that one of the main reasons for the", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "above phenomenon is the absence of conversations covering object identification for the salient objects and the objects in the background from the instruction-tuning data for MLLMs. To overcome this issue, we prepare the COCO Segmentation Text (COST) dataset with COCO [36] images and create sentences using the output from an image segmentation model [25] to obtain an image-text dataset to train and evaluate MLLMs for object perception MLLMs. Moreover, we also introduce a segmentation map as a control image input to the MLLM for better performance and quantify object perception performance with a count score (CS) and a hallucination score (HS).", + "bbox": [ + 496, + 599, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. COST to Identify Objects with MLLMs", + "text_level": 1, + "bbox": [ + 498, + 797, + 841, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that image segmentation methods [10, 25] can accurately identify salient (foreground objects like people, cars, etc.) and background objects (like sky, wall, etc.) in a given scene. Guided by this finding, we use images from the COCO [36] dataset and obtain the correspond", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27995", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19db383ea4f7748686b1f99a722a44cbac71d98a8070d6b4663bc02331396284.jpg", + "image_caption": [ + "Figure 4. Adapting Multimodal LLMs for accurate object perception with VCoder. (a) We add our VCoder as an adapter to the LLaVA-1.5 [39] and feed perception modalities as extra control inputs for improved object perception performance. During training, we freeze the components from LLaVA-1.5 (ImCoder, MLP, and LLM) to retain the original reasoning performance. (b) Using depth map and segmentation map as the control inputs to VCoder for the object order perception task." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 651, + 398 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3de7d34acd71cfecd0a37e856f4715f32d83ead567f5eb97af23dfbd466b284.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 893, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing segmentation outputs from OneFormer [25], a state-of-the-art image segmentation model. Next, we extract the object (class) names and counts from the segmentation outputs and convert them into a sentence form for the ground-truth answer: \"The objects present in the image are: \\([CNT_1][OBJ_1],[CNT_2][OBJ_2],\\ldots,[CNT_N][OBJ_N]\". with [OBJ_i] representing the object name and \\([CNT_i]\\) representing the count (if greater than one) for the \\(i^{th}\\) object in the image. We prompt GPT-4 [48] to collect a bucket of questions for three different object identification tasks: semantic, instance, and panoptic, corresponding to the three different image segmentation tasks. Finally, as shown in Fig. 3, we organize the images from COCO, segmentation maps from OneFormer, questions from GPT-4, and sentences containing object information into a question-answer format to construct our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object identification task.", + "bbox": [ + 75, + 472, + 472, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistically, we prompt GPT-4 [48] to return 20 questions for each question bucket (panoptic, semantic, and instance). In total, we used 280k images from the train2017, test2017, and unlabeled2017 splits of the COCO [36] dataset and corresponding segmentation outputs from OneFormer [25] to form the visual component of the COST training dataset. Similarly, we prepare a COST validation split using the 5k images from the val2017 split of the COCO dataset.", + "bbox": [ + 75, + 747, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that a similar approach can extend the COST", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset to other perception modalities. In this work, we incorporate the depth map modality into our COST dataset for the object order perception task. Particularly, we leverage the publicly available DINOv2 [49] DPT [54] model to obtain depth maps for COCO images and use the panoptic mask (from OneFormer [25]) to estimate the depth order of objects in an image. We format the obtained ordering of objects into the text with the template: \"The depth order for objects present in the image is: $[OBJ_1], [OBJ_2], \\ldots, [OBJ_J]$ ,\" with $[OBJ_j]$ representing the $j^{th}$ object name. To maintain relative ordering among objects belonging to the same class, we append a count number to the second and later objects, as shown in the bottom right of Fig. 3 for person and person-2. Similar to the previous setting, we prompt GPT-4 [48] to return 20 questions for the object order perception task. We provide a detailed flow of obtaining ground-truth object orders in the appendix.", + "bbox": [ + 496, + 472, + 893, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. VCoder for Multimodal LLMs", + "text_level": 1, + "bbox": [ + 500, + 739, + 772, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We notice that existing open-source Multimodal LLMs generally use the ViT [15] from CLIP [53] as the image encoder (ImCoder) during instruction tuning. We reason that the ViT focuses mainly on salient objects because it is trained against captions, which leave out information about background regions. We argue that identifying objects in the background is critical for a Multimodal LLM to become skilled at perception. To overcome this limitation, we introduce a segmentation map as a control input [45, 70] into", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27996", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our Multimodal LLM. Specifically, we use the segmentation map from OneFormer [25] and project it to the LLM's embedding space using a pretrained ViT [15] (from CLIP [53]) as a SegCoder and a two-layer MLP [39] which we collectively refer to as our Versatile enCoder (VCoder). This extra control from the segmentation map results in considerable performance gains on the object identification task.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4a, our VCoder adapted MLLM takes three sets of inputs: perception modalities as control inputs fed into the VCoder, an RGB image fed into an Image enCoder (and MLP), and the text question. The RGB image and text are tokenized to the and tokens, respectively. VCoder is flexible at handling various perception modalities with a unique token for each modality. For example, the segmentation map and depth map inputs are tokenized to and tokens, respectively. Similarly, one can incorporate more modalities with modality-specific tokens. Finally, all tokenized embeddings are concatenated and fed into the LLM.", + "bbox": [ + 75, + 199, + 470, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We treat our VCoder as an adapter, added to our base MLLM, LLaVA-1.5 [39] to obtain the final MLLM framework for experiments. Note that we only train the MLP components in the VCoder on the COST dataset. We decided to keep all other parameters fixed during training to keep the reasoning ability unaffected while achieving improved object perception performance.", + "bbox": [ + 75, + 382, + 468, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Evaluating MLLMs for Object Identification", + "text_level": 1, + "bbox": [ + 76, + 503, + 455, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the availability of various metrics [35, 42, 55] to measure object hallucination in vision-language models, no existing metric considers the explicit object counts while calculating their hallucination scores. We argue that object counts returned by an MLLM are a critical component that should not be overlooked while evaluating object identification performance. Therefore, we propose evaluating object identification performance in MLLMs using two metrics: count-score (CS) and hallucination-score (HS).", + "bbox": [ + 75, + 529, + 468, + 666 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {G}: \\mathrm {C N T} _ {1} ^ {G}; \\dots ; \\mathrm {O B J} _ {N} ^ {G}: \\mathrm {C N T} _ {N} ^ {G} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 676, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {P}: \\mathrm {C N T} _ {1} ^ {P}; \\dots ; \\mathrm {O B J} _ {M} ^ {P}: \\mathrm {C N T} _ {M} ^ {P} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 699, + 429, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 5, given a ground-truth sentence $(G)$ and an MLLM predicted response $(P)$ , we first extract the object words (nouns) and their corresponding count from both text samples and represent them in a dictionary form with keys as the object noun and the value as the corresponding object's count as shown in Eq. (1) with $N$ and $M$ representing the number of different object nouns in the $G$ and $P$ respectively. Next, we perform one-to-one matching between the counts for keys with $G_{\\mathrm{dict}}$ and $P_{\\mathrm{dict}}$ as the reference for Count Score (CS) and Hallucination Score (HS), respectively, as shown in Eq. (2).", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa2c0010a20b8a024492580b0046c5ff61415493ffd30cc162a25359859df1a0.jpg", + "image_caption": [ + "Figure 5. Evaluation Metrics for Object Identification. We compare the object counts in the ground truth and prediction to calculate a count score (CS) and a hallucination score (HS)." + ], + "image_footnote": [], + "bbox": [ + 522, + 87, + 870, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C S} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\left\\{ \\begin{array}{l l} \\frac {\\min \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)}{\\max \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)} & \\text {i f I (O B J} _ {i} ^ {G}, P _ {\\text {d i c t}}) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 353, + 854, + 396 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H S} = \\frac {1 0 0}{M} \\sum_ {j = 1} ^ {M} \\left\\{ \\begin{array}{l l} 1 - \\frac {\\operatorname* {m i n} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)}{\\operatorname* {m a x} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)} & \\text {i f} I (\\mathrm {O B J} _ {j} ^ {P}, G _ {\\text {d i c t}}) \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 400, + 887, + 448 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathrm {O B J}, D) = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f O B J i s i n k e y s} (D) \\\\ \\text {F a l s e} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 450, + 890, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Count Score (CS). It represents the percentage of correct object counts predicted by the MLLM with respect to the ground-truth sentence. The higher the CS, the better.", + "bbox": [ + 498, + 515, + 890, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hallucination Score (HS). It represents the percentage of extra object counts predicted by the MLLM that do not exist in the ground-truth sentence. The lower the HS, the better.", + "bbox": [ + 498, + 560, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that due to the one-to-one word-matching nature of our evaluation, we manually define a mapping between the categories in COCO [36] and their synonyms [43, 55]. For example, we replace words like man, woman, child, boy, girl, etc. with person in the prediction before evaluation.", + "bbox": [ + 498, + 606, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 694, + 632, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use LLaVA-1.5 [39] as our base MLLM. LLaVA-1.5 uses CLIP-ViT-L-336px [53] as the image encoder (ImCoder) with a two-layer MLP as projection and Vicuna1.5 [73] as the LLM. Inside our VCoder, we also use a CLIP-ViT-L-336px to encode the control inputs and project the features into the LLM embedding space using modality-specific two-layer MLPs. We resize the visual inputs to $336 \\times 336$ resolution (corresponds to 576 tokens) for our MLLM. During training, we load the instruction-tuned weights from LLaVA-1.5 and keep those frozen while only tuning the MLP component of our VCoder. We use the publicly available OneFormer [25] model trained on", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27997", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/51490ee5a2756050485e12e3b8680ca6dd3d16eeeab8bd5623b1577aa0f03950.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInput TokensSemanticInstancePanoptic
CS (↑)HS (↓)CS (↑)HS (↓)CS (↑)HS (↓)
Closed Model, Open API
GPT-4V [48]⟨img⟩ + ⟨query⟩————38.483.0
Existing Open-Source Multimodal LLMs
MiniGPT-4 LLaMA-2-7b [74]⟨img⟩ + ⟨query⟩6.292.25.697.76.294.9
InstructBLIP Vicuna-7b [14]⟨img⟩ + ⟨query⟩14.285.825.391.917.591.2
LLaVA-1.5-7b [39]⟨img⟩ + ⟨query⟩30.660.150.375.938.767.3
LLaVA-1.5-13b [39]⟨img⟩ + ⟨query⟩25.069.349.975.035.868.6
CogVLM-17b [64]⟨img⟩ + ⟨query⟩33.467.543.586.240.675.9
Baselines trained on the COST dataset
COST IT LLaVA-1.5-7b⟨img⟩ + ⟨query⟩78.722.167.530.371.928.2
Soft-Prompted LLaVA-1.5-7b⟨prompt⟩ + ⟨img⟩ + ⟨query⟩36.256.718.472.226.863.0
ImCoder LLaVA-1.5-7b⟨img⟩ + ⟨img⟩ + ⟨query⟩78.922.764.029.470.827.9
VCoder augmented LLaVA-1.5
VCoder LLaVA-1.5-7b⟨seg⟩ + ⟨img⟩ + ⟨query⟩88.610.471.126.986.012.8
VCoder LLaVA-1.5-13b⟨seg⟩ + ⟨img⟩ + ⟨query⟩89.010.073.325.087.211.6
", + "bbox": [ + 81, + 85, + 890, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison to baseline Multimodal LLMs on the COST validation dataset for Object Identification. We compare our VCoder to existing off-the-shelf baseline MLLMs: MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64]. We also train three different variants of LLaVA-1.5 on the COST dataset: COST IT mixes the COST training data with the instruction tuning data; Soft-Prompted uses a set of learnable tokens, and ImCoder uses an RGB image as the control input. Our VCoder adapted LLaVA-1.5 performs the best on all three object perception tasks. Note: $\\langle \\cdot \\rangle$ denotes input tokens to LLM with seg representing segmentation map, img representing RGB image, prompt representing learnable prompt, and query representing the user question. We also evaluate the performance of GPT-4V [48] on the COST dataset using the publicly accessible paid API released by OpenAI. Our VCoder-adapted LLaVA-1.5 shows the best performance on object identification among all MLLMs.", + "bbox": [ + 75, + 406, + 893, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COCO [36] with DiNAT-L [19, 20] backbone to obtain the segmentation map. For getting depth maps, we use the publicly available ViT-L/14 distilled variant of DINOv2 [49] DPT [54] trained on the NYUd [46] dataset. In this section, we discuss our results on the object identification task.", + "bbox": [ + 75, + 536, + 470, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 625, + 294, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. We train our VCoder-adapted LLaVA-1.5 framework for two epochs on the COST training dataset with a batch size 256 and a learning rate of $1e^{-3}$ . For other training hyperparameters, we follow the settings used during the instruction-tuning stage in LLaVA-1.5 [39]. Following [25], we uniformly sample each object identification task (semantic, instance, and panoptic) during training. We also use the corresponding segmentation map from OneFormer [25] as input to the VCoder during training and inference. On 8 A100 GPUs, it takes 8 and 14 hours to train our VCoder with the 7b and 13b variants of LLaVA-1.5 as the base MLLM, respectively.", + "bbox": [ + 75, + 652, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Details. We evaluate all MLLMs on the COST validation set. We separately evaluate semantic, instance, and panoptic object identification tasks while randomly sampling questions from the corresponding task's question", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bucket. Note that for evaluating all off-the-shelf MLLMs, we experiment with various prompts and finally use the prompt: \"[QUESTION]. Return the answer in the paragraph format: 'The objects present in the image are: ...' and then list the objects with their count in word format (if greater than 1) in front of them, like 'two people'\"; where [QUESTION] is the randomly sampled question from the object identification task bucket.", + "bbox": [ + 496, + 536, + 893, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 667, + 638, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare the performance of VCoder to open-source Multimodal LLMs, namely, MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64] on the COST validation set in Tab. 1. Furthermore, we provide three additional baselines, trained for two epochs:", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COST IT LLaVA-1.5: We mix the COST training data with the instruction tuning data used in LLaVA-1.5 [39] and fine-tune a LLaVA-1.5 [39] model from scratch.", + "bbox": [ + 496, + 763, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft-Prompted LLaVA-1.5: We pretend 576 learnable tokens $\\langle \\langle prompt\\rangle \\rangle$ to the LLM input and tune only the $\\langle prompt\\rangle$ parameters on the COST training dataset.", + "bbox": [ + 496, + 809, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImCoder LLaVA-1.5: We use an RGB image as the control input and train on the COST training dataset.", + "bbox": [ + 496, + 854, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 1, we notice that all existing MLLMs", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27998", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/067ae9138f730bd434f24050a087b69e70726aade0eef943f45eb09f13cfc147.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDepth Score (↓)
LLaVA-1.5-7b [39]166.1
LLaVA-1.5-13b [39]227.2
VCoder-DS LLaVA-1.5-7b65.9
VCoder-DS LLaVA-1.5-13b63.3
", + "bbox": [ + 135, + 88, + 405, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Performance on Object Order Perception. Our VCoder LLaVA-1.5 considerably outperforms LLaVA-1.5 [39], owing to the usage of control inputs and training on the COST dataset.", + "bbox": [ + 75, + 167, + 468, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perform poorly on our COST validation set, demonstrating their inability to count and identify objects accurately. Note that existing MLLMs perform relatively better on instance object identification, reaffirming our claim that MLLMs are better at detecting salient objects than background objects. Although the baselines trained on the COST dataset perform relatively better, they still lag in performance compared to the VCoder. Notably, a segmentation map performs considerably better than using an RGB image as the control input, proving the segmentation map's vitality.", + "bbox": [ + 75, + 218, + 468, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to GPT-4V [48]. We utilize OpenAI's newly released gpt-4-vision-preview $^1$ API to obtain responses from GPT-4V. Our experiments show that GPT-4V's responses are consistent across all object identification tasks, closely aligning with the panoptic identification task. Therefore, we compare our VCoder to GPT-4V only on the panoptic object identification to reduce API requests due to a daily limit of 500 API requests during this project. As shown in Tab. 1, GPT-4V [48] lags behind our VCoder by a considerable margin, reaffirming our claim that existing MLLMs cannot perform accurate object-level perception.", + "bbox": [ + 75, + 369, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Object Order Perception with MLLMs", + "text_level": 1, + "bbox": [ + 76, + 549, + 426, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 4, multiple perception modalities can be leveraged to improve object perception in MLLMs with our VCoder. This section presents our experiments with our VCoder using the segmentation and depth maps as the control inputs. We term the resulting MLLM as VCoder-DS LLaVA-1.5. Intuitively, predicting the object order implicitly means identifying the objects in an image. Therefore, for the object order perception task (Fig. 4b), we use both and inputs, while only the input as the additional control for object identification.", + "bbox": [ + 75, + 574, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, we use a mixture of datasets, including the object identification and object order perception components from the COST dataset. We also use about 200k image-conversation (along with the corresponding segmentation map obtained using OneFormer [25]) pairs randomly sampled from the instruction tuning data used in LLaVA1.5 [39]. We train our VCoder for one epoch following the same hyperparameter settings mentioned in Sec. 4.", + "bbox": [ + 75, + 726, + 468, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 2, our VCoder-DS LLaVA-1.5 significantly outperforms the base MLLM, LLaVA-1.5 [39] on", + "bbox": [ + 75, + 847, + 468, + 877 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the COST validation set. For quantitative evaluation of MLLMs on the object order perception task, we calculate a depth score (DS) using the absolute difference between the position of objects in the ground truth and prediction.", + "bbox": [ + 500, + 90, + 890, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 164, + 620, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the improved object perception performance after training our VCoder on the COST dataset, certain limitations remain to be addressed for future work. Firstly, we build our COST dataset using OneFormer [25], which can only perceive objects belonging to a limited number of categories due to being trained on a closed-set vocabulary dataset [36]. For real-world applications, it is imperative to develop an object perception benchmark for MLLMs covering many more classes with varying granularity. Secondly, the count, hallucination, and depth scores use one-to-one word matching, which requires defining a mapping between synonymous words. It will be promising to explore ways to overcome using synonym mappings.", + "bbox": [ + 496, + 190, + 890, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 398, + 617, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work analyzes the object-level perception skills of Multimodal Large Language Models (VLMMs). Although MLLMs are good visual reasoners, they need to improve at the simple yet fundamental task of object perception. To improve object perception ability in MLLMs, we propose the COST dataset for training and evaluating MLLMs at the object perception task. We benchmark different off-the-shelf MLLMs and GPT-4V on our COST dataset and observe their lousy performance. Consequently, we propose using perception modalities as control inputs and a Versatile vision enCoders (VCoder) as an adapter for projecting the control inputs to the LLM embedding space. Our VCoder can easily be extended to leverage various modalities as the control inputs depending on the task. To quantify the object-level perception ability in MLLMs, we introduce a Count-Score (CS), a Hallucination-Score (HS), and a Depth-Score (DS). We adapted LLaVA-1.5 with VCoder, only trained the VCoder on our COST dataset, and demonstrated its improved performance at the object perception task while retaining the reasoning performance. We hope our work can inspire the research community to focus on developing object perception datasets for MLLMs and develop vision systems that are equally good at perception and reasoning in the future.", + "bbox": [ + 496, + 424, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was in part supported by IARPA under Contract No. 2022-21102100004, NSF CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. We also thank the ML Center at Georgia Tech for generously supporting this work.", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1https://platform.openai.com/docs/guides/vision", + "bbox": [ + 94, + 886, + 465, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27999", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2015. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 3", + "[3] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. Falcon-40B: an open large language model with state-of-the-art performance. arXiv, 2023. 1", + "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv, 2023. 1, 3", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv, 2023. 3", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2", + "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 3", + "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 3", + "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformer for universal image segmentation. In CVPR, 2022.3,4", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 1, 3", + "[12] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.3", + "[13] Timothy F Cootes, Gareth J Edwards, and Christopher J Taylor. Active appearance models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681-685, 2001. 3", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 3, 7", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 5, 6", + "[16] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in neural information processing systems, pages 2366-2374, 2014. 3", + "[17] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv, 2023. 3", + "[18] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 3", + "[19] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 7", + "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 7", + "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 3", + "[22] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. ArXiv, abs/2302.14045, 2023. 1, 3", + "[23] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. CVPR, 2019. 3", + "[24] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Ji-achen Li, Steven Walton, and Humphrey Shi. Semask: Se" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28000", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mentally masking transformer backbones for effective semantic segmentation. arXiv, 2021. 3", + "[25] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8", + "[26] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Yadong Mu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arXiv, 2023. 1", + "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2023. 1", + "[28] Deanna Kuhn. The Skills of Argument. Cambridge University Press, 1991. 1", + "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth International Conference on 3D Vision (3DV), pages 239-248. IEEE, 2016. 3", + "[30] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE. 3", + "[31] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. 2023. 3", + "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 3", + "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2", + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3", + "[35] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 3, 6", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5, 6, 7, 8", + "[37] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3", + "[38] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 3, 5, 6, 7, 8", + "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3", + "[41] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3", + "[42] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv, 2023. 6", + "[43] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In CVPR, 2018. 6", + "[44] H. Moravec. Mind children: The future of robot and human intelligence. Harvard University Press, 1988. 4", + "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv, 2023. 3, 5", + "[46] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7", + "[47] OpenAI. Chatgpt. https://chat.openai.com/, 2022.1", + "[48] OpenAI. Gpt-4 technical report, 2023. 1, 2, 3, 4, 5, 7, 8", + "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4, 5, 7", + "[50] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011. 2", + "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. ArXiv, abs/2306, 2023. 1", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. arXiv, 2021. 3, 5, 6", + "[54] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 3, 4, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28001", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In EMNLP, 2018. 3, 6", + "[56] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops 2021, 2021. 2", + "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv, 2023. 1", + "[58] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 3", + "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv, 2023. 3", + "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv, 2023. 1, 3", + "[61] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 3", + "[62] Z. Tu, Xiangrong Chen, Alan Yuille, and Song Zhu. Image parsing: Unifying segmentation, detection, and recognition. In IJCV, 2005. 3", + "[63] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In CVPR, 2001. 3", + "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv, 2023. 7", + "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and ef" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3", + "[66] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 3", + "[67] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 3", + "[68] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 3", + "[69] Yan Zeng, Hanbo Zhang, Jiani Zheng, Jiangnan Xia, Guoqiang Wei, Yang Wei, Yuchen Zhang, and Tao Kong. What matters in training a gpt4-style language model with multimodal inputs? arXiv preprint arXiv:2307.02469, 2023. 3", + "[70] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5", + "[71] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv, 2023. 3", + "[72] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv, 2023. 1", + "[73] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6", + "[74] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv, 2023. 1, 3, 7", + "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv, 2020. 3" + ], + "bbox": [ + 501, + 92, + 893, + 685 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28002", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_model.json b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aac2f8322c395fc928fc8b3291f037232006c69e --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_model.json @@ -0,0 +1,3269 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "VCoder: Versatile Vision Encoders for Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.18, + 0.677, + 0.2 + ], + "angle": 0, + "content": "Jitesh Jain\\(^{1}\\) Jianwei Yang\\(^{2}\\) Humphrey Shi\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.201, + 0.766, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)SHI Labs @ Georgia Tech \\(^{2}\\)Microsoft Research, Redmond \\(^{3}\\)Picsart AI Research (PAIR)" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.22, + 0.633, + 0.233 + ], + "angle": 0, + "content": "https://github.com/SHI-Labs/VCoder" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Humans possess the remarkable skill of Visual Perception, the ability to see and understand the seen, helping them make sense of the visual world and, in turn, reason. Multimodal Large Language Models (MLLM) have recently achieved impressive performance on vision-language tasks ranging from visual question-answering and image captioning to visual reasoning and image generation. However, when prompted to identify or count (perceive) the entities in a given image, existing MLLM systems fail. Working towards developing an accurate MLLM system for perception and reasoning, we propose using Versatile vision en-Coders (VCoder) as perception eyes for Multimodal LLMs. We feed the VCoder with perception modalities such as segmentation or depth maps, improving the MLLM's perception abilities. Secondly, we leverage the images from COCO and outputs from off-the-shelf vision perception models to create our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object perception task. Thirdly, we introduce metrics to assess the object perception abilities in MLLMs on our COST dataset. Lastly, we provide extensive experimental evidence proving the VCoder's improved object-level perception skills over existing Multimodal LLMs, including GPT-4V. We open-source our dataset, code, and models to promote research." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.681, + 0.209, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.708, + 0.45, + 0.737 + ], + "angle": 0, + "content": "'Perception is the soil; reasoning, the seed. Without fertile ground, the seed cannot flourish.'" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.738, + 0.449, + 0.753 + ], + "angle": 0, + "content": "(GPT-4 [48], 2023)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The ability to think and reason is one of the most remarkable traits that help humans function daily. Generally, understanding the environment precedes the act of thinking and reasoning [28]. Following the success of ChatGPT-like instruction following AI agents [3, 11, 47, 48, 60] at language understanding and reasoning, researchers have leveraged LLMs to develop instruct frameworks [14, 40, 68, 74] that can understand vision and language inputs in an effort to imitate human perception and reasoning ability. We" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.268, + 0.885, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.436, + 0.55, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.453, + 0.664, + 0.468 + ], + "angle": 0, + "content": "at is happening in the image?" + }, + { + "type": "text", + "bbox": [ + 0.742, + 0.442, + 0.819, + 0.454 + ], + "angle": 0, + "content": "Difficulty Level" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.472, + 0.534, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.536, + 0.477, + 0.58, + 0.489 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.495, + 0.845, + 0.531 + ], + "angle": 0, + "content": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.539, + 0.534, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.536, + 0.543, + 0.579, + 0.554 + ], + "angle": 0, + "content": "VCoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.872, + 0.596 + ], + "angle": 0, + "content": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.606, + 0.55, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.624, + 0.625, + 0.645 + ], + "angle": 0, + "content": "Can you count the number of people in the image?" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.651, + 0.614, + 0.662 + ], + "angle": 0, + "content": "Difficulty Level" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.608, + 0.666, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.667, + 0.611, + 0.712, + 0.623 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.644, + 0.631, + 0.737, + 0.666 + ], + "angle": 0, + "content": "Yes, there are ten people visible in the image." + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.609, + 0.774, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.775, + 0.611, + 0.819, + 0.623 + ], + "angle": 0, + "content": "VCoder" + }, + { + "type": "text", + "bbox": [ + 0.756, + 0.633, + 0.867, + 0.655 + ], + "angle": 0, + "content": "Yes, there are a total of 14 people in the image." + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Figure 1. GPT-4V [48] (access date: Dec 16, 2023) returns impressive responses when prompted to describe complex visual scenes. However, it fails at the simple task of counting in the same scene. Our VCoder returns the correct count of people." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.901 + ], + "angle": 0, + "content": "refer to such systems as Multimodal LLMs (MLLM). Although MLLMs exhibit the ability to perform complex vision-language tasks like visual captioning [2, 4, 39], image generation [26, 27, 57], visual reasoning and grounding [22, 51, 72], they often display sub-par performance at simple tasks like counting objects (Fig. 1). As shown in Fig. 2, MLLMs output incorrect object counts (people, chairs) and hallucinate about the presence (handbag, couch) of certain objects when prompted to identify entities in a visual input. The perception performance is much worse" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27992" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.292, + 0.094, + 0.434, + 0.11 + ], + "angle": 0, + "content": "Object Perception" + }, + { + "type": "header", + "bbox": [ + 0.66, + 0.094, + 0.879, + 0.11 + ], + "angle": 0, + "content": "General Question-Answering" + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.127, + 0.121, + 0.14 + ], + "angle": 0, + "content": "USER" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.144, + 0.263, + 0.156 + ], + "angle": 0, + "content": "What objects can be seen in the image?" + }, + { + "type": "title", + "bbox": [ + 0.274, + 0.127, + 0.308, + 0.139 + ], + "angle": 0, + "content": "USER" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.145, + 0.451, + 0.155 + ], + "angle": 0, + "content": "What objects can be seen in the image?" + }, + { + "type": "title", + "bbox": [ + 0.461, + 0.127, + 0.495, + 0.139 + ], + "angle": 0, + "content": "USER" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.145, + 0.638, + 0.155 + ], + "angle": 0, + "content": "What objects can be seen in the image?" + }, + { + "type": "title", + "bbox": [ + 0.654, + 0.127, + 0.689, + 0.139 + ], + "angle": 0, + "content": "USER" + }, + { + "type": "text", + "bbox": [ + 0.658, + 0.145, + 0.799, + 0.155 + ], + "angle": 0, + "content": "What is happening in the image?" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.16, + 0.268, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.16, + 0.455, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.459, + 0.16, + 0.642, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.16, + 0.887, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.259, + 0.105, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.262, + 0.147, + 0.272 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.278, + 0.265, + 0.334 + ], + "angle": 0, + "content": "The objects present in the image are: eleven people, one large potted plant, two ceiling lamps, one exit sign, and various pieces of furniture that appear to be office desks and chairs." + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.26, + 0.308, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.309, + 0.262, + 0.334, + 0.272 + ], + "angle": 0, + "content": "T-4V" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.28, + 0.451, + 0.339 + ], + "angle": 0, + "content": "The objects present in the image are: two Corgi dogs, two umbrellas, multiple chairs, two speakers, and one palm tree. The setting is an outdoor area with a beachfront view." + }, + { + "type": "title", + "bbox": [ + 0.481, + 0.273, + 0.521, + 0.284 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.289, + 0.636, + 0.339 + ], + "angle": 0, + "content": "The objects present in the image are: two people (1 male and 1 female), a table, windows, and buildings seen through the windows." + }, + { + "type": "title", + "bbox": [ + 0.673, + 0.268, + 0.713, + 0.279 + ], + "angle": 0, + "content": "GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.283, + 0.879, + 0.344 + ], + "angle": 0, + "content": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.343, + 0.105, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.344, + 0.158, + 0.354 + ], + "angle": 0, + "content": "LLaVA-1.5" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.36, + 0.258, + 0.397 + ], + "angle": 0, + "content": "The objects present in the image are: 12 people, 1 potted plant, 1 TV, 2 chairs, 1 handbag." + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.35, + 0.309, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.31, + 0.351, + 0.346, + 0.361 + ], + "angle": 0, + "content": "vA-1.5" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.368, + 0.452, + 0.404 + ], + "angle": 0, + "content": "The objects present in the image are: 2 dogs, 3 couches, 3 umbrellas, 1 table, and 2 chairs." + }, + { + "type": "title", + "bbox": [ + 0.483, + 0.347, + 0.531, + 0.359 + ], + "angle": 0, + "content": "LaVA-1.5" + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.365, + 0.64, + 0.402 + ], + "angle": 0, + "content": "The objects present in the image are: man, woman, ties (2), suit (1), glass wall (1), building (1), desk (1)." + }, + { + "type": "title", + "bbox": [ + 0.673, + 0.354, + 0.726, + 0.364 + ], + "angle": 0, + "content": "LLaVA-1.5" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.371, + 0.88, + 0.408 + ], + "angle": 0, + "content": "In the image, a group of people is gathered in a room, sitting around a table. They are dressed in costumes, possibly for a party or a themed event." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.405, + 0.104, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.406, + 0.144, + 0.417 + ], + "angle": 0, + "content": "VCoder" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.423, + 0.261, + 0.471 + ], + "angle": 0, + "content": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, three chairs." + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.412, + 0.308, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.309, + 0.415, + 0.331, + 0.425 + ], + "angle": 0, + "content": "oder" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.43, + 0.449, + 0.48 + ], + "angle": 0, + "content": "The objects present in the image are: sky, tree, pillow, mountain, water, grass, bed, two dogs, umbrella, potted plant, dining table, six chairs." + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.41, + 0.518, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.428, + 0.635, + 0.452 + ], + "angle": 0, + "content": "The objects present in the image are: window, wall, two people, tie." + }, + { + "type": "title", + "bbox": [ + 0.673, + 0.415, + 0.712, + 0.427 + ], + "angle": 0, + "content": "VCoder" + }, + { + "type": "text", + "bbox": [ + 0.654, + 0.431, + 0.877, + 0.495 + ], + "angle": 0, + "content": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.481, + 0.189, + 0.492 + ], + "angle": 0, + "content": "OneFormer (Oracle)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.496, + 0.261, + 0.544 + ], + "angle": 0, + "content": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, chair, TV." + }, + { + "type": "title", + "bbox": [ + 0.275, + 0.489, + 0.376, + 0.499 + ], + "angle": 0, + "content": "OneFormer (Oracle)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.506, + 0.64, + 0.542 + ], + "angle": 0, + "content": "The objects present in the image are: sky, tree, sea, pillow, wall, bed, two dogs, dining table, six chairs, grass, umbrella, potted plant, mountain." + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.505, + 0.813, + 0.516 + ], + "angle": 0, + "content": "Text in red denotes wrong count." + }, + { + "type": "text", + "bbox": [ + 0.658, + 0.518, + 0.818, + 0.528 + ], + "angle": 0, + "content": "Text in pink denotes hallucination." + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.53, + 0.862, + 0.54 + ], + "angle": 0, + "content": "Text in blue denotes correct object counts." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.556, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Figure 2. MLLMs counting and identifying objects. As shown in the first column, GPT-4V [48] (access date: Dec 16, 2023) and LLaVA-1.5 [39] both fail at counting people. Moreover, LLaVA-1.5 [39] misses background entities like window, wall, etc. and hallucinates about the presence of a handbag. VCoder can predict the people counts and other background entities accurately except chairs. Similarly, in the second column, GPT-4V and LLaVA-1.5 fail at counting chairs while the VCoder matches the Oracle's performance. Notably, all MLLMs can perceive objects accurately for a non-cluttered image in the third column, with LLaVA-1.5 failing at counting ties. Our VCoder can also accurately perform general question-answering tasks, as shown in the fourth column. We treat OneFormer [25] as the Oracle for object perception. Red text represents counting mistakes; pink text represents hallucination; blue text represents correct object perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.469, + 0.718 + ], + "angle": 0, + "content": "when the scenes are cluttered with many entities. Consequently, a natural question arises: \"How to develop MLLM systems that respond to perception questions accurately?\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This work aims to improve Multimodal LLMs at the simple yet fundamental object-level perception skills, including counting. Our motivation stems from the intuition that one can only describe and reason about a visual scene with the correct understanding of the entities in the image. In our effort to develop an accurate Multimodal LLM perception system, we face three significant challenges: (i) the scarcity of a vision-language dataset focused on the object perception task; (ii) existing open-sourced Multimodal LLMs usually use the ViT from CLIP [52] with an RGB image as input as the visual component that majorly focuses only on salient objects, and (iii) the absence of evaluation metrics to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.718 + ], + "angle": 0, + "content": "quantitatively measure Multimodal LLMs' object perception and in particular, counting skills. We list our efforts to overcome the issues above in the following paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The contemporary vision-language models [14, 34, 52] owe their success to the availability of large-scale image-text datasets [7, 50, 56]. However, these datasets are more focused on image captioning [33] and VQA [1] tasks, making them unfit for training Multimodal LLMs for basic perception skills like object identification and counting. To overcome the scarcity of fundamental perception-focused image-text data, we leverage images from the COCO [36] dataset and use predictions from off-the-shelf visual perception models [25, 49, 54] to prepare a COCO Segmentation Text (COST) dataset comprising of question-answer pairs about the objects (background and foreground) present in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27993" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.401, + 0.106 + ], + "angle": 0, + "content": "each image. We provide more details in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.289 + ], + "angle": 0, + "content": "Inspired by diffusion models that add various perception \"control\" or \"context\" images [45, 66, 67, 70] as auxiliary inputs to aid image generation, we propose feeding extra perception modalities as control inputs through additional vision encoders, which we term as our Versatile vision encoders (VCoder). In this work, we focus on the task of object perception and leverage a segmentation map, depth map, or both as the control inputs; however, the same design can be extended to other modalities. Our VCoder projects the control inputs' information into the LLM's space as shown in Fig. 4. We hypothesize that this added control helps the MLLM improve its object perception ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.468, + 0.41 + ], + "angle": 0, + "content": "Lastly, owing to the absence of metrics to quantify the counting ability in MLLMs, we propose computing a count score (CS) using one-to-one matching of object words in the ground truth and MLLM's answer. We also compute a hallucination score (HS) based on the extra objects in the MLLM's response that are absent from the ground truth. Similarly, we introduce a depth score (DS) to quantify the object order prediction performance in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.468, + 0.486 + ], + "angle": 0, + "content": "Among the open-source MLLMs, we choose LLaVA-1.5 [39] as our base MLLM. Our extensive experimental analysis demonstrates the importance of our COST dataset and VCoder LLaVA-1.5's improved perception ability. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.488, + 0.468, + 0.533 + ], + "angle": 0, + "content": "- We propose using extra (perception) control inputs and feeding those to a Versatile enCoder (VCoder) for improved object perception performance." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.534, + 0.468, + 0.593 + ], + "angle": 0, + "content": "- We introduce a COCO Segmentation Text (COST) dataset to train and evaluate Multimodal LLM systems on the fundamental object-level perception tasks of object identification, counting, and order prediction." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.594, + 0.468, + 0.682 + ], + "angle": 0, + "content": "- Furthermore, to quantify the object perception ability in MLLMs, we propose calculating a count score (CS), a hallucination score (HS) and a depth score (DS). Our experiments show that the VCoder-adapted LLaVA1.5 outperforms the baseline MLLMs on all metrics when validated on the COST dataset." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.488, + 0.468, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.218, + 0.715 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.248, + 0.741 + ], + "angle": 0, + "content": "2.1. Visual Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The fundamental nature of visual perception makes it a critical component in MLLM systems. Perception can be divided into sub-tasks, including dense prediction tasks like image segmentation [24, 41, 62] and depth estimation [16, 18, 54], and sparse prediction tasks like object detection [6, 63] and pose estimation [13, 58]. In the DL era, initial methods tackled the perception task using CNN based methods [8, 9, 21, 29, 30, 58] with recent methods shifting to the use of transformer-based architectures [10, 18, 25, 54, 65, 75]. In this work, we tackle the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "task of object-level perception, mainly focusing on predicting names, counts, and order of objects using MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.132, + 0.795, + 0.148 + ], + "angle": 0, + "content": "2.2. Visual Understanding with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Using LLMs for vision applications is not a new concept. In a nutshell, developing Multimodal LLMs involves projecting [2, 4, 34, 61] the features from a vision encoder [15, 53] to the embedding space of a language model (LLM) [11, 59, 60], and, visual instruction-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.504 + ], + "angle": 0, + "content": "LLaVA [40] proposed a pipeline to convert existing image-text data into dialog format and then finetuned a CLIP [52] and LLaMA [59] model end-to-end on their collected dataset showing one of the earliest evidence of visual-language instruction tuning. Concurrent to LLaVA, MiniGPT-4 [74] used the visual encoder from BLIP2 [34] and used a linear layer for projecting visual features into Vicuna's [11] feature space. InstructBLIP [14] open-sourced a collection of 16 different datasets covering various vision tasks like VQA, reasoning, captioning, classification, etc., and finetuned a BLIP2 model on their dataset. mPLUGOwl [68] proposed using a vision abstractor and finetuning the vision encoder. More recently, LLaVA-1.5 [39] proposed using an MLP as the projector and finetuned on academic instruction datasets to achieve state-of-the-art performance on various benchmarks [12, 17, 23, 35]. Among various open-source MLLMs [5, 22, 31, 32, 69], we chose LLaVA-1.5 as our baseline due to its superior performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.816, + 0.53 + ], + "angle": 0, + "content": "2.3. Perception Hallucination in MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Since the introduction of LLMs, there has been a comprehensive study about their ability to hallucinate [71] in the NLP community. However, the phenomenon of hallucination in Multimodal LLMs has received comparatively less attention. LRV-Instruction [38] introduced a new instruction-tuning dataset containing 400k visual instructions to prevent hallucination in MLLMs and measured performance treating responses from GPT-4 [48] as ground truths. More recently, HallusionBench [37] quantitatively benchmarked various failure modes in MLLMs that lead to hallucinations based primarily on logical consistency and reasoning. Unlike these works that tried to benchmark MLLMs mainly on VQA-type tasks, this paper focuses on the object-level hallucination in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The two closest works to our objective are POPE [35] and CHAIR [55]. On the one hand, POPE [35] tried to measure hallucination in MLLMs using a binary “Yes”-“No” answer policy in response to questions based on the absence or presence of an object in the image. On the other hand, CHAIR [55] focused on measuring hallucination in image captioning based on only words and not counts for the objects. In our work, we consider not only object words but also the corresponding count to compute an object-level count score and hallucination score." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27994" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.095, + 0.336, + 0.108 + ], + "angle": 0, + "content": "Semantic Object Identification" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.111, + 0.237, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.111, + 0.397, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.084, + 0.23, + 0.382, + 0.257 + ], + "angle": 0, + "content": "USER: What objects can be seen in the image? Perceive as done for semantic segmentation." + }, + { + "type": "image_footnote", + "bbox": [ + 0.086, + 0.272, + 0.387, + 0.3 + ], + "angle": 0, + "content": "ASSISTANT: The objects present in the image are: grass, person, tree, house, car, frisbee, fence, window" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.23, + 0.387, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.095, + 0.664, + 0.108 + ], + "angle": 0, + "content": "Instance Object Identification" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.111, + 0.569, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.111, + 0.728, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.42, + 0.23, + 0.718, + 0.257 + ], + "angle": 0, + "content": "USER: What objects can be seen in the image? Perceive as done for instance segmentation." + }, + { + "type": "image_footnote", + "bbox": [ + 0.418, + 0.273, + 0.704, + 0.3 + ], + "angle": 0, + "content": "ASSISTANT: The objects present in the image are: two people, car, frisbee." + }, + { + "type": "list", + "bbox": [ + 0.418, + 0.23, + 0.718, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.106, + 0.85, + 0.132 + ], + "angle": 0, + "content": "from DINov2 DPT" + }, + { + "type": "text", + "bbox": [ + 0.778, + 0.148, + 0.842, + 0.173 + ], + "angle": 0, + "content": "from OneFormer" + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.191, + 0.85, + 0.215 + ], + "angle": 0, + "content": " from COCO dataset" + }, + { + "type": "text", + "bbox": [ + 0.756, + 0.233, + 0.865, + 0.26 + ], + "angle": 0, + "content": "question text from prompting GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.758, + 0.274, + 0.863, + 0.3 + ], + "angle": 0, + "content": "answer text from OneFormer output" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.314, + 0.237, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.314, + 0.398, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.085, + 0.433, + 0.382, + 0.46 + ], + "angle": 0, + "content": "USER: What objects can be seen in the image? Perceive as done for panoptic segmentation." + }, + { + "type": "image_footnote", + "bbox": [ + 0.085, + 0.475, + 0.389, + 0.504 + ], + "angle": 0, + "content": "ASSISTANT: The objects present in the image are: grass, tree, house, two people, frisbee, car." + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.433, + 0.389, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.314, + 0.57, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.314, + 0.728, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.314, + 0.888, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.416, + 0.433, + 0.862, + 0.462 + ], + "angle": 0, + "content": "USER: Could you list the objects in the image in order of their perceived distance from the foreground to the background?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.416, + 0.475, + 0.858, + 0.504 + ], + "angle": 0, + "content": "ASSISTANT: The depth order for objects present in the image is: frisbee, person, person-2, car, tree, house, grass." + }, + { + "type": "list", + "bbox": [ + 0.416, + 0.433, + 0.862, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.512, + 0.334, + 0.526 + ], + "angle": 0, + "content": "Panoptic Object Identification" + }, + { + "type": "title", + "bbox": [ + 0.567, + 0.512, + 0.726, + 0.527 + ], + "angle": 0, + "content": "Object Order Perception" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.533, + 0.894, + 0.588 + ], + "angle": 0, + "content": "Figure 3. Organization of the COST dataset. We incorporate the images from COCO [36], the questions from GPT-4 [48], and the segmentation outputs from OneFormer [25] in a question-answer format for training and evaluating MLLMs on the object identification task. We also extend COST to the object order perception task by incorporating depth map outputs from DINOv2 [49] DPT [54]. COST can be extended to more object-level tasks by similarly incorporating other modalities (for example, keypoint maps)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.6, + 0.393, + 0.616 + ], + "angle": 0, + "content": "3. Object Identification with MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.627, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Suppose you are invited to a Halloween party and want to bring candies for every person at that party. You ask your friend to send you a picture (Fig. 1) of the party room so that you can estimate the number of people and the number of candies you need to buy. In a hurry, you ask GPT-4V [48]: \"Can you count the number of people in the image?\", and it responds: \"Yes, there are ten people visible in the image.\" Excited, you arrive at the party with ten candies but wait, you see fourteen people! Confused, you look at the image your friend sent you, and you can count fourteen people in that image, realizing that GPT-4V fails at the simple task of counting the people in the picture. At the same time, it can accurately describe the happening of a Halloween party in the image (Fig. 1). We refer to the phenomenon of Multimodal LLMs failing at simple visual perception tasks while succeeding at complex visual reasoning tasks as Moravec's Paradox [44] in perception." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We hypothesize that one of the main reasons for the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.782 + ], + "angle": 0, + "content": "above phenomenon is the absence of conversations covering object identification for the salient objects and the objects in the background from the instruction-tuning data for MLLMs. To overcome this issue, we prepare the COCO Segmentation Text (COST) dataset with COCO [36] images and create sentences using the output from an image segmentation model [25] to obtain an image-text dataset to train and evaluate MLLMs for object perception MLLMs. Moreover, we also introduce a segmentation map as a control image input to the MLLM for better performance and quantify object perception performance with a count score (CS) and a hallucination score (HS)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.799, + 0.842, + 0.815 + ], + "angle": 0, + "content": "3.1. COST to Identify Objects with MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We find that image segmentation methods [10, 25] can accurately identify salient (foreground objects like people, cars, etc.) and background objects (like sky, wall, etc.) in a given scene. Guided by this finding, we use images from the COCO [36] dataset and obtain the correspond" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27995" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.652, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.091, + 0.895, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.893, + 0.463 + ], + "angle": 0, + "content": "Figure 4. Adapting Multimodal LLMs for accurate object perception with VCoder. (a) We add our VCoder as an adapter to the LLaVA-1.5 [39] and feed perception modalities as extra control inputs for improved object perception performance. During training, we freeze the components from LLaVA-1.5 (ImCoder, MLP, and LLM) to retain the original reasoning performance. (b) Using depth map and segmentation map as the control inputs to VCoder for the object order perception task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.473, + 0.744 + ], + "angle": 0, + "content": "ing segmentation outputs from OneFormer [25], a state-of-the-art image segmentation model. Next, we extract the object (class) names and counts from the segmentation outputs and convert them into a sentence form for the ground-truth answer: \"The objects present in the image are: \\([CNT_1][OBJ_1],[CNT_2][OBJ_2],\\ldots,[CNT_N][OBJ_N]\". with \\[OBJ_i\\] representing the object name and \\([CNT_i]\\) representing the count (if greater than one) for the \\(i^{th}\\) object in the image. We prompt GPT-4 [48] to collect a bucket of questions for three different object identification tasks: semantic, instance, and panoptic, corresponding to the three different image segmentation tasks. Finally, as shown in Fig. 3, we organize the images from COCO, segmentation maps from OneFormer, questions from GPT-4, and sentences containing object information into a question-answer format to construct our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object identification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Statistically, we prompt GPT-4 [48] to return 20 questions for each question bucket (panoptic, semantic, and instance). In total, we used 280k images from the train2017, test2017, and unlabeled2017 splits of the COCO [36] dataset and corresponding segmentation outputs from OneFormer [25] to form the visual component of the COST training dataset. Similarly, we prepare a COST validation split using the 5k images from the val2017 split of the COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Note that a similar approach can extend the COST" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.732 + ], + "angle": 0, + "content": "dataset to other perception modalities. In this work, we incorporate the depth map modality into our COST dataset for the object order perception task. Particularly, we leverage the publicly available DINOv2 [49] DPT [54] model to obtain depth maps for COCO images and use the panoptic mask (from OneFormer [25]) to estimate the depth order of objects in an image. We format the obtained ordering of objects into the text with the template: \"The depth order for objects present in the image is: \\([OBJ_1], [OBJ_2], \\ldots, [OBJ_J]\\),\" with \\([OBJ_j]\\) representing the \\(j^{th}\\) object name. To maintain relative ordering among objects belonging to the same class, we append a count number to the second and later objects, as shown in the bottom right of Fig. 3 for person and person-2. Similar to the previous setting, we prompt GPT-4 [48] to return 20 questions for the object order perception task. We provide a detailed flow of obtaining ground-truth object orders in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.773, + 0.756 + ], + "angle": 0, + "content": "3.2. VCoder for Multimodal LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We notice that existing open-source Multimodal LLMs generally use the ViT [15] from CLIP [53] as the image encoder (ImCoder) during instruction tuning. We reason that the ViT focuses mainly on salient objects because it is trained against captions, which leave out information about background regions. We argue that identifying objects in the background is critical for a Multimodal LLM to become skilled at perception. To overcome this limitation, we introduce a segmentation map as a control input [45, 70] into" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27996" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.198 + ], + "angle": 0, + "content": "our Multimodal LLM. Specifically, we use the segmentation map from OneFormer [25] and project it to the LLM's embedding space using a pretrained ViT [15] (from CLIP [53]) as a SegCoder and a two-layer MLP [39] which we collectively refer to as our Versatile enCoder (VCoder). This extra control from the segmentation map results in considerable performance gains on the object identification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.38 + ], + "angle": 0, + "content": "As shown in Fig. 4a, our VCoder adapted MLLM takes three sets of inputs: perception modalities as control inputs fed into the VCoder, an RGB image fed into an Image enCoder (and MLP), and the text question. The RGB image and text are tokenized to the and tokens, respectively. VCoder is flexible at handling various perception modalities with a unique token for each modality. For example, the segmentation map and depth map inputs are tokenized to and tokens, respectively. Similarly, one can incorporate more modalities with modality-specific tokens. Finally, all tokenized embeddings are concatenated and fed into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.489 + ], + "angle": 0, + "content": "We treat our VCoder as an adapter, added to our base MLLM, LLaVA-1.5 [39] to obtain the final MLLM framework for experiments. Note that we only train the MLP components in the VCoder on the COST dataset. We decided to keep all other parameters fixed during training to keep the reasoning ability unaffected while achieving improved object perception performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.457, + 0.521 + ], + "angle": 0, + "content": "3.3. Evaluating MLLMs for Object Identification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.667 + ], + "angle": 0, + "content": "Despite the availability of various metrics [35, 42, 55] to measure object hallucination in vision-language models, no existing metric considers the explicit object counts while calculating their hallucination scores. We argue that object counts returned by an MLLM are a critical component that should not be overlooked while evaluating object identification performance. Therefore, we propose evaluating object identification performance in MLLMs using two metrics: count-score (CS) and hallucination-score (HS)." + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.678, + 0.47, + 0.704 + ], + "angle": 0, + "content": "\\[\nG _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {G}: \\mathrm {C N T} _ {1} ^ {G}; \\dots ; \\mathrm {O B J} _ {N} ^ {G}: \\mathrm {C N T} _ {N} ^ {G} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.7, + 0.43, + 0.718 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {P}: \\mathrm {C N T} _ {1} ^ {P}; \\dots ; \\mathrm {O B J} _ {M} ^ {P}: \\mathrm {C N T} _ {M} ^ {P} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 5, given a ground-truth sentence \\((G)\\) and an MLLM predicted response \\((P)\\), we first extract the object words (nouns) and their corresponding count from both text samples and represent them in a dictionary form with keys as the object noun and the value as the corresponding object's count as shown in Eq. (1) with \\(N\\) and \\(M\\) representing the number of different object nouns in the \\(G\\) and \\(P\\) respectively. Next, we perform one-to-one matching between the counts for keys with \\(G_{\\mathrm{dict}}\\) and \\(P_{\\mathrm{dict}}\\) as the reference for Count Score (CS) and Hallucination Score (HS), respectively, as shown in Eq. (2)." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.088, + 0.872, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.288, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 5. Evaluation Metrics for Object Identification. We compare the object counts in the ground truth and prediction to calculate a count score (CS) and a hallucination score (HS)." + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.354, + 0.856, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C S} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\left\\{ \\begin{array}{l l} \\frac {\\min \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)}{\\max \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)} & \\text {i f I (O B J} _ {i} ^ {G}, P _ {\\text {d i c t}}) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.401, + 0.888, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H S} = \\frac {1 0 0}{M} \\sum_ {j = 1} ^ {M} \\left\\{ \\begin{array}{l l} 1 - \\frac {\\operatorname* {m i n} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)}{\\operatorname* {m a x} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)} & \\text {i f} I (\\mathrm {O B J} _ {j} ^ {P}, G _ {\\text {d i c t}}) \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.452, + 0.892, + 0.506 + ], + "angle": 0, + "content": "\\[\nI (\\mathrm {O B J}, D) = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f O B J i s i n k e y s} (D) \\\\ \\text {F a l s e} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Count Score (CS). It represents the percentage of correct object counts predicted by the MLLM with respect to the ground-truth sentence. The higher the CS, the better." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Hallucination Score (HS). It represents the percentage of extra object counts predicted by the MLLM that do not exist in the ground-truth sentence. The lower the HS, the better." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.607, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Note that due to the one-to-one word-matching nature of our evaluation, we manually define a mapping between the categories in COCO [36] and their synonyms [43, 55]. For example, we replace words like man, woman, child, boy, girl, etc. with person in the prediction before evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.633, + 0.711 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use LLaVA-1.5 [39] as our base MLLM. LLaVA-1.5 uses CLIP-ViT-L-336px [53] as the image encoder (ImCoder) with a two-layer MLP as projection and Vicuna1.5 [73] as the LLM. Inside our VCoder, we also use a CLIP-ViT-L-336px to encode the control inputs and project the features into the LLM embedding space using modality-specific two-layer MLPs. We resize the visual inputs to \\(336 \\times 336\\) resolution (corresponds to 576 tokens) for our MLLM. During training, we load the instruction-tuned weights from LLaVA-1.5 and keep those frozen while only tuning the MLP component of our VCoder. We use the publicly available OneFormer [25] model trained on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27997" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.086, + 0.891, + 0.403 + ], + "angle": 0, + "content": "
MethodInput TokensSemanticInstancePanoptic
CS (↑)HS (↓)CS (↑)HS (↓)CS (↑)HS (↓)
Closed Model, Open API
GPT-4V [48]⟨img⟩ + ⟨query⟩————38.483.0
Existing Open-Source Multimodal LLMs
MiniGPT-4 LLaMA-2-7b [74]⟨img⟩ + ⟨query⟩6.292.25.697.76.294.9
InstructBLIP Vicuna-7b [14]⟨img⟩ + ⟨query⟩14.285.825.391.917.591.2
LLaVA-1.5-7b [39]⟨img⟩ + ⟨query⟩30.660.150.375.938.767.3
LLaVA-1.5-13b [39]⟨img⟩ + ⟨query⟩25.069.349.975.035.868.6
CogVLM-17b [64]⟨img⟩ + ⟨query⟩33.467.543.586.240.675.9
Baselines trained on the COST dataset
COST IT LLaVA-1.5-7b⟨img⟩ + ⟨query⟩78.722.167.530.371.928.2
Soft-Prompted LLaVA-1.5-7b⟨prompt⟩ + ⟨img⟩ + ⟨query⟩36.256.718.472.226.863.0
ImCoder LLaVA-1.5-7b⟨img⟩ + ⟨img⟩ + ⟨query⟩78.922.764.029.470.827.9
VCoder augmented LLaVA-1.5
VCoder LLaVA-1.5-7b⟨seg⟩ + ⟨img⟩ + ⟨query⟩88.610.471.126.986.012.8
VCoder LLaVA-1.5-13b⟨seg⟩ + ⟨img⟩ + ⟨query⟩89.010.073.325.087.211.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.407, + 0.895, + 0.518 + ], + "angle": 0, + "content": "Table 1. Comparison to baseline Multimodal LLMs on the COST validation dataset for Object Identification. We compare our VCoder to existing off-the-shelf baseline MLLMs: MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64]. We also train three different variants of LLaVA-1.5 on the COST dataset: COST IT mixes the COST training data with the instruction tuning data; Soft-Prompted uses a set of learnable tokens, and ImCoder uses an RGB image as the control input. Our VCoder adapted LLaVA-1.5 performs the best on all three object perception tasks. Note: \\(\\langle \\cdot \\rangle\\) denotes input tokens to LLM with seg representing segmentation map, img representing RGB image, prompt representing learnable prompt, and query representing the user question. We also evaluate the performance of GPT-4V [48] on the COST dataset using the publicly accessible paid API released by OpenAI. Our VCoder-adapted LLaVA-1.5 shows the best performance on object identification among all MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.613 + ], + "angle": 0, + "content": "COCO [36] with DiNAT-L [19, 20] backbone to obtain the segmentation map. For getting depth maps, we use the publicly available ViT-L/14 distilled variant of DINOv2 [49] DPT [54] trained on the NYUd [46] dataset. In this section, we discuss our results on the object identification task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.626, + 0.295, + 0.643 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Training Details. We train our VCoder-adapted LLaVA-1.5 framework for two epochs on the COST training dataset with a batch size 256 and a learning rate of \\(1e^{-3}\\). For other training hyperparameters, we follow the settings used during the instruction-tuning stage in LLaVA-1.5 [39]. Following [25], we uniformly sample each object identification task (semantic, instance, and panoptic) during training. We also use the corresponding segmentation map from OneFormer [25] as input to the VCoder during training and inference. On 8 A100 GPUs, it takes 8 and 14 hours to train our VCoder with the 7b and 13b variants of LLaVA-1.5 as the base MLLM, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation Details. We evaluate all MLLMs on the COST validation set. We separately evaluate semantic, instance, and panoptic object identification tasks while randomly sampling questions from the corresponding task's question" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.659 + ], + "angle": 0, + "content": "bucket. Note that for evaluating all off-the-shelf MLLMs, we experiment with various prompts and finally use the prompt: \"[QUESTION]. Return the answer in the paragraph format: 'The objects present in the image are: ...' and then list the objects with their count in word format (if greater than 1) in front of them, like 'two people'\"; where [QUESTION] is the randomly sampled question from the object identification task bucket." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.64, + 0.684 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Baselines. We compare the performance of VCoder to open-source Multimodal LLMs, namely, MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64] on the COST validation set in Tab. 1. Furthermore, we provide three additional baselines, trained for two epochs:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.809 + ], + "angle": 0, + "content": "COST IT LLaVA-1.5: We mix the COST training data with the instruction tuning data used in LLaVA-1.5 [39] and fine-tune a LLaVA-1.5 [39] model from scratch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Soft-Prompted LLaVA-1.5: We pretend 576 learnable tokens \\(\\langle \\langle prompt\\rangle \\rangle\\) to the LLM input and tune only the \\(\\langle prompt\\rangle\\) parameters on the COST training dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.885 + ], + "angle": 0, + "content": "ImCoder LLaVA-1.5: We use an RGB image as the control input and train on the COST training dataset." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1, we notice that all existing MLLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27998" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.406, + 0.164 + ], + "angle": 0, + "content": "
MethodDepth Score (↓)
LLaVA-1.5-7b [39]166.1
LLaVA-1.5-13b [39]227.2
VCoder-DS LLaVA-1.5-7b65.9
VCoder-DS LLaVA-1.5-13b63.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.21 + ], + "angle": 0, + "content": "Table 2. Performance on Object Order Perception. Our VCoder LLaVA-1.5 considerably outperforms LLaVA-1.5 [39], owing to the usage of control inputs and training on the COST dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.469, + 0.369 + ], + "angle": 0, + "content": "perform poorly on our COST validation set, demonstrating their inability to count and identify objects accurately. Note that existing MLLMs perform relatively better on instance object identification, reaffirming our claim that MLLMs are better at detecting salient objects than background objects. Although the baselines trained on the COST dataset perform relatively better, they still lag in performance compared to the VCoder. Notably, a segmentation map performs considerably better than using an RGB image as the control input, proving the segmentation map's vitality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Comparison to GPT-4V [48]. We utilize OpenAI's newly released gpt-4-vision-preview\\(^1\\) API to obtain responses from GPT-4V. Our experiments show that GPT-4V's responses are consistent across all object identification tasks, closely aligning with the panoptic identification task. Therefore, we compare our VCoder to GPT-4V only on the panoptic object identification to reduce API requests due to a daily limit of 500 API requests during this project. As shown in Tab. 1, GPT-4V [48] lags behind our VCoder by a considerable margin, reaffirming our claim that existing MLLMs cannot perform accurate object-level perception." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.427, + 0.568 + ], + "angle": 0, + "content": "5. Object Order Perception with MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.726 + ], + "angle": 0, + "content": "As shown in Fig. 4, multiple perception modalities can be leveraged to improve object perception in MLLMs with our VCoder. This section presents our experiments with our VCoder using the segmentation and depth maps as the control inputs. We term the resulting MLLM as VCoder-DS LLaVA-1.5. Intuitively, predicting the object order implicitly means identifying the objects in an image. Therefore, for the object order perception task (Fig. 4b), we use both and inputs, while only the input as the additional control for object identification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.469, + 0.847 + ], + "angle": 0, + "content": "During training, we use a mixture of datasets, including the object identification and object order perception components from the COST dataset. We also use about 200k image-conversation (along with the corresponding segmentation map obtained using OneFormer [25]) pairs randomly sampled from the instruction tuning data used in LLaVA1.5 [39]. We train our VCoder for one epoch following the same hyperparameter settings mentioned in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.848, + 0.469, + 0.878 + ], + "angle": 0, + "content": "As shown in Tab. 2, our VCoder-DS LLaVA-1.5 significantly outperforms the base MLLM, LLaVA-1.5 [39] on" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the COST validation set. For quantitative evaluation of MLLMs on the object order perception task, we calculate a depth score (DS) using the absolute difference between the position of objects in the ground truth and prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.165, + 0.622, + 0.181 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Despite the improved object perception performance after training our VCoder on the COST dataset, certain limitations remain to be addressed for future work. Firstly, we build our COST dataset using OneFormer [25], which can only perceive objects belonging to a limited number of categories due to being trained on a closed-set vocabulary dataset [36]. For real-world applications, it is imperative to develop an object perception benchmark for MLLMs covering many more classes with varying granularity. Secondly, the count, hallucination, and depth scores use one-to-one word matching, which requires defining a mapping between synonymous words. It will be promising to explore ways to overcome using synonym mappings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.4, + 0.618, + 0.416 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.789 + ], + "angle": 0, + "content": "This work analyzes the object-level perception skills of Multimodal Large Language Models (VLMMs). Although MLLMs are good visual reasoners, they need to improve at the simple yet fundamental task of object perception. To improve object perception ability in MLLMs, we propose the COST dataset for training and evaluating MLLMs at the object perception task. We benchmark different off-the-shelf MLLMs and GPT-4V on our COST dataset and observe their lousy performance. Consequently, we propose using perception modalities as control inputs and a Versatile vision enCoders (VCoder) as an adapter for projecting the control inputs to the LLM embedding space. Our VCoder can easily be extended to leverage various modalities as the control inputs depending on the task. To quantify the object-level perception ability in MLLMs, we introduce a Count-Score (CS), a Hallucination-Score (HS), and a Depth-Score (DS). We adapted LLaVA-1.5 with VCoder, only trained the VCoder on our COST dataset, and demonstrated its improved performance at the object perception task while retaining the reasoning performance. We hope our work can inspire the research community to focus on developing object perception datasets for MLLMs and develop vision systems that are equally good at perception and reasoning in the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work was in part supported by IARPA under Contract No. 2022-21102100004, NSF CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. We also thank the ML Center at Georgia Tech for generously supporting this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.466, + 0.9 + ], + "angle": 0, + "content": "1https://platform.openai.com/docs/guides/vision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27999" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.3, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[3] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. Falcon-40B: an open large language model with state-of-the-art performance. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.471, + 0.652 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.655, + 0.47, + 0.696 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.699, + 0.47, + 0.753 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.757, + 0.469, + 0.811 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.815, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "transformer for universal image segmentation. In CVPR, 2022.3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[12] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[13] Timothy F Cootes, Gareth J Edwards, and Christopher J Taylor. Active appearance models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681-685, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[16] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in neural information processing systems, pages 2366-2374, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[17] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[18] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[19] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[22] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. ArXiv, abs/2302.14045, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[23] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Ji-achen Li, Steven Walton, and Humphrey Shi. Semask: Se" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28000" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "mentally masking transformer backbones for effective semantic segmentation. arXiv, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[25] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[26] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Yadong Mu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[28] Deanna Kuhn. The Skills of Argument. Cambridge University Press, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.312, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth International Conference on 3D Vision (3DV), pages 239-248. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.384, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[30] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[31] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.472, + 0.469, + 0.526 + ], + "angle": 0, + "content": "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[35] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[37] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.891, + 0.164 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[41] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[42] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.891, + 0.295 + ], + "angle": 0, + "content": "[43] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.298, + 0.891, + 0.325 + ], + "angle": 0, + "content": "[44] H. Moravec. Mind children: The future of robot and human intelligence. Harvard University Press, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.328, + 0.892, + 0.382 + ], + "angle": 0, + "content": "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[46] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.429, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[47] OpenAI. Chatgpt. https://chat.openai.com/, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.865, + 0.473 + ], + "angle": 0, + "content": "[48] OpenAI. Gpt-4 technical report, 2023. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[50] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. ArXiv, abs/2306, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. arXiv, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[54] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28001" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[55] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In EMNLP, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops 2021, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[58] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[61] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[62] Z. Tu, Xiangrong Chen, Alan Yuille, and Song Zhu. Image parsing: Unifying segmentation, detection, and recognition. In IJCV, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[63] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In CVPR, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and ef" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[66] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[67] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[68] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[69] Yan Zeng, Hanbo Zhang, Jiani Zheng, Jiangnan Xia, Guoqiang Wei, Yang Wei, Yuchen Zhang, and Tao Kong. What matters in training a gpt4-style language model with multimodal inputs? arXiv preprint arXiv:2307.02469, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[70] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[71] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[72] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[73] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[74] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28002" + } + ] +] \ No newline at end of file diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_origin.pdf b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79483215d774aa55ecfc0adffaf689246802e0b2 --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/817f63e6-6ca6-4fee-9f1a-257768563a6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:445affc5b3e2319a2f91998a2e9f5e9bd11f93070e86629c65e2ff5041d8b70b +size 4040225 diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/full.md b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bebaf606804e9be6b07b87b2c5b22ca966e7ef9a --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/full.md @@ -0,0 +1,456 @@ +# VCoder: Versatile Vision Encoders for Multimodal Large Language Models + +Jitesh Jain $^{1}$ Jianwei Yang $^{2}$ Humphrey Shi $^{1,3}$ + +$^{1}$ SHI Labs @ Georgia Tech $^{2}$ Microsoft Research, Redmond $^{3}$ Picsart AI Research (PAIR) + +https://github.com/SHI-Labs/VCoder + +# Abstract + +Humans possess the remarkable skill of Visual Perception, the ability to see and understand the seen, helping them make sense of the visual world and, in turn, reason. Multimodal Large Language Models (MLLM) have recently achieved impressive performance on vision-language tasks ranging from visual question-answering and image captioning to visual reasoning and image generation. However, when prompted to identify or count (perceive) the entities in a given image, existing MLLM systems fail. Working towards developing an accurate MLLM system for perception and reasoning, we propose using Versatile vision en-Coders (VCoder) as perception eyes for Multimodal LLMs. We feed the VCoder with perception modalities such as segmentation or depth maps, improving the MLLM's perception abilities. Secondly, we leverage the images from COCO and outputs from off-the-shelf vision perception models to create our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object perception task. Thirdly, we introduce metrics to assess the object perception abilities in MLLMs on our COST dataset. Lastly, we provide extensive experimental evidence proving the VCoder's improved object-level perception skills over existing Multimodal LLMs, including GPT-4V. We open-source our dataset, code, and models to promote research. + +# 1. Introduction + +'Perception is the soil; reasoning, the seed. Without fertile ground, the seed cannot flourish.' + +(GPT-4 [48], 2023) + +The ability to think and reason is one of the most remarkable traits that help humans function daily. Generally, understanding the environment precedes the act of thinking and reasoning [28]. Following the success of ChatGPT-like instruction following AI agents [3, 11, 47, 48, 60] at language understanding and reasoning, researchers have leveraged LLMs to develop instruct frameworks [14, 40, 68, 74] that can understand vision and language inputs in an effort to imitate human perception and reasoning ability. We + +![](images/a33cf8d5cc97945d86c635dc43762b30a586590bd13807c0dc74ef2fda3a96b2.jpg) + +![](images/7e452c7f4f23ac429c598f27d17519c50e8933bca261b74aa876d83c01f1c714.jpg) + +at is happening in the image? + +Difficulty Level + +![](images/f1575724ae16c41b02d01f25dee5ba0c162882520f5ba01dcbc765d7af2a9ec4.jpg) + +# GPT-4V + +Characters from "Friends" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume. + +![](images/81344a50af2718e479e7c1e048a6000cb55be66574ef99b11bf39adc76755975.jpg) +Figure 1. GPT-4V [48] (access date: Dec 16, 2023) returns impressive responses when prompted to describe complex visual scenes. However, it fails at the simple task of counting in the same scene. Our VCoder returns the correct count of people. + +# VCoder + +The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party. + +![](images/7986ddffcc6976567fabd9c85ad0d55109c9d17abdc5a5bae8cbdbd9c18909f6.jpg) + +Can you count the number of people in the image? + +Difficulty Level + +![](images/a465be8f016c7c2fee0dac85a91c639e77ac5f6cc354fa3880e2db2a16bf65ca.jpg) + +# GPT-4V + +Yes, there are ten people visible in the image. + +![](images/0d8d0cadd883301e5c7d9b3144bd1721d42d9a3425657cc76202caf637b6ef2d.jpg) + +# VCoder + +Yes, there are a total of 14 people in the image. + +refer to such systems as Multimodal LLMs (MLLM). Although MLLMs exhibit the ability to perform complex vision-language tasks like visual captioning [2, 4, 39], image generation [26, 27, 57], visual reasoning and grounding [22, 51, 72], they often display sub-par performance at simple tasks like counting objects (Fig. 1). As shown in Fig. 2, MLLMs output incorrect object counts (people, chairs) and hallucinate about the presence (handbag, couch) of certain objects when prompted to identify entities in a visual input. The perception performance is much worse + +# USER + +What objects can be seen in the image? + +# USER + +What objects can be seen in the image? + +# USER + +What objects can be seen in the image? + +# USER + +What is happening in the image? + +![](images/f64e7236a0404dc1fadf4fe07010d4e9e5a2108a61a75e673e718f9f6690caf3.jpg) +Figure 2. MLLMs counting and identifying objects. As shown in the first column, GPT-4V [48] (access date: Dec 16, 2023) and LLaVA-1.5 [39] both fail at counting people. Moreover, LLaVA-1.5 [39] misses background entities like window, wall, etc. and hallucinates about the presence of a handbag. VCoder can predict the people counts and other background entities accurately except chairs. Similarly, in the second column, GPT-4V and LLaVA-1.5 fail at counting chairs while the VCoder matches the Oracle's performance. Notably, all MLLMs can perceive objects accurately for a non-cluttered image in the third column, with LLaVA-1.5 failing at counting ties. Our VCoder can also accurately perform general question-answering tasks, as shown in the fourth column. We treat OneFormer [25] as the Oracle for object perception. Red text represents counting mistakes; pink text represents hallucination; blue text represents correct object perception. + +![](images/e239ad6519c95bb5f6688e29f8e8aa194e00a04c870ee0001134d8c7b3383520.jpg) + +![](images/fd6c711c84170ee68b6f4a19c7e1db48db20e232e1ea5c28cf980aadf48814c7.jpg) + +![](images/2e63190aef579ad86efb7a24543562ac404b65f177bc38939acb5afb9030d0ff.jpg) + +![](images/c614f92a5cf193e12ec5f9dbbb292af9e0a2c0657bc7537ffcbad6306ab6e6de.jpg) + +# GPT-4V + +The objects present in the image are: eleven people, one large potted plant, two ceiling lamps, one exit sign, and various pieces of furniture that appear to be office desks and chairs. + +![](images/b65306bffef51c66058ccede5b60e585c2c534a8d4ac2cc63e00ae3ad33b2016.jpg) + +# T-4V + +The objects present in the image are: two Corgi dogs, two umbrellas, multiple chairs, two speakers, and one palm tree. The setting is an outdoor area with a beachfront view. + +# GPT-4V + +The objects present in the image are: two people (1 male and 1 female), a table, windows, and buildings seen through the windows. + +# GPT-4V + +Characters from "Friends" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume. + +![](images/7552bb6475bf959d9117a786e069a22478c1f3ab3f8e31956a9a693e1676967d.jpg) + +# LLaVA-1.5 + +The objects present in the image are: 12 people, 1 potted plant, 1 TV, 2 chairs, 1 handbag. + +![](images/45973119da94a73ddf1a687610e895582eca4e3f4fce7bb4baaba606c1cdfc17.jpg) + +# vA-1.5 + +The objects present in the image are: 2 dogs, 3 couches, 3 umbrellas, 1 table, and 2 chairs. + +# LaVA-1.5 + +The objects present in the image are: man, woman, ties (2), suit (1), glass wall (1), building (1), desk (1). + +# LLaVA-1.5 + +In the image, a group of people is gathered in a room, sitting around a table. They are dressed in costumes, possibly for a party or a themed event. + +![](images/e80386a4a303de7610ddb8bb5e3b51ee4b58f1973e284548425be081b994e518.jpg) + +# VCoder + +The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, three chairs. + +![](images/2e62166f332d6bf694d9ccb74d469d6c899c75c7f02b6fa208f5a556d777aeeb.jpg) + +# oder + +The objects present in the image are: sky, tree, pillow, mountain, water, grass, bed, two dogs, umbrella, potted plant, dining table, six chairs. + +![](images/d2131fbc61eef778ea71423e0776188fcb7f442143baa7784752142b9b6d90bc.jpg) + +The objects present in the image are: window, wall, two people, tie. + +# VCoder + +The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party. + +# OneFormer (Oracle) + +The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, chair, TV. + +# OneFormer (Oracle) + +The objects present in the image are: sky, tree, sea, pillow, wall, bed, two dogs, dining table, six chairs, grass, umbrella, potted plant, mountain. + +Text in red denotes wrong count. + +Text in pink denotes hallucination. + +Text in blue denotes correct object counts. + +when the scenes are cluttered with many entities. Consequently, a natural question arises: "How to develop MLLM systems that respond to perception questions accurately?" + +This work aims to improve Multimodal LLMs at the simple yet fundamental object-level perception skills, including counting. Our motivation stems from the intuition that one can only describe and reason about a visual scene with the correct understanding of the entities in the image. In our effort to develop an accurate Multimodal LLM perception system, we face three significant challenges: (i) the scarcity of a vision-language dataset focused on the object perception task; (ii) existing open-sourced Multimodal LLMs usually use the ViT from CLIP [52] with an RGB image as input as the visual component that majorly focuses only on salient objects, and (iii) the absence of evaluation metrics to + +quantitatively measure Multimodal LLMs' object perception and in particular, counting skills. We list our efforts to overcome the issues above in the following paragraphs. + +The contemporary vision-language models [14, 34, 52] owe their success to the availability of large-scale image-text datasets [7, 50, 56]. However, these datasets are more focused on image captioning [33] and VQA [1] tasks, making them unfit for training Multimodal LLMs for basic perception skills like object identification and counting. To overcome the scarcity of fundamental perception-focused image-text data, we leverage images from the COCO [36] dataset and use predictions from off-the-shelf visual perception models [25, 49, 54] to prepare a COCO Segmentation Text (COST) dataset comprising of question-answer pairs about the objects (background and foreground) present in + +each image. We provide more details in Sec. 3.1. + +Inspired by diffusion models that add various perception "control" or "context" images [45, 66, 67, 70] as auxiliary inputs to aid image generation, we propose feeding extra perception modalities as control inputs through additional vision encoders, which we term as our Versatile vision encoders (VCoder). In this work, we focus on the task of object perception and leverage a segmentation map, depth map, or both as the control inputs; however, the same design can be extended to other modalities. Our VCoder projects the control inputs' information into the LLM's space as shown in Fig. 4. We hypothesize that this added control helps the MLLM improve its object perception ability. + +Lastly, owing to the absence of metrics to quantify the counting ability in MLLMs, we propose computing a count score (CS) using one-to-one matching of object words in the ground truth and MLLM's answer. We also compute a hallucination score (HS) based on the extra objects in the MLLM's response that are absent from the ground truth. Similarly, we introduce a depth score (DS) to quantify the object order prediction performance in MLLMs. + +Among the open-source MLLMs, we choose LLaVA-1.5 [39] as our base MLLM. Our extensive experimental analysis demonstrates the importance of our COST dataset and VCoder LLaVA-1.5's improved perception ability. To summarize, our contributions are as follows: + +- We propose using extra (perception) control inputs and feeding those to a Versatile enCoder (VCoder) for improved object perception performance. +- We introduce a COCO Segmentation Text (COST) dataset to train and evaluate Multimodal LLM systems on the fundamental object-level perception tasks of object identification, counting, and order prediction. +- Furthermore, to quantify the object perception ability in MLLMs, we propose calculating a count score (CS), a hallucination score (HS) and a depth score (DS). Our experiments show that the VCoder-adapted LLaVA1.5 outperforms the baseline MLLMs on all metrics when validated on the COST dataset. + +# 2. Related Work + +# 2.1. Visual Perception + +The fundamental nature of visual perception makes it a critical component in MLLM systems. Perception can be divided into sub-tasks, including dense prediction tasks like image segmentation [24, 41, 62] and depth estimation [16, 18, 54], and sparse prediction tasks like object detection [6, 63] and pose estimation [13, 58]. In the DL era, initial methods tackled the perception task using CNN based methods [8, 9, 21, 29, 30, 58] with recent methods shifting to the use of transformer-based architectures [10, 18, 25, 54, 65, 75]. In this work, we tackle the + +task of object-level perception, mainly focusing on predicting names, counts, and order of objects using MLLMs. + +# 2.2. Visual Understanding with LLMs + +Using LLMs for vision applications is not a new concept. In a nutshell, developing Multimodal LLMs involves projecting [2, 4, 34, 61] the features from a vision encoder [15, 53] to the embedding space of a language model (LLM) [11, 59, 60], and, visual instruction-tuning. + +LLaVA [40] proposed a pipeline to convert existing image-text data into dialog format and then finetuned a CLIP [52] and LLaMA [59] model end-to-end on their collected dataset showing one of the earliest evidence of visual-language instruction tuning. Concurrent to LLaVA, MiniGPT-4 [74] used the visual encoder from BLIP2 [34] and used a linear layer for projecting visual features into Vicuna's [11] feature space. InstructBLIP [14] open-sourced a collection of 16 different datasets covering various vision tasks like VQA, reasoning, captioning, classification, etc., and finetuned a BLIP2 model on their dataset. mPLUGOwl [68] proposed using a vision abstractor and finetuning the vision encoder. More recently, LLaVA-1.5 [39] proposed using an MLP as the projector and finetuned on academic instruction datasets to achieve state-of-the-art performance on various benchmarks [12, 17, 23, 35]. Among various open-source MLLMs [5, 22, 31, 32, 69], we chose LLaVA-1.5 as our baseline due to its superior performance. + +# 2.3. Perception Hallucination in MLLMs + +Since the introduction of LLMs, there has been a comprehensive study about their ability to hallucinate [71] in the NLP community. However, the phenomenon of hallucination in Multimodal LLMs has received comparatively less attention. LRV-Instruction [38] introduced a new instruction-tuning dataset containing 400k visual instructions to prevent hallucination in MLLMs and measured performance treating responses from GPT-4 [48] as ground truths. More recently, HallusionBench [37] quantitatively benchmarked various failure modes in MLLMs that lead to hallucinations based primarily on logical consistency and reasoning. Unlike these works that tried to benchmark MLLMs mainly on VQA-type tasks, this paper focuses on the object-level hallucination in MLLMs. + +The two closest works to our objective are POPE [35] and CHAIR [55]. On the one hand, POPE [35] tried to measure hallucination in MLLMs using a binary “Yes”-“No” answer policy in response to questions based on the absence or presence of an object in the image. On the other hand, CHAIR [55] focused on measuring hallucination in image captioning based on only words and not counts for the objects. In our work, we consider not only object words but also the corresponding count to compute an object-level count score and hallucination score. + +Semantic Object Identification +![](images/96879fa2cf51322e094aa1136a5a31d22cf2f59569536a42fc019c52e93e991e.jpg) +USER: What objects can be seen in the image? Perceive as done for semantic segmentation. + +![](images/066a464a56fe331f09262e6ea63fed3bb02b376a606cff4c5810fc0201e25758.jpg) +USER: What objects can be seen in the image? Perceive as done for instance segmentation. + +![](images/a21f3b6a4be5b31abed17f36fa7580bb491c5dfd0575e73ff072c8194d0913f4.jpg) +Instance Object Identification + +![](images/1e5db06204c7bfcd8ec256e39f554b30237b895027b5b649bd81ab2606bbb964.jpg) + +from DINov2 DPT + +from OneFormer + + from COCO dataset + +question text from prompting GPT-4 + +answer text from OneFormer output + +Figure 3. Organization of the COST dataset. We incorporate the images from COCO [36], the questions from GPT-4 [48], and the segmentation outputs from OneFormer [25] in a question-answer format for training and evaluating MLLMs on the object identification task. We also extend COST to the object order perception task by incorporating depth map outputs from DINOv2 [49] DPT [54]. COST can be extended to more object-level tasks by similarly incorporating other modalities (for example, keypoint maps). +![](images/5a6f9272957ecdc8c391e59afc6476a399d10228ec0cf0f137297cc9322b9393.jpg) +ASSISTANT: The objects present in the image are: grass, person, tree, house, car, frisbee, fence, window + +![](images/e1d2bfb79a938f31da3ce43d62fc342a520b25002afa321309f128449e25a256.jpg) +USER: What objects can be seen in the image? Perceive as done for panoptic segmentation. + +![](images/dafc120373e7522f7fda861a262d307ffbe5e49e08a2f6645ef0921ebdb298e8.jpg) +ASSISTANT: The objects present in the image are: two people, car, frisbee. +ASSISTANT: The depth order for objects present in the image is: frisbee, person, person-2, car, tree, house, grass. + +![](images/700d946e4e38e0865fdac7be190cf9e63df81028dbf61a2ee66c8a71f705564c.jpg) +ASSISTANT: The objects present in the image are: grass, tree, house, two people, frisbee, car. + +![](images/39a475e679b05d7d4bfeefdc95cd9e528e683c640ca2cf64889ab66cc65005db.jpg) +USER: Could you list the objects in the image in order of their perceived distance from the foreground to the background? + +# Panoptic Object Identification + +# Object Order Perception + +# 3. Object Identification with MLLMs + +Suppose you are invited to a Halloween party and want to bring candies for every person at that party. You ask your friend to send you a picture (Fig. 1) of the party room so that you can estimate the number of people and the number of candies you need to buy. In a hurry, you ask GPT-4V [48]: "Can you count the number of people in the image?", and it responds: "Yes, there are ten people visible in the image." Excited, you arrive at the party with ten candies but wait, you see fourteen people! Confused, you look at the image your friend sent you, and you can count fourteen people in that image, realizing that GPT-4V fails at the simple task of counting the people in the picture. At the same time, it can accurately describe the happening of a Halloween party in the image (Fig. 1). We refer to the phenomenon of Multimodal LLMs failing at simple visual perception tasks while succeeding at complex visual reasoning tasks as Moravec's Paradox [44] in perception. + +We hypothesize that one of the main reasons for the + +above phenomenon is the absence of conversations covering object identification for the salient objects and the objects in the background from the instruction-tuning data for MLLMs. To overcome this issue, we prepare the COCO Segmentation Text (COST) dataset with COCO [36] images and create sentences using the output from an image segmentation model [25] to obtain an image-text dataset to train and evaluate MLLMs for object perception MLLMs. Moreover, we also introduce a segmentation map as a control image input to the MLLM for better performance and quantify object perception performance with a count score (CS) and a hallucination score (HS). + +# 3.1. COST to Identify Objects with MLLMs + +We find that image segmentation methods [10, 25] can accurately identify salient (foreground objects like people, cars, etc.) and background objects (like sky, wall, etc.) in a given scene. Guided by this finding, we use images from the COCO [36] dataset and obtain the correspond + +![](images/19db383ea4f7748686b1f99a722a44cbac71d98a8070d6b4663bc02331396284.jpg) +Figure 4. Adapting Multimodal LLMs for accurate object perception with VCoder. (a) We add our VCoder as an adapter to the LLaVA-1.5 [39] and feed perception modalities as extra control inputs for improved object perception performance. During training, we freeze the components from LLaVA-1.5 (ImCoder, MLP, and LLM) to retain the original reasoning performance. (b) Using depth map and segmentation map as the control inputs to VCoder for the object order perception task. + +![](images/f3de7d34acd71cfecd0a37e856f4715f32d83ead567f5eb97af23dfbd466b284.jpg) + +ing segmentation outputs from OneFormer [25], a state-of-the-art image segmentation model. Next, we extract the object (class) names and counts from the segmentation outputs and convert them into a sentence form for the ground-truth answer: "The objects present in the image are: \([CNT_1][OBJ_1],[CNT_2][OBJ_2],\ldots,[CNT_N][OBJ_N]". with [OBJ_i] representing the object name and \([CNT_i]\) representing the count (if greater than one) for the \(i^{th}\) object in the image. We prompt GPT-4 [48] to collect a bucket of questions for three different object identification tasks: semantic, instance, and panoptic, corresponding to the three different image segmentation tasks. Finally, as shown in Fig. 3, we organize the images from COCO, segmentation maps from OneFormer, questions from GPT-4, and sentences containing object information into a question-answer format to construct our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object identification task. + +Statistically, we prompt GPT-4 [48] to return 20 questions for each question bucket (panoptic, semantic, and instance). In total, we used 280k images from the train2017, test2017, and unlabeled2017 splits of the COCO [36] dataset and corresponding segmentation outputs from OneFormer [25] to form the visual component of the COST training dataset. Similarly, we prepare a COST validation split using the 5k images from the val2017 split of the COCO dataset. + +Note that a similar approach can extend the COST + +dataset to other perception modalities. In this work, we incorporate the depth map modality into our COST dataset for the object order perception task. Particularly, we leverage the publicly available DINOv2 [49] DPT [54] model to obtain depth maps for COCO images and use the panoptic mask (from OneFormer [25]) to estimate the depth order of objects in an image. We format the obtained ordering of objects into the text with the template: "The depth order for objects present in the image is: $[OBJ_1], [OBJ_2], \ldots, [OBJ_J]$ ," with $[OBJ_j]$ representing the $j^{th}$ object name. To maintain relative ordering among objects belonging to the same class, we append a count number to the second and later objects, as shown in the bottom right of Fig. 3 for person and person-2. Similar to the previous setting, we prompt GPT-4 [48] to return 20 questions for the object order perception task. We provide a detailed flow of obtaining ground-truth object orders in the appendix. + +# 3.2. VCoder for Multimodal LLMs + +We notice that existing open-source Multimodal LLMs generally use the ViT [15] from CLIP [53] as the image encoder (ImCoder) during instruction tuning. We reason that the ViT focuses mainly on salient objects because it is trained against captions, which leave out information about background regions. We argue that identifying objects in the background is critical for a Multimodal LLM to become skilled at perception. To overcome this limitation, we introduce a segmentation map as a control input [45, 70] into + +our Multimodal LLM. Specifically, we use the segmentation map from OneFormer [25] and project it to the LLM's embedding space using a pretrained ViT [15] (from CLIP [53]) as a SegCoder and a two-layer MLP [39] which we collectively refer to as our Versatile enCoder (VCoder). This extra control from the segmentation map results in considerable performance gains on the object identification task. + +As shown in Fig. 4a, our VCoder adapted MLLM takes three sets of inputs: perception modalities as control inputs fed into the VCoder, an RGB image fed into an Image enCoder (and MLP), and the text question. The RGB image and text are tokenized to the and tokens, respectively. VCoder is flexible at handling various perception modalities with a unique token for each modality. For example, the segmentation map and depth map inputs are tokenized to and tokens, respectively. Similarly, one can incorporate more modalities with modality-specific tokens. Finally, all tokenized embeddings are concatenated and fed into the LLM. + +We treat our VCoder as an adapter, added to our base MLLM, LLaVA-1.5 [39] to obtain the final MLLM framework for experiments. Note that we only train the MLP components in the VCoder on the COST dataset. We decided to keep all other parameters fixed during training to keep the reasoning ability unaffected while achieving improved object perception performance. + +# 3.3. Evaluating MLLMs for Object Identification + +Despite the availability of various metrics [35, 42, 55] to measure object hallucination in vision-language models, no existing metric considers the explicit object counts while calculating their hallucination scores. We argue that object counts returned by an MLLM are a critical component that should not be overlooked while evaluating object identification performance. Therefore, we propose evaluating object identification performance in MLLMs using two metrics: count-score (CS) and hallucination-score (HS). + +$$ +G _ {\text {d i c t}} = \left\{\mathrm {O B J} _ {1} ^ {G}: \mathrm {C N T} _ {1} ^ {G}; \dots ; \mathrm {O B J} _ {N} ^ {G}: \mathrm {C N T} _ {N} ^ {G} \right\} \tag {1} +$$ + +$$ +P _ {\text {d i c t}} = \left\{\mathrm {O B J} _ {1} ^ {P}: \mathrm {C N T} _ {1} ^ {P}; \dots ; \mathrm {O B J} _ {M} ^ {P}: \mathrm {C N T} _ {M} ^ {P} \right\} +$$ + +As shown in Fig. 5, given a ground-truth sentence $(G)$ and an MLLM predicted response $(P)$ , we first extract the object words (nouns) and their corresponding count from both text samples and represent them in a dictionary form with keys as the object noun and the value as the corresponding object's count as shown in Eq. (1) with $N$ and $M$ representing the number of different object nouns in the $G$ and $P$ respectively. Next, we perform one-to-one matching between the counts for keys with $G_{\mathrm{dict}}$ and $P_{\mathrm{dict}}$ as the reference for Count Score (CS) and Hallucination Score (HS), respectively, as shown in Eq. (2). + +![](images/fa2c0010a20b8a024492580b0046c5ff61415493ffd30cc162a25359859df1a0.jpg) +Figure 5. Evaluation Metrics for Object Identification. We compare the object counts in the ground truth and prediction to calculate a count score (CS) and a hallucination score (HS). + +$$ +\mathbf {C S} = \frac {1 0 0}{N} \sum_ {i = 1} ^ {N} \left\{ \begin{array}{l l} \frac {\min \left(\mathrm {C N T} _ {i} ^ {G} , \mathrm {C N T} _ {i} ^ {P}\right)}{\max \left(\mathrm {C N T} _ {i} ^ {G} , \mathrm {C N T} _ {i} ^ {P}\right)} & \text {i f I (O B J} _ {i} ^ {G}, P _ {\text {d i c t}}) \\ 0 & \text {o t h e r w i s e} \end{array} \right. +$$ + +$$ +\mathbf {H S} = \frac {1 0 0}{M} \sum_ {j = 1} ^ {M} \left\{ \begin{array}{l l} 1 - \frac {\operatorname* {m i n} \left(\mathrm {C N T} _ {j} ^ {P} , \mathrm {C N T} _ {j} ^ {G}\right)}{\operatorname* {m a x} \left(\mathrm {C N T} _ {j} ^ {P} , \mathrm {C N T} _ {j} ^ {G}\right)} & \text {i f} I (\mathrm {O B J} _ {j} ^ {P}, G _ {\text {d i c t}}) \\ 1 & \text {o t h e r w i s e} \end{array} \right. +$$ + +$$ +I (\mathrm {O B J}, D) = \left\{ \begin{array}{l l} \text {T r u e} & \text {i f O B J i s i n k e y s} (D) \\ \text {F a l s e} & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +Count Score (CS). It represents the percentage of correct object counts predicted by the MLLM with respect to the ground-truth sentence. The higher the CS, the better. + +Hallucination Score (HS). It represents the percentage of extra object counts predicted by the MLLM that do not exist in the ground-truth sentence. The lower the HS, the better. + +Note that due to the one-to-one word-matching nature of our evaluation, we manually define a mapping between the categories in COCO [36] and their synonyms [43, 55]. For example, we replace words like man, woman, child, boy, girl, etc. with person in the prediction before evaluation. + +# 4. Experiments + +We use LLaVA-1.5 [39] as our base MLLM. LLaVA-1.5 uses CLIP-ViT-L-336px [53] as the image encoder (ImCoder) with a two-layer MLP as projection and Vicuna1.5 [73] as the LLM. Inside our VCoder, we also use a CLIP-ViT-L-336px to encode the control inputs and project the features into the LLM embedding space using modality-specific two-layer MLPs. We resize the visual inputs to $336 \times 336$ resolution (corresponds to 576 tokens) for our MLLM. During training, we load the instruction-tuned weights from LLaVA-1.5 and keep those frozen while only tuning the MLP component of our VCoder. We use the publicly available OneFormer [25] model trained on + +
MethodInput TokensSemanticInstancePanoptic
CS (↑)HS (↓)CS (↑)HS (↓)CS (↑)HS (↓)
Closed Model, Open API
GPT-4V [48]⟨img⟩ + ⟨query⟩————38.483.0
Existing Open-Source Multimodal LLMs
MiniGPT-4 LLaMA-2-7b [74]⟨img⟩ + ⟨query⟩6.292.25.697.76.294.9
InstructBLIP Vicuna-7b [14]⟨img⟩ + ⟨query⟩14.285.825.391.917.591.2
LLaVA-1.5-7b [39]⟨img⟩ + ⟨query⟩30.660.150.375.938.767.3
LLaVA-1.5-13b [39]⟨img⟩ + ⟨query⟩25.069.349.975.035.868.6
CogVLM-17b [64]⟨img⟩ + ⟨query⟩33.467.543.586.240.675.9
Baselines trained on the COST dataset
COST IT LLaVA-1.5-7b⟨img⟩ + ⟨query⟩78.722.167.530.371.928.2
Soft-Prompted LLaVA-1.5-7b⟨prompt⟩ + ⟨img⟩ + ⟨query⟩36.256.718.472.226.863.0
ImCoder LLaVA-1.5-7b⟨img⟩ + ⟨img⟩ + ⟨query⟩78.922.764.029.470.827.9
VCoder augmented LLaVA-1.5
VCoder LLaVA-1.5-7b⟨seg⟩ + ⟨img⟩ + ⟨query⟩88.610.471.126.986.012.8
VCoder LLaVA-1.5-13b⟨seg⟩ + ⟨img⟩ + ⟨query⟩89.010.073.325.087.211.6
+ +Table 1. Comparison to baseline Multimodal LLMs on the COST validation dataset for Object Identification. We compare our VCoder to existing off-the-shelf baseline MLLMs: MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64]. We also train three different variants of LLaVA-1.5 on the COST dataset: COST IT mixes the COST training data with the instruction tuning data; Soft-Prompted uses a set of learnable tokens, and ImCoder uses an RGB image as the control input. Our VCoder adapted LLaVA-1.5 performs the best on all three object perception tasks. Note: $\langle \cdot \rangle$ denotes input tokens to LLM with seg representing segmentation map, img representing RGB image, prompt representing learnable prompt, and query representing the user question. We also evaluate the performance of GPT-4V [48] on the COST dataset using the publicly accessible paid API released by OpenAI. Our VCoder-adapted LLaVA-1.5 shows the best performance on object identification among all MLLMs. + +COCO [36] with DiNAT-L [19, 20] backbone to obtain the segmentation map. For getting depth maps, we use the publicly available ViT-L/14 distilled variant of DINOv2 [49] DPT [54] trained on the NYUd [46] dataset. In this section, we discuss our results on the object identification task. + +# 4.1. Implementation Details + +Training Details. We train our VCoder-adapted LLaVA-1.5 framework for two epochs on the COST training dataset with a batch size 256 and a learning rate of $1e^{-3}$ . For other training hyperparameters, we follow the settings used during the instruction-tuning stage in LLaVA-1.5 [39]. Following [25], we uniformly sample each object identification task (semantic, instance, and panoptic) during training. We also use the corresponding segmentation map from OneFormer [25] as input to the VCoder during training and inference. On 8 A100 GPUs, it takes 8 and 14 hours to train our VCoder with the 7b and 13b variants of LLaVA-1.5 as the base MLLM, respectively. + +Evaluation Details. We evaluate all MLLMs on the COST validation set. We separately evaluate semantic, instance, and panoptic object identification tasks while randomly sampling questions from the corresponding task's question + +bucket. Note that for evaluating all off-the-shelf MLLMs, we experiment with various prompts and finally use the prompt: "[QUESTION]. Return the answer in the paragraph format: 'The objects present in the image are: ...' and then list the objects with their count in word format (if greater than 1) in front of them, like 'two people'"; where [QUESTION] is the randomly sampled question from the object identification task bucket. + +# 4.2. Main Results + +Baselines. We compare the performance of VCoder to open-source Multimodal LLMs, namely, MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64] on the COST validation set in Tab. 1. Furthermore, we provide three additional baselines, trained for two epochs: + +COST IT LLaVA-1.5: We mix the COST training data with the instruction tuning data used in LLaVA-1.5 [39] and fine-tune a LLaVA-1.5 [39] model from scratch. + +Soft-Prompted LLaVA-1.5: We pretend 576 learnable tokens $\langle \langle prompt\rangle \rangle$ to the LLM input and tune only the $\langle prompt\rangle$ parameters on the COST training dataset. + +ImCoder LLaVA-1.5: We use an RGB image as the control input and train on the COST training dataset. + +As shown in Tab. 1, we notice that all existing MLLMs + +
MethodDepth Score (↓)
LLaVA-1.5-7b [39]166.1
LLaVA-1.5-13b [39]227.2
VCoder-DS LLaVA-1.5-7b65.9
VCoder-DS LLaVA-1.5-13b63.3
+ +Table 2. Performance on Object Order Perception. Our VCoder LLaVA-1.5 considerably outperforms LLaVA-1.5 [39], owing to the usage of control inputs and training on the COST dataset. + +perform poorly on our COST validation set, demonstrating their inability to count and identify objects accurately. Note that existing MLLMs perform relatively better on instance object identification, reaffirming our claim that MLLMs are better at detecting salient objects than background objects. Although the baselines trained on the COST dataset perform relatively better, they still lag in performance compared to the VCoder. Notably, a segmentation map performs considerably better than using an RGB image as the control input, proving the segmentation map's vitality. + +Comparison to GPT-4V [48]. We utilize OpenAI's newly released gpt-4-vision-preview $^1$ API to obtain responses from GPT-4V. Our experiments show that GPT-4V's responses are consistent across all object identification tasks, closely aligning with the panoptic identification task. Therefore, we compare our VCoder to GPT-4V only on the panoptic object identification to reduce API requests due to a daily limit of 500 API requests during this project. As shown in Tab. 1, GPT-4V [48] lags behind our VCoder by a considerable margin, reaffirming our claim that existing MLLMs cannot perform accurate object-level perception. + +# 5. Object Order Perception with MLLMs + +As shown in Fig. 4, multiple perception modalities can be leveraged to improve object perception in MLLMs with our VCoder. This section presents our experiments with our VCoder using the segmentation and depth maps as the control inputs. We term the resulting MLLM as VCoder-DS LLaVA-1.5. Intuitively, predicting the object order implicitly means identifying the objects in an image. Therefore, for the object order perception task (Fig. 4b), we use both and inputs, while only the input as the additional control for object identification. + +During training, we use a mixture of datasets, including the object identification and object order perception components from the COST dataset. We also use about 200k image-conversation (along with the corresponding segmentation map obtained using OneFormer [25]) pairs randomly sampled from the instruction tuning data used in LLaVA1.5 [39]. We train our VCoder for one epoch following the same hyperparameter settings mentioned in Sec. 4. + +As shown in Tab. 2, our VCoder-DS LLaVA-1.5 significantly outperforms the base MLLM, LLaVA-1.5 [39] on + +the COST validation set. For quantitative evaluation of MLLMs on the object order perception task, we calculate a depth score (DS) using the absolute difference between the position of objects in the ground truth and prediction. + +# 6. Limitations + +Despite the improved object perception performance after training our VCoder on the COST dataset, certain limitations remain to be addressed for future work. Firstly, we build our COST dataset using OneFormer [25], which can only perceive objects belonging to a limited number of categories due to being trained on a closed-set vocabulary dataset [36]. For real-world applications, it is imperative to develop an object perception benchmark for MLLMs covering many more classes with varying granularity. Secondly, the count, hallucination, and depth scores use one-to-one word matching, which requires defining a mapping between synonymous words. It will be promising to explore ways to overcome using synonym mappings. + +# 7. Conclusion + +This work analyzes the object-level perception skills of Multimodal Large Language Models (VLMMs). Although MLLMs are good visual reasoners, they need to improve at the simple yet fundamental task of object perception. To improve object perception ability in MLLMs, we propose the COST dataset for training and evaluating MLLMs at the object perception task. We benchmark different off-the-shelf MLLMs and GPT-4V on our COST dataset and observe their lousy performance. Consequently, we propose using perception modalities as control inputs and a Versatile vision enCoders (VCoder) as an adapter for projecting the control inputs to the LLM embedding space. Our VCoder can easily be extended to leverage various modalities as the control inputs depending on the task. To quantify the object-level perception ability in MLLMs, we introduce a Count-Score (CS), a Hallucination-Score (HS), and a Depth-Score (DS). We adapted LLaVA-1.5 with VCoder, only trained the VCoder on our COST dataset, and demonstrated its improved performance at the object perception task while retaining the reasoning performance. We hope our work can inspire the research community to focus on developing object perception datasets for MLLMs and develop vision systems that are equally good at perception and reasoning in the future. + +Acknowledgements. This work was in part supported by IARPA under Contract No. 2022-21102100004, NSF CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. We also thank the ML Center at Georgia Tech for generously supporting this work. + +# References + +[1] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2015. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 3 +[3] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. Falcon-40B: an open large language model with state-of-the-art performance. arXiv, 2023. 1 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv, 2023. 1, 3 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv, 2023. 3 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2 +[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 3 +[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 3 +[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask + +transformer for universal image segmentation. In CVPR, 2022.3,4 +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 1, 3 +[12] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.3 +[13] Timothy F Cootes, Gareth J Edwards, and Christopher J Taylor. Active appearance models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681-685, 2001. 3 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 3, 7 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 5, 6 +[16] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in neural information processing systems, pages 2366-2374, 2014. 3 +[17] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv, 2023. 3 +[18] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 3 +[19] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 7 +[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 7 +[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 3 +[22] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. ArXiv, abs/2302.14045, 2023. 1, 3 +[23] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. CVPR, 2019. 3 +[24] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Ji-achen Li, Steven Walton, and Humphrey Shi. Semask: Se + +mentally masking transformer backbones for effective semantic segmentation. arXiv, 2021. 3 +[25] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8 +[26] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Yadong Mu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arXiv, 2023. 1 +[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2023. 1 +[28] Deanna Kuhn. The Skills of Argument. Cambridge University Press, 1991. 1 +[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth International Conference on 3D Vision (3DV), pages 239-248. IEEE, 2016. 3 +[30] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE. 3 +[31] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. 2023. 3 +[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 3 +[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2 +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3 +[35] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 3, 6 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5, 6, 7, 8 +[37] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3 +[38] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3 + +[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 3, 5, 6, 7, 8 +[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3 +[41] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3 +[42] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv, 2023. 6 +[43] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In CVPR, 2018. 6 +[44] H. Moravec. Mind children: The future of robot and human intelligence. Harvard University Press, 1988. 4 +[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv, 2023. 3, 5 +[46] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7 +[47] OpenAI. Chatgpt. https://chat.openai.com/, 2022.1 +[48] OpenAI. Gpt-4 technical report, 2023. 1, 2, 3, 4, 5, 7, 8 +[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4, 5, 7 +[50] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011. 2 +[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. ArXiv, abs/2306, 2023. 1 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3 +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. arXiv, 2021. 3, 5, 6 +[54] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 3, 4, 5, 7 + +[55] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In EMNLP, 2018. 3, 6 +[56] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops 2021, 2021. 2 +[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv, 2023. 1 +[58] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 3 +[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv, 2023. 3 +[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv, 2023. 1, 3 +[61] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 3 +[62] Z. Tu, Xiangrong Chen, Alan Yuille, and Song Zhu. Image parsing: Unifying segmentation, detection, and recognition. In IJCV, 2005. 3 +[63] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In CVPR, 2001. 3 +[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv, 2023. 7 +[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and ef + +ficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3 +[66] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking "text" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 3 +[67] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 3 +[68] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 3 +[69] Yan Zeng, Hanbo Zhang, Jiani Zheng, Jiangnan Xia, Guoqiang Wei, Yang Wei, Yuchen Zhang, and Tao Kong. What matters in training a gpt4-style language model with multimodal inputs? arXiv preprint arXiv:2307.02469, 2023. 3 +[70] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5 +[71] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv, 2023. 3 +[72] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv, 2023. 1 +[73] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6 +[74] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv, 2023. 1, 3, 7 +[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv, 2020. 3 \ No newline at end of file diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/images.zip b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b64a06c4480bb0ea51f9103f2a37c3fa54ec8187 --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e33e09e1c4f074f6ee2eaae876af7d1554be3f24fd86dbf81409030eb73b6706 +size 566344 diff --git a/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/layout.json b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c4119d6acf7414e291dc96ebb2be692314529406 --- /dev/null +++ b/2024/VCoder_ Versatile Vision Encoders for Multimodal Large Language Models/layout.json @@ -0,0 +1,10448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "text", + "content": "VCoder: Versatile Vision Encoders for Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "spans": [ + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "text", + "content": "Jitesh Jain" + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "text", + "content": " Jianwei Yang" + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "text", + "content": " Humphrey Shi" + }, + { + "bbox": [ + 179, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "text", + "content": "SHI Labs @ Georgia Tech " + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "text", + "content": "Microsoft Research, Redmond " + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 159, + 468, + 171 + ], + "type": "text", + "content": "Picsart AI Research (PAIR)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 174, + 387, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 174, + 387, + 184 + ], + "spans": [ + { + "bbox": [ + 201, + 174, + 387, + 184 + ], + "type": "text", + "content": "https://github.com/SHI-Labs/VCoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 525 + ], + "type": "text", + "content": "Humans possess the remarkable skill of Visual Perception, the ability to see and understand the seen, helping them make sense of the visual world and, in turn, reason. Multimodal Large Language Models (MLLM) have recently achieved impressive performance on vision-language tasks ranging from visual question-answering and image captioning to visual reasoning and image generation. However, when prompted to identify or count (perceive) the entities in a given image, existing MLLM systems fail. Working towards developing an accurate MLLM system for perception and reasoning, we propose using Versatile vision en-Coders (VCoder) as perception eyes for Multimodal LLMs. We feed the VCoder with perception modalities such as segmentation or depth maps, improving the MLLM's perception abilities. Secondly, we leverage the images from COCO and outputs from off-the-shelf vision perception models to create our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object perception task. Thirdly, we introduce metrics to assess the object perception abilities in MLLMs on our COST dataset. Lastly, we provide extensive experimental evidence proving the VCoder's improved object-level perception skills over existing Multimodal LLMs, including GPT-4V. We open-source our dataset, code, and models to promote research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 539, + 127, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 539, + 127, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 539, + 127, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 560, + 275, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 560, + 275, + 583 + ], + "spans": [ + { + "bbox": [ + 58, + 560, + 275, + 583 + ], + "type": "text", + "content": "'Perception is the soil; reasoning, the seed. Without fertile ground, the seed cannot flourish.'" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 195, + 584, + 274, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 584, + 274, + 596 + ], + "spans": [ + { + "bbox": [ + 195, + 584, + 274, + 596 + ], + "type": "text", + "content": "(GPT-4 [48], 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "The ability to think and reason is one of the most remarkable traits that help humans function daily. Generally, understanding the environment precedes the act of thinking and reasoning [28]. Following the success of ChatGPT-like instruction following AI agents [3, 11, 47, 48, 60] at language understanding and reasoning, researchers have leveraged LLMs to develop instruct frameworks [14, 40, 68, 74] that can understand vision and language inputs in an effort to imitate human perception and reasoning ability. We" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 212, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 310, + 212, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 310, + 212, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 310, + 212, + 541, + 342 + ], + "type": "image", + "image_path": "a33cf8d5cc97945d86c635dc43762b30a586590bd13807c0dc74ef2fda3a96b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 345, + 336, + 357 + ], + "blocks": [ + { + "bbox": [ + 312, + 345, + 336, + 357 + ], + "lines": [ + { + "bbox": [ + 312, + 345, + 336, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 345, + 336, + 357 + ], + "type": "image", + "image_path": "7e452c7f4f23ac429c598f27d17519c50e8933bca261b74aa876d83c01f1c714.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 358, + 406, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 358, + 406, + 370 + ], + "spans": [ + { + "bbox": [ + 328, + 358, + 406, + 370 + ], + "type": "text", + "content": "at is happening in the image?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 454, + 350, + 501, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 350, + 501, + 359 + ], + "spans": [ + { + "bbox": [ + 454, + 350, + 501, + 359 + ], + "type": "text", + "content": "Difficulty Level" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 373, + 326, + 388 + ], + "blocks": [ + { + "bbox": [ + 312, + 373, + 326, + 388 + ], + "lines": [ + { + "bbox": [ + 312, + 373, + 326, + 388 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 326, + 388 + ], + "type": "image", + "image_path": "f1575724ae16c41b02d01f25dee5ba0c162882520f5ba01dcbc765d7af2a9ec4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 377, + 354, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 377, + 354, + 387 + ], + "spans": [ + { + "bbox": [ + 328, + 377, + 354, + 387 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 392, + 517, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 392, + 517, + 420 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 517, + 420 + ], + "type": "text", + "content": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 426, + 326, + 440 + ], + "blocks": [ + { + "bbox": [ + 312, + 426, + 326, + 440 + ], + "lines": [ + { + "bbox": [ + 312, + 426, + 326, + 440 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 326, + 440 + ], + "type": "image", + "image_path": "81344a50af2718e479e7c1e048a6000cb55be66574ef99b11bf39adc76755975.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 537, + 545, + 582 + ], + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 582 + ], + "type": "text", + "content": "Figure 1. GPT-4V [48] (access date: Dec 16, 2023) returns impressive responses when prompted to describe complex visual scenes. However, it fails at the simple task of counting in the same scene. Our VCoder returns the correct count of people." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 328, + 430, + 354, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 430, + 354, + 438 + ], + "spans": [ + { + "bbox": [ + 328, + 430, + 354, + 438 + ], + "type": "text", + "content": "VCoder" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 442, + 533, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 533, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 533, + 472 + ], + "type": "text", + "content": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 312, + 479, + 336, + 492 + ], + "blocks": [ + { + "bbox": [ + 312, + 479, + 336, + 492 + ], + "lines": [ + { + "bbox": [ + 312, + 479, + 336, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 336, + 492 + ], + "type": "image", + "image_path": "7986ddffcc6976567fabd9c85ad0d55109c9d17abdc5a5bae8cbdbd9c18909f6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 494, + 382, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 382, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 382, + 510 + ], + "type": "text", + "content": "Can you count the number of people in the image?" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 329, + 515, + 375, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 515, + 375, + 524 + ], + "spans": [ + { + "bbox": [ + 329, + 515, + 375, + 524 + ], + "type": "text", + "content": "Difficulty Level" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 392, + 481, + 407, + 494 + ], + "blocks": [ + { + "bbox": [ + 392, + 481, + 407, + 494 + ], + "lines": [ + { + "bbox": [ + 392, + 481, + 407, + 494 + ], + "spans": [ + { + "bbox": [ + 392, + 481, + 407, + 494 + ], + "type": "image", + "image_path": "a465be8f016c7c2fee0dac85a91c639e77ac5f6cc354fa3880e2db2a16bf65ca.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 408, + 483, + 435, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 483, + 435, + 493 + ], + "spans": [ + { + "bbox": [ + 408, + 483, + 435, + 493 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 394, + 499, + 451, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 499, + 451, + 527 + ], + "spans": [ + { + "bbox": [ + 394, + 499, + 451, + 527 + ], + "type": "text", + "content": "Yes, there are ten people visible in the image." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 458, + 482, + 473, + 495 + ], + "blocks": [ + { + "bbox": [ + 458, + 482, + 473, + 495 + ], + "lines": [ + { + "bbox": [ + 458, + 482, + 473, + 495 + ], + "spans": [ + { + "bbox": [ + 458, + 482, + 473, + 495 + ], + "type": "image", + "image_path": "0d8d0cadd883301e5c7d9b3144bd1721d42d9a3425657cc76202caf637b6ef2d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 474, + 483, + 501, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 483, + 501, + 493 + ], + "spans": [ + { + "bbox": [ + 474, + 483, + 501, + 493 + ], + "type": "text", + "content": "VCoder" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 462, + 501, + 530, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 501, + 530, + 518 + ], + "spans": [ + { + "bbox": [ + 462, + 501, + 530, + 518 + ], + "type": "text", + "content": "Yes, there are a total of 14 people in the image." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "refer to such systems as Multimodal LLMs (MLLM). Although MLLMs exhibit the ability to perform complex vision-language tasks like visual captioning [2, 4, 39], image generation [26, 27, 57], visual reasoning and grounding [22, 51, 72], they often display sub-par performance at simple tasks like counting objects (Fig. 1). As shown in Fig. 2, MLLMs output incorrect object counts (people, chairs) and hallucinate about the presence (handbag, couch) of certain objects when prompted to identify entities in a visual input. The perception performance is much worse" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27992" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 100, + 74, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 100, + 74, + 110 + ], + "spans": [ + { + "bbox": [ + 51, + 100, + 74, + 110 + ], + "type": "text", + "content": "USER" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 114, + 160, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 114, + 160, + 123 + ], + "spans": [ + { + "bbox": [ + 54, + 114, + 160, + 123 + ], + "type": "text", + "content": "What objects can be seen in the image?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 100, + 188, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 100, + 188, + 110 + ], + "spans": [ + { + "bbox": [ + 167, + 100, + 188, + 110 + ], + "type": "text", + "content": "USER" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 114, + 276, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 114, + 276, + 122 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 276, + 122 + ], + "type": "text", + "content": "What objects can be seen in the image?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 282, + 100, + 302, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 100, + 302, + 110 + ], + "spans": [ + { + "bbox": [ + 282, + 100, + 302, + 110 + ], + "type": "text", + "content": "USER" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 284, + 114, + 390, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 114, + 390, + 122 + ], + "spans": [ + { + "bbox": [ + 284, + 114, + 390, + 122 + ], + "type": "text", + "content": "What objects can be seen in the image?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 400, + 100, + 421, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 100, + 421, + 110 + ], + "spans": [ + { + "bbox": [ + 400, + 100, + 421, + 110 + ], + "type": "text", + "content": "USER" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 402, + 114, + 488, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 114, + 488, + 122 + ], + "spans": [ + { + "bbox": [ + 402, + 114, + 488, + 122 + ], + "type": "text", + "content": "What is happening in the image?" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 126, + 164, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 126, + 164, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 126, + 164, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 126, + 164, + 201 + ], + "type": "image", + "image_path": "f64e7236a0404dc1fadf4fe07010d4e9e5a2108a61a75e673e718f9f6690caf3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 440, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 440, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 545, + 517 + ], + "type": "text", + "content": "Figure 2. MLLMs counting and identifying objects. As shown in the first column, GPT-4V [48] (access date: Dec 16, 2023) and LLaVA-1.5 [39] both fail at counting people. Moreover, LLaVA-1.5 [39] misses background entities like window, wall, etc. and hallucinates about the presence of a handbag. VCoder can predict the people counts and other background entities accurately except chairs. Similarly, in the second column, GPT-4V and LLaVA-1.5 fail at counting chairs while the VCoder matches the Oracle's performance. Notably, all MLLMs can perceive objects accurately for a non-cluttered image in the third column, with LLaVA-1.5 failing at counting ties. Our VCoder can also accurately perform general question-answering tasks, as shown in the fourth column. We treat OneFormer [25] as the Oracle for object perception. Red text represents counting mistakes; pink text represents hallucination; blue text represents correct object perception." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 166, + 126, + 278, + 201 + ], + "blocks": [ + { + "bbox": [ + 166, + 126, + 278, + 201 + ], + "lines": [ + { + "bbox": [ + 166, + 126, + 278, + 201 + ], + "spans": [ + { + "bbox": [ + 166, + 126, + 278, + 201 + ], + "type": "image", + "image_path": "e239ad6519c95bb5f6688e29f8e8aa194e00a04c870ee0001134d8c7b3383520.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 280, + 126, + 392, + 211 + ], + "blocks": [ + { + "bbox": [ + 280, + 126, + 392, + 211 + ], + "lines": [ + { + "bbox": [ + 280, + 126, + 392, + 211 + ], + "spans": [ + { + "bbox": [ + 280, + 126, + 392, + 211 + ], + "type": "image", + "image_path": "fd6c711c84170ee68b6f4a19c7e1db48db20e232e1ea5c28cf980aadf48814c7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 398, + 126, + 542, + 208 + ], + "blocks": [ + { + "bbox": [ + 398, + 126, + 542, + 208 + ], + "lines": [ + { + "bbox": [ + 398, + 126, + 542, + 208 + ], + "spans": [ + { + "bbox": [ + 398, + 126, + 542, + 208 + ], + "type": "image", + "image_path": "2e63190aef579ad86efb7a24543562ac404b65f177bc38939acb5afb9030d0ff.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 51, + 205, + 64, + 215 + ], + "blocks": [ + { + "bbox": [ + 51, + 205, + 64, + 215 + ], + "lines": [ + { + "bbox": [ + 51, + 205, + 64, + 215 + ], + "spans": [ + { + "bbox": [ + 51, + 205, + 64, + 215 + ], + "type": "image", + "image_path": "c614f92a5cf193e12ec5f9dbbb292af9e0a2c0657bc7537ffcbad6306ab6e6de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 64, + 207, + 89, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 207, + 89, + 215 + ], + "spans": [ + { + "bbox": [ + 64, + 207, + 89, + 215 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 52, + 220, + 162, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 220, + 162, + 264 + ], + "spans": [ + { + "bbox": [ + 52, + 220, + 162, + 264 + ], + "type": "text", + "content": "The objects present in the image are: eleven people, one large potted plant, two ceiling lamps, one exit sign, and various pieces of furniture that appear to be office desks and chairs." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 165, + 205, + 188, + 215 + ], + "blocks": [ + { + "bbox": [ + 165, + 205, + 188, + 215 + ], + "lines": [ + { + "bbox": [ + 165, + 205, + 188, + 215 + ], + "spans": [ + { + "bbox": [ + 165, + 205, + 188, + 215 + ], + "type": "image", + "image_path": "b65306bffef51c66058ccede5b60e585c2c534a8d4ac2cc63e00ae3ad33b2016.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 189, + 207, + 204, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 207, + 204, + 215 + ], + "spans": [ + { + "bbox": [ + 189, + 207, + 204, + 215 + ], + "type": "text", + "content": "T-4V" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 167, + 221, + 276, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 221, + 276, + 268 + ], + "spans": [ + { + "bbox": [ + 167, + 221, + 276, + 268 + ], + "type": "text", + "content": "The objects present in the image are: two Corgi dogs, two umbrellas, multiple chairs, two speakers, and one palm tree. The setting is an outdoor area with a beachfront view." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 216, + 318, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 216, + 318, + 224 + ], + "spans": [ + { + "bbox": [ + 294, + 216, + 318, + 224 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 282, + 228, + 389, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 228, + 389, + 268 + ], + "spans": [ + { + "bbox": [ + 282, + 228, + 389, + 268 + ], + "type": "text", + "content": "The objects present in the image are: two people (1 male and 1 female), a table, windows, and buildings seen through the windows." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 411, + 212, + 436, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 212, + 436, + 220 + ], + "spans": [ + { + "bbox": [ + 411, + 212, + 436, + 220 + ], + "type": "text", + "content": "GPT-4V" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 400, + 224, + 537, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 224, + 537, + 272 + ], + "spans": [ + { + "bbox": [ + 400, + 224, + 537, + 272 + ], + "type": "text", + "content": "Characters from \"Friends\" are gathered in a living room, wearing various Halloween costumes, possibly celebrating or attending a Halloween party. Chandler stands out in a pink bunny costume." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 51, + 271, + 64, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 271, + 64, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 271, + 64, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 271, + 64, + 281 + ], + "type": "image", + "image_path": "7552bb6475bf959d9117a786e069a22478c1f3ab3f8e31956a9a693e1676967d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 64, + 272, + 96, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 272, + 96, + 280 + ], + "spans": [ + { + "bbox": [ + 64, + 272, + 96, + 280 + ], + "type": "text", + "content": "LLaVA-1.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 52, + 285, + 157, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 285, + 157, + 314 + ], + "spans": [ + { + "bbox": [ + 52, + 285, + 157, + 314 + ], + "type": "text", + "content": "The objects present in the image are: 12 people, 1 potted plant, 1 TV, 2 chairs, 1 handbag." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 165, + 277, + 189, + 286 + ], + "blocks": [ + { + "bbox": [ + 165, + 277, + 189, + 286 + ], + "lines": [ + { + "bbox": [ + 165, + 277, + 189, + 286 + ], + "spans": [ + { + "bbox": [ + 165, + 277, + 189, + 286 + ], + "type": "image", + "image_path": "45973119da94a73ddf1a687610e895582eca4e3f4fce7bb4baaba606c1cdfc17.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 189, + 277, + 211, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 277, + 211, + 285 + ], + "spans": [ + { + "bbox": [ + 189, + 277, + 211, + 285 + ], + "type": "text", + "content": "vA-1.5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 167, + 291, + 276, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 291, + 276, + 319 + ], + "spans": [ + { + "bbox": [ + 167, + 291, + 276, + 319 + ], + "type": "text", + "content": "The objects present in the image are: 2 dogs, 3 couches, 3 umbrellas, 1 table, and 2 chairs." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 274, + 324, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 274, + 324, + 284 + ], + "spans": [ + { + "bbox": [ + 295, + 274, + 324, + 284 + ], + "type": "text", + "content": "LaVA-1.5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 282, + 289, + 391, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 289, + 391, + 318 + ], + "spans": [ + { + "bbox": [ + 282, + 289, + 391, + 318 + ], + "type": "text", + "content": "The objects present in the image are: man, woman, ties (2), suit (1), glass wall (1), building (1), desk (1)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 411, + 280, + 444, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 280, + 444, + 288 + ], + "spans": [ + { + "bbox": [ + 411, + 280, + 444, + 288 + ], + "type": "text", + "content": "LLaVA-1.5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 400, + 293, + 538, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 293, + 538, + 323 + ], + "spans": [ + { + "bbox": [ + 400, + 293, + 538, + 323 + ], + "type": "text", + "content": "In the image, a group of people is gathered in a room, sitting around a table. They are dressed in costumes, possibly for a party or a themed event." + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 51, + 320, + 63, + 331 + ], + "blocks": [ + { + "bbox": [ + 51, + 320, + 63, + 331 + ], + "lines": [ + { + "bbox": [ + 51, + 320, + 63, + 331 + ], + "spans": [ + { + "bbox": [ + 51, + 320, + 63, + 331 + ], + "type": "image", + "image_path": "e80386a4a303de7610ddb8bb5e3b51ee4b58f1973e284548425be081b994e518.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 64, + 321, + 88, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 321, + 88, + 330 + ], + "spans": [ + { + "bbox": [ + 64, + 321, + 88, + 330 + ], + "type": "text", + "content": "VCoder" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 53, + 335, + 159, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 159, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 159, + 373 + ], + "type": "text", + "content": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, three chairs." + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 165, + 326, + 188, + 336 + ], + "blocks": [ + { + "bbox": [ + 165, + 326, + 188, + 336 + ], + "lines": [ + { + "bbox": [ + 165, + 326, + 188, + 336 + ], + "spans": [ + { + "bbox": [ + 165, + 326, + 188, + 336 + ], + "type": "image", + "image_path": "2e62166f332d6bf694d9ccb74d469d6c899c75c7f02b6fa208f5a556d777aeeb.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 189, + 328, + 202, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 328, + 202, + 336 + ], + "spans": [ + { + "bbox": [ + 189, + 328, + 202, + 336 + ], + "type": "text", + "content": "oder" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 167, + 340, + 274, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 340, + 274, + 380 + ], + "spans": [ + { + "bbox": [ + 167, + 340, + 274, + 380 + ], + "type": "text", + "content": "The objects present in the image are: sky, tree, pillow, mountain, water, grass, bed, two dogs, umbrella, potted plant, dining table, six chairs." + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 295, + 324, + 317, + 334 + ], + "blocks": [ + { + "bbox": [ + 295, + 324, + 317, + 334 + ], + "lines": [ + { + "bbox": [ + 295, + 324, + 317, + 334 + ], + "spans": [ + { + "bbox": [ + 295, + 324, + 317, + 334 + ], + "type": "image", + "image_path": "d2131fbc61eef778ea71423e0776188fcb7f442143baa7784752142b9b6d90bc.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 282, + 338, + 388, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 338, + 388, + 357 + ], + "spans": [ + { + "bbox": [ + 282, + 338, + 388, + 357 + ], + "type": "text", + "content": "The objects present in the image are: window, wall, two people, tie." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 411, + 328, + 435, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 328, + 435, + 338 + ], + "spans": [ + { + "bbox": [ + 411, + 328, + 435, + 338 + ], + "type": "text", + "content": "VCoder" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 400, + 341, + 536, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 341, + 536, + 392 + ], + "spans": [ + { + "bbox": [ + 400, + 341, + 536, + 392 + ], + "type": "text", + "content": "The image shows a group of people gathered in what appears to be a living room or a dining area, dressed in various silly, quirky costumes. They are sitting at a dining table with chairs, engaged in a social event or party." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 53, + 380, + 115, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 115, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 115, + 389 + ], + "type": "text", + "content": "OneFormer (Oracle)" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 53, + 392, + 159, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 159, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 159, + 430 + ], + "type": "text", + "content": "The objects present in the image are: ceiling, floor, wall, window, cabinet, door, light, table, rug, thirteen people, potted plant, chair, TV." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 168, + 387, + 230, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 387, + 230, + 395 + ], + "spans": [ + { + "bbox": [ + 168, + 387, + 230, + 395 + ], + "type": "text", + "content": "OneFormer (Oracle)" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 167, + 400, + 391, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 400, + 391, + 429 + ], + "spans": [ + { + "bbox": [ + 167, + 400, + 391, + 429 + ], + "type": "text", + "content": "The objects present in the image are: sky, tree, sea, pillow, wall, bed, two dogs, dining table, six chairs, grass, umbrella, potted plant, mountain." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 402, + 399, + 497, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 399, + 497, + 408 + ], + "spans": [ + { + "bbox": [ + 402, + 399, + 497, + 408 + ], + "type": "text", + "content": "Text in red denotes wrong count." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 402, + 410, + 500, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 410, + 500, + 418 + ], + "spans": [ + { + "bbox": [ + 402, + 410, + 500, + 418 + ], + "type": "text", + "content": "Text in pink denotes hallucination." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 402, + 419, + 527, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 419, + 527, + 427 + ], + "spans": [ + { + "bbox": [ + 402, + 419, + 527, + 427 + ], + "type": "text", + "content": "Text in blue denotes correct object counts." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 568 + ], + "type": "text", + "content": "when the scenes are cluttered with many entities. Consequently, a natural question arises: \"How to develop MLLM systems that respond to perception questions accurately?\"" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "This work aims to improve Multimodal LLMs at the simple yet fundamental object-level perception skills, including counting. Our motivation stems from the intuition that one can only describe and reason about a visual scene with the correct understanding of the entities in the image. In our effort to develop an accurate Multimodal LLM perception system, we face three significant challenges: (i) the scarcity of a vision-language dataset focused on the object perception task; (ii) existing open-sourced Multimodal LLMs usually use the ViT from CLIP [52] with an RGB image as input as the visual component that majorly focuses only on salient objects, and (iii) the absence of evaluation metrics to" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 533, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 568 + ], + "type": "text", + "content": "quantitatively measure Multimodal LLMs' object perception and in particular, counting skills. We list our efforts to overcome the issues above in the following paragraphs." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "The contemporary vision-language models [14, 34, 52] owe their success to the availability of large-scale image-text datasets [7, 50, 56]. However, these datasets are more focused on image captioning [33] and VQA [1] tasks, making them unfit for training Multimodal LLMs for basic perception skills like object identification and counting. To overcome the scarcity of fundamental perception-focused image-text data, we leverage images from the COCO [36] dataset and use predictions from off-the-shelf visual perception models [25, 49, 54] to prepare a COCO Segmentation Text (COST) dataset comprising of question-answer pairs about the objects (background and foreground) present in" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 178, + 74, + 265, + 87 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 74, + 265, + 87 + ], + "spans": [ + { + "bbox": [ + 178, + 74, + 265, + 87 + ], + "type": "text", + "content": "Object Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 403, + 74, + 537, + 87 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 74, + 537, + 87 + ], + "spans": [ + { + "bbox": [ + 403, + 74, + 537, + 87 + ], + "type": "text", + "content": "General Question-Answering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27993" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 245, + 83 + ], + "type": "text", + "content": "each image. We provide more details in Sec. 3.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 228 + ], + "type": "text", + "content": "Inspired by diffusion models that add various perception \"control\" or \"context\" images [45, 66, 67, 70] as auxiliary inputs to aid image generation, we propose feeding extra perception modalities as control inputs through additional vision encoders, which we term as our Versatile vision encoders (VCoder). In this work, we focus on the task of object perception and leverage a segmentation map, depth map, or both as the control inputs; however, the same design can be extended to other modalities. Our VCoder projects the control inputs' information into the LLM's space as shown in Fig. 4. We hypothesize that this added control helps the MLLM improve its object perception ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 286, + 324 + ], + "type": "text", + "content": "Lastly, owing to the absence of metrics to quantify the counting ability in MLLMs, we propose computing a count score (CS) using one-to-one matching of object words in the ground truth and MLLM's answer. We also compute a hallucination score (HS) based on the extra objects in the MLLM's response that are absent from the ground truth. Similarly, we introduce a depth score (DS) to quantify the object order prediction performance in MLLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 326, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 286, + 384 + ], + "type": "text", + "content": "Among the open-source MLLMs, we choose LLaVA-1.5 [39] as our base MLLM. Our extensive experimental analysis demonstrates the importance of our COST dataset and VCoder LLaVA-1.5's improved perception ability. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 386, + 286, + 540 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 386, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 386, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 58, + 386, + 286, + 422 + ], + "type": "text", + "content": "- We propose using extra (perception) control inputs and feeding those to a Versatile enCoder (VCoder) for improved object perception performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 422, + 286, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 422, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 58, + 422, + 286, + 469 + ], + "type": "text", + "content": "- We introduce a COCO Segmentation Text (COST) dataset to train and evaluate Multimodal LLM systems on the fundamental object-level perception tasks of object identification, counting, and order prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 470, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 470, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 58, + 470, + 286, + 540 + ], + "type": "text", + "content": "- Furthermore, to quantify the object perception ability in MLLMs, we propose calculating a count score (CS), a hallucination score (HS) and a depth score (DS). Our experiments show that the VCoder-adapted LLaVA1.5 outperforms the baseline MLLMs on all metrics when validated on the COST dataset." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 554, + 133, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 133, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 133, + 566 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 151, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 151, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 151, + 586 + ], + "type": "text", + "content": "2.1. Visual Perception" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "The fundamental nature of visual perception makes it a critical component in MLLM systems. Perception can be divided into sub-tasks, including dense prediction tasks like image segmentation [24, 41, 62] and depth estimation [16, 18, 54], and sparse prediction tasks like object detection [6, 63] and pose estimation [13, 58]. In the DL era, initial methods tackled the perception task using CNN based methods [8, 9, 21, 29, 30, 58] with recent methods shifting to the use of transformer-based architectures [10, 18, 25, 54, 65, 75]. In this work, we tackle the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "task of object-level perception, mainly focusing on predicting names, counts, and order of objects using MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 104, + 486, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 104, + 486, + 117 + ], + "spans": [ + { + "bbox": [ + 305, + 104, + 486, + 117 + ], + "type": "text", + "content": "2.2. Visual Understanding with LLMs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 123, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 182 + ], + "type": "text", + "content": "Using LLMs for vision applications is not a new concept. In a nutshell, developing Multimodal LLMs involves projecting [2, 4, 34, 61] the features from a vision encoder [15, 53] to the embedding space of a language model (LLM) [11, 59, 60], and, visual instruction-tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 183, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 399 + ], + "type": "text", + "content": "LLaVA [40] proposed a pipeline to convert existing image-text data into dialog format and then finetuned a CLIP [52] and LLaMA [59] model end-to-end on their collected dataset showing one of the earliest evidence of visual-language instruction tuning. Concurrent to LLaVA, MiniGPT-4 [74] used the visual encoder from BLIP2 [34] and used a linear layer for projecting visual features into Vicuna's [11] feature space. InstructBLIP [14] open-sourced a collection of 16 different datasets covering various vision tasks like VQA, reasoning, captioning, classification, etc., and finetuned a BLIP2 model on their dataset. mPLUGOwl [68] proposed using a vision abstractor and finetuning the vision encoder. More recently, LLaVA-1.5 [39] proposed using an MLP as the projector and finetuned on academic instruction datasets to achieve state-of-the-art performance on various benchmarks [12, 17, 23, 35]. Among various open-source MLLMs [5, 22, 31, 32, 69], we chose LLaVA-1.5 as our baseline due to its superior performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 407, + 499, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 499, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 499, + 419 + ], + "type": "text", + "content": "2.3. Perception Hallucination in MLLMs" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 593 + ], + "type": "text", + "content": "Since the introduction of LLMs, there has been a comprehensive study about their ability to hallucinate [71] in the NLP community. However, the phenomenon of hallucination in Multimodal LLMs has received comparatively less attention. LRV-Instruction [38] introduced a new instruction-tuning dataset containing 400k visual instructions to prevent hallucination in MLLMs and measured performance treating responses from GPT-4 [48] as ground truths. More recently, HallusionBench [37] quantitatively benchmarked various failure modes in MLLMs that lead to hallucinations based primarily on logical consistency and reasoning. Unlike these works that tried to benchmark MLLMs mainly on VQA-type tasks, this paper focuses on the object-level hallucination in MLLMs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "The two closest works to our objective are POPE [35] and CHAIR [55]. On the one hand, POPE [35] tried to measure hallucination in MLLMs using a binary “Yes”-“No” answer policy in response to questions based on the absence or presence of an object in the image. On the other hand, CHAIR [55] focused on measuring hallucination in image captioning based on only words and not counts for the objects. In our work, we consider not only object words but also the corresponding count to compute an object-level count score and hallucination score." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27994" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 87, + 145, + 174 + ], + "blocks": [ + { + "bbox": [ + 86, + 75, + 205, + 85 + ], + "lines": [ + { + "bbox": [ + 86, + 75, + 205, + 85 + ], + "spans": [ + { + "bbox": [ + 86, + 75, + 205, + 85 + ], + "type": "text", + "content": "Semantic Object Identification" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 87, + 145, + 174 + ], + "lines": [ + { + "bbox": [ + 50, + 87, + 145, + 174 + ], + "spans": [ + { + "bbox": [ + 50, + 87, + 145, + 174 + ], + "type": "image", + "image_path": "96879fa2cf51322e094aa1136a5a31d22cf2f59569536a42fc019c52e93e991e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 182, + 233, + 203 + ], + "lines": [ + { + "bbox": [ + 51, + 182, + 233, + 203 + ], + "spans": [ + { + "bbox": [ + 51, + 182, + 233, + 203 + ], + "type": "text", + "content": "USER: What objects can be seen in the image? Perceive as done for semantic segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 147, + 87, + 242, + 174 + ], + "blocks": [ + { + "bbox": [ + 147, + 87, + 242, + 174 + ], + "lines": [ + { + "bbox": [ + 147, + 87, + 242, + 174 + ], + "spans": [ + { + "bbox": [ + 147, + 87, + 242, + 174 + ], + "type": "image", + "image_path": "066a464a56fe331f09262e6ea63fed3bb02b376a606cff4c5810fc0201e25758.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 182, + 439, + 203 + ], + "lines": [ + { + "bbox": [ + 257, + 182, + 439, + 203 + ], + "spans": [ + { + "bbox": [ + 257, + 182, + 439, + 203 + ], + "type": "text", + "content": "USER: What objects can be seen in the image? Perceive as done for instance segmentation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 87, + 348, + 174 + ], + "blocks": [ + { + "bbox": [ + 290, + 75, + 406, + 85 + ], + "lines": [ + { + "bbox": [ + 290, + 75, + 406, + 85 + ], + "spans": [ + { + "bbox": [ + 290, + 75, + 406, + 85 + ], + "type": "text", + "content": "Instance Object Identification" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 87, + 348, + 174 + ], + "lines": [ + { + "bbox": [ + 252, + 87, + 348, + 174 + ], + "spans": [ + { + "bbox": [ + 252, + 87, + 348, + 174 + ], + "type": "image", + "image_path": "a21f3b6a4be5b31abed17f36fa7580bb491c5dfd0575e73ff072c8194d0913f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 350, + 87, + 445, + 174 + ], + "blocks": [ + { + "bbox": [ + 350, + 87, + 445, + 174 + ], + "lines": [ + { + "bbox": [ + 350, + 87, + 445, + 174 + ], + "spans": [ + { + "bbox": [ + 350, + 87, + 445, + 174 + ], + "type": "image", + "image_path": "1e5db06204c7bfcd8ec256e39f554b30237b895027b5b649bd81ab2606bbb964.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 471, + 83, + 520, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 83, + 520, + 104 + ], + "spans": [ + { + "bbox": [ + 471, + 83, + 520, + 104 + ], + "type": "text", + "content": "from DINov2 DPT" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 476, + 117, + 515, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 117, + 515, + 137 + ], + "spans": [ + { + "bbox": [ + 476, + 117, + 515, + 137 + ], + "type": "text", + "content": "from OneFormer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 471, + 151, + 520, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 151, + 520, + 170 + ], + "spans": [ + { + "bbox": [ + 471, + 151, + 520, + 170 + ], + "type": "text", + "content": " from COCO dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 462, + 184, + 529, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 184, + 529, + 205 + ], + "spans": [ + { + "bbox": [ + 462, + 184, + 529, + 205 + ], + "type": "text", + "content": "question text from prompting GPT-4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 463, + 217, + 528, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 463, + 217, + 528, + 237 + ], + "spans": [ + { + "bbox": [ + 463, + 217, + 528, + 237 + ], + "type": "text", + "content": "answer text from OneFormer output" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 50, + 248, + 145, + 335 + ], + "blocks": [ + { + "bbox": [ + 52, + 215, + 236, + 237 + ], + "lines": [ + { + "bbox": [ + 52, + 215, + 236, + 237 + ], + "spans": [ + { + "bbox": [ + 52, + 215, + 236, + 237 + ], + "type": "text", + "content": "ASSISTANT: The objects present in the image are: grass, person, tree, house, car, frisbee, fence, window" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 50, + 248, + 145, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 248, + 145, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 248, + 145, + 335 + ], + "type": "image", + "image_path": "5a6f9272957ecdc8c391e59afc6476a399d10228ec0cf0f137297cc9322b9393.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 422, + 547, + 465 + ], + "lines": [ + { + "bbox": [ + 45, + 422, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 547, + 465 + ], + "type": "text", + "content": "Figure 3. Organization of the COST dataset. We incorporate the images from COCO [36], the questions from GPT-4 [48], and the segmentation outputs from OneFormer [25] in a question-answer format for training and evaluating MLLMs on the object identification task. We also extend COST to the object order perception task by incorporating depth map outputs from DINOv2 [49] DPT [54]. COST can be extended to more object-level tasks by similarly incorporating other modalities (for example, keypoint maps)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 147, + 248, + 243, + 335 + ], + "blocks": [ + { + "bbox": [ + 147, + 248, + 243, + 335 + ], + "lines": [ + { + "bbox": [ + 147, + 248, + 243, + 335 + ], + "spans": [ + { + "bbox": [ + 147, + 248, + 243, + 335 + ], + "type": "image", + "image_path": "e1d2bfb79a938f31da3ce43d62fc342a520b25002afa321309f128449e25a256.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 342, + 233, + 364 + ], + "lines": [ + { + "bbox": [ + 52, + 342, + 233, + 364 + ], + "spans": [ + { + "bbox": [ + 52, + 342, + 233, + 364 + ], + "type": "text", + "content": "USER: What objects can be seen in the image? Perceive as done for panoptic segmentation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 252, + 248, + 348, + 335 + ], + "blocks": [ + { + "bbox": [ + 255, + 216, + 430, + 237 + ], + "lines": [ + { + "bbox": [ + 255, + 216, + 430, + 237 + ], + "spans": [ + { + "bbox": [ + 255, + 216, + 430, + 237 + ], + "type": "text", + "content": "ASSISTANT: The objects present in the image are: two people, car, frisbee." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 252, + 248, + 348, + 335 + ], + "lines": [ + { + "bbox": [ + 252, + 248, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 252, + 248, + 348, + 335 + ], + "type": "image", + "image_path": "dafc120373e7522f7fda861a262d307ffbe5e49e08a2f6645ef0921ebdb298e8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 376, + 525, + 399 + ], + "lines": [ + { + "bbox": [ + 254, + 376, + 525, + 399 + ], + "spans": [ + { + "bbox": [ + 254, + 376, + 525, + 399 + ], + "type": "text", + "content": "ASSISTANT: The depth order for objects present in the image is: frisbee, person, person-2, car, tree, house, grass." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 350, + 248, + 445, + 335 + ], + "blocks": [ + { + "bbox": [ + 52, + 376, + 238, + 399 + ], + "lines": [ + { + "bbox": [ + 52, + 376, + 238, + 399 + ], + "spans": [ + { + "bbox": [ + 52, + 376, + 238, + 399 + ], + "type": "text", + "content": "ASSISTANT: The objects present in the image are: grass, tree, house, two people, frisbee, car." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 350, + 248, + 445, + 335 + ], + "lines": [ + { + "bbox": [ + 350, + 248, + 445, + 335 + ], + "spans": [ + { + "bbox": [ + 350, + 248, + 445, + 335 + ], + "type": "image", + "image_path": "700d946e4e38e0865fdac7be190cf9e63df81028dbf61a2ee66c8a71f705564c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 447, + 248, + 543, + 335 + ], + "blocks": [ + { + "bbox": [ + 447, + 248, + 543, + 335 + ], + "lines": [ + { + "bbox": [ + 447, + 248, + 543, + 335 + ], + "spans": [ + { + "bbox": [ + 447, + 248, + 543, + 335 + ], + "type": "image", + "image_path": "39a475e679b05d7d4bfeefdc95cd9e528e683c640ca2cf64889ab66cc65005db.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 342, + 527, + 365 + ], + "lines": [ + { + "bbox": [ + 254, + 342, + 527, + 365 + ], + "spans": [ + { + "bbox": [ + 254, + 342, + 527, + 365 + ], + "type": "text", + "content": "USER: Could you list the objects in the image in order of their perceived distance from the foreground to the background?" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 24 + }, + { + "bbox": [ + 86, + 405, + 204, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 405, + 204, + 416 + ], + "spans": [ + { + "bbox": [ + 86, + 405, + 204, + 416 + ], + "type": "text", + "content": "Panoptic Object Identification" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 347, + 405, + 444, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 405, + 444, + 417 + ], + "spans": [ + { + "bbox": [ + 347, + 405, + 444, + 417 + ], + "type": "text", + "content": "Object Order Perception" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 475, + 240, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 240, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 240, + 487 + ], + "type": "text", + "content": "3. Object Identification with MLLMs" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 45, + 496, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 496, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 496, + 287, + 700 + ], + "type": "text", + "content": "Suppose you are invited to a Halloween party and want to bring candies for every person at that party. You ask your friend to send you a picture (Fig. 1) of the party room so that you can estimate the number of people and the number of candies you need to buy. In a hurry, you ask GPT-4V [48]: \"Can you count the number of people in the image?\", and it responds: \"Yes, there are ten people visible in the image.\" Excited, you arrive at the party with ten candies but wait, you see fourteen people! Confused, you look at the image your friend sent you, and you can count fourteen people in that image, realizing that GPT-4V fails at the simple task of counting the people in the picture. At the same time, it can accurately describe the happening of a Halloween party in the image (Fig. 1). We refer to the phenomenon of Multimodal LLMs failing at simple visual perception tasks while succeeding at complex visual reasoning tasks as Moravec's Paradox [44] in perception." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We hypothesize that one of the main reasons for the" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 475, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 619 + ], + "type": "text", + "content": "above phenomenon is the absence of conversations covering object identification for the salient objects and the objects in the background from the instruction-tuning data for MLLMs. To overcome this issue, we prepare the COCO Segmentation Text (COST) dataset with COCO [36] images and create sentences using the output from an image segmentation model [25] to obtain an image-text dataset to train and evaluate MLLMs for object perception MLLMs. Moreover, we also introduce a segmentation map as a control image input to the MLLM for better performance and quantify object perception performance with a count score (CS) and a hallucination score (HS)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 632, + 515, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 515, + 645 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 515, + 645 + ], + "type": "text", + "content": "3.1. COST to Identify Objects with MLLMs" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We find that image segmentation methods [10, 25] can accurately identify salient (foreground objects like people, cars, etc.) and background objects (like sky, wall, etc.) in a given scene. Guided by this finding, we use images from the COCO [36] dataset and obtain the correspond" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27995" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 399, + 316 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 399, + 316 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 399, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 399, + 316 + ], + "type": "image", + "image_path": "19db383ea4f7748686b1f99a722a44cbac71d98a8070d6b4663bc02331396284.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 546, + 366 + ], + "type": "text", + "content": "Figure 4. Adapting Multimodal LLMs for accurate object perception with VCoder. (a) We add our VCoder as an adapter to the LLaVA-1.5 [39] and feed perception modalities as extra control inputs for improved object perception performance. During training, we freeze the components from LLaVA-1.5 (ImCoder, MLP, and LLM) to retain the original reasoning performance. (b) Using depth map and segmentation map as the control inputs to VCoder for the object order perception task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 547, + 315 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 547, + 315 + ], + "type": "image", + "image_path": "f3de7d34acd71cfecd0a37e856f4715f32d83ead567f5eb97af23dfbd466b284.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 374, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 289, + 589 + ], + "type": "text", + "content": "ing segmentation outputs from OneFormer [25], a state-of-the-art image segmentation model. Next, we extract the object (class) names and counts from the segmentation outputs and convert them into a sentence form for the ground-truth answer: \"The objects present in the image are: \\([CNT_1][OBJ_1],[CNT_2][OBJ_2],\\ldots,[CNT_N][OBJ_N]\". with [OBJ_i] representing the object name and \\([CNT_i]\\) representing the count (if greater than one) for the \\(i^{th}\\) object in the image. We prompt GPT-4 [48] to collect a bucket of questions for three different object identification tasks: semantic, instance, and panoptic, corresponding to the three different image segmentation tasks. Finally, as shown in Fig. 3, we organize the images from COCO, segmentation maps from OneFormer, questions from GPT-4, and sentences containing object information into a question-answer format to construct our COCO Segmentation Text (COST) dataset for training and evaluating MLLMs on the object identification task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "type": "text", + "content": "Statistically, we prompt GPT-4 [48] to return 20 questions for each question bucket (panoptic, semantic, and instance). In total, we used 280k images from the train2017, test2017, and unlabeled2017 splits of the COCO [36] dataset and corresponding segmentation outputs from OneFormer [25] to form the visual component of the COST training dataset. Similarly, we prepare a COST validation split using the 5k images from the val2017 split of the COCO dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Note that a similar approach can extend the COST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "text", + "content": "dataset to other perception modalities. In this work, we incorporate the depth map modality into our COST dataset for the object order perception task. Particularly, we leverage the publicly available DINOv2 [49] DPT [54] model to obtain depth maps for COCO images and use the panoptic mask (from OneFormer [25]) to estimate the depth order of objects in an image. We format the obtained ordering of objects into the text with the template: \"The depth order for objects present in the image is: " + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "inline_equation", + "content": "[OBJ_1], [OBJ_2], \\ldots, [OBJ_J]" + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "text", + "content": ",\" with " + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "inline_equation", + "content": "[OBJ_j]" + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "text", + "content": " representing the " + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 374, + 547, + 579 + ], + "type": "text", + "content": " object name. To maintain relative ordering among objects belonging to the same class, we append a count number to the second and later objects, as shown in the bottom right of Fig. 3 for person and person-2. Similar to the previous setting, we prompt GPT-4 [48] to return 20 questions for the object order perception task. We provide a detailed flow of obtaining ground-truth object orders in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "type": "text", + "content": "3.2. VCoder for Multimodal LLMs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We notice that existing open-source Multimodal LLMs generally use the ViT [15] from CLIP [53] as the image encoder (ImCoder) during instruction tuning. We reason that the ViT focuses mainly on salient objects because it is trained against captions, which leave out information about background regions. We argue that identifying objects in the background is critical for a Multimodal LLM to become skilled at perception. To overcome this limitation, we introduce a segmentation map as a control input [45, 70] into" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27996" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "our Multimodal LLM. Specifically, we use the segmentation map from OneFormer [25] and project it to the LLM's embedding space using a pretrained ViT [15] (from CLIP [53]) as a SegCoder and a two-layer MLP [39] which we collectively refer to as our Versatile enCoder (VCoder). This extra control from the segmentation map results in considerable performance gains on the object identification task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 300 + ], + "type": "text", + "content": "As shown in Fig. 4a, our VCoder adapted MLLM takes three sets of inputs: perception modalities as control inputs fed into the VCoder, an RGB image fed into an Image enCoder (and MLP), and the text question. The RGB image and text are tokenized to the and tokens, respectively. VCoder is flexible at handling various perception modalities with a unique token for each modality. For example, the segmentation map and depth map inputs are tokenized to and tokens, respectively. Similarly, one can incorporate more modalities with modality-specific tokens. Finally, all tokenized embeddings are concatenated and fed into the LLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 387 + ], + "type": "text", + "content": "We treat our VCoder as an adapter, added to our base MLLM, LLaVA-1.5 [39] to obtain the final MLLM framework for experiments. Note that we only train the MLP components in the VCoder on the COST dataset. We decided to keep all other parameters fixed during training to keep the reasoning ability unaffected while achieving improved object perception performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 279, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 279, + 412 + ], + "type": "text", + "content": "3.3. Evaluating MLLMs for Object Identification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 419, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 528 + ], + "type": "text", + "content": "Despite the availability of various metrics [35, 42, 55] to measure object hallucination in vision-language models, no existing metric considers the explicit object counts while calculating their hallucination scores. We argue that object counts returned by an MLLM are a critical component that should not be overlooked while evaluating object identification performance. Therefore, we propose evaluating object identification performance in MLLMs using two metrics: count-score (CS) and hallucination-score (HS)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 536, + 287, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 536, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 70, + 536, + 287, + 557 + ], + "type": "interline_equation", + "content": "G _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {G}: \\mathrm {C N T} _ {1} ^ {G}; \\dots ; \\mathrm {O B J} _ {N} ^ {G}: \\mathrm {C N T} _ {N} ^ {G} \\right\\} \\tag {1}", + "image_path": "57f7d5778d542a4da22fddd9bcbef080289bb36aa83b4a30efc4c24cceebe236.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 554, + 263, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 554, + 263, + 568 + ], + "spans": [ + { + "bbox": [ + 74, + 554, + 263, + 568 + ], + "type": "interline_equation", + "content": "P _ {\\text {d i c t}} = \\left\\{\\mathrm {O B J} _ {1} ^ {P}: \\mathrm {C N T} _ {1} ^ {P}; \\dots ; \\mathrm {O B J} _ {M} ^ {P}: \\mathrm {C N T} _ {M} ^ {P} \\right\\}", + "image_path": "9f8f1b575f4bfb423918fd2ddcf94fc60ed70be3cceb279ea05fc52137441fb2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "As shown in Fig. 5, given a ground-truth sentence " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "(G)" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " and an MLLM predicted response " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "(P)" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ", we first extract the object words (nouns) and their corresponding count from both text samples and represent them in a dictionary form with keys as the object noun and the value as the corresponding object's count as shown in Eq. (1) with " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " representing the number of different object nouns in the " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " respectively. Next, we perform one-to-one matching between the counts for keys with " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{dict}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{dict}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " as the reference for Count Score (CS) and Hallucination Score (HS), respectively, as shown in Eq. (2)." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 69, + 533, + 224 + ], + "blocks": [ + { + "bbox": [ + 320, + 69, + 533, + 224 + ], + "lines": [ + { + "bbox": [ + 320, + 69, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 320, + 69, + 533, + 224 + ], + "type": "image", + "image_path": "fa2c0010a20b8a024492580b0046c5ff61415493ffd30cc162a25359859df1a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 262 + ], + "type": "text", + "content": "Figure 5. Evaluation Metrics for Object Identification. We compare the object counts in the ground truth and prediction to calculate a count score (CS) and a hallucination score (HS)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 280, + 523, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 523, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 523, + 314 + ], + "type": "interline_equation", + "content": "\\mathbf {C S} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\left\\{ \\begin{array}{l l} \\frac {\\min \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)}{\\max \\left(\\mathrm {C N T} _ {i} ^ {G} , \\mathrm {C N T} _ {i} ^ {P}\\right)} & \\text {i f I (O B J} _ {i} ^ {G}, P _ {\\text {d i c t}}) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "5a60371252f6b3cc107a24b26845173bf4e3574fa79466d91ca3c8db4e219b62.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 317, + 543, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 543, + 355 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 543, + 355 + ], + "type": "interline_equation", + "content": "\\mathbf {H S} = \\frac {1 0 0}{M} \\sum_ {j = 1} ^ {M} \\left\\{ \\begin{array}{l l} 1 - \\frac {\\operatorname* {m i n} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)}{\\operatorname* {m a x} \\left(\\mathrm {C N T} _ {j} ^ {P} , \\mathrm {C N T} _ {j} ^ {G}\\right)} & \\text {i f} I (\\mathrm {O B J} _ {j} ^ {P}, G _ {\\text {d i c t}}) \\\\ 1 & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "0a615dd2435941c537cfa2bd072072696597efbbcc73d3de476c89c4ecf3b94c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "type": "interline_equation", + "content": "I (\\mathrm {O B J}, D) = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f O B J i s i n k e y s} (D) \\\\ \\text {F a l s e} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "b7efde675e57cb4781cae7bac778e8ad372cd24d7766c0e2d94c289a8fa49ada.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 408, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 443 + ], + "type": "text", + "content": "Count Score (CS). It represents the percentage of correct object counts predicted by the MLLM with respect to the ground-truth sentence. The higher the CS, the better." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "type": "text", + "content": "Hallucination Score (HS). It represents the percentage of extra object counts predicted by the MLLM that do not exist in the ground-truth sentence. The lower the HS, the better." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 480, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 480, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 480, + 545, + 540 + ], + "type": "text", + "content": "Note that due to the one-to-one word-matching nature of our evaluation, we manually define a mapping between the categories in COCO [36] and their synonyms [43, 55]. For example, we replace words like man, woman, child, boy, girl, etc. with person in the prediction before evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 550, + 387, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 387, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 387, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We use LLaVA-1.5 [39] as our base MLLM. LLaVA-1.5 uses CLIP-ViT-L-336px [53] as the image encoder (ImCoder) with a two-layer MLP as projection and Vicuna1.5 [73] as the LLM. Inside our VCoder, we also use a CLIP-ViT-L-336px to encode the control inputs and project the features into the LLM embedding space using modality-specific two-layer MLPs. We resize the visual inputs to " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " resolution (corresponds to 576 tokens) for our MLLM. During training, we load the instruction-tuned weights from LLaVA-1.5 and keep those frozen while only tuning the MLP component of our VCoder. We use the publicly available OneFormer [25] model trained on" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27997" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 68, + 545, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 319 + ], + "type": "table", + "html": "
MethodInput TokensSemanticInstancePanoptic
CS (↑)HS (↓)CS (↑)HS (↓)CS (↑)HS (↓)
Closed Model, Open API
GPT-4V [48]⟨img⟩ + ⟨query⟩————38.483.0
Existing Open-Source Multimodal LLMs
MiniGPT-4 LLaMA-2-7b [74]⟨img⟩ + ⟨query⟩6.292.25.697.76.294.9
InstructBLIP Vicuna-7b [14]⟨img⟩ + ⟨query⟩14.285.825.391.917.591.2
LLaVA-1.5-7b [39]⟨img⟩ + ⟨query⟩30.660.150.375.938.767.3
LLaVA-1.5-13b [39]⟨img⟩ + ⟨query⟩25.069.349.975.035.868.6
CogVLM-17b [64]⟨img⟩ + ⟨query⟩33.467.543.586.240.675.9
Baselines trained on the COST dataset
COST IT LLaVA-1.5-7b⟨img⟩ + ⟨query⟩78.722.167.530.371.928.2
Soft-Prompted LLaVA-1.5-7b⟨prompt⟩ + ⟨img⟩ + ⟨query⟩36.256.718.472.226.863.0
ImCoder LLaVA-1.5-7b⟨img⟩ + ⟨img⟩ + ⟨query⟩78.922.764.029.470.827.9
VCoder augmented LLaVA-1.5
VCoder LLaVA-1.5-7b⟨seg⟩ + ⟨img⟩ + ⟨query⟩88.610.471.126.986.012.8
VCoder LLaVA-1.5-13b⟨seg⟩ + ⟨img⟩ + ⟨query⟩89.010.073.325.087.211.6
", + "image_path": "51490ee5a2756050485e12e3b8680ca6dd3d16eeeab8bd5623b1577aa0f03950.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 547, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 547, + 410 + ], + "type": "text", + "content": "Table 1. Comparison to baseline Multimodal LLMs on the COST validation dataset for Object Identification. We compare our VCoder to existing off-the-shelf baseline MLLMs: MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64]. We also train three different variants of LLaVA-1.5 on the COST dataset: COST IT mixes the COST training data with the instruction tuning data; Soft-Prompted uses a set of learnable tokens, and ImCoder uses an RGB image as the control input. Our VCoder adapted LLaVA-1.5 performs the best on all three object perception tasks. Note: " + }, + { + "bbox": [ + 46, + 322, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 46, + 322, + 547, + 410 + ], + "type": "text", + "content": " denotes input tokens to LLM with seg representing segmentation map, img representing RGB image, prompt representing learnable prompt, and query representing the user question. We also evaluate the performance of GPT-4V [48] on the COST dataset using the publicly accessible paid API released by OpenAI. Our VCoder-adapted LLaVA-1.5 shows the best performance on object identification among all MLLMs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 425, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 485 + ], + "type": "text", + "content": "COCO [36] with DiNAT-L [19, 20] backbone to obtain the segmentation map. For getting depth maps, we use the publicly available ViT-L/14 distilled variant of DINOv2 [49] DPT [54] trained on the NYUd [46] dataset. In this section, we discuss our results on the object identification task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 495, + 180, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 180, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 180, + 509 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 662 + ], + "type": "text", + "content": "Training Details. We train our VCoder-adapted LLaVA-1.5 framework for two epochs on the COST training dataset with a batch size 256 and a learning rate of " + }, + { + "bbox": [ + 46, + 517, + 287, + 662 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 517, + 287, + 662 + ], + "type": "text", + "content": ". For other training hyperparameters, we follow the settings used during the instruction-tuning stage in LLaVA-1.5 [39]. Following [25], we uniformly sample each object identification task (semantic, instance, and panoptic) during training. We also use the corresponding segmentation map from OneFormer [25] as input to the VCoder during training and inference. On 8 A100 GPUs, it takes 8 and 14 hours to train our VCoder with the 7b and 13b variants of LLaVA-1.5 as the base MLLM, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Evaluation Details. We evaluate all MLLMs on the COST validation set. We separately evaluate semantic, instance, and panoptic object identification tasks while randomly sampling questions from the corresponding task's question" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 425, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 521 + ], + "type": "text", + "content": "bucket. Note that for evaluating all off-the-shelf MLLMs, we experiment with various prompts and finally use the prompt: \"[QUESTION]. Return the answer in the paragraph format: 'The objects present in the image are: ...' and then list the objects with their count in word format (if greater than 1) in front of them, like 'two people'\"; where [QUESTION] is the randomly sampled question from the object identification task bucket." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 391, + 541 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "Baselines. We compare the performance of VCoder to open-source Multimodal LLMs, namely, MiniGPT-4 [74], InstructBLIP [14], LLaVA-1.5 [39], and CogVLM [64] on the COST validation set in Tab. 1. Furthermore, we provide three additional baselines, trained for two epochs:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "text", + "content": "COST IT LLaVA-1.5: We mix the COST training data with the instruction tuning data used in LLaVA-1.5 [39] and fine-tune a LLaVA-1.5 [39] model from scratch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "content": "Soft-Prompted LLaVA-1.5: We pretend 576 learnable tokens " + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\langle \\langle prompt\\rangle \\rangle" + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "content": " to the LLM input and tune only the " + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\langle prompt\\rangle" + }, + { + "bbox": [ + 304, + 641, + 545, + 677 + ], + "type": "text", + "content": " parameters on the COST training dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 700 + ], + "type": "text", + "content": "ImCoder LLaVA-1.5: We use an RGB image as the control input and train on the COST training dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1, we notice that all existing MLLMs" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27998" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 248, + 129 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 248, + 129 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 248, + 129 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 248, + 129 + ], + "type": "table", + "html": "
MethodDepth Score (↓)
LLaVA-1.5-7b [39]166.1
LLaVA-1.5-13b [39]227.2
VCoder-DS LLaVA-1.5-7b65.9
VCoder-DS LLaVA-1.5-13b63.3
", + "image_path": "067ae9138f730bd434f24050a087b69e70726aade0eef943f45eb09f13cfc147.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 166 + ], + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 166 + ], + "type": "text", + "content": "Table 2. Performance on Object Order Perception. Our VCoder LLaVA-1.5 considerably outperforms LLaVA-1.5 [39], owing to the usage of control inputs and training on the COST dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 173, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 287, + 292 + ], + "type": "text", + "content": "perform poorly on our COST validation set, demonstrating their inability to count and identify objects accurately. Note that existing MLLMs perform relatively better on instance object identification, reaffirming our claim that MLLMs are better at detecting salient objects than background objects. Although the baselines trained on the COST dataset perform relatively better, they still lag in performance compared to the VCoder. Notably, a segmentation map performs considerably better than using an RGB image as the control input, proving the segmentation map's vitality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 424 + ], + "type": "text", + "content": "Comparison to GPT-4V [48]. We utilize OpenAI's newly released gpt-4-vision-preview" + }, + { + "bbox": [ + 46, + 293, + 287, + 424 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 46, + 293, + 287, + 424 + ], + "type": "text", + "content": " API to obtain responses from GPT-4V. Our experiments show that GPT-4V's responses are consistent across all object identification tasks, closely aligning with the panoptic identification task. Therefore, we compare our VCoder to GPT-4V only on the panoptic object identification to reduce API requests due to a daily limit of 500 API requests during this project. As shown in Tab. 1, GPT-4V [48] lags behind our VCoder by a considerable margin, reaffirming our claim that existing MLLMs cannot perform accurate object-level perception." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 435, + 261, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 261, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 261, + 449 + ], + "type": "text", + "content": "5. Object Order Perception with MLLMs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": "As shown in Fig. 4, multiple perception modalities can be leveraged to improve object perception in MLLMs with our VCoder. This section presents our experiments with our VCoder using the segmentation and depth maps as the control inputs. We term the resulting MLLM as VCoder-DS LLaVA-1.5. Intuitively, predicting the object order implicitly means identifying the objects in an image. Therefore, for the object order perception task (Fig. 4b), we use both and inputs, while only the input as the additional control for object identification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 575, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 670 + ], + "type": "text", + "content": "During training, we use a mixture of datasets, including the object identification and object order perception components from the COST dataset. We also use about 200k image-conversation (along with the corresponding segmentation map obtained using OneFormer [25]) pairs randomly sampled from the instruction tuning data used in LLaVA1.5 [39]. We train our VCoder for one epoch following the same hyperparameter settings mentioned in Sec. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 671, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 287, + 695 + ], + "type": "text", + "content": "As shown in Tab. 2, our VCoder-DS LLaVA-1.5 significantly outperforms the base MLLM, LLaVA-1.5 [39] on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "the COST validation set. For quantitative evaluation of MLLMs on the object order perception task, we calculate a depth score (DS) using the absolute difference between the position of objects in the ground truth and prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 130, + 380, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 130, + 380, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 380, + 143 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 151, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 307 + ], + "type": "text", + "content": "Despite the improved object perception performance after training our VCoder on the COST dataset, certain limitations remain to be addressed for future work. Firstly, we build our COST dataset using OneFormer [25], which can only perceive objects belonging to a limited number of categories due to being trained on a closed-set vocabulary dataset [36]. For real-world applications, it is imperative to develop an object perception benchmark for MLLMs covering many more classes with varying granularity. Secondly, the count, hallucination, and depth scores use one-to-one word matching, which requires defining a mapping between synonymous words. It will be promising to explore ways to overcome using synonym mappings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 316, + 378, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 378, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 378, + 329 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 624 + ], + "type": "text", + "content": "This work analyzes the object-level perception skills of Multimodal Large Language Models (VLMMs). Although MLLMs are good visual reasoners, they need to improve at the simple yet fundamental task of object perception. To improve object perception ability in MLLMs, we propose the COST dataset for training and evaluating MLLMs at the object perception task. We benchmark different off-the-shelf MLLMs and GPT-4V on our COST dataset and observe their lousy performance. Consequently, we propose using perception modalities as control inputs and a Versatile vision enCoders (VCoder) as an adapter for projecting the control inputs to the LLM embedding space. Our VCoder can easily be extended to leverage various modalities as the control inputs depending on the task. To quantify the object-level perception ability in MLLMs, we introduce a Count-Score (CS), a Hallucination-Score (HS), and a Depth-Score (DS). We adapted LLaVA-1.5 with VCoder, only trained the VCoder on our COST dataset, and demonstrated its improved performance at the object perception task while retaining the reasoning performance. We hope our work can inspire the research community to focus on developing object perception datasets for MLLMs and develop vision systems that are equally good at perception and reasoning in the future." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work was in part supported by IARPA under Contract No. 2022-21102100004, NSF CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. We also thank the ML Center at Georgia Tech for generously supporting this work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 285, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 285, + 712 + ], + "type": "text", + "content": "1https://platform.openai.com/docs/guides/vision" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27999" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 235 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 237, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 237, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 237, + 288, + 303 + ], + "type": "text", + "content": "[3] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. Falcon-40B: an open large language model with state-of-the-art performance. arXiv, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 306, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 381 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 384, + 288, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 516 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 518, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 518, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 518, + 287, + 551 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 553, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 553, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 53, + 553, + 287, + 596 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 599, + 287, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 599, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 53, + 599, + 287, + 642 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 645, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 645, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 645, + 287, + 689 + ], + "type": "text", + "content": "[9] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "transformer for universal image segmentation. In CVPR, 2022.3,4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[12] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[13] Timothy F Cootes, Gareth J Edwards, and Christopher J Taylor. Active appearance models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681-685, 2001. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[16] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Advances in neural information processing systems, pages 2366-2374, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "text", + "content": "[17] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 465, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 498 + ], + "type": "text", + "content": "[18] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 499, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 520 + ], + "type": "text", + "content": "[19] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv:2209.15001, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "text", + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 580, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 655 + ], + "type": "text", + "content": "[22] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. ArXiv, abs/2302.14045, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[23] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. CVPR, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[24] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Ji-achen Li, Steven Walton, and Humphrey Shi. Semask: Se" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28000" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "mentally masking transformer backbones for effective semantic segmentation. arXiv, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[25] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "text", + "content": "[26] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Yadong Mu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arXiv, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "text", + "content": "[27] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 245 + ], + "type": "text", + "content": "[28] Deanna Kuhn. The Skills of Argument. Cambridge University Press, 1991. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 287, + 300 + ], + "type": "text", + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth International Conference on 3D Vision (3DV), pages 239-248. IEEE, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "type": "text", + "content": "[30] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "text", + "content": "[31] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning. 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "type": "text", + "content": "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "text", + "content": "[35] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "text", + "content": "[37] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[38] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 132, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 163 + ], + "type": "text", + "content": "[41] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "type": "text", + "content": "[42] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 233 + ], + "type": "text", + "content": "[43] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In CVPR, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 236, + 545, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 257 + ], + "type": "text", + "content": "[44] H. Moravec. Mind children: The future of robot and human intelligence. Harvard University Press, 1988. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "type": "text", + "content": "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "text", + "content": "[46] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 339, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 339, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 545, + 360 + ], + "type": "text", + "content": "[47] OpenAI. Chatgpt. https://chat.openai.com/, 2022.1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 363, + 529, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 529, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 529, + 374 + ], + "type": "text", + "content": "[48] OpenAI. Gpt-4 technical report, 2023. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 376, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 474 + ], + "type": "text", + "content": "[49] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "text", + "content": "[50] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "text", + "content": "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. ArXiv, abs/2306, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. arXiv, 2021. 3, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[54] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28001" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "text", + "content": "[55] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In EMNLP, 2018. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops 2021, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[57] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 250 + ], + "type": "text", + "content": "[58] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 316 + ], + "type": "text", + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 546 + ], + "type": "text", + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "text", + "content": "[61] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "text", + "content": "[62] Z. Tu, Xiangrong Chen, Alan Yuille, and Song Zhu. Image parsing: Unifying segmentation, detection, and recognition. In IJCV, 2005. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 635 + ], + "type": "text", + "content": "[63] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In CVPR, 2001. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. arXiv, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and ef" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 543 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "text", + "content": "[66] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking \"text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[67] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "type": "text", + "content": "[68] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[69] Yan Zeng, Hanbo Zhang, Jiani Zheng, Jiangnan Xia, Guoqiang Wei, Yang Wei, Yuchen Zhang, and Tao Kong. What matters in training a gpt4-style language model with multimodal inputs? arXiv preprint arXiv:2307.02469, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[70] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "text", + "content": "[71] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 408 + ], + "type": "text", + "content": "[72] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "text", + "content": "[73] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 465, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 509 + ], + "type": "text", + "content": "[74] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv, 2023. 1, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv, 2020. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28002" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_content_list.json b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40eac41141af332b8c153e40673e828e51227f92 --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_content_list.json @@ -0,0 +1,1864 @@ +[ + { + "type": "text", + "text": "VGGSfM: Visual Geometry Grounded Deep Structure From Motion", + "text_level": 1, + "bbox": [ + 138, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianyuan Wang $^{1,2}$", + "bbox": [ + 129, + 179, + 272, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikita Karaev 1,2", + "bbox": [ + 315, + 180, + 450, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Rupprecht1", + "bbox": [ + 493, + 180, + 661, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Novotny2", + "bbox": [ + 705, + 180, + 836, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Geometry Group, University of Oxford", + "bbox": [ + 204, + 215, + 578, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta AI", + "bbox": [ + 681, + 215, + 759, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://vggsfm.github.io", + "bbox": [ + 362, + 250, + 604, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 312, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structure-from-motion (SfM) is a long-standing problem in the computer vision community, which aims to reconstruct the camera poses and 3D structure of a scene from a set of unconstrained 2D images. Classical frameworks solve this problem in an incremental manner by detecting and matching keypoints, registering images, triangulating 3D points, and conducting bundle adjustment. Recent research efforts have predominantly revolved around harnessing the power of deep learning techniques to enhance specific elements (e.g., keypoint matching), but are still based on the original, non-differentiable pipeline. Instead, we propose a new deep pipeline VGGsfM, where each component is fully differentiable and thus can be trained in an end-to-end manner. To this end, we introduce new mechanisms and simplifications. First, we build on recent advances in deep 2D point tracking to extract reliable pixel-accurate tracks, which eliminates the need for chaining pairwise matches. Furthermore, we recover all cameras simultaneously based on the image and track features instead of gradually registering cameras. Finally, we optimise the cameras and triangulate 3D points via a differentiable bundle adjustment layer. We attain state-of-the-art performance on three popular datasets, CO3D, IMC Phototourism, and ETH3D.", + "bbox": [ + 76, + 319, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 694, + 207, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing the camera parameters and the 3D structure of a scene from a set of unconstrained 2D images is a long-standing problem in the computer vision community. Among many other applications [8, 27, 29, 53, 79], it has recently emerged as an important component of learning neural fields [10, 28, 34, 37, 48, 76]. The problem is usually solved via the Structure-from-Motion (SfM) framework which estimates the 3D point cloud (Structure) and the parameters of each camera (Motion) in the scene. State-of-the-art methods [25, 39] follow the incremental SfM paradigm whose origins can be traced back to the early 2000s [23, 58]. It usually begins with a small set of", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/442374b47f7f05fbfd05936a49654d12e904f4edefe9b5e38edb91c9d73bc601.jpg", + "image_caption": [ + "Figure 1. Reconstruction of In-the-wild Photos with VGGsFm, displaying estimated point clouds (in blue) and cameras (orange)." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 691, + 753 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/936d5f7f73bd041253a5fa48db6febd10820525639041254582ea1446a59bce0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 287, + 888, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correspondence-rich images as initialization, and gradually adds more views into the reconstruction, through keypoint detection, matching, verification, image registration, triangulation, bundle adjustment (BA), and so on [2, 20, 26, 59].", + "bbox": [ + 496, + 808, + 890, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research efforts have predominantly revolved around leveraging the power of deep learning techniques to", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21686", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae5f31ae48b8f47b1f507ddcca7d928cd84069cdca49fbdcc7f7e8df73fbb707.jpg", + "image_caption": [ + "Figure 2. Overview of VGGs fM. Our method extracts 2D tracks from input images, reconstructs cameras using image and track features, initializes a point cloud based on these tracks and camera parameters, and applies a bundle adjustment layer for reconstruction refinement. The whole framework is fully differentiable and designed for end-to-end training." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 893, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhance specific elements within the original pipeline while preserving the incremental SfM framework as a whole. For instance, SuperPoint and SuperGlue [14, 57] focus on improving keypoint detection and matching. Pixel-perfect SfM [39] proposes deep feature-metric refinement to adjust both keypoints and bundles. Detector-free feature matching methods [12, 64, 75] bypass early keypoint detection by means of attention, which is powerful in poorly textured scenes. Detector-free SfM [25] builds a coarse SfM model through quantized detector-free matches and then iteratively refines it with multi-view consistency constraints. These advancements successfully combine deep learning approaches (such as deep feature matching) with well-established hand-engineered components, such as the incremental camera registration of COLMAP [59].", + "bbox": [ + 75, + 354, + 472, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The widespread success of end-to-end training warrants the question of what benefits it can bring to long-standing frameworks such as SfM. Naturally, it is often difficult to assess the merits of new approaches when compared with decades of continuous improvements. Nonetheless, in this paper, we answer this question by introducing a fully-differentiable SfM pipeline, dubbed Visual Geometry Grounded Deep Structure From Motion (VGGSfM), which trains in an end-to-end manner. We find that this allows the pipeline to be simpler than prior frameworks while achieving better or comparable performance. Training end-to-end allows each component to generate outputs that facilitate the task of its successor.", + "bbox": [ + 75, + 582, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To build a fully-differentiable pipeline, we make several substantial changes to the SfM procedures and overall obtain better performance. Specifically, our model builds on recent advances in deep 2D point tracking [15, 16, 22, 32] to directly extract reliable pixel-accurate tracks. This simplifies the correspondence estimation step in traditional SfM, which first estimates pairwise matches and then connects them into tracks. Then, based on the image and track fea", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures, VGGsFm estimates all cameras jointly via a Transformer [72], and subsequently all 3D points. Different from Incremental SfM, this approach is simpler and easier to differentiate as it does not depend on a discrete, combinatorial correspondence chaining step. Finally, for bundle adjustment, we replace the commonly employed non-differentiable Ceres solver [3] with the fully differentiable second-order Theseus solver [54].", + "bbox": [ + 496, + 354, + 890, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, we fuse all the SfM components into a single fully differentiable reconstruction function $f$ . Besides that, in our experiments, we also show that the individual modules perform well in isolation. Ultimately, end-to-end training yields another performance improvement, that surpasses the performance of isolated components.", + "bbox": [ + 496, + 476, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate VGGsFm for the task of camera pose estimation on the CO3Dv2 [55] and IMC Phototourism [31] datasets, and for 3D triangulation on the ETH3D [60] dataset. Our method attains strong performance on all benchmarks. At the same time, we conduct in-the-wild reconstruction to validate the generalization ability of our proposed framework, as shown in Fig. 1.", + "bbox": [ + 496, + 568, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 691, + 640, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structure from Motion is a fundamental problem in computer vision and has been investigated for decades [23, 52, 53]. The classical pipelines usually solve the SfM problem in a global [13, 50, 80] or incremental [2, 19, 59, 63, 81] manner. Both of which are usually based on pairwise image keypoint matching. Incremental SfM is arguably the most widely adopted strategy (e.g., the popular framework COLMAP [59]). Therefore, in the following sections, we refer to incremental SfM as \"classical\" or \"traditional\" SfM. We defer the discussion of global SfM to the supplementary.", + "bbox": [ + 496, + 717, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional SfM frameworks often start by detecting keypoints and feature descriptors [4, 42, 43, 46]. They then", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21687", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "search for image pairs with overlapping frusta by matching these keypoints across different images (e.g., with a nearest-neighbour search) [2, 41, 59]. These image pairs are further verified via two-view epipolar geometry or homography [23] through RANSAC [18]. Then, a pair or a small set of images is carefully selected for initialization. New images are gradually registered by solving the Perspective- $n$ -Point (PnP) problem [44], followed by triangulating 3D points, and bundle adjustment [69]. This process is iterated until all the frames are either registered or discarded. 2D correspondences (multi-view tracks) are the basis of the whole process, however, they are usually simply constructed by chaining two-view matches [59].", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many deep-learning approaches have been proposed to enhance this framework. For example, [14, 70, 84] provide better keypoint detection and [11, 30, 40, 57, 61] focus on matching. Furthermore, detector-free matching methods [12, 64, 75] propose to avoid sparse keypoint detection by building semi-dense matches via self and cross attention. [7] solves the subproblems of SfM by graph attention networks. Some studies improve the performance of RANSAC by making it trainable [5, 6, 77]. Recent state-of-the-art methods are PixSfM [39] and the concurrent Detector-free SfM (DFSfM) [25]. PixSfM refines the tracks and structure estimated by COLMAP through feature-metric keypoint adjustment and feature-metric bundle adjustment. Detector-free SfM proposes to first build a coarse SfM model using detector-free matches and COLMAP (or other frameworks), and then to iteratively refine the tracks and the structure of the coarse model by enforcing multi-view consistency.", + "bbox": [ + 75, + 289, + 472, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, fully differentiable SfM pipelines have also been explored. They usually use deep neural networks to regress camera poses and depths [65, 66, 68, 71, 73, 78, 87]. Although using an approximation of bundle adjustment [65, 68, 78], these methods suffer from limited generalizability and scalability (very few input frames) [65, 71, 73, 78, 87], or rely on temporal relationship [66, 68]. Meanwhile, some methods are category-specific [45, 82, 83]. The recent efforts on deep camera pose estimation can scale up to more than 50 frames, but they do not reconstruct the scene [36, 62, 74, 85].", + "bbox": [ + 75, + 547, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point tracking. Since VGGSfM proposes a novel point track predictor, next, we review recent advances in this field. Inspired by the optical-flow architecture of RAFT [67], PIPs [22] revisited point tracking, a task related to Particle Video [56], and proposed a highly accurate tracker of isolated points in a video. TAP-Vid [15] (i.e., \"Tracking Any Point\") introduced a benchmark for point tracking and a baseline model, which was later improved in TAPIR [16] by integrating the iterative update mechanism from PIPs. PointOdyssey [86] simplified PIPs and proposed a benchmark for the long-term version of point tracking. CoTracker [32] closed the gap between single point track", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing and dense Optical Flow with joint point tracking. However, these works are designed for videos, i.e. temporally-ordered sequences of frames. In our point tracker, given the input frames are unordered, we do not assume a temporal relationship between input frames. We therefore process all frames jointly, avoiding windowed inference of [32]. Since SfM relies on highly accurate correspondences, our tracks are further refined in a coarse-to-fine manner to achieve subpixel accuracy.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 243, + 591, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the components of VG-GSfM and how they are composed in a fully differentiable pipeline. An overview of our framework is shown in Fig. 2.", + "bbox": [ + 496, + 268, + 890, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setting Given a set of free-form images observing a scene, VGGSfM estimates their corresponding camera parameters and the 3D scene shape represented as a point cloud. Formally, given a tuple $\\mathcal{I} = (I_1,\\dots,I_{N_I})$ of $N_{I}\\in \\mathbb{N}$ RGB images $I_{i}\\in \\mathbb{R}^{3\\times H\\times W}$ , VGGSfM estimates the corresponding camera projection matrices $\\mathcal{P} = (P_{1},\\ldots ,P_{N_{I}}|P_{i}\\subset \\mathbb{R}^{3\\times 4})$ and the scene cloud $X = \\{\\mathbf{x}^j\\}_{j = 1}^{N_{\\mathbf{x}}} \\text{ of } N_{\\mathbf{x}}\\in \\mathbb{N}$ 3D points $\\mathbf{x}^j\\in \\mathbb{R}^3$ . Each projection matrix $P_{i}$ consists of extrinsic (pose) $g_{i}\\in \\mathbb{SE}(3)$ and intrinsics $K_{i}\\in \\mathbb{R}^{3\\times 3}$ .", + "bbox": [ + 496, + 318, + 893, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A 3D point $\\mathbf{x}^j$ can be projected to the $i$ -th camera yielding a 2D screen coordinate $\\mathbf{y}_i^j = P_i(\\mathbf{x}^j) \\sim \\lambda K_i \\hat{\\mathbf{x}}_i^j$ ; $\\lambda \\in \\mathbb{R}_+$ , where $\\hat{\\mathbf{x}}_i^j = g_i \\mathbf{x}^j$ is the world coordinate $\\mathbf{x}^j$ expressed in view-coordinates of the $i$ -th camera. The projection of the point $\\mathbf{x}^j$ to all input cameras is a track $T^j = ((y_1^j, v_1^j) \\dots, (y_{N_I}^j, v_{N_I}^j))$ consisting of $N_I$ matching 2D points $\\mathbf{y}_i^j \\in \\mathbb{R}^2$ , and their corresponding binary indicators $v_i^j \\in \\{0, 1\\}$ denoting visibility of the $j$ -th point in the $i$ -th camera. We denote $\\mathcal{T}_i = \\{T_i^1, \\dots, T_i^{N_T}\\}$ as the set of all tracks $T_i^j$ in the $i$ -th camera.", + "bbox": [ + 496, + 472, + 893, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Method overview", + "text_level": 1, + "bbox": [ + 498, + 643, + 669, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VGGsM implements SfM via a single function $f_{\\theta}$", + "bbox": [ + 500, + 667, + 839, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\mathcal {I}) = \\mathcal {P}, X \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 694, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "accepting the set of $N_I$ scene images $\\mathcal{I}$ and outputting the camera parameters $\\mathcal{P}$ and the scene point cloud $X$ . Importantly, $f_{\\theta}$ is fully differentiable and, as such, its parameters $\\theta$ are learned by minimizing the training loss $\\mathcal{L}$ :", + "bbox": [ + 496, + 720, + 890, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {\\star} = \\arg \\min _ {\\theta} \\sum_ {s = 1} ^ {S} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathcal {I} _ {s}\\right), \\mathcal {P} _ {s} ^ {\\star}, \\mathcal {T} _ {s} ^ {\\star}, X _ {s} ^ {\\star}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 789, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "summing over $S \\in \\mathbb{N}$ training image sets $\\mathcal{I}_s$ annotated with ground-truth cameras $\\mathcal{P}_s^\\star$ , tracks $\\mathcal{T}_s^\\star$ , and point clouds $X_s^\\star$ . We defer the details of $\\mathcal{L}$ to Sec. 3.5 and, in the following paragraphs, discuss the architecture of $f_{\\theta}$ .", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21688", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reconstruction function Following traditional SfM [59], VGGSfM decomposes the reconstruction function $f_{\\theta}$ into four seamless stages: 1) point tracking $\\mathbb{T}$ , 2) initial camera estimator $\\mathfrak{T}_{\\mathcal{P}}$ , 3) triangulator $\\mathfrak{T}_X$ and, 4) Bundle Adjustment BA, as follows:", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} = \\mathrm {T} (\\mathcal {I})\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 175, + 289, + 191 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} (\\mathcal {I}, \\mathcal {T}) \\\\ \\hat {\\hat {\\tau}} = \\mathfrak {T} _ {\\hat {\\tau}} (\\mathcal {\\hat {T}}, \\hat {\\hat {\\tau}}) \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 195, + 468, + 222 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\mathfrak {T} _ {X} (\\mathcal {T}, \\hat {\\mathcal {P}})\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 214, + 325, + 234 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P}, X = _ {\\mathrm {B A}} (\\mathcal {T}, \\hat {\\mathcal {P}}, \\hat {X}).\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 238, + 349, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The track predictor $\\mathbb{T}$ , hereafter referred to as \"tracker\", estimates 2D tracks $\\mathcal{T}$ given input images $\\mathcal{I}$ . Subsequently, $\\mathfrak{T}_{\\mathcal{P}}$ and $\\mathfrak{T}_X$ provide initial cameras $\\mathcal{P}$ and an initial point cloud $\\hat{X}$ respectively. Finally, BA enhances accuracy by refining the cameras and 3D points together.", + "bbox": [ + 76, + 261, + 468, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tracking", + "text_level": 1, + "bbox": [ + 76, + 345, + 184, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Establishing precise 2D correspondences is important for achieving accurate 3D reconstruction. Traditionally, SfM frameworks first estimate pairwise image-to-image correspondences that are later chained into multi-image tracks $T$ [25, 39, 59]. Here, typically only point-pair matching benefits from learned components [14, 40, 57, 70], while the chaining of pairwise correspondences remains a hand-engineered process.", + "bbox": [ + 75, + 368, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead, VGGSfM significantly simplifies SfM correspondence tracking by employing a deep feed-forward tracking function. It accepts a collection of images and directly outputs a set of reliable point trajectories across all images. We achieve this by exploiting recent advances in video point tracking methods [15, 16, 22, 32]. Although developed for video-point tracking, these methods are inherently appropriate for SfM which requires a very compact set of highly accurate tracks (e.g., dense optical flow is too memory-demanding). Furthermore, point trackers mitigate the potential errors (sometimes called drift) caused by chaining of pairwise matches. However, as we describe below, our design differs from video trackers because SfM, which accepts free-form images, cannot assume temporal smoothness or ordering, and requires sub-pixel accuracy.", + "bbox": [ + 75, + 489, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tracker architecture The design of our tracker $\\mathbb{T}$ follows [22, 32], and is illustrated in Fig. 3. More specifically, given $N_{T}$ query points $\\{\\hat{\\mathbf{y}}_i^1,\\dots \\hat{\\mathbf{y}}_i^{N_T}\\}$ in a frame $I_{i}$ we bilinearly sample their corresponding query descriptors $\\{\\mathbf{m}_i^1,\\dots ,\\mathbf{m}_i^{N_T}\\}$ from image feature maps output by a 2D CNN. Then, each query descriptor is correlated with the feature maps of all $N_{I}$ frames at different spatial resolutions, which constructs a cost-volume pyramid. Flattening the latter yields tokens $V\\in \\mathbb{R}^{N_T\\times N_I\\times C}$ , where $C$ is the total number of elements in the cost-volume pyramid. Feeding the tokens to a Transformer, we obtain tracks $\\mathcal{T} = \\{T^j\\}_{j = 1}^{N_T}$ . Recall that each track $T^j$ comprises the", + "bbox": [ + 76, + 719, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/76d43148b58f308457f880f40b291753d2cd49c2f7b25858f343e5d406bd36e2.jpg", + "image_caption": [ + "Figure 3. Architecture of Tracker T. We adopt a coarse-to-fine design for the tracker. The coarse tracker first locates the approximate positions of corresponding points, and the fine tracker then refines these initial predictions." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "set of $N_I$ tracked 2D locations $\\mathbf{y}_i^j$ together with predicted visibility indicators $v_i^j$ .", + "bbox": [ + 498, + 306, + 890, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that, differently from [22, 32], our tracker does not assume temporal continuity. Therefore, we avoid the sliding window approach and, instead, attend to all the input frames together. Furthermore, unlike in [32], we predict each track independently of others. This allows to track a larger number of points at test time leading to increased density of reconstructed point clouds.", + "bbox": [ + 496, + 340, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Predicting tracking confidence In SfM, it is crucial to filter out any outlier correspondences as they can negatively impact the subsequent reconstruction stages. To this end, we enhance the tracker with the ability to estimate confidence for each track-point prediction.", + "bbox": [ + 496, + 450, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, we leverage the aleatoric uncertainty [33, 51] model which predicts variance $\\sigma_{i}^{j}$ together with each 2D track point $\\mathbf{y}_i^j$ , so that the resulting normal distribution $\\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)$ tightly peaks around each ground-truth 2D track point $\\mathbf{y}_i^{j\\star}$ . Hence, during training, the $\\ell_1 / \\ell_2$ loss, originally used in video point tracking, is replaced with the (negated) logarithm of the latter probability evaluated at each ground truth point $\\mathbf{y}_i^{j\\star}$ . Once trained, the confidence measure $1 / \\sigma_{i}^{j}$ is proportional to the inverse of the predicted variance. In practice, we assume a diagonal covariance matrix resulting in horizontal and vertical uncertainties $\\sigma_{i}^{j}$ .", + "bbox": [ + 496, + 527, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-fine tracking Moreover, since SfM requires highly accurate (pixel or sub-pixel level) correspondences, we employ a coarse-to-fine point-tracking strategy. As described above, we first coarsely track image points using feature maps that fully cover the input images $I$ . Then, we form $P \\times P$ patches by cropping input images around the coarse point estimates and execute the tracking again to obtain a sub-pixel estimate. Recall that, differently from the chained matching of traditional SfM, our tracker is fully differentiable. This enables back-propagating the gradient of the training loss $\\mathcal{L}$ through the whole framework to the tracker parameters. This reinforces the synergy between the tracking and the ensuing stages, which are described next.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21689", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learnable camera & point initialization", + "text_level": 1, + "bbox": [ + 76, + 90, + 418, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed above, a traditional SfM pipeline [39, 59] usually relies on an an incremental loop, which often initializes with a correspondence-rich image pair, gradually registers new frames, enlarges the point cloud, and conducts joint optimization (e.g., BA). However, although the framework has been fortified in robustness and accuracy through decades of improvements, this cumulative process has inevitably led to increased complexity. Furthermore, Incremental SfM is largely non-differentiable which complicates end-to-end learning from annotated data.", + "bbox": [ + 75, + 112, + 467, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, in pursuit of simplicity and differentiability, our method departs from the classical SfM scheme. Inspired by recent advances in deep camera pose estimators [36, 74, 85], we propose to initialize the cameras and the point cloud with a pair of deep Transformer [72] networks. Importantly, we register all cameras and reconstruct all scene points collectively in a non-incremental differentiable fashion.", + "bbox": [ + 75, + 265, + 467, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable camera initializer The predictor of initial cameras $\\hat{\\mathcal{P}}$ is implemented as a deep Transformer architecture $\\mathfrak{T}_{\\mathcal{P}}$ :", + "bbox": [ + 75, + 373, + 467, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} \\left(\\left\\{\\phi \\left(I _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I} \\right\\}, \\left\\{d ^ {\\mathcal {P}} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 422, + 467, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It accepts a set of tokens comprising global ResNet50 [24] features $\\phi(I_i)$ of input images $\\mathcal{I}$ , and the set of descriptors $d^{\\mathcal{P}}(\\mathbf{y}_i^j)$ of track points $\\mathbf{y}_i^j \\in T_i \\in \\mathcal{T}$ . Here, each track descriptor is output by an auxiliary branch of the tracker $\\mathbb{T}$ . Given these inputs, $\\mathfrak{T}_{\\mathcal{P}}$ first applies cross-attention between the global image feature (query) and the track-descriptor (key-value) pairs yielding $N_I = |\\mathcal{I}|$ tokens per scene. The output of cross-attention is then concatenated with an embedding of a preliminary camera estimated by the 8-point algorithm taking the correspondences between track points $\\mathbf{y}_i^j$ as input. Finally, we feed this concatenation to a Transformer trunk resulting in the initial cameras $\\hat{\\mathcal{P}}$ .", + "bbox": [ + 75, + 455, + 467, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable triangulation Given initial cameras $\\hat{\\mathcal{P}}$ and 2D tracks $\\mathcal{T}$ , the triangulator outputs the initial point cloud $\\hat{X}$ . Similar to the camera predictor, the triangulator is a Transformer $\\mathfrak{T}_X$", + "bbox": [ + 75, + 641, + 467, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\mathfrak {T} _ {X} \\left(\\left\\{d ^ {X} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 707, + 467, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accepting descriptors $d^{X}(\\mathbf{y}_{i}^{j})$ each comprising a tracker feature, and a positional harmonic embedding [47] of points $\\bar{\\mathbf{x}}^j\\in \\bar{X}$ from a preliminary point cloud $\\bar{X}$ . The preliminary point cloud is formed via closed-form multi-view Direct Linear Transform (DLT) 3D triangulation [23] of the tracks $\\mathcal{T}$ given the initial cameras $\\hat{\\mathcal{P}}$ . Please refer to the supplementary for a detailed description of both initializers.", + "bbox": [ + 75, + 733, + 467, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Bundle adjustment", + "text_level": 1, + "bbox": [ + 76, + 847, + 259, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the tracks $\\mathcal{T}$ (Sec. 3.2), initial cameras $\\hat{\\mathcal{P}}$ , and the initial point cloud $\\hat{X}$ (Sec. 3.3), Bundle Adjustment BA mini", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mizes the reprojection loss $\\mathcal{L}_{\\mathrm{BA}}[1,2,23,59]$", + "bbox": [ + 500, + 90, + 797, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX,\\mathcal{P} = \\operatorname {BA}(\\mathcal{T},\\hat{X},\\hat{\\mathcal{P}}) = \\operatorname *{arg min}_{X,\\mathcal{P}}\\mathcal{L}_{\\operatorname {BA}}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 116, + 820, + 143 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B A}} = \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {\\mathbf {x}}} v _ {i} ^ {j} \\| P _ {i} \\left(\\mathbf {x} ^ {j}\\right) - y _ {i} ^ {j} \\|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 147, + 888, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "summing over all reprojection errors $\\| P_i(\\mathbf{x}^j) - \\mathbf{y}_i^j\\|$ each comprising the distance between the projection $P_{i}(\\mathbf{x}^{j})$ of the point cloud $\\mathbf{x}^j\\in X$ to camera $P_{i}\\in \\mathcal{P}$ , and the $i$ -th 2D point $\\mathbf{y}_i^j\\in T^j$ of the track $T^j$ . Additionally, the error terms are filtered out if the corresponding points have low visibility, low confidence, or do not fit the geometric constraints defined by [59]. Points with large reprojections errors are also filtered [59, 63, 81]. More details are provided in the supplementary material.", + "bbox": [ + 500, + 202, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Levenberg-Marquardt Following common practice [39, 59], we minimize Eq. (6) with second-order Levenberg-Marquardt (LM) optimizer [49]. However, optimizing the main training loss (Eq. (2)) via backpropagation requires differentiability of Eq. (6) which is nontrivial. Therefore, we leverage the recently proposed Theiss library [54] which exploits the implicit function theorem to backpropagate through deep networks with nested optimization loops.", + "bbox": [ + 496, + 342, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Method details", + "text_level": 1, + "bbox": [ + 500, + 487, + 650, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera parameterization Each camera pose $P \\in \\mathcal{P}$ is parameterized with 8-degrees of freedom: the quaternion $q(R) \\in \\mathbb{R}^4$ of the rotation $R \\in \\mathbb{SO}(3)$ and the translation $\\mathbf{t} \\in \\mathbb{R}^3$ components of $P$ 's extrinsics $g \\in \\mathbb{SE}(3)$ , and the logarithm $\\ln(\\mathfrak{f}) \\in \\mathbb{R}$ of the camera focal length $\\mathfrak{f} \\in \\mathbb{R}^+$ . Given these values, the $3 \\times 4$ pose matrix is defined as $P = KR[\\mathbb{I}_3|\\mathbf{t}]$ , with the calibration matrix $K = [\\mathfrak{f},0,p_x;0,\\mathfrak{f},p_y;0,0,1] \\subset \\mathbb{R}^{3 \\times 3}$ (row major order). Following standard practice [39, 59], we set the principal-point coordinates $p_x, p_y \\in \\mathbb{R}$ to the center of the image.", + "bbox": [ + 496, + 511, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training loss The training loss $\\mathcal{L}$ (Eq. (2)) is defined as:", + "bbox": [ + 500, + 665, + 883, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(f _ {\\theta} (\\mathcal {I}), \\mathcal {P} ^ {\\star}, \\mathcal {T} ^ {\\star}, X ^ {\\star}\\right) = \\sum_ {j = 1} ^ {N _ {T}} \\left| \\mathbf {x} ^ {\\star j} - \\mathbf {x} ^ {j} \\right| _ {\\epsilon} + \\left| \\mathbf {x} ^ {\\star j} - \\hat {\\mathbf {x}} ^ {j} \\right| _ {\\epsilon} + \\\\ + \\sum_ {i = 1} ^ {N _ {I}} e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, P _ {i}\\right) + e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, \\hat {P} _ {i}\\right) - \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {T}} \\log \\mathcal {N} \\left(\\mathbf {y} _ {i} ^ {j \\star} \\mid \\mathbf {y} _ {i} ^ {j}, \\sigma_ {i} ^ {j}\\right) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 691, + 888, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $|\\mathbf{x}^{\\star j} - \\mathbf{x}^{j}|_{\\epsilon}$ and $|\\mathbf{x}^{\\star j} - \\hat{\\mathbf{x}}^{j}|_{\\epsilon}$ evaluate the $\\epsilon$ -thresholded pseudo-Huber loss $|\\cdot|_{\\epsilon}$ [9] between the ground truth 3D points $\\mathbf{x}^{\\star j}$ and the initial and BA-refined 3D points $\\hat{\\mathbf{x}}^{j} \\in X$ , $\\mathbf{x}^{j} \\in \\hat{X}$ respectively. The camera errors $e_{\\mathcal{P}}(P_{i}^{\\star}, P_{i})$ and $e_{\\mathcal{P}}(P_{i}^{\\star}, \\hat{P}_{i})$ compare the predicted initial pose $\\hat{P}_{i} \\in \\hat{\\mathcal{P}}$ and bundle-adjusted camera pose $P_{i} \\in \\mathcal{P}$ to the ground truth camera annotation $P_{i}^{\\star} \\in \\mathcal{P}_{i}^{\\star}$ . Here, $e_{\\mathcal{P}}(P, P')$ is defined as", + "bbox": [ + 496, + 792, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21690", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae02d9a9a590f9d23b4ab5a802f8a36d70ca2ba5bbd5951131f454bdb08ad79d.jpg", + "image_caption": [ + "Figure 4. Camera and point-cloud reconstructions of VGGsM on Co3D (left) and IMC Phototourism (right)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 196, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a51b61737b1376a155571c60a5f65263fcebdf3c48abbb88549c37f2144676e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 88, + 395, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c82164bf713cf0f8e4366d30c8135a3cb7ccdc4fc13aef0281abad9bd47c1dc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 89, + 591, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a5d9372aa6c42dd286a9a73473d33a9784a1a476ce6cf8e18ce3bea01eb6ebe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 88, + 689, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af2dbb6dcff635d97285e09f1f562c8ba09dd4a1c53793afba87a29c2375cf15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 89, + 887, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 80, + 244, + 194, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11d7d1c06220d36f6bd6e598cc22094c18313e9802b83f1ce23d1f47f2850690.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 244, + 395, + 398 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca1f83117bc6914ddc7ade570ebb5cca3debc3b35976cc2e10efd7174979c992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 244, + 591, + 398 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff9e74907372f44fedef4f1315f3cead9c279c3ff58d90db2f22f47372d642f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 244, + 689, + 397 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3009b743da56e9602210e16f47d34eb09f908072ae60a07846635dacfa36c6e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 244, + 887, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Huber-loss $|\\cdot|_{\\epsilon}$ between the parameterizations of poses $P, P'$ . Finally, $\\log \\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)$ computes the likelihood of a ground-truth track point $\\mathbf{y}_i^{j\\star} \\in T_i^\\star$ under a probabilistic track-point estimate defined by a 2D gaussian with mean and variance predictions $\\mathbf{y}_i^j$ and $\\sigma_i^j$ respectively (i.e. the aleatoric uncertainty model described in Sec. 3.2).", + "bbox": [ + 76, + 450, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 561, + 207, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets together with the protocols for training and evaluation. Then, we provide comparison to existing methods and ablation studies.", + "bbox": [ + 75, + 589, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following prior work [25, 39, 74], we evaluate camera pose estimation on Co3Dv2 [55] and IMC Phototourism datasets [31], and 3D triangulation on ETH3D [60]. Co3D is an object-centric dataset comprising turntable-style videos from 51 categories of MS COCO [38]. IMC Phototourism, provided by Image Matching Challenge [31], contains 8 testing scenes and 3 validation scenes of famous landmarks. Generally, the Co3D scenes have much wider baselines, making them challenging for traditional frameworks such as COLMAP, while the IMC samples often have sufficiently overlapping fursta, which is where COLMAP excels. ETH3D provides highly accurate point clouds (captured by laser scanner) for 13 indoor and outdoor scenes, and hence is suitable for the evaluation of triangulation.", + "bbox": [ + 75, + 638, + 468, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. For the model evaluated on IMC Phototourism and ETH3D, we follow the protocol of [25, 39, 57] and train on the MegaDepth dataset [35]. MegaDepth con", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tains 1M crowd-sourced images depicting 196 tourism landmarks, auto-annotated with SfM tools. Hyper-parameters are tuned using the IMC validation set. As in prior work [25, 40], some scenes of MegaDepth are excluded from training due to their low quality or due to overlap with the IMC test set. For Co3Dv2, we conduct training and evaluation on 41 categories as in [36, 74, 85].", + "bbox": [ + 496, + 450, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We chose a multi-stage training strategy for better stability. We first train the tracker T on the synthetic Kubric dataset [21] following the training protocol of [15, 32]. Then, the tracker is fine-tuned solely on Co3D or MegaDepth, depending on the test dataset. Subsequently, we train the camera initializer, with the tracker frozen. Next, the triangulator is trained with the aforementioned two components held frozen. Finally, all components are trained end-to-end. In all stages, we randomly sample a training batch of 3 to 30 frames.", + "bbox": [ + 496, + 561, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing. Given input test frames $\\mathcal{I}$ , we first select the query frame by identifying the image that is closest to all others based on the cosine similarity between global descriptors extracted by an off-the-shelf ResNet50 [24]. Then, we extract SuperPoint and SIFT keypoints from the query frame to serve as the query points for the tracker $\\mathbb{T}$ . Although our method can track any query point, it performs better when the queries are distinctive. To improve accuracy, we iterate the whole reconstruction function $f_{\\theta}$ multiple times until reaching sub-pixel BA reprojection error $\\mathcal{L}_{\\mathrm{BA}}$ . After the first iteration, the query image for each subsequent iteration is the one that is farthest from the", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21691", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aafcf0ddb6d3eb1def661e0bba51f4556223b33b0003c81c16cba03e33de7a63.jpg", + "image_caption": [ + "Query" + ], + "image_footnote": [], + "bbox": [ + 78, + 106, + 240, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80075a9003b24193276a93f9a5956a0e05e136ba6187903daa1aecf422f056b3.jpg", + "image_caption": [ + "Figure 5. Qualitative Evaluation of Tracking. In each row, the left-most frame contains the query image with query points (crosses). The predicted track points $\\mathbf{y}_i^j$ (dots) are shown to the right. The top-right part also highlights our track-point confidence predictions (described in Sec. 3.2), illustrated as ellipses with extent proportional to the predicted variance $\\sigma_i^j$ . Note how the uncertainty corresponds to an expectation, e.g., a keypoint covering a vertical stripe has higher uncertainty along the $y$ -axis." + ], + "image_footnote": [], + "bbox": [ + 112, + 205, + 204, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71c979bdebe60df942874b8dae91efdf8f940377ad6bb37136b2e8737ab02f1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 106, + 419, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d98486beba8ef5e294d995b23c8d7716b7b309d66fbd06445ac1778f35257e1b.jpg", + "image_caption": [ + "Tracking Results" + ], + "image_footnote": [], + "bbox": [ + 423, + 106, + 584, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37b5585f4a0a05ba48ce1022a036e51ad609ccef2806fb207417d0382f55567e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 106, + 750, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2d9614eef7562234ca71e2aeb1c5cac6a76d2546f17eeddde3ca15d67b6270f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 95, + 893, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2568c0bf9252837084175c3737cdf7ed89569449b9f2aa8d346e4a4e0d697c4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 205, + 431, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b56803d0e810aa036875e36d69780fa9e1acb0b086af5f3b4a2af733ec63603f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 205, + 584, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38e8a6110c2ca96bd8346c0e7a57f7b9b15dbf2a058617ae51db700f05497954.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 205, + 743, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32de1d2a21c79c0f5b2f33537f1d3b62f79ee87ca9e35ad9e1bdaf66c776a044.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 205, + 898, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b5f6d3ab871d7a691b02de27a45780637b573e4901b0710aed7fc0ecfa2635b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Co3D DatasetIncremental SfMDeep
MethodCOLMAP (SP + SG)PixSfM (SP + SG)RelPose[85]PoseReg[74]RelPose++[36]PoseDiffusion[74]Ours w/o JointOurs
RRE@15°31.633.757.153.282.380.588.292.1
RTE@15°27.332.9-49.177.279.883.488.3
AUC@30°25.330.1-45.065.166.570.774.0
", + "bbox": [ + 96, + 380, + 870, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b614ce7de853b292172c90a8c138b9393701c0eeecedb68a0b4663db46b8541.jpg", + "table_caption": [ + "Table 1. Camera Pose Estimation on Co3D, where the proposed method outperforms previous methods by a large margin. Ours w/o Joint indicates the variant of our framework without training all components jointly." + ], + "table_footnote": [], + "table_body": "
IMC DatasetMethodAUC@3°AUC@5°AUC@10°
DeepDeepSFM10.2719.3631.35
PoseDiffusion12.3123.1736.82
Incremental SfMCOLMAP (SIFT+NN)23.5832.6644.79
PixSfM (SIFT + NN)25.5434.8046.73
PixSfM (LoFTR)44.0656.1669.61
PixSfM (SP + SG)45.1957.2270.47
DFSfM (LoFTR)46.5558.7472.19
DeepOurs w/o Joint38.2351.6068.35
Ours45.2358.8973.92
", + "bbox": [ + 81, + 518, + 467, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Camera Pose Estimation on IMC. Our method achieves better accuracy than state-of-the-art Incremental SfM approaches on 2 out of 3 AUC thresholds.", + "bbox": [ + 75, + 655, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "query image of the previous iteration (as measured by the ResNet50 cosine similarity). The re-projections of the point cloud from the current iteration initialize the tracking in the next iteration.", + "bbox": [ + 75, + 727, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Camera pose estimation", + "text_level": 1, + "bbox": [ + 76, + 800, + 299, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [25, 31, 39, 74], we report the metric area-under-curve (AUC) to evaluate camera pose accuracy. In Co3D, similar to [74], we also report the relative rotation error (RRE) and relative translation error (RTE). More specifically, for every pair of input frames, we compute the angular", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "translation and rotation error, which are later compared to a threshold yielding accuracies RTE and RRE respectively. For a range of thresholds, AUC picks the lower between RRE and RTE, and outputs the area under the accuracy-threshold curve. The results on IMC and CO3D are presented in Tab. 1 and Tab. 2 respectively. For a fair comparison on IMC, we finetune DeepSFM [78] and PoseDiffusion [74] on MegaDepth using their open-source code. The results of Incremental SfM methods are copied from Detector-free SfM [25].", + "bbox": [ + 496, + 521, + 893, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results indicate that VGGSfM outperforms existing methods by a large margin (+9 accuracy points for each metric) on the CO3D dataset. Here, traditional SfM pipelines suffer because of the wide baselines between test frames. On IMC, with a good overlap between views, traditional SfM remains superior to recent data-driven deep-learning pipelines [74, 78]. Our end-to-end trained VGGSfM, however, outperforms all other methods on AUC@10 and AUC@5, and ranks second on AUC@3, convincingly demonstrating its ability to perform well in both narrow- and wide-baseline regimes. The accuracy and completeness of our point clouds can be further qualitatively evaluated in Fig. 4.", + "bbox": [ + 496, + 703, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21692", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d55c3f00fb44fed494b689d711a3de6d52d9884c90830a134adc12d7bab576cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)Completeness (%)
1cm2cm5cm1cm2cm5cm
PatchFlow (LoFTR)66.7378.7389.933.4811.3430.96
PixSfM (LoFTR)74.4284.0892.632.919.3927.31
PixSfM (SIFT + NN)76.1885.6093.160.170.713.29
PixSfM (SP + SG)79.0187.0493.800.752.7711.28
DFSfM (LoFTR)80.3889.0195.833.7311.0729.54
Ours80.6289.4996.524.5213.1133.96
", + "bbox": [ + 84, + 88, + 467, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3D triangulation", + "text_level": 1, + "bbox": [ + 76, + 275, + 240, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the accuracy and completeness of 3D triangulation on the ETH3D dataset [60] using the same protocol as [17, 25, 39], which triangulates points with fixed camera poses and intrinsics. Results are shown in Tab. 3, where metrics are averaged over all scenes. Our VGGsFm achieves better accuracy and completeness than all baselines (PatchFlow [17], PixSfM [39], and DFSfM [25]), regardless of which keypoint detection or matching method they use. This is especially obvious from the completeness attained at the $5\\mathrm{cm}$ threshold, with our $33.96\\%$ compared to $29.54\\%$ of the best prior work.", + "bbox": [ + 75, + 299, + 470, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 474, + 225, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "End-to-end Training. As reported in Tab. 1 and Tab. 2, the end-to-end joint training of the whole framework is important for achieving state-of-the-art performance. Specifically, comparing VGGsF to an ablation which lacks end-to-end training (Ours w/o joint) we record an improvement from $70.7\\%$ AUC@30 to $74.0\\%$ on the Co3D dataset, and $68.35\\%$ AUC@10 to $73.92\\%$ on the IMC dataset. This demonstrates the benefits of our fully-differentiable design, and the synergy between its components.", + "bbox": [ + 75, + 497, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tracking or Pairwise Matching. We compare the performance of our predicted tracks to pairwise matching methods on the IMC dataset. Specifically, we split our 2D tracks into pairwise matches and feed these matches to PixSfM. Also, we construct 2D tracks by pairwise matching (based on the open-source implementation of PixSfM) and feed them to our framework. It is worth noting that tracks from pairwise matching have a lot of \"holes\" because pairwise matching cannot guarantee proper point tracking. We fix these holes by setting their locations as the point locations in the query frame, marking them as invisible. At the same time, for fair comparison, cameras are still initialized using our tracks, because SP+SG cannot provide track features to our camera initializer. The results are shown in Tab. 4. Although COLMAP (the basis of PixSfM) is designed and carefully engineered for pairwise matching, our tracks achieve a slightly better result than the state-of-the-art matching op", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/526c4b5c968c6a3ddaafda2bd845ac0d7ee9e8970df11e6c593d0ca97cb541bc.jpg", + "table_caption": [ + "Table 3. 3D Triangulation on ETH3D [60] reporting the accuracy and completeness metrics at different thresholds." + ], + "table_footnote": [], + "table_body": "
PixSfM(SP + SG)PixSfM(Our Tracks)Ours(SP + LG)Ours
AUC@10°70.4770.6268.7873.92
", + "bbox": [ + 503, + 88, + 890, + 122 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/737c144d44999c8bfe6a1e2075a1fa85167e26a147792fd53d1fca19fa14764b.jpg", + "table_caption": [ + "Table 4. Tracking or Pairwise Matching. We respectively provide tracks predicted by our tracker or matches estimated by SP+SG to PixSfM and to our VGGSfM." + ], + "table_footnote": [], + "table_body": "
PoseDiffusionOur Camera Initialized
DLT62.1869.42
Our Triangulator66.3773.92
", + "bbox": [ + 542, + 186, + 849, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Study for Camera Initializer and Triangulator. A clear performance drop is observed when replacing our camera initializer by deep camera prediction method PoseDiffusion, or replacing triangulator by DLT.", + "bbox": [ + 498, + 244, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion $\\mathrm{SP + SG}$ . Instead, directly feeding $\\mathrm{SP + SG}$ tracks to our framework leads to a performance drop. We attribute this to the fact that using $\\mathrm{SP + SG}$ tracks cannot benefit from the joint training. We also provide a qualitative evaluation of our tracking accuracy in Fig. 5 on the IMC dataset.", + "bbox": [ + 496, + 327, + 890, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Camera Controller and Triangulator. We also validate the design of our camera controller $\\mathfrak{T}_{\\mathcal{P}}$ and triangulator $\\mathfrak{T}_X$ on the IMC dataset. As reported in Tab. 5, AUC@10 drops clearly if we replace them with alternatives, proving that they provide sufficiently accurate initialization for our global bundle adjustment BA, without the need for incremental camera registration.", + "bbox": [ + 496, + 406, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coarse-to-fine Tracking. As discussed above, accurate correspondences are important for structure from motion. We demonstrate the significance of our coarse-to-fine tracking mechanism for the method performance. By conducting an ablation study where the fine tracker is removed, we observe a significant performance drop on the IMC dataset, with AUC@10 dropping from $73.92\\%$ to $62.30\\%$ .", + "bbox": [ + 496, + 515, + 890, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 647, + 617, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented VGGSfM, a fully differentiable SfM approach. We find that even long-standing pipelines, such as Structure-from-Motion, benefit from a learned adaptation between their components. This allows VGGSfM to be simpler than traditional SfM frameworks while achieving better performance across benchmark datasets. Moreover, our framework is fully implemented in Python, which will allow for easy modification and improvements in the future. While VGGSfM already achieves good performance, it cannot yet compete with established pipelines in all application domains. For example, it currently lacks the capability to process thousands of images as in traditional SfM frameworks. Nonetheless, we find differentiable SfM a promising direction of research, and our approach lays the foundation for further advances.", + "bbox": [ + 496, + 672, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Noah Snavely, Steven M Seitz, and Richard Szeliski. Bundle adjustment in the large. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 29-42. Springer, 2010. 5", + "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1, 2, 3, 5", + "[3] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2022. 2", + "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 2", + "[5] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4322-4331, 2019. 3", + "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 3", + "[7] Lucas Brynte, José Pedro Iglesias, Carl Olsson, and Fredrik Kahl. Learning structure-from-motion with graph attention networks. arXiv preprint arXiv:2308.15984, 2023. 3", + "[8] Jonathan L. Carrivick, Mark W. Smith, and Duncan J. Quincey. Structure from Motion in the Geosciences. John Wiley & Sons, 2016. 1", + "[9] Pierre Charbonnier, Laure Blanc-féaud, Gilles Aubert, and Michel Barlaud. Deterministic edge-preserving regularization in computed imaging. IEEE Trans. Image Processing, 6:298-311, 1997. 5", + "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1", + "[11] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 3", + "[12] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2, 3", + "[13] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. In Proceedings of the IEEE International Conference on Computer Vision, pages 864-872, 2015. 2", + "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 4", + "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurlPS Datasets Track, 2022. 2, 3, 4, 6", + "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 2, 3, 4", + "[17] Mihai Dusmanu, Johannes L Schonberger, and Marc Pollefeys. Multi-view optimization of local feature geometry. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 670-686. Springer, 2020. 8", + "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3", + "[19] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 368-381. Springer, 2010. 2", + "[20] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Towards Internet-scale multi-view stereo. In Proc. CVPR. IEEE, 2010. 1", + "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliaasacchi. Kubric: a scalable dataset generator. 2022. 6", + "[22] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2, 3, 4", + "[23] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 1, 2, 3, 5", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. arXiv preprint arXiv:1512.03385, 2015. 5, 6", + "[25] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In *arxiv*, 2023. 1, 2, 3, 4, 6, 7, 8", + "[26] Jared Heinly, Johannes L. Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world* in six days $(\\ast)$ (as captured by the yahoo 100 million image dataset). In" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21694", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1", + "[27] Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Piermattei, James O'Connor, and Jacqueline Rosette. Structure from motion photogrammetry in forestry: A review. Current Forestry Reports, 5:155-168, 2019. 1", + "[28] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. ArXiv, 2212.04492, 2022. 1", + "[29] San Jiang, Cheng Jiang, and Wanshou Jiang. Efficient structure from motion for large-scale UAV images: A review and a comparison of sfm tools. ISPRS Journal of Photogrammetry and Remote Sensing, 167:230-251, 2020. 1", + "[30] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3", + "[31] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 2, 6, 7", + "[32] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023. 2, 3, 4, 6", + "[33] Alex Kendall and Yarin Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Proc. NeurIPS, 2017. 4", + "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1", + "[35] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 6", + "[36] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3, 5, 6, 7", + "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 1", + "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In Proc. ECCV, 2014. 6", + "[39] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. arXiv.cs, abs/2108.08291, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[40] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023. 3, 4, 6" + ], + "bbox": [ + 78, + 92, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yin Lou, Noah Snavely, and Johannes Gehrke. Matchminer: Efficient spanning structure mining in large image collections. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12, pages 45-58. Springer, 2012. 3", + "[42] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 2", + "[43] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 2", + "[44] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 3", + "[45] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3", + "[46] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust Wide Baseline Stereo from Maximally Stable Extremal Regions. In Proc. BMVC, 2002. 2", + "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Proc. ECCV, 2020. 5", + "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[49] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 5", + "[50] Pierre Moulon, Pascal Monasse, and Renaud Marlet. Global fusion of relative motions for robust, accurate and scalable structure from motion. In Proceedings of the IEEE international conference on computer vision, pages 3248-3255, 2013. 2", + "[51] David Novotny, Diane Larlus, and Andrea Vedaldi. Learning 3d object categories by looking around them. In Proc. ICCV, 2017. 4", + "[52] John Oliensis. A critique of structure-from-motion algorithms. Computer Vision and Image Understanding, 80(2): 172-214, 2000. 2", + "[53] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1, 2", + "[54] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, et al. Theses: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 35:3801-3818, 2022. 2, 5", + "[55] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21695", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2, 6", + "[56] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 3", + "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 4, 6", + "[58] Frederik Schaffalitzky and Andrew Zisserman. Multi-view Matching for Unordered Image Sets, or \"How Do I Organize My Holiday Snaps?\" In Proc. ECCV, 2002. 1", + "[59] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3, 4, 5", + "[60] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 2, 6, 8", + "[61] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 3", + "[62] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 3", + "[63] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5", + "[64] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3", + "[65] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. arXiv preprint arXiv:1806.04807, 2018. 3", + "[66] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018.3", + "[67] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 3", + "[68] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Bill Triggs, Philip F. McLauchlan, Richard I. Hartley, and Andrew W. Fitzgibbon. Bundle Adjustment - A Modern Synthesis. In Proc. ICCV Workshop, 2000. 3", + "[70] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 3, 4", + "[71] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 3", + "[72] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. NeurIPS, 2017. 2, 5", + "[73] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 3", + "[74] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3, 5, 6, 7", + "[75] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2, 3", + "[76] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1", + "[77] Tong Wei, Yash Patel, Alexander Shekhotsov, Jiri Matas, and Daniel Barath. Generalized differentiable ransac. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17649-17660, 2023. 3", + "[78] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 230–247. Springer, 2020. 3, 7", + "[79] Matthew J Westoby, James Brasington, Niel F Glasser, Michael J Hambrey, and Jennifer M Reynolds. 'structure-from-motion'photogrammetry: A low-cost, effective tool for geoscience applications. Geomorphology, 179:300-314, 2012. 1", + "[80] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 61-75. Springer, 2014. 2", + "[81] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 3", + "[83] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3d animals in the wild. 2023. 3", + "[84] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned Invariant Feature Transform. In Proc. ECCV, 2016. 3", + "[85] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3, 5, 6, 7", + "[86] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 3", + "[87] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 430 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21697", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_model.json b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..67f84a4dd5f8d74a42ba61c91d946116a91a0cba --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_model.json @@ -0,0 +1,2776 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "VGGSfM: Visual Geometry Grounded Deep Structure From Motion" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.273, + 0.2 + ], + "angle": 0, + "content": "Jianyuan Wang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.181, + 0.452, + 0.198 + ], + "angle": 0, + "content": "Nikita Karaev 1,2" + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.181, + 0.663, + 0.199 + ], + "angle": 0, + "content": "Christian Rupprecht1" + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.181, + 0.838, + 0.199 + ], + "angle": 0, + "content": "David Novotny2" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.216, + 0.58, + 0.235 + ], + "angle": 0, + "content": "Visual Geometry Group, University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.216, + 0.761, + 0.233 + ], + "angle": 0, + "content": "Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.25, + 0.605, + 0.265 + ], + "angle": 0, + "content": "https://vggsfm.github.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.288, + 0.313, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.474, + 0.668 + ], + "angle": 0, + "content": "Structure-from-motion (SfM) is a long-standing problem in the computer vision community, which aims to reconstruct the camera poses and 3D structure of a scene from a set of unconstrained 2D images. Classical frameworks solve this problem in an incremental manner by detecting and matching keypoints, registering images, triangulating 3D points, and conducting bundle adjustment. Recent research efforts have predominantly revolved around harnessing the power of deep learning techniques to enhance specific elements (e.g., keypoint matching), but are still based on the original, non-differentiable pipeline. Instead, we propose a new deep pipeline VGGsfM, where each component is fully differentiable and thus can be trained in an end-to-end manner. To this end, we introduce new mechanisms and simplifications. First, we build on recent advances in deep 2D point tracking to extract reliable pixel-accurate tracks, which eliminates the need for chaining pairwise matches. Furthermore, we recover all cameras simultaneously based on the image and track features instead of gradually registering cameras. Finally, we optimise the cameras and triangulate 3D points via a differentiable bundle adjustment layer. We attain state-of-the-art performance on three popular datasets, CO3D, IMC Phototourism, and ETH3D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.208, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Reconstructing the camera parameters and the 3D structure of a scene from a set of unconstrained 2D images is a long-standing problem in the computer vision community. Among many other applications [8, 27, 29, 53, 79], it has recently emerged as an important component of learning neural fields [10, 28, 34, 37, 48, 76]. The problem is usually solved via the Structure-from-Motion (SfM) framework which estimates the 3D point cloud (Structure) and the parameters of each camera (Motion) in the scene. State-of-the-art methods [25, 39] follow the incremental SfM paradigm whose origins can be traced back to the early 2000s [23, 58]. It usually begins with a small set of" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.288, + 0.692, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.289, + 0.89, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.764, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Figure 1. Reconstruction of In-the-wild Photos with VGGsFm, displaying estimated point clouds (in blue) and cameras (orange)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.869 + ], + "angle": 0, + "content": "correspondence-rich images as initialization, and gradually adds more views into the reconstruction, through keypoint detection, matching, verification, image registration, triangulation, bundle adjustment (BA), and so on [2, 20, 26, 59]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent research efforts have predominantly revolved around leveraging the power of deep learning techniques to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.895, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Figure 2. Overview of VGGs fM. Our method extracts 2D tracks from input images, reconstructs cameras using image and track features, initializes a point cloud based on these tracks and camera parameters, and applies a bundle adjustment layer for reconstruction refinement. The whole framework is fully differentiable and designed for end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.473, + 0.581 + ], + "angle": 0, + "content": "enhance specific elements within the original pipeline while preserving the incremental SfM framework as a whole. For instance, SuperPoint and SuperGlue [14, 57] focus on improving keypoint detection and matching. Pixel-perfect SfM [39] proposes deep feature-metric refinement to adjust both keypoints and bundles. Detector-free feature matching methods [12, 64, 75] bypass early keypoint detection by means of attention, which is powerful in poorly textured scenes. Detector-free SfM [25] builds a coarse SfM model through quantized detector-free matches and then iteratively refines it with multi-view consistency constraints. These advancements successfully combine deep learning approaches (such as deep feature matching) with well-established hand-engineered components, such as the incremental camera registration of COLMAP [59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The widespread success of end-to-end training warrants the question of what benefits it can bring to long-standing frameworks such as SfM. Naturally, it is often difficult to assess the merits of new approaches when compared with decades of continuous improvements. Nonetheless, in this paper, we answer this question by introducing a fully-differentiable SfM pipeline, dubbed Visual Geometry Grounded Deep Structure From Motion (VGGSfM), which trains in an end-to-end manner. We find that this allows the pipeline to be simpler than prior frameworks while achieving better or comparable performance. Training end-to-end allows each component to generate outputs that facilitate the task of its successor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To build a fully-differentiable pipeline, we make several substantial changes to the SfM procedures and overall obtain better performance. Specifically, our model builds on recent advances in deep 2D point tracking [15, 16, 22, 32] to directly extract reliable pixel-accurate tracks. This simplifies the correspondence estimation step in traditional SfM, which first estimates pairwise matches and then connects them into tracks. Then, based on the image and track fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.476 + ], + "angle": 0, + "content": "tures, VGGsFm estimates all cameras jointly via a Transformer [72], and subsequently all 3D points. Different from Incremental SfM, this approach is simpler and easier to differentiate as it does not depend on a discrete, combinatorial correspondence chaining step. Finally, for bundle adjustment, we replace the commonly employed non-differentiable Ceres solver [3] with the fully differentiable second-order Theseus solver [54]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Hence, we fuse all the SfM components into a single fully differentiable reconstruction function \\( f \\). Besides that, in our experiments, we also show that the individual modules perform well in isolation. Ultimately, end-to-end training yields another performance improvement, that surpasses the performance of isolated components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.676 + ], + "angle": 0, + "content": "We evaluate VGGsFm for the task of camera pose estimation on the CO3Dv2 [55] and IMC Phototourism [31] datasets, and for 3D triangulation on the ETH3D [60] dataset. Our method attains strong performance on all benchmarks. At the same time, we conduct in-the-wild reconstruction to validate the generalization ability of our proposed framework, as shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.642, + 0.708 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Structure from Motion is a fundamental problem in computer vision and has been investigated for decades [23, 52, 53]. The classical pipelines usually solve the SfM problem in a global [13, 50, 80] or incremental [2, 19, 59, 63, 81] manner. Both of which are usually based on pairwise image keypoint matching. Incremental SfM is arguably the most widely adopted strategy (e.g., the popular framework COLMAP [59]). Therefore, in the following sections, we refer to incremental SfM as \"classical\" or \"traditional\" SfM. We defer the discussion of global SfM to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Traditional SfM frameworks often start by detecting keypoints and feature descriptors [4, 42, 43, 46]. They then" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21687" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "search for image pairs with overlapping frusta by matching these keypoints across different images (e.g., with a nearest-neighbour search) [2, 41, 59]. These image pairs are further verified via two-view epipolar geometry or homography [23] through RANSAC [18]. Then, a pair or a small set of images is carefully selected for initialization. New images are gradually registered by solving the Perspective-\\(n\\)-Point (PnP) problem [44], followed by triangulating 3D points, and bundle adjustment [69]. This process is iterated until all the frames are either registered or discarded. 2D correspondences (multi-view tracks) are the basis of the whole process, however, they are usually simply constructed by chaining two-view matches [59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.547 + ], + "angle": 0, + "content": "Many deep-learning approaches have been proposed to enhance this framework. For example, [14, 70, 84] provide better keypoint detection and [11, 30, 40, 57, 61] focus on matching. Furthermore, detector-free matching methods [12, 64, 75] propose to avoid sparse keypoint detection by building semi-dense matches via self and cross attention. [7] solves the subproblems of SfM by graph attention networks. Some studies improve the performance of RANSAC by making it trainable [5, 6, 77]. Recent state-of-the-art methods are PixSfM [39] and the concurrent Detector-free SfM (DFSfM) [25]. PixSfM refines the tracks and structure estimated by COLMAP through feature-metric keypoint adjustment and feature-metric bundle adjustment. Detector-free SfM proposes to first build a coarse SfM model using detector-free matches and COLMAP (or other frameworks), and then to iteratively refine the tracks and the structure of the coarse model by enforcing multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Recently, fully differentiable SfM pipelines have also been explored. They usually use deep neural networks to regress camera poses and depths [65, 66, 68, 71, 73, 78, 87]. Although using an approximation of bundle adjustment [65, 68, 78], these methods suffer from limited generalizability and scalability (very few input frames) [65, 71, 73, 78, 87], or rely on temporal relationship [66, 68]. Meanwhile, some methods are category-specific [45, 82, 83]. The recent efforts on deep camera pose estimation can scale up to more than 50 frames, but they do not reconstruct the scene [36, 62, 74, 85]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Point tracking. Since VGGSfM proposes a novel point track predictor, next, we review recent advances in this field. Inspired by the optical-flow architecture of RAFT [67], PIPs [22] revisited point tracking, a task related to Particle Video [56], and proposed a highly accurate tracker of isolated points in a video. TAP-Vid [15] (i.e., \"Tracking Any Point\") introduced a benchmark for point tracking and a baseline model, which was later improved in TAPIR [16] by integrating the iterative update mechanism from PIPs. PointOdyssey [86] simplified PIPs and proposed a benchmark for the long-term version of point tracking. CoTracker [32] closed the gap between single point track" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "ing and dense Optical Flow with joint point tracking. However, these works are designed for videos, i.e. temporally-ordered sequences of frames. In our point tracker, given the input frames are unordered, we do not assume a temporal relationship between input frames. We therefore process all frames jointly, avoiding windowed inference of [32]. Since SfM relies on highly accurate correspondences, our tracks are further refined in a coarse-to-fine manner to achieve subpixel accuracy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.244, + 0.593, + 0.26 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.316 + ], + "angle": 0, + "content": "In this section, we describe the components of VG-GSfM and how they are composed in a fully differentiable pipeline. An overview of our framework is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.472 + ], + "angle": 0, + "content": "Problem setting Given a set of free-form images observing a scene, VGGSfM estimates their corresponding camera parameters and the 3D scene shape represented as a point cloud. Formally, given a tuple \\(\\mathcal{I} = (I_1,\\dots,I_{N_I})\\) of \\(N_{I}\\in \\mathbb{N}\\) RGB images \\(I_{i}\\in \\mathbb{R}^{3\\times H\\times W}\\), VGGSfM estimates the corresponding camera projection matrices \\(\\mathcal{P} = (P_{1},\\ldots ,P_{N_{I}}|P_{i}\\subset \\mathbb{R}^{3\\times 4})\\) and the scene cloud \\(X = \\{\\mathbf{x}^j\\}_{j = 1}^{N_{\\mathbf{x}}} \\text{ of } N_{\\mathbf{x}}\\in \\mathbb{N} \\) 3D points \\(\\mathbf{x}^j\\in \\mathbb{R}^3\\). Each projection matrix \\(P_{i}\\) consists of extrinsic (pose) \\(g_{i}\\in \\mathbb{SE}(3)\\) and intrinsics \\(K_{i}\\in \\mathbb{R}^{3\\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.635 + ], + "angle": 0, + "content": "A 3D point \\(\\mathbf{x}^j\\) can be projected to the \\(i\\)-th camera yielding a 2D screen coordinate \\(\\mathbf{y}_i^j = P_i(\\mathbf{x}^j) \\sim \\lambda K_i \\hat{\\mathbf{x}}_i^j\\); \\(\\lambda \\in \\mathbb{R}_+\\), where \\(\\hat{\\mathbf{x}}_i^j = g_i \\mathbf{x}^j\\) is the world coordinate \\(\\mathbf{x}^j\\) expressed in view-coordinates of the \\(i\\)-th camera. The projection of the point \\(\\mathbf{x}^j\\) to all input cameras is a track \\(T^j = ((y_1^j, v_1^j) \\dots, (y_{N_I}^j, v_{N_I}^j))\\) consisting of \\(N_I\\) matching 2D points \\(\\mathbf{y}_i^j \\in \\mathbb{R}^2\\), and their corresponding binary indicators \\(v_i^j \\in \\{0, 1\\}\\) denoting visibility of the \\(j\\)-th point in the \\(i\\)-th camera. We denote \\(\\mathcal{T}_i = \\{T_i^1, \\dots, T_i^{N_T}\\}\\) as the set of all tracks \\(T_i^j\\) in the \\(i\\)-th camera." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.645, + 0.671, + 0.66 + ], + "angle": 0, + "content": "3.1. Method overview" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.669, + 0.84, + 0.686 + ], + "angle": 0, + "content": "VGGsM implements SfM via a single function \\( f_{\\theta} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.695, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\mathcal {I}) = \\mathcal {P}, X \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.892, + 0.782 + ], + "angle": 0, + "content": "accepting the set of \\(N_I\\) scene images \\(\\mathcal{I}\\) and outputting the camera parameters \\(\\mathcal{P}\\) and the scene point cloud \\(X\\). Importantly, \\(f_{\\theta}\\) is fully differentiable and, as such, its parameters \\(\\theta\\) are learned by minimizing the training loss \\(\\mathcal{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.79, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\theta^ {\\star} = \\arg \\min _ {\\theta} \\sum_ {s = 1} ^ {S} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathcal {I} _ {s}\\right), \\mathcal {P} _ {s} ^ {\\star}, \\mathcal {T} _ {s} ^ {\\star}, X _ {s} ^ {\\star}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "summing over \\(S \\in \\mathbb{N}\\) training image sets \\(\\mathcal{I}_s\\) annotated with ground-truth cameras \\(\\mathcal{P}_s^\\star\\), tracks \\(\\mathcal{T}_s^\\star\\), and point clouds \\(X_s^\\star\\). We defer the details of \\(\\mathcal{L}\\) to Sec. 3.5 and, in the following paragraphs, discuss the architecture of \\(f_{\\theta}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.168 + ], + "angle": 0, + "content": "The reconstruction function Following traditional SfM [59], VGGSfM decomposes the reconstruction function \\( f_{\\theta} \\) into four seamless stages: 1) point tracking \\( \\mathbb{T} \\), 2) initial camera estimator \\( \\mathfrak{T}_{\\mathcal{P}} \\), 3) triangulator \\( \\mathfrak{T}_X \\) and, 4) Bundle Adjustment BA, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.176, + 0.29, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} = \\mathrm {T} (\\mathcal {I})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.196, + 0.469, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} (\\mathcal {I}, \\mathcal {T}) \\\\ \\hat {\\hat {\\tau}} = \\mathfrak {T} _ {\\hat {\\tau}} (\\mathcal {\\hat {T}}, \\hat {\\hat {\\tau}}) \\end{array} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.215, + 0.326, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\mathfrak {T} _ {X} (\\mathcal {T}, \\hat {\\mathcal {P}})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.239, + 0.35, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P}, X = _ {\\mathrm {B A}} (\\mathcal {T}, \\hat {\\mathcal {P}}, \\hat {X}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.262, + 0.47, + 0.339 + ], + "angle": 0, + "content": "The track predictor \\(\\mathbb{T}\\), hereafter referred to as \"tracker\", estimates 2D tracks \\(\\mathcal{T}\\) given input images \\(\\mathcal{I}\\). Subsequently, \\(\\mathfrak{T}_{\\mathcal{P}}\\) and \\(\\mathfrak{T}_X\\) provide initial cameras \\(\\mathcal{P}\\) and an initial point cloud \\(\\hat{X}\\) respectively. Finally, BA enhances accuracy by refining the cameras and 3D points together." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.346, + 0.185, + 0.363 + ], + "angle": 0, + "content": "3.2. Tracking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.489 + ], + "angle": 0, + "content": "Establishing precise 2D correspondences is important for achieving accurate 3D reconstruction. Traditionally, SfM frameworks first estimate pairwise image-to-image correspondences that are later chained into multi-image tracks \\( T \\) [25, 39, 59]. Here, typically only point-pair matching benefits from learned components [14, 40, 57, 70], while the chaining of pairwise correspondences remains a hand-engineered process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Instead, VGGSfM significantly simplifies SfM correspondence tracking by employing a deep feed-forward tracking function. It accepts a collection of images and directly outputs a set of reliable point trajectories across all images. We achieve this by exploiting recent advances in video point tracking methods [15, 16, 22, 32]. Although developed for video-point tracking, these methods are inherently appropriate for SfM which requires a very compact set of highly accurate tracks (e.g., dense optical flow is too memory-demanding). Furthermore, point trackers mitigate the potential errors (sometimes called drift) caused by chaining of pairwise matches. However, as we describe below, our design differs from video trackers because SfM, which accepts free-form images, cannot assume temporal smoothness or ordering, and requires sub-pixel accuracy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Tracker architecture The design of our tracker \\(\\mathbb{T}\\) follows [22, 32], and is illustrated in Fig. 3. More specifically, given \\(N_{T}\\) query points \\(\\{\\hat{\\mathbf{y}}_i^1,\\dots \\hat{\\mathbf{y}}_i^{N_T}\\}\\) in a frame \\(I_{i}\\) we bilinearly sample their corresponding query descriptors \\(\\{\\mathbf{m}_i^1,\\dots ,\\mathbf{m}_i^{N_T}\\}\\) from image feature maps output by a 2D CNN. Then, each query descriptor is correlated with the feature maps of all \\(N_{I}\\) frames at different spatial resolutions, which constructs a cost-volume pyramid. Flattening the latter yields tokens \\(V\\in \\mathbb{R}^{N_T\\times N_I\\times C}\\), where \\(C\\) is the total number of elements in the cost-volume pyramid. Feeding the tokens to a Transformer, we obtain tracks \\(\\mathcal{T} = \\{T^j\\}_{j = 1}^{N_T}\\). Recall that each track \\(T^j\\) comprises the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.225, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Architecture of Tracker T. We adopt a coarse-to-fine design for the tracker. The coarse tracker first locates the approximate positions of corresponding points, and the fine tracker then refines these initial predictions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.342 + ], + "angle": 0, + "content": "set of \\(N_I\\) tracked 2D locations \\(\\mathbf{y}_i^j\\) together with predicted visibility indicators \\(v_i^j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.448 + ], + "angle": 0, + "content": "It is worth noting that, differently from [22, 32], our tracker does not assume temporal continuity. Therefore, we avoid the sliding window approach and, instead, attend to all the input frames together. Furthermore, unlike in [32], we predict each track independently of others. This allows to track a larger number of points at test time leading to increased density of reconstructed point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Predicting tracking confidence In SfM, it is crucial to filter out any outlier correspondences as they can negatively impact the subsequent reconstruction stages. To this end, we enhance the tracker with the ability to estimate confidence for each track-point prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.702 + ], + "angle": 0, + "content": "More specifically, we leverage the aleatoric uncertainty [33, 51] model which predicts variance \\(\\sigma_{i}^{j}\\) together with each 2D track point \\(\\mathbf{y}_i^j\\), so that the resulting normal distribution \\(\\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)\\) tightly peaks around each ground-truth 2D track point \\(\\mathbf{y}_i^{j\\star}\\). Hence, during training, the \\(\\ell_1 / \\ell_2\\) loss, originally used in video point tracking, is replaced with the (negated) logarithm of the latter probability evaluated at each ground truth point \\(\\mathbf{y}_i^{j\\star}\\). Once trained, the confidence measure \\(1 / \\sigma_{i}^{j}\\) is proportional to the inverse of the predicted variance. In practice, we assume a diagonal covariance matrix resulting in horizontal and vertical uncertainties \\(\\sigma_{i}^{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Coarse-to-fine tracking Moreover, since SfM requires highly accurate (pixel or sub-pixel level) correspondences, we employ a coarse-to-fine point-tracking strategy. As described above, we first coarsely track image points using feature maps that fully cover the input images \\( I \\). Then, we form \\( P \\times P \\) patches by cropping input images around the coarse point estimates and execute the tracking again to obtain a sub-pixel estimate. Recall that, differently from the chained matching of traditional SfM, our tracker is fully differentiable. This enables back-propagating the gradient of the training loss \\( \\mathcal{L} \\) through the whole framework to the tracker parameters. This reinforces the synergy between the tracking and the ensuing stages, which are described next." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21689" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.419, + 0.108 + ], + "angle": 0, + "content": "3.3. Learnable camera & point initialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.468, + 0.265 + ], + "angle": 0, + "content": "As discussed above, a traditional SfM pipeline [39, 59] usually relies on an an incremental loop, which often initializes with a correspondence-rich image pair, gradually registers new frames, enlarges the point cloud, and conducts joint optimization (e.g., BA). However, although the framework has been fortified in robustness and accuracy through decades of improvements, this cumulative process has inevitably led to increased complexity. Furthermore, Incremental SfM is largely non-differentiable which complicates end-to-end learning from annotated data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.468, + 0.37 + ], + "angle": 0, + "content": "Thus, in pursuit of simplicity and differentiability, our method departs from the classical SfM scheme. Inspired by recent advances in deep camera pose estimators [36, 74, 85], we propose to initialize the cameras and the point cloud with a pair of deep Transformer [72] networks. Importantly, we register all cameras and reconstruct all scene points collectively in a non-incremental differentiable fashion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.468, + 0.419 + ], + "angle": 0, + "content": "Learnable camera initializer The predictor of initial cameras \\(\\hat{\\mathcal{P}}\\) is implemented as a deep Transformer architecture \\(\\mathfrak{T}_{\\mathcal{P}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.424, + 0.468, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} \\left(\\left\\{\\phi \\left(I _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I} \\right\\}, \\left\\{d ^ {\\mathcal {P}} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.468, + 0.639 + ], + "angle": 0, + "content": "It accepts a set of tokens comprising global ResNet50 [24] features \\(\\phi(I_i)\\) of input images \\(\\mathcal{I}\\), and the set of descriptors \\(d^{\\mathcal{P}}(\\mathbf{y}_i^j)\\) of track points \\(\\mathbf{y}_i^j \\in T_i \\in \\mathcal{T}\\). Here, each track descriptor is output by an auxiliary branch of the tracker \\(\\mathbb{T}\\). Given these inputs, \\(\\mathfrak{T}_{\\mathcal{P}}\\) first applies cross-attention between the global image feature (query) and the track-descriptor (key-value) pairs yielding \\(N_I = |\\mathcal{I}|\\) tokens per scene. The output of cross-attention is then concatenated with an embedding of a preliminary camera estimated by the 8-point algorithm taking the correspondences between track points \\(\\mathbf{y}_i^j\\) as input. Finally, we feed this concatenation to a Transformer trunk resulting in the initial cameras \\(\\hat{\\mathcal{P}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.703 + ], + "angle": 0, + "content": "Learnable triangulation Given initial cameras \\(\\hat{\\mathcal{P}}\\) and 2D tracks \\(\\mathcal{T}\\), the triangulator outputs the initial point cloud \\(\\hat{X}\\). Similar to the camera predictor, the triangulator is a Transformer \\(\\mathfrak{T}_X\\)" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.708, + 0.468, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\mathfrak {T} _ {X} \\left(\\left\\{d ^ {X} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.84 + ], + "angle": 0, + "content": "accepting descriptors \\(d^{X}(\\mathbf{y}_{i}^{j})\\) each comprising a tracker feature, and a positional harmonic embedding [47] of points \\(\\bar{\\mathbf{x}}^j\\in \\bar{X}\\) from a preliminary point cloud \\(\\bar{X}\\). The preliminary point cloud is formed via closed-form multi-view Direct Linear Transform (DLT) 3D triangulation [23] of the tracks \\(\\mathcal{T}\\) given the initial cameras \\(\\hat{\\mathcal{P}}\\). Please refer to the supplementary for a detailed description of both initializers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.26, + 0.864 + ], + "angle": 0, + "content": "3.4. Bundle adjustment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Given the tracks \\(\\mathcal{T}\\) (Sec. 3.2), initial cameras \\(\\hat{\\mathcal{P}}\\), and the initial point cloud \\(\\hat{X}\\) (Sec. 3.3), Bundle Adjustment BA mini" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.799, + 0.107 + ], + "angle": 0, + "content": "mizes the reprojection loss \\(\\mathcal{L}_{\\mathrm{BA}}[1,2,23,59]\\)" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.117, + 0.821, + 0.145 + ], + "angle": 0, + "content": "\\[\nX,\\mathcal{P} = \\operatorname {BA}(\\mathcal{T},\\hat{X},\\hat{\\mathcal{P}}) = \\operatorname *{arg min}_{X,\\mathcal{P}}\\mathcal{L}_{\\operatorname {BA}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.148, + 0.89, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B A}} = \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {\\mathbf {x}}} v _ {i} ^ {j} \\| P _ {i} \\left(\\mathbf {x} ^ {j}\\right) - y _ {i} ^ {j} \\|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.203, + 0.892, + 0.34 + ], + "angle": 0, + "content": "summing over all reprojection errors \\(\\| P_i(\\mathbf{x}^j) - \\mathbf{y}_i^j\\|\\) each comprising the distance between the projection \\(P_{i}(\\mathbf{x}^{j})\\) of the point cloud \\(\\mathbf{x}^j\\in X\\) to camera \\(P_{i}\\in \\mathcal{P}\\), and the \\(i\\)-th 2D point \\(\\mathbf{y}_i^j\\in T^j\\) of the track \\(T^j\\). Additionally, the error terms are filtered out if the corresponding points have low visibility, low confidence, or do not fit the geometric constraints defined by [59]. Points with large reprojections errors are also filtered [59, 63, 81]. More details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Differentiable Levenberg-Marquardt Following common practice [39, 59], we minimize Eq. (6) with second-order Levenberg-Marquardt (LM) optimizer [49]. However, optimizing the main training loss (Eq. (2)) via backpropagation requires differentiability of Eq. (6) which is nontrivial. Therefore, we leverage the recently proposed Theiss library [54] which exploits the implicit function theorem to backpropagate through deep networks with nested optimization loops." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.488, + 0.651, + 0.503 + ], + "angle": 0, + "content": "3.5. Method details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Camera parameterization Each camera pose \\( P \\in \\mathcal{P} \\) is parameterized with 8-degrees of freedom: the quaternion \\( q(R) \\in \\mathbb{R}^4 \\) of the rotation \\( R \\in \\mathbb{SO}(3) \\) and the translation \\( \\mathbf{t} \\in \\mathbb{R}^3 \\) components of \\( P \\)'s extrinsics \\( g \\in \\mathbb{SE}(3) \\), and the logarithm \\( \\ln(\\mathfrak{f}) \\in \\mathbb{R} \\) of the camera focal length \\( \\mathfrak{f} \\in \\mathbb{R}^+ \\). Given these values, the \\( 3 \\times 4 \\) pose matrix is defined as \\( P = KR[\\mathbb{I}_3|\\mathbf{t}] \\), with the calibration matrix \\( K = [\\mathfrak{f},0,p_x;0,\\mathfrak{f},p_y;0,0,1] \\subset \\mathbb{R}^{3 \\times 3} \\) (row major order). Following standard practice [39, 59], we set the principal-point coordinates \\( p_x, p_y \\in \\mathbb{R} \\) to the center of the image." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.666, + 0.884, + 0.681 + ], + "angle": 0, + "content": "Training loss The training loss \\(\\mathcal{L}\\) (Eq. (2)) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.693, + 0.89, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(f _ {\\theta} (\\mathcal {I}), \\mathcal {P} ^ {\\star}, \\mathcal {T} ^ {\\star}, X ^ {\\star}\\right) = \\sum_ {j = 1} ^ {N _ {T}} \\left| \\mathbf {x} ^ {\\star j} - \\mathbf {x} ^ {j} \\right| _ {\\epsilon} + \\left| \\mathbf {x} ^ {\\star j} - \\hat {\\mathbf {x}} ^ {j} \\right| _ {\\epsilon} + \\\\ + \\sum_ {i = 1} ^ {N _ {I}} e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, P _ {i}\\right) + e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, \\hat {P} _ {i}\\right) - \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {T}} \\log \\mathcal {N} \\left(\\mathbf {y} _ {i} ^ {j \\star} \\mid \\mathbf {y} _ {i} ^ {j}, \\sigma_ {i} ^ {j}\\right) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here, \\(|\\mathbf{x}^{\\star j} - \\mathbf{x}^{j}|_{\\epsilon}\\) and \\(|\\mathbf{x}^{\\star j} - \\hat{\\mathbf{x}}^{j}|_{\\epsilon}\\) evaluate the \\(\\epsilon\\)-thresholded pseudo-Huber loss \\(|\\cdot|_{\\epsilon}\\) [9] between the ground truth 3D points \\(\\mathbf{x}^{\\star j}\\) and the initial and BA-refined 3D points \\(\\hat{\\mathbf{x}}^{j} \\in X\\), \\(\\mathbf{x}^{j} \\in \\hat{X}\\) respectively. The camera errors \\(e_{\\mathcal{P}}(P_{i}^{\\star}, P_{i})\\) and \\(e_{\\mathcal{P}}(P_{i}^{\\star}, \\hat{P}_{i})\\) compare the predicted initial pose \\(\\hat{P}_{i} \\in \\hat{\\mathcal{P}}\\) and bundle-adjusted camera pose \\(P_{i} \\in \\mathcal{P}\\) to the ground truth camera annotation \\(P_{i}^{\\star} \\in \\mathcal{P}_{i}^{\\star}\\). Here, \\(e_{\\mathcal{P}}(P, P')\\) is defined as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21690" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.197, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.089, + 0.396, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.09, + 0.593, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.089, + 0.691, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.09, + 0.888, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.246, + 0.196, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.246, + 0.396, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.246, + 0.593, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.246, + 0.691, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.246, + 0.888, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.41, + 0.817, + 0.424 + ], + "angle": 0, + "content": "Figure 4. Camera and point-cloud reconstructions of VGGsM on Co3D (left) and IMC Phototourism (right)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.451, + 0.47, + 0.543 + ], + "angle": 0, + "content": "the Huber-loss \\(|\\cdot|_{\\epsilon}\\) between the parameterizations of poses \\(P, P'\\). Finally, \\(\\log \\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)\\) computes the likelihood of a ground-truth track point \\(\\mathbf{y}_i^{j\\star} \\in T_i^\\star\\) under a probabilistic track-point estimate defined by a 2D gaussian with mean and variance predictions \\(\\mathbf{y}_i^j\\) and \\(\\sigma_i^j\\) respectively (i.e. the aleatoric uncertainty model described in Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.208, + 0.579 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.469, + 0.635 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets together with the protocols for training and evaluation. Then, we provide comparison to existing methods and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.469, + 0.851 + ], + "angle": 0, + "content": "Datasets. Following prior work [25, 39, 74], we evaluate camera pose estimation on Co3Dv2 [55] and IMC Phototourism datasets [31], and 3D triangulation on ETH3D [60]. Co3D is an object-centric dataset comprising turntable-style videos from 51 categories of MS COCO [38]. IMC Phototourism, provided by Image Matching Challenge [31], contains 8 testing scenes and 3 validation scenes of famous landmarks. Generally, the Co3D scenes have much wider baselines, making them challenging for traditional frameworks such as COLMAP, while the IMC samples often have sufficiently overlapping fursta, which is where COLMAP excels. ETH3D provides highly accurate point clouds (captured by laser scanner) for 13 indoor and outdoor scenes, and hence is suitable for the evaluation of triangulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Training. For the model evaluated on IMC Phototourism and ETH3D, we follow the protocol of [25, 39, 57] and train on the MegaDepth dataset [35]. MegaDepth con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.557 + ], + "angle": 0, + "content": "tains 1M crowd-sourced images depicting 196 tourism landmarks, auto-annotated with SfM tools. Hyper-parameters are tuned using the IMC validation set. As in prior work [25, 40], some scenes of MegaDepth are excluded from training due to their low quality or due to overlap with the IMC test set. For Co3Dv2, we conduct training and evaluation on 41 categories as in [36, 74, 85]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.712 + ], + "angle": 0, + "content": "We chose a multi-stage training strategy for better stability. We first train the tracker T on the synthetic Kubric dataset [21] following the training protocol of [15, 32]. Then, the tracker is fine-tuned solely on Co3D or MegaDepth, depending on the test dataset. Subsequently, we train the camera initializer, with the tracker frozen. Next, the triangulator is trained with the aforementioned two components held frozen. Finally, all components are trained end-to-end. In all stages, we randomly sample a training batch of 3 to 30 frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Testing. Given input test frames \\(\\mathcal{I}\\), we first select the query frame by identifying the image that is closest to all others based on the cosine similarity between global descriptors extracted by an off-the-shelf ResNet50 [24]. Then, we extract SuperPoint and SIFT keypoints from the query frame to serve as the query points for the tracker \\(\\mathbb{T}\\). Although our method can track any query point, it performs better when the queries are distinctive. To improve accuracy, we iterate the whole reconstruction function \\(f_{\\theta}\\) multiple times until reaching sub-pixel BA reprojection error \\(\\mathcal{L}_{\\mathrm{BA}}\\). After the first iteration, the query image for each subsequent iteration is the one that is farthest from the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21691" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.09, + 0.182, + 0.105 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.107, + 0.241, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.206, + 0.205, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.09, + 0.647, + 0.105 + ], + "angle": 0, + "content": "Tracking Results" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.107, + 0.42, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.107, + 0.585, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.107, + 0.75, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.097, + 0.895, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.206, + 0.433, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.207, + 0.585, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.207, + 0.744, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.207, + 0.899, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.894, + 0.369 + ], + "angle": 0, + "content": "Figure 5. Qualitative Evaluation of Tracking. In each row, the left-most frame contains the query image with query points (crosses). The predicted track points \\(\\mathbf{y}_i^j\\) (dots) are shown to the right. The top-right part also highlights our track-point confidence predictions (described in Sec. 3.2), illustrated as ellipses with extent proportional to the predicted variance \\(\\sigma_i^j\\). Note how the uncertainty corresponds to an expectation, e.g., a keypoint covering a vertical stripe has higher uncertainty along the \\(y\\)-axis." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.381, + 0.871, + 0.456 + ], + "angle": 0, + "content": "
Co3D DatasetIncremental SfMDeep
MethodCOLMAP (SP + SG)PixSfM (SP + SG)RelPose[85]PoseReg[74]RelPose++[36]PoseDiffusion[74]Ours w/o JointOurs
RRE@15°31.633.757.153.282.380.588.292.1
RTE@15°27.332.9-49.177.279.883.488.3
AUC@30°25.330.1-45.065.166.570.774.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.467, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Table 1. Camera Pose Estimation on Co3D, where the proposed method outperforms previous methods by a large margin. Ours w/o Joint indicates the variant of our framework without training all components jointly." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.519, + 0.468, + 0.646 + ], + "angle": 0, + "content": "
IMC DatasetMethodAUC@3°AUC@5°AUC@10°
DeepDeepSFM10.2719.3631.35
PoseDiffusion12.3123.1736.82
Incremental SfMCOLMAP (SIFT+NN)23.5832.6644.79
PixSfM (SIFT + NN)25.5434.8046.73
PixSfM (LoFTR)44.0656.1669.61
PixSfM (SP + SG)45.1957.2270.47
DFSfM (LoFTR)46.5558.7472.19
DeepOurs w/o Joint38.2351.6068.35
Ours45.2358.8973.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Table 2. Camera Pose Estimation on IMC. Our method achieves better accuracy than state-of-the-art Incremental SfM approaches on 2 out of 3 AUC thresholds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.788 + ], + "angle": 0, + "content": "query image of the previous iteration (as measured by the ResNet50 cosine similarity). The re-projections of the point cloud from the current iteration initialize the tracking in the next iteration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.3, + 0.817 + ], + "angle": 0, + "content": "4.1. Camera pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Following [25, 31, 39, 74], we report the metric area-under-curve (AUC) to evaluate camera pose accuracy. In Co3D, similar to [74], we also report the relative rotation error (RRE) and relative translation error (RTE). More specifically, for every pair of input frames, we compute the angular" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.674 + ], + "angle": 0, + "content": "translation and rotation error, which are later compared to a threshold yielding accuracies RTE and RRE respectively. For a range of thresholds, AUC picks the lower between RRE and RTE, and outputs the area under the accuracy-threshold curve. The results on IMC and CO3D are presented in Tab. 1 and Tab. 2 respectively. For a fair comparison on IMC, we finetune DeepSFM [78] and PoseDiffusion [74] on MegaDepth using their open-source code. The results of Incremental SfM methods are copied from Detector-free SfM [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Results indicate that VGGSfM outperforms existing methods by a large margin (+9 accuracy points for each metric) on the CO3D dataset. Here, traditional SfM pipelines suffer because of the wide baselines between test frames. On IMC, with a good overlap between views, traditional SfM remains superior to recent data-driven deep-learning pipelines [74, 78]. Our end-to-end trained VGGSfM, however, outperforms all other methods on AUC@10 and AUC@5, and ranks second on AUC@3, convincingly demonstrating its ability to perform well in both narrow- and wide-baseline regimes. The accuracy and completeness of our point clouds can be further qualitatively evaluated in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.468, + 0.205 + ], + "angle": 0, + "content": "
MethodAccuracy (%)Completeness (%)
1cm2cm5cm1cm2cm5cm
PatchFlow (LoFTR)66.7378.7389.933.4811.3430.96
PixSfM (LoFTR)74.4284.0892.632.919.3927.31
PixSfM (SIFT + NN)76.1885.6093.160.170.713.29
PixSfM (SP + SG)79.0187.0493.800.752.7711.28
DFSfM (LoFTR)80.3889.0195.833.7311.0729.54
Ours80.6289.4996.524.5213.1133.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Table 3. 3D Triangulation on ETH3D [60] reporting the accuracy and completeness metrics at different thresholds." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.276, + 0.241, + 0.292 + ], + "angle": 0, + "content": "4.2.3D triangulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.466 + ], + "angle": 0, + "content": "We evaluate the accuracy and completeness of 3D triangulation on the ETH3D dataset [60] using the same protocol as [17, 25, 39], which triangulates points with fixed camera poses and intrinsics. Results are shown in Tab. 3, where metrics are averaged over all scenes. Our VGGsFm achieves better accuracy and completeness than all baselines (PatchFlow [17], PixSfM [39], and DFSfM [25]), regardless of which keypoint detection or matching method they use. This is especially obvious from the completeness attained at the \\(5\\mathrm{cm}\\) threshold, with our \\(33.96\\%\\) compared to \\(29.54\\%\\) of the best prior work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.475, + 0.227, + 0.491 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.635 + ], + "angle": 0, + "content": "End-to-end Training. As reported in Tab. 1 and Tab. 2, the end-to-end joint training of the whole framework is important for achieving state-of-the-art performance. Specifically, comparing VGGsF to an ablation which lacks end-to-end training (Ours w/o joint) we record an improvement from \\(70.7\\%\\) AUC@30 to \\(74.0\\%\\) on the Co3D dataset, and \\(68.35\\%\\) AUC@10 to \\(73.92\\%\\) on the IMC dataset. This demonstrates the benefits of our fully-differentiable design, and the synergy between its components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Tracking or Pairwise Matching. We compare the performance of our predicted tracks to pairwise matching methods on the IMC dataset. Specifically, we split our 2D tracks into pairwise matches and feed these matches to PixSfM. Also, we construct 2D tracks by pairwise matching (based on the open-source implementation of PixSfM) and feed them to our framework. It is worth noting that tracks from pairwise matching have a lot of \"holes\" because pairwise matching cannot guarantee proper point tracking. We fix these holes by setting their locations as the point locations in the query frame, marking them as invisible. At the same time, for fair comparison, cameras are still initialized using our tracks, because SP+SG cannot provide track features to our camera initializer. The results are shown in Tab. 4. Although COLMAP (the basis of PixSfM) is designed and carefully engineered for pairwise matching, our tracks achieve a slightly better result than the state-of-the-art matching op" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.891, + 0.123 + ], + "angle": 0, + "content": "
PixSfM(SP + SG)PixSfM(Our Tracks)Ours(SP + LG)Ours
AUC@10°70.4770.6268.7873.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.133, + 0.892, + 0.174 + ], + "angle": 0, + "content": "Table 4. Tracking or Pairwise Matching. We respectively provide tracks predicted by our tracker or matches estimated by SP+SG to PixSfM and to our VGGSfM." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.188, + 0.851, + 0.236 + ], + "angle": 0, + "content": "
PoseDiffusionOur Camera Initialized
DLT62.1869.42
Our Triangulator66.3773.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.246, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Table 5. Ablation Study for Camera Initializer and Triangulator. A clear performance drop is observed when replacing our camera initializer by deep camera prediction method PoseDiffusion, or replacing triangulator by DLT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.404 + ], + "angle": 0, + "content": "tion \\(\\mathrm{SP + SG}\\). Instead, directly feeding \\(\\mathrm{SP + SG}\\) tracks to our framework leads to a performance drop. We attribute this to the fact that using \\(\\mathrm{SP + SG}\\) tracks cannot benefit from the joint training. We also provide a qualitative evaluation of our tracking accuracy in Fig. 5 on the IMC dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Camera Controller and Triangulator. We also validate the design of our camera controller \\(\\mathfrak{T}_{\\mathcal{P}}\\) and triangulator \\(\\mathfrak{T}_X\\) on the IMC dataset. As reported in Tab. 5, AUC@10 drops clearly if we replace them with alternatives, proving that they provide sufficiently accurate initialization for our global bundle adjustment BA, without the need for incremental camera registration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Coarse-to-fine Tracking. As discussed above, accurate correspondences are important for structure from motion. We demonstrate the significance of our coarse-to-fine tracking mechanism for the method performance. By conducting an ablation study where the fine tracker is removed, we observe a significant performance drop on the IMC dataset, with AUC@10 dropping from \\(73.92\\%\\) to \\(62.30\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.619, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this paper, we have presented VGGSfM, a fully differentiable SfM approach. We find that even long-standing pipelines, such as Structure-from-Motion, benefit from a learned adaptation between their components. This allows VGGSfM to be simpler than traditional SfM frameworks while achieving better performance across benchmark datasets. Moreover, our framework is fully implemented in Python, which will allow for easy modification and improvements in the future. While VGGSfM already achieves good performance, it cannot yet compete with established pipelines in all application domains. For example, it currently lacks the capability to process thousands of images as in traditional SfM frameworks. Nonetheless, we find differentiable SfM a promising direction of research, and our approach lays the foundation for further advances." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Noah Snavely, Steven M Seitz, and Richard Szeliski. Bundle adjustment in the large. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 29-42. Springer, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[3] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.469, + 0.386 + ], + "angle": 0, + "content": "[5] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4322-4331, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[7] Lucas Brynte, José Pedro Iglesias, Carl Olsson, and Fredrik Kahl. Learning structure-from-motion with graph attention networks. arXiv preprint arXiv:2308.15984, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[8] Jonathan L. Carrivick, Mark W. Smith, and Duncan J. Quincey. Structure from Motion in the Geosciences. John Wiley & Sons, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[9] Pierre Charbonnier, Laure Blanc-féaud, Gilles Aubert, and Michel Barlaud. Deterministic edge-preserving regularization in computed imaging. IEEE Trans. Image Processing, 6:298-311, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[11] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[12] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[13] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. In Proceedings of the IEEE International Conference on Computer Vision, pages 864-872, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurlPS Datasets Track, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Mihai Dusmanu, Johannes L Schonberger, and Marc Pollefeys. Multi-view optimization of local feature geometry. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 670-686. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[19] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 368-381. Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[20] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Towards Internet-scale multi-view stereo. In Proc. CVPR. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliaasacchi. Kubric: a scalable dataset generator. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[22] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[23] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. arXiv preprint arXiv:1512.03385, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In *arxiv*, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Jared Heinly, Johannes L. Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world* in six days \\((\\ast)\\) (as captured by the yahoo 100 million image dataset). In" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21694" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[27] Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Piermattei, James O'Connor, and Jacqueline Rosette. Structure from motion photogrammetry in forestry: A review. Current Forestry Reports, 5:155-168, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.223 + ], + "angle": 0, + "content": "[28] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. ArXiv, 2212.04492, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[29] San Jiang, Cheng Jiang, and Wanshou Jiang. Efficient structure from motion for large-scale UAV images: A review and a comparison of sfm tools. ISPRS Journal of Photogrammetry and Remote Sensing, 167:230-251, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[30] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[31] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[32] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Alex Kendall and Yarin Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Proc. NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[35] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[36] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In Proc. ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[39] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. arXiv.cs, abs/2108.08291, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023. 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[41] Yin Lou, Noah Snavely, and Johannes Gehrke. Matchminer: Efficient spanning structure mining in large image collections. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12, pages 45-58. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[42] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[43] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[44] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[45] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[46] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust Wide Baseline Stereo from Maximally Stable Extremal Regions. In Proc. BMVC, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Proc. ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[49] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[50] Pierre Moulon, Pascal Monasse, and Renaud Marlet. Global fusion of relative motions for robust, accurate and scalable structure from motion. In Proceedings of the IEEE international conference on computer vision, pages 3248-3255, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[51] David Novotny, Diane Larlus, and Andrea Vedaldi. Learning 3d object categories by looking around them. In Proc. ICCV, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[52] John Oliensis. A critique of structure-from-motion algorithms. Computer Vision and Image Understanding, 80(2): 172-214, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[53] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[54] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, et al. Theses: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 35:3801-3818, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21695" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[56] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[58] Frederik Schaffalitzky and Andrew Zisserman. Multi-view Matching for Unordered Image Sets, or \"How Do I Organize My Holiday Snaps?\" In Proc. ECCV, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[59] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[60] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[61] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[62] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[63] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[64] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[65] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. arXiv preprint arXiv:1806.04807, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[66] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[67] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[68] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[69] Bill Triggs, Philip F. McLauchlan, Richard I. Hartley, and Andrew W. Fitzgibbon. Bundle Adjustment - A Modern Synthesis. In Proc. ICCV Workshop, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[70] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[71] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[72] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. NeurIPS, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[73] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[74] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[75] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[76] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[77] Tong Wei, Yash Patel, Alexander Shekhotsov, Jiri Matas, and Daniel Barath. Generalized differentiable ransac. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17649-17660, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[78] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 230–247. Springer, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[79] Matthew J Westoby, James Brasington, Niel F Glasser, Michael J Hambrey, and Jennifer M Reynolds. 'structure-from-motion'photogrammetry: A low-cost, effective tool for geoscience applications. Geomorphology, 179:300-314, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[80] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 61-75. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[81] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21696" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[82] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[83] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3d animals in the wild. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[84] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned Invariant Feature Transform. In Proc. ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[85] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[86] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[87] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21697" + } + ] +] \ No newline at end of file diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_origin.pdf b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4339af813bafef21fe2cfe4159b110614e22394c --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/d2ee9e08-83a5-42e7-ad6b-08e958fae9fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca6cb510989a0b5cae42b9d1d6442c578b029166e040a0b0186fd1e50e9897b3 +size 9156017 diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/full.md b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d308549eb660dabd84aace408b6f17dc756232e --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/full.md @@ -0,0 +1,387 @@ +# VGGSfM: Visual Geometry Grounded Deep Structure From Motion + +Jianyuan Wang $^{1,2}$ + +Nikita Karaev 1,2 + +Christian Rupprecht1 + +David Novotny2 + +Visual Geometry Group, University of Oxford + +Meta AI + +https://vggsfm.github.io + +# Abstract + +Structure-from-motion (SfM) is a long-standing problem in the computer vision community, which aims to reconstruct the camera poses and 3D structure of a scene from a set of unconstrained 2D images. Classical frameworks solve this problem in an incremental manner by detecting and matching keypoints, registering images, triangulating 3D points, and conducting bundle adjustment. Recent research efforts have predominantly revolved around harnessing the power of deep learning techniques to enhance specific elements (e.g., keypoint matching), but are still based on the original, non-differentiable pipeline. Instead, we propose a new deep pipeline VGGsfM, where each component is fully differentiable and thus can be trained in an end-to-end manner. To this end, we introduce new mechanisms and simplifications. First, we build on recent advances in deep 2D point tracking to extract reliable pixel-accurate tracks, which eliminates the need for chaining pairwise matches. Furthermore, we recover all cameras simultaneously based on the image and track features instead of gradually registering cameras. Finally, we optimise the cameras and triangulate 3D points via a differentiable bundle adjustment layer. We attain state-of-the-art performance on three popular datasets, CO3D, IMC Phototourism, and ETH3D. + +# 1. Introduction + +Reconstructing the camera parameters and the 3D structure of a scene from a set of unconstrained 2D images is a long-standing problem in the computer vision community. Among many other applications [8, 27, 29, 53, 79], it has recently emerged as an important component of learning neural fields [10, 28, 34, 37, 48, 76]. The problem is usually solved via the Structure-from-Motion (SfM) framework which estimates the 3D point cloud (Structure) and the parameters of each camera (Motion) in the scene. State-of-the-art methods [25, 39] follow the incremental SfM paradigm whose origins can be traced back to the early 2000s [23, 58]. It usually begins with a small set of + +![](images/442374b47f7f05fbfd05936a49654d12e904f4edefe9b5e38edb91c9d73bc601.jpg) +Figure 1. Reconstruction of In-the-wild Photos with VGGsFm, displaying estimated point clouds (in blue) and cameras (orange). + +![](images/936d5f7f73bd041253a5fa48db6febd10820525639041254582ea1446a59bce0.jpg) + +correspondence-rich images as initialization, and gradually adds more views into the reconstruction, through keypoint detection, matching, verification, image registration, triangulation, bundle adjustment (BA), and so on [2, 20, 26, 59]. + +Recent research efforts have predominantly revolved around leveraging the power of deep learning techniques to + +![](images/ae5f31ae48b8f47b1f507ddcca7d928cd84069cdca49fbdcc7f7e8df73fbb707.jpg) +Figure 2. Overview of VGGs fM. Our method extracts 2D tracks from input images, reconstructs cameras using image and track features, initializes a point cloud based on these tracks and camera parameters, and applies a bundle adjustment layer for reconstruction refinement. The whole framework is fully differentiable and designed for end-to-end training. + +enhance specific elements within the original pipeline while preserving the incremental SfM framework as a whole. For instance, SuperPoint and SuperGlue [14, 57] focus on improving keypoint detection and matching. Pixel-perfect SfM [39] proposes deep feature-metric refinement to adjust both keypoints and bundles. Detector-free feature matching methods [12, 64, 75] bypass early keypoint detection by means of attention, which is powerful in poorly textured scenes. Detector-free SfM [25] builds a coarse SfM model through quantized detector-free matches and then iteratively refines it with multi-view consistency constraints. These advancements successfully combine deep learning approaches (such as deep feature matching) with well-established hand-engineered components, such as the incremental camera registration of COLMAP [59]. + +The widespread success of end-to-end training warrants the question of what benefits it can bring to long-standing frameworks such as SfM. Naturally, it is often difficult to assess the merits of new approaches when compared with decades of continuous improvements. Nonetheless, in this paper, we answer this question by introducing a fully-differentiable SfM pipeline, dubbed Visual Geometry Grounded Deep Structure From Motion (VGGSfM), which trains in an end-to-end manner. We find that this allows the pipeline to be simpler than prior frameworks while achieving better or comparable performance. Training end-to-end allows each component to generate outputs that facilitate the task of its successor. + +To build a fully-differentiable pipeline, we make several substantial changes to the SfM procedures and overall obtain better performance. Specifically, our model builds on recent advances in deep 2D point tracking [15, 16, 22, 32] to directly extract reliable pixel-accurate tracks. This simplifies the correspondence estimation step in traditional SfM, which first estimates pairwise matches and then connects them into tracks. Then, based on the image and track fea + +tures, VGGsFm estimates all cameras jointly via a Transformer [72], and subsequently all 3D points. Different from Incremental SfM, this approach is simpler and easier to differentiate as it does not depend on a discrete, combinatorial correspondence chaining step. Finally, for bundle adjustment, we replace the commonly employed non-differentiable Ceres solver [3] with the fully differentiable second-order Theseus solver [54]. + +Hence, we fuse all the SfM components into a single fully differentiable reconstruction function $f$ . Besides that, in our experiments, we also show that the individual modules perform well in isolation. Ultimately, end-to-end training yields another performance improvement, that surpasses the performance of isolated components. + +We evaluate VGGsFm for the task of camera pose estimation on the CO3Dv2 [55] and IMC Phototourism [31] datasets, and for 3D triangulation on the ETH3D [60] dataset. Our method attains strong performance on all benchmarks. At the same time, we conduct in-the-wild reconstruction to validate the generalization ability of our proposed framework, as shown in Fig. 1. + +# 2. Related Work + +Structure from Motion is a fundamental problem in computer vision and has been investigated for decades [23, 52, 53]. The classical pipelines usually solve the SfM problem in a global [13, 50, 80] or incremental [2, 19, 59, 63, 81] manner. Both of which are usually based on pairwise image keypoint matching. Incremental SfM is arguably the most widely adopted strategy (e.g., the popular framework COLMAP [59]). Therefore, in the following sections, we refer to incremental SfM as "classical" or "traditional" SfM. We defer the discussion of global SfM to the supplementary. + +Traditional SfM frameworks often start by detecting keypoints and feature descriptors [4, 42, 43, 46]. They then + +search for image pairs with overlapping frusta by matching these keypoints across different images (e.g., with a nearest-neighbour search) [2, 41, 59]. These image pairs are further verified via two-view epipolar geometry or homography [23] through RANSAC [18]. Then, a pair or a small set of images is carefully selected for initialization. New images are gradually registered by solving the Perspective- $n$ -Point (PnP) problem [44], followed by triangulating 3D points, and bundle adjustment [69]. This process is iterated until all the frames are either registered or discarded. 2D correspondences (multi-view tracks) are the basis of the whole process, however, they are usually simply constructed by chaining two-view matches [59]. + +Many deep-learning approaches have been proposed to enhance this framework. For example, [14, 70, 84] provide better keypoint detection and [11, 30, 40, 57, 61] focus on matching. Furthermore, detector-free matching methods [12, 64, 75] propose to avoid sparse keypoint detection by building semi-dense matches via self and cross attention. [7] solves the subproblems of SfM by graph attention networks. Some studies improve the performance of RANSAC by making it trainable [5, 6, 77]. Recent state-of-the-art methods are PixSfM [39] and the concurrent Detector-free SfM (DFSfM) [25]. PixSfM refines the tracks and structure estimated by COLMAP through feature-metric keypoint adjustment and feature-metric bundle adjustment. Detector-free SfM proposes to first build a coarse SfM model using detector-free matches and COLMAP (or other frameworks), and then to iteratively refine the tracks and the structure of the coarse model by enforcing multi-view consistency. + +Recently, fully differentiable SfM pipelines have also been explored. They usually use deep neural networks to regress camera poses and depths [65, 66, 68, 71, 73, 78, 87]. Although using an approximation of bundle adjustment [65, 68, 78], these methods suffer from limited generalizability and scalability (very few input frames) [65, 71, 73, 78, 87], or rely on temporal relationship [66, 68]. Meanwhile, some methods are category-specific [45, 82, 83]. The recent efforts on deep camera pose estimation can scale up to more than 50 frames, but they do not reconstruct the scene [36, 62, 74, 85]. + +Point tracking. Since VGGSfM proposes a novel point track predictor, next, we review recent advances in this field. Inspired by the optical-flow architecture of RAFT [67], PIPs [22] revisited point tracking, a task related to Particle Video [56], and proposed a highly accurate tracker of isolated points in a video. TAP-Vid [15] (i.e., "Tracking Any Point") introduced a benchmark for point tracking and a baseline model, which was later improved in TAPIR [16] by integrating the iterative update mechanism from PIPs. PointOdyssey [86] simplified PIPs and proposed a benchmark for the long-term version of point tracking. CoTracker [32] closed the gap between single point track + +ing and dense Optical Flow with joint point tracking. However, these works are designed for videos, i.e. temporally-ordered sequences of frames. In our point tracker, given the input frames are unordered, we do not assume a temporal relationship between input frames. We therefore process all frames jointly, avoiding windowed inference of [32]. Since SfM relies on highly accurate correspondences, our tracks are further refined in a coarse-to-fine manner to achieve subpixel accuracy. + +# 3. Method + +In this section, we describe the components of VG-GSfM and how they are composed in a fully differentiable pipeline. An overview of our framework is shown in Fig. 2. + +Problem setting Given a set of free-form images observing a scene, VGGSfM estimates their corresponding camera parameters and the 3D scene shape represented as a point cloud. Formally, given a tuple $\mathcal{I} = (I_1,\dots,I_{N_I})$ of $N_{I}\in \mathbb{N}$ RGB images $I_{i}\in \mathbb{R}^{3\times H\times W}$ , VGGSfM estimates the corresponding camera projection matrices $\mathcal{P} = (P_{1},\ldots ,P_{N_{I}}|P_{i}\subset \mathbb{R}^{3\times 4})$ and the scene cloud $X = \{\mathbf{x}^j\}_{j = 1}^{N_{\mathbf{x}}} \text{ of } N_{\mathbf{x}}\in \mathbb{N}$ 3D points $\mathbf{x}^j\in \mathbb{R}^3$ . Each projection matrix $P_{i}$ consists of extrinsic (pose) $g_{i}\in \mathbb{SE}(3)$ and intrinsics $K_{i}\in \mathbb{R}^{3\times 3}$ . + +A 3D point $\mathbf{x}^j$ can be projected to the $i$ -th camera yielding a 2D screen coordinate $\mathbf{y}_i^j = P_i(\mathbf{x}^j) \sim \lambda K_i \hat{\mathbf{x}}_i^j$ ; $\lambda \in \mathbb{R}_+$ , where $\hat{\mathbf{x}}_i^j = g_i \mathbf{x}^j$ is the world coordinate $\mathbf{x}^j$ expressed in view-coordinates of the $i$ -th camera. The projection of the point $\mathbf{x}^j$ to all input cameras is a track $T^j = ((y_1^j, v_1^j) \dots, (y_{N_I}^j, v_{N_I}^j))$ consisting of $N_I$ matching 2D points $\mathbf{y}_i^j \in \mathbb{R}^2$ , and their corresponding binary indicators $v_i^j \in \{0, 1\}$ denoting visibility of the $j$ -th point in the $i$ -th camera. We denote $\mathcal{T}_i = \{T_i^1, \dots, T_i^{N_T}\}$ as the set of all tracks $T_i^j$ in the $i$ -th camera. + +# 3.1. Method overview + +VGGsM implements SfM via a single function $f_{\theta}$ + +$$ +f _ {\theta} (\mathcal {I}) = \mathcal {P}, X \tag {1} +$$ + +accepting the set of $N_I$ scene images $\mathcal{I}$ and outputting the camera parameters $\mathcal{P}$ and the scene point cloud $X$ . Importantly, $f_{\theta}$ is fully differentiable and, as such, its parameters $\theta$ are learned by minimizing the training loss $\mathcal{L}$ : + +$$ +\theta^ {\star} = \arg \min _ {\theta} \sum_ {s = 1} ^ {S} \mathcal {L} \left(f _ {\theta} \left(\mathcal {I} _ {s}\right), \mathcal {P} _ {s} ^ {\star}, \mathcal {T} _ {s} ^ {\star}, X _ {s} ^ {\star}\right), \tag {2} +$$ + +summing over $S \in \mathbb{N}$ training image sets $\mathcal{I}_s$ annotated with ground-truth cameras $\mathcal{P}_s^\star$ , tracks $\mathcal{T}_s^\star$ , and point clouds $X_s^\star$ . We defer the details of $\mathcal{L}$ to Sec. 3.5 and, in the following paragraphs, discuss the architecture of $f_{\theta}$ . + +The reconstruction function Following traditional SfM [59], VGGSfM decomposes the reconstruction function $f_{\theta}$ into four seamless stages: 1) point tracking $\mathbb{T}$ , 2) initial camera estimator $\mathfrak{T}_{\mathcal{P}}$ , 3) triangulator $\mathfrak{T}_X$ and, 4) Bundle Adjustment BA, as follows: + +$$ +\mathcal {T} = \mathrm {T} (\mathcal {I}) +$$ + +$$ +\begin{array}{l} \hat {\mathcal {P}} = \mathfrak {T} _ {\mathcal {P}} (\mathcal {I}, \mathcal {T}) \\ \hat {\hat {\tau}} = \mathfrak {T} _ {\hat {\tau}} (\mathcal {\hat {T}}, \hat {\hat {\tau}}) \end{array} \tag {3} +$$ + +$$ +\hat {X} = \mathfrak {T} _ {X} (\mathcal {T}, \hat {\mathcal {P}}) +$$ + +$$ +\mathcal {P}, X = _ {\mathrm {B A}} (\mathcal {T}, \hat {\mathcal {P}}, \hat {X}). +$$ + +The track predictor $\mathbb{T}$ , hereafter referred to as "tracker", estimates 2D tracks $\mathcal{T}$ given input images $\mathcal{I}$ . Subsequently, $\mathfrak{T}_{\mathcal{P}}$ and $\mathfrak{T}_X$ provide initial cameras $\mathcal{P}$ and an initial point cloud $\hat{X}$ respectively. Finally, BA enhances accuracy by refining the cameras and 3D points together. + +# 3.2. Tracking + +Establishing precise 2D correspondences is important for achieving accurate 3D reconstruction. Traditionally, SfM frameworks first estimate pairwise image-to-image correspondences that are later chained into multi-image tracks $T$ [25, 39, 59]. Here, typically only point-pair matching benefits from learned components [14, 40, 57, 70], while the chaining of pairwise correspondences remains a hand-engineered process. + +Instead, VGGSfM significantly simplifies SfM correspondence tracking by employing a deep feed-forward tracking function. It accepts a collection of images and directly outputs a set of reliable point trajectories across all images. We achieve this by exploiting recent advances in video point tracking methods [15, 16, 22, 32]. Although developed for video-point tracking, these methods are inherently appropriate for SfM which requires a very compact set of highly accurate tracks (e.g., dense optical flow is too memory-demanding). Furthermore, point trackers mitigate the potential errors (sometimes called drift) caused by chaining of pairwise matches. However, as we describe below, our design differs from video trackers because SfM, which accepts free-form images, cannot assume temporal smoothness or ordering, and requires sub-pixel accuracy. + +Tracker architecture The design of our tracker $\mathbb{T}$ follows [22, 32], and is illustrated in Fig. 3. More specifically, given $N_{T}$ query points $\{\hat{\mathbf{y}}_i^1,\dots \hat{\mathbf{y}}_i^{N_T}\}$ in a frame $I_{i}$ we bilinearly sample their corresponding query descriptors $\{\mathbf{m}_i^1,\dots ,\mathbf{m}_i^{N_T}\}$ from image feature maps output by a 2D CNN. Then, each query descriptor is correlated with the feature maps of all $N_{I}$ frames at different spatial resolutions, which constructs a cost-volume pyramid. Flattening the latter yields tokens $V\in \mathbb{R}^{N_T\times N_I\times C}$ , where $C$ is the total number of elements in the cost-volume pyramid. Feeding the tokens to a Transformer, we obtain tracks $\mathcal{T} = \{T^j\}_{j = 1}^{N_T}$ . Recall that each track $T^j$ comprises the + +![](images/76d43148b58f308457f880f40b291753d2cd49c2f7b25858f343e5d406bd36e2.jpg) +Figure 3. Architecture of Tracker T. We adopt a coarse-to-fine design for the tracker. The coarse tracker first locates the approximate positions of corresponding points, and the fine tracker then refines these initial predictions. + +set of $N_I$ tracked 2D locations $\mathbf{y}_i^j$ together with predicted visibility indicators $v_i^j$ . + +It is worth noting that, differently from [22, 32], our tracker does not assume temporal continuity. Therefore, we avoid the sliding window approach and, instead, attend to all the input frames together. Furthermore, unlike in [32], we predict each track independently of others. This allows to track a larger number of points at test time leading to increased density of reconstructed point clouds. + +Predicting tracking confidence In SfM, it is crucial to filter out any outlier correspondences as they can negatively impact the subsequent reconstruction stages. To this end, we enhance the tracker with the ability to estimate confidence for each track-point prediction. + +More specifically, we leverage the aleatoric uncertainty [33, 51] model which predicts variance $\sigma_{i}^{j}$ together with each 2D track point $\mathbf{y}_i^j$ , so that the resulting normal distribution $\mathcal{N}(\mathbf{y}_i^{j\star}|\mathbf{y}_i^j,\sigma_i^j)$ tightly peaks around each ground-truth 2D track point $\mathbf{y}_i^{j\star}$ . Hence, during training, the $\ell_1 / \ell_2$ loss, originally used in video point tracking, is replaced with the (negated) logarithm of the latter probability evaluated at each ground truth point $\mathbf{y}_i^{j\star}$ . Once trained, the confidence measure $1 / \sigma_{i}^{j}$ is proportional to the inverse of the predicted variance. In practice, we assume a diagonal covariance matrix resulting in horizontal and vertical uncertainties $\sigma_{i}^{j}$ . + +Coarse-to-fine tracking Moreover, since SfM requires highly accurate (pixel or sub-pixel level) correspondences, we employ a coarse-to-fine point-tracking strategy. As described above, we first coarsely track image points using feature maps that fully cover the input images $I$ . Then, we form $P \times P$ patches by cropping input images around the coarse point estimates and execute the tracking again to obtain a sub-pixel estimate. Recall that, differently from the chained matching of traditional SfM, our tracker is fully differentiable. This enables back-propagating the gradient of the training loss $\mathcal{L}$ through the whole framework to the tracker parameters. This reinforces the synergy between the tracking and the ensuing stages, which are described next. + +# 3.3. Learnable camera & point initialization + +As discussed above, a traditional SfM pipeline [39, 59] usually relies on an an incremental loop, which often initializes with a correspondence-rich image pair, gradually registers new frames, enlarges the point cloud, and conducts joint optimization (e.g., BA). However, although the framework has been fortified in robustness and accuracy through decades of improvements, this cumulative process has inevitably led to increased complexity. Furthermore, Incremental SfM is largely non-differentiable which complicates end-to-end learning from annotated data. + +Thus, in pursuit of simplicity and differentiability, our method departs from the classical SfM scheme. Inspired by recent advances in deep camera pose estimators [36, 74, 85], we propose to initialize the cameras and the point cloud with a pair of deep Transformer [72] networks. Importantly, we register all cameras and reconstruct all scene points collectively in a non-incremental differentiable fashion. + +Learnable camera initializer The predictor of initial cameras $\hat{\mathcal{P}}$ is implemented as a deep Transformer architecture $\mathfrak{T}_{\mathcal{P}}$ : + +$$ +\hat {\mathcal {P}} = \mathfrak {T} _ {\mathcal {P}} \left(\left\{\phi \left(I _ {i}\right) \mid I _ {i} \in \mathcal {I} \right\}, \left\{d ^ {\mathcal {P}} \left(y _ {i} ^ {j}\right) \mid \forall T _ {i} \in \mathcal {T} \forall y _ {i} ^ {j} \in T _ {i} \right\}\right). \tag {4} +$$ + +It accepts a set of tokens comprising global ResNet50 [24] features $\phi(I_i)$ of input images $\mathcal{I}$ , and the set of descriptors $d^{\mathcal{P}}(\mathbf{y}_i^j)$ of track points $\mathbf{y}_i^j \in T_i \in \mathcal{T}$ . Here, each track descriptor is output by an auxiliary branch of the tracker $\mathbb{T}$ . Given these inputs, $\mathfrak{T}_{\mathcal{P}}$ first applies cross-attention between the global image feature (query) and the track-descriptor (key-value) pairs yielding $N_I = |\mathcal{I}|$ tokens per scene. The output of cross-attention is then concatenated with an embedding of a preliminary camera estimated by the 8-point algorithm taking the correspondences between track points $\mathbf{y}_i^j$ as input. Finally, we feed this concatenation to a Transformer trunk resulting in the initial cameras $\hat{\mathcal{P}}$ . + +Learnable triangulation Given initial cameras $\hat{\mathcal{P}}$ and 2D tracks $\mathcal{T}$ , the triangulator outputs the initial point cloud $\hat{X}$ . Similar to the camera predictor, the triangulator is a Transformer $\mathfrak{T}_X$ + +$$ +\hat {X} = \mathfrak {T} _ {X} \left(\left\{d ^ {X} \left(y _ {i} ^ {j}\right) \mid \forall T _ {i} \in \mathcal {T} \forall y _ {i} ^ {j} \in T _ {i} \right\}\right) \tag {5} +$$ + +accepting descriptors $d^{X}(\mathbf{y}_{i}^{j})$ each comprising a tracker feature, and a positional harmonic embedding [47] of points $\bar{\mathbf{x}}^j\in \bar{X}$ from a preliminary point cloud $\bar{X}$ . The preliminary point cloud is formed via closed-form multi-view Direct Linear Transform (DLT) 3D triangulation [23] of the tracks $\mathcal{T}$ given the initial cameras $\hat{\mathcal{P}}$ . Please refer to the supplementary for a detailed description of both initializers. + +# 3.4. Bundle adjustment + +Given the tracks $\mathcal{T}$ (Sec. 3.2), initial cameras $\hat{\mathcal{P}}$ , and the initial point cloud $\hat{X}$ (Sec. 3.3), Bundle Adjustment BA mini + +mizes the reprojection loss $\mathcal{L}_{\mathrm{BA}}[1,2,23,59]$ + +$$ +X,\mathcal{P} = \operatorname {BA}(\mathcal{T},\hat{X},\hat{\mathcal{P}}) = \operatorname *{arg min}_{X,\mathcal{P}}\mathcal{L}_{\operatorname {BA}} +$$ + +$$ +\mathcal {L} _ {\mathrm {B A}} = \sum_ {i = 1} ^ {N _ {I}} \sum_ {j = 1} ^ {N _ {\mathbf {x}}} v _ {i} ^ {j} \| P _ {i} \left(\mathbf {x} ^ {j}\right) - y _ {i} ^ {j} \|, \tag {6} +$$ + +summing over all reprojection errors $\| P_i(\mathbf{x}^j) - \mathbf{y}_i^j\|$ each comprising the distance between the projection $P_{i}(\mathbf{x}^{j})$ of the point cloud $\mathbf{x}^j\in X$ to camera $P_{i}\in \mathcal{P}$ , and the $i$ -th 2D point $\mathbf{y}_i^j\in T^j$ of the track $T^j$ . Additionally, the error terms are filtered out if the corresponding points have low visibility, low confidence, or do not fit the geometric constraints defined by [59]. Points with large reprojections errors are also filtered [59, 63, 81]. More details are provided in the supplementary material. + +Differentiable Levenberg-Marquardt Following common practice [39, 59], we minimize Eq. (6) with second-order Levenberg-Marquardt (LM) optimizer [49]. However, optimizing the main training loss (Eq. (2)) via backpropagation requires differentiability of Eq. (6) which is nontrivial. Therefore, we leverage the recently proposed Theiss library [54] which exploits the implicit function theorem to backpropagate through deep networks with nested optimization loops. + +# 3.5. Method details + +Camera parameterization Each camera pose $P \in \mathcal{P}$ is parameterized with 8-degrees of freedom: the quaternion $q(R) \in \mathbb{R}^4$ of the rotation $R \in \mathbb{SO}(3)$ and the translation $\mathbf{t} \in \mathbb{R}^3$ components of $P$ 's extrinsics $g \in \mathbb{SE}(3)$ , and the logarithm $\ln(\mathfrak{f}) \in \mathbb{R}$ of the camera focal length $\mathfrak{f} \in \mathbb{R}^+$ . Given these values, the $3 \times 4$ pose matrix is defined as $P = KR[\mathbb{I}_3|\mathbf{t}]$ , with the calibration matrix $K = [\mathfrak{f},0,p_x;0,\mathfrak{f},p_y;0,0,1] \subset \mathbb{R}^{3 \times 3}$ (row major order). Following standard practice [39, 59], we set the principal-point coordinates $p_x, p_y \in \mathbb{R}$ to the center of the image. + +Training loss The training loss $\mathcal{L}$ (Eq. (2)) is defined as: + +$$ +\begin{array}{l} \mathcal {L} \left(f _ {\theta} (\mathcal {I}), \mathcal {P} ^ {\star}, \mathcal {T} ^ {\star}, X ^ {\star}\right) = \sum_ {j = 1} ^ {N _ {T}} \left| \mathbf {x} ^ {\star j} - \mathbf {x} ^ {j} \right| _ {\epsilon} + \left| \mathbf {x} ^ {\star j} - \hat {\mathbf {x}} ^ {j} \right| _ {\epsilon} + \\ + \sum_ {i = 1} ^ {N _ {I}} e _ {\mathcal {P}} \left(P _ {i} ^ {\star}, P _ {i}\right) + e _ {\mathcal {P}} \left(P _ {i} ^ {\star}, \hat {P} _ {i}\right) - \sum_ {i = 1} ^ {N _ {I}} \sum_ {j = 1} ^ {N _ {T}} \log \mathcal {N} \left(\mathbf {y} _ {i} ^ {j \star} \mid \mathbf {y} _ {i} ^ {j}, \sigma_ {i} ^ {j}\right) \tag {7} \\ \end{array} +$$ + +Here, $|\mathbf{x}^{\star j} - \mathbf{x}^{j}|_{\epsilon}$ and $|\mathbf{x}^{\star j} - \hat{\mathbf{x}}^{j}|_{\epsilon}$ evaluate the $\epsilon$ -thresholded pseudo-Huber loss $|\cdot|_{\epsilon}$ [9] between the ground truth 3D points $\mathbf{x}^{\star j}$ and the initial and BA-refined 3D points $\hat{\mathbf{x}}^{j} \in X$ , $\mathbf{x}^{j} \in \hat{X}$ respectively. The camera errors $e_{\mathcal{P}}(P_{i}^{\star}, P_{i})$ and $e_{\mathcal{P}}(P_{i}^{\star}, \hat{P}_{i})$ compare the predicted initial pose $\hat{P}_{i} \in \hat{\mathcal{P}}$ and bundle-adjusted camera pose $P_{i} \in \mathcal{P}$ to the ground truth camera annotation $P_{i}^{\star} \in \mathcal{P}_{i}^{\star}$ . Here, $e_{\mathcal{P}}(P, P')$ is defined as + +![](images/ae02d9a9a590f9d23b4ab5a802f8a36d70ca2ba5bbd5951131f454bdb08ad79d.jpg) +Figure 4. Camera and point-cloud reconstructions of VGGsM on Co3D (left) and IMC Phototourism (right). + +![](images/a51b61737b1376a155571c60a5f65263fcebdf3c48abbb88549c37f2144676e1.jpg) + +![](images/c82164bf713cf0f8e4366d30c8135a3cb7ccdc4fc13aef0281abad9bd47c1dc1.jpg) + +![](images/4a5d9372aa6c42dd286a9a73473d33a9784a1a476ce6cf8e18ce3bea01eb6ebe.jpg) + +![](images/af2dbb6dcff635d97285e09f1f562c8ba09dd4a1c53793afba87a29c2375cf15.jpg) + +![](images/11d7d1c06220d36f6bd6e598cc22094c18313e9802b83f1ce23d1f47f2850690.jpg) + +![](images/ca1f83117bc6914ddc7ade570ebb5cca3debc3b35976cc2e10efd7174979c992.jpg) + +![](images/ff9e74907372f44fedef4f1315f3cead9c279c3ff58d90db2f22f47372d642f2.jpg) + +![](images/3009b743da56e9602210e16f47d34eb09f908072ae60a07846635dacfa36c6e0.jpg) + +the Huber-loss $|\cdot|_{\epsilon}$ between the parameterizations of poses $P, P'$ . Finally, $\log \mathcal{N}(\mathbf{y}_i^{j\star}|\mathbf{y}_i^j,\sigma_i^j)$ computes the likelihood of a ground-truth track point $\mathbf{y}_i^{j\star} \in T_i^\star$ under a probabilistic track-point estimate defined by a 2D gaussian with mean and variance predictions $\mathbf{y}_i^j$ and $\sigma_i^j$ respectively (i.e. the aleatoric uncertainty model described in Sec. 3.2). + +# 4. Experiments + +In this section, we first introduce the datasets together with the protocols for training and evaluation. Then, we provide comparison to existing methods and ablation studies. + +Datasets. Following prior work [25, 39, 74], we evaluate camera pose estimation on Co3Dv2 [55] and IMC Phototourism datasets [31], and 3D triangulation on ETH3D [60]. Co3D is an object-centric dataset comprising turntable-style videos from 51 categories of MS COCO [38]. IMC Phototourism, provided by Image Matching Challenge [31], contains 8 testing scenes and 3 validation scenes of famous landmarks. Generally, the Co3D scenes have much wider baselines, making them challenging for traditional frameworks such as COLMAP, while the IMC samples often have sufficiently overlapping fursta, which is where COLMAP excels. ETH3D provides highly accurate point clouds (captured by laser scanner) for 13 indoor and outdoor scenes, and hence is suitable for the evaluation of triangulation. + +Training. For the model evaluated on IMC Phototourism and ETH3D, we follow the protocol of [25, 39, 57] and train on the MegaDepth dataset [35]. MegaDepth con + +tains 1M crowd-sourced images depicting 196 tourism landmarks, auto-annotated with SfM tools. Hyper-parameters are tuned using the IMC validation set. As in prior work [25, 40], some scenes of MegaDepth are excluded from training due to their low quality or due to overlap with the IMC test set. For Co3Dv2, we conduct training and evaluation on 41 categories as in [36, 74, 85]. + +We chose a multi-stage training strategy for better stability. We first train the tracker T on the synthetic Kubric dataset [21] following the training protocol of [15, 32]. Then, the tracker is fine-tuned solely on Co3D or MegaDepth, depending on the test dataset. Subsequently, we train the camera initializer, with the tracker frozen. Next, the triangulator is trained with the aforementioned two components held frozen. Finally, all components are trained end-to-end. In all stages, we randomly sample a training batch of 3 to 30 frames. + +Testing. Given input test frames $\mathcal{I}$ , we first select the query frame by identifying the image that is closest to all others based on the cosine similarity between global descriptors extracted by an off-the-shelf ResNet50 [24]. Then, we extract SuperPoint and SIFT keypoints from the query frame to serve as the query points for the tracker $\mathbb{T}$ . Although our method can track any query point, it performs better when the queries are distinctive. To improve accuracy, we iterate the whole reconstruction function $f_{\theta}$ multiple times until reaching sub-pixel BA reprojection error $\mathcal{L}_{\mathrm{BA}}$ . After the first iteration, the query image for each subsequent iteration is the one that is farthest from the + +![](images/aafcf0ddb6d3eb1def661e0bba51f4556223b33b0003c81c16cba03e33de7a63.jpg) +Query + +![](images/80075a9003b24193276a93f9a5956a0e05e136ba6187903daa1aecf422f056b3.jpg) +Figure 5. Qualitative Evaluation of Tracking. In each row, the left-most frame contains the query image with query points (crosses). The predicted track points $\mathbf{y}_i^j$ (dots) are shown to the right. The top-right part also highlights our track-point confidence predictions (described in Sec. 3.2), illustrated as ellipses with extent proportional to the predicted variance $\sigma_i^j$ . Note how the uncertainty corresponds to an expectation, e.g., a keypoint covering a vertical stripe has higher uncertainty along the $y$ -axis. + +![](images/71c979bdebe60df942874b8dae91efdf8f940377ad6bb37136b2e8737ab02f1c.jpg) + +![](images/d98486beba8ef5e294d995b23c8d7716b7b309d66fbd06445ac1778f35257e1b.jpg) +Tracking Results + +![](images/37b5585f4a0a05ba48ce1022a036e51ad609ccef2806fb207417d0382f55567e.jpg) + +![](images/f2d9614eef7562234ca71e2aeb1c5cac6a76d2546f17eeddde3ca15d67b6270f.jpg) + +![](images/2568c0bf9252837084175c3737cdf7ed89569449b9f2aa8d346e4a4e0d697c4f.jpg) + +![](images/b56803d0e810aa036875e36d69780fa9e1acb0b086af5f3b4a2af733ec63603f.jpg) + +![](images/38e8a6110c2ca96bd8346c0e7a57f7b9b15dbf2a058617ae51db700f05497954.jpg) + +![](images/32de1d2a21c79c0f5b2f33537f1d3b62f79ee87ca9e35ad9e1bdaf66c776a044.jpg) + +
Co3D DatasetIncremental SfMDeep
MethodCOLMAP (SP + SG)PixSfM (SP + SG)RelPose[85]PoseReg[74]RelPose++[36]PoseDiffusion[74]Ours w/o JointOurs
RRE@15°31.633.757.153.282.380.588.292.1
RTE@15°27.332.9-49.177.279.883.488.3
AUC@30°25.330.1-45.065.166.570.774.0
+ +Table 1. Camera Pose Estimation on Co3D, where the proposed method outperforms previous methods by a large margin. Ours w/o Joint indicates the variant of our framework without training all components jointly. + +
IMC DatasetMethodAUC@3°AUC@5°AUC@10°
DeepDeepSFM10.2719.3631.35
PoseDiffusion12.3123.1736.82
Incremental SfMCOLMAP (SIFT+NN)23.5832.6644.79
PixSfM (SIFT + NN)25.5434.8046.73
PixSfM (LoFTR)44.0656.1669.61
PixSfM (SP + SG)45.1957.2270.47
DFSfM (LoFTR)46.5558.7472.19
DeepOurs w/o Joint38.2351.6068.35
Ours45.2358.8973.92
+ +Table 2. Camera Pose Estimation on IMC. Our method achieves better accuracy than state-of-the-art Incremental SfM approaches on 2 out of 3 AUC thresholds. + +query image of the previous iteration (as measured by the ResNet50 cosine similarity). The re-projections of the point cloud from the current iteration initialize the tracking in the next iteration. + +# 4.1. Camera pose estimation + +Following [25, 31, 39, 74], we report the metric area-under-curve (AUC) to evaluate camera pose accuracy. In Co3D, similar to [74], we also report the relative rotation error (RRE) and relative translation error (RTE). More specifically, for every pair of input frames, we compute the angular + +translation and rotation error, which are later compared to a threshold yielding accuracies RTE and RRE respectively. For a range of thresholds, AUC picks the lower between RRE and RTE, and outputs the area under the accuracy-threshold curve. The results on IMC and CO3D are presented in Tab. 1 and Tab. 2 respectively. For a fair comparison on IMC, we finetune DeepSFM [78] and PoseDiffusion [74] on MegaDepth using their open-source code. The results of Incremental SfM methods are copied from Detector-free SfM [25]. + +Results indicate that VGGSfM outperforms existing methods by a large margin (+9 accuracy points for each metric) on the CO3D dataset. Here, traditional SfM pipelines suffer because of the wide baselines between test frames. On IMC, with a good overlap between views, traditional SfM remains superior to recent data-driven deep-learning pipelines [74, 78]. Our end-to-end trained VGGSfM, however, outperforms all other methods on AUC@10 and AUC@5, and ranks second on AUC@3, convincingly demonstrating its ability to perform well in both narrow- and wide-baseline regimes. The accuracy and completeness of our point clouds can be further qualitatively evaluated in Fig. 4. + +
MethodAccuracy (%)Completeness (%)
1cm2cm5cm1cm2cm5cm
PatchFlow (LoFTR)66.7378.7389.933.4811.3430.96
PixSfM (LoFTR)74.4284.0892.632.919.3927.31
PixSfM (SIFT + NN)76.1885.6093.160.170.713.29
PixSfM (SP + SG)79.0187.0493.800.752.7711.28
DFSfM (LoFTR)80.3889.0195.833.7311.0729.54
Ours80.6289.4996.524.5213.1133.96
+ +# 4.2.3D triangulation + +We evaluate the accuracy and completeness of 3D triangulation on the ETH3D dataset [60] using the same protocol as [17, 25, 39], which triangulates points with fixed camera poses and intrinsics. Results are shown in Tab. 3, where metrics are averaged over all scenes. Our VGGsFm achieves better accuracy and completeness than all baselines (PatchFlow [17], PixSfM [39], and DFSfM [25]), regardless of which keypoint detection or matching method they use. This is especially obvious from the completeness attained at the $5\mathrm{cm}$ threshold, with our $33.96\%$ compared to $29.54\%$ of the best prior work. + +# 4.3. Ablation study + +End-to-end Training. As reported in Tab. 1 and Tab. 2, the end-to-end joint training of the whole framework is important for achieving state-of-the-art performance. Specifically, comparing VGGsF to an ablation which lacks end-to-end training (Ours w/o joint) we record an improvement from $70.7\%$ AUC@30 to $74.0\%$ on the Co3D dataset, and $68.35\%$ AUC@10 to $73.92\%$ on the IMC dataset. This demonstrates the benefits of our fully-differentiable design, and the synergy between its components. + +Tracking or Pairwise Matching. We compare the performance of our predicted tracks to pairwise matching methods on the IMC dataset. Specifically, we split our 2D tracks into pairwise matches and feed these matches to PixSfM. Also, we construct 2D tracks by pairwise matching (based on the open-source implementation of PixSfM) and feed them to our framework. It is worth noting that tracks from pairwise matching have a lot of "holes" because pairwise matching cannot guarantee proper point tracking. We fix these holes by setting their locations as the point locations in the query frame, marking them as invisible. At the same time, for fair comparison, cameras are still initialized using our tracks, because SP+SG cannot provide track features to our camera initializer. The results are shown in Tab. 4. Although COLMAP (the basis of PixSfM) is designed and carefully engineered for pairwise matching, our tracks achieve a slightly better result than the state-of-the-art matching op + +Table 3. 3D Triangulation on ETH3D [60] reporting the accuracy and completeness metrics at different thresholds. + +
PixSfM(SP + SG)PixSfM(Our Tracks)Ours(SP + LG)Ours
AUC@10°70.4770.6268.7873.92
+ +Table 4. Tracking or Pairwise Matching. We respectively provide tracks predicted by our tracker or matches estimated by SP+SG to PixSfM and to our VGGSfM. + +
PoseDiffusionOur Camera Initialized
DLT62.1869.42
Our Triangulator66.3773.92
+ +Table 5. Ablation Study for Camera Initializer and Triangulator. A clear performance drop is observed when replacing our camera initializer by deep camera prediction method PoseDiffusion, or replacing triangulator by DLT. + +tion $\mathrm{SP + SG}$ . Instead, directly feeding $\mathrm{SP + SG}$ tracks to our framework leads to a performance drop. We attribute this to the fact that using $\mathrm{SP + SG}$ tracks cannot benefit from the joint training. We also provide a qualitative evaluation of our tracking accuracy in Fig. 5 on the IMC dataset. + +Camera Controller and Triangulator. We also validate the design of our camera controller $\mathfrak{T}_{\mathcal{P}}$ and triangulator $\mathfrak{T}_X$ on the IMC dataset. As reported in Tab. 5, AUC@10 drops clearly if we replace them with alternatives, proving that they provide sufficiently accurate initialization for our global bundle adjustment BA, without the need for incremental camera registration. + +Coarse-to-fine Tracking. As discussed above, accurate correspondences are important for structure from motion. We demonstrate the significance of our coarse-to-fine tracking mechanism for the method performance. By conducting an ablation study where the fine tracker is removed, we observe a significant performance drop on the IMC dataset, with AUC@10 dropping from $73.92\%$ to $62.30\%$ . + +# 5. Conclusion + +In this paper, we have presented VGGSfM, a fully differentiable SfM approach. We find that even long-standing pipelines, such as Structure-from-Motion, benefit from a learned adaptation between their components. This allows VGGSfM to be simpler than traditional SfM frameworks while achieving better performance across benchmark datasets. Moreover, our framework is fully implemented in Python, which will allow for easy modification and improvements in the future. While VGGSfM already achieves good performance, it cannot yet compete with established pipelines in all application domains. For example, it currently lacks the capability to process thousands of images as in traditional SfM frameworks. Nonetheless, we find differentiable SfM a promising direction of research, and our approach lays the foundation for further advances. + +# References + +[1] Sameer Agarwal, Noah Snavely, Steven M Seitz, and Richard Szeliski. Bundle adjustment in the large. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 29-42. Springer, 2010. 5 +[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1, 2, 3, 5 +[3] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2022. 2 +[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 2 +[5] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4322-4331, 2019. 3 +[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 3 +[7] Lucas Brynte, José Pedro Iglesias, Carl Olsson, and Fredrik Kahl. Learning structure-from-motion with graph attention networks. arXiv preprint arXiv:2308.15984, 2023. 3 +[8] Jonathan L. Carrivick, Mark W. Smith, and Duncan J. Quincey. Structure from Motion in the Geosciences. John Wiley & Sons, 2016. 1 +[9] Pierre Charbonnier, Laure Blanc-féaud, Gilles Aubert, and Michel Barlaud. Deterministic edge-preserving regularization in computed imaging. IEEE Trans. Image Processing, 6:298-311, 1997. 5 +[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1 +[11] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 3 +[12] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2, 3 +[13] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. In Proceedings of the IEEE International Conference on Computer Vision, pages 864-872, 2015. 2 +[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on + +computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 4 +[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurlPS Datasets Track, 2022. 2, 3, 4, 6 +[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 2, 3, 4 +[17] Mihai Dusmanu, Johannes L Schonberger, and Marc Pollefeys. Multi-view optimization of local feature geometry. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 670-686. Springer, 2020. 8 +[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3 +[19] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 368-381. Springer, 2010. 2 +[20] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Towards Internet-scale multi-view stereo. In Proc. CVPR. IEEE, 2010. 1 +[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliaasacchi. Kubric: a scalable dataset generator. 2022. 6 +[22] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2, 3, 4 +[23] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 1, 2, 3, 5 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. arXiv preprint arXiv:1512.03385, 2015. 5, 6 +[25] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In *arxiv*, 2023. 1, 2, 3, 4, 6, 7, 8 +[26] Jared Heinly, Johannes L. Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world* in six days $(\ast)$ (as captured by the yahoo 100 million image dataset). In + +Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1 +[27] Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Piermattei, James O'Connor, and Jacqueline Rosette. Structure from motion photogrammetry in forestry: A review. Current Forestry Reports, 5:155-168, 2019. 1 +[28] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. ArXiv, 2212.04492, 2022. 1 +[29] San Jiang, Cheng Jiang, and Wanshou Jiang. Efficient structure from motion for large-scale UAV images: A review and a comparison of sfm tools. ISPRS Journal of Photogrammetry and Remote Sensing, 167:230-251, 2020. 1 +[30] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3 +[31] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 2, 6, 7 +[32] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023. 2, 3, 4, 6 +[33] Alex Kendall and Yarin Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Proc. NeurIPS, 2017. 4 +[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1 +[35] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 6 +[36] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3, 5, 6, 7 +[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 1 +[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In Proc. ECCV, 2014. 6 +[39] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. arXiv.cs, abs/2108.08291, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[40] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023. 3, 4, 6 + +[41] Yin Lou, Noah Snavely, and Johannes Gehrke. Matchminer: Efficient spanning structure mining in large image collections. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12, pages 45-58. Springer, 2012. 3 +[42] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 2 +[43] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 2 +[44] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 3 +[45] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3 +[46] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust Wide Baseline Stereo from Maximally Stable Extremal Regions. In Proc. BMVC, 2002. 2 +[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Proc. ECCV, 2020. 5 +[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[49] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 5 +[50] Pierre Moulon, Pascal Monasse, and Renaud Marlet. Global fusion of relative motions for robust, accurate and scalable structure from motion. In Proceedings of the IEEE international conference on computer vision, pages 3248-3255, 2013. 2 +[51] David Novotny, Diane Larlus, and Andrea Vedaldi. Learning 3d object categories by looking around them. In Proc. ICCV, 2017. 4 +[52] John Oliensis. A critique of structure-from-motion algorithms. Computer Vision and Image Understanding, 80(2): 172-214, 2000. 2 +[53] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1, 2 +[54] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, et al. Theses: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 35:3801-3818, 2022. 2, 5 +[55] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of + +real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2, 6 +[56] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 3 +[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 4, 6 +[58] Frederik Schaffalitzky and Andrew Zisserman. Multi-view Matching for Unordered Image Sets, or "How Do I Organize My Holiday Snaps?" In Proc. ECCV, 2002. 1 +[59] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3, 4, 5 +[60] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 2, 6, 8 +[61] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 3 +[62] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 3 +[63] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5 +[64] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3 +[65] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. arXiv preprint arXiv:1806.04807, 2018. 3 +[66] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018.3 +[67] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 3 +[68] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 3 + +[69] Bill Triggs, Philip F. McLauchlan, Richard I. Hartley, and Andrew W. Fitzgibbon. Bundle Adjustment - A Modern Synthesis. In Proc. ICCV Workshop, 2000. 3 +[70] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 3, 4 +[71] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 3 +[72] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. NeurIPS, 2017. 2, 5 +[73] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 3 +[74] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3, 5, 6, 7 +[75] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2, 3 +[76] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1 +[77] Tong Wei, Yash Patel, Alexander Shekhotsov, Jiri Matas, and Daniel Barath. Generalized differentiable ransac. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17649-17660, 2023. 3 +[78] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 230–247. Springer, 2020. 3, 7 +[79] Matthew J Westoby, James Brasington, Niel F Glasser, Michael J Hambrey, and Jennifer M Reynolds. 'structure-from-motion'photogrammetry: A low-cost, effective tool for geoscience applications. Geomorphology, 179:300-314, 2012. 1 +[80] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 61-75. Springer, 2014. 2 +[81] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2, 5 + +[82] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 3 +[83] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3d animals in the wild. 2023. 3 +[84] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned Invariant Feature Transform. In Proc. ECCV, 2016. 3 +[85] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3, 5, 6, 7 +[86] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 3 +[87] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3 \ No newline at end of file diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/images.zip b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e570097de51d09b9996c76becce6847f3c5bce01 --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7073962ec52635aa8096a7fe571c9e1a89f1d699d9775cddcd7925a5c2cb11d2 +size 665181 diff --git a/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/layout.json b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec3884107d1dfbcee9768c028c4683ce32704783 --- /dev/null +++ b/2024/VGGSfM_ Visual Geometry Grounded Deep Structure From Motion/layout.json @@ -0,0 +1,11867 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "text", + "content": "VGGSfM: Visual Geometry Grounded Deep Structure From Motion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 167, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 167, + 158 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 167, + 158 + ], + "type": "text", + "content": "Jianyuan Wang" + }, + { + "bbox": [ + 79, + 142, + 167, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 143, + 276, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 143, + 276, + 156 + ], + "spans": [ + { + "bbox": [ + 193, + 143, + 276, + 156 + ], + "type": "text", + "content": "Nikita Karaev 1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 143, + 405, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 143, + 405, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 143, + 405, + 157 + ], + "type": "text", + "content": "Christian Rupprecht1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 432, + 143, + 512, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 143, + 512, + 157 + ], + "spans": [ + { + "bbox": [ + 432, + 143, + 512, + 157 + ], + "type": "text", + "content": "David Novotny2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 171, + 354, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 171, + 354, + 186 + ], + "spans": [ + { + "bbox": [ + 125, + 171, + 354, + 186 + ], + "type": "text", + "content": "Visual Geometry Group, University of Oxford" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 417, + 171, + 465, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 171, + 465, + 184 + ], + "spans": [ + { + "bbox": [ + 417, + 171, + 465, + 184 + ], + "type": "text", + "content": "Meta AI" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 198, + 370, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 198, + 370, + 209 + ], + "spans": [ + { + "bbox": [ + 222, + 198, + 370, + 209 + ], + "type": "text", + "content": "https://vggsfm.github.io" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 228, + 191, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 191, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 191, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 253, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 290, + 529 + ], + "type": "text", + "content": "Structure-from-motion (SfM) is a long-standing problem in the computer vision community, which aims to reconstruct the camera poses and 3D structure of a scene from a set of unconstrained 2D images. Classical frameworks solve this problem in an incremental manner by detecting and matching keypoints, registering images, triangulating 3D points, and conducting bundle adjustment. Recent research efforts have predominantly revolved around harnessing the power of deep learning techniques to enhance specific elements (e.g., keypoint matching), but are still based on the original, non-differentiable pipeline. Instead, we propose a new deep pipeline VGGsfM, where each component is fully differentiable and thus can be trained in an end-to-end manner. To this end, we introduce new mechanisms and simplifications. First, we build on recent advances in deep 2D point tracking to extract reliable pixel-accurate tracks, which eliminates the need for chaining pairwise matches. Furthermore, we recover all cameras simultaneously based on the image and track features instead of gradually registering cameras. Finally, we optimise the cameras and triangulate 3D points via a differentiable bundle adjustment layer. We attain state-of-the-art performance on three popular datasets, CO3D, IMC Phototourism, and ETH3D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Reconstructing the camera parameters and the 3D structure of a scene from a set of unconstrained 2D images is a long-standing problem in the computer vision community. Among many other applications [8, 27, 29, 53, 79], it has recently emerged as an important component of learning neural fields [10, 28, 34, 37, 48, 76]. The problem is usually solved via the Structure-from-Motion (SfM) framework which estimates the 3D point cloud (Structure) and the parameters of each camera (Motion) in the scene. State-of-the-art methods [25, 39] follow the incremental SfM paradigm whose origins can be traced back to the early 2000s [23, 58]. It usually begins with a small set of" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 423, + 597 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 423, + 597 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 423, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 423, + 597 + ], + "type": "image", + "image_path": "442374b47f7f05fbfd05936a49654d12e904f4edefe9b5e38edb91c9d73bc601.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "text", + "content": "Figure 1. Reconstruction of In-the-wild Photos with VGGsFm, displaying estimated point clouds (in blue) and cameras (orange)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 423, + 228, + 544, + 596 + ], + "blocks": [ + { + "bbox": [ + 423, + 228, + 544, + 596 + ], + "lines": [ + { + "bbox": [ + 423, + 228, + 544, + 596 + ], + "spans": [ + { + "bbox": [ + 423, + 228, + 544, + 596 + ], + "type": "image", + "image_path": "936d5f7f73bd041253a5fa48db6febd10820525639041254582ea1446a59bce0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "type": "text", + "content": "correspondence-rich images as initialization, and gradually adds more views into the reconstruction, through keypoint detection, matching, verification, image registration, triangulation, bundle adjustment (BA), and so on [2, 20, 26, 59]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Recent research efforts have predominantly revolved around leveraging the power of deep learning techniques to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21686" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 547, + 218 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 547, + 218 + ], + "type": "image", + "image_path": "ae5f31ae48b8f47b1f507ddcca7d928cd84069cdca49fbdcc7f7e8df73fbb707.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 261 + ], + "type": "text", + "content": "Figure 2. Overview of VGGs fM. Our method extracts 2D tracks from input images, reconstructs cameras using image and track features, initializes a point cloud based on these tracks and camera parameters, and applies a bundle adjustment layer for reconstruction refinement. The whole framework is fully differentiable and designed for end-to-end training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 460 + ], + "type": "text", + "content": "enhance specific elements within the original pipeline while preserving the incremental SfM framework as a whole. For instance, SuperPoint and SuperGlue [14, 57] focus on improving keypoint detection and matching. Pixel-perfect SfM [39] proposes deep feature-metric refinement to adjust both keypoints and bundles. Detector-free feature matching methods [12, 64, 75] bypass early keypoint detection by means of attention, which is powerful in poorly textured scenes. Detector-free SfM [25] builds a coarse SfM model through quantized detector-free matches and then iteratively refines it with multi-view consistency constraints. These advancements successfully combine deep learning approaches (such as deep feature matching) with well-established hand-engineered components, such as the incremental camera registration of COLMAP [59]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 461, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 616 + ], + "type": "text", + "content": "The widespread success of end-to-end training warrants the question of what benefits it can bring to long-standing frameworks such as SfM. Naturally, it is often difficult to assess the merits of new approaches when compared with decades of continuous improvements. Nonetheless, in this paper, we answer this question by introducing a fully-differentiable SfM pipeline, dubbed Visual Geometry Grounded Deep Structure From Motion (VGGSfM), which trains in an end-to-end manner. We find that this allows the pipeline to be simpler than prior frameworks while achieving better or comparable performance. Training end-to-end allows each component to generate outputs that facilitate the task of its successor." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "To build a fully-differentiable pipeline, we make several substantial changes to the SfM procedures and overall obtain better performance. Specifically, our model builds on recent advances in deep 2D point tracking [15, 16, 22, 32] to directly extract reliable pixel-accurate tracks. This simplifies the correspondence estimation step in traditional SfM, which first estimates pairwise matches and then connects them into tracks. Then, based on the image and track fea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "content": "tures, VGGsFm estimates all cameras jointly via a Transformer [72], and subsequently all 3D points. Different from Incremental SfM, this approach is simpler and easier to differentiate as it does not depend on a discrete, combinatorial correspondence chaining step. Finally, for bundle adjustment, we replace the commonly employed non-differentiable Ceres solver [3] with the fully differentiable second-order Theseus solver [54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "content": "Hence, we fuse all the SfM components into a single fully differentiable reconstruction function " + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "content": ". Besides that, in our experiments, we also show that the individual modules perform well in isolation. Ultimately, end-to-end training yields another performance improvement, that surpasses the performance of isolated components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 450, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 535 + ], + "type": "text", + "content": "We evaluate VGGsFm for the task of camera pose estimation on the CO3Dv2 [55] and IMC Phototourism [31] datasets, and for 3D triangulation on the ETH3D [60] dataset. Our method attains strong performance on all benchmarks. At the same time, we conduct in-the-wild reconstruction to validate the generalization ability of our proposed framework, as shown in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 392, + 560 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "type": "text", + "content": "Structure from Motion is a fundamental problem in computer vision and has been investigated for decades [23, 52, 53]. The classical pipelines usually solve the SfM problem in a global [13, 50, 80] or incremental [2, 19, 59, 63, 81] manner. Both of which are usually based on pairwise image keypoint matching. Incremental SfM is arguably the most widely adopted strategy (e.g., the popular framework COLMAP [59]). Therefore, in the following sections, we refer to incremental SfM as \"classical\" or \"traditional\" SfM. We defer the discussion of global SfM to the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Traditional SfM frameworks often start by detecting keypoints and feature descriptors [4, 42, 43, 46]. They then" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21687" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "search for image pairs with overlapping frusta by matching these keypoints across different images (e.g., with a nearest-neighbour search) [2, 41, 59]. These image pairs are further verified via two-view epipolar geometry or homography [23] through RANSAC [18]. Then, a pair or a small set of images is carefully selected for initialization. New images are gradually registered by solving the Perspective-" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "-Point (PnP) problem [44], followed by triangulating 3D points, and bundle adjustment [69]. This process is iterated until all the frames are either registered or discarded. 2D correspondences (multi-view tracks) are the basis of the whole process, however, they are usually simply constructed by chaining two-view matches [59]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 433 + ], + "type": "text", + "content": "Many deep-learning approaches have been proposed to enhance this framework. For example, [14, 70, 84] provide better keypoint detection and [11, 30, 40, 57, 61] focus on matching. Furthermore, detector-free matching methods [12, 64, 75] propose to avoid sparse keypoint detection by building semi-dense matches via self and cross attention. [7] solves the subproblems of SfM by graph attention networks. Some studies improve the performance of RANSAC by making it trainable [5, 6, 77]. Recent state-of-the-art methods are PixSfM [39] and the concurrent Detector-free SfM (DFSfM) [25]. PixSfM refines the tracks and structure estimated by COLMAP through feature-metric keypoint adjustment and feature-metric bundle adjustment. Detector-free SfM proposes to first build a coarse SfM model using detector-free matches and COLMAP (or other frameworks), and then to iteratively refine the tracks and the structure of the coarse model by enforcing multi-view consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 434, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 567 + ], + "type": "text", + "content": "Recently, fully differentiable SfM pipelines have also been explored. They usually use deep neural networks to regress camera poses and depths [65, 66, 68, 71, 73, 78, 87]. Although using an approximation of bundle adjustment [65, 68, 78], these methods suffer from limited generalizability and scalability (very few input frames) [65, 71, 73, 78, 87], or rely on temporal relationship [66, 68]. Meanwhile, some methods are category-specific [45, 82, 83]. The recent efforts on deep camera pose estimation can scale up to more than 50 frames, but they do not reconstruct the scene [36, 62, 74, 85]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Point tracking. Since VGGSfM proposes a novel point track predictor, next, we review recent advances in this field. Inspired by the optical-flow architecture of RAFT [67], PIPs [22] revisited point tracking, a task related to Particle Video [56], and proposed a highly accurate tracker of isolated points in a video. TAP-Vid [15] (i.e., \"Tracking Any Point\") introduced a benchmark for point tracking and a baseline model, which was later improved in TAPIR [16] by integrating the iterative update mechanism from PIPs. PointOdyssey [86] simplified PIPs and proposed a benchmark for the long-term version of point tracking. CoTracker [32] closed the gap between single point track" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "ing and dense Optical Flow with joint point tracking. However, these works are designed for videos, i.e. temporally-ordered sequences of frames. In our point tracker, given the input frames are unordered, we do not assume a temporal relationship between input frames. We therefore process all frames jointly, avoiding windowed inference of [32]. Since SfM relies on highly accurate correspondences, our tracks are further refined in a coarse-to-fine manner to achieve subpixel accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 193, + 362, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 193, + 362, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 362, + 205 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 213, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 250 + ], + "type": "text", + "content": "In this section, we describe the components of VG-GSfM and how they are composed in a fully differentiable pipeline. An overview of our framework is shown in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": "Problem setting Given a set of free-form images observing a scene, VGGSfM estimates their corresponding camera parameters and the 3D scene shape represented as a point cloud. Formally, given a tuple " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = (I_1,\\dots,I_{N_I})" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "N_{I}\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " RGB images " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "I_{i}\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": ", VGGSfM estimates the corresponding camera projection matrices " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = (P_{1},\\ldots ,P_{N_{I}}|P_{i}\\subset \\mathbb{R}^{3\\times 4})" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " and the scene cloud " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "X = \\{\\mathbf{x}^j\\}_{j = 1}^{N_{\\mathbf{x}}} \\text{ of } N_{\\mathbf{x}}\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " 3D points " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": ". Each projection matrix " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " consists of extrinsic (pose) " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "g_{i}\\in \\mathbb{SE}(3)" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": " and intrinsics " + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "inline_equation", + "content": "K_{i}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 304, + 252, + 547, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "A 3D point " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " can be projected to the " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "-th camera yielding a 2D screen coordinate " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j = P_i(\\mathbf{x}^j) \\sim \\lambda K_i \\hat{\\mathbf{x}}_i^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^j = g_i \\mathbf{x}^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " is the world coordinate " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " expressed in view-coordinates of the " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "-th camera. The projection of the point " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " to all input cameras is a track " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "T^j = ((y_1^j, v_1^j) \\dots, (y_{N_I}^j, v_{N_I}^j))" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " matching 2D points " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": ", and their corresponding binary indicators " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "v_i^j \\in \\{0, 1\\}" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " denoting visibility of the " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "-th point in the " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "-th camera. We denote " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i = \\{T_i^1, \\dots, T_i^{N_T}\\}" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " as the set of all tracks " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "T_i^j" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 374, + 547, + 502 + ], + "type": "text", + "content": "-th camera." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 510, + 410, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 410, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 410, + 522 + ], + "type": "text", + "content": "3.1. Method overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 529, + 514, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 514, + 543 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 514, + 543 + ], + "type": "text", + "content": "VGGsM implements SfM via a single function " + }, + { + "bbox": [ + 306, + 529, + 514, + 543 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 394, + 550, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 550, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 394, + 550, + 545, + 563 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\mathcal {I}) = \\mathcal {P}, X \\tag {1}", + "image_path": "9a001e69cb68fb53c5218fc783c122bb27b13e308619565989778aace2bc662c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": "accepting the set of " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " scene images " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " and outputting the camera parameters " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " and the scene point cloud " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": ". Importantly, " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " is fully differentiable and, as such, its parameters " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " are learned by minimizing the training loss " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 625, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 339, + 625, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\theta^ {\\star} = \\arg \\min _ {\\theta} \\sum_ {s = 1} ^ {S} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathcal {I} _ {s}\\right), \\mathcal {P} _ {s} ^ {\\star}, \\mathcal {T} _ {s} ^ {\\star}, X _ {s} ^ {\\star}\\right), \\tag {2}", + "image_path": "ef0de5e335ba09906c60ad06a8fc8f0b9de3805d5b23d2bb71c913871f27059b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "summing over " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " training image sets " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_s" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " annotated with ground-truth cameras " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s^\\star" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ", tracks " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s^\\star" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ", and point clouds " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_s^\\star" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". We defer the details of " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " to Sec. 3.5 and, in the following paragraphs, discuss the architecture of " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21688" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "The reconstruction function Following traditional SfM [59], VGGSfM decomposes the reconstruction function " + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": " into four seamless stages: 1) point tracking " + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": ", 2) initial camera estimator " + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_{\\mathcal{P}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": ", 3) triangulator " + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_X" + }, + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": " and, 4) Bundle Adjustment BA, as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 139, + 177, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 177, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 177, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {T} = \\mathrm {T} (\\mathcal {I})", + "image_path": "41f1c77b27ae2a8a06af714998e758a2d50616a07316105b09422d4f2af9a53e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 155, + 287, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 155, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 155, + 287, + 176 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} (\\mathcal {I}, \\mathcal {T}) \\\\ \\hat {\\hat {\\tau}} = \\mathfrak {T} _ {\\hat {\\tau}} (\\mathcal {\\hat {T}}, \\hat {\\hat {\\tau}}) \\end{array} \\tag {3}", + "image_path": "66f45272ee301b9803b065525a8535a491cb4ea9f6a0d465572fda042d28bd52.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 170, + 199, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 199, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 199, + 186 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\mathfrak {T} _ {X} (\\mathcal {T}, \\hat {\\mathcal {P}})", + "image_path": "e9cbd7ceaad7103ab824646a09044ce3ede61b291ddf08344412db36d8545a9e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 189, + 214, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 189, + 214, + 203 + ], + "spans": [ + { + "bbox": [ + 120, + 189, + 214, + 203 + ], + "type": "interline_equation", + "content": "\\mathcal {P}, X = _ {\\mathrm {B A}} (\\mathcal {T}, \\hat {\\mathcal {P}}, \\hat {X}).", + "image_path": "256411977a2727276afbac1e79018d1ccd03d46252085661b94dcb45864c55db.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": "The track predictor " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": ", hereafter referred to as \"tracker\", estimates 2D tracks " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": " given input images " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_{\\mathcal{P}}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_X" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": " provide initial cameras " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": " and an initial point cloud " + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 207, + 287, + 268 + ], + "type": "text", + "content": " respectively. Finally, BA enhances accuracy by refining the cameras and 3D points together." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 274, + 113, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 113, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 113, + 287 + ], + "type": "text", + "content": "3.2. Tracking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": "Establishing precise 2D correspondences is important for achieving accurate 3D reconstruction. Traditionally, SfM frameworks first estimate pairwise image-to-image correspondences that are later chained into multi-image tracks " + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": " [25, 39, 59]. Here, typically only point-pair matching benefits from learned components [14, 40, 57, 70], while the chaining of pairwise correspondences remains a hand-engineered process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 388, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 568 + ], + "type": "text", + "content": "Instead, VGGSfM significantly simplifies SfM correspondence tracking by employing a deep feed-forward tracking function. It accepts a collection of images and directly outputs a set of reliable point trajectories across all images. We achieve this by exploiting recent advances in video point tracking methods [15, 16, 22, 32]. Although developed for video-point tracking, these methods are inherently appropriate for SfM which requires a very compact set of highly accurate tracks (e.g., dense optical flow is too memory-demanding). Furthermore, point trackers mitigate the potential errors (sometimes called drift) caused by chaining of pairwise matches. However, as we describe below, our design differs from video trackers because SfM, which accepts free-form images, cannot assume temporal smoothness or ordering, and requires sub-pixel accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": "Tracker architecture The design of our tracker " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " follows [22, 32], and is illustrated in Fig. 3. More specifically, given " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " query points " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{y}}_i^1,\\dots \\hat{\\mathbf{y}}_i^{N_T}\\}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " in a frame " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " we bilinearly sample their corresponding query descriptors " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_i^1,\\dots ,\\mathbf{m}_i^{N_T}\\}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " from image feature maps output by a 2D CNN. Then, each query descriptor is correlated with the feature maps of all " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{I}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " frames at different spatial resolutions, which constructs a cost-volume pyramid. Flattening the latter yields tokens " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{N_T\\times N_I\\times C}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " is the total number of elements in the cost-volume pyramid. Feeding the tokens to a Transformer, we obtain tracks " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{T^j\\}_{j = 1}^{N_T}" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": ". Recall that each track " + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 47, + 570, + 288, + 715 + ], + "type": "text", + "content": " comprises the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "type": "image", + "image_path": "76d43148b58f308457f880f40b291753d2cd49c2f7b25858f343e5d406bd36e2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 222 + ], + "type": "text", + "content": "Figure 3. Architecture of Tracker T. We adopt a coarse-to-fine design for the tracker. The coarse tracker first locates the approximate positions of corresponding points, and the fine tracker then refines these initial predictions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "text", + "content": "set of " + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "text", + "content": " tracked 2D locations " + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j" + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "text", + "content": " together with predicted visibility indicators " + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "inline_equation", + "content": "v_i^j" + }, + { + "bbox": [ + 305, + 243, + 545, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "text", + "content": "It is worth noting that, differently from [22, 32], our tracker does not assume temporal continuity. Therefore, we avoid the sliding window approach and, instead, attend to all the input frames together. Furthermore, unlike in [32], we predict each track independently of others. This allows to track a larger number of points at test time leading to increased density of reconstructed point clouds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 357, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 418 + ], + "type": "text", + "content": "Predicting tracking confidence In SfM, it is crucial to filter out any outlier correspondences as they can negatively impact the subsequent reconstruction stages. To this end, we enhance the tracker with the ability to estimate confidence for each track-point prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": "More specifically, we leverage the aleatoric uncertainty [33, 51] model which predicts variance " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{j}" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": " together with each 2D track point " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": ", so that the resulting normal distribution " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": " tightly peaks around each ground-truth 2D track point " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{j\\star}" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": ". Hence, during training, the " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\ell_1 / \\ell_2" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": " loss, originally used in video point tracking, is replaced with the (negated) logarithm of the latter probability evaluated at each ground truth point " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{j\\star}" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": ". Once trained, the confidence measure " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "1 / \\sigma_{i}^{j}" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": " is proportional to the inverse of the predicted variance. In practice, we assume a diagonal covariance matrix resulting in horizontal and vertical uncertainties " + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{j}" + }, + { + "bbox": [ + 304, + 418, + 545, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Coarse-to-fine tracking Moreover, since SfM requires highly accurate (pixel or sub-pixel level) correspondences, we employ a coarse-to-fine point-tracking strategy. As described above, we first coarsely track image points using feature maps that fully cover the input images " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ". Then, we form " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " patches by cropping input images around the coarse point estimates and execute the tracking again to obtain a sub-pixel estimate. Recall that, differently from the chained matching of traditional SfM, our tracker is fully differentiable. This enables back-propagating the gradient of the training loss " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " through the whole framework to the tracker parameters. This reinforces the synergy between the tracking and the ensuing stages, which are described next." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21689" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "type": "text", + "content": "3.3. Learnable camera & point initialization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 286, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 286, + 209 + ], + "type": "text", + "content": "As discussed above, a traditional SfM pipeline [39, 59] usually relies on an an incremental loop, which often initializes with a correspondence-rich image pair, gradually registers new frames, enlarges the point cloud, and conducts joint optimization (e.g., BA). However, although the framework has been fortified in robustness and accuracy through decades of improvements, this cumulative process has inevitably led to increased complexity. Furthermore, Incremental SfM is largely non-differentiable which complicates end-to-end learning from annotated data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 286, + 293 + ], + "type": "text", + "content": "Thus, in pursuit of simplicity and differentiability, our method departs from the classical SfM scheme. Inspired by recent advances in deep camera pose estimators [36, 74, 85], we propose to initialize the cameras and the point cloud with a pair of deep Transformer [72] networks. Importantly, we register all cameras and reconstruct all scene points collectively in a non-incremental differentiable fashion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "text", + "content": "Learnable camera initializer The predictor of initial cameras " + }, + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "text", + "content": " is implemented as a deep Transformer architecture " + }, + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 296, + 286, + 331 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 335, + 286, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 335, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 52, + 335, + 286, + 361 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {P}} = \\mathfrak {T} _ {\\mathcal {P}} \\left(\\left\\{\\phi \\left(I _ {i}\\right) \\mid I _ {i} \\in \\mathcal {I} \\right\\}, \\left\\{d ^ {\\mathcal {P}} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right). \\tag {4}", + "image_path": "aafb75279b132969e3ee8e4348ef5a3f6ade48b35aa451d4a184aa34a0852c84.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": "It accepts a set of tokens comprising global ResNet50 [24] features " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\phi(I_i)" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": " of input images " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": ", and the set of descriptors " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "d^{\\mathcal{P}}(\\mathbf{y}_i^j)" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": " of track points " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j \\in T_i \\in \\mathcal{T}" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": ". Here, each track descriptor is output by an auxiliary branch of the tracker " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": ". Given these inputs, " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": " first applies cross-attention between the global image feature (query) and the track-descriptor (key-value) pairs yielding " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "N_I = |\\mathcal{I}|" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": " tokens per scene. The output of cross-attention is then concatenated with an embedding of a preliminary camera estimated by the 8-point algorithm taking the correspondences between track points " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": " as input. Finally, we feed this concatenation to a Transformer trunk resulting in the initial cameras " + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 361, + 286, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "text", + "content": "Learnable triangulation Given initial cameras " + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "text", + "content": " and 2D tracks " + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "text", + "content": ", the triangulator outputs the initial point cloud " + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "text", + "content": ". Similar to the camera predictor, the triangulator is a Transformer " + }, + { + "bbox": [ + 46, + 508, + 286, + 556 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_X" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 560, + 286, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 560, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 86, + 560, + 286, + 576 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\mathfrak {T} _ {X} \\left(\\left\\{d ^ {X} \\left(y _ {i} ^ {j}\\right) \\mid \\forall T _ {i} \\in \\mathcal {T} \\forall y _ {i} ^ {j} \\in T _ {i} \\right\\}\\right) \\tag {5}", + "image_path": "3ebe357ab4f1701cfe807cd05240526f0201455761ddeb33bafcbaa466d6668d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": "accepting descriptors " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "d^{X}(\\mathbf{y}_{i}^{j})" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": " each comprising a tracker feature, and a positional harmonic embedding [47] of points " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^j\\in \\bar{X}" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": " from a preliminary point cloud " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": ". The preliminary point cloud is formed via closed-form multi-view Direct Linear Transform (DLT) 3D triangulation [23] of the tracks " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": " given the initial cameras " + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": ". Please refer to the supplementary for a detailed description of both initializers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 159, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 159, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 159, + 684 + ], + "type": "text", + "content": "3.4. Bundle adjustment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Given the tracks " + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": " (Sec. 3.2), initial cameras " + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": ", and the initial point cloud " + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": " (Sec. 3.3), Bundle Adjustment BA mini" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "type": "text", + "content": "mizes the reprojection loss " + }, + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BA}}[1,2,23,59]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 92, + 502, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 92, + 502, + 114 + ], + "spans": [ + { + "bbox": [ + 348, + 92, + 502, + 114 + ], + "type": "interline_equation", + "content": "X,\\mathcal{P} = \\operatorname {BA}(\\mathcal{T},\\hat{X},\\hat{\\mathcal{P}}) = \\operatorname *{arg min}_{X,\\mathcal{P}}\\mathcal{L}_{\\operatorname {BA}}", + "image_path": "5430f52a309bb1eb9c798472a7d966583306f80c8287107151ccd078c2a79c6e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 117, + 544, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 117, + 544, + 150 + ], + "spans": [ + { + "bbox": [ + 353, + 117, + 544, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B A}} = \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {\\mathbf {x}}} v _ {i} ^ {j} \\| P _ {i} \\left(\\mathbf {x} ^ {j}\\right) - y _ {i} ^ {j} \\|, \\tag {6}", + "image_path": "a86e1ac917fe8e6e2f8c119741a9cefb46384535135e61493e760fe65d2c64a8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": "summing over all reprojection errors " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\| P_i(\\mathbf{x}^j) - \\mathbf{y}_i^j\\|" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": " each comprising the distance between the projection " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "P_{i}(\\mathbf{x}^{j})" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": " of the point cloud " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j\\in X" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": " to camera " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "P_{i}\\in \\mathcal{P}" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": "-th 2D point " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j\\in T^j" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": " of the track " + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 306, + 160, + 545, + 269 + ], + "type": "text", + "content": ". Additionally, the error terms are filtered out if the corresponding points have low visibility, low confidence, or do not fit the geometric constraints defined by [59]. Points with large reprojections errors are also filtered [59, 63, 81]. More details are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 380 + ], + "type": "text", + "content": "Differentiable Levenberg-Marquardt Following common practice [39, 59], we minimize Eq. (6) with second-order Levenberg-Marquardt (LM) optimizer [49]. However, optimizing the main training loss (Eq. (2)) via backpropagation requires differentiability of Eq. (6) which is nontrivial. Therefore, we leverage the recently proposed Theiss library [54] which exploits the implicit function theorem to backpropagate through deep networks with nested optimization loops." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 386, + 398, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 398, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 398, + 398 + ], + "type": "text", + "content": "3.5. Method details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": "Camera parameterization Each camera pose " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "P \\in \\mathcal{P}" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " is parameterized with 8-degrees of freedom: the quaternion " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "q(R) \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " of the rotation " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{SO}(3)" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " and the translation " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " components of " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": "'s extrinsics " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "g \\in \\mathbb{SE}(3)" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": ", and the logarithm " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\ln(\\mathfrak{f}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " of the camera focal length " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathfrak{f} \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": ". Given these values, the " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "3 \\times 4" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " pose matrix is defined as " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "P = KR[\\mathbb{I}_3|\\mathbf{t}]" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": ", with the calibration matrix " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "K = [\\mathfrak{f},0,p_x;0,\\mathfrak{f},p_y;0,0,1] \\subset \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " (row major order). Following standard practice [39, 59], we set the principal-point coordinates " + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "inline_equation", + "content": "p_x, p_y \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 405, + 545, + 525 + ], + "type": "text", + "content": " to the center of the image." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 527, + 541, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 541, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 541, + 539 + ], + "type": "text", + "content": "Training loss The training loss " + }, + { + "bbox": [ + 306, + 527, + 541, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 306, + 527, + 541, + 539 + ], + "type": "text", + "content": " (Eq. (2)) is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 548, + 544, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 544, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 544, + 628 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(f _ {\\theta} (\\mathcal {I}), \\mathcal {P} ^ {\\star}, \\mathcal {T} ^ {\\star}, X ^ {\\star}\\right) = \\sum_ {j = 1} ^ {N _ {T}} \\left| \\mathbf {x} ^ {\\star j} - \\mathbf {x} ^ {j} \\right| _ {\\epsilon} + \\left| \\mathbf {x} ^ {\\star j} - \\hat {\\mathbf {x}} ^ {j} \\right| _ {\\epsilon} + \\\\ + \\sum_ {i = 1} ^ {N _ {I}} e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, P _ {i}\\right) + e _ {\\mathcal {P}} \\left(P _ {i} ^ {\\star}, \\hat {P} _ {i}\\right) - \\sum_ {i = 1} ^ {N _ {I}} \\sum_ {j = 1} ^ {N _ {T}} \\log \\mathcal {N} \\left(\\mathbf {y} _ {i} ^ {j \\star} \\mid \\mathbf {y} _ {i} ^ {j}, \\sigma_ {i} ^ {j}\\right) \\tag {7} \\\\ \\end{array}", + "image_path": "58e0c4b332a876915d29d09be3b5a498220359cb0dd2313775bd3256ca2e6088.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "|\\mathbf{x}^{\\star j} - \\mathbf{x}^{j}|_{\\epsilon}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "|\\mathbf{x}^{\\star j} - \\hat{\\mathbf{x}}^{j}|_{\\epsilon}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " evaluate the " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": "-thresholded pseudo-Huber loss " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "|\\cdot|_{\\epsilon}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " [9] between the ground truth 3D points " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\star j}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " and the initial and BA-refined 3D points " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}^{j} \\in X" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{j} \\in \\hat{X}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " respectively. The camera errors " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_{\\mathcal{P}}(P_{i}^{\\star}, P_{i})" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_{\\mathcal{P}}(P_{i}^{\\star}, \\hat{P}_{i})" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " compare the predicted initial pose " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{P}_{i} \\in \\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " and bundle-adjusted camera pose " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_{i} \\in \\mathcal{P}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " to the ground truth camera annotation " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_{i}^{\\star} \\in \\mathcal{P}_{i}^{\\star}" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_{\\mathcal{P}}(P, P')" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21690" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 120, + 317 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 120, + 317 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 120, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 120, + 317 + ], + "type": "image", + "image_path": "ae02d9a9a590f9d23b4ab5a802f8a36d70ca2ba5bbd5951131f454bdb08ad79d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 324, + 500, + 335 + ], + "lines": [ + { + "bbox": [ + 91, + 324, + 500, + 335 + ], + "spans": [ + { + "bbox": [ + 91, + 324, + 500, + 335 + ], + "type": "text", + "content": "Figure 4. Camera and point-cloud reconstructions of VGGsM on Co3D (left) and IMC Phototourism (right)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 70, + 242, + 194 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 242, + 194 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 242, + 194 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 242, + 194 + ], + "type": "image", + "image_path": "a51b61737b1376a155571c60a5f65263fcebdf3c48abbb88549c37f2144676e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 242, + 71, + 362, + 194 + ], + "blocks": [ + { + "bbox": [ + 242, + 71, + 362, + 194 + ], + "lines": [ + { + "bbox": [ + 242, + 71, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 242, + 71, + 362, + 194 + ], + "type": "image", + "image_path": "c82164bf713cf0f8e4366d30c8135a3cb7ccdc4fc13aef0281abad9bd47c1dc1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 362, + 70, + 422, + 194 + ], + "blocks": [ + { + "bbox": [ + 362, + 70, + 422, + 194 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 422, + 194 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 422, + 194 + ], + "type": "image", + "image_path": "4a5d9372aa6c42dd286a9a73473d33a9784a1a476ce6cf8e18ce3bea01eb6ebe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 422, + 71, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 422, + 71, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 422, + 71, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 422, + 71, + 543, + 194 + ], + "type": "image", + "image_path": "af2dbb6dcff635d97285e09f1f562c8ba09dd4a1c53793afba87a29c2375cf15.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 194, + 119, + 255 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 119, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 119, + 255 + ], + "type": "image", + "image_path": "48c57885b7f7b1ef7c0df23a73e60f6e2d2690f59b48f160574186765eb0c7b7.jpg" + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 120, + 194, + 242, + 316 + ], + "blocks": [ + { + "bbox": [ + 120, + 194, + 242, + 316 + ], + "lines": [ + { + "bbox": [ + 120, + 194, + 242, + 316 + ], + "spans": [ + { + "bbox": [ + 120, + 194, + 242, + 316 + ], + "type": "image", + "image_path": "11d7d1c06220d36f6bd6e598cc22094c18313e9802b83f1ce23d1f47f2850690.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 242, + 194, + 362, + 316 + ], + "blocks": [ + { + "bbox": [ + 242, + 194, + 362, + 316 + ], + "lines": [ + { + "bbox": [ + 242, + 194, + 362, + 316 + ], + "spans": [ + { + "bbox": [ + 242, + 194, + 362, + 316 + ], + "type": "image", + "image_path": "ca1f83117bc6914ddc7ade570ebb5cca3debc3b35976cc2e10efd7174979c992.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 194, + 422, + 315 + ], + "blocks": [ + { + "bbox": [ + 362, + 194, + 422, + 315 + ], + "lines": [ + { + "bbox": [ + 362, + 194, + 422, + 315 + ], + "spans": [ + { + "bbox": [ + 362, + 194, + 422, + 315 + ], + "type": "image", + "image_path": "ff9e74907372f44fedef4f1315f3cead9c279c3ff58d90db2f22f47372d642f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 422, + 194, + 543, + 315 + ], + "blocks": [ + { + "bbox": [ + 422, + 194, + 543, + 315 + ], + "lines": [ + { + "bbox": [ + 422, + 194, + 543, + 315 + ], + "spans": [ + { + "bbox": [ + 422, + 194, + 543, + 315 + ], + "type": "image", + "image_path": "3009b743da56e9602210e16f47d34eb09f908072ae60a07846635dacfa36c6e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": "the Huber-loss " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "|\\cdot|_{\\epsilon}" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": " between the parameterizations of poses " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "P, P'" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\log \\mathcal{N}(\\mathbf{y}_i^{j\\star}|\\mathbf{y}_i^j,\\sigma_i^j)" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": " computes the likelihood of a ground-truth track point " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{j\\star} \\in T_i^\\star" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": " under a probabilistic track-point estimate defined by a 2D gaussian with mean and variance predictions " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\sigma_i^j" + }, + { + "bbox": [ + 47, + 357, + 287, + 430 + ], + "type": "text", + "content": " respectively (i.e. the aleatoric uncertainty model described in Sec. 3.2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 445, + 127, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 127, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 127, + 458 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "text", + "content": "In this section, we first introduce the datasets together with the protocols for training and evaluation. Then, we provide comparison to existing methods and ablation studies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 506, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 673 + ], + "type": "text", + "content": "Datasets. Following prior work [25, 39, 74], we evaluate camera pose estimation on Co3Dv2 [55] and IMC Phototourism datasets [31], and 3D triangulation on ETH3D [60]. Co3D is an object-centric dataset comprising turntable-style videos from 51 categories of MS COCO [38]. IMC Phototourism, provided by Image Matching Challenge [31], contains 8 testing scenes and 3 validation scenes of famous landmarks. Generally, the Co3D scenes have much wider baselines, making them challenging for traditional frameworks such as COLMAP, while the IMC samples often have sufficiently overlapping fursta, which is where COLMAP excels. ETH3D provides highly accurate point clouds (captured by laser scanner) for 13 indoor and outdoor scenes, and hence is suitable for the evaluation of triangulation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Training. For the model evaluated on IMC Phototourism and ETH3D, we follow the protocol of [25, 39, 57] and train on the MegaDepth dataset [35]. MegaDepth con" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 357, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 441 + ], + "type": "text", + "content": "tains 1M crowd-sourced images depicting 196 tourism landmarks, auto-annotated with SfM tools. Hyper-parameters are tuned using the IMC validation set. As in prior work [25, 40], some scenes of MegaDepth are excluded from training due to their low quality or due to overlap with the IMC test set. For Co3Dv2, we conduct training and evaluation on 41 categories as in [36, 74, 85]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 445, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 563 + ], + "type": "text", + "content": "We chose a multi-stage training strategy for better stability. We first train the tracker T on the synthetic Kubric dataset [21] following the training protocol of [15, 32]. Then, the tracker is fine-tuned solely on Co3D or MegaDepth, depending on the test dataset. Subsequently, we train the camera initializer, with the tracker frozen. Next, the triangulator is trained with the aforementioned two components held frozen. Finally, all components are trained end-to-end. In all stages, we randomly sample a training batch of 3 to 30 frames." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Testing. Given input test frames " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", we first select the query frame by identifying the image that is closest to all others based on the cosine similarity between global descriptors extracted by an off-the-shelf ResNet50 [24]. Then, we extract SuperPoint and SIFT keypoints from the query frame to serve as the query points for the tracker " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Although our method can track any query point, it performs better when the queries are distinctive. To improve accuracy, we iterate the whole reconstruction function " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " multiple times until reaching sub-pixel BA reprojection error " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BA}}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". After the first iteration, the query image for each subsequent iteration is the one that is farthest from the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21691" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 84, + 147, + 159 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 111, + 83 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 111, + 83 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 111, + 83 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 84, + 147, + 159 + ], + "lines": [ + { + "bbox": [ + 48, + 84, + 147, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 147, + 159 + ], + "type": "image", + "image_path": "aafcf0ddb6d3eb1def661e0bba51f4556223b33b0003c81c16cba03e33de7a63.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 163, + 125, + 237 + ], + "blocks": [ + { + "bbox": [ + 69, + 163, + 125, + 237 + ], + "lines": [ + { + "bbox": [ + 69, + 163, + 125, + 237 + ], + "spans": [ + { + "bbox": [ + 69, + 163, + 125, + 237 + ], + "type": "image", + "image_path": "80075a9003b24193276a93f9a5956a0e05e136ba6187903daa1aecf422f056b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "text", + "content": "Figure 5. Qualitative Evaluation of Tracking. In each row, the left-most frame contains the query image with query points (crosses). The predicted track points " + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^j" + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "text", + "content": " (dots) are shown to the right. The top-right part also highlights our track-point confidence predictions (described in Sec. 3.2), illustrated as ellipses with extent proportional to the predicted variance " + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "inline_equation", + "content": "\\sigma_i^j" + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "text", + "content": ". Note how the uncertainty corresponds to an expectation, e.g., a keypoint covering a vertical stripe has higher uncertainty along the " + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "text", + "content": "-axis." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 158, + 84, + 257, + 159 + ], + "blocks": [ + { + "bbox": [ + 158, + 84, + 257, + 159 + ], + "lines": [ + { + "bbox": [ + 158, + 84, + 257, + 159 + ], + "spans": [ + { + "bbox": [ + 158, + 84, + 257, + 159 + ], + "type": "image", + "image_path": "71c979bdebe60df942874b8dae91efdf8f940377ad6bb37136b2e8737ab02f1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 259, + 84, + 358, + 159 + ], + "blocks": [ + { + "bbox": [ + 324, + 71, + 395, + 83 + ], + "lines": [ + { + "bbox": [ + 324, + 71, + 395, + 83 + ], + "spans": [ + { + "bbox": [ + 324, + 71, + 395, + 83 + ], + "type": "text", + "content": "Tracking Results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 259, + 84, + 358, + 159 + ], + "lines": [ + { + "bbox": [ + 259, + 84, + 358, + 159 + ], + "spans": [ + { + "bbox": [ + 259, + 84, + 358, + 159 + ], + "type": "image", + "image_path": "d98486beba8ef5e294d995b23c8d7716b7b309d66fbd06445ac1778f35257e1b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 360, + 84, + 459, + 159 + ], + "blocks": [ + { + "bbox": [ + 360, + 84, + 459, + 159 + ], + "lines": [ + { + "bbox": [ + 360, + 84, + 459, + 159 + ], + "spans": [ + { + "bbox": [ + 360, + 84, + 459, + 159 + ], + "type": "image", + "image_path": "37b5585f4a0a05ba48ce1022a036e51ad609ccef2806fb207417d0382f55567e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 461, + 76, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 461, + 76, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 461, + 76, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 461, + 76, + 547, + 159 + ], + "type": "image", + "image_path": "f2d9614eef7562234ca71e2aeb1c5cac6a76d2546f17eeddde3ca15d67b6270f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 157, + 163, + 264, + 237 + ], + "blocks": [ + { + "bbox": [ + 157, + 163, + 264, + 237 + ], + "lines": [ + { + "bbox": [ + 157, + 163, + 264, + 237 + ], + "spans": [ + { + "bbox": [ + 157, + 163, + 264, + 237 + ], + "type": "image", + "image_path": "2568c0bf9252837084175c3737cdf7ed89569449b9f2aa8d346e4a4e0d697c4f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 265, + 163, + 358, + 237 + ], + "blocks": [ + { + "bbox": [ + 265, + 163, + 358, + 237 + ], + "lines": [ + { + "bbox": [ + 265, + 163, + 358, + 237 + ], + "spans": [ + { + "bbox": [ + 265, + 163, + 358, + 237 + ], + "type": "image", + "image_path": "b56803d0e810aa036875e36d69780fa9e1acb0b086af5f3b4a2af733ec63603f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 163, + 455, + 237 + ], + "blocks": [ + { + "bbox": [ + 359, + 163, + 455, + 237 + ], + "lines": [ + { + "bbox": [ + 359, + 163, + 455, + 237 + ], + "spans": [ + { + "bbox": [ + 359, + 163, + 455, + 237 + ], + "type": "image", + "image_path": "38e8a6110c2ca96bd8346c0e7a57f7b9b15dbf2a058617ae51db700f05497954.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 455, + 163, + 550, + 237 + ], + "blocks": [ + { + "bbox": [ + 455, + 163, + 550, + 237 + ], + "lines": [ + { + "bbox": [ + 455, + 163, + 550, + 237 + ], + "spans": [ + { + "bbox": [ + 455, + 163, + 550, + 237 + ], + "type": "image", + "image_path": "32de1d2a21c79c0f5b2f33537f1d3b62f79ee87ca9e35ad9e1bdaf66c776a044.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 59, + 301, + 533, + 361 + ], + "blocks": [ + { + "bbox": [ + 59, + 301, + 533, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 301, + 533, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 301, + 533, + 361 + ], + "type": "table", + "html": "
Co3D DatasetIncremental SfMDeep
MethodCOLMAP (SP + SG)PixSfM (SP + SG)RelPose[85]PoseReg[74]RelPose++[36]PoseDiffusion[74]Ours w/o JointOurs
RRE@15°31.633.757.153.282.380.588.292.1
RTE@15°27.332.9-49.177.279.883.488.3
AUC@30°25.330.1-45.065.166.570.774.0
", + "image_path": "5b5f6d3ab871d7a691b02de27a45780637b573e4901b0710aed7fc0ecfa2635b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 50, + 411, + 286, + 511 + ], + "blocks": [ + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "type": "text", + "content": "Table 1. Camera Pose Estimation on Co3D, where the proposed method outperforms previous methods by a large margin. Ours w/o Joint indicates the variant of our framework without training all components jointly." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 411, + 286, + 511 + ], + "lines": [ + { + "bbox": [ + 50, + 411, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 50, + 411, + 286, + 511 + ], + "type": "table", + "html": "
IMC DatasetMethodAUC@3°AUC@5°AUC@10°
DeepDeepSFM10.2719.3631.35
PoseDiffusion12.3123.1736.82
Incremental SfMCOLMAP (SIFT+NN)23.5832.6644.79
PixSfM (SIFT + NN)25.5434.8046.73
PixSfM (LoFTR)44.0656.1669.61
PixSfM (SP + SG)45.1957.2270.47
DFSfM (LoFTR)46.5558.7472.19
DeepOurs w/o Joint38.2351.6068.35
Ours45.2358.8973.92
", + "image_path": "8b614ce7de853b292172c90a8c138b9393701c0eeecedb68a0b4663db46b8541.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "type": "text", + "content": "Table 2. Camera Pose Estimation on IMC. Our method achieves better accuracy than state-of-the-art Incremental SfM approaches on 2 out of 3 AUC thresholds." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": "query image of the previous iteration (as measured by the ResNet50 cosine similarity). The re-projections of the point cloud from the current iteration initialize the tracking in the next iteration." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 634, + 183, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 183, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 183, + 647 + ], + "type": "text", + "content": "4.1. Camera pose estimation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Following [25, 31, 39, 74], we report the metric area-under-curve (AUC) to evaluate camera pose accuracy. In Co3D, similar to [74], we also report the relative rotation error (RRE) and relative translation error (RTE). More specifically, for every pair of input frames, we compute the angular" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "type": "text", + "content": "translation and rotation error, which are later compared to a threshold yielding accuracies RTE and RRE respectively. For a range of thresholds, AUC picks the lower between RRE and RTE, and outputs the area under the accuracy-threshold curve. The results on IMC and CO3D are presented in Tab. 1 and Tab. 2 respectively. For a fair comparison on IMC, we finetune DeepSFM [78] and PoseDiffusion [74] on MegaDepth using their open-source code. The results of Incremental SfM methods are copied from Detector-free SfM [25]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "content": "Results indicate that VGGSfM outperforms existing methods by a large margin (+9 accuracy points for each metric) on the CO3D dataset. Here, traditional SfM pipelines suffer because of the wide baselines between test frames. On IMC, with a good overlap between views, traditional SfM remains superior to recent data-driven deep-learning pipelines [74, 78]. Our end-to-end trained VGGSfM, however, outperforms all other methods on AUC@10 and AUC@5, and ranks second on AUC@3, convincingly demonstrating its ability to perform well in both narrow- and wide-baseline regimes. The accuracy and completeness of our point clouds can be further qualitatively evaluated in Fig. 4." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21692" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 286, + 162 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 286, + 162 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 286, + 162 + ], + "type": "table", + "html": "
MethodAccuracy (%)Completeness (%)
1cm2cm5cm1cm2cm5cm
PatchFlow (LoFTR)66.7378.7389.933.4811.3430.96
PixSfM (LoFTR)74.4284.0892.632.919.3927.31
PixSfM (SIFT + NN)76.1885.6093.160.170.713.29
PixSfM (SP + SG)79.0187.0493.800.752.7711.28
DFSfM (LoFTR)80.3889.0195.833.7311.0729.54
Ours80.6289.4996.524.5213.1133.96
", + "image_path": "d55c3f00fb44fed494b689d711a3de6d52d9884c90830a134adc12d7bab576cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 218, + 147, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 147, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 147, + 231 + ], + "type": "text", + "content": "4.2.3D triangulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "text", + "content": "We evaluate the accuracy and completeness of 3D triangulation on the ETH3D dataset [60] using the same protocol as [17, 25, 39], which triangulates points with fixed camera poses and intrinsics. Results are shown in Tab. 3, where metrics are averaged over all scenes. Our VGGsFm achieves better accuracy and completeness than all baselines (PatchFlow [17], PixSfM [39], and DFSfM [25]), regardless of which keypoint detection or matching method they use. This is especially obvious from the completeness attained at the " + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "text", + "content": " threshold, with our " + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "inline_equation", + "content": "33.96\\%" + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "inline_equation", + "content": "29.54\\%" + }, + { + "bbox": [ + 46, + 237, + 288, + 369 + ], + "type": "text", + "content": " of the best prior work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 138, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 138, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 138, + 388 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": "End-to-end Training. As reported in Tab. 1 and Tab. 2, the end-to-end joint training of the whole framework is important for achieving state-of-the-art performance. Specifically, comparing VGGsF to an ablation which lacks end-to-end training (Ours w/o joint) we record an improvement from " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "70.7\\%" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": " AUC@30 to " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": " on the Co3D dataset, and " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "68.35\\%" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": " AUC@10 to " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "73.92\\%" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": " on the IMC dataset. This demonstrates the benefits of our fully-differentiable design, and the synergy between its components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Tracking or Pairwise Matching. We compare the performance of our predicted tracks to pairwise matching methods on the IMC dataset. Specifically, we split our 2D tracks into pairwise matches and feed these matches to PixSfM. Also, we construct 2D tracks by pairwise matching (based on the open-source implementation of PixSfM) and feed them to our framework. It is worth noting that tracks from pairwise matching have a lot of \"holes\" because pairwise matching cannot guarantee proper point tracking. We fix these holes by setting their locations as the point locations in the query frame, marking them as invisible. At the same time, for fair comparison, cameras are still initialized using our tracks, because SP+SG cannot provide track features to our camera initializer. The results are shown in Tab. 4. Although COLMAP (the basis of PixSfM) is designed and carefully engineered for pairwise matching, our tracks achieve a slightly better result than the state-of-the-art matching op" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 545, + 97 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "Table 3. 3D Triangulation on ETH3D [60] reporting the accuracy and completeness metrics at different thresholds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 70, + 545, + 97 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 97 + ], + "type": "table", + "html": "
PixSfM(SP + SG)PixSfM(Our Tracks)Ours(SP + LG)Ours
AUC@10°70.4770.6268.7873.92
", + "image_path": "526c4b5c968c6a3ddaafda2bd845ac0d7ee9e8970df11e6c593d0ca97cb541bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 148, + 520, + 186 + ], + "blocks": [ + { + "bbox": [ + 305, + 105, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 105, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 105, + 545, + 137 + ], + "type": "text", + "content": "Table 4. Tracking or Pairwise Matching. We respectively provide tracks predicted by our tracker or matches estimated by SP+SG to PixSfM and to our VGGSfM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 148, + 520, + 186 + ], + "lines": [ + { + "bbox": [ + 332, + 148, + 520, + 186 + ], + "spans": [ + { + "bbox": [ + 332, + 148, + 520, + 186 + ], + "type": "table", + "html": "
PoseDiffusionOur Camera Initialized
DLT62.1869.42
Our Triangulator66.3773.92
", + "image_path": "737c144d44999c8bfe6a1e2075a1fa85167e26a147792fd53d1fca19fa14764b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 194, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 239 + ], + "type": "text", + "content": "Table 5. Ablation Study for Camera Initializer and Triangulator. A clear performance drop is observed when replacing our camera initializer by deep camera prediction method PoseDiffusion, or replacing triangulator by DLT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "content": "tion " + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{SP + SG}" + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "content": ". Instead, directly feeding " + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{SP + SG}" + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "content": " tracks to our framework leads to a performance drop. We attribute this to the fact that using " + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{SP + SG}" + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "content": " tracks cannot benefit from the joint training. We also provide a qualitative evaluation of our tracking accuracy in Fig. 5 on the IMC dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "text", + "content": "Camera Controller and Triangulator. We also validate the design of our camera controller " + }, + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_{\\mathcal{P}}" + }, + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "text", + "content": " and triangulator " + }, + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathfrak{T}_X" + }, + { + "bbox": [ + 304, + 322, + 545, + 407 + ], + "type": "text", + "content": " on the IMC dataset. As reported in Tab. 5, AUC@10 drops clearly if we replace them with alternatives, proving that they provide sufficiently accurate initialization for our global bundle adjustment BA, without the need for incremental camera registration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "content": "Coarse-to-fine Tracking. As discussed above, accurate correspondences are important for structure from motion. We demonstrate the significance of our coarse-to-fine tracking mechanism for the method performance. By conducting an ablation study where the fine tracker is removed, we observe a significant performance drop on the IMC dataset, with AUC@10 dropping from " + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "inline_equation", + "content": "73.92\\%" + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "inline_equation", + "content": "62.30\\%" + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 712 + ], + "type": "text", + "content": "In this paper, we have presented VGGSfM, a fully differentiable SfM approach. We find that even long-standing pipelines, such as Structure-from-Motion, benefit from a learned adaptation between their components. This allows VGGSfM to be simpler than traditional SfM frameworks while achieving better performance across benchmark datasets. Moreover, our framework is fully implemented in Python, which will allow for easy modification and improvements in the future. While VGGSfM already achieves good performance, it cannot yet compete with established pipelines in all application domains. For example, it currently lacks the capability to process thousands of images as in traditional SfM frameworks. Nonetheless, we find differentiable SfM a promising direction of research, and our approach lays the foundation for further advances." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21693" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Noah Snavely, Steven M Seitz, and Richard Szeliski. Bundle adjustment in the large. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 29-42. Springer, 2010. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 202 + ], + "type": "text", + "content": "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 287, + 224 + ], + "type": "text", + "content": "[3] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "type": "text", + "content": "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "text", + "content": "[5] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4322-4331, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 361 + ], + "type": "text", + "content": "[6] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684-6692, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "text", + "content": "[7] Lucas Brynte, José Pedro Iglesias, Carl Olsson, and Fredrik Kahl. Learning structure-from-motion with graph attention networks. arXiv preprint arXiv:2308.15984, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "text", + "content": "[8] Jonathan L. Carrivick, Mark W. Smith, and Duncan J. Quincey. Structure from Motion in the Geosciences. John Wiley & Sons, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "text", + "content": "[9] Pierre Charbonnier, Laure Blanc-féaud, Gilles Aubert, and Michel Barlaud. Deterministic edge-preserving regularization in computed imaging. IEEE Trans. Image Processing, 6:298-311, 1997. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 520 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 577 + ], + "type": "text", + "content": "[11] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "text", + "content": "[12] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. In Proceedings of the IEEE International Conference on Computer Vision, pages 864-872, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurlPS Datasets Track, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[17] Mihai Dusmanu, Johannes L Schonberger, and Marc Pollefeys. Multi-view optimization of local feature geometry. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 670-686. Springer, 2020. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 386 + ], + "type": "text", + "content": "[19] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 368-381. Springer, 2010. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[20] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Towards Internet-scale multi-view stereo. In Proc. CVPR. IEEE, 2010. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 542 + ], + "type": "text", + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliaasacchi. Kubric: a scalable dataset generator. 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "type": "text", + "content": "[22] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "type": "text", + "content": "[23] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 1, 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. arXiv preprint arXiv:1512.03385, 2015. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[25] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure from motion. In *arxiv*, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Jared Heinly, Johannes L. Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world* in six days " + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\ast)" + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": " (as captured by the yahoo 100 million image dataset). In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21694" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[27] Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Piermattei, James O'Connor, and Jacqueline Rosette. Structure from motion photogrammetry in forestry: A review. Current Forestry Reports, 5:155-168, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "type": "text", + "content": "[28] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. ArXiv, 2212.04492, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 221 + ], + "type": "text", + "content": "[29] San Jiang, Cheng Jiang, and Wanshou Jiang. Efficient structure from motion for large-scale UAV images: A review and a comparison of sfm tools. ISPRS Journal of Photogrammetry and Remote Sensing, 167:230-251, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 277 + ], + "type": "text", + "content": "[30] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 334 + ], + "type": "text", + "content": "[31] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 335, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 369 + ], + "type": "text", + "content": "[32] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[33] Alex Kendall and Yarin Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Proc. NeurIPS, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 449 + ], + "type": "text", + "content": "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[35] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "text", + "content": "[36] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In Proc. ECCV, 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[39] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. arXiv.cs, abs/2108.08291, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[40] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023. 3, 4, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[41] Yin Lou, Noah Snavely, and Johannes Gehrke. Matchminer: Efficient spanning structure mining in large image collections. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12, pages 45-58. Springer, 2012. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 161 + ], + "type": "text", + "content": "[42] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "text", + "content": "[43] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[44] Xiao Xin Lu. A review of solutions for perspective-n-point problem in camera pose estimation. In Journal of Physics: Conference Series, page 052009. IOP Publishing, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[45] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[46] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust Wide Baseline Stereo from Maximally Stable Extremal Regions. In Proc. BMVC, 2002. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "text", + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Proc. ECCV, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "text", + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[49] Jorge J Moré. The levenberg-marquardt algorithm: implementation and theory. In Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977, pages 105-116. Springer, 2006. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "text", + "content": "[50] Pierre Moulon, Pascal Monasse, and Renaud Marlet. Global fusion of relative motions for robust, accurate and scalable structure from motion. In Proceedings of the IEEE international conference on computer vision, pages 3248-3255, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "type": "text", + "content": "[51] David Novotny, Diane Larlus, and Andrea Vedaldi. Learning 3d object categories by looking around them. In Proc. ICCV, 2017. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[52] John Oliensis. A critique of structure-from-motion algorithms. Computer Vision and Image Understanding, 80(2): 172-214, 2000. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "text", + "content": "[53] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "text", + "content": "[54] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, et al. Theses: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 35:3801-3818, 2022. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[55] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21695" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[56] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 196 + ], + "type": "text", + "content": "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[58] Frederik Schaffalitzky and Andrew Zisserman. Multi-view Matching for Unordered Image Sets, or \"How Do I Organize My Holiday Snaps?\" In Proc. ECCV, 2002. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[59] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 342 + ], + "type": "text", + "content": "[60] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "text", + "content": "[61] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[62] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349-21359, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[63] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[64] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 577 + ], + "type": "text", + "content": "[65] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. arXiv preprint arXiv:1806.04807, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "text", + "content": "[66] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[67] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[68] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[69] Bill Triggs, Philip F. McLauchlan, Richard I. Hartley, and Andrew W. Fitzgibbon. Bundle Adjustment - A Modern Synthesis. In Proc. ICCV Workshop, 2000. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[70] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "text", + "content": "[71] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "text", + "content": "[72] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. NeurIPS, 2017. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "text", + "content": "[73] Jianyuan Wang, Yiran Zhong, Yuchao Dai, Stan Birchfield, Kaihao Zhang, Nikolai Smolyanskiy, and Hongdong Li. Deep two-view structure-from-motion revisited. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8953-8962, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 545, + 370 + ], + "type": "text", + "content": "[74] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[75] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, pages 2746-2762, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[76] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 514 + ], + "type": "text", + "content": "[77] Tong Wei, Yash Patel, Alexander Shekhotsov, Jiri Matas, and Daniel Barath. Generalized differentiable ransac. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17649-17660, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "type": "text", + "content": "[78] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 230–247. Springer, 2020. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "text", + "content": "[79] Matthew J Westoby, James Brasington, Niel F Glasser, Michael J Hambrey, and Jennifer M Reynolds. 'structure-from-motion'photogrammetry: A low-cost, effective tool for geoscience applications. Geomorphology, 179:300-314, 2012. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 678 + ], + "type": "text", + "content": "[80] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 61-75. Springer, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[81] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21696" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[82] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "text", + "content": "[83] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. MagicPony: Learning articulated 3d animals in the wild. 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "text", + "content": "[84] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned Invariant Feature Transform. In Proc. ECCV, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "text", + "content": "[85] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 296 + ], + "type": "text", + "content": "[86] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "type": "text", + "content": "[87] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21697" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_content_list.json b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1195c9bec61ae0704cb563121a76c34dede72e00 --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_content_list.json @@ -0,0 +1,1911 @@ +[ + { + "type": "text", + "text": "VILA: On Pre-training for Visual Language Models", + "text_level": 1, + "bbox": [ + 222, + 130, + 750, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ji Lin $^{1,2*†}$ Hongxu Yin $^{1*}$ Wei Ping $^{1}$ Pavlo Molchanov $^{1}$ Mohammad Shoeybi $^{1}$ Song Han $^{1,2}$ $^{1}$ NVIDIA ${}^{2}$ MIT", + "bbox": [ + 202, + 178, + 759, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ViLA https://github.com/Efficient-Large-Model/VILA", + "bbox": [ + 264, + 236, + 705, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pretraining data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM.", + "bbox": [ + 75, + 334, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 728, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have demonstrated superior capabilities for natural language tasks [3, 7, 9, 14, 15, 18, 30, 43, 48, 54-56]. Augmenting LLMs to support visual inputs allows the final model to inherit some of the appealing properties like instruction following, zero-shot generalization, and few-shot in-context learning (ICL), empowering various visual language tasks [1, 2, 5, 8, 13, 19, 33, 37, 68]. The central challenge of unifying vision and language for", + "bbox": [ + 75, + 753, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27534e03254453598560de2a3d0d2ed6661a6a83ec42c3d343e5beda0bcfdcd4.jpg", + "image_caption": [ + "Figure 1. VILA's enhanced visual-language pre-training consistently improves the downstream task accuracy under a comparison to recent methods [7, 17, 37]." + ], + "image_footnote": [], + "bbox": [ + 511, + 284, + 875, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "collaborative inference resides in connecting the LLM and the vision foundation model (e.g., a CLIP encoder): both foundation models are usually pre-trained individually, before aligned via vision-language joint training. Most of the efforts in this field have been focusing on improving the visual language instruction-tuning process, i.e., supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF) [36, 37, 52]. However, there lacks a thorough study of the pre-training process, where the model is trained on image-text datasets/corpora at scale [10, 49, 69]. This process is costly but critical for the modality alignment.", + "bbox": [ + 496, + 595, + 893, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to explore different design options for enhanced visual language model pre-training. In particular, we aim to answer \"How do various design choices in visual language model pre-training impact the downstream performance?\" We followed the pre-training + SFT pipeline and ablated different design options for pre-training overseeing dataset properties and training protocols. We discover several findings: (1) Freezing the LLM during pretraining can achieve a decent zero-shot performance, but", + "bbox": [ + 496, + 763, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Work done during an internship at NVIDIA.", + "bbox": [ + 99, + 886, + 460, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26689", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebb73adf99d692f0cd74a3d6f8a92538fad46860aa9bb5e398b19b8a0b96f011.jpg", + "image_caption": [ + "Figure 2. We study auto-regressive visual language model, where images are tokenized and fed to the input of LLMs. We find updating the LLM is essential for in-context learning capabilities, and interleaved corpus like [69] helps pre-training. Joint SFT with text-only data helps maintain the text-only capabilities." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 478, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "not in-context learning (ICL) capability, whereas updating the LLMs encourages deep embedding alignment, which we found is important for ICL; (2) Interleaved visual language data is essential for pre-training, that provides accurate gradient update and maintains text-only capability; (3) Adding in text-only instruction data during SFT can further remedy text-only degradation and boost visual language task accuracy.", + "bbox": [ + 75, + 327, + 468, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce practical guidance to design Visual Language models, dubbed VILA. Without bells and whistles, VILA outperforms the state-of-the-art model [36] by noticeable margins across a wide range of vision language tasks (Figure 1), thanks to the help of improved pre-training. Moreover, we observe that the pre-training process unlocked several interesting capabilities for the model, such as (i) multi-image reasoning (despite the model only sees single image-text pairs during SFT), (ii) stronger in-context learning capabilities, and (iii) enhanced world knowledge. We hope our findings can provide a good pre-training recipe for future visual language models.", + "bbox": [ + 75, + 449, + 470, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 76, + 642, + 205, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model architecture. Multi-modal LLMs can be generally categorized into two settings: cross-attention-based [5, 33] and auto-regressive-based [2, 19, 37]. The latter VLM family tokenizes images into visual tokens, which are concatenated with textual tokens and fed as the input to LLMs (i.e., treating visual input as a foreign language). It is a natural extension of text-only LLMs by augmenting the input with visual embeddings and can handle arbitrary interleaved image-text inputs. In this study, we focus on the pre-training of auto-regressive VLMs due to its flexibility and popularity. As shown in Figure 2, auto-regressive VLMs consists of three components: a visual encoder, an LLM, and a projector that bridges the embeddings from the two modalities. The projector can be a simple linear layer [37] or more capable Transformer blocks [6, 17] – we will compare their efficacy in our experiments. The model takes visual and text input", + "bbox": [ + 75, + 657, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and generates text outputs.", + "bbox": [ + 500, + 92, + 676, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training stages. Following common practice [6, 19, 37], we study how to augment a pre-trained text-only LLM with visual input support. The training can be categorized into three stages:", + "bbox": [ + 496, + 119, + 893, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "0. **Projector initialization.** The LLM and ViT are separately pre-trained, while the projector is usually initialized from random weights. Therefore, we first pre-train the projector while freezing both ViT and LLMs on image-caption pairs following existing literature [17, 33, 37].", + "1. visual language pre-training. We then pre-train the model (LLM and the projector) on visual language corpus. We consider two types of corpus: interleaved image-text corpus (e.g., MMC4 [69]) and image-text pairs (e.g., COYO [10] and LAION [49]). We focus the study of this work on the pre-training process, which are most costly and important for visual language alignment.", + "2. Visual instruction-tuning. Finally, we further perform instruction tuning of the pre-trained model on visual language instruction datasets. We convert existing visual language datasets into FLAN [59] style (i.e., with dataset-specific prompts) following [17]. Please find the data blend of the visual instruction data in the supplementary." + ], + "bbox": [ + 496, + 181, + 893, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Evaluations.** During our ablation study, we evaluate the fine-tuned model on 4 visual language tasks: accuracy for OKVQA [42] and TextVQA [50], and CIDEr score for COCO [35] and Flickr [62]. We evaluate both 0-shot and 4-shot performance, which reflects the models' in-context learning capability.", + "bbox": [ + 496, + 470, + 892, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. On Pre-training for Visual Language Models", + "text_level": 1, + "bbox": [ + 498, + 577, + 890, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss practical design choices and learned lessons for the visual language pre-training process.", + "bbox": [ + 496, + 604, + 893, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Updating LLM is Essential", + "text_level": 1, + "bbox": [ + 498, + 646, + 743, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-tuning vs. prompt tuning. There are two popular ways to augment a pre-trained text-only LM with visual inputs: fine-tune LLMs on the visual input tokens [19, 37], or freeze the LLM and train only the visual input projector as prompt tuning [17, 33]. The latter is attractive since freezing the LLMs prevents the degradation of the pre-trained text-only LLM. Nonetheless, we found updating the base LLM is essential to inherit some of the appealing LLM properties like in-context learning.", + "bbox": [ + 496, + 662, + 893, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the idea, we compare the two training protocols in Table 1. We use a Transformer block for the projector instead of a single linear layer [37] in setting a-c, which provides enough capacity when freezing LLMs. We use MMC4-core $[69]^*$ for the comparison. We observed that:", + "bbox": [ + 496, + 799, + 893, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*We downloaded only 25M of 30M images amid some expired URLs.", + "bbox": [ + 517, + 886, + 887, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26690", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e902d9ae2ce39ab62a7ea2fc341aa46636839292a712a681c8a6d08e7af83ad2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Ablation study on whether to train LLM or freeze LLM and only perform prompt tuning during visual language pre-training (PreT). Interestingly, freezing the LLM during pre-training does not hurt the 0-shot accuracy, but leads to worse in-context learning capability (worse 4-shot). Using a simple linear projector forces the LLM to learn more and leads to better generalization. We report accuracy for VQA datasets (OKVQA, TextVQA) and CIDEr score for captioning (COCO and Flickr). Note: we used a different evaluation setting just for ablation study; the absolute value in this setting is lower and should not be compared against other work." + ], + "table_body": "
PreTSFTProjectileOKVQATextVQACOCOFlickrAverage
TrainLLM?0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot
(a)XXTransformer10.419.214.823.117.460.211.047.413.437.5
(b)X✓Transformer47.147.737.236.6109.488.073.658.166.857.6
(c)✓✓Transformer44.849.838.538.8112.3113.571.572.966.868.8
(d)✓✓Linear45.250.339.740.2115.7118.574.274.768.770.9
", + "bbox": [ + 88, + 88, + 877, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb9a16df2a643fb262bb55a9c4b08fc72f95f9297f083535f9a60604ed5f4d8f.jpg", + "image_caption": [ + "Figure 3. Prompt-tuning to support visual tokens can only enable shallow alignment, while fine-tuning the LLM leads to alignment at deeper layers. From configuration (b) to (d) (as in Table 1), the alignment improves at deeper layer, so as ICL accuracy (4-shot)." + ], + "image_footnote": [], + "bbox": [ + 107, + 284, + 433, + 439 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Training only the projector during SFT leads to poor performance (setting a), despite using a high-capacity design. It is rewarding to fine-tune LLM during SFT.", + "(2) Interestingly, freezing the LLM during pre-training does not affect 0-shot performance, but degrades in-context learning capabilities (i.e., 4-shot, comparing setting b and c). The gap is even larger for captioning datasets (COCO & Flickr) since they are out-of-distribution (the instruction tuning data is mostly VQA-alike, see supplementary), showing the worse generalization capability when freezing LLMs.", + "(3) When using a small-capacity projector (a linear layer instead of a Transformer block), the accuracy is slightly better (comparing c and d). We hypothesize a simpler projector forces the LLM to learn more on handling visual inputs, leading to better generalization." + ], + "bbox": [ + 75, + 512, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The deep embedding alignment hypothesis. To understand why fine-tuning LLM is beneficial, we hypothesize that it is important to align the distribution of visual and textual latent embeddings (especially in the deeper layers), so that the model can seamlessly model the interaction between the two modalities. It is essential if we want to inherit some of the good properties of LLM like in-context learning for visual language applications.", + "bbox": [ + 75, + 750, + 470, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To verify the idea, we calculate the Chamfer distance of visual and textual embeddings in different layers to measure", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/689ba363ccef0585a91033c231720bd5d9618375d8f0ed923d50a804ca551df9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Two image-text corpus considered for pre-training. The COYO captions are generally very short, which has a different distribution compared to the text-only corpus for LLM training. We sample each data source to contain 25M images by choosing samples with high CLIP similarities." + ], + "table_body": "
DatasetTypeText Src.#img/sample#tok./img
MMC4 [69]InterleaveHTML4.0122.5
COYO [10]Img-text pairAlt-text122.7
", + "bbox": [ + 500, + 282, + 895, + 342 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0f0f26a67643d4c03fd2ea00b878cb56cf449d8fb750190eed0e4275a775eefa.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Pre-training on MMC4 data provides better visual language accuracy (0-shot and few-shot) and smaller degradation on text-only accuracy compared to caption data (COYO). The benefits comes from the interleave nature but not the better text distribution (MMC4 vs. MMC4-pairs). Blending interleaved and caption data provides a better diversity and downstream accuracy." + ], + "table_body": "
Pre-train DataVLM acc (avg)MMLU acc.
0-shot4-shot
Llama-2--46.0%
COYO51.1%50.3%28.8% (-17.2%)
MMC4-pairs46.4%44.5%32.4% (-13.6%)
MMC468.7%70.9%40.7% (-5.3%)
MMC4+COYO69.0%71.3%40.2% (-5.8%)
", + "bbox": [ + 521, + 436, + 870, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "how well they align in Figure 3. We calculate the pairwise cosine similarity to exclude the affect of magnitude. From configuration (b) to (d), the similarity of deeper layer goes higher, so as the 4-shot accuracy in Table 1, showing the positive relationship between deep embedding alignment and in-context learning.", + "bbox": [ + 496, + 684, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the observations, we fine-tune the LLM during both pre-training and instruction-tuning in later studies, and use a simple linear projection layer.", + "bbox": [ + 496, + 775, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Interleaved Visual Language Corpus Helps Pre-training", + "text_level": 1, + "bbox": [ + 498, + 830, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to \"augment\" the LLM to support visual input, instead of training a model that only works well on visual", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26691", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7d481958cc4c790af80703c2561cfdda658bb50d21d459440399a781ef17e21.jpg", + "image_caption": [ + "Figure 4. A sample from MMC4 [69] dataset consisting of interleaved images and text segments. The images are placed before the corresponding text. The text are weakly conditioned on images: only colored text can be better inferred with the help of images." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 168, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b97409618609b8ed58bd87e4b7804e973e6a0857ddb3bc4f48267edd734704a.jpg", + "image_caption": [], + "image_footnote": [ + "...,\"Easy to grow, and fruitful, these are a favorite amongst those who like to have their own at home garden.\", , \"There is a large number of different kinds of tomatoes, meaning that there is one for every different size garden.\", \"You can choose from a variety of these plants to accomplish different food missions, be them small cucumbers for pickling or larger ones for slicing.\", ..., , \"These tender green beans are durable and quite at home growing in the hot season air.\", \"To get the most out of planting them, space the seeds 4 to 6 inches apart and allow 3 feet between rows, ..." + ], + "bbox": [ + 169, + 90, + 256, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd1a94b4a45456d974e21204bb803f015894c921fc39734db13a38a88618ee65.jpg", + "image_caption": [ + "Figure 5. Loss is lower when pre-training on MMC4 compared to MMC4-pairs (samples broken into image-text pairs), since the text segments provide more information for language modeling." + ], + "image_footnote": [], + "bbox": [ + 86, + 196, + 455, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "language inputs. Therefore, it is essential to preserve the text-only capabilities of LLMs. We found that data blending is a key factor, both for pre-training and instruction tuning.", + "bbox": [ + 75, + 367, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training dataset options. Most of the VLM pretraining [33, 37, 58] relies on image-text pairs (i.e., image and captions) due to the wide availability and large diversity (e.g., LAION [49], COYO [10]). On the other hand, interleaved image-text datasets (MMC4 [69], M3W [5]) follow a more similar distribution compared to the text-only corpus and is found to be important in Flamingo-style model training [5]. We hypothesize that the interleaved dataset is even more important for VLMs when LLM backbone is updated to accommodate the visual input. For a better understanding of the two data types, we compare statistics in Table 2: COYO suffers from a short text distribution since the accompanying text is taken from alt-text. We subsample the COYO dataset by ranking CLIP similarities and keep only 25M images (a similar size as MMC4-core).", + "bbox": [ + 75, + 422, + 470, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the same pre-training + SFT process and ablate different pre-training corpus. We compare the 0-shot and few-shot visual language accuracy as well as text-only accuracy (MMLU [26]) in Table 3. Due to space limit, we report the average accuracy over four datasets (as in Table 1).", + "bbox": [ + 75, + 648, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interleaved data is essential. We notice using image-text pairs (i.e., COYO) for pre-training can lead to catastrophic forgetting. The text-only accuracy (MMLU) degrades by $17.2\\%$ . The visual language accuracy is also much worse compared to MMC4 pre-training. Noticeably, the 4-shot accuracy is even worse than 0-shot, showing the model cannot properly do in-context learning for visual language inputs (probably because it never sees more than one image during pre-training). We hypothesize the catastrophic forgetting is due to the distribution of text-based captions, which are generally very short and concise.", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the contrary, dataset like MMC4 has a much closer distribution compared to text-only corpus (e.g., C4 [48]). When using the interleaved data for pre-training, the degradation on MMLU is only $\\sim 5\\%$ . The degradation would be even smaller when using a larger base LLM [19]. With proper instruction tuning (Section 3.3), this degradation can be fully recovered. It also promotes visual in-context learning, leading to a higher 4-shot accuracy compared to 0-shot.", + "bbox": [ + 496, + 198, + 893, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interleave data structure matters, but not the text distribution. We further question whether the benefits come from the better text distribution (e.g., longer) or from the interleave nature. To ablate this, we construct a new MMC4 variant by only keeping the images and their corresponding text segments, without considering the interleave nature, denoted as \"MMC4-pairs\". For example an MMC4 sample may look like:", + "bbox": [ + 496, + 328, + 893, + 449 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n< \\text {t x t 1} > < \\text {i m 1} > < \\text {t x t 2} > < \\text {t x t 3} > < \\text {i m 2} > < \\text {t x t 4} >\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 455, + 846, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It will be converted into two MMC4-pairs samples:", + "bbox": [ + 503, + 479, + 849, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n< \\mathbf {i m} 1 > < \\mathbf {t x t} 2 >, < \\mathbf {i m} 2 > < \\mathbf {t x t} 4 >\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 503, + 803, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, training on MMC4-pairs does not lead to a satisfactory result: it slightly reduces the degradation on MMLU due to a longer text distribution, but the VLM accuracy is even lower compared to pre-training on COYO; there is also no in-context improvement. We hypothesize the MMC4 samples do not have a very strict image-text correspondence; the image only provides marginal information for text modeling (i.e., most of the information is still from pure text modeling; an example is provided in Figure 4). It is also demonstrated by the loss curves in Figure 5, where training on the interleave corpus leads to a much lower loss, indicating the full text segments provides more information. Therefore, the interleaved data structure is critical, allowing the model to pick up the image-related information, without over-forcing it to learn unrelated text modeling.", + "bbox": [ + 496, + 527, + 893, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data blending improves pre-training. Training on image-text pairs only led to a sharp degradation on text-only accuracy (more than $17\\%$ ). Luckily, blending the interleaved corpus and image-text pairs allows us to introduce more diversity in the corpus, while also preventing the severe degradation. Training on MMC4+COYO further boosts the accuracy on visual language benchmarks (the gain is larger when we perform joint SFT, as we will show later (Table 4).", + "bbox": [ + 496, + 763, + 890, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$^{\\dagger}$ We followed [69] to match the image and text segments by CLIP scores.", + "bbox": [ + 514, + 886, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26692", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9633c0d723bfee93ed173248fb51b402c9c49175b34798f4fc32452826f0d844.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Joint SFT (Vis. + Text) not only bridges the degradation of text-only capability (MMLU acc.), but also improves the performance on visual-language tasks (both zero-shot and few-shot)." + ], + "table_body": "
PT dataSFT dataVLM acc. (avg)MMLU acc.
0-shot4-shot
Llama-2---46.0%
MMC4Visual68.7%70.9%40.7% (-5.3%)
MMC4+COYOVisual69.0%71.3%40.2% (-5.8%)
Llama-2Text--51.2%
MMC4Vis.+Text71.0%72.1%51.4% (+0.2%)
MMC4+COYOVis.+Text72.3%73.6%50.9% (-0.3%)
", + "bbox": [ + 78, + 88, + 470, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Recover LLM Degradation with Joint SFT", + "text_level": 1, + "bbox": [ + 76, + 272, + 441, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the interleave data helps maintain the text-only capability, there is still a $5\\%$ accuracy drop. A potential approach is to maintain the text-only capability would be to add in text-only corpus (the one used in the LLM pre-training). However, such text corpus are usually proprietary even for open-source models; it is also unclear how to subsample the data to match the scale of vision-language corpus. Luckily, we found the text-only capabilities are temporarily hidden, but not forgotten. Adding in text-only data during SFT can help bridge the degradation, despite using a much smaller scale compared to the text pre-training corpora (trillion scale).", + "bbox": [ + 75, + 297, + 473, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint supervised fine-tuning. The common way for instruction tuning is to fine-tune the model on some visual language datasets (VQA/Caption style [17] or GPT-generated [37]). We found blending in text-only instruction data can simultaneously (i) recover the degradation in text-only accuracy, and (ii) improve the visual language accuracy. To this end, we also blended in 1M text-only instruction tuning data sampled from FLAN [16], which we termed as joint SFT. We provide the comparison in Table 4.", + "bbox": [ + 75, + 482, + 470, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can see that blending in the text-only SFT data not only bridges the degradation on text-only capability (the MMLU accuracy is on par compared to the original Llama-2 model fine-tuned on the same text-only instruction data), but also improves the visual language capability. We hypothesize that the text-only instruction data improves the model's instruction-following capability, which is also important for visual language tasks. Interestingly, the benefits of blending in COYO data is more significant with joint SFT. We believe that with joint SFT, the model no longer suffers from the text-only degradation when pre-trained with short captions, thus unlocking the full benefits from the better visual diversity.", + "bbox": [ + 75, + 619, + 470, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 819, + 210, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Scaling up VLM pre-training", + "text_level": 1, + "bbox": [ + 76, + 845, + 339, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We scale up the training of VLM in the following aspects to form our final model:", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Higher image resolution. Above ablation studies used the OpenAI CLIP-L [46] with $224 \\times 224$ resolutions as the visual encoder. We now use $336 \\times 336$ image resolutions to include more visual details for the model, which can help tasks requiring fine-grained details (e.g., TextVQA [50]).", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Larger LLMs. By default, we used Llama-2 [56] 7B for ablation study. We also scaled to a larger LLM backbone (e.g., Llama-2 [56] 13B) to further improve the performance.", + "bbox": [ + 496, + 175, + 893, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training data. We used both interleaved image-text data and image-text pairs for pre-training (we sample roughly 1:1 image proportions) to improve the data diversity. The total the pre-training corpus contains about 50M images. It is smaller than the billion-scale pre-training data [5, 13, 58], but already demonstrates impressive improvements on downstream tasks.", + "bbox": [ + 496, + 229, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SFT data. We also include a better SFT data blend from LLaVA-1.5 [36], which is more diverse (e.g., contains reference-based annotations) and has high-quality prompt. The new SFT data blend can significantly improve the downstream evaluation metrics. We include details the Appendix.", + "bbox": [ + 496, + 345, + 893, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations. Due to the limited compute budget, we have not been able to further scale up the size of the pre-training corpus to billion-scale, which we leave as future work. Nonetheless, pre-training on 50M images already demonstrated significant performance improvement.", + "bbox": [ + 496, + 430, + 893, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 498, + 513, + 720, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "visual language tasks. We perform a comprehensive comparison with state-of-the-art models on 12 visual language benchmarks in Table 5. Compared to existing models (e.g., LLaVA-1.5 [36]), our model achieves consistent improvements over most datasets at different model sizes under a head-to-head setting (using the same prompts and base LLM; Vicuna-1.5 is based on Llama-2). Remarkably, we 7B model is able to outperform LLaVA-1.5 13B on VisWiz [25] and TextVQA [50] by a large margin thanks to the pre-training. Our 7B model even outperforms the 13B LLaVA model on these datasets. Our model also has multi-lingual capability despite the vision-language instruction data is in English, outperforming LLaVA-1.5 on MMBench-Chinese benchmark. Our results demonstrate the benefits of vision-language pre-training on downstream tasks, even when using a high-quality instruction tuning dataset [36].", + "bbox": [ + 496, + 529, + 893, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-only performance. Our goal is to augment an LLM to support visual inputs. It is essential that the model can retain the text-only capability. Therefore, we further evaluate the text-only performance of the models under three benchmarks: MMLU [26], BBH [53], and DROP [21] in Table 6. We did not choose benchmarks like MT-Bench [67] since text instruction tuning is not the focus of the work. Overall, our model achieves performance as Llama-2 fine-tuned with", + "bbox": [ + 496, + 779, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26693", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b85835f552d1844a11f4f3267c15a14c321370cb64bc6aa056ae10ef883de4f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Comparison with state-of-the-art methods on 12 visual-language benchmarks. Our models consistently outperform LLaVA-1.5 under a head-to-head comparison, using the same prompts and the same base LLM (Vicuna-1.5 is based on Llama-2), showing the effectiveness of visual-language pre-training. We mark the best performance **bold** and the second-best **underlined**. Benchmark names are abbreviated due to space limits. VQA-v2 [24]; GQA [28]; VisWiz [25]; SQA†: ScienceQA-IMG [39]; VQAT: TextVQA [50]; POPE [34]; MME [23]; MMB: MMBench [38]; MMBCN: MMBench-Chinese [38]; SEED: SEED-Bench [31]; LLaVAW: LLaVA-Bench (In-the-Wild) [37]; MM-Vet [63]. *The training images of the datasets are observed during training. We also tried adding the ShareGPT4V [12] to the SFT blend on top of VILA-13B (last row), leading to a significant improvement on LLaVA-Bench and MM-Vet (marked in green)." + ], + "table_body": "
MethodLLMRes.PTIT\\(VQA^{v2}\\)GQAVisWiz\\(SQA^I\\)\\(VQAT^T\\)POPEMMEMMB\\(MMBC^{CN}\\)SEED\\(LLaVA^W\\)MM-Vet
BLIP-2 [33]Vicuna-13B224129M-41.04119.66142.585.31293.8--46.438.122.4
InstructBLIP [17]Vicuna-7B224129M1.2M-49.234.560.550.1--3623.753.460.926.2
InstructBLIP [17]Vicuna-13B224129M1.2M-49.533.463.150.778.91212.8---58.225.6
Shikra [11]Vicuna-13B224600K5.5M77.4*------58.8----
IDEFICS-9B [29]LLaMA-7B224353M1M50.938.435.5-25.9--48.225.2---
IDEFICS-80B [29]LLaMA-65B224353M1M60.045.236.0-30.9--54.538.1---
Qwen-VL [8]Qwen-7B4481.4B50M78.8*59.3*35.267.163.8--38.27.456.3--
Qwen-VL-Chat [8]Qwen-7B4481.4B50M78.2*57.5*38.968.261.5-1487.560.656.758.2--
LLaVA-1.5 [36]Vicuna-1.5-7B3360.6M0.7M78.5*62.0*50.066.858.285.91510.764.358.358.663.430.5
LLaVA-1.5 [36]Vicuna-1.5-13B3360.6M0.7M80.0*63.3*53.671.661.385.91531.367.763.661.670.735.4
VILA-7B (ours)Llama-2-7B33650M1M79.9*62.3*57.868.264.485.51533.068.961.761.169.734.9
VILA-13B (ours)Llama-2-13B33650M1M80.8*63.3*60.673.766.684.21570.170.364.362.873.038.8
+ShareGPT4VLlama-2-13B33650M1M80.6*63.2*62.473.165.384.81556.570.865.461.478.445.7
", + "bbox": [ + 81, + 88, + 903, + 265 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cbb7935cb5b9b73613802e3adf737d63edaba054e2f9cf5ab8921c52d2a60fdc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. VILA maintains competitive accuracy on text-only benchmarks. There is a small gap compared to the text-only model under 7B; but the accuracy is even better under 13B." + ], + "table_body": "
SizeModelMMLU [26]BBH [53]DROP [21]
7BLlama-246.0%32.0%31.7%
Llama-2+SFT51.8%39.3%53.1%
Vicuna-1.549.8%36.9%29.2%
VILA50.8%38.5%52.7%
13BLlama-255.7%37.5%41.6%
Llama-2+SFT54.3%43.2%59.2%
Vicuna-1.555.8%38.4%43.6%
VILA56.0%44.2%63.6%
", + "bbox": [ + 84, + 395, + 460, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the same text SFT data: the accuracy of our 7B model is a bit lower, while the 13B is higher. We suspect the smaller model may suffer from a larger text performance degradation during the pre-training, as observed in [19].", + "bbox": [ + 75, + 609, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 681, + 287, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we study how visual language pre-training enables new capabilities for the model. Part of the image samples are taken from [5, 13, 60].", + "bbox": [ + 75, + 705, + 468, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-image reasoning. Thanks to the pre-training, our model has the ability to reason over multiple images, despite the SFT data is composed of single-image samples. We provide two examples in Figure 6. In the first example, our model is able to figure out the common object (i.e., a flamingo) across the three images and the different art styles of each one, while the LLaVA model failed. The LLaVA model hallucinates and cannot distinguish the information from different input images. In the second example, our", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model is able to find one of the differences (i.e., the head-wear) out of the two.", + "bbox": [ + 498, + 398, + 893, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-context learning. In-context learning is an important characteristic of LLMs, allowing people to prompt the LLM with few-shot samples to enable new tasks. We provide in-context learning samples in Figure 7. Interestingly LLaVA-1.5 [36] can also perform in-context learning to some extent, despite only being trained on single-image-text-paired samples. We believe the capability is inherited from text-only pre-training of the base LLM. Nonetheless, our model outperforms LLaVA-1.5 for in-context learning: LLaVA-1.5 failed the first sample due to limited OCR capability, and failed the third example by repeating the semantics.", + "bbox": [ + 496, + 450, + 893, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Chain-of-Thoughts (CoT). Our model is able to perform chain-of-thought reasoning given visual inputs. As shown in Figure 8, VILA is able to perform complex CoT reasoning over the input images (multi-image or single-image) when adding \"Think step-by-step\" to the end of the prompt. We believe the CoT capability is inherited from text-only SFT, despite there is no such samples from the visual language instruction data.", + "bbox": [ + 496, + 637, + 895, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Better world knowledge. Since our model is pre-trained on a large-scale corpus, it has better understanding of world knowledge. We perform a case study by prompting the model to recognize the locations of some famous landmarks (please see supplementary due to space limits). VILA can correctly recognize 4 out of the 4 samples, while LLaVA-1.5 only gets 2 out of the 4, demonstrating the effectiveness of the pre-training. Samples are taken from [60].", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26694", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: Image 1: \\nImage 2: \\nImage 3: \\n", + "bbox": [ + 104, + 95, + 457, + 108 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What is the common thing about these three images?", + "bbox": [ + 107, + 108, + 359, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA-1.5: The common theme among these three images is the presence of a pink flamingo. In each image, the flamingo is standing in the water, with one image showing the bird in the water, another image showing the bird's legs, and the third image showing the bird's head. The flamingo is the central subject in all three images, and its distinctive pink color and unique appearance make it the focal point of each photo.", + "bbox": [ + 106, + 119, + 643, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a6c4e05e035aa5095cc9877430e7a1eb3eddaaccd73583bd10dda611184a804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 99, + 715, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81ef5fec8e35979932ce96f0dc846883b309d2a590a7fda7640cb7ae95a660d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 99, + 789, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee40e8e5a4945c2c6b1340757f4cb815f01c14880aa7c8842f9455e89a31122c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 99, + 857, + 150 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ours: The common theme among these three images is the presence of a flamingo. In the first image, the flamingo is a 3D model, while in the second image, it is a pink cartoon. In the third image, it is a real flamingo standing in the water. All three images showcase the bird in different styles and formats, but they all share the same subject.", + "bbox": [ + 104, + 164, + 779, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: Image 1: \\nImage 2: \\n", + "bbox": [ + 106, + 200, + 356, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tell me about the differences between image 1 and image 2.", + "bbox": [ + 107, + 213, + 390, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA-1.5: In image 1, there is a small girl wearing a pink dress and a crown. She is smiling and appears to be happy. In image 2, the girl is wearing a pink dress and a crown, but she is not smiling. The main difference between the two images is the girl's facial expression, with one showing a smile and the other not showing any expression.", + "bbox": [ + 106, + 224, + 681, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1701258bc25ee6cbaf68205e7a3ab9d1defa761da6acd5b28796d3d66a6fb1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 202, + 767, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/145e93d879fafb3d77c6d058099893008b3b8c07005fc051d97c4a8990cf85d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 202, + 841, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ours. The two images feature a cartoon girl wearing a pink dress. In the first image, the girl is wearing a crown, while in the second image, she is wearing a tiara. Both images showcase the girl in a princess-like attire, but the difference lies in the type of headwear she is wearing. The crown is a more traditional and regal accessory, while the tiara is a more playful and common accessory associated with princesses in popular culture.", + "bbox": [ + 104, + 258, + 854, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/270b3b736a1b61b8b88310e6a581b7a6856ffaded777f185305a6591a36bf3ff.jpg", + "image_caption": [ + "Input few shot examples + target image", + "Underground." + ], + "image_footnote": [], + "bbox": [ + 109, + 349, + 173, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df5d0c291a9b2ce4f73ddda4035435bfe81f6f7e03a17ca26c999971f20016d8.jpg", + "image_caption": [ + "Congress." + ], + "image_footnote": [], + "bbox": [ + 184, + 349, + 246, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d262444857fb4095041897039a483f3d5b79b79880641c58e1ac600c0579c40.jpg", + "image_caption": [ + "?", + "Output", + "LLaVA-1.5: Soulemes", + "Ours: Soulomes." + ], + "image_footnote": [], + "bbox": [ + 256, + 349, + 320, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87dca514e52b07032f828d8a841a7aa95cd7a243e12d3507d127edeb0a8271b7.jpg", + "image_caption": [ + "$2 + 1 = 3$" + ], + "image_footnote": [], + "bbox": [ + 109, + 415, + 173, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ec856ed2fb5f9dc331ea38f43821e022bf7f3dab913a5612ec009ed6fd3d323.jpg", + "image_caption": [ + "$5 + 6 = 11$" + ], + "image_footnote": [], + "bbox": [ + 184, + 415, + 246, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdd0dac968f9f032ec085ba05d87f02cd800c5777f64ebcfd29bbdc78d490066.jpg", + "image_caption": [ + "?", + "LLaVA-1.5: $3 \\times 6 = 18$", + "Ours: $3 \\times 6 = 18$" + ], + "image_footnote": [], + "bbox": [ + 256, + 415, + 320, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/902749e88ea5c794423892338c8098798ccc69ba6d0fda5e1dff66866c4490ca.jpg", + "image_caption": [ + "Romanticism", + "Figure 7. VILA has better in-context learning capability thanks to interleaved image text pretraining rather than single image-text pairs. We feed two image+text pairs and a third image as the context to prompt the VLM. LLaVA failed the first sample due to limited OCR capability, and failed the third examples by repeating the second sample semantic." + ], + "image_footnote": [], + "bbox": [ + 109, + 465, + 173, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/787e50c20c377486f4d9a95d6d81301ab01030e131518bd207d6a8f1aa8690e1.jpg", + "image_caption": [ + "Surrealism" + ], + "image_footnote": [], + "bbox": [ + 184, + 465, + 246, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f3f287ea3eb502a0a7a5d066a589f5b9420772267c589b26b15ded607e2126f.jpg", + "image_caption": [ + "?", + "LLaVA-1.5: Surrealism", + "Ours: Impressionism" + ], + "image_footnote": [], + "bbox": [ + 256, + 465, + 320, + 513 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6878f0c25cffb7131418dacdfb15013246fe72f5f14f4ff3ef6b8c59f9bb85d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ResolutionProjector#TokensEvaluation ↑
OKVQATextVQACOCO
224linear25649.9%41.6%116.0
336linear57649.7%49.8%117.7
336downsample14449.3%45.6%115.7
", + "bbox": [ + 78, + 619, + 478, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Improving the image resolution from 224 to 336 can significantly improve TextVQA accuracy. The raw resolution matters more than #tokens; high-resolution with token downsampling works better than low-resolution. We report accuracy for OKVQA and TextVQA, and CIDEr for COCO. Note: the evaluation protocol is different from Table 5 and can only be compared within the table.", + "bbox": [ + 75, + 720, + 470, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Other Learnings.", + "text_level": 1, + "bbox": [ + 76, + 824, + 245, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image resolution matters, not #tokens. We chose an image resolution of $336^{2}$ since it provides more fine-grained details compared to $224^{2}$ , leading to improved accuracy on tasks like TextVQA [50]. As shown in Table 7, increasing", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/762bac7201380454e14b7bd27cd958bc992333275a2b3c06429e5687034e7e51.jpg", + "image_caption": [ + "Figure 6. Our model VILA can reason over multiple images thanks to the pre-training process. The samples are taken from [5, 60].", + "Example: Visual Chain-of-Though (CoT)" + ], + "image_footnote": [], + "bbox": [ + 527, + 344, + 643, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d1812cc2c8be4f7f591e77bd6d6114593dd177c5afa0425b097f03cfb70cc82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 344, + 785, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: Photo: \\nMenu: \\nHow much should I pay for all the beer on the table according to the price on the menu?", + "bbox": [ + 521, + 414, + 857, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": " Answer: According to the menu, the price for the beer on the table is $6.", + "bbox": [ + 521, + 436, + 861, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: ... + Think step-by-step.", + "text_level": 1, + "bbox": [ + 521, + 455, + 683, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": " Answer: The beer on the table is Magna, which is priced at $6.00. There are two bottles of Magna on the table. Therefore, the total cost of the beer should be $6.00 x 2 = $12.00.", + "bbox": [ + 519, + 467, + 862, + 498 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00f75f8675192043637d340b0f48ae3523938c2fc46ceffa077137af0ed44139.jpg", + "image_caption": [ + "Figure 8. Our model is able to perform chain-of-thought reasoning given visual inputs. It is able to generate the correct answer when adding \"Think step-by-step\" to the prompt. Zoom in for a better view of the image details. Samples from [19, 60]." + ], + "image_footnote": [], + "bbox": [ + 527, + 503, + 627, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: \\nI am getting just two custom pizza for me and my friend. How much should I pay in total according to the image?", + "bbox": [ + 632, + 522, + 856, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Answer: According to the image, you should pay $11.99 for two custom pizzas.", + "bbox": [ + 632, + 555, + 846, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: ... + Let's think step-by-step.", + "text_level": 1, + "bbox": [ + 521, + 584, + 705, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": " Answer: The image shows a chalkboard menu with a special for today, which is to create your own pizza. The price for this special is $9.99. Since you are getting two custom pizzas, you should pay $9.99 x 2 = $19.98 for the two pizzas. So, the answer is $19.98.", + "bbox": [ + 519, + 597, + 854, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the resolution from 224 to 336 can improve the TextVQA accuracy from $41.6\\%$ to $49.8\\%$ . However, a higher resolution leads to more tokens per image ( $336^2$ corresponds to 576 tokens/image) and a higher computational cost. It also limits the number of demonstrations for in-context learning.", + "bbox": [ + 496, + 717, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Luckily, we find that the raw resolution matters more than the #visual tokens/image. We can use different projector designs to compress the visual tokens. Here we try a \"downsample\" projector, which simply concatenates every $2 \\times 2$ tokens into a single one and use a linear layer to fuse the information. It reduces the #tokens to 144 under the 336 resolution, that is even smaller than the 224+linear", + "bbox": [ + 496, + 795, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/95c5706fb6a76056bd245b004e1bef0fa09c6a23032ea8cae2c912e41d8e38fc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Directly fine-tuning the LLM during pre-training leads to better VLM accuracy and in-context learning capabilities. It also enjoys a smaller model size. Both settings are pre-trained on the MMC4-core dataset [69]." + ], + "table_body": "
#ParamVLM acc. (avg)
0-shot4-shot
Visual Expert [58]1.9×67.0%64.8%
Fine-tune1×71.0%72.1%
", + "bbox": [ + 112, + 88, + 431, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "setup. Nonetheless, the TextVQA accuracy is higher ( $\\sim 46\\%$ vs. $41.6\\%$ ), despite still $3\\%$ worse compared to $336+$ linear setup, showing a large redundancy in the image tokens. The gap on other datasets such as OKVQA and COCO is smaller since they usually require higher-level semantics.", + "bbox": [ + 75, + 229, + 468, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our main results, we did not use any token compression methods to provide the best accuracy despite this encouraging observation, and leave it to future work.", + "bbox": [ + 75, + 306, + 470, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to frozen LLMs with visual experts. Another interesting method for retaining the text capabilities of LLMs during the pre-training is to freeze the base LLM and add an extra visual expert to process the visual tokens [58]. The definition of expert is similar to MoE frameworks, but with a manual routing mechanism according to token types. Since the base LLM is frozen, the model fully retains the original functionality for text-only inputs during pre-training. However, we find that directly fine-tuning the LLM during visual language pre-training still leads to a better VLM accuracy and in-context learning capability (Table 8). Adding an extra visual expert also leads to near $2 \\times$ model size increase, which is not friendly for edge deployment. Therefore, we chose to directly fine-tune the base LLM.", + "bbox": [ + 75, + 368, + 472, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to PEFT/LoRA. In addition to visual experts, we also provide extra results when performing LoRA tuning with rank 64 (7B model) in Table 9. Fine-tuning LLM outperforms LoRA tuning by a large margin.", + "bbox": [ + 75, + 595, + 470, + 657 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9204013aef3140f470654b2f5d6f134e06cb5a5987141d79252a5103e948759.jpg", + "table_caption": [], + "table_footnote": [ + "Table 9. Fine-tuning LLM consistently outperforms LoRA tuning." + ], + "table_body": "
VQA v2GQAVisWizVQATMMBLLaVA W
LoRA (r=64)69.454.348.450.060.351.2
Fine-tune (ours)79.962.357.864.468.969.7
", + "bbox": [ + 78, + 670, + 468, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reformatting the interleaved structure. For additional insights we also reformatted the MMC4 dataset to be $\\langle \\text{im1} \\rangle < \\text{im2} \\rangle < \\text{txt1} \\rangle < \\text{txt2} \\rangle$ instead of the $\\langle \\text{im1} \\rangle < \\text{txt1} \\rangle < \\text{im2} \\rangle < \\text{txt2} \\rangle$ and evaluated the model under the setting in Table 1. We observed that the reformatted MMC4 degrades the average 0-shot accuracy (on the 4 benchmarks) by $4.4\\%$ , and degrades the average 4-shot accuracy by $37.5\\%$ . The disorder breaks in-context learning capability, showing the importance of interleaved data.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 498, + 89, + 640, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large language models (LLMs). LLMs based on Transformers [57] have fundamentally changed the language processing field. They are achieving increasing capabilities by scaling up the model size and the pre-training corpus [1, 9, 15, 18, 20, 22, 27, 47, 51]. It is believed that most the capability of the LLM is obtained from the large-scale pre-training process, which are later unlocked through instruction tuning [16, 43, 44]. There is a growing effort from the open-source community to build a strong base LLM [55, 56, 65], the conversational variants [14, 54] and the parameter-efficient finetuned versions of large LLMs [40, 64]. In this work, we start with the base Llama-2 model [56].", + "bbox": [ + 496, + 106, + 893, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual language models (VLMs). VLMs are LLMs augmented with visual inputs to provide a unified interface for visual language tasks. There are two main designs for VLMs: 1. cross-attention based, where the LLM is frozen while the visual information is fused into intermediate embeddings with a cross-attention mechanism [5, 6]; 2. auto-regressive based, where the visual input is tokenized and fed to the LLM alongside text tokens [2, 4, 7, 13, 19, 33, 37, 61, 68]. The latter is a natural extension by treating visual inputs as a foreign language. VLMs are also instruction-tuned so that they can better follow human instructions or perform conversations [17, 37, 52]. In this work, we study the pre-training process of the auto-regressive VLMs due to their flexibility when handling multi-modal inputs.", + "bbox": [ + 496, + 311, + 893, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following text-only LLMs, people also study different training recipes for VLMs. Some work freezes the LLM and train auxiliary components [5, 32, 33, 58], others fine-tune the LLM to enable visual capabilities [13, 19, 66]. There is also usage of different data corpora, including image-text pairs [13, 19, 32, 37], interleaved datasets [6], video-text pairs [41], visual-grounded annotations [36, 45], etc. In this work, we provide a holistic ablation of different design choices for the pre-training stage.", + "bbox": [ + 496, + 522, + 893, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 670, + 619, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has explored effective pretraining design options to augment LLMs towards vision tasks. Leveraging full strength of LLM learning, interleaved-nature of image-text data, and careful text data re-blending, VILA has surpassed state-of-the-art methods for vision tasks while preserving text-only capabilities. VILA has also depicted strong reasoning capability for multi-image analysis, in-context learning and zero/few-shot tasks. We hope our paper can help spur further research on VLM pretraining.", + "bbox": [ + 496, + 695, + 893, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Bryan Catanzaro for fruitful discussions. We appreciate the help from Zhuolin Yang, Guilin Liu, Lukas Voegtle, Philipp Fischer, Karan Sapra and Timo Roman on dataset preparation and feedback.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] GPT-4 technical report. Technical report, OpenAI, 2023. https://arxiv.org/abs/2303.08774.", + "[2] Fuyu-8B: A multimodal architecture for AI agents. https://www.adept.ai/blog/fuyu-8b, 2023.", + "[3] Yi-34B large language model. https://huggingface.co/01-ai/Yi-34B, 2023.", + "[4] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023.", + "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022.", + "[6] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023.", + "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. Technical report, Alibaba Group, 2023. https://arxiv.org/abs/2303.08774.", + "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020.", + "[10] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/ coyo-dataset, 2022.", + "[11] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023.", + "[12] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023.", + "[13] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023." + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality, 2023.", + "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", + "[16] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. arXiv preprint arXiv:2210.11416, 2022.", + "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023.", + "[18] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019.", + "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.", + "[20] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022.", + "[21] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL, 2019.", + "[22] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022.", + "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.", + "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.", + "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 3608-3617, 2018.", + "[26] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. CoRR, abs/2009.03300, 2020.", + "[27] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022.", + "[28] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019.", + "[29] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.", + "[30] Siddharth Karamcheti, Laurel Orr, Jason Bolton, Tianyi Zhang, Karan Goel, Avanika Narayan, Rishi Bommasani, Deepak Narayanan, Tatsunori Hashimoto, Dan Jurafsky, et al. Mistral—a journey towards reproducible language model training, 2021.", + "[31] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023.", + "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022.", + "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023.", + "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023.", + "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014.", + "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023.", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. 2023.", + "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023.", + "[39] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022.", + "[40] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024.", + "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023.", + "[42] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019.", + "[43] OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt, 2023. Accessed: 2023.", + "[44] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022.", + "[45] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023.", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[47] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021.", + "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020.", + "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022.", + "[50] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019.", + "[51] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu," + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26698", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, et al. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model. arXiv preprint arXiv:2201.11990, 2022.", + "[52] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023.", + "[53] Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022.", + "[54] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.", + "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[56] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.", + "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023.", + "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021.", + "[60] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 2023.", + "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023.", + "[62] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014.", + "[63] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.", + "[64] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023.", + "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022.", + "[66] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023.", + "[67] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023.", + "[68] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.", + "[69] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023." + ], + "bbox": [ + 501, + 92, + 893, + 529 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26699", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_model.json b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d33e75fc83ea60752eed2eddbe0c4c54c30a24ae --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_model.json @@ -0,0 +1,2939 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.75, + 0.154 + ], + "angle": 0, + "content": "VILA: On Pre-training for Visual Language Models" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.179, + 0.761, + 0.233 + ], + "angle": 0, + "content": "Ji Lin\\(^{1,2*†}\\) Hongxu Yin\\(^{1*}\\) Wei Ping\\(^{1}\\) Pavlo Molchanov\\(^{1}\\) Mohammad Shoeybi\\(^{1}\\) Song Han\\(^{1,2}\\) \n\\(^{1}\\)NVIDIA \\({}^{2}\\)MIT" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.237, + 0.706, + 0.251 + ], + "angle": 0, + "content": "ViLA https://github.com/Efficient-Large-Model/VILA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pretraining data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.729, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Large language models (LLMs) have demonstrated superior capabilities for natural language tasks [3, 7, 9, 14, 15, 18, 30, 43, 48, 54-56]. Augmenting LLMs to support visual inputs allows the final model to inherit some of the appealing properties like instruction following, zero-shot generalization, and few-shot in-context learning (ICL), empowering various visual language tasks [1, 2, 5, 8, 13, 19, 33, 37, 68]. The central challenge of unifying vision and language for" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.285, + 0.877, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.535, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Figure 1. VILA's enhanced visual-language pre-training consistently improves the downstream task accuracy under a comparison to recent methods [7, 17, 37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.763 + ], + "angle": 0, + "content": "collaborative inference resides in connecting the LLM and the vision foundation model (e.g., a CLIP encoder): both foundation models are usually pre-trained individually, before aligned via vision-language joint training. Most of the efforts in this field have been focusing on improving the visual language instruction-tuning process, i.e., supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF) [36, 37, 52]. However, there lacks a thorough study of the pre-training process, where the model is trained on image-text datasets/corpora at scale [10, 49, 69]. This process is costly but critical for the modality alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.902 + ], + "angle": 0, + "content": "In this work, we aim to explore different design options for enhanced visual language model pre-training. In particular, we aim to answer \"How do various design choices in visual language model pre-training impact the downstream performance?\" We followed the pre-training + SFT pipeline and ablated different design options for pre-training overseeing dataset properties and training protocols. We discover several findings: (1) Freezing the LLM during pretraining can achieve a decent zero-shot performance, but" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.462, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Work done during an internship at NVIDIA." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.48, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.246, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Figure 2. We study auto-regressive visual language model, where images are tokenized and fed to the input of LLMs. We find updating the LLM is essential for in-context learning capabilities, and interleaved corpus like [69] helps pre-training. Joint SFT with text-only data helps maintain the text-only capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.448 + ], + "angle": 0, + "content": "not in-context learning (ICL) capability, whereas updating the LLMs encourages deep embedding alignment, which we found is important for ICL; (2) Interleaved visual language data is essential for pre-training, that provides accurate gradient update and maintains text-only capability; (3) Adding in text-only instruction data during SFT can further remedy text-only degradation and boost visual language task accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.471, + 0.631 + ], + "angle": 0, + "content": "We introduce practical guidance to design Visual Language models, dubbed VILA. Without bells and whistles, VILA outperforms the state-of-the-art model [36] by noticeable margins across a wide range of vision language tasks (Figure 1), thanks to the help of improved pre-training. Moreover, we observe that the pre-training process unlocked several interesting capabilities for the model, such as (i) multi-image reasoning (despite the model only sees single image-text pairs during SFT), (ii) stronger in-context learning capabilities, and (iii) enhanced world knowledge. We hope our findings can provide a good pre-training recipe for future visual language models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.206, + 0.659 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Model architecture. Multi-modal LLMs can be generally categorized into two settings: cross-attention-based [5, 33] and auto-regressive-based [2, 19, 37]. The latter VLM family tokenizes images into visual tokens, which are concatenated with textual tokens and fed as the input to LLMs (i.e., treating visual input as a foreign language). It is a natural extension of text-only LLMs by augmenting the input with visual embeddings and can handle arbitrary interleaved image-text inputs. In this study, we focus on the pre-training of auto-regressive VLMs due to its flexibility and popularity. As shown in Figure 2, auto-regressive VLMs consists of three components: a visual encoder, an LLM, and a projector that bridges the embeddings from the two modalities. The projector can be a simple linear layer [37] or more capable Transformer blocks [6, 17] – we will compare their efficacy in our experiments. The model takes visual and text input" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.678, + 0.107 + ], + "angle": 0, + "content": "and generates text outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.894, + 0.181 + ], + "angle": 0, + "content": "Training stages. Following common practice [6, 19, 37], we study how to augment a pre-trained text-only LLM with visual input support. The training can be categorized into three stages:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.894, + 0.258 + ], + "angle": 0, + "content": "0. **Projector initialization.** The LLM and ViT are separately pre-trained, while the projector is usually initialized from random weights. Therefore, we first pre-train the projector while freezing both ViT and LLMs on image-caption pairs following existing literature [17, 33, 37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.894, + 0.365 + ], + "angle": 0, + "content": "1. visual language pre-training. We then pre-train the model (LLM and the projector) on visual language corpus. We consider two types of corpus: interleaved image-text corpus (e.g., MMC4 [69]) and image-text pairs (e.g., COYO [10] and LAION [49]). We focus the study of this work on the pre-training process, which are most costly and important for visual language alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.894, + 0.457 + ], + "angle": 0, + "content": "2. Visual instruction-tuning. Finally, we further perform instruction tuning of the pre-trained model on visual language instruction datasets. We convert existing visual language datasets into FLAN [59] style (i.e., with dataset-specific prompts) following [17]. Please find the data blend of the visual instruction data in the supplementary." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.183, + 0.894, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.563 + ], + "angle": 0, + "content": "**Evaluations.** During our ablation study, we evaluate the fine-tuned model on 4 visual language tasks: accuracy for OKVQA [42] and TextVQA [50], and CIDEr score for COCO [35] and Flickr [62]. We evaluate both 0-shot and 4-shot performance, which reflects the models' in-context learning capability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.596 + ], + "angle": 0, + "content": "3. On Pre-training for Visual Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.894, + 0.636 + ], + "angle": 0, + "content": "In this section, we discuss practical design choices and learned lessons for the visual language pre-training process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.647, + 0.744, + 0.662 + ], + "angle": 0, + "content": "3.1. Updating LLM is Essential" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.894, + 0.798 + ], + "angle": 0, + "content": "Fine-tuning vs. prompt tuning. There are two popular ways to augment a pre-trained text-only LM with visual inputs: fine-tune LLMs on the visual input tokens [19, 37], or freeze the LLM and train only the visual input projector as prompt tuning [17, 33]. The latter is attractive since freezing the LLMs prevents the degradation of the pre-trained text-only LLM. Nonetheless, we found updating the base LLM is essential to inherit some of the appealing LLM properties like in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.894, + 0.875 + ], + "angle": 0, + "content": "To verify the idea, we compare the two training protocols in Table 1. We use a Transformer block for the projector instead of a single linear layer [37] in setting a-c, which provides enough capacity when freezing LLMs. We use MMC4-core \\([69]^*\\) for the comparison. We observed that:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.888, + 0.901 + ], + "angle": 0, + "content": "*We downloaded only 25M of 30M images amid some expired URLs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26690" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.878, + 0.201 + ], + "angle": 0, + "content": "
PreTSFTProjectileOKVQATextVQACOCOFlickrAverage
TrainLLM?0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot
(a)XXTransformer10.419.214.823.117.460.211.047.413.437.5
(b)X✓Transformer47.147.737.236.6109.488.073.658.166.857.6
(c)✓✓Transformer44.849.838.538.8112.3113.571.572.966.868.8
(d)✓✓Linear45.250.339.740.2115.7118.574.274.768.770.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.203, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Table 1. Ablation study on whether to train LLM or freeze LLM and only perform prompt tuning during visual language pre-training (PreT). Interestingly, freezing the LLM during pre-training does not hurt the 0-shot accuracy, but leads to worse in-context learning capability (worse 4-shot). Using a simple linear projector forces the LLM to learn more and leads to better generalization. We report accuracy for VQA datasets (OKVQA, TextVQA) and CIDEr score for captioning (COCO and Flickr). Note: we used a different evaluation setting just for ablation study; the absolute value in this setting is lower and should not be compared against other work." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.285, + 0.434, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Figure 3. Prompt-tuning to support visual tokens can only enable shallow alignment, while fine-tuning the LLM leads to alignment at deeper layers. From configuration (b) to (d) (as in Table 1), the alignment improves at deeper layer, so as ICL accuracy (4-shot)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.559 + ], + "angle": 0, + "content": "(1) Training only the projector during SFT leads to poor performance (setting a), despite using a high-capacity design. It is rewarding to fine-tune LLM during SFT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.664 + ], + "angle": 0, + "content": "(2) Interestingly, freezing the LLM during pre-training does not affect 0-shot performance, but degrades in-context learning capabilities (i.e., 4-shot, comparing setting b and c). The gap is even larger for captioning datasets (COCO & Flickr) since they are out-of-distribution (the instruction tuning data is mostly VQA-alike, see supplementary), showing the worse generalization capability when freezing LLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.74 + ], + "angle": 0, + "content": "(3) When using a small-capacity projector (a linear layer instead of a Transformer block), the accuracy is slightly better (comparing c and d). We hypothesize a simpler projector forces the LLM to learn more on handling visual inputs, leading to better generalization." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.87 + ], + "angle": 0, + "content": "The deep embedding alignment hypothesis. To understand why fine-tuning LLM is beneficial, we hypothesize that it is important to align the distribution of visual and textual latent embeddings (especially in the deeper layers), so that the model can seamlessly model the interaction between the two modalities. It is essential if we want to inherit some of the good properties of LLM like in-context learning for visual language applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To verify the idea, we calculate the Chamfer distance of visual and textual embeddings in different layers to measure" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.284, + 0.897, + 0.343 + ], + "angle": 0, + "content": "
DatasetTypeText Src.#img/sample#tok./img
MMC4 [69]InterleaveHTML4.0122.5
COYO [10]Img-text pairAlt-text122.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.352, + 0.894, + 0.422 + ], + "angle": 0, + "content": "Table 2. Two image-text corpus considered for pre-training. The COYO captions are generally very short, which has a different distribution compared to the text-only corpus for LLM training. We sample each data source to contain 25M images by choosing samples with high CLIP similarities." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.437, + 0.872, + 0.564 + ], + "angle": 0, + "content": "
Pre-train DataVLM acc (avg)MMLU acc.
0-shot4-shot
Llama-2--46.0%
COYO51.1%50.3%28.8% (-17.2%)
MMC4-pairs46.4%44.5%32.4% (-13.6%)
MMC468.7%70.9%40.7% (-5.3%)
MMC4+COYO69.0%71.3%40.2% (-5.8%)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.573, + 0.894, + 0.658 + ], + "angle": 0, + "content": "Table 3. Pre-training on MMC4 data provides better visual language accuracy (0-shot and few-shot) and smaller degradation on text-only accuracy compared to caption data (COYO). The benefits comes from the interleave nature but not the better text distribution (MMC4 vs. MMC4-pairs). Blending interleaved and caption data provides a better diversity and downstream accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.775 + ], + "angle": 0, + "content": "how well they align in Figure 3. We calculate the pairwise cosine similarity to exclude the affect of magnitude. From configuration (b) to (d), the similarity of deeper layer goes higher, so as the 4-shot accuracy in Table 1, showing the positive relationship between deep embedding alignment and in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Given the observations, we fine-tune the LLM during both pre-training and instruction-tuning in later studies, and use a simple linear projection layer." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.832, + 0.892, + 0.863 + ], + "angle": 0, + "content": "3.2. Interleaved Visual Language Corpus Helps Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our goal is to \"augment\" the LLM to support visual input, instead of training a model that only works well on visual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.169, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.092, + 0.258, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.258, + 0.089, + 0.89, + 0.155 + ], + "angle": 0, + "content": "...,\"Easy to grow, and fruitful, these are a favorite amongst those who like to have their own at home garden.\", , \"There is a large number of different kinds of tomatoes, meaning that there is one for every different size garden.\", \"You can choose from a variety of these plants to accomplish different food missions, be them small cucumbers for pickling or larger ones for slicing.\", ..., , \"These tender green beans are durable and quite at home growing in the hot season air.\", \"To get the most out of planting them, space the seeds 4 to 6 inches apart and allow 3 feet between rows, ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.163, + 0.89, + 0.191 + ], + "angle": 0, + "content": "Figure 4. A sample from MMC4 [69] dataset consisting of interleaved images and text segments. The images are placed before the corresponding text. The text are weakly conditioned on images: only colored text can be better inferred with the help of images." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.197, + 0.456, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.312, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 5. Loss is lower when pre-training on MMC4 compared to MMC4-pairs (samples broken into image-text pairs), since the text segments provide more information for language modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.414 + ], + "angle": 0, + "content": "language inputs. Therefore, it is essential to preserve the text-only capabilities of LLMs. We found that data blending is a key factor, both for pre-training and instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.649 + ], + "angle": 0, + "content": "Pre-training dataset options. Most of the VLM pretraining [33, 37, 58] relies on image-text pairs (i.e., image and captions) due to the wide availability and large diversity (e.g., LAION [49], COYO [10]). On the other hand, interleaved image-text datasets (MMC4 [69], M3W [5]) follow a more similar distribution compared to the text-only corpus and is found to be important in Flamingo-style model training [5]. We hypothesize that the interleaved dataset is even more important for VLMs when LLM backbone is updated to accommodate the visual input. For a better understanding of the two data types, we compare statistics in Table 2: COYO suffers from a short text distribution since the accompanying text is taken from alt-text. We subsample the COYO dataset by ranking CLIP similarities and keep only 25M images (a similar size as MMC4-core)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.725 + ], + "angle": 0, + "content": "We follow the same pre-training + SFT process and ablate different pre-training corpus. We compare the 0-shot and few-shot visual language accuracy as well as text-only accuracy (MMLU [26]) in Table 3. Due to space limit, we report the average accuracy over four datasets (as in Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Interleaved data is essential. We notice using image-text pairs (i.e., COYO) for pre-training can lead to catastrophic forgetting. The text-only accuracy (MMLU) degrades by \\(17.2\\%\\). The visual language accuracy is also much worse compared to MMC4 pre-training. Noticeably, the 4-shot accuracy is even worse than 0-shot, showing the model cannot properly do in-context learning for visual language inputs (probably because it never sees more than one image during pre-training). We hypothesize the catastrophic forgetting is due to the distribution of text-based captions, which are generally very short and concise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.895, + 0.32 + ], + "angle": 0, + "content": "On the contrary, dataset like MMC4 has a much closer distribution compared to text-only corpus (e.g., C4 [48]). When using the interleaved data for pre-training, the degradation on MMLU is only \\(\\sim 5\\%\\). The degradation would be even smaller when using a larger base LLM [19]. With proper instruction tuning (Section 3.3), this degradation can be fully recovered. It also promotes visual in-context learning, leading to a higher 4-shot accuracy compared to 0-shot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.895, + 0.45 + ], + "angle": 0, + "content": "Interleave data structure matters, but not the text distribution. We further question whether the benefits come from the better text distribution (e.g., longer) or from the interleave nature. To ablate this, we construct a new MMC4 variant by only keeping the images and their corresponding text segments, without considering the interleave nature, denoted as \"MMC4-pairs\". For example an MMC4 sample may look like:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.457, + 0.848, + 0.469 + ], + "angle": 0, + "content": "\\[\n< \\text {t x t 1} > < \\text {i m 1} > < \\text {t x t 2} > < \\text {t x t 3} > < \\text {i m 2} > < \\text {t x t 4} >\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.481, + 0.851, + 0.497 + ], + "angle": 0, + "content": "It will be converted into two MMC4-pairs samples:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.504, + 0.804, + 0.516 + ], + "angle": 0, + "content": "\\[\n< \\mathbf {i m} 1 > < \\mathbf {t x t} 2 >, < \\mathbf {i m} 2 > < \\mathbf {t x t} 4 >\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.756 + ], + "angle": 0, + "content": "However, training on MMC4-pairs does not lead to a satisfactory result: it slightly reduces the degradation on MMLU due to a longer text distribution, but the VLM accuracy is even lower compared to pre-training on COYO; there is also no in-context improvement. We hypothesize the MMC4 samples do not have a very strict image-text correspondence; the image only provides marginal information for text modeling (i.e., most of the information is still from pure text modeling; an example is provided in Figure 4). It is also demonstrated by the loss curves in Figure 5, where training on the interleave corpus leads to a much lower loss, indicating the full text segments provides more information. Therefore, the interleaved data structure is critical, allowing the model to pick up the image-related information, without over-forcing it to learn unrelated text modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Data blending improves pre-training. Training on image-text pairs only led to a sharp degradation on text-only accuracy (more than \\(17\\%\\)). Luckily, blending the interleaved corpus and image-text pairs allows us to introduce more diversity in the corpus, while also preventing the severe degradation. Training on MMC4+COYO further boosts the accuracy on visual language benchmarks (the gain is larger when we perform joint SFT, as we will show later (Table 4)." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.887, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)We followed [69] to match the image and text segments by CLIP scores." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.219 + ], + "angle": 0, + "content": "
PT dataSFT dataVLM acc. (avg)MMLU acc.
0-shot4-shot
Llama-2---46.0%
MMC4Visual68.7%70.9%40.7% (-5.3%)
MMC4+COYOVisual69.0%71.3%40.2% (-5.8%)
Llama-2Text--51.2%
MMC4Vis.+Text71.0%72.1%51.4% (+0.2%)
MMC4+COYOVis.+Text72.3%73.6%50.9% (-0.3%)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.221, + 0.472, + 0.264 + ], + "angle": 0, + "content": "Table 4. Joint SFT (Vis. + Text) not only bridges the degradation of text-only capability (MMLU acc.), but also improves the performance on visual-language tasks (both zero-shot and few-shot)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.273, + 0.442, + 0.289 + ], + "angle": 0, + "content": "3.3. Recover LLM Degradation with Joint SFT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.474, + 0.466 + ], + "angle": 0, + "content": "Despite the interleave data helps maintain the text-only capability, there is still a \\(5\\%\\) accuracy drop. A potential approach is to maintain the text-only capability would be to add in text-only corpus (the one used in the LLM pre-training). However, such text corpus are usually proprietary even for open-source models; it is also unclear how to subsample the data to match the scale of vision-language corpus. Luckily, we found the text-only capabilities are temporarily hidden, but not forgotten. Adding in text-only data during SFT can help bridge the degradation, despite using a much smaller scale compared to the text pre-training corpora (trillion scale)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.472, + 0.619 + ], + "angle": 0, + "content": "Joint supervised fine-tuning. The common way for instruction tuning is to fine-tune the model on some visual language datasets (VQA/Caption style [17] or GPT-generated [37]). We found blending in text-only instruction data can simultaneously (i) recover the degradation in text-only accuracy, and (ii) improve the visual language accuracy. To this end, we also blended in 1M text-only instruction tuning data sampled from FLAN [16], which we termed as joint SFT. We provide the comparison in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.472, + 0.803 + ], + "angle": 0, + "content": "We can see that blending in the text-only SFT data not only bridges the degradation on text-only capability (the MMLU accuracy is on par compared to the original Llama-2 model fine-tuned on the same text-only instruction data), but also improves the visual language capability. We hypothesize that the text-only instruction data improves the model's instruction-following capability, which is also important for visual language tasks. Interestingly, the benefits of blending in COYO data is more significant with joint SFT. We believe that with joint SFT, the model no longer suffers from the text-only degradation when pre-trained with short captions, thus unlocking the full benefits from the better visual diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.82, + 0.212, + 0.836 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.34, + 0.864 + ], + "angle": 0, + "content": "4.1. Scaling up VLM pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We scale up the training of VLM in the following aspects to form our final model:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.169 + ], + "angle": 0, + "content": "Higher image resolution. Above ablation studies used the OpenAI CLIP-L [46] with \\(224 \\times 224\\) resolutions as the visual encoder. We now use \\(336 \\times 336\\) image resolutions to include more visual details for the model, which can help tasks requiring fine-grained details (e.g., TextVQA [50])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.895, + 0.223 + ], + "angle": 0, + "content": "Larger LLMs. By default, we used Llama-2 [56] 7B for ablation study. We also scaled to a larger LLM backbone (e.g., Llama-2 [56] 13B) to further improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.336 + ], + "angle": 0, + "content": "Pre-training data. We used both interleaved image-text data and image-text pairs for pre-training (we sample roughly 1:1 image proportions) to improve the data diversity. The total the pre-training corpus contains about 50M images. It is smaller than the billion-scale pre-training data [5, 13, 58], but already demonstrates impressive improvements on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.895, + 0.422 + ], + "angle": 0, + "content": "SFT data. We also include a better SFT data blend from LLaVA-1.5 [36], which is more diverse (e.g., contains reference-based annotations) and has high-quality prompt. The new SFT data blend can significantly improve the downstream evaluation metrics. We include details the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.507 + ], + "angle": 0, + "content": "Limitations. Due to the limited compute budget, we have not been able to further scale up the size of the pre-training corpus to billion-scale, which we leave as future work. Nonetheless, pre-training on 50M images already demonstrated significant performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.722, + 0.53 + ], + "angle": 0, + "content": "4.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.772 + ], + "angle": 0, + "content": "visual language tasks. We perform a comprehensive comparison with state-of-the-art models on 12 visual language benchmarks in Table 5. Compared to existing models (e.g., LLaVA-1.5 [36]), our model achieves consistent improvements over most datasets at different model sizes under a head-to-head setting (using the same prompts and base LLM; Vicuna-1.5 is based on Llama-2). Remarkably, we 7B model is able to outperform LLaVA-1.5 13B on VisWiz [25] and TextVQA [50] by a large margin thanks to the pre-training. Our 7B model even outperforms the 13B LLaVA model on these datasets. Our model also has multi-lingual capability despite the vision-language instruction data is in English, outperforming LLaVA-1.5 on MMBench-Chinese benchmark. Our results demonstrate the benefits of vision-language pre-training on downstream tasks, even when using a high-quality instruction tuning dataset [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Text-only performance. Our goal is to augment an LLM to support visual inputs. It is essential that the model can retain the text-only capability. Therefore, we further evaluate the text-only performance of the models under three benchmarks: MMLU [26], BBH [53], and DROP [21] in Table 6. We did not choose benchmarks like MT-Bench [67] since text instruction tuning is not the focus of the work. Overall, our model achieves performance as Llama-2 fine-tuned with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26693" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.905, + 0.266 + ], + "angle": 0, + "content": "
MethodLLMRes.PTIT\\(VQA^{v2}\\)GQAVisWiz\\(SQA^I\\)\\(VQAT^T\\)POPEMMEMMB\\(MMBC^{CN}\\)SEED\\(LLaVA^W\\)MM-Vet
BLIP-2 [33]Vicuna-13B224129M-41.04119.66142.585.31293.8--46.438.122.4
InstructBLIP [17]Vicuna-7B224129M1.2M-49.234.560.550.1--3623.753.460.926.2
InstructBLIP [17]Vicuna-13B224129M1.2M-49.533.463.150.778.91212.8---58.225.6
Shikra [11]Vicuna-13B224600K5.5M77.4*------58.8----
IDEFICS-9B [29]LLaMA-7B224353M1M50.938.435.5-25.9--48.225.2---
IDEFICS-80B [29]LLaMA-65B224353M1M60.045.236.0-30.9--54.538.1---
Qwen-VL [8]Qwen-7B4481.4B50M78.8*59.3*35.267.163.8--38.27.456.3--
Qwen-VL-Chat [8]Qwen-7B4481.4B50M78.2*57.5*38.968.261.5-1487.560.656.758.2--
LLaVA-1.5 [36]Vicuna-1.5-7B3360.6M0.7M78.5*62.0*50.066.858.285.91510.764.358.358.663.430.5
LLaVA-1.5 [36]Vicuna-1.5-13B3360.6M0.7M80.0*63.3*53.671.661.385.91531.367.763.661.670.735.4
VILA-7B (ours)Llama-2-7B33650M1M79.9*62.3*57.868.264.485.51533.068.961.761.169.734.9
VILA-13B (ours)Llama-2-13B33650M1M80.8*63.3*60.673.766.684.21570.170.364.362.873.038.8
+ShareGPT4VLlama-2-13B33650M1M80.6*63.2*62.473.165.384.81556.570.865.461.478.445.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.374 + ], + "angle": 0, + "content": "Table 5. Comparison with state-of-the-art methods on 12 visual-language benchmarks. Our models consistently outperform LLaVA-1.5 under a head-to-head comparison, using the same prompts and the same base LLM (Vicuna-1.5 is based on Llama-2), showing the effectiveness of visual-language pre-training. We mark the best performance **bold** and the second-best **underlined**. Benchmark names are abbreviated due to space limits. VQA-v2 [24]; GQA [28]; VisWiz [25]; SQA†: ScienceQA-IMG [39]; VQAT: TextVQA [50]; POPE [34]; MME [23]; MMB: MMBench [38]; MMBCN: MMBench-Chinese [38]; SEED: SEED-Bench [31]; LLaVAW: LLaVA-Bench (In-the-Wild) [37]; MM-Vet [63]. *The training images of the datasets are observed during training. We also tried adding the ShareGPT4V [12] to the SFT blend on top of VILA-13B (last row), leading to a significant improvement on LLaVA-Bench and MM-Vet (marked in green)." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.396, + 0.462, + 0.543 + ], + "angle": 0, + "content": "
SizeModelMMLU [26]BBH [53]DROP [21]
7BLlama-246.0%32.0%31.7%
Llama-2+SFT51.8%39.3%53.1%
Vicuna-1.549.8%36.9%29.2%
VILA50.8%38.5%52.7%
13BLlama-255.7%37.5%41.6%
Llama-2+SFT54.3%43.2%59.2%
Vicuna-1.555.8%38.4%43.6%
VILA56.0%44.2%63.6%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.553, + 0.472, + 0.596 + ], + "angle": 0, + "content": "Table 6. VILA maintains competitive accuracy on text-only benchmarks. There is a small gap compared to the text-only model under 7B; but the accuracy is even better under 13B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.673 + ], + "angle": 0, + "content": "the same text SFT data: the accuracy of our 7B model is a bit lower, while the 13B is higher. We suspect the smaller model may suffer from a larger text performance degradation during the pre-training, as observed in [19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.683, + 0.288, + 0.699 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Here we study how visual language pre-training enables new capabilities for the model. Part of the image samples are taken from [5, 13, 60]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multi-image reasoning. Thanks to the pre-training, our model has the ability to reason over multiple images, despite the SFT data is composed of single-image samples. We provide two examples in Figure 6. In the first example, our model is able to figure out the common object (i.e., a flamingo) across the three images and the different art styles of each one, while the LLaVA model failed. The LLaVA model hallucinates and cannot distinguish the information from different input images. In the second example, our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.895, + 0.429 + ], + "angle": 0, + "content": "model is able to find one of the differences (i.e., the head-wear) out of the two." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.618 + ], + "angle": 0, + "content": "In-context learning. In-context learning is an important characteristic of LLMs, allowing people to prompt the LLM with few-shot samples to enable new tasks. We provide in-context learning samples in Figure 7. Interestingly LLaVA-1.5 [36] can also perform in-context learning to some extent, despite only being trained on single-image-text-paired samples. We believe the capability is inherited from text-only pre-training of the base LLM. Nonetheless, our model outperforms LLaVA-1.5 for in-context learning: LLaVA-1.5 failed the first sample due to limited OCR capability, and failed the third example by repeating the semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.897, + 0.76 + ], + "angle": 0, + "content": "Visual Chain-of-Thoughts (CoT). Our model is able to perform chain-of-thought reasoning given visual inputs. As shown in Figure 8, VILA is able to perform complex CoT reasoning over the input images (multi-image or single-image) when adding \"Think step-by-step\" to the end of the prompt. We believe the CoT capability is inherited from text-only SFT, despite there is no such samples from the visual language instruction data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Better world knowledge. Since our model is pre-trained on a large-scale corpus, it has better understanding of world knowledge. We perform a case study by prompting the model to recognize the locations of some famous landmarks (please see supplementary due to space limits). VILA can correctly recognize 4 out of the 4 samples, while LLaVA-1.5 only gets 2 out of the 4, demonstrating the effectiveness of the pre-training. Samples are taken from [60]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.106, + 0.097, + 0.458, + 0.109 + ], + "angle": 0, + "content": "Prompt: Image 1: \\nImage 2: \\nImage 3: \\n" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.109, + 0.36, + 0.12 + ], + "angle": 0, + "content": "What is the common thing about these three images?" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.121, + 0.645, + 0.164 + ], + "angle": 0, + "content": "LLaVA-1.5: The common theme among these three images is the presence of a pink flamingo. In each image, the flamingo is standing in the water, with one image showing the bird in the water, another image showing the bird's legs, and the third image showing the bird's head. The flamingo is the central subject in all three images, and its distinctive pink color and unique appearance make it the focal point of each photo." + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.101, + 0.716, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.101, + 0.79, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.101, + 0.859, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.165, + 0.781, + 0.197 + ], + "angle": 0, + "content": "Ours: The common theme among these three images is the presence of a flamingo. In the first image, the flamingo is a 3D model, while in the second image, it is a pink cartoon. In the third image, it is a real flamingo standing in the water. All three images showcase the bird in different styles and formats, but they all share the same subject." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.202, + 0.357, + 0.214 + ], + "angle": 0, + "content": "Prompt: Image 1: \\nImage 2: \\n" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.214, + 0.392, + 0.225 + ], + "angle": 0, + "content": "Tell me about the differences between image 1 and image 2." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.226, + 0.682, + 0.258 + ], + "angle": 0, + "content": "LLaVA-1.5: In image 1, there is a small girl wearing a pink dress and a crown. She is smiling and appears to be happy. In image 2, the girl is wearing a pink dress and a crown, but she is not smiling. The main difference between the two images is the girl's facial expression, with one showing a smile and the other not showing any expression." + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.203, + 0.769, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.203, + 0.842, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.259, + 0.855, + 0.291 + ], + "angle": 0, + "content": "Ours. The two images feature a cartoon girl wearing a pink dress. In the first image, the girl is wearing a crown, while in the second image, she is wearing a tiara. Both images showcase the girl in a princess-like attire, but the difference lies in the type of headwear she is wearing. The crown is a more traditional and regal accessory, while the tiara is a more playful and common accessory associated with princesses in popular culture." + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.297, + 0.871, + 0.312 + ], + "angle": 0, + "content": "Figure 6. Our model VILA can reason over multiple images thanks to the pre-training process. The samples are taken from [5, 60]." + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.329, + 0.307, + 0.34 + ], + "angle": 0, + "content": "Input few shot examples + target image" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.35, + 0.174, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.393, + 0.174, + 0.404 + ], + "angle": 0, + "content": "Underground." + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.35, + 0.247, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.393, + 0.238, + 0.405 + ], + "angle": 0, + "content": "Congress." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.35, + 0.321, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.393, + 0.293, + 0.401 + ], + "angle": 0, + "content": "?" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.33, + 0.406, + 0.34 + ], + "angle": 0, + "content": "Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.364, + 0.439, + 0.374 + ], + "angle": 0, + "content": "LLaVA-1.5: Soulemes" + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.374, + 0.411, + 0.384 + ], + "angle": 0, + "content": "Ours: Soulomes." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.416, + 0.174, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.444, + 0.158, + 0.453 + ], + "angle": 0, + "content": "\\(2 + 1 = 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.416, + 0.247, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.444, + 0.231, + 0.453 + ], + "angle": 0, + "content": "\\(5 + 6 = 11\\)" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.416, + 0.321, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.444, + 0.293, + 0.451 + ], + "angle": 0, + "content": "?" + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.419, + 0.432, + 0.429 + ], + "angle": 0, + "content": "LLaVA-1.5: \\(3 \\times 6 = 18\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.429, + 0.403, + 0.439 + ], + "angle": 0, + "content": "Ours: \\(3 \\times 6 = 18\\)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.467, + 0.174, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.516, + 0.174, + 0.525 + ], + "angle": 0, + "content": "Romanticism" + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.467, + 0.247, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.516, + 0.24, + 0.525 + ], + "angle": 0, + "content": "Surrealism" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.467, + 0.321, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.517, + 0.293, + 0.524 + ], + "angle": 0, + "content": "?" + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.482, + 0.444, + 0.493 + ], + "angle": 0, + "content": "LLaVA-1.5: Surrealism" + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.494, + 0.433, + 0.503 + ], + "angle": 0, + "content": "Ours: Impressionism" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Figure 7. VILA has better in-context learning capability thanks to interleaved image text pretraining rather than single image-text pairs. We feed two image+text pairs and a third image as the context to prompt the VLM. LLaVA failed the first sample due to limited OCR capability, and failed the third examples by repeating the second sample semantic." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.621, + 0.479, + 0.711 + ], + "angle": 0, + "content": "
ResolutionProjector#TokensEvaluation ↑
OKVQATextVQACOCO
224linear25649.9%41.6%116.0
336linear57649.7%49.8%117.7
336downsample14449.3%45.6%115.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.471, + 0.805 + ], + "angle": 0, + "content": "Table 7. Improving the image resolution from 224 to 336 can significantly improve TextVQA accuracy. The raw resolution matters more than #tokens; high-resolution with token downsampling works better than low-resolution. We report accuracy for OKVQA and TextVQA, and CIDEr for COCO. Note: the evaluation protocol is different from Table 5 and can only be compared within the table." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.825, + 0.246, + 0.84 + ], + "angle": 0, + "content": "4.4. Other Learnings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Image resolution matters, not #tokens. We chose an image resolution of \\(336^{2}\\) since it provides more fine-grained details compared to \\(224^{2}\\), leading to improved accuracy on tasks like TextVQA [50]. As shown in Table 7, increasing" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.329, + 0.735, + 0.34 + ], + "angle": 0, + "content": "Example: Visual Chain-of-Though (CoT)" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.345, + 0.645, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.345, + 0.787, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.415, + 0.859, + 0.436 + ], + "angle": 0, + "content": "Prompt: Photo: \\nMenu: \\nHow much should I pay for all the beer on the table according to the price on the menu?" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.437, + 0.862, + 0.45 + ], + "angle": 0, + "content": " Answer: According to the menu, the price for the beer on the table is $6." + }, + { + "type": "title", + "bbox": [ + 0.522, + 0.456, + 0.684, + 0.468 + ], + "angle": 0, + "content": "Prompt: ... + Think step-by-step." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.468, + 0.863, + 0.499 + ], + "angle": 0, + "content": " Answer: The beer on the table is Magna, which is priced at $6.00. There are two bottles of Magna on the table. Therefore, the total cost of the beer should be $6.00 x 2 = $12.00." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.505, + 0.629, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.523, + 0.857, + 0.555 + ], + "angle": 0, + "content": "Prompt: \\nI am getting just two custom pizza for me and my friend. How much should I pay in total according to the image?" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.556, + 0.847, + 0.578 + ], + "angle": 0, + "content": "Answer: According to the image, you should pay $11.99 for two custom pizzas." + }, + { + "type": "title", + "bbox": [ + 0.522, + 0.585, + 0.706, + 0.597 + ], + "angle": 0, + "content": "Prompt: ... + Let's think step-by-step." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.598, + 0.856, + 0.64 + ], + "angle": 0, + "content": " Answer: The image shows a chalkboard menu with a special for today, which is to create your own pizza. The price for this special is $9.99. Since you are getting two custom pizzas, you should pay $9.99 x 2 = $19.98 for the two pizzas. So, the answer is $19.98." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Figure 8. Our model is able to perform chain-of-thought reasoning given visual inputs. It is able to generate the correct answer when adding \"Think step-by-step\" to the prompt. Zoom in for a better view of the image details. Samples from [19, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.794 + ], + "angle": 0, + "content": "the resolution from 224 to 336 can improve the TextVQA accuracy from \\(41.6\\%\\) to \\(49.8\\%\\). However, a higher resolution leads to more tokens per image (\\(336^2\\) corresponds to 576 tokens/image) and a higher computational cost. It also limits the number of demonstrations for in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Luckily, we find that the raw resolution matters more than the #visual tokens/image. We can use different projector designs to compress the visual tokens. Here we try a \"downsample\" projector, which simply concatenates every \\(2 \\times 2\\) tokens into a single one and use a linear layer to fuse the information. It reduces the #tokens to 144 under the 336 resolution, that is even smaller than the 224+linear" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.432, + 0.162 + ], + "angle": 0, + "content": "
#ParamVLM acc. (avg)
0-shot4-shot
Visual Expert [58]1.9×67.0%64.8%
Fine-tune1×71.0%72.1%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.166, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Table 8. Directly fine-tuning the LLM during pre-training leads to better VLM accuracy and in-context learning capabilities. It also enjoys a smaller model size. Both settings are pre-trained on the MMC4-core dataset [69]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.307 + ], + "angle": 0, + "content": "setup. Nonetheless, the TextVQA accuracy is higher (\\(\\sim 46\\%\\) vs. \\(41.6\\%\\)), despite still \\(3\\%\\) worse compared to \\(336+\\) linear setup, showing a large redundancy in the image tokens. The gap on other datasets such as OKVQA and COCO is smaller since they usually require higher-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.472, + 0.355 + ], + "angle": 0, + "content": "In our main results, we did not use any token compression methods to provide the best accuracy despite this encouraging observation, and leave it to future work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.473, + 0.581 + ], + "angle": 0, + "content": "Comparison to frozen LLMs with visual experts. Another interesting method for retaining the text capabilities of LLMs during the pre-training is to freeze the base LLM and add an extra visual expert to process the visual tokens [58]. The definition of expert is similar to MoE frameworks, but with a manual routing mechanism according to token types. Since the base LLM is frozen, the model fully retains the original functionality for text-only inputs during pre-training. However, we find that directly fine-tuning the LLM during visual language pre-training still leads to a better VLM accuracy and in-context learning capability (Table 8). Adding an extra visual expert also leads to near \\(2 \\times\\) model size increase, which is not friendly for edge deployment. Therefore, we chose to directly fine-tune the base LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.472, + 0.658 + ], + "angle": 0, + "content": "Comparison to PEFT/LoRA. In addition to visual experts, we also provide extra results when performing LoRA tuning with rank 64 (7B model) in Table 9. Fine-tuning LLM outperforms LoRA tuning by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.671, + 0.47, + 0.727 + ], + "angle": 0, + "content": "
VQA v2GQAVisWizVQATMMBLLaVA W
LoRA (r=64)69.454.348.450.060.351.2
Fine-tune (ours)79.962.357.864.468.969.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.737, + 0.465, + 0.751 + ], + "angle": 0, + "content": "Table 9. Fine-tuning LLM consistently outperforms LoRA tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Reformatting the interleaved structure. For additional insights we also reformatted the MMC4 dataset to be \\( \\langle \\text{im1} \\rangle < \\text{im2} \\rangle < \\text{txt1} \\rangle < \\text{txt2} \\rangle \\) instead of the \\( \\langle \\text{im1} \\rangle < \\text{txt1} \\rangle < \\text{im2} \\rangle < \\text{txt2} \\rangle \\) and evaluated the model under the setting in Table 1. We observed that the reformatted MMC4 degrades the average 0-shot accuracy (on the 4 benchmarks) by \\( 4.4\\% \\), and degrades the average 4-shot accuracy by \\( 37.5\\% \\). The disorder breaks in-context learning capability, showing the importance of interleaved data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Large language models (LLMs). LLMs based on Transformers [57] have fundamentally changed the language processing field. They are achieving increasing capabilities by scaling up the model size and the pre-training corpus [1, 9, 15, 18, 20, 22, 27, 47, 51]. It is believed that most the capability of the LLM is obtained from the large-scale pre-training process, which are later unlocked through instruction tuning [16, 43, 44]. There is a growing effort from the open-source community to build a strong base LLM [55, 56, 65], the conversational variants [14, 54] and the parameter-efficient finetuned versions of large LLMs [40, 64]. In this work, we start with the base Llama-2 model [56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.523 + ], + "angle": 0, + "content": "Visual language models (VLMs). VLMs are LLMs augmented with visual inputs to provide a unified interface for visual language tasks. There are two main designs for VLMs: 1. cross-attention based, where the LLM is frozen while the visual information is fused into intermediate embeddings with a cross-attention mechanism [5, 6]; 2. auto-regressive based, where the visual input is tokenized and fed to the LLM alongside text tokens [2, 4, 7, 13, 19, 33, 37, 61, 68]. The latter is a natural extension by treating visual inputs as a foreign language. VLMs are also instruction-tuned so that they can better follow human instructions or perform conversations [17, 37, 52]. In this work, we study the pre-training process of the auto-regressive VLMs due to their flexibility when handling multi-modal inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.66 + ], + "angle": 0, + "content": "Following text-only LLMs, people also study different training recipes for VLMs. Some work freezes the LLM and train auxiliary components [5, 32, 33, 58], others fine-tune the LLM to enable visual capabilities [13, 19, 66]. There is also usage of different data corpora, including image-text pairs [13, 19, 32, 37], interleaved datasets [6], video-text pairs [41], visual-grounded annotations [36, 45], etc. In this work, we provide a holistic ablation of different design choices for the pre-training stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.62, + 0.687 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.895, + 0.833 + ], + "angle": 0, + "content": "This paper has explored effective pretraining design options to augment LLMs towards vision tasks. Leveraging full strength of LLM learning, interleaved-nature of image-text data, and careful text data re-blending, VILA has surpassed state-of-the-art methods for vision tasks while preserving text-only capabilities. VILA has also depicted strong reasoning capability for multi-image analysis, in-context learning and zero/few-shot tasks. We hope our paper can help spur further research on VLM pretraining." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. We thank Bryan Catanzaro for fruitful discussions. We appreciate the help from Zhuolin Yang, Guilin Liu, Lukas Voegtle, Philipp Fischer, Karan Sapra and Timo Roman on dataset preparation and feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26696" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.143 + ], + "angle": 0, + "content": "[1] GPT-4 technical report. Technical report, OpenAI, 2023. https://arxiv.org/abs/2303.08774." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.472, + 0.172 + ], + "angle": 0, + "content": "[2] Fuyu-8B: A multimodal architecture for AI agents. https://www.adept.ai/blog/fuyu-8b, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.473, + 0.2 + ], + "angle": 0, + "content": "[3] Yi-34B large language model. https://huggingface.co/01-ai/Yi-34B, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[4] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.471, + 0.397 + ], + "angle": 0, + "content": "[6] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. Technical report, Alibaba Group, 2023. https://arxiv.org/abs/2303.08774." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.454, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[10] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/ coyo-dataset, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[11] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[12] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.163 + ], + "angle": 0, + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[16] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. arXiv preprint arXiv:2210.11416, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[18] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[20] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[21] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[22] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.894, + 0.772 + ], + "angle": 0, + "content": "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26697" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 3608-3617, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[26] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. CoRR, abs/2009.03300, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[27] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[28] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[29] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[30] Siddharth Karamcheti, Laurel Orr, Jason Bolton, Tianyi Zhang, Karan Goel, Avanika Narayan, Rishi Bommasani, Deepak Narayanan, Tatsunori Hashimoto, Dan Jurafsky, et al. Mistral—a journey towards reproducible language model training, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[31] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[39] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.133 + ], + "angle": 0, + "content": "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[40] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[42] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.894, + 0.356 + ], + "angle": 0, + "content": "[43] OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt, 2023. Accessed: 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.894, + 0.439 + ], + "angle": 0, + "content": "[44] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[45] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.893, + 0.579 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[47] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.893, + 0.718 + ], + "angle": 0, + "content": "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.893, + 0.802 + ], + "angle": 0, + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[50] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[51] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, et al. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model. arXiv preprint arXiv:2201.11990, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[52] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[53] Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[54] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[56] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[60] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[62] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[63] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[64] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[66] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[67] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[68] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[69] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26699" + } + ] +] \ No newline at end of file diff --git a/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_origin.pdf b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b3ff99cad68a151026a775a7ddbf5aafb515ee8b --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/f648f3d5-06a0-470c-8e61-c14e811bec54_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87a45c9addad1cc2b6eec422920bd49aa8c2f3387df76b36154f7ae767be0c35 +size 5036596 diff --git a/2024/VILA_ On Pre-training for Visual Language Models/full.md b/2024/VILA_ On Pre-training for Visual Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eacaf4c8cc8b5f3128ed08bf83b36533408e230e --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/full.md @@ -0,0 +1,389 @@ +# VILA: On Pre-training for Visual Language Models + +Ji Lin $^{1,2*†}$ Hongxu Yin $^{1*}$ Wei Ping $^{1}$ Pavlo Molchanov $^{1}$ Mohammad Shoeybi $^{1}$ Song Han $^{1,2}$ $^{1}$ NVIDIA ${}^{2}$ MIT + +ViLA https://github.com/Efficient-Large-Model/VILA + +# Abstract + +Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pretraining data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM. + +# 1. Introduction + +Large language models (LLMs) have demonstrated superior capabilities for natural language tasks [3, 7, 9, 14, 15, 18, 30, 43, 48, 54-56]. Augmenting LLMs to support visual inputs allows the final model to inherit some of the appealing properties like instruction following, zero-shot generalization, and few-shot in-context learning (ICL), empowering various visual language tasks [1, 2, 5, 8, 13, 19, 33, 37, 68]. The central challenge of unifying vision and language for + +![](images/27534e03254453598560de2a3d0d2ed6661a6a83ec42c3d343e5beda0bcfdcd4.jpg) +Figure 1. VILA's enhanced visual-language pre-training consistently improves the downstream task accuracy under a comparison to recent methods [7, 17, 37]. + +collaborative inference resides in connecting the LLM and the vision foundation model (e.g., a CLIP encoder): both foundation models are usually pre-trained individually, before aligned via vision-language joint training. Most of the efforts in this field have been focusing on improving the visual language instruction-tuning process, i.e., supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF) [36, 37, 52]. However, there lacks a thorough study of the pre-training process, where the model is trained on image-text datasets/corpora at scale [10, 49, 69]. This process is costly but critical for the modality alignment. + +In this work, we aim to explore different design options for enhanced visual language model pre-training. In particular, we aim to answer "How do various design choices in visual language model pre-training impact the downstream performance?" We followed the pre-training + SFT pipeline and ablated different design options for pre-training overseeing dataset properties and training protocols. We discover several findings: (1) Freezing the LLM during pretraining can achieve a decent zero-shot performance, but + +![](images/ebb73adf99d692f0cd74a3d6f8a92538fad46860aa9bb5e398b19b8a0b96f011.jpg) +Figure 2. We study auto-regressive visual language model, where images are tokenized and fed to the input of LLMs. We find updating the LLM is essential for in-context learning capabilities, and interleaved corpus like [69] helps pre-training. Joint SFT with text-only data helps maintain the text-only capabilities. + +not in-context learning (ICL) capability, whereas updating the LLMs encourages deep embedding alignment, which we found is important for ICL; (2) Interleaved visual language data is essential for pre-training, that provides accurate gradient update and maintains text-only capability; (3) Adding in text-only instruction data during SFT can further remedy text-only degradation and boost visual language task accuracy. + +We introduce practical guidance to design Visual Language models, dubbed VILA. Without bells and whistles, VILA outperforms the state-of-the-art model [36] by noticeable margins across a wide range of vision language tasks (Figure 1), thanks to the help of improved pre-training. Moreover, we observe that the pre-training process unlocked several interesting capabilities for the model, such as (i) multi-image reasoning (despite the model only sees single image-text pairs during SFT), (ii) stronger in-context learning capabilities, and (iii) enhanced world knowledge. We hope our findings can provide a good pre-training recipe for future visual language models. + +# 2. Background + +Model architecture. Multi-modal LLMs can be generally categorized into two settings: cross-attention-based [5, 33] and auto-regressive-based [2, 19, 37]. The latter VLM family tokenizes images into visual tokens, which are concatenated with textual tokens and fed as the input to LLMs (i.e., treating visual input as a foreign language). It is a natural extension of text-only LLMs by augmenting the input with visual embeddings and can handle arbitrary interleaved image-text inputs. In this study, we focus on the pre-training of auto-regressive VLMs due to its flexibility and popularity. As shown in Figure 2, auto-regressive VLMs consists of three components: a visual encoder, an LLM, and a projector that bridges the embeddings from the two modalities. The projector can be a simple linear layer [37] or more capable Transformer blocks [6, 17] – we will compare their efficacy in our experiments. The model takes visual and text input + +and generates text outputs. + +Training stages. Following common practice [6, 19, 37], we study how to augment a pre-trained text-only LLM with visual input support. The training can be categorized into three stages: + +0. **Projector initialization.** The LLM and ViT are separately pre-trained, while the projector is usually initialized from random weights. Therefore, we first pre-train the projector while freezing both ViT and LLMs on image-caption pairs following existing literature [17, 33, 37]. +1. visual language pre-training. We then pre-train the model (LLM and the projector) on visual language corpus. We consider two types of corpus: interleaved image-text corpus (e.g., MMC4 [69]) and image-text pairs (e.g., COYO [10] and LAION [49]). We focus the study of this work on the pre-training process, which are most costly and important for visual language alignment. +2. Visual instruction-tuning. Finally, we further perform instruction tuning of the pre-trained model on visual language instruction datasets. We convert existing visual language datasets into FLAN [59] style (i.e., with dataset-specific prompts) following [17]. Please find the data blend of the visual instruction data in the supplementary. + +**Evaluations.** During our ablation study, we evaluate the fine-tuned model on 4 visual language tasks: accuracy for OKVQA [42] and TextVQA [50], and CIDEr score for COCO [35] and Flickr [62]. We evaluate both 0-shot and 4-shot performance, which reflects the models' in-context learning capability. + +# 3. On Pre-training for Visual Language Models + +In this section, we discuss practical design choices and learned lessons for the visual language pre-training process. + +# 3.1. Updating LLM is Essential + +Fine-tuning vs. prompt tuning. There are two popular ways to augment a pre-trained text-only LM with visual inputs: fine-tune LLMs on the visual input tokens [19, 37], or freeze the LLM and train only the visual input projector as prompt tuning [17, 33]. The latter is attractive since freezing the LLMs prevents the degradation of the pre-trained text-only LLM. Nonetheless, we found updating the base LLM is essential to inherit some of the appealing LLM properties like in-context learning. + +To verify the idea, we compare the two training protocols in Table 1. We use a Transformer block for the projector instead of a single linear layer [37] in setting a-c, which provides enough capacity when freezing LLMs. We use MMC4-core $[69]^*$ for the comparison. We observed that: + +
PreTSFTProjectileOKVQATextVQACOCOFlickrAverage
TrainLLM?0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot
(a)XXTransformer10.419.214.823.117.460.211.047.413.437.5
(b)X✓Transformer47.147.737.236.6109.488.073.658.166.857.6
(c)✓✓Transformer44.849.838.538.8112.3113.571.572.966.868.8
(d)✓✓Linear45.250.339.740.2115.7118.574.274.768.770.9
+ +Table 1. Ablation study on whether to train LLM or freeze LLM and only perform prompt tuning during visual language pre-training (PreT). Interestingly, freezing the LLM during pre-training does not hurt the 0-shot accuracy, but leads to worse in-context learning capability (worse 4-shot). Using a simple linear projector forces the LLM to learn more and leads to better generalization. We report accuracy for VQA datasets (OKVQA, TextVQA) and CIDEr score for captioning (COCO and Flickr). Note: we used a different evaluation setting just for ablation study; the absolute value in this setting is lower and should not be compared against other work. + +![](images/fb9a16df2a643fb262bb55a9c4b08fc72f95f9297f083535f9a60604ed5f4d8f.jpg) +Figure 3. Prompt-tuning to support visual tokens can only enable shallow alignment, while fine-tuning the LLM leads to alignment at deeper layers. From configuration (b) to (d) (as in Table 1), the alignment improves at deeper layer, so as ICL accuracy (4-shot). + +(1) Training only the projector during SFT leads to poor performance (setting a), despite using a high-capacity design. It is rewarding to fine-tune LLM during SFT. +(2) Interestingly, freezing the LLM during pre-training does not affect 0-shot performance, but degrades in-context learning capabilities (i.e., 4-shot, comparing setting b and c). The gap is even larger for captioning datasets (COCO & Flickr) since they are out-of-distribution (the instruction tuning data is mostly VQA-alike, see supplementary), showing the worse generalization capability when freezing LLMs. +(3) When using a small-capacity projector (a linear layer instead of a Transformer block), the accuracy is slightly better (comparing c and d). We hypothesize a simpler projector forces the LLM to learn more on handling visual inputs, leading to better generalization. + +The deep embedding alignment hypothesis. To understand why fine-tuning LLM is beneficial, we hypothesize that it is important to align the distribution of visual and textual latent embeddings (especially in the deeper layers), so that the model can seamlessly model the interaction between the two modalities. It is essential if we want to inherit some of the good properties of LLM like in-context learning for visual language applications. + +To verify the idea, we calculate the Chamfer distance of visual and textual embeddings in different layers to measure + +
DatasetTypeText Src.#img/sample#tok./img
MMC4 [69]InterleaveHTML4.0122.5
COYO [10]Img-text pairAlt-text122.7
+ +Table 2. Two image-text corpus considered for pre-training. The COYO captions are generally very short, which has a different distribution compared to the text-only corpus for LLM training. We sample each data source to contain 25M images by choosing samples with high CLIP similarities. + +
Pre-train DataVLM acc (avg)MMLU acc.
0-shot4-shot
Llama-2--46.0%
COYO51.1%50.3%28.8% (-17.2%)
MMC4-pairs46.4%44.5%32.4% (-13.6%)
MMC468.7%70.9%40.7% (-5.3%)
MMC4+COYO69.0%71.3%40.2% (-5.8%)
+ +Table 3. Pre-training on MMC4 data provides better visual language accuracy (0-shot and few-shot) and smaller degradation on text-only accuracy compared to caption data (COYO). The benefits comes from the interleave nature but not the better text distribution (MMC4 vs. MMC4-pairs). Blending interleaved and caption data provides a better diversity and downstream accuracy. + +how well they align in Figure 3. We calculate the pairwise cosine similarity to exclude the affect of magnitude. From configuration (b) to (d), the similarity of deeper layer goes higher, so as the 4-shot accuracy in Table 1, showing the positive relationship between deep embedding alignment and in-context learning. + +Given the observations, we fine-tune the LLM during both pre-training and instruction-tuning in later studies, and use a simple linear projection layer. + +# 3.2. Interleaved Visual Language Corpus Helps Pre-training + +Our goal is to "augment" the LLM to support visual input, instead of training a model that only works well on visual + +![](images/d7d481958cc4c790af80703c2561cfdda658bb50d21d459440399a781ef17e21.jpg) +Figure 4. A sample from MMC4 [69] dataset consisting of interleaved images and text segments. The images are placed before the corresponding text. The text are weakly conditioned on images: only colored text can be better inferred with the help of images. + +![](images/2b97409618609b8ed58bd87e4b7804e973e6a0857ddb3bc4f48267edd734704a.jpg) +...,"Easy to grow, and fruitful, these are a favorite amongst those who like to have their own at home garden.", , "There is a large number of different kinds of tomatoes, meaning that there is one for every different size garden.", "You can choose from a variety of these plants to accomplish different food missions, be them small cucumbers for pickling or larger ones for slicing.", ..., , "These tender green beans are durable and quite at home growing in the hot season air.", "To get the most out of planting them, space the seeds 4 to 6 inches apart and allow 3 feet between rows, ... + +![](images/bd1a94b4a45456d974e21204bb803f015894c921fc39734db13a38a88618ee65.jpg) +Figure 5. Loss is lower when pre-training on MMC4 compared to MMC4-pairs (samples broken into image-text pairs), since the text segments provide more information for language modeling. + +language inputs. Therefore, it is essential to preserve the text-only capabilities of LLMs. We found that data blending is a key factor, both for pre-training and instruction tuning. + +Pre-training dataset options. Most of the VLM pretraining [33, 37, 58] relies on image-text pairs (i.e., image and captions) due to the wide availability and large diversity (e.g., LAION [49], COYO [10]). On the other hand, interleaved image-text datasets (MMC4 [69], M3W [5]) follow a more similar distribution compared to the text-only corpus and is found to be important in Flamingo-style model training [5]. We hypothesize that the interleaved dataset is even more important for VLMs when LLM backbone is updated to accommodate the visual input. For a better understanding of the two data types, we compare statistics in Table 2: COYO suffers from a short text distribution since the accompanying text is taken from alt-text. We subsample the COYO dataset by ranking CLIP similarities and keep only 25M images (a similar size as MMC4-core). + +We follow the same pre-training + SFT process and ablate different pre-training corpus. We compare the 0-shot and few-shot visual language accuracy as well as text-only accuracy (MMLU [26]) in Table 3. Due to space limit, we report the average accuracy over four datasets (as in Table 1). + +Interleaved data is essential. We notice using image-text pairs (i.e., COYO) for pre-training can lead to catastrophic forgetting. The text-only accuracy (MMLU) degrades by $17.2\%$ . The visual language accuracy is also much worse compared to MMC4 pre-training. Noticeably, the 4-shot accuracy is even worse than 0-shot, showing the model cannot properly do in-context learning for visual language inputs (probably because it never sees more than one image during pre-training). We hypothesize the catastrophic forgetting is due to the distribution of text-based captions, which are generally very short and concise. + +On the contrary, dataset like MMC4 has a much closer distribution compared to text-only corpus (e.g., C4 [48]). When using the interleaved data for pre-training, the degradation on MMLU is only $\sim 5\%$ . The degradation would be even smaller when using a larger base LLM [19]. With proper instruction tuning (Section 3.3), this degradation can be fully recovered. It also promotes visual in-context learning, leading to a higher 4-shot accuracy compared to 0-shot. + +Interleave data structure matters, but not the text distribution. We further question whether the benefits come from the better text distribution (e.g., longer) or from the interleave nature. To ablate this, we construct a new MMC4 variant by only keeping the images and their corresponding text segments, without considering the interleave nature, denoted as "MMC4-pairs". For example an MMC4 sample may look like: + +$$ +< \text {t x t 1} > < \text {i m 1} > < \text {t x t 2} > < \text {t x t 3} > < \text {i m 2} > < \text {t x t 4} > +$$ + +It will be converted into two MMC4-pairs samples: + +$$ +< \mathbf {i m} 1 > < \mathbf {t x t} 2 >, < \mathbf {i m} 2 > < \mathbf {t x t} 4 > +$$ + +However, training on MMC4-pairs does not lead to a satisfactory result: it slightly reduces the degradation on MMLU due to a longer text distribution, but the VLM accuracy is even lower compared to pre-training on COYO; there is also no in-context improvement. We hypothesize the MMC4 samples do not have a very strict image-text correspondence; the image only provides marginal information for text modeling (i.e., most of the information is still from pure text modeling; an example is provided in Figure 4). It is also demonstrated by the loss curves in Figure 5, where training on the interleave corpus leads to a much lower loss, indicating the full text segments provides more information. Therefore, the interleaved data structure is critical, allowing the model to pick up the image-related information, without over-forcing it to learn unrelated text modeling. + +Data blending improves pre-training. Training on image-text pairs only led to a sharp degradation on text-only accuracy (more than $17\%$ ). Luckily, blending the interleaved corpus and image-text pairs allows us to introduce more diversity in the corpus, while also preventing the severe degradation. Training on MMC4+COYO further boosts the accuracy on visual language benchmarks (the gain is larger when we perform joint SFT, as we will show later (Table 4). + +$^{\dagger}$ We followed [69] to match the image and text segments by CLIP scores. + +
PT dataSFT dataVLM acc. (avg)MMLU acc.
0-shot4-shot
Llama-2---46.0%
MMC4Visual68.7%70.9%40.7% (-5.3%)
MMC4+COYOVisual69.0%71.3%40.2% (-5.8%)
Llama-2Text--51.2%
MMC4Vis.+Text71.0%72.1%51.4% (+0.2%)
MMC4+COYOVis.+Text72.3%73.6%50.9% (-0.3%)
+ +Table 4. Joint SFT (Vis. + Text) not only bridges the degradation of text-only capability (MMLU acc.), but also improves the performance on visual-language tasks (both zero-shot and few-shot). + +# 3.3. Recover LLM Degradation with Joint SFT + +Despite the interleave data helps maintain the text-only capability, there is still a $5\%$ accuracy drop. A potential approach is to maintain the text-only capability would be to add in text-only corpus (the one used in the LLM pre-training). However, such text corpus are usually proprietary even for open-source models; it is also unclear how to subsample the data to match the scale of vision-language corpus. Luckily, we found the text-only capabilities are temporarily hidden, but not forgotten. Adding in text-only data during SFT can help bridge the degradation, despite using a much smaller scale compared to the text pre-training corpora (trillion scale). + +Joint supervised fine-tuning. The common way for instruction tuning is to fine-tune the model on some visual language datasets (VQA/Caption style [17] or GPT-generated [37]). We found blending in text-only instruction data can simultaneously (i) recover the degradation in text-only accuracy, and (ii) improve the visual language accuracy. To this end, we also blended in 1M text-only instruction tuning data sampled from FLAN [16], which we termed as joint SFT. We provide the comparison in Table 4. + +We can see that blending in the text-only SFT data not only bridges the degradation on text-only capability (the MMLU accuracy is on par compared to the original Llama-2 model fine-tuned on the same text-only instruction data), but also improves the visual language capability. We hypothesize that the text-only instruction data improves the model's instruction-following capability, which is also important for visual language tasks. Interestingly, the benefits of blending in COYO data is more significant with joint SFT. We believe that with joint SFT, the model no longer suffers from the text-only degradation when pre-trained with short captions, thus unlocking the full benefits from the better visual diversity. + +# 4. Experiments + +# 4.1. Scaling up VLM pre-training + +We scale up the training of VLM in the following aspects to form our final model: + +Higher image resolution. Above ablation studies used the OpenAI CLIP-L [46] with $224 \times 224$ resolutions as the visual encoder. We now use $336 \times 336$ image resolutions to include more visual details for the model, which can help tasks requiring fine-grained details (e.g., TextVQA [50]). + +Larger LLMs. By default, we used Llama-2 [56] 7B for ablation study. We also scaled to a larger LLM backbone (e.g., Llama-2 [56] 13B) to further improve the performance. + +Pre-training data. We used both interleaved image-text data and image-text pairs for pre-training (we sample roughly 1:1 image proportions) to improve the data diversity. The total the pre-training corpus contains about 50M images. It is smaller than the billion-scale pre-training data [5, 13, 58], but already demonstrates impressive improvements on downstream tasks. + +SFT data. We also include a better SFT data blend from LLaVA-1.5 [36], which is more diverse (e.g., contains reference-based annotations) and has high-quality prompt. The new SFT data blend can significantly improve the downstream evaluation metrics. We include details the Appendix. + +Limitations. Due to the limited compute budget, we have not been able to further scale up the size of the pre-training corpus to billion-scale, which we leave as future work. Nonetheless, pre-training on 50M images already demonstrated significant performance improvement. + +# 4.2.Quantitative Evaluation + +visual language tasks. We perform a comprehensive comparison with state-of-the-art models on 12 visual language benchmarks in Table 5. Compared to existing models (e.g., LLaVA-1.5 [36]), our model achieves consistent improvements over most datasets at different model sizes under a head-to-head setting (using the same prompts and base LLM; Vicuna-1.5 is based on Llama-2). Remarkably, we 7B model is able to outperform LLaVA-1.5 13B on VisWiz [25] and TextVQA [50] by a large margin thanks to the pre-training. Our 7B model even outperforms the 13B LLaVA model on these datasets. Our model also has multi-lingual capability despite the vision-language instruction data is in English, outperforming LLaVA-1.5 on MMBench-Chinese benchmark. Our results demonstrate the benefits of vision-language pre-training on downstream tasks, even when using a high-quality instruction tuning dataset [36]. + +Text-only performance. Our goal is to augment an LLM to support visual inputs. It is essential that the model can retain the text-only capability. Therefore, we further evaluate the text-only performance of the models under three benchmarks: MMLU [26], BBH [53], and DROP [21] in Table 6. We did not choose benchmarks like MT-Bench [67] since text instruction tuning is not the focus of the work. Overall, our model achieves performance as Llama-2 fine-tuned with + +
MethodLLMRes.PTIT\(VQA^{v2}\)GQAVisWiz\(SQA^I\)\(VQAT^T\)POPEMMEMMB\(MMBC^{CN}\)SEED\(LLaVA^W\)MM-Vet
BLIP-2 [33]Vicuna-13B224129M-41.04119.66142.585.31293.8--46.438.122.4
InstructBLIP [17]Vicuna-7B224129M1.2M-49.234.560.550.1--3623.753.460.926.2
InstructBLIP [17]Vicuna-13B224129M1.2M-49.533.463.150.778.91212.8---58.225.6
Shikra [11]Vicuna-13B224600K5.5M77.4*------58.8----
IDEFICS-9B [29]LLaMA-7B224353M1M50.938.435.5-25.9--48.225.2---
IDEFICS-80B [29]LLaMA-65B224353M1M60.045.236.0-30.9--54.538.1---
Qwen-VL [8]Qwen-7B4481.4B50M78.8*59.3*35.267.163.8--38.27.456.3--
Qwen-VL-Chat [8]Qwen-7B4481.4B50M78.2*57.5*38.968.261.5-1487.560.656.758.2--
LLaVA-1.5 [36]Vicuna-1.5-7B3360.6M0.7M78.5*62.0*50.066.858.285.91510.764.358.358.663.430.5
LLaVA-1.5 [36]Vicuna-1.5-13B3360.6M0.7M80.0*63.3*53.671.661.385.91531.367.763.661.670.735.4
VILA-7B (ours)Llama-2-7B33650M1M79.9*62.3*57.868.264.485.51533.068.961.761.169.734.9
VILA-13B (ours)Llama-2-13B33650M1M80.8*63.3*60.673.766.684.21570.170.364.362.873.038.8
+ShareGPT4VLlama-2-13B33650M1M80.6*63.2*62.473.165.384.81556.570.865.461.478.445.7
+ +Table 5. Comparison with state-of-the-art methods on 12 visual-language benchmarks. Our models consistently outperform LLaVA-1.5 under a head-to-head comparison, using the same prompts and the same base LLM (Vicuna-1.5 is based on Llama-2), showing the effectiveness of visual-language pre-training. We mark the best performance **bold** and the second-best **underlined**. Benchmark names are abbreviated due to space limits. VQA-v2 [24]; GQA [28]; VisWiz [25]; SQA†: ScienceQA-IMG [39]; VQAT: TextVQA [50]; POPE [34]; MME [23]; MMB: MMBench [38]; MMBCN: MMBench-Chinese [38]; SEED: SEED-Bench [31]; LLaVAW: LLaVA-Bench (In-the-Wild) [37]; MM-Vet [63]. *The training images of the datasets are observed during training. We also tried adding the ShareGPT4V [12] to the SFT blend on top of VILA-13B (last row), leading to a significant improvement on LLaVA-Bench and MM-Vet (marked in green). + +
SizeModelMMLU [26]BBH [53]DROP [21]
7BLlama-246.0%32.0%31.7%
Llama-2+SFT51.8%39.3%53.1%
Vicuna-1.549.8%36.9%29.2%
VILA50.8%38.5%52.7%
13BLlama-255.7%37.5%41.6%
Llama-2+SFT54.3%43.2%59.2%
Vicuna-1.555.8%38.4%43.6%
VILA56.0%44.2%63.6%
+ +Table 6. VILA maintains competitive accuracy on text-only benchmarks. There is a small gap compared to the text-only model under 7B; but the accuracy is even better under 13B. + +the same text SFT data: the accuracy of our 7B model is a bit lower, while the 13B is higher. We suspect the smaller model may suffer from a larger text performance degradation during the pre-training, as observed in [19]. + +# 4.3. Qualitative Evaluation + +Here we study how visual language pre-training enables new capabilities for the model. Part of the image samples are taken from [5, 13, 60]. + +Multi-image reasoning. Thanks to the pre-training, our model has the ability to reason over multiple images, despite the SFT data is composed of single-image samples. We provide two examples in Figure 6. In the first example, our model is able to figure out the common object (i.e., a flamingo) across the three images and the different art styles of each one, while the LLaVA model failed. The LLaVA model hallucinates and cannot distinguish the information from different input images. In the second example, our + +model is able to find one of the differences (i.e., the head-wear) out of the two. + +In-context learning. In-context learning is an important characteristic of LLMs, allowing people to prompt the LLM with few-shot samples to enable new tasks. We provide in-context learning samples in Figure 7. Interestingly LLaVA-1.5 [36] can also perform in-context learning to some extent, despite only being trained on single-image-text-paired samples. We believe the capability is inherited from text-only pre-training of the base LLM. Nonetheless, our model outperforms LLaVA-1.5 for in-context learning: LLaVA-1.5 failed the first sample due to limited OCR capability, and failed the third example by repeating the semantics. + +Visual Chain-of-Thoughts (CoT). Our model is able to perform chain-of-thought reasoning given visual inputs. As shown in Figure 8, VILA is able to perform complex CoT reasoning over the input images (multi-image or single-image) when adding "Think step-by-step" to the end of the prompt. We believe the CoT capability is inherited from text-only SFT, despite there is no such samples from the visual language instruction data. + +Better world knowledge. Since our model is pre-trained on a large-scale corpus, it has better understanding of world knowledge. We perform a case study by prompting the model to recognize the locations of some famous landmarks (please see supplementary due to space limits). VILA can correctly recognize 4 out of the 4 samples, while LLaVA-1.5 only gets 2 out of the 4, demonstrating the effectiveness of the pre-training. Samples are taken from [60]. + +Prompt: Image 1: \nImage 2: \nImage 3: \n + +What is the common thing about these three images? + +LLaVA-1.5: The common theme among these three images is the presence of a pink flamingo. In each image, the flamingo is standing in the water, with one image showing the bird in the water, another image showing the bird's legs, and the third image showing the bird's head. The flamingo is the central subject in all three images, and its distinctive pink color and unique appearance make it the focal point of each photo. + +![](images/1a6c4e05e035aa5095cc9877430e7a1eb3eddaaccd73583bd10dda611184a804.jpg) + +![](images/81ef5fec8e35979932ce96f0dc846883b309d2a590a7fda7640cb7ae95a660d7.jpg) + +![](images/ee40e8e5a4945c2c6b1340757f4cb815f01c14880aa7c8842f9455e89a31122c.jpg) + +Ours: The common theme among these three images is the presence of a flamingo. In the first image, the flamingo is a 3D model, while in the second image, it is a pink cartoon. In the third image, it is a real flamingo standing in the water. All three images showcase the bird in different styles and formats, but they all share the same subject. + +Prompt: Image 1: \nImage 2: \n + +Tell me about the differences between image 1 and image 2. + +LLaVA-1.5: In image 1, there is a small girl wearing a pink dress and a crown. She is smiling and appears to be happy. In image 2, the girl is wearing a pink dress and a crown, but she is not smiling. The main difference between the two images is the girl's facial expression, with one showing a smile and the other not showing any expression. + +![](images/d1701258bc25ee6cbaf68205e7a3ab9d1defa761da6acd5b28796d3d66a6fb1f.jpg) + +![](images/145e93d879fafb3d77c6d058099893008b3b8c07005fc051d97c4a8990cf85d4.jpg) + +Ours. The two images feature a cartoon girl wearing a pink dress. In the first image, the girl is wearing a crown, while in the second image, she is wearing a tiara. Both images showcase the girl in a princess-like attire, but the difference lies in the type of headwear she is wearing. The crown is a more traditional and regal accessory, while the tiara is a more playful and common accessory associated with princesses in popular culture. + +![](images/270b3b736a1b61b8b88310e6a581b7a6856ffaded777f185305a6591a36bf3ff.jpg) +Input few shot examples + target image +Underground. + +![](images/df5d0c291a9b2ce4f73ddda4035435bfe81f6f7e03a17ca26c999971f20016d8.jpg) +Congress. + +![](images/7d262444857fb4095041897039a483f3d5b79b79880641c58e1ac600c0579c40.jpg) +? +Output +LLaVA-1.5: Soulemes +Ours: Soulomes. + +![](images/87dca514e52b07032f828d8a841a7aa95cd7a243e12d3507d127edeb0a8271b7.jpg) +$2 + 1 = 3$ + +![](images/6ec856ed2fb5f9dc331ea38f43821e022bf7f3dab913a5612ec009ed6fd3d323.jpg) +$5 + 6 = 11$ + +![](images/bdd0dac968f9f032ec085ba05d87f02cd800c5777f64ebcfd29bbdc78d490066.jpg) +? +LLaVA-1.5: $3 \times 6 = 18$ +Ours: $3 \times 6 = 18$ + +![](images/902749e88ea5c794423892338c8098798ccc69ba6d0fda5e1dff66866c4490ca.jpg) +Romanticism +Figure 7. VILA has better in-context learning capability thanks to interleaved image text pretraining rather than single image-text pairs. We feed two image+text pairs and a third image as the context to prompt the VLM. LLaVA failed the first sample due to limited OCR capability, and failed the third examples by repeating the second sample semantic. + +![](images/787e50c20c377486f4d9a95d6d81301ab01030e131518bd207d6a8f1aa8690e1.jpg) +Surrealism + +![](images/2f3f287ea3eb502a0a7a5d066a589f5b9420772267c589b26b15ded607e2126f.jpg) +? +LLaVA-1.5: Surrealism +Ours: Impressionism + +
ResolutionProjector#TokensEvaluation ↑
OKVQATextVQACOCO
224linear25649.9%41.6%116.0
336linear57649.7%49.8%117.7
336downsample14449.3%45.6%115.7
+ +Table 7. Improving the image resolution from 224 to 336 can significantly improve TextVQA accuracy. The raw resolution matters more than #tokens; high-resolution with token downsampling works better than low-resolution. We report accuracy for OKVQA and TextVQA, and CIDEr for COCO. Note: the evaluation protocol is different from Table 5 and can only be compared within the table. + +# 4.4. Other Learnings. + +Image resolution matters, not #tokens. We chose an image resolution of $336^{2}$ since it provides more fine-grained details compared to $224^{2}$ , leading to improved accuracy on tasks like TextVQA [50]. As shown in Table 7, increasing + +![](images/762bac7201380454e14b7bd27cd958bc992333275a2b3c06429e5687034e7e51.jpg) +Figure 6. Our model VILA can reason over multiple images thanks to the pre-training process. The samples are taken from [5, 60]. +Example: Visual Chain-of-Though (CoT) + +![](images/9d1812cc2c8be4f7f591e77bd6d6114593dd177c5afa0425b097f03cfb70cc82.jpg) + +Prompt: Photo: \nMenu: \nHow much should I pay for all the beer on the table according to the price on the menu? + +Answer: According to the menu, the price for the beer on the table is $6. + +# Prompt: ... + Think step-by-step. + +Answer: The beer on the table is Magna, which is priced at $6.00. There are two bottles of Magna on the table. Therefore, the total cost of the beer should be $6.00 x 2 = $12.00. + +![](images/00f75f8675192043637d340b0f48ae3523938c2fc46ceffa077137af0ed44139.jpg) +Figure 8. Our model is able to perform chain-of-thought reasoning given visual inputs. It is able to generate the correct answer when adding "Think step-by-step" to the prompt. Zoom in for a better view of the image details. Samples from [19, 60]. + +Prompt: \nI am getting just two custom pizza for me and my friend. How much should I pay in total according to the image? + +Answer: According to the image, you should pay $11.99 for two custom pizzas. + +# Prompt: ... + Let's think step-by-step. + +Answer: The image shows a chalkboard menu with a special for today, which is to create your own pizza. The price for this special is $9.99. Since you are getting two custom pizzas, you should pay $9.99 x 2 = $19.98 for the two pizzas. So, the answer is $19.98. + +the resolution from 224 to 336 can improve the TextVQA accuracy from $41.6\%$ to $49.8\%$ . However, a higher resolution leads to more tokens per image ( $336^2$ corresponds to 576 tokens/image) and a higher computational cost. It also limits the number of demonstrations for in-context learning. + +Luckily, we find that the raw resolution matters more than the #visual tokens/image. We can use different projector designs to compress the visual tokens. Here we try a "downsample" projector, which simply concatenates every $2 \times 2$ tokens into a single one and use a linear layer to fuse the information. It reduces the #tokens to 144 under the 336 resolution, that is even smaller than the 224+linear + +
#ParamVLM acc. (avg)
0-shot4-shot
Visual Expert [58]1.9×67.0%64.8%
Fine-tune1×71.0%72.1%
+ +Table 8. Directly fine-tuning the LLM during pre-training leads to better VLM accuracy and in-context learning capabilities. It also enjoys a smaller model size. Both settings are pre-trained on the MMC4-core dataset [69]. + +setup. Nonetheless, the TextVQA accuracy is higher ( $\sim 46\%$ vs. $41.6\%$ ), despite still $3\%$ worse compared to $336+$ linear setup, showing a large redundancy in the image tokens. The gap on other datasets such as OKVQA and COCO is smaller since they usually require higher-level semantics. + +In our main results, we did not use any token compression methods to provide the best accuracy despite this encouraging observation, and leave it to future work. + +Comparison to frozen LLMs with visual experts. Another interesting method for retaining the text capabilities of LLMs during the pre-training is to freeze the base LLM and add an extra visual expert to process the visual tokens [58]. The definition of expert is similar to MoE frameworks, but with a manual routing mechanism according to token types. Since the base LLM is frozen, the model fully retains the original functionality for text-only inputs during pre-training. However, we find that directly fine-tuning the LLM during visual language pre-training still leads to a better VLM accuracy and in-context learning capability (Table 8). Adding an extra visual expert also leads to near $2 \times$ model size increase, which is not friendly for edge deployment. Therefore, we chose to directly fine-tune the base LLM. + +Comparison to PEFT/LoRA. In addition to visual experts, we also provide extra results when performing LoRA tuning with rank 64 (7B model) in Table 9. Fine-tuning LLM outperforms LoRA tuning by a large margin. + +
VQA v2GQAVisWizVQATMMBLLaVA W
LoRA (r=64)69.454.348.450.060.351.2
Fine-tune (ours)79.962.357.864.468.969.7
+ +Table 9. Fine-tuning LLM consistently outperforms LoRA tuning. + +Reformatting the interleaved structure. For additional insights we also reformatted the MMC4 dataset to be $\langle \text{im1} \rangle < \text{im2} \rangle < \text{txt1} \rangle < \text{txt2} \rangle$ instead of the $\langle \text{im1} \rangle < \text{txt1} \rangle < \text{im2} \rangle < \text{txt2} \rangle$ and evaluated the model under the setting in Table 1. We observed that the reformatted MMC4 degrades the average 0-shot accuracy (on the 4 benchmarks) by $4.4\%$ , and degrades the average 4-shot accuracy by $37.5\%$ . The disorder breaks in-context learning capability, showing the importance of interleaved data. + +# 5. Related Work + +Large language models (LLMs). LLMs based on Transformers [57] have fundamentally changed the language processing field. They are achieving increasing capabilities by scaling up the model size and the pre-training corpus [1, 9, 15, 18, 20, 22, 27, 47, 51]. It is believed that most the capability of the LLM is obtained from the large-scale pre-training process, which are later unlocked through instruction tuning [16, 43, 44]. There is a growing effort from the open-source community to build a strong base LLM [55, 56, 65], the conversational variants [14, 54] and the parameter-efficient finetuned versions of large LLMs [40, 64]. In this work, we start with the base Llama-2 model [56]. + +Visual language models (VLMs). VLMs are LLMs augmented with visual inputs to provide a unified interface for visual language tasks. There are two main designs for VLMs: 1. cross-attention based, where the LLM is frozen while the visual information is fused into intermediate embeddings with a cross-attention mechanism [5, 6]; 2. auto-regressive based, where the visual input is tokenized and fed to the LLM alongside text tokens [2, 4, 7, 13, 19, 33, 37, 61, 68]. The latter is a natural extension by treating visual inputs as a foreign language. VLMs are also instruction-tuned so that they can better follow human instructions or perform conversations [17, 37, 52]. In this work, we study the pre-training process of the auto-regressive VLMs due to their flexibility when handling multi-modal inputs. + +Following text-only LLMs, people also study different training recipes for VLMs. Some work freezes the LLM and train auxiliary components [5, 32, 33, 58], others fine-tune the LLM to enable visual capabilities [13, 19, 66]. There is also usage of different data corpora, including image-text pairs [13, 19, 32, 37], interleaved datasets [6], video-text pairs [41], visual-grounded annotations [36, 45], etc. In this work, we provide a holistic ablation of different design choices for the pre-training stage. + +# 6. Conclusion + +This paper has explored effective pretraining design options to augment LLMs towards vision tasks. Leveraging full strength of LLM learning, interleaved-nature of image-text data, and careful text data re-blending, VILA has surpassed state-of-the-art methods for vision tasks while preserving text-only capabilities. VILA has also depicted strong reasoning capability for multi-image analysis, in-context learning and zero/few-shot tasks. We hope our paper can help spur further research on VLM pretraining. + +Acknowledgements. We thank Bryan Catanzaro for fruitful discussions. We appreciate the help from Zhuolin Yang, Guilin Liu, Lukas Voegtle, Philipp Fischer, Karan Sapra and Timo Roman on dataset preparation and feedback. + +# References + +[1] GPT-4 technical report. Technical report, OpenAI, 2023. https://arxiv.org/abs/2303.08774. +[2] Fuyu-8B: A multimodal architecture for AI agents. https://www.adept.ai/blog/fuyu-8b, 2023. +[3] Yi-34B large language model. https://huggingface.co/01-ai/Yi-34B, 2023. +[4] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023. +[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. +[6] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023. +[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. Technical report, Alibaba Group, 2023. https://arxiv.org/abs/2303.08774. +[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. +[10] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/ coyo-dataset, 2022. +[11] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. +[12] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. +[13] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. + +[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality, 2023. +[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. +[16] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. arXiv preprint arXiv:2210.11416, 2022. +[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. +[18] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. +[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. +[20] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. +[21] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL, 2019. +[22] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. +[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. +[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. +[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on + +computer vision and pattern recognition, pages 3608-3617, 2018. +[26] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. CoRR, abs/2009.03300, 2020. +[27] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022. +[28] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. +[29] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023. +[30] Siddharth Karamcheti, Laurel Orr, Jason Bolton, Tianyi Zhang, Karan Goel, Avanika Narayan, Rishi Bommasani, Deepak Narayanan, Tatsunori Hashimoto, Dan Jurafsky, et al. Mistral—a journey towards reproducible language model training, 2021. +[31] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. +[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. +[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. +[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. 2023. +[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. +[39] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via + +thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. +[40] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024. +[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. +[42] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. +[43] OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt, 2023. Accessed: 2023. +[44] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. +[45] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[47] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021. +[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. +[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. +[50] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. +[51] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, + +Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, et al. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model. arXiv preprint arXiv:2201.11990, 2022. +[52] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. +[53] Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022. +[54] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023. +[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[56] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. +[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. +[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. +[60] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 2023. +[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. +[62] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. +[63] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. + +Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. +[64] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. +[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022. +[66] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. +[67] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. +[68] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. +[69] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023. \ No newline at end of file diff --git a/2024/VILA_ On Pre-training for Visual Language Models/images.zip b/2024/VILA_ On Pre-training for Visual Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..390847d98603ebb608e53f533af6420748dd3647 --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7858c4cf91dafd24f26c703f290377a58fac960da5734d33c1cd15449997231 +size 511664 diff --git a/2024/VILA_ On Pre-training for Visual Language Models/layout.json b/2024/VILA_ On Pre-training for Visual Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5a4c3e6c184936449495aa4fc10021cc2a34fc1 --- /dev/null +++ b/2024/VILA_ On Pre-training for Visual Language Models/layout.json @@ -0,0 +1,9755 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 459, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 459, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 459, + 121 + ], + "type": "text", + "content": "VILA: On Pre-training for Visual Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "spans": [ + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": "Ji Lin" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1,2*†}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": " Hongxu Yin" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": " Wei Ping" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": " Pavlo Molchanov" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": " Mohammad Shoeybi" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": " Song Han" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 124, + 141, + 465, + 184 + ], + "type": "text", + "content": "MIT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 187, + 432, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 187, + 432, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 187, + 432, + 198 + ], + "type": "text", + "content": "ViLA https://github.com/Efficient-Large-Model/VILA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 265, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 290, + 553 + ], + "type": "text", + "content": "Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pretraining data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 597, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 288, + 693 + ], + "type": "text", + "content": "Large language models (LLMs) have demonstrated superior capabilities for natural language tasks [3, 7, 9, 14, 15, 18, 30, 43, 48, 54-56]. Augmenting LLMs to support visual inputs allows the final model to inherit some of the appealing properties like instruction following, zero-shot generalization, and few-shot in-context learning (ICL), empowering various visual language tasks [1, 2, 5, 8, 13, 19, 33, 37, 68]. The central challenge of unifying vision and language for" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 225, + 536, + 422 + ], + "blocks": [ + { + "bbox": [ + 313, + 225, + 536, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 225, + 536, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 536, + 422 + ], + "type": "image", + "image_path": "27534e03254453598560de2a3d0d2ed6661a6a83ec42c3d343e5beda0bcfdcd4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 423, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 423, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 547, + 456 + ], + "type": "text", + "content": "Figure 1. VILA's enhanced visual-language pre-training consistently improves the downstream task accuracy under a comparison to recent methods [7, 17, 37]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 472, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 604 + ], + "type": "text", + "content": "collaborative inference resides in connecting the LLM and the vision foundation model (e.g., a CLIP encoder): both foundation models are usually pre-trained individually, before aligned via vision-language joint training. Most of the efforts in this field have been focusing on improving the visual language instruction-tuning process, i.e., supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF) [36, 37, 52]. However, there lacks a thorough study of the pre-training process, where the model is trained on image-text datasets/corpora at scale [10, 49, 69]. This process is costly but critical for the modality alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "In this work, we aim to explore different design options for enhanced visual language model pre-training. In particular, we aim to answer \"How do various design choices in visual language model pre-training impact the downstream performance?\" We followed the pre-training + SFT pipeline and ablated different design options for pre-training overseeing dataset properties and training protocols. We discover several findings: (1) Freezing the LLM during pretraining can achieve a decent zero-shot performance, but" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 282, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Work done during an internship at NVIDIA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26689" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 293, + 186 + ], + "type": "image", + "image_path": "ebb73adf99d692f0cd74a3d6f8a92538fad46860aa9bb5e398b19b8a0b96f011.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 288, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 251 + ], + "type": "text", + "content": "Figure 2. We study auto-regressive visual language model, where images are tokenized and fed to the input of LLMs. We find updating the LLM is essential for in-context learning capabilities, and interleaved corpus like [69] helps pre-training. Joint SFT with text-only data helps maintain the text-only capabilities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 259, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 354 + ], + "type": "text", + "content": "not in-context learning (ICL) capability, whereas updating the LLMs encourages deep embedding alignment, which we found is important for ICL; (2) Interleaved visual language data is essential for pre-training, that provides accurate gradient update and maintains text-only capability; (3) Adding in text-only instruction data during SFT can further remedy text-only degradation and boost visual language task accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "content": "We introduce practical guidance to design Visual Language models, dubbed VILA. Without bells and whistles, VILA outperforms the state-of-the-art model [36] by noticeable margins across a wide range of vision language tasks (Figure 1), thanks to the help of improved pre-training. Moreover, we observe that the pre-training process unlocked several interesting capabilities for the model, such as (i) multi-image reasoning (despite the model only sees single image-text pairs during SFT), (ii) stronger in-context learning capabilities, and (iii) enhanced world knowledge. We hope our findings can provide a good pre-training recipe for future visual language models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 509, + 126, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 126, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 126, + 521 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": "Model architecture. Multi-modal LLMs can be generally categorized into two settings: cross-attention-based [5, 33] and auto-regressive-based [2, 19, 37]. The latter VLM family tokenizes images into visual tokens, which are concatenated with textual tokens and fed as the input to LLMs (i.e., treating visual input as a foreign language). It is a natural extension of text-only LLMs by augmenting the input with visual embeddings and can handle arbitrary interleaved image-text inputs. In this study, we focus on the pre-training of auto-regressive VLMs due to its flexibility and popularity. As shown in Figure 2, auto-regressive VLMs consists of three components: a visual encoder, an LLM, and a projector that bridges the embeddings from the two modalities. The projector can be a simple linear layer [37] or more capable Transformer blocks [6, 17] – we will compare their efficacy in our experiments. The model takes visual and text input" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 73, + 414, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 414, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 414, + 84 + ], + "type": "text", + "content": "and generates text outputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 95, + 547, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 547, + 143 + ], + "type": "text", + "content": "Training stages. Following common practice [6, 19, 37], we study how to augment a pre-trained text-only LLM with visual input support. The training can be categorized into three stages:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 547, + 361 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 304, + 144, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 204 + ], + "type": "text", + "content": "0. **Projector initialization.** The LLM and ViT are separately pre-trained, while the projector is usually initialized from random weights. Therefore, we first pre-train the projector while freezing both ViT and LLMs on image-caption pairs following existing literature [17, 33, 37]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 289 + ], + "type": "text", + "content": "1. visual language pre-training. We then pre-train the model (LLM and the projector) on visual language corpus. We consider two types of corpus: interleaved image-text corpus (e.g., MMC4 [69]) and image-text pairs (e.g., COYO [10] and LAION [49]). We focus the study of this work on the pre-training process, which are most costly and important for visual language alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "content": "2. Visual instruction-tuning. Finally, we further perform instruction tuning of the pre-trained model on visual language instruction datasets. We convert existing visual language datasets into FLAN [59] style (i.e., with dataset-specific prompts) following [17]. Please find the data blend of the visual instruction data in the supplementary." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 373, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 445 + ], + "type": "text", + "content": "**Evaluations.** During our ablation study, we evaluate the fine-tuned model on 4 visual language tasks: accuracy for OKVQA [42] and TextVQA [50], and CIDEr score for COCO [35] and Flickr [62]. We evaluate both 0-shot and 4-shot performance, which reflects the models' in-context learning capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 457, + 545, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 472 + ], + "type": "text", + "content": "3. On Pre-training for Visual Language Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 479, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 503 + ], + "type": "text", + "content": "In this section, we discuss practical design choices and learned lessons for the visual language pre-training process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 512, + 455, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 455, + 524 + ], + "type": "text", + "content": "3.1. Updating LLM is Essential" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 525, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 632 + ], + "type": "text", + "content": "Fine-tuning vs. prompt tuning. There are two popular ways to augment a pre-trained text-only LM with visual inputs: fine-tune LLMs on the visual input tokens [19, 37], or freeze the LLM and train only the visual input projector as prompt tuning [17, 33]. The latter is attractive since freezing the LLMs prevents the degradation of the pre-trained text-only LLM. Nonetheless, we found updating the base LLM is essential to inherit some of the appealing LLM properties like in-context learning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 633, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 547, + 693 + ], + "type": "text", + "content": "To verify the idea, we compare the two training protocols in Table 1. We use a Transformer block for the projector instead of a single linear layer [37] in setting a-c, which provides enough capacity when freezing LLMs. We use MMC4-core " + }, + { + "bbox": [ + 304, + 633, + 547, + 693 + ], + "type": "inline_equation", + "content": "[69]^*" + }, + { + "bbox": [ + 304, + 633, + 547, + 693 + ], + "type": "text", + "content": " for the comparison. We observed that:" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 543, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 543, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 543, + 713 + ], + "type": "text", + "content": "*We downloaded only 25M of 30M images amid some expired URLs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26690" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 537, + 159 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 537, + 159 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 537, + 159 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 537, + 159 + ], + "type": "table", + "html": "
PreTSFTProjectileOKVQATextVQACOCOFlickrAverage
TrainLLM?0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot0-shot4-shot
(a)XXTransformer10.419.214.823.117.460.211.047.413.437.5
(b)X✓Transformer47.147.737.236.6109.488.073.658.166.857.6
(c)✓✓Transformer44.849.838.538.8112.3113.571.572.966.868.8
(d)✓✓Linear45.250.339.740.2115.7118.574.274.768.770.9
", + "image_path": "e902d9ae2ce39ab62a7ea2fc341aa46636839292a712a681c8a6d08e7af83ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 160, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 547, + 216 + ], + "type": "text", + "content": "Table 1. Ablation study on whether to train LLM or freeze LLM and only perform prompt tuning during visual language pre-training (PreT). Interestingly, freezing the LLM during pre-training does not hurt the 0-shot accuracy, but leads to worse in-context learning capability (worse 4-shot). Using a simple linear projector forces the LLM to learn more and leads to better generalization. We report accuracy for VQA datasets (OKVQA, TextVQA) and CIDEr score for captioning (COCO and Flickr). Note: we used a different evaluation setting just for ablation study; the absolute value in this setting is lower and should not be compared against other work." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 225, + 265, + 348 + ], + "blocks": [ + { + "bbox": [ + 66, + 225, + 265, + 348 + ], + "lines": [ + { + "bbox": [ + 66, + 225, + 265, + 348 + ], + "spans": [ + { + "bbox": [ + 66, + 225, + 265, + 348 + ], + "type": "image", + "image_path": "fb9a16df2a643fb262bb55a9c4b08fc72f95f9297f083535f9a60604ed5f4d8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 396 + ], + "type": "text", + "content": "Figure 3. Prompt-tuning to support visual tokens can only enable shallow alignment, while fine-tuning the LLM leads to alignment at deeper layers. From configuration (b) to (d) (as in Table 1), the alignment improves at deeper layer, so as ICL accuracy (4-shot)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 406, + 287, + 586 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 46, + 406, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 442 + ], + "type": "text", + "content": "(1) Training only the projector during SFT leads to poor performance (setting a), despite using a high-capacity design. It is rewarding to fine-tune LLM during SFT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "type": "text", + "content": "(2) Interestingly, freezing the LLM during pre-training does not affect 0-shot performance, but degrades in-context learning capabilities (i.e., 4-shot, comparing setting b and c). The gap is even larger for captioning datasets (COCO & Flickr) since they are out-of-distribution (the instruction tuning data is mostly VQA-alike, see supplementary), showing the worse generalization capability when freezing LLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "type": "text", + "content": "(3) When using a small-capacity projector (a linear layer instead of a Transformer block), the accuracy is slightly better (comparing c and d). We hypothesize a simpler projector forces the LLM to learn more on handling visual inputs, leading to better generalization." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "content": "The deep embedding alignment hypothesis. To understand why fine-tuning LLM is beneficial, we hypothesize that it is important to align the distribution of visual and textual latent embeddings (especially in the deeper layers), so that the model can seamlessly model the interaction between the two modalities. It is essential if we want to inherit some of the good properties of LLM like in-context learning for visual language applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "To verify the idea, we calculate the Chamfer distance of visual and textual embeddings in different layers to measure" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 224, + 548, + 271 + ], + "blocks": [ + { + "bbox": [ + 306, + 224, + 548, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 224, + 548, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 548, + 271 + ], + "type": "table", + "html": "
DatasetTypeText Src.#img/sample#tok./img
MMC4 [69]InterleaveHTML4.0122.5
COYO [10]Img-text pairAlt-text122.7
", + "image_path": "689ba363ccef0585a91033c231720bd5d9618375d8f0ed923d50a804ca551df9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 278, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 547, + 334 + ], + "type": "text", + "content": "Table 2. Two image-text corpus considered for pre-training. The COYO captions are generally very short, which has a different distribution compared to the text-only corpus for LLM training. We sample each data source to contain 25M images by choosing samples with high CLIP similarities." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 319, + 346, + 533, + 446 + ], + "blocks": [ + { + "bbox": [ + 319, + 346, + 533, + 446 + ], + "lines": [ + { + "bbox": [ + 319, + 346, + 533, + 446 + ], + "spans": [ + { + "bbox": [ + 319, + 346, + 533, + 446 + ], + "type": "table", + "html": "
Pre-train DataVLM acc (avg)MMLU acc.
0-shot4-shot
Llama-2--46.0%
COYO51.1%50.3%28.8% (-17.2%)
MMC4-pairs46.4%44.5%32.4% (-13.6%)
MMC468.7%70.9%40.7% (-5.3%)
MMC4+COYO69.0%71.3%40.2% (-5.8%)
", + "image_path": "0f0f26a67643d4c03fd2ea00b878cb56cf449d8fb750190eed0e4275a775eefa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 453, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 547, + 521 + ], + "type": "text", + "content": "Table 3. Pre-training on MMC4 data provides better visual language accuracy (0-shot and few-shot) and smaller degradation on text-only accuracy compared to caption data (COYO). The benefits comes from the interleave nature but not the better text distribution (MMC4 vs. MMC4-pairs). Blending interleaved and caption data provides a better diversity and downstream accuracy." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "type": "text", + "content": "how well they align in Figure 3. We calculate the pairwise cosine similarity to exclude the affect of magnitude. From configuration (b) to (d), the similarity of deeper layer goes higher, so as the 4-shot accuracy in Table 1, showing the positive relationship between deep embedding alignment and in-context learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 614, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 651 + ], + "type": "text", + "content": "Given the observations, we fine-tune the LLM during both pre-training and instruction-tuning in later studies, and use a simple linear projection layer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "text", + "content": "3.2. Interleaved Visual Language Corpus Helps Pre-training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our goal is to \"augment\" the LLM to support visual input, instead of training a model that only works well on visual" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26691" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 103, + 119 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 103, + 119 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 103, + 119 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 103, + 119 + ], + "type": "image", + "image_path": "d7d481958cc4c790af80703c2561cfdda658bb50d21d459440399a781ef17e21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 129, + 544, + 151 + ], + "lines": [ + { + "bbox": [ + 46, + 129, + 544, + 151 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 544, + 151 + ], + "type": "text", + "content": "Figure 4. A sample from MMC4 [69] dataset consisting of interleaved images and text segments. The images are placed before the corresponding text. The text are weakly conditioned on images: only colored text can be better inferred with the help of images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 72, + 157, + 119 + ], + "blocks": [ + { + "bbox": [ + 104, + 72, + 157, + 119 + ], + "lines": [ + { + "bbox": [ + 104, + 72, + 157, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 72, + 157, + 119 + ], + "type": "image", + "image_path": "2b97409618609b8ed58bd87e4b7804e973e6a0857ddb3bc4f48267edd734704a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 70, + 544, + 122 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 544, + 122 + ], + "type": "text", + "content": "...,\"Easy to grow, and fruitful, these are a favorite amongst those who like to have their own at home garden.\", , \"There is a large number of different kinds of tomatoes, meaning that there is one for every different size garden.\", \"You can choose from a variety of these plants to accomplish different food missions, be them small cucumbers for pickling or larger ones for slicing.\", ..., , \"These tender green beans are durable and quite at home growing in the hot season air.\", \"To get the most out of planting them, space the seeds 4 to 6 inches apart and allow 3 feet between rows, ..." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 156, + 279, + 246 + ], + "blocks": [ + { + "bbox": [ + 53, + 156, + 279, + 246 + ], + "lines": [ + { + "bbox": [ + 53, + 156, + 279, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 279, + 246 + ], + "type": "image", + "image_path": "bd1a94b4a45456d974e21204bb803f015894c921fc39734db13a38a88618ee65.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "type": "text", + "content": "Figure 5. Loss is lower when pre-training on MMC4 compared to MMC4-pairs (samples broken into image-text pairs), since the text segments provide more information for language modeling." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "content": "language inputs. Therefore, it is essential to preserve the text-only capabilities of LLMs. We found that data blending is a key factor, both for pre-training and instruction tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 335, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 514 + ], + "type": "text", + "content": "Pre-training dataset options. Most of the VLM pretraining [33, 37, 58] relies on image-text pairs (i.e., image and captions) due to the wide availability and large diversity (e.g., LAION [49], COYO [10]). On the other hand, interleaved image-text datasets (MMC4 [69], M3W [5]) follow a more similar distribution compared to the text-only corpus and is found to be important in Flamingo-style model training [5]. We hypothesize that the interleaved dataset is even more important for VLMs when LLM backbone is updated to accommodate the visual input. For a better understanding of the two data types, we compare statistics in Table 2: COYO suffers from a short text distribution since the accompanying text is taken from alt-text. We subsample the COYO dataset by ranking CLIP similarities and keep only 25M images (a similar size as MMC4-core)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "content": "We follow the same pre-training + SFT process and ablate different pre-training corpus. We compare the 0-shot and few-shot visual language accuracy as well as text-only accuracy (MMLU [26]) in Table 3. Due to space limit, we report the average accuracy over four datasets (as in Table 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Interleaved data is essential. We notice using image-text pairs (i.e., COYO) for pre-training can lead to catastrophic forgetting. The text-only accuracy (MMLU) degrades by " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "17.2\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". The visual language accuracy is also much worse compared to MMC4 pre-training. Noticeably, the 4-shot accuracy is even worse than 0-shot, showing the model cannot properly do in-context learning for visual language inputs (probably because it never sees more than one image during pre-training). We hypothesize the catastrophic forgetting is due to the distribution of text-based captions, which are generally very short and concise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "text", + "content": "On the contrary, dataset like MMC4 has a much closer distribution compared to text-only corpus (e.g., C4 [48]). When using the interleaved data for pre-training, the degradation on MMLU is only " + }, + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\sim 5\\%" + }, + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "text", + "content": ". The degradation would be even smaller when using a larger base LLM [19]. With proper instruction tuning (Section 3.3), this degradation can be fully recovered. It also promotes visual in-context learning, leading to a higher 4-shot accuracy compared to 0-shot." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 260, + 547, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 547, + 356 + ], + "type": "text", + "content": "Interleave data structure matters, but not the text distribution. We further question whether the benefits come from the better text distribution (e.g., longer) or from the interleave nature. To ablate this, we construct a new MMC4 variant by only keeping the images and their corresponding text segments, without considering the interleave nature, denoted as \"MMC4-pairs\". For example an MMC4 sample may look like:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 361, + 518, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 361, + 518, + 371 + ], + "spans": [ + { + "bbox": [ + 332, + 361, + 518, + 371 + ], + "type": "interline_equation", + "content": "< \\text {t x t 1} > < \\text {i m 1} > < \\text {t x t 2} > < \\text {t x t 3} > < \\text {i m 2} > < \\text {t x t 4} >", + "image_path": "e73247ff48472cea819df41e1c402989a6d057e81630669abcf113ee0e9f6526.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 380, + 520, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 380, + 520, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 520, + 393 + ], + "type": "text", + "content": "It will be converted into two MMC4-pairs samples:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 399, + 492, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 399, + 492, + 408 + ], + "spans": [ + { + "bbox": [ + 359, + 399, + 492, + 408 + ], + "type": "interline_equation", + "content": "< \\mathbf {i m} 1 > < \\mathbf {t x t} 2 >, < \\mathbf {i m} 2 > < \\mathbf {t x t} 4 >", + "image_path": "67204c4ff0cc5c7e2fd5d8dd7e7d7111e9283c4ac4251c71dbd873304aec58a6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 418, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 598 + ], + "type": "text", + "content": "However, training on MMC4-pairs does not lead to a satisfactory result: it slightly reduces the degradation on MMLU due to a longer text distribution, but the VLM accuracy is even lower compared to pre-training on COYO; there is also no in-context improvement. We hypothesize the MMC4 samples do not have a very strict image-text correspondence; the image only provides marginal information for text modeling (i.e., most of the information is still from pure text modeling; an example is provided in Figure 4). It is also demonstrated by the loss curves in Figure 5, where training on the interleave corpus leads to a much lower loss, indicating the full text segments provides more information. Therefore, the interleaved data structure is critical, allowing the model to pick up the image-related information, without over-forcing it to learn unrelated text modeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": "Data blending improves pre-training. Training on image-text pairs only led to a sharp degradation on text-only accuracy (more than " + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": "). Luckily, blending the interleaved corpus and image-text pairs allows us to introduce more diversity in the corpus, while also preventing the severe degradation. Training on MMC4+COYO further boosts the accuracy on visual language benchmarks (the gain is larger when we perform joint SFT, as we will show later (Table 4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 315, + 702, + 545, + 712 + ], + "type": "text", + "content": "We followed [69] to match the image and text segments by CLIP scores." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26692" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 173 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 173 + ], + "type": "table", + "html": "
PT dataSFT dataVLM acc. (avg)MMLU acc.
0-shot4-shot
Llama-2---46.0%
MMC4Visual68.7%70.9%40.7% (-5.3%)
MMC4+COYOVisual69.0%71.3%40.2% (-5.8%)
Llama-2Text--51.2%
MMC4Vis.+Text71.0%72.1%51.4% (+0.2%)
MMC4+COYOVis.+Text72.3%73.6%50.9% (-0.3%)
", + "image_path": "9633c0d723bfee93ed173248fb51b402c9c49175b34798f4fc32452826f0d844.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "type": "text", + "content": "Table 4. Joint SFT (Vis. + Text) not only bridges the degradation of text-only capability (MMLU acc.), but also improves the performance on visual-language tasks (both zero-shot and few-shot)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 216, + 270, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 270, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 270, + 228 + ], + "type": "text", + "content": "3.3. Recover LLM Degradation with Joint SFT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 290, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 290, + 369 + ], + "type": "text", + "content": "Despite the interleave data helps maintain the text-only capability, there is still a " + }, + { + "bbox": [ + 46, + 236, + 290, + 369 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 236, + 290, + 369 + ], + "type": "text", + "content": " accuracy drop. A potential approach is to maintain the text-only capability would be to add in text-only corpus (the one used in the LLM pre-training). However, such text corpus are usually proprietary even for open-source models; it is also unclear how to subsample the data to match the scale of vision-language corpus. Luckily, we found the text-only capabilities are temporarily hidden, but not forgotten. Adding in text-only data during SFT can help bridge the degradation, despite using a much smaller scale compared to the text pre-training corpora (trillion scale)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 382, + 288, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 490 + ], + "type": "text", + "content": "Joint supervised fine-tuning. The common way for instruction tuning is to fine-tune the model on some visual language datasets (VQA/Caption style [17] or GPT-generated [37]). We found blending in text-only instruction data can simultaneously (i) recover the degradation in text-only accuracy, and (ii) improve the visual language accuracy. To this end, we also blended in 1M text-only instruction tuning data sampled from FLAN [16], which we termed as joint SFT. We provide the comparison in Table 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 635 + ], + "type": "text", + "content": "We can see that blending in the text-only SFT data not only bridges the degradation on text-only capability (the MMLU accuracy is on par compared to the original Llama-2 model fine-tuned on the same text-only instruction data), but also improves the visual language capability. We hypothesize that the text-only instruction data improves the model's instruction-following capability, which is also important for visual language tasks. Interestingly, the benefits of blending in COYO data is more significant with joint SFT. We believe that with joint SFT, the model no longer suffers from the text-only degradation when pre-trained with short captions, thus unlocking the full benefits from the better visual diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 649, + 129, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 129, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 129, + 662 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 208, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 208, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 208, + 684 + ], + "type": "text", + "content": "4.1. Scaling up VLM pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We scale up the training of VLM in the following aspects to form our final model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "Higher image resolution. Above ablation studies used the OpenAI CLIP-L [46] with " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " resolutions as the visual encoder. We now use " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " image resolutions to include more visual details for the model, which can help tasks requiring fine-grained details (e.g., TextVQA [50])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 139, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 547, + 176 + ], + "type": "text", + "content": "Larger LLMs. By default, we used Llama-2 [56] 7B for ablation study. We also scaled to a larger LLM backbone (e.g., Llama-2 [56] 13B) to further improve the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "type": "text", + "content": "Pre-training data. We used both interleaved image-text data and image-text pairs for pre-training (we sample roughly 1:1 image proportions) to improve the data diversity. The total the pre-training corpus contains about 50M images. It is smaller than the billion-scale pre-training data [5, 13, 58], but already demonstrates impressive improvements on downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 274, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 334 + ], + "type": "text", + "content": "SFT data. We also include a better SFT data blend from LLaVA-1.5 [36], which is more diverse (e.g., contains reference-based annotations) and has high-quality prompt. The new SFT data blend can significantly improve the downstream evaluation metrics. We include details the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 341, + 547, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 401 + ], + "type": "text", + "content": "Limitations. Due to the limited compute budget, we have not been able to further scale up the size of the pre-training corpus to billion-scale, which we leave as future work. Nonetheless, pre-training on 50M images already demonstrated significant performance improvement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 407, + 441, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 441, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 441, + 419 + ], + "type": "text", + "content": "4.2.Quantitative Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 419, + 547, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 611 + ], + "type": "text", + "content": "visual language tasks. We perform a comprehensive comparison with state-of-the-art models on 12 visual language benchmarks in Table 5. Compared to existing models (e.g., LLaVA-1.5 [36]), our model achieves consistent improvements over most datasets at different model sizes under a head-to-head setting (using the same prompts and base LLM; Vicuna-1.5 is based on Llama-2). Remarkably, we 7B model is able to outperform LLaVA-1.5 13B on VisWiz [25] and TextVQA [50] by a large margin thanks to the pre-training. Our 7B model even outperforms the 13B LLaVA model on these datasets. Our model also has multi-lingual capability despite the vision-language instruction data is in English, outperforming LLaVA-1.5 on MMBench-Chinese benchmark. Our results demonstrate the benefits of vision-language pre-training on downstream tasks, even when using a high-quality instruction tuning dataset [36]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": "Text-only performance. Our goal is to augment an LLM to support visual inputs. It is essential that the model can retain the text-only capability. Therefore, we further evaluate the text-only performance of the models under three benchmarks: MMLU [26], BBH [53], and DROP [21] in Table 6. We did not choose benchmarks like MT-Bench [67] since text instruction tuning is not the focus of the work. Overall, our model achieves performance as Llama-2 fine-tuned with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26693" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 553, + 210 + ], + "type": "table", + "html": "
MethodLLMRes.PTIT\\(VQA^{v2}\\)GQAVisWiz\\(SQA^I\\)\\(VQAT^T\\)POPEMMEMMB\\(MMBC^{CN}\\)SEED\\(LLaVA^W\\)MM-Vet
BLIP-2 [33]Vicuna-13B224129M-41.04119.66142.585.31293.8--46.438.122.4
InstructBLIP [17]Vicuna-7B224129M1.2M-49.234.560.550.1--3623.753.460.926.2
InstructBLIP [17]Vicuna-13B224129M1.2M-49.533.463.150.778.91212.8---58.225.6
Shikra [11]Vicuna-13B224600K5.5M77.4*------58.8----
IDEFICS-9B [29]LLaMA-7B224353M1M50.938.435.5-25.9--48.225.2---
IDEFICS-80B [29]LLaMA-65B224353M1M60.045.236.0-30.9--54.538.1---
Qwen-VL [8]Qwen-7B4481.4B50M78.8*59.3*35.267.163.8--38.27.456.3--
Qwen-VL-Chat [8]Qwen-7B4481.4B50M78.2*57.5*38.968.261.5-1487.560.656.758.2--
LLaVA-1.5 [36]Vicuna-1.5-7B3360.6M0.7M78.5*62.0*50.066.858.285.91510.764.358.358.663.430.5
LLaVA-1.5 [36]Vicuna-1.5-13B3360.6M0.7M80.0*63.3*53.671.661.385.91531.367.763.661.670.735.4
VILA-7B (ours)Llama-2-7B33650M1M79.9*62.3*57.868.264.485.51533.068.961.761.169.734.9
VILA-13B (ours)Llama-2-13B33650M1M80.8*63.3*60.673.766.684.21570.170.364.362.873.038.8
+ShareGPT4VLlama-2-13B33650M1M80.6*63.2*62.473.165.384.81556.570.865.461.478.445.7
", + "image_path": "2b85835f552d1844a11f4f3267c15a14c321370cb64bc6aa056ae10ef883de4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 296 + ], + "type": "text", + "content": "Table 5. Comparison with state-of-the-art methods on 12 visual-language benchmarks. Our models consistently outperform LLaVA-1.5 under a head-to-head comparison, using the same prompts and the same base LLM (Vicuna-1.5 is based on Llama-2), showing the effectiveness of visual-language pre-training. We mark the best performance **bold** and the second-best **underlined**. Benchmark names are abbreviated due to space limits. VQA-v2 [24]; GQA [28]; VisWiz [25]; SQA†: ScienceQA-IMG [39]; VQAT: TextVQA [50]; POPE [34]; MME [23]; MMB: MMBench [38]; MMBCN: MMBench-Chinese [38]; SEED: SEED-Bench [31]; LLaVAW: LLaVA-Bench (In-the-Wild) [37]; MM-Vet [63]. *The training images of the datasets are observed during training. We also tried adding the ShareGPT4V [12] to the SFT blend on top of VILA-13B (last row), leading to a significant improvement on LLaVA-Bench and MM-Vet (marked in green)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 313, + 282, + 430 + ], + "blocks": [ + { + "bbox": [ + 52, + 313, + 282, + 430 + ], + "lines": [ + { + "bbox": [ + 52, + 313, + 282, + 430 + ], + "spans": [ + { + "bbox": [ + 52, + 313, + 282, + 430 + ], + "type": "table", + "html": "
SizeModelMMLU [26]BBH [53]DROP [21]
7BLlama-246.0%32.0%31.7%
Llama-2+SFT51.8%39.3%53.1%
Vicuna-1.549.8%36.9%29.2%
VILA50.8%38.5%52.7%
13BLlama-255.7%37.5%41.6%
Llama-2+SFT54.3%43.2%59.2%
Vicuna-1.555.8%38.4%43.6%
VILA56.0%44.2%63.6%
", + "image_path": "cbb7935cb5b9b73613802e3adf737d63edaba054e2f9cf5ab8921c52d2a60fdc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 437, + 288, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 472 + ], + "type": "text", + "content": "Table 6. VILA maintains competitive accuracy on text-only benchmarks. There is a small gap compared to the text-only model under 7B; but the accuracy is even better under 13B." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 533 + ], + "type": "text", + "content": "the same text SFT data: the accuracy of our 7B model is a bit lower, while the 13B is higher. We suspect the smaller model may suffer from a larger text performance degradation during the pre-training, as observed in [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 176, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 176, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 176, + 553 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 559, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 595 + ], + "type": "text", + "content": "Here we study how visual language pre-training enables new capabilities for the model. Part of the image samples are taken from [5, 13, 60]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Multi-image reasoning. Thanks to the pre-training, our model has the ability to reason over multiple images, despite the SFT data is composed of single-image samples. We provide two examples in Figure 6. In the first example, our model is able to figure out the common object (i.e., a flamingo) across the three images and the different art styles of each one, while the LLaVA model failed. The LLaVA model hallucinates and cannot distinguish the information from different input images. In the second example, our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 316, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 547, + 339 + ], + "type": "text", + "content": "model is able to find one of the differences (i.e., the head-wear) out of the two." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "type": "text", + "content": "In-context learning. In-context learning is an important characteristic of LLMs, allowing people to prompt the LLM with few-shot samples to enable new tasks. We provide in-context learning samples in Figure 7. Interestingly LLaVA-1.5 [36] can also perform in-context learning to some extent, despite only being trained on single-image-text-paired samples. We believe the capability is inherited from text-only pre-training of the base LLM. Nonetheless, our model outperforms LLaVA-1.5 for in-context learning: LLaVA-1.5 failed the first sample due to limited OCR capability, and failed the third example by repeating the semantics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "type": "text", + "content": "Visual Chain-of-Thoughts (CoT). Our model is able to perform chain-of-thought reasoning given visual inputs. As shown in Figure 8, VILA is able to perform complex CoT reasoning over the input images (multi-image or single-image) when adding \"Think step-by-step\" to the end of the prompt. We believe the CoT capability is inherited from text-only SFT, despite there is no such samples from the visual language instruction data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Better world knowledge. Since our model is pre-trained on a large-scale corpus, it has better understanding of world knowledge. We perform a case study by prompting the model to recognize the locations of some famous landmarks (please see supplementary due to space limits). VILA can correctly recognize 4 out of the 4 samples, while LLaVA-1.5 only gets 2 out of the 4, demonstrating the effectiveness of the pre-training. Samples are taken from [60]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26694" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 76, + 280, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 76, + 280, + 86 + ], + "spans": [ + { + "bbox": [ + 64, + 76, + 280, + 86 + ], + "type": "text", + "content": "Prompt: Image 1: \\nImage 2: \\nImage 3: \\n" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 86, + 220, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 86, + 220, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 86, + 220, + 95 + ], + "type": "text", + "content": "What is the common thing about these three images?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 95, + 394, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 95, + 394, + 129 + ], + "spans": [ + { + "bbox": [ + 65, + 95, + 394, + 129 + ], + "type": "text", + "content": "LLaVA-1.5: The common theme among these three images is the presence of a pink flamingo. In each image, the flamingo is standing in the water, with one image showing the bird in the water, another image showing the bird's legs, and the third image showing the bird's head. The flamingo is the central subject in all three images, and its distinctive pink color and unique appearance make it the focal point of each photo." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 79, + 438, + 119 + ], + "blocks": [ + { + "bbox": [ + 399, + 79, + 438, + 119 + ], + "lines": [ + { + "bbox": [ + 399, + 79, + 438, + 119 + ], + "spans": [ + { + "bbox": [ + 399, + 79, + 438, + 119 + ], + "type": "image", + "image_path": "1a6c4e05e035aa5095cc9877430e7a1eb3eddaaccd73583bd10dda611184a804.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 443, + 79, + 483, + 119 + ], + "blocks": [ + { + "bbox": [ + 443, + 79, + 483, + 119 + ], + "lines": [ + { + "bbox": [ + 443, + 79, + 483, + 119 + ], + "spans": [ + { + "bbox": [ + 443, + 79, + 483, + 119 + ], + "type": "image", + "image_path": "81ef5fec8e35979932ce96f0dc846883b309d2a590a7fda7640cb7ae95a660d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 489, + 79, + 525, + 119 + ], + "blocks": [ + { + "bbox": [ + 489, + 79, + 525, + 119 + ], + "lines": [ + { + "bbox": [ + 489, + 79, + 525, + 119 + ], + "spans": [ + { + "bbox": [ + 489, + 79, + 525, + 119 + ], + "type": "image", + "image_path": "ee40e8e5a4945c2c6b1340757f4cb815f01c14880aa7c8842f9455e89a31122c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 130, + 477, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 130, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 64, + 130, + 477, + 156 + ], + "type": "text", + "content": "Ours: The common theme among these three images is the presence of a flamingo. In the first image, the flamingo is a 3D model, while in the second image, it is a pink cartoon. In the third image, it is a real flamingo standing in the water. All three images showcase the bird in different styles and formats, but they all share the same subject." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 159, + 218, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 159, + 218, + 169 + ], + "spans": [ + { + "bbox": [ + 65, + 159, + 218, + 169 + ], + "type": "text", + "content": "Prompt: Image 1: \\nImage 2: \\n" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 169, + 239, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 169, + 239, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 169, + 239, + 178 + ], + "type": "text", + "content": "Tell me about the differences between image 1 and image 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 178, + 417, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 178, + 417, + 204 + ], + "spans": [ + { + "bbox": [ + 65, + 178, + 417, + 204 + ], + "type": "text", + "content": "LLaVA-1.5: In image 1, there is a small girl wearing a pink dress and a crown. She is smiling and appears to be happy. In image 2, the girl is wearing a pink dress and a crown, but she is not smiling. The main difference between the two images is the girl's facial expression, with one showing a smile and the other not showing any expression." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 160, + 470, + 204 + ], + "blocks": [ + { + "bbox": [ + 429, + 160, + 470, + 204 + ], + "lines": [ + { + "bbox": [ + 429, + 160, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 429, + 160, + 470, + 204 + ], + "type": "image", + "image_path": "d1701258bc25ee6cbaf68205e7a3ab9d1defa761da6acd5b28796d3d66a6fb1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 160, + 515, + 204 + ], + "blocks": [ + { + "bbox": [ + 474, + 160, + 515, + 204 + ], + "lines": [ + { + "bbox": [ + 474, + 160, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 474, + 160, + 515, + 204 + ], + "type": "image", + "image_path": "145e93d879fafb3d77c6d058099893008b3b8c07005fc051d97c4a8990cf85d4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 205, + 523, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 205, + 523, + 230 + ], + "spans": [ + { + "bbox": [ + 64, + 205, + 523, + 230 + ], + "type": "text", + "content": "Ours. The two images feature a cartoon girl wearing a pink dress. In the first image, the girl is wearing a crown, while in the second image, she is wearing a tiara. Both images showcase the girl in a princess-like attire, but the difference lies in the type of headwear she is wearing. The crown is a more traditional and regal accessory, while the tiara is a more playful and common accessory associated with princesses in popular culture." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 67, + 277, + 106, + 310 + ], + "blocks": [ + { + "bbox": [ + 67, + 260, + 187, + 269 + ], + "lines": [ + { + "bbox": [ + 67, + 260, + 187, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 260, + 187, + 269 + ], + "type": "text", + "content": "Input few shot examples + target image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 277, + 106, + 310 + ], + "lines": [ + { + "bbox": [ + 67, + 277, + 106, + 310 + ], + "spans": [ + { + "bbox": [ + 67, + 277, + 106, + 310 + ], + "type": "image", + "image_path": "270b3b736a1b61b8b88310e6a581b7a6856ffaded777f185305a6591a36bf3ff.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 311, + 106, + 319 + ], + "lines": [ + { + "bbox": [ + 66, + 311, + 106, + 319 + ], + "spans": [ + { + "bbox": [ + 66, + 311, + 106, + 319 + ], + "type": "text", + "content": "Underground." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 113, + 277, + 151, + 310 + ], + "blocks": [ + { + "bbox": [ + 113, + 277, + 151, + 310 + ], + "lines": [ + { + "bbox": [ + 113, + 277, + 151, + 310 + ], + "spans": [ + { + "bbox": [ + 113, + 277, + 151, + 310 + ], + "type": "image", + "image_path": "df5d0c291a9b2ce4f73ddda4035435bfe81f6f7e03a17ca26c999971f20016d8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 311, + 145, + 320 + ], + "lines": [ + { + "bbox": [ + 116, + 311, + 145, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 311, + 145, + 320 + ], + "type": "text", + "content": "Congress." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 157, + 277, + 196, + 310 + ], + "blocks": [ + { + "bbox": [ + 157, + 277, + 196, + 310 + ], + "lines": [ + { + "bbox": [ + 157, + 277, + 196, + 310 + ], + "spans": [ + { + "bbox": [ + 157, + 277, + 196, + 310 + ], + "type": "image", + "image_path": "7d262444857fb4095041897039a483f3d5b79b79880641c58e1ac600c0579c40.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 311, + 179, + 317 + ], + "lines": [ + { + "bbox": [ + 176, + 311, + 179, + 317 + ], + "spans": [ + { + "bbox": [ + 176, + 311, + 179, + 317 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 261, + 248, + 269 + ], + "lines": [ + { + "bbox": [ + 225, + 261, + 248, + 269 + ], + "spans": [ + { + "bbox": [ + 225, + 261, + 248, + 269 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 288, + 268, + 296 + ], + "lines": [ + { + "bbox": [ + 202, + 288, + 268, + 296 + ], + "spans": [ + { + "bbox": [ + 202, + 288, + 268, + 296 + ], + "type": "text", + "content": "LLaVA-1.5: Soulemes" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 296, + 251, + 304 + ], + "lines": [ + { + "bbox": [ + 202, + 296, + 251, + 304 + ], + "spans": [ + { + "bbox": [ + 202, + 296, + 251, + 304 + ], + "type": "text", + "content": "Ours: Soulomes." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 67, + 329, + 106, + 350 + ], + "blocks": [ + { + "bbox": [ + 67, + 329, + 106, + 350 + ], + "lines": [ + { + "bbox": [ + 67, + 329, + 106, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 329, + 106, + 350 + ], + "type": "image", + "image_path": "87dca514e52b07032f828d8a841a7aa95cd7a243e12d3507d127edeb0a8271b7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 351, + 96, + 358 + ], + "lines": [ + { + "bbox": [ + 77, + 351, + 96, + 358 + ], + "spans": [ + { + "bbox": [ + 77, + 351, + 96, + 358 + ], + "type": "inline_equation", + "content": "2 + 1 = 3" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 113, + 329, + 151, + 350 + ], + "blocks": [ + { + "bbox": [ + 113, + 329, + 151, + 350 + ], + "lines": [ + { + "bbox": [ + 113, + 329, + 151, + 350 + ], + "spans": [ + { + "bbox": [ + 113, + 329, + 151, + 350 + ], + "type": "image", + "image_path": "6ec856ed2fb5f9dc331ea38f43821e022bf7f3dab913a5612ec009ed6fd3d323.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 351, + 141, + 358 + ], + "lines": [ + { + "bbox": [ + 119, + 351, + 141, + 358 + ], + "spans": [ + { + "bbox": [ + 119, + 351, + 141, + 358 + ], + "type": "inline_equation", + "content": "5 + 6 = 11" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 157, + 329, + 196, + 350 + ], + "blocks": [ + { + "bbox": [ + 157, + 329, + 196, + 350 + ], + "lines": [ + { + "bbox": [ + 157, + 329, + 196, + 350 + ], + "spans": [ + { + "bbox": [ + 157, + 329, + 196, + 350 + ], + "type": "image", + "image_path": "bdd0dac968f9f032ec085ba05d87f02cd800c5777f64ebcfd29bbdc78d490066.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 351, + 179, + 357 + ], + "lines": [ + { + "bbox": [ + 176, + 351, + 179, + 357 + ], + "spans": [ + { + "bbox": [ + 176, + 351, + 179, + 357 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 204, + 331, + 264, + 339 + ], + "lines": [ + { + "bbox": [ + 204, + 331, + 264, + 339 + ], + "spans": [ + { + "bbox": [ + 204, + 331, + 264, + 339 + ], + "type": "text", + "content": "LLaVA-1.5: " + }, + { + "bbox": [ + 204, + 331, + 264, + 339 + ], + "type": "inline_equation", + "content": "3 \\times 6 = 18" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 204, + 339, + 246, + 347 + ], + "lines": [ + { + "bbox": [ + 204, + 339, + 246, + 347 + ], + "spans": [ + { + "bbox": [ + 204, + 339, + 246, + 347 + ], + "type": "text", + "content": "Ours: " + }, + { + "bbox": [ + 204, + 339, + 246, + 347 + ], + "type": "inline_equation", + "content": "3 \\times 6 = 18" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 67, + 369, + 106, + 407 + ], + "blocks": [ + { + "bbox": [ + 67, + 369, + 106, + 407 + ], + "lines": [ + { + "bbox": [ + 67, + 369, + 106, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 369, + 106, + 407 + ], + "type": "image", + "image_path": "902749e88ea5c794423892338c8098798ccc69ba6d0fda5e1dff66866c4490ca.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 408, + 106, + 415 + ], + "lines": [ + { + "bbox": [ + 67, + 408, + 106, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 408, + 106, + 415 + ], + "type": "text", + "content": "Romanticism" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 421, + 287, + 486 + ], + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 486 + ], + "type": "text", + "content": "Figure 7. VILA has better in-context learning capability thanks to interleaved image text pretraining rather than single image-text pairs. We feed two image+text pairs and a third image as the context to prompt the VLM. LLaVA failed the first sample due to limited OCR capability, and failed the third examples by repeating the second sample semantic." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 113, + 369, + 151, + 407 + ], + "blocks": [ + { + "bbox": [ + 113, + 369, + 151, + 407 + ], + "lines": [ + { + "bbox": [ + 113, + 369, + 151, + 407 + ], + "spans": [ + { + "bbox": [ + 113, + 369, + 151, + 407 + ], + "type": "image", + "image_path": "787e50c20c377486f4d9a95d6d81301ab01030e131518bd207d6a8f1aa8690e1.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 408, + 146, + 415 + ], + "lines": [ + { + "bbox": [ + 115, + 408, + 146, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 146, + 415 + ], + "type": "text", + "content": "Surrealism" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 157, + 369, + 196, + 407 + ], + "blocks": [ + { + "bbox": [ + 157, + 369, + 196, + 407 + ], + "lines": [ + { + "bbox": [ + 157, + 369, + 196, + 407 + ], + "spans": [ + { + "bbox": [ + 157, + 369, + 196, + 407 + ], + "type": "image", + "image_path": "2f3f287ea3eb502a0a7a5d066a589f5b9420772267c589b26b15ded607e2126f.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 409, + 179, + 415 + ], + "lines": [ + { + "bbox": [ + 176, + 409, + 179, + 415 + ], + "spans": [ + { + "bbox": [ + 176, + 409, + 179, + 415 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 381, + 271, + 390 + ], + "lines": [ + { + "bbox": [ + 203, + 381, + 271, + 390 + ], + "spans": [ + { + "bbox": [ + 203, + 381, + 271, + 390 + ], + "type": "text", + "content": "LLaVA-1.5: Surrealism" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 391, + 264, + 398 + ], + "lines": [ + { + "bbox": [ + 203, + 391, + 264, + 398 + ], + "spans": [ + { + "bbox": [ + 203, + 391, + 264, + 398 + ], + "type": "text", + "content": "Ours: Impressionism" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 48, + 491, + 293, + 563 + ], + "blocks": [ + { + "bbox": [ + 48, + 491, + 293, + 563 + ], + "lines": [ + { + "bbox": [ + 48, + 491, + 293, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 293, + 563 + ], + "type": "table", + "html": "
ResolutionProjector#TokensEvaluation ↑
OKVQATextVQACOCO
224linear25649.9%41.6%116.0
336linear57649.7%49.8%117.7
336downsample14449.3%45.6%115.7
", + "image_path": "e6878f0c25cffb7131418dacdfb15013246fe72f5f14f4ff3ef6b8c59f9bb85d.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 571, + 288, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 288, + 637 + ], + "type": "text", + "content": "Table 7. Improving the image resolution from 224 to 336 can significantly improve TextVQA accuracy. The raw resolution matters more than #tokens; high-resolution with token downsampling works better than low-resolution. We report accuracy for OKVQA and TextVQA, and CIDEr for COCO. Note: the evaluation protocol is different from Table 5 and can only be compared within the table." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 47, + 653, + 150, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 150, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 150, + 665 + ], + "type": "text", + "content": "4.4. Other Learnings." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "Image resolution matters, not #tokens. We chose an image resolution of " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "336^{2}" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " since it provides more fine-grained details compared to " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": ", leading to improved accuracy on tasks like TextVQA [50]. As shown in Table 7, increasing" + } + ] + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 323, + 273, + 394, + 328 + ], + "blocks": [ + { + "bbox": [ + 59, + 235, + 533, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 235, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 235, + 533, + 247 + ], + "type": "text", + "content": "Figure 6. Our model VILA can reason over multiple images thanks to the pre-training process. The samples are taken from [5, 60]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 260, + 449, + 269 + ], + "lines": [ + { + "bbox": [ + 322, + 260, + 449, + 269 + ], + "spans": [ + { + "bbox": [ + 322, + 260, + 449, + 269 + ], + "type": "text", + "content": "Example: Visual Chain-of-Though (CoT)" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 273, + 394, + 328 + ], + "lines": [ + { + "bbox": [ + 323, + 273, + 394, + 328 + ], + "spans": [ + { + "bbox": [ + 323, + 273, + 394, + 328 + ], + "type": "image", + "image_path": "762bac7201380454e14b7bd27cd958bc992333275a2b3c06429e5687034e7e51.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 398, + 273, + 481, + 327 + ], + "blocks": [ + { + "bbox": [ + 398, + 273, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 398, + 273, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 398, + 273, + 481, + 327 + ], + "type": "image", + "image_path": "9d1812cc2c8be4f7f591e77bd6d6114593dd177c5afa0425b097f03cfb70cc82.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 319, + 328, + 525, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 328, + 525, + 345 + ], + "spans": [ + { + "bbox": [ + 319, + 328, + 525, + 345 + ], + "type": "text", + "content": "Prompt: Photo: \\nMenu: \\nHow much should I pay for all the beer on the table according to the price on the menu?" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 319, + 346, + 527, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 346, + 527, + 356 + ], + "spans": [ + { + "bbox": [ + 319, + 346, + 527, + 356 + ], + "type": "text", + "content": " Answer: According to the menu, the price for the beer on the table is $6." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 319, + 361, + 418, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 361, + 418, + 370 + ], + "spans": [ + { + "bbox": [ + 319, + 361, + 418, + 370 + ], + "type": "text", + "content": "Prompt: ... + Think step-by-step." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 318, + 370, + 528, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 370, + 528, + 395 + ], + "spans": [ + { + "bbox": [ + 318, + 370, + 528, + 395 + ], + "type": "text", + "content": " Answer: The beer on the table is Magna, which is priced at $6.00. There are two bottles of Magna on the table. Therefore, the total cost of the beer should be $6.00 x 2 = $12.00." + } + ] + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 323, + 399, + 384, + 461 + ], + "blocks": [ + { + "bbox": [ + 323, + 399, + 384, + 461 + ], + "lines": [ + { + "bbox": [ + 323, + 399, + 384, + 461 + ], + "spans": [ + { + "bbox": [ + 323, + 399, + 384, + 461 + ], + "type": "image", + "image_path": "00f75f8675192043637d340b0f48ae3523938c2fc46ceffa077137af0ed44139.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 513, + 545, + 558 + ], + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 558 + ], + "type": "text", + "content": "Figure 8. Our model is able to perform chain-of-thought reasoning given visual inputs. It is able to generate the correct answer when adding \"Think step-by-step\" to the prompt. Zoom in for a better view of the image details. Samples from [19, 60]." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "bbox": [ + 387, + 414, + 524, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 414, + 524, + 439 + ], + "spans": [ + { + "bbox": [ + 387, + 414, + 524, + 439 + ], + "type": "text", + "content": "Prompt: \\nI am getting just two custom pizza for me and my friend. How much should I pay in total according to the image?" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 387, + 440, + 518, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 440, + 518, + 457 + ], + "spans": [ + { + "bbox": [ + 387, + 440, + 518, + 457 + ], + "type": "text", + "content": "Answer: According to the image, you should pay $11.99 for two custom pizzas." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 319, + 463, + 432, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 463, + 432, + 472 + ], + "spans": [ + { + "bbox": [ + 319, + 463, + 432, + 472 + ], + "type": "text", + "content": "Prompt: ... + Let's think step-by-step." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 318, + 473, + 523, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 473, + 523, + 506 + ], + "spans": [ + { + "bbox": [ + 318, + 473, + 523, + 506 + ], + "type": "text", + "content": " Answer: The image shows a chalkboard menu with a special for today, which is to create your own pizza. The price for this special is $9.99. Since you are getting two custom pizzas, you should pay $9.99 x 2 = $19.98 for the two pizzas. So, the answer is $19.98." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": "the resolution from 224 to 336 can improve the TextVQA accuracy from " + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "inline_equation", + "content": "41.6\\%" + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": ". However, a higher resolution leads to more tokens per image (" + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "inline_equation", + "content": "336^2" + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": " corresponds to 576 tokens/image) and a higher computational cost. It also limits the number of demonstrations for in-context learning." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": "Luckily, we find that the raw resolution matters more than the #visual tokens/image. We can use different projector designs to compress the visual tokens. Here we try a \"downsample\" projector, which simply concatenates every " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " tokens into a single one and use a linear layer to fuse the information. It reduces the #tokens to 144 under the 336 resolution, that is even smaller than the 224+linear" + } + ] + } + ], + "index": 59 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26695" + } + ] + } + ], + "index": 60 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 264, + 128 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 264, + 128 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 264, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 264, + 128 + ], + "type": "table", + "html": "
#ParamVLM acc. (avg)
0-shot4-shot
Visual Expert [58]1.9×67.0%64.8%
Fine-tune1×71.0%72.1%
", + "image_path": "95c5706fb6a76056bd245b004e1bef0fa09c6a23032ea8cae2c912e41d8e38fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 131, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 131, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 131, + 287, + 175 + ], + "type": "text", + "content": "Table 8. Directly fine-tuning the LLM during pre-training leads to better VLM accuracy and in-context learning capabilities. It also enjoys a smaller model size. Both settings are pre-trained on the MMC4-core dataset [69]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "content": "setup. Nonetheless, the TextVQA accuracy is higher (" + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\sim 46\\%" + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "inline_equation", + "content": "41.6\\%" + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "content": "), despite still " + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "content": " worse compared to " + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "inline_equation", + "content": "336+" + }, + { + "bbox": [ + 46, + 182, + 287, + 243 + ], + "type": "text", + "content": " linear setup, showing a large redundancy in the image tokens. The gap on other datasets such as OKVQA and COCO is smaller since they usually require higher-level semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "type": "text", + "content": "In our main results, we did not use any token compression methods to provide the best accuracy despite this encouraging observation, and leave it to future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 292, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 289, + 460 + ], + "type": "text", + "content": "Comparison to frozen LLMs with visual experts. Another interesting method for retaining the text capabilities of LLMs during the pre-training is to freeze the base LLM and add an extra visual expert to process the visual tokens [58]. The definition of expert is similar to MoE frameworks, but with a manual routing mechanism according to token types. Since the base LLM is frozen, the model fully retains the original functionality for text-only inputs during pre-training. However, we find that directly fine-tuning the LLM during visual language pre-training still leads to a better VLM accuracy and in-context learning capability (Table 8). Adding an extra visual expert also leads to near " + }, + { + "bbox": [ + 46, + 292, + 289, + 460 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 292, + 289, + 460 + ], + "type": "text", + "content": " model size increase, which is not friendly for edge deployment. Therefore, we chose to directly fine-tune the base LLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 521 + ], + "type": "text", + "content": "Comparison to PEFT/LoRA. In addition to visual experts, we also provide extra results when performing LoRA tuning with rank 64 (7B model) in Table 9. Fine-tuning LLM outperforms LoRA tuning by a large margin." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 531, + 287, + 575 + ], + "blocks": [ + { + "bbox": [ + 48, + 531, + 287, + 575 + ], + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 575 + ], + "type": "table", + "html": "
VQA v2GQAVisWizVQATMMBLLaVA W
LoRA (r=64)69.454.348.450.060.351.2
Fine-tune (ours)79.962.357.864.468.969.7
", + "image_path": "a9204013aef3140f470654b2f5d6f134e06cb5a5987141d79252a5103e948759.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 583, + 284, + 594 + ], + "lines": [ + { + "bbox": [ + 47, + 583, + 284, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 284, + 594 + ], + "type": "text", + "content": "Table 9. Fine-tuning LLM consistently outperforms LoRA tuning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Reformatting the interleaved structure. For additional insights we also reformatted the MMC4 dataset to be " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\text{im1} \\rangle < \\text{im2} \\rangle < \\text{txt1} \\rangle < \\text{txt2} \\rangle" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " instead of the " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\text{im1} \\rangle < \\text{txt1} \\rangle < \\text{im2} \\rangle < \\text{txt2} \\rangle" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and evaluated the model under the setting in Table 1. We observed that the reformatted MMC4 degrades the average 0-shot accuracy (on the 4 benchmarks) by " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", and degrades the average 4-shot accuracy by " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ". The disorder breaks in-context learning capability, showing the importance of interleaved data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 392, + 83 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 84, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 547, + 240 + ], + "type": "text", + "content": "Large language models (LLMs). LLMs based on Transformers [57] have fundamentally changed the language processing field. They are achieving increasing capabilities by scaling up the model size and the pre-training corpus [1, 9, 15, 18, 20, 22, 27, 47, 51]. It is believed that most the capability of the LLM is obtained from the large-scale pre-training process, which are later unlocked through instruction tuning [16, 43, 44]. There is a growing effort from the open-source community to build a strong base LLM [55, 56, 65], the conversational variants [14, 54] and the parameter-efficient finetuned versions of large LLMs [40, 64]. In this work, we start with the base Llama-2 model [56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 547, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 414 + ], + "type": "text", + "content": "Visual language models (VLMs). VLMs are LLMs augmented with visual inputs to provide a unified interface for visual language tasks. There are two main designs for VLMs: 1. cross-attention based, where the LLM is frozen while the visual information is fused into intermediate embeddings with a cross-attention mechanism [5, 6]; 2. auto-regressive based, where the visual input is tokenized and fed to the LLM alongside text tokens [2, 4, 7, 13, 19, 33, 37, 61, 68]. The latter is a natural extension by treating visual inputs as a foreign language. VLMs are also instruction-tuned so that they can better follow human instructions or perform conversations [17, 37, 52]. In this work, we study the pre-training process of the auto-regressive VLMs due to their flexibility when handling multi-modal inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 547, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 522 + ], + "type": "text", + "content": "Following text-only LLMs, people also study different training recipes for VLMs. Some work freezes the LLM and train auxiliary components [5, 32, 33, 58], others fine-tune the LLM to enable visual capabilities [13, 19, 66]. There is also usage of different data corpora, including image-text pairs [13, 19, 32, 37], interleaved datasets [6], video-text pairs [41], visual-grounded annotations [36, 45], etc. In this work, we provide a holistic ablation of different design choices for the pre-training stage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 531, + 379, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 379, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 379, + 544 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 551, + 547, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 547, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 547, + 659 + ], + "type": "text", + "content": "This paper has explored effective pretraining design options to augment LLMs towards vision tasks. Leveraging full strength of LLM learning, interleaved-nature of image-text data, and careful text data re-blending, VILA has surpassed state-of-the-art methods for vision tasks while preserving text-only capabilities. VILA has also depicted strong reasoning capability for multi-image analysis, in-context learning and zero/few-shot tasks. We hope our paper can help spur further research on VLM pretraining." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgements. We thank Bryan Catanzaro for fruitful discussions. We appreciate the help from Zhuolin Yang, Guilin Liu, Lukas Voegtle, Philipp Fischer, Karan Sapra and Timo Roman on dataset preparation and feedback." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26696" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "type": "text", + "content": "[1] GPT-4 technical report. Technical report, OpenAI, 2023. https://arxiv.org/abs/2303.08774." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 136 + ], + "type": "text", + "content": "[2] Fuyu-8B: A multimodal architecture for AI agents. https://www.adept.ai/blog/fuyu-8b, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 289, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 289, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 289, + 158 + ], + "type": "text", + "content": "[3] Yi-34B large language model. https://huggingface.co/01-ai/Yi-34B, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "text", + "content": "[4] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 192, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 257 + ], + "type": "text", + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 259, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 288, + 314 + ], + "type": "text", + "content": "[6] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 358 + ], + "type": "text", + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. Technical report, Alibaba Group, 2023. https://arxiv.org/abs/2303.08774." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 359, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 359, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 359, + 288, + 403 + ], + "type": "text", + "content": "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 524 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "text", + "content": "[10] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/ coyo-dataset, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "text", + "content": "[11] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "text", + "content": "[12] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[13] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "text", + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "text", + "content": " chatgpt quality, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 239 + ], + "type": "text", + "content": "[16] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. arXiv preprint arXiv:2210.11416, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "text", + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 342 + ], + "type": "text", + "content": "[18] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 397 + ], + "type": "text", + "content": "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 464 + ], + "type": "text", + "content": "[20] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "type": "text", + "content": "[21] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proc. of NAACL, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 547, + 555 + ], + "type": "text", + "content": "[22] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 547, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 547, + 611 + ], + "type": "text", + "content": "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 669 + ], + "type": "text", + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 714 + ], + "type": "text", + "content": "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26697" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 3608-3617, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "text", + "content": "[26] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. CoRR, abs/2009.03300, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[27] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[28] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "text", + "content": "[29] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "text", + "content": "[30] Siddharth Karamcheti, Laurel Orr, Jason Bolton, Tianyi Zhang, Karan Goel, Avanika Narayan, Rishi Bommasani, Deepak Narayanan, Tatsunori Hashimoto, Dan Jurafsky, et al. Mistral—a journey towards reproducible language model training, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "text", + "content": "[31] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "text", + "content": "[32] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[34] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[39] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "type": "text", + "content": "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "type": "text", + "content": "[40] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "type": "text", + "content": "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 195, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 546, + 249 + ], + "type": "text", + "content": "[42] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 251, + 547, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 281 + ], + "type": "text", + "content": "[43] OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt, 2023. Accessed: 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 283, + 547, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 547, + 347 + ], + "type": "text", + "content": "[44] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 350, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 546, + 392 + ], + "type": "text", + "content": "[45] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 394, + 546, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 546, + 458 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "text", + "content": "[47] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 546, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 546, + 568 + ], + "type": "text", + "content": "[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 546, + 635 + ], + "type": "text", + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "type": "text", + "content": "[50] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 546, + 713 + ], + "type": "text", + "content": "[51] Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26698" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 117 + ], + "type": "text", + "content": "Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, et al. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model. arXiv preprint arXiv:2201.11990, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[52] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[53] Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[54] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "text", + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "type": "text", + "content": "[56] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "text", + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[60] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[61] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[62] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[63] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 419 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[64] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "type": "text", + "content": "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "type": "text", + "content": "[66] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "type": "text", + "content": "[67] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "type": "text", + "content": "[68] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 364, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 547, + 419 + ], + "type": "text", + "content": "[69] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26699" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_content_list.json b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48685673ce9e77157aba1ad9ed40f87b0b7502cc --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "VINECS: Video-based Neural Character Skinning", + "text_level": 1, + "bbox": [ + 230, + 130, + 738, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhouyingcheng Liao $^{1,2}$", + "bbox": [ + 86, + 178, + 274, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vladislav Golyanik1", + "bbox": [ + 315, + 180, + 475, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Habermann1,3", + "bbox": [ + 519, + 179, + 678, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Theobalt1,3", + "bbox": [ + 718, + 179, + 887, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Max Planck Institute for Informatics, Saarland Informatics Campus 2The University of Hong Kong", + "bbox": [ + 189, + 200, + 810, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Saarbrücken Research Center for Visual Computing, Interaction and AI", + "bbox": [ + 272, + 218, + 702, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rigging and skinning clothed human avatars is a challenging task and traditionally requires a lot of manual work and expertise. Recent methods addressing it either generalize across different characters or focus on capturing the dynamics of a single character observed under different pose configurations. However, the former methods typically predict solely static skinning weights, which perform poorly for highly articulated poses, and the latter ones either require dense 3D character scans in different poses or cannot generate an explicit mesh with vertex correspondence over time. To address these challenges, we propose a fully automated approach for creating a fully rigged character with pose-dependent skinning weights, which can be solely learned from multi-view video. Therefore, we first acquire a rigged template, which is then statically skinned. Next, a coordinate-based MLP learns a skinning weights field parameterized over the position in a canonical pose space and the respective pose. Moreover, we introduce our pose- and view-dependent appearance field allowing us to differentiably render and supervise the posed mesh using multi-view imagery. We show that our approach outperforms state-of-the-art while not relying on dense 4D scans. More details can be found on our project page1.", + "bbox": [ + 75, + 300, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 660, + 209, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating 3D characters is a long-standing and challenging problem in computer graphics and vision. Traditionally, rigging, the process of positioning a sparse skeletal structure inside a dense 3D mesh, and skinning, the process of associating vertices with the skeleton, require large amounts of manual effort by experienced artists. This is not only expensive and time-consuming but also inherently difficult since one set of skinning weights is typically not ideal for all sorts of articulated character poses.", + "bbox": [ + 75, + 685, + 468, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are also methods for automated rigging and skinning. Some recent works [33, 34, 41, 59] propose a learning-based solution for obtaining skinning and rigging", + "bbox": [ + 75, + 821, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6086a2dbb211e552a96dea27ad4f1f682c93d54696c16f3a7fdbb66d2cf5e476.jpg", + "image_caption": [ + "Figure 1. We propose the first end-to-end trainable method for generating a dense and rigged 3D character mesh including learned pose-dependent skinning weights solely from multiview videos. The results demonstrate that our model can generate visually-plausible geometry without requiring manual user input." + ], + "image_footnote": [], + "bbox": [ + 532, + 267, + 861, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "given a single 3D character mesh. However, they do not consider pose-dependent skinning correctives and lead to the typical appearance artifacts. Therefore, other techniques focus on pose-dependent skinning formulations accounting for pose-dependent variations and deformations [3, 30, 40, 52, 57]. Noteworthy, none of these methods tried to learn pose-dependent skinning weight correctives purely from $2D$ image data. Instead, they typically assume a 3D mesh is given, and most of them solely work on a fixed mesh resolution, which cannot be easily adjusted.", + "bbox": [ + 496, + 532, + 893, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose VINECS, i.e., Video-based NEural Character Skinning. Given solely a multi-view video of a human actor, our method creates a fully rigged, skinned, and textured 3D template (including hands) along with pose-dependent skinning weights while no manual editing is required. Moreover, VINECS enables multi-resolution character skinning since skinning weights can be sampled over a continuous 3D canonical space, which is not limited to any specific mesh resolution.", + "bbox": [ + 496, + 686, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For each character, we select one canonical pose frame, for which we extract a dense and textured 3D template mesh using implicit surface reconstruction techniques [55]. Given the template and the respective pose obtained from an off-the-shelf multi-view markerless capture system [53],", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://people.mpi-inf.mpg.de/~mhaberma/projects/2023-Vinecs", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1377", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "we leverage an auto-rigging and skinning method, Pinocchio [7], for obtaining an initial set of skinning weights. Note that Pinocchio produces only a static set of skinning weights, which leads to the typical geometry artifacts when poses deviate too much from the canonical pose. Therefore, we propose a coordinate-based multi-layer perceptron (MLP)—which takes as input a 3D point on the surface of the template in canonical pose and the skeletal pose—and predicts the pose-dependent skinning weights at that point. Since we want to learn the skinning weights solely from multi-view imagery, we design dedicated losses consisting of a silhouette loss and a rendering loss. Specifically, we find that obtaining a well-behaved convergence is challenging only using a static texture for the rendering loss. Thus, we adopt the rendering formulation with an albedo network and a shadow network, which predict pose- and view-dependent appearance of our 3D character resulting in an improved convergence and skinning weights prediction. In summary, our technical contributions are as follows:", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An end-to-end trainable method for animation-ready explicit character creation, involving static template generation, rigging, and pose-dependent skinning solely using 2D multi-view videos (as few as 10 views).", + "- A coordinate-based and pose-dependent skinning formulation, which enables multi-resolution skinning.", + "- A dedicated character appearance formulation and differentiable rendering enabling weakly supervised learning based on multi-view videos." + ], + "bbox": [ + 76, + 378, + 468, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare VINECS to the state-of-the-art and our results confirm that our method is a clear step toward accurate and automated creation of animatable 3D characters.", + "bbox": [ + 76, + 516, + 468, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Potential Impact. The focus of this work is not to model deformable geometry [13, 14, 20, 32] or photorealistic appearance [15, 16, 25, 35, 44, 47, 48] of humans but on learning 3D characters with pose-dependent skinning from 2D observations only. We believe our method can greatly benefit many downstream tasks as many existing works leverage naked body models [36] for space canonicalization and feature assignment, fundamentally limiting them to clothing types of the same topology as the human body. As our rigging and skinning are accurate, easy-to-use, and robust to any type of clothing, our approach can stimulate research on the modeling of geometry and appearance of loose apparel.", + "bbox": [ + 75, + 561, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 753, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Static Skinning. Linear Blend Skinning (LBS) [28, 38] linearly blends rigid transformations of each bone to obtain a posed geometry. However, this formulation can lead to a collapse of geometry near the skeleton joints, also called candy-wappers effect. To overcome this, several follow-up works have been proposed using spherical representations [22], sweep surfaces [17], curve-based skeleton parameterizations [61], Dual Quaternions [23], or by optimiz", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing the centers of rotation [27]. Other works focused more on computing the skinning weights itself by modeling the weight distribution as a heat diffusion process [7], as illumination [58], or by voxelizing the mesh to increase robustness [9, 10]. Later on, deformation editing was also modeled using a Laplacian energy formulation [18], and user-defined splines on the mesh enable intuitive editing of skinning-based deformation behaviors [6]. They all only assume a single geometry and skeleton are given.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, data-driven methods assume multiple instances of the same character under different poses or a large collection of rigged and skinned characters. Earlier works [19, 26] requires meshes under different poses and then automatically detect skeleton joints and compute the skinning weights. Recently, NeuroSkinning [34] was proposed as the first learning-based method predicting a static set of skinning weights given the mesh and a corresponding skeleton. RigNet [59] instead jointly predicts the skeleton as well as a set of skinning weights solely given a 3D mesh as input. Therefore, they collected a large dataset of 3D characters including artist-designed skeletons and skinning weights. SkinningNet [41] proposed a two-stream graph convolutional network architecture for improved skinning weight prediction given a mesh and the skeleton. Yang et al. [60] proposed to predict the 3D shape together with the rigging from a single image. Liao et al. [33] introduced a learning-based method which jointly learns to predict skinning weights and transfer poses.", + "bbox": [ + 496, + 229, + 893, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What all the above methods have in common is that they do not explicitly account for dynamic or pose-dependent effects, such as muscle bulging or coarse cloth deformations. In contrast, our proposed method learns pose-dependent skinning weights, which alleviate typical artifacts that arise when using a static set of skinning weights. Moreover, it can also represent coarse deformations induced by the skeletal pose, such as coarse cloth deformations.", + "bbox": [ + 496, + 518, + 893, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Methods. More closely related to our method are approaches that account for dynamic or pose-dependent effects, which are discussed next. There are works [11] that allow an artist to add pose-dependent effects by painting onto specialized textures or by adding so-called helper bones [42]. Alternatively, Kavan et al. [21] proposed to find the optimal skinning weights by minimizing an elastic energy and adding joint-based deformers. Interestingly, those methods do not assume any dynamic observations of the character while still modeling dynamic effects. However, they either require manual editing [11, 42] or can mostly model elasticity-based dynamics [21].", + "bbox": [ + 496, + 641, + 893, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to the static methods, several dynamic methods assume a (large) dataset of character meshes is given. Earlier works [3, 40, 46, 57] account for dynamic effects by capturing a set of range scans depicting the subject in different poses. Given a new pose, a set of pose-dependent blend", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1378", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "weights is estimated, which linearly combines the set of example poses in the dataset or alternatively uses linear and radial basis functions [52]. However, since the complexity of computation typically linearly scales with the number of example poses in the dataset, those methods cannot scale up to a very large number of examples (e.g., 20000 pose frames), which is in stark contrast to our approach that can deal with such an amount of data. There are also parametric naked human body models [4, 5, 36] learned by applying PCA over a large set of human scans, which can account for different shape- and pose-dependent deformations. In contrast, we are building a highly-detailed rigged and skinned character that can wear (loose) clothing in a completely automated fashion. Li et al. [30] predict static skinning weights and pose-dependent neural blend shapes. They assume a dataset of dense meshes is provided, whereas we purely learn pose-dependent skinning weights from videos. SCANimate [49] also requires dense point clouds, and learns a static set of skinning weights, while pose-dependent shapes are predicted by a pose-aware implicit field. SNARF [8] represents the character as a pose-dependent neural implicit surface in the canonical pose space while the learned skinning weights are pose-agnostic. Again, they require point cloud data while our method can be supervised directly on the video. In summary, none of the above methods that consider pose-dependent effects investigated learning solely from 2D imagery. Thus, we are the first method demonstrating that dynamic effects in the form of pose-dependent skinning weights can be learned only from videos.", + "bbox": [ + 76, + 90, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, some works were proposed to skin and rig implicit surface representations learn an implicit human avatar. ARAH [56] learns an implicit human model using neural rendering. However, as it relies on a SMPL template and an inverse root finding to convert from view space to canonical space, it easily fails when the garment deviates much from the body. Importantly, it also does not have the vertex correspondence over time. TAVA [31] avoids these problems by getting rid of a template, which, however, leads to poor geometry recovery mainly due to their density-based geometry representation. While these methods can reconstruct an implicit avatar, they do not learn a high-quality skinning weight, which is important to traditional graphics pipeline.", + "bbox": [ + 76, + 529, + 472, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 739, + 169, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given synchronized and calibrated multi-view RGB videos of a human character in motion, our goal is to learn an animatable 3D human model with dynamic skinning, without any manual process, such as meshing, rigging, and skinning. To this end, we first extract the static geometry of the actor from one canonical pose frame (typically depicting the actor in a T-pose) of the recording and automatically compute initial skinning weights (Sec. 3.1). However, the initial skinning weights are calculated only based on", + "bbox": [ + 76, + 763, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the static geometry of the human character, and, as studied in earlier works, one set of weights might be ideal for one pose while it can lead to strong artifacts in other poses. Thus, we propose a skinning network, called SkinNet, to predict the dynamic skinning weights as a function of the skeletal pose and 3D coordinates in the canonical space (Sec. 3.2). Given the skeletal pose, the static geometry, and the pose-dependent skinning weights, we leverage Linear Blend Skinning (LBS) to obtain the posed geometry. We then use a differentiable renderer to weakly supervise the skinning weights solely based on multi-view imagery. Thereafter, we propose a novel appearance field (Sec. 3.3), which is composed of an albedo component and a pose-and view-dependent shadow/shading component. In addition, we propose several regularization losses ensuring that the skinning weights are also robust to out-of-distribution poses and that the final animated geometry looks visually plausible (Sec. 3.4). An overview is shown in Fig. 2.", + "bbox": [ + 496, + 90, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Preparation. For a human subject, we record a multi-view video of the performance using $C$ calibrated and synchronized cameras with a total length of $F$ frames. For all frames $\\mathcal{I}_{c,f}$ , we calculate the foreground segmentation using color keying or background matting [51] and compute respective distance transform images $\\mathcal{D}_{c,f}$ . Here, $c$ and $f$ refer to the respective camera and frame index. We capture the size of the human skeleton and the motion using markerless motion capture [53]. The skeleton motion is parameterized by a root rotation $\\alpha_{f} \\in \\mathbb{R}^{3}$ , a root translation $\\mathbf{t}_f \\in \\mathbb{R}^3$ , and the joint angles $\\rho_f \\in \\mathbb{R}^D$ . The full pose vector is defined as $\\theta_f = [\\alpha_f, \\mathbf{t}_f, \\rho_f]^T \\in \\mathbb{R}^{3 + 3 + D}$ .", + "bbox": [ + 496, + 364, + 892, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. To represent a variable in the canonical space, we use a bar over the letter, i.e., $\\bar{\\mathbf{x}}$ . For any point in canonical and global space, we use $\\bar{\\mathbf{x}}$ and $\\mathbf{x}$ , respectively, and mesh vertices are denoted as $\\bar{\\mathbf{v}}_i$ and $\\mathbf{v}_i$ where $i$ is the $i$ th vertex. Without loss of generality, we assume the frame $f$ is fixed in the following (if not explicitly mentioned otherwise) and omit the subscript for better readability.", + "bbox": [ + 496, + 547, + 893, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Canonical Character Model", + "text_level": 1, + "bbox": [ + 500, + 665, + 751, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the first step, we acquire a canonical model of the actor. Recently, implicit surface and coordinate-based representations [45] have gained considerable attention due to their flexibility and comparably simple integration into learning frameworks, and they were also leveraged in the context of pose-dependent human deformation modeling [8]. Moreover, we compared the reconstruction accuracy on the first frame $f = 0$ between classical multi-view stereo and implicit reconstruction method called NeuS [55] in the supplemental material. Interestingly, NeuS has a much higher level of detail and comparably less noise on the surface. Thus, we use NeuS [55] to reconstruct the mesh in the first frame, also called the canonical frame/space.", + "bbox": [ + 496, + 688, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit vs. Explicit Canonical Model. In stark con", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1379", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca837b81452c319e118eb6c32059b87417436f8fb1c276cb015464647221369d.jpg", + "image_caption": [ + "Figure 2. Method overview. Our method, VINECS, is a fully automated solution to template generation, rigging, and pose-dependent skinning solely using multi-view videos. We first recover the skeletal poses using markerless motion capture [53] and a static template mesh using implicit surface reconstruction [55]. An initial skinning is obtained by a heat diffusion process [7]. Since one fixed set of skinning weights can lead to artifacts when significantly changing the character pose, we learn pose-dependent skinning weights using a coordinate-based MLP. An auxiliary appearance field is applied to better supervise the learning guide the training of pose-dependent skinning. Note that we do not intend to learn a detailed appearance and only the components in the blue box are used during inference." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 857, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trast to Chen et al. [8], we run Marching Cubes [37] to convert the implicit surface into an explicit mesh and obtain per-vertex colors from NeuS by setting the viewing direction to the normal direction of the mesh. The reason for choosing an explicit mesh in canonical space is three-fold: 1.) It avoids running the rather slow Marching Cubes algorithm for every potentially posed mesh. 2.) We avoid the backward skinning [8], which boils down to a correspondence search that has to be solved using iterative optimization. 3.) We can maintain correspondence on the surface across poses, which is especially useful when texturing and other editing have to be applied. Thus, we aim for a hybrid strategy for template acquisition, i.e., using the state-of-the-art method for implicit surface reconstruction and then converting it to an explicit mesh.", + "bbox": [ + 75, + 388, + 472, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meshing and Initial Skinning. During training, we downsample the template mesh to $N$ vertices where $\\bar{\\mathbf{v}}_i\\in \\mathbb{R}^3$ denotes the ith vertex. Given the template mesh and the corresponding posed skeleton, i.e. $\\theta_{f}$ , we utilize Pinocchio [7] to obtain an initial set of skinning weights $\\mathbf{w}_{\\mathrm{init},i}\\in \\mathbb{R}^{J}$ for each vertex $i$ (see Fig. 4). Here, $J$ denotes the number of skinning joints. Given these weights and a skeletal pose $\\theta$ , we can transform the vertex $\\bar{\\mathbf{v}}_i$ from the canonical space to the posed space using LBS [28]:", + "bbox": [ + 75, + 614, + 472, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} = L B S \\left(\\bar {\\mathbf {v}} _ {i}, \\mathbf {w} _ {\\text {i n i t}, i}, \\theta\\right): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {J} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {3}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 755, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh Parsing. Since different body parts and materials (skin vs. clothing) have different deformation behaviors, we compute per-vertex human parsing labels. We render the 3D model into all camera views and apply a 2D human parsing approach [29] on each view. We then perform max-voting to obtain the per-vertex human parsing labels. Note that so far, we obtained a segmented, rigged, and statically skinned character model without any manual user input.", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pose-dependent Skinning Field", + "text_level": 1, + "bbox": [ + 498, + 387, + 774, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So far, the initial skinning weights obtained in Sec. 3.1 have two main limitations: 1.) They are solely computed on a single static pose leading to artifacts such as candy wrappers and mesh distortions if the new pose significantly differs from the canonical one. 2.) The set of weights is computed per-vertex. Thus, the resolution of the mesh has to be fixed beforehand and any change in the mesh requires a new computation of skinning weights or some complicated weight transfer from one mesh to the other.", + "bbox": [ + 496, + 410, + 890, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To resolve this, we propose SkinNet $f_{\\omega}$ , which is a coordinate-based MLP where $\\omega$ denotes the trainable network weights. Given any 3D point $\\bar{\\mathbf{x}}$ in the canonical space, $f_{\\omega}$ predicts its skinning weight $\\mathbf{w}$ conditioned on the normalized human pose $\\tilde{\\theta}$ , which can be formulated as:", + "bbox": [ + 496, + 547, + 893, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {\\tilde {\\theta}} = f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\boldsymbol {\\theta}}): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {J}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 633, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{x}}$ and $\\tilde{\\theta}$ are concatenated before being fed into the MLP. Here, the normalized pose $\\tilde{\\theta}$ is obtained by neglecting the global translation as well as the yaw angle, i.e., the rotation around the spine. The intuition behind is that the skinning weights do not depend on where in global 3D space the person performs a respective pose, and, similarly, the skinning weight prediction should be agnostic to which direction the person is facing. Thus, we mask those degrees of freedom before feeding the pose vector into the network. To obtain the transformed point $\\mathbf{x}$ in posed space, one can insert Eq. (2) into Eq. (1) resulting in:", + "bbox": [ + 496, + 662, + 893, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = L B S (\\bar {\\mathbf {x}}, f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\theta}), \\theta). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 840, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Revisiting the drawbacks of initial skinning discussed above, the SkinNet now supports pose-dependent skinning", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1380", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "weights such that the artifacts of static skinning can be reduced. Moreover, since SkinNet is a coordinate-based MLP, it supports arbitrary mesh resolution since each vertex is queried independently, which is in stark contrast to an architecture that outputs all skinning weights at once and where the final tensor shape is fixed to a specific mesh resolution.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Auxiliary Appearance Field", + "text_level": 1, + "bbox": [ + 76, + 191, + 326, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can now pose the geometry with our pose-dependent skinning weight formulation. However, since we want to supervise SkinNet solely on multi-view images, we also have to model the appearance of the actor. A naive choice would be to just leverage the static vertex colors acquired from NeuS. However, we found that this texture contains baked-in shadows and wrinkles, and cannot account for the pose- and view-dependent appearance changes harming the training of the SkinNet. We also refer to our ablation studies in Sec. 4.3. To address this issue, we propose an auxiliary appearance field, which consists of an albedo field predicting a pose- and view-agnostic color for a given point $\\bar{\\mathbf{x}}$ in the canonical space and a shadow field predicting the pose- and view-dependent shadow/shading properties of $\\bar{\\mathbf{x}}$ . Important to note here is that we are not interested in creating a highly photorealistic appearance model of the human, but we mainly use it as an auxiliary tool to better supervise the learning of the skinning weight network.", + "bbox": [ + 75, + 215, + 470, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More specifically, the albedo field, referred to as AlbedoNet, is an MLP that predicts a static albedo value $a_{\\omega'}(\\bar{\\mathbf{x}}) = \\mathbf{a} \\in \\mathbb{R}^3$ for a given 3D sample point in the canonical space. Here, $\\omega'$ are the trainable weights of the albedo network. Moreover, the shadow field is also parameterized by an MLP $r_{\\omega''}(\\bar{\\mathbf{x}}, \\theta, \\mathbf{n}, \\mathbf{d}) = s \\in \\mathbb{R}^+$ , called ShadowNet, which predicts a scalar multiplier accounting for color changes due to shadows and shading. Note that we assume that such effects happen uniformly across all color channels and, thus, we are only predicting a single scalar instead of a scaling factor per color channel. In addition to taking the point in canonical space, the shadow field also takes the skeletal pose $\\theta$ as input as well as the surface normal $\\mathbf{n}$ of $\\mathbf{x}$ in global space and the viewing direction $\\mathbf{d}$ , which allows to potentially model pose- and view-dependent lighting effects such as shadows and shading. $\\omega''$ denotes the respective network weights. Details about the architectures are provided in the supplemental document. The final color $\\mathbf{c} \\in \\mathbb{R}^3$ for point $\\mathbf{x}$ is computed as", + "bbox": [ + 75, + 488, + 468, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = a _ {\\omega^ {\\prime}} (\\bar {\\mathbf {x}}) r _ {\\omega^ {\\prime \\prime}} (\\bar {\\mathbf {x}}, \\theta , \\mathbf {n}, \\mathbf {d}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 787, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-view Video-based Supervision", + "text_level": 1, + "bbox": [ + 76, + 816, + 393, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we describe our dedicated supervision strategy that solely requires multi-view imagery. We first introduce the individual loss terms and then describe our training strategy. Importantly, we can obtain a posed template mesh", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{V} \\in \\mathbb{R}^{N \\times 3}$ and the pose- and view-dependent per-vertex colors $\\mathbf{C}_c \\in \\mathbb{R}^{N \\times 3}$ by evaluating Eq. (3) and Eq. (4) for all vertices $\\bar{\\mathbf{v}}_i$ and views $c$ . Moreover, the following losses that are applied to the deformed geometry and its appearance can directly backpropagate into the respective networks since our formulation is fully differentiable.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Silhouette Loss. This loss [15] ensures that the projected posed geometry projects onto the foreground masks for all views (see supp. material). Since our pose-dependent skinning field Eq. (3) is differentiable with respect to the SkinNet weights $\\omega$ , the silhouette loss can provide supervision and learn pose-dependent skinning by matching the posed model silhouette and the image silhouettes.", + "bbox": [ + 496, + 181, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering Loss. Although the silhouette loss ensures that the posed geometry matches the multi-view silhouettes, it cannot supervise drifts on the visual hull. Supervision directly from the RGB images can help with resolving this issue since appearance features can constrain drifts in the image plane. Thus, we introduce the rendering loss", + "bbox": [ + 496, + 287, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d}} = \\sum_ {c = 1} ^ {C} \\| \\Pi_ {c} (\\mathbf {V}, \\mathbf {C} _ {c}) - \\mathcal {I} _ {c} \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 388, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which leverages a differentiable renderer $\\Pi_c$ based on the one of Habermann et al. [15] to render the posed and colored model into all camera views $c$ that is then compared to the ground truth image $\\mathcal{I}_c$ . Importantly, this loss can supervise both the pose-dependent skinning field as well as the pose- and view-dependent appearance field.", + "bbox": [ + 496, + 436, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Laplacian Loss. Since the data terms alone can still lead to degenerate results, we also introduce some regularization on the geometry and the skinning weights. First, we regularize the posed geometry with a Laplacian loss", + "bbox": [ + 496, + 527, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l a p}} = \\sum_ {i = 1} ^ {N} \\| \\mathbf {v} _ {i} - \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {k \\in \\mathcal {N} _ {i}} \\mathbf {v} _ {k} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 598, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_i$ is the indices of one-ring neighborhood of vertex $i$ . This regularizer ensures locally smooth geometry.", + "bbox": [ + 496, + 647, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Skinning Regularization.** We found that SkinNet can overfit the training poses and respective deformations by predicting non-local skinning weights, e.g., the foot joint has a non-zero skinning weight for vertices on the shoulder. This, however, harms generalization to novel poses at test time. To prevent this, we regularize the skinning weights by constraining their values based on the geodesic distance $d_{\\mathrm{geo}}$ between the vertex and the initial body part it belongs to. This can be formalized with the following loss", + "bbox": [ + 496, + 678, + 890, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s k i n}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\mathbf {w} _ {i, j} \\left(\\frac {\\min _ {k \\in \\mathscr {A} _ {j}} d _ {\\text {g e o}} (\\bar {\\mathbf {v}} _ {i} , \\bar {\\mathbf {v}} _ {k})}{d _ {\\text {g e o m a x}}}\\right) ^ {r}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 823, + 890, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathscr {A} _ {j} = \\left\\{k \\mid j = \\underset {t} {\\operatorname {a r g m a x}} \\mathbf {w} _ {\\text {i n i t}, k, t} \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 878, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1381", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_j$ is the set of vertices, which have the highest skinning weight with respect to joint $j$ . $r = 3$ is a hyperparameter which controls how penalization increases as the geodesic distance increases. $d_{\\mathrm{geomax}}$ is the maximum geodesic distance between any of the vertex pairs.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part-wise Regularization. In contrast to the clothed areas, the skin part has rather static skinning. Thus, we use the initial skinning to regularize the predicted skinning weights of the skin part, i.e., the set $\\mathcal{G}$ of vertices, which have the parsing label skin. However, even for skin vertices, we found that the initial skinning weights around the joints are not accurate while the rigid parts (i.e., the vertices away from joints) usually have correct initial skinning. We identify the set of rigid vertices as $\\mathcal{R} = \\{k|\\max (\\mathbf{w}_{\\mathrm{init},k}) > u\\}$ where $u = 0.95$ . Thus, our part loss", + "bbox": [ + 75, + 167, + 468, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathscr {L} _ {\\text {p a r t}} = \\sum_ {i \\in \\mathscr {R} \\cap \\mathscr {G}} \\| f _ {\\omega} (\\bar {\\mathbf {v}} _ {i}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {w} _ {\\text {i n i t}, i} \\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 325, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "enforces that the predicted skinning weights are close to the initial ones iff the human parsing label of a vertex is the skin $(\\mathcal{G})$ and the vertex belongs to the rigid part $(\\mathcal{R})$ .", + "bbox": [ + 75, + 364, + 468, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy. Our training proceeds in four stages. First, we train SkinNet without using the rendering loss $\\mathcal{L}_{\\mathrm{rend}}$ . This step is required to ensure that the posed mesh roughly overlaps with the foreground masks. Next, the AlbedoNet is trained using only the rendering loss $\\mathcal{L}_{\\mathrm{rend}}$ while the SkinNet weights $\\omega$ are fixed. Importantly, the AlbedoNet is not pose-dependent, but we optimize the weights $\\omega'$ across all frames. Afterward, the AlbedoNet and the SkinNet are fixed and we train the ShadowNet using the rendering loss $\\mathcal{L}_{\\mathrm{rend}}$ . Again we train on all training frames. Lastly, since we now have a pose- and view-dependent appearance model available, we refine the SkinNet weights including the rendering loss. We found this gives the best performance, as validated by our results.", + "bbox": [ + 75, + 410, + 468, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 633, + 163, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate our method on 3 subjects wearing tight and loose clothing, i.e., pants or skirts. Two of them are from the DynaCap dataset [15] and they are referred as $D2$ and $D5$ . $D2$ is wearing short pants and a T-shirt while $D5$ is in a skirt. The total numbers of cameras are 94 and 101 for those two sequences and our method can be trained using as few as 10 cameras. There are around 19000 training frames per subject depicting the actor in various articulated poses. For evaluation, there is a separate testing set of around 7000 frames per subject. For convenience, we keep 1 out of 10 frames for training and evaluation. Additionally, we captured a new subject, $V6$ , wearing a T-shirt and trousers using 116 cameras. We captured a separate training (17730 frames) and testing sequence (5000 frames). Importantly, for this sequence, we recover hand gestures and a respective 3D model including hands.", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. For evaluation, we reconstruct the dense 3D geometry for the testing frames using multi-view stereo reconstruction [2]. We do not use NeuS here as it is very time-consuming. Those dense meshes serve as ground truth scans and we compare the posed meshes recovered by our method as well as competing methods in terms of the Chamfer distance and the Hausdorff distance going from the recovered mesh to the ground truth scan (M2S) and vice versa (S2M). We report the average results over all testing frames. All metrics are reported in centimeter scale.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing Methods. We compare our work to our re-implementation of Pinocchio [7], which proposed a heat diffusion process to obtain a static set of skinning weights. Moreover, we compare to SCANimate [49] and SNARF [8], which, in stark contrast to our work, require dense point clouds for training. They learn static skinning weights and a pose-aware implicit shape. Since our approach mostly focuses on skinning, we compare to SCANimate with and without pose-aware shape, referred to as SCANimate and SCANimate* in the following. Finally, we compare to RigNet [59], which can jointly predict a rigging skeleton and static skinning weights given a template. Unfortunately, SkinningNet [41] and NeuroSkinning [34] do not provide their code and, thus, we cannot compare to them. For more details and comparisons to other (slightly less) related works [31, 56], we refer to the supplemental document.", + "bbox": [ + 496, + 243, + 892, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 494, + 683, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 3, we qualitatively evaluate our approach. In the first column, we show the reconstructed static geometry for different subjects. In the following columns, we show the characters in novel poses from the test set overlayed onto the respective images. Note that our posed character nicely matches the reference image, which confirms that our pose-dependent skinning indeed generalizes to novel poses and results in accurately posed meshes. In Fig. 4, we further visualize the effect of the pose-dependent skinning weights obtained by our approach. The initial skinning weights are fixed across poses leading to the typical skinning artifacts and wrong deformation behavior. In contrast, our proposed pose-dependent skinning weights adapt according to the pose and as a result the deformations look more natural.", + "bbox": [ + 496, + 518, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare our results to previous works. Note that we achieve the most accurate results when compared to works only focusing on character skinning. This validates the necessity of pose-dependent skinning weights since other works rely on static ones. In this table, SCAN-imate and SNARF predict pose-aware shapes while assuming dense point clouds are given. Since it is not the scope of this work to predict pose-dependent non-rigid deformations and our method solely learns from video data, we placed", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1382", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/719713ab2f42937e14871c69ffd912807906a8f24c90a9b142a6bc4127eab9a4.jpg", + "image_caption": [ + "Figure 3. Qualitative results. From left to right: Recovered static character template. Reference image showing a test pose. Our posed character overlayed onto the reference image. The precise overlay of the posed geometry confirms the effectiveness of our approach." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 834, + 290 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/62fa1d9ba254515436c5ca583ee449fa3449b29ed305ec001bc0318ca4e63ece.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Quantitative Geometry Comparison
SubjectD2D5V6
MethodChamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓
Pinnochio [7]3.7602.1621.5995.0772.8112.2673.3581.8711.487
SCANimate* [49]3.7502.0991.6505.4532.9652.4883.5021.8901.612
RigNet [59]3.5992.0781.5214.9892.8362.1533.3691.8941.475
Ours3.0341.7461.2884.5122.4422.0702.9931.7191.274
SCANimate [49]2.8421.6461.1964.9822.2842.6983.1541.7141.440
SNARF [8]2.5911.4521.1394.3202.2882.0322.7601.4961.263
", + "bbox": [ + 98, + 332, + 870, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b7bfe1688536b0c2107190fe42e6468994f3c962a4c90860a9099dfaf15cc45.jpg", + "image_caption": [ + "Figure 4. Qualitative results showing the pose-dependent skinning weights. As a reference, we show the result using the initial static skinning weights. Notably, using static skinning weights leads to artifacts and worse deformation. When using our pose-dependent skinning, one can see a clear improvement in the final result and we also show how the weights change over different poses." + ], + "image_footnote": [], + "bbox": [ + 76, + 518, + 429, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "them separately. Nonetheless, it is worth noting that our approach is close to SCANImate and SNARF in terms of accuracy and better than SCANImate for subject $D5$ and $V6$ even though our training does not require point clouds.", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also visually compare to those works in Fig. 5. We can see qualitatively our method achieves a lower error especially in joint regions where most of the skinning artifacts", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c2c5e053f2fd2086dc5513b3a9a6299cbc47c643bfff42d85ba7006ca99e658.jpg", + "table_caption": [ + "Table 1. We compare our method to previous works on 3 subjects. Our method obtains better results than other skinning-based approaches. It is even close to SCANimate and SNARF, which require dense pointclouds while we solely use multi-view video. Importantly, they focus on learning pose-aware shapes while our goal is to obtain a fully rigged and skinned character in a completely automated fashion." + ], + "table_footnote": [], + "table_body": "
Ablation Study on D2
MethodChamfer↓M2S↓S2M↓
Initial weights3.7612.1621.599
Static weights3.3541.9351.418
w/o Lrend3.1161.7831.333
w/ albedo-only3.0471.7501.297
w/ NeuS appearance3.1521.7861.365
w/ NeuS albedo3.1371.7951.342
w/ displacements3.1911.8051.386
Ours (10 views)3.1291.7911.337
Ours3.0341.7461.288
", + "bbox": [ + 513, + 518, + 879, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. We compare our method against baselines. All designs improve the accuracy confirming the effectiveness of VINECS.", + "bbox": [ + 500, + 669, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "become visible. This can be nicely seen in the error maps, which visualize the per-vertex error. Although SCAnimate and SNARF perform well on naked body and tight clothes regions, they are worse than ours around loose clothing, as they rely on SMPL while our method does not have such a restriction. Further, the posed meshes clearly show fewer artifacts for our approach compared to others. Note that the entire character including the geometry, rigged skeleton, and the pose-dependent skinning can be obtained solely from multi-view video without any manual user input.", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1383", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ebc3df62fe87c9fcb61b4eb00b466dfd019df2bef5405e00db6ebc93889705f.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison. For each method, we visualize the recovered posed geometry as well as the per-vertex error map when comparing the ground truth in terms of M2S. Note that our method consistently shows the lowest error and also has the least visual artifacts." + ], + "image_footnote": [], + "bbox": [ + 125, + 88, + 841, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/147be2e9b316f2464c834dd70fcdb6a9eb6118b251bc3dfa1049bb34874b0779.jpg", + "image_caption": [ + "Figure 6. We compare the initial skinning, the static skinning optimized using our supervision scheme, ablating the rendering loss, and our pose-dependent skinning. We outperform all baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 318, + 450, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation", + "text_level": 1, + "bbox": [ + 76, + 525, + 181, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we evaluate our design choices in Tab. 2. First, we evaluate the effect of pose-dependent skinning weights. Therefore, we compare our final result to the initial skinning, which we obtain by leveraging Pinocchio [7] and a version of our method where we use our supervision scheme, but optimize a static set of skinning weights, i.e., the weights are fixed across all poses. From the results, we can see that pose-dependent weights significantly improve the accuracy over the two baselines proofing the effectiveness of our formulation. This is also visualized in Fig. 6.", + "bbox": [ + 75, + 551, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another key aspect when learning those weights solely from video data is our proposed supervision scheme. When we do not employ the dense rendering loss or soley use the albedo without shadow estimation, the performance drops. Moreover, we also evaluate the accuracy when using a dense rendering loss with the static appearance from NeuS and using NeuS appearance as albedo while the shadow network is learned. We found that our proposed rendering scheme performs best. To validate the effectiveness of training pose-dependent skinning, we replace the SkinNet by a network predicting per-vertex displacement, which generalizes worse than SkinNet. We train our model with sparse cameras (10 views), and the performance is still satisfying.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 319, + 746, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although we propose the first method for learning pose-dependent skinning solely from multi-view video, our method still has some limitations, which opens up interesting directions for future work. Currently, we query the SkinNet for every vertex in canonical space, which may be inefficient when many points are sampled. Thus, future work could involve exploring more efficient architectures like hashgrids [43]. While hand and body can be faithfully animated, there is no control over the facial expression. Thus, extending our automated character creation pipeline with some parametric model for facial expressions might be a promising direction. Also, a joint consideration of rigging, skinning, and pose tracking may further improve the quality of the 3D animation-ready characters. Lastly, some artifacts still remain since skinning alone cannot account for high-frequency deformations. Combining VINECS with downstream tasks such as human surface tracking/modeling could alleviate such artifacts.", + "bbox": [ + 496, + 353, + 892, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We proposed the first method for creating a fully rigged and skinned character solely from multi-view video without any manual processing. To this end, we first acquire a canonical template using an implicit surface reconstruction, automated initial skinning and mesh parsing. Most importantly, we learn a pose-dependent skinning field from a multi-view video, which greatly reduces artifacts. We show how such a skinning field can be learned in a weakly supervised manner by an appearance model and differentiable rendering. Our work demonstrates that animatable character creation can be fully automated while maintaining high-quality geometry, rigging, and skinning.", + "bbox": [ + 496, + 635, + 892, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Heming Zhu and Diogo Luvizon for their help. This project was also supported by Saarbrücken Research Center for Visual Computing, Interaction, and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784).", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1384", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 2", + "[2] Agisoft. PhotoScan. http://www.agisoft.com, 2016. 6, 1, 3", + "[3] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. ACM Trans. Graph., 21(3):612-619, 2002. 1, 2", + "[4] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a Correlated Model of Identity and Pose-Dependent Body Shape Variation for Real-Time Synthesis. In ACM SIGGRAPH / Eurographics Symposium on Computer Animation. The Eurographics Association, 2006. 3", + "[5] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. ACM Transactions on Graphics, 24(3):408-416, 2005. 3", + "[6] Seungbae Bang and Sung-Hee Lee. Spline interface for intuitive skinning weight editing. ACM Trans. Graph., 37(5), 2018. 2", + "[7] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Trans. Graph., 26(3), 2007, 2, 4, 6, 7, 8, 3", + "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 3, 4, 6, 7", + "[9] Olivier Dionne and Martin de Lasa. Geodesic voxel binding for production character meshes. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 173-180, New York, NY, USA, 2013. Association for Computing Machinery. 2", + "[10] Olivier Dionne and Martin de Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE Transactions on Visualization and Computer Graphics, 20 (10):1367-1378, 2014. 2", + "[11] Sven Forstmann, Jun Ohya, Artus Krohn-Grimberghe, and Ryan McDougall. Deformation styles for spline-based skeletal animation. In Proceedings of the 2007 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 141-150, Goslar, DEU, 2007. Eurographics Association. 2", + "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 1", + "[13] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2", + "[14] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 1:1, 2020. 2", + "[15] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Trans. Graph., 40(4), 2021. 2, 5, 6, 1", + "[16] Marc Habermann, Lingjie Liu, Weipeng Xu, Gerard Pons-Moll, Michael Zollhoefer, and Christian Theobalt. Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech., 6(3), 2023. 2", + "[17] Dae-Eun Hyun, Seung-Hyun Yoon, Jung-Woo Chang, Joon-Kyung Seong, Myung-Soo Kim, and Bert Jüttler. Sweep-based human deformation. The Visual Computer, 21:542-550, 2005. 2", + "[18] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine-Hornung. Bounded biharmonic weights for real-time deformation. Commun. ACM, 57(4):99-106, 2014. 2", + "[19] Doug L. James and Christopher D. Twigg. Skinning mesh animations. ACM Trans. Graph., 24(3):399-407, 2005. 2", + "[20] Yue Jiang, Marc Habermann, Vladislav Golyanik, and Christian Theobalt. Hifecap: Monocular high-fidelity and expressive capture of human performances. In BMVC, 2022. 2", + "[21] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Trans. Graph., 31 (6), 2012. 2", + "[22] Ladislav Kavan and Jiří Žára. Spherical blend skinning: A real-time deformation of articulated models. In Proceedings of the 2005 Symposium on Interactive 3D Graphics and Games, page 9–16, New York, NY, USA, 2005. Association for Computing Machinery. 2", + "[23] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46. ACM, 2007. 2", + "[24] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2014. 3", + "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems, 2023. 2", + "[26] Binh Huy Le and Zhigang Deng. Robust and accurate skeletal rigging from mesh sequences. ACM Trans. Graph., 33(4), 2014. 2", + "[27] Binh Huy Le and Jessica K. Hodgins. Real-time skeletal skinning with optimized centers of rotation. ACM Trans. Graph., 35(4), 2016. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1385", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] J. P. Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 165-172, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2, 4", + "[29] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4, 1", + "[30] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 1, 3", + "[31] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. 2022. 3, 6, 1, 2", + "[32] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. 2020. 2", + "[33] Zhouyingcheng Liao, Jimei Yang, Jun Saito, Gerard Pons-Moll, and Yang Zhou. Skeleton-free pose transfer for stylized 3d characters. In European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2", + "[34] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Trans. Graph., 38(4), 2019. 1, 2, 6", + "[35] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6), 2021. 2", + "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3", + "[37] William Lorensen and Harvey Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21:163-, 1987. 4, 3", + "[38] N. Magnenat-Thalmann, A. Laperri'ere, and D. Thalmann. Joint-dependent local deformations for hand animation and object grasping. In Proceedings of Graphics Interface '88, pages 26-33. Canadian Man-Computer Communications Society, 1988. 2", + "[39] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 1", + "[40] Alex Mohr and Michael Gleicher. Building efficient, accurate character skins from examples. ACM Trans. Graph., 22 (3):562-568, 2003. 1, 2", + "[41] A. Mosella-Montoro and J. Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18572-18581, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1, 2, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Tomohiko Mukai and Shigeru Kuriyama. Efficient dynamic skinning with low-rank helper bone controllers. ACM Trans. Graph., 35(4), 2016. 2", + "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 8", + "[44] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. 2023. 2", + "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[46] Sang II Park and Jessica K. Hodgins. Capturing and animating skin deformation in human motion. ACM Trans. Graph., 25(3):881-889, 2006. 2", + "[47] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Animatable neural radiance fields for human body modeling. ICCV, 2021. 2", + "[48] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 1(1): 9054-9063, 2021. 2", + "[49] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7", + "[50] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. Advances in neural information processing systems, 29, 2016. 1", + "[51] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[52] Peter-Pike J. Sloan, Charles F. Rose, and Michael F. Cohen. Shape by example. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, pages 135-143, New York, NY, USA, 2001. Association for Computing Machinery. 1, 3", + "[53] TheCaptury. The Captury. http://www.thecapture.com/, 2020. 1, 3, 4", + "[54] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cirmrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algo" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1386", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. 3", + "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 1, 3, 4", + "[56] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, page 1–19, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 6, 1, 2", + "[57] Xiaohuan Corina Wang and Cary Phillips. Multi-weight enveloping: Least-squares approximation techniques for skin animation. In Proceedings of the 2002 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 129–138, New York, NY, USA, 2002. Association for Computing Machinery. 1, 2", + "[58] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects, pages 63-71, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2", + "[59] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. ACM Trans. on Graphics, 39, 2020. 1, 2, 6, 7, 3", + "[60] Ji Yang, Xinxin Zuo, Sen Wang, Zhenbo Yu, Xingyu Li, Bingbing Ni, Minglun Gong, and Li Cheng. Object wakeup: 3d object rigging from a single image. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 311-327. Springer, 2022. 2", + "[61] Xiaosong Yang, Arun Somasekharan, and Jian Jun Zhang. Curve skeleton skinning for human and creature characters. Computer Animation and Virtual Worlds, 17, 2006. 2" + ], + "bbox": [ + 78, + 90, + 468, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1387", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_model.json b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a80f0296767f39cfc843733861bb10ce6b684bef --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_model.json @@ -0,0 +1,2114 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.231, + 0.131, + 0.74, + 0.154 + ], + "angle": 0, + "content": "VINECS: Video-based Neural Character Skinning" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.179, + 0.275, + 0.199 + ], + "angle": 0, + "content": "Zhouyingcheng Liao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.181, + 0.477, + 0.199 + ], + "angle": 0, + "content": "Vladislav Golyanik1" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.18, + 0.679, + 0.198 + ], + "angle": 0, + "content": "Marc Habermann1,3" + }, + { + "type": "text", + "bbox": [ + 0.72, + 0.18, + 0.888, + 0.198 + ], + "angle": 0, + "content": "Christian Theobalt1,3" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.201, + 0.812, + 0.217 + ], + "angle": 0, + "content": "1Max Planck Institute for Informatics, Saarland Informatics Campus 2The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.219, + 0.704, + 0.234 + ], + "angle": 0, + "content": "\\(^{3}\\)Saarbrücken Research Center for Visual Computing, Interaction and AI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Rigging and skinning clothed human avatars is a challenging task and traditionally requires a lot of manual work and expertise. Recent methods addressing it either generalize across different characters or focus on capturing the dynamics of a single character observed under different pose configurations. However, the former methods typically predict solely static skinning weights, which perform poorly for highly articulated poses, and the latter ones either require dense 3D character scans in different poses or cannot generate an explicit mesh with vertex correspondence over time. To address these challenges, we propose a fully automated approach for creating a fully rigged character with pose-dependent skinning weights, which can be solely learned from multi-view video. Therefore, we first acquire a rigged template, which is then statically skinned. Next, a coordinate-based MLP learns a skinning weights field parameterized over the position in a canonical pose space and the respective pose. Moreover, we introduce our pose- and view-dependent appearance field allowing us to differentiably render and supervise the posed mesh using multi-view imagery. We show that our approach outperforms state-of-the-art while not relying on dense 4D scans. More details can be found on our project page1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.21, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Animating 3D characters is a long-standing and challenging problem in computer graphics and vision. Traditionally, rigging, the process of positioning a sparse skeletal structure inside a dense 3D mesh, and skinning, the process of associating vertices with the skeleton, require large amounts of manual effort by experienced artists. This is not only expensive and time-consuming but also inherently difficult since one set of skinning weights is typically not ideal for all sorts of articulated character poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.868 + ], + "angle": 0, + "content": "There are also methods for automated rigging and skinning. Some recent works [33, 34, 41, 59] propose a learning-based solution for obtaining skinning and rigging" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.268, + 0.862, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.46, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Figure 1. We propose the first end-to-end trainable method for generating a dense and rigged 3D character mesh including learned pose-dependent skinning weights solely from multiview videos. The results demonstrate that our model can generate visually-plausible geometry without requiring manual user input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.685 + ], + "angle": 0, + "content": "given a single 3D character mesh. However, they do not consider pose-dependent skinning correctives and lead to the typical appearance artifacts. Therefore, other techniques focus on pose-dependent skinning formulations accounting for pose-dependent variations and deformations [3, 30, 40, 52, 57]. Noteworthy, none of these methods tried to learn pose-dependent skinning weight correctives purely from \\(2D\\) image data. Instead, they typically assume a 3D mesh is given, and most of them solely work on a fixed mesh resolution, which cannot be easily adjusted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.824 + ], + "angle": 0, + "content": "To overcome these limitations, we propose VINECS, i.e., Video-based NEural Character Skinning. Given solely a multi-view video of a human actor, our method creates a fully rigged, skinned, and textured 3D template (including hands) along with pose-dependent skinning weights while no manual editing is required. Moreover, VINECS enables multi-resolution character skinning since skinning weights can be sampled over a continuous 3D canonical space, which is not limited to any specific mesh resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For each character, we select one canonical pose frame, for which we extract a dense and textured 3D template mesh using implicit surface reconstruction techniques [55]. Given the template and the respective pose obtained from an off-the-shelf multi-view markerless capture system [53]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "1https://people.mpi-inf.mpg.de/~mhaberma/projects/2023-Vinecs" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "we leverage an auto-rigging and skinning method, Pinocchio [7], for obtaining an initial set of skinning weights. Note that Pinocchio produces only a static set of skinning weights, which leads to the typical geometry artifacts when poses deviate too much from the canonical pose. Therefore, we propose a coordinate-based multi-layer perceptron (MLP)—which takes as input a 3D point on the surface of the template in canonical pose and the skeletal pose—and predicts the pose-dependent skinning weights at that point. Since we want to learn the skinning weights solely from multi-view imagery, we design dedicated losses consisting of a silhouette loss and a rendering loss. Specifically, we find that obtaining a well-behaved convergence is challenging only using a static texture for the rendering loss. Thus, we adopt the rendering formulation with an albedo network and a shadow network, which predict pose- and view-dependent appearance of our 3D character resulting in an improved convergence and skinning weights prediction. In summary, our technical contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.379, + 0.47, + 0.438 + ], + "angle": 0, + "content": "- An end-to-end trainable method for animation-ready explicit character creation, involving static template generation, rigging, and pose-dependent skinning solely using 2D multi-view videos (as few as 10 views)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.44, + 0.469, + 0.47 + ], + "angle": 0, + "content": "- A coordinate-based and pose-dependent skinning formulation, which enables multi-resolution skinning." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.471, + 0.469, + 0.516 + ], + "angle": 0, + "content": "- A dedicated character appearance formulation and differentiable rendering enabling weakly supervised learning based on multi-view videos." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.379, + 0.47, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.517, + 0.47, + 0.561 + ], + "angle": 0, + "content": "We compare VINECS to the state-of-the-art and our results confirm that our method is a clear step toward accurate and automated creation of animatable 3D characters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.473, + 0.745 + ], + "angle": 0, + "content": "Potential Impact. The focus of this work is not to model deformable geometry [13, 14, 20, 32] or photorealistic appearance [15, 16, 25, 35, 44, 47, 48] of humans but on learning 3D characters with pose-dependent skinning from 2D observations only. We believe our method can greatly benefit many downstream tasks as many existing works leverage naked body models [36] for space canonicalization and feature assignment, fundamentally limiting them to clothing types of the same topology as the human body. As our rigging and skinning are accurate, easy-to-use, and robust to any type of clothing, our approach can stimulate research on the modeling of geometry and appearance of loose apparel." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.22, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Static Skinning. Linear Blend Skinning (LBS) [28, 38] linearly blends rigid transformations of each bone to obtain a posed geometry. However, this formulation can lead to a collapse of geometry near the skeleton joints, also called candy-wappers effect. To overcome this, several follow-up works have been proposed using spherical representations [22], sweep surfaces [17], curve-based skeleton parameterizations [61], Dual Quaternions [23], or by optimiz" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "ing the centers of rotation [27]. Other works focused more on computing the skinning weights itself by modeling the weight distribution as a heat diffusion process [7], as illumination [58], or by voxelizing the mesh to increase robustness [9, 10]. Later on, deformation editing was also modeled using a Laplacian energy formulation [18], and user-defined splines on the mesh enable intuitive editing of skinning-based deformation behaviors [6]. They all only assume a single geometry and skeleton are given." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.518 + ], + "angle": 0, + "content": "In contrast, data-driven methods assume multiple instances of the same character under different poses or a large collection of rigged and skinned characters. Earlier works [19, 26] requires meshes under different poses and then automatically detect skeleton joints and compute the skinning weights. Recently, NeuroSkinning [34] was proposed as the first learning-based method predicting a static set of skinning weights given the mesh and a corresponding skeleton. RigNet [59] instead jointly predicts the skeleton as well as a set of skinning weights solely given a 3D mesh as input. Therefore, they collected a large dataset of 3D characters including artist-designed skeletons and skinning weights. SkinningNet [41] proposed a two-stream graph convolutional network architecture for improved skinning weight prediction given a mesh and the skeleton. Yang et al. [60] proposed to predict the 3D shape together with the rigging from a single image. Liao et al. [33] introduced a learning-based method which jointly learns to predict skinning weights and transfer poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.894, + 0.64 + ], + "angle": 0, + "content": "What all the above methods have in common is that they do not explicitly account for dynamic or pose-dependent effects, such as muscle bulging or coarse cloth deformations. In contrast, our proposed method learns pose-dependent skinning weights, which alleviate typical artifacts that arise when using a static set of skinning weights. Moreover, it can also represent coarse deformations induced by the skeletal pose, such as coarse cloth deformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.824 + ], + "angle": 0, + "content": "Dynamic Methods. More closely related to our method are approaches that account for dynamic or pose-dependent effects, which are discussed next. There are works [11] that allow an artist to add pose-dependent effects by painting onto specialized textures or by adding so-called helper bones [42]. Alternatively, Kavan et al. [21] proposed to find the optimal skinning weights by minimizing an elastic energy and adding joint-based deformers. Interestingly, those methods do not assume any dynamic observations of the character while still modeling dynamic effects. However, they either require manual editing [11, 42] or can mostly model elasticity-based dynamics [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Similar to the static methods, several dynamic methods assume a (large) dataset of character meshes is given. Earlier works [3, 40, 46, 57] account for dynamic effects by capturing a set of range scans depicting the subject in different poses. Given a new pose, a set of pose-dependent blend" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "1378" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.529 + ], + "angle": 0, + "content": "weights is estimated, which linearly combines the set of example poses in the dataset or alternatively uses linear and radial basis functions [52]. However, since the complexity of computation typically linearly scales with the number of example poses in the dataset, those methods cannot scale up to a very large number of examples (e.g., 20000 pose frames), which is in stark contrast to our approach that can deal with such an amount of data. There are also parametric naked human body models [4, 5, 36] learned by applying PCA over a large set of human scans, which can account for different shape- and pose-dependent deformations. In contrast, we are building a highly-detailed rigged and skinned character that can wear (loose) clothing in a completely automated fashion. Li et al. [30] predict static skinning weights and pose-dependent neural blend shapes. They assume a dataset of dense meshes is provided, whereas we purely learn pose-dependent skinning weights from videos. SCANimate [49] also requires dense point clouds, and learns a static set of skinning weights, while pose-dependent shapes are predicted by a pose-aware implicit field. SNARF [8] represents the character as a pose-dependent neural implicit surface in the canonical pose space while the learned skinning weights are pose-agnostic. Again, they require point cloud data while our method can be supervised directly on the video. In summary, none of the above methods that consider pose-dependent effects investigated learning solely from 2D imagery. Thus, we are the first method demonstrating that dynamic effects in the form of pose-dependent skinning weights can be learned only from videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.473, + 0.727 + ], + "angle": 0, + "content": "Recently, some works were proposed to skin and rig implicit surface representations learn an implicit human avatar. ARAH [56] learns an implicit human model using neural rendering. However, as it relies on a SMPL template and an inverse root finding to convert from view space to canonical space, it easily fails when the garment deviates much from the body. Importantly, it also does not have the vertex correspondence over time. TAVA [31] avoids these problems by getting rid of a template, which, however, leads to poor geometry recovery mainly due to their density-based geometry representation. While these methods can reconstruct an implicit avatar, they do not learn a high-quality skinning weight, which is important to traditional graphics pipeline." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.17, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given synchronized and calibrated multi-view RGB videos of a human character in motion, our goal is to learn an animatable 3D human model with dynamic skinning, without any manual process, such as meshing, rigging, and skinning. To this end, we first extract the static geometry of the actor from one canonical pose frame (typically depicting the actor in a T-pose) of the recording and automatically compute initial skinning weights (Sec. 3.1). However, the initial skinning weights are calculated only based on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.363 + ], + "angle": 0, + "content": "the static geometry of the human character, and, as studied in earlier works, one set of weights might be ideal for one pose while it can lead to strong artifacts in other poses. Thus, we propose a skinning network, called SkinNet, to predict the dynamic skinning weights as a function of the skeletal pose and 3D coordinates in the canonical space (Sec. 3.2). Given the skeletal pose, the static geometry, and the pose-dependent skinning weights, we leverage Linear Blend Skinning (LBS) to obtain the posed geometry. We then use a differentiable renderer to weakly supervise the skinning weights solely based on multi-view imagery. Thereafter, we propose a novel appearance field (Sec. 3.3), which is composed of an albedo component and a pose-and view-dependent shadow/shading component. In addition, we propose several regularization losses ensuring that the skinning weights are also robust to out-of-distribution poses and that the final animated geometry looks visually plausible (Sec. 3.4). An overview is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Data Preparation. For a human subject, we record a multi-view video of the performance using \\(C\\) calibrated and synchronized cameras with a total length of \\(F\\) frames. For all frames \\(\\mathcal{I}_{c,f}\\), we calculate the foreground segmentation using color keying or background matting [51] and compute respective distance transform images \\(\\mathcal{D}_{c,f}\\). Here, \\(c\\) and \\(f\\) refer to the respective camera and frame index. We capture the size of the human skeleton and the motion using markerless motion capture [53]. The skeleton motion is parameterized by a root rotation \\(\\alpha_{f} \\in \\mathbb{R}^{3}\\), a root translation \\(\\mathbf{t}_f \\in \\mathbb{R}^3\\), and the joint angles \\(\\rho_f \\in \\mathbb{R}^D\\). The full pose vector is defined as \\(\\theta_f = [\\alpha_f, \\mathbf{t}_f, \\rho_f]^T \\in \\mathbb{R}^{3 + 3 + D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.894, + 0.655 + ], + "angle": 0, + "content": "Notation. To represent a variable in the canonical space, we use a bar over the letter, i.e., \\(\\bar{\\mathbf{x}}\\). For any point in canonical and global space, we use \\(\\bar{\\mathbf{x}}\\) and \\(\\mathbf{x}\\), respectively, and mesh vertices are denoted as \\(\\bar{\\mathbf{v}}_i\\) and \\(\\mathbf{v}_i\\) where \\(i\\) is the \\(i\\)th vertex. Without loss of generality, we assume the frame \\(f\\) is fixed in the following (if not explicitly mentioned otherwise) and omit the subscript for better readability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.753, + 0.681 + ], + "angle": 0, + "content": "3.1. Canonical Character Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.884 + ], + "angle": 0, + "content": "As the first step, we acquire a canonical model of the actor. Recently, implicit surface and coordinate-based representations [45] have gained considerable attention due to their flexibility and comparably simple integration into learning frameworks, and they were also leveraged in the context of pose-dependent human deformation modeling [8]. Moreover, we compared the reconstruction accuracy on the first frame \\( f = 0 \\) between classical multi-view stereo and implicit reconstruction method called NeuS [55] in the supplemental material. Interestingly, NeuS has a much higher level of detail and comparably less noise on the surface. Thus, we use NeuS [55] to reconstruct the mesh in the first frame, also called the canonical frame/space." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implicit vs. Explicit Canonical Model. In stark con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1379" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.091, + 0.858, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Figure 2. Method overview. Our method, VINECS, is a fully automated solution to template generation, rigging, and pose-dependent skinning solely using multi-view videos. We first recover the skeletal poses using markerless motion capture [53] and a static template mesh using implicit surface reconstruction [55]. An initial skinning is obtained by a heat diffusion process [7]. Since one fixed set of skinning weights can lead to artifacts when significantly changing the character pose, we learn pose-dependent skinning weights using a coordinate-based MLP. An auxiliary appearance field is applied to better supervise the learning guide the training of pose-dependent skinning. Note that we do not intend to learn a detailed appearance and only the components in the blue box are used during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.473, + 0.614 + ], + "angle": 0, + "content": "trast to Chen et al. [8], we run Marching Cubes [37] to convert the implicit surface into an explicit mesh and obtain per-vertex colors from NeuS by setting the viewing direction to the normal direction of the mesh. The reason for choosing an explicit mesh in canonical space is three-fold: 1.) It avoids running the rather slow Marching Cubes algorithm for every potentially posed mesh. 2.) We avoid the backward skinning [8], which boils down to a correspondence search that has to be solved using iterative optimization. 3.) We can maintain correspondence on the surface across poses, which is especially useful when texturing and other editing have to be applied. Thus, we aim for a hybrid strategy for template acquisition, i.e., using the state-of-the-art method for implicit surface reconstruction and then converting it to an explicit mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.473, + 0.751 + ], + "angle": 0, + "content": "Meshing and Initial Skinning. During training, we downsample the template mesh to \\(N\\) vertices where \\(\\bar{\\mathbf{v}}_i\\in \\mathbb{R}^3\\) denotes the ith vertex. Given the template mesh and the corresponding posed skeleton, i.e. \\(\\theta_{f}\\), we utilize Pinocchio [7] to obtain an initial set of skinning weights \\(\\mathbf{w}_{\\mathrm{init},i}\\in \\mathbb{R}^{J}\\) for each vertex \\(i\\) (see Fig. 4). Here, \\(J\\) denotes the number of skinning joints. Given these weights and a skeletal pose \\(\\theta\\), we can transform the vertex \\(\\bar{\\mathbf{v}}_i\\) from the canonical space to the posed space using LBS [28]:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.756, + 0.47, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} = L B S \\left(\\bar {\\mathbf {v}} _ {i}, \\mathbf {w} _ {\\text {i n i t}, i}, \\theta\\right): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {J} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {3}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Mesh Parsing. Since different body parts and materials (skin vs. clothing) have different deformation behaviors, we compute per-vertex human parsing labels. We render the 3D model into all camera views and apply a 2D human parsing approach [29] on each view. We then perform max-voting to obtain the per-vertex human parsing labels. Note that so far, we obtained a segmented, rigged, and statically skinned character model without any manual user input." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.388, + 0.776, + 0.405 + ], + "angle": 0, + "content": "3.2. Pose-dependent Skinning Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.546 + ], + "angle": 0, + "content": "So far, the initial skinning weights obtained in Sec. 3.1 have two main limitations: 1.) They are solely computed on a single static pose leading to artifacts such as candy wrappers and mesh distortions if the new pose significantly differs from the canonical one. 2.) The set of weights is computed per-vertex. Thus, the resolution of the mesh has to be fixed beforehand and any change in the mesh requires a new computation of skinning weights or some complicated weight transfer from one mesh to the other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.894, + 0.624 + ], + "angle": 0, + "content": "To resolve this, we propose SkinNet \\( f_{\\omega} \\), which is a coordinate-based MLP where \\( \\omega \\) denotes the trainable network weights. Given any 3D point \\( \\bar{\\mathbf{x}} \\) in the canonical space, \\( f_{\\omega} \\) predicts its skinning weight \\( \\mathbf{w} \\) conditioned on the normalized human pose \\( \\tilde{\\theta} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.634, + 0.892, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {\\tilde {\\theta}} = f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\boldsymbol {\\theta}}): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {J}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.831 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{x}}\\) and \\(\\tilde{\\theta}\\) are concatenated before being fed into the MLP. Here, the normalized pose \\(\\tilde{\\theta}\\) is obtained by neglecting the global translation as well as the yaw angle, i.e., the rotation around the spine. The intuition behind is that the skinning weights do not depend on where in global 3D space the person performs a respective pose, and, similarly, the skinning weight prediction should be agnostic to which direction the person is facing. Thus, we mask those degrees of freedom before feeding the pose vector into the network. To obtain the transformed point \\(\\mathbf{x}\\) in posed space, one can insert Eq. (2) into Eq. (1) resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.842, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = L B S (\\bar {\\mathbf {x}}, f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\theta}), \\theta). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Revisiting the drawbacks of initial skinning discussed above, the SkinNet now supports pose-dependent skinning" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "weights such that the artifacts of static skinning can be reduced. Moreover, since SkinNet is a coordinate-based MLP, it supports arbitrary mesh resolution since each vertex is queried independently, which is in stark contrast to an architecture that outputs all skinning weights at once and where the final tensor shape is fixed to a specific mesh resolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.193, + 0.327, + 0.209 + ], + "angle": 0, + "content": "3.3. Auxiliary Appearance Field" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.471, + 0.488 + ], + "angle": 0, + "content": "We can now pose the geometry with our pose-dependent skinning weight formulation. However, since we want to supervise SkinNet solely on multi-view images, we also have to model the appearance of the actor. A naive choice would be to just leverage the static vertex colors acquired from NeuS. However, we found that this texture contains baked-in shadows and wrinkles, and cannot account for the pose- and view-dependent appearance changes harming the training of the SkinNet. We also refer to our ablation studies in Sec. 4.3. To address this issue, we propose an auxiliary appearance field, which consists of an albedo field predicting a pose- and view-agnostic color for a given point \\(\\bar{\\mathbf{x}}\\) in the canonical space and a shadow field predicting the pose- and view-dependent shadow/shading properties of \\(\\bar{\\mathbf{x}}\\). Important to note here is that we are not interested in creating a highly photorealistic appearance model of the human, but we mainly use it as an auxiliary tool to better supervise the learning of the skinning weight network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.777 + ], + "angle": 0, + "content": "More specifically, the albedo field, referred to as AlbedoNet, is an MLP that predicts a static albedo value \\(a_{\\omega'}(\\bar{\\mathbf{x}}) = \\mathbf{a} \\in \\mathbb{R}^3\\) for a given 3D sample point in the canonical space. Here, \\(\\omega'\\) are the trainable weights of the albedo network. Moreover, the shadow field is also parameterized by an MLP \\(r_{\\omega''}(\\bar{\\mathbf{x}}, \\theta, \\mathbf{n}, \\mathbf{d}) = s \\in \\mathbb{R}^+\\), called ShadowNet, which predicts a scalar multiplier accounting for color changes due to shadows and shading. Note that we assume that such effects happen uniformly across all color channels and, thus, we are only predicting a single scalar instead of a scaling factor per color channel. In addition to taking the point in canonical space, the shadow field also takes the skeletal pose \\(\\theta\\) as input as well as the surface normal \\(\\mathbf{n}\\) of \\(\\mathbf{x}\\) in global space and the viewing direction \\(\\mathbf{d}\\), which allows to potentially model pose- and view-dependent lighting effects such as shadows and shading. \\(\\omega''\\) denotes the respective network weights. Details about the architectures are provided in the supplemental document. The final color \\(\\mathbf{c} \\in \\mathbb{R}^3\\) for point \\(\\mathbf{x}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.789, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = a _ {\\omega^ {\\prime}} (\\bar {\\mathbf {x}}) r _ {\\omega^ {\\prime \\prime}} (\\bar {\\mathbf {x}}, \\theta , \\mathbf {n}, \\mathbf {d}). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.394, + 0.833 + ], + "angle": 0, + "content": "3.4. Multi-view Video-based Supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Next, we describe our dedicated supervision strategy that solely requires multi-view imagery. We first introduce the individual loss terms and then describe our training strategy. Importantly, we can obtain a posed template mesh" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\(\\mathbf{V} \\in \\mathbb{R}^{N \\times 3}\\) and the pose- and view-dependent per-vertex colors \\(\\mathbf{C}_c \\in \\mathbb{R}^{N \\times 3}\\) by evaluating Eq. (3) and Eq. (4) for all vertices \\(\\bar{\\mathbf{v}}_i\\) and views \\(c\\). Moreover, the following losses that are applied to the deformed geometry and its appearance can directly backpropagate into the respective networks since our formulation is fully differentiable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Silhouette Loss. This loss [15] ensures that the projected posed geometry projects onto the foreground masks for all views (see supp. material). Since our pose-dependent skinning field Eq. (3) is differentiable with respect to the SkinNet weights \\(\\omega\\), the silhouette loss can provide supervision and learn pose-dependent skinning by matching the posed model silhouette and the image silhouettes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Rendering Loss. Although the silhouette loss ensures that the posed geometry matches the multi-view silhouettes, it cannot supervise drifts on the visual hull. Supervision directly from the RGB images can help with resolving this issue since appearance features can constrain drifts in the image plane. Thus, we introduce the rendering loss" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.39, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d}} = \\sum_ {c = 1} ^ {C} \\| \\Pi_ {c} (\\mathbf {V}, \\mathbf {C} _ {c}) - \\mathcal {I} _ {c} \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.527 + ], + "angle": 0, + "content": "which leverages a differentiable renderer \\(\\Pi_c\\) based on the one of Habermann et al. [15] to render the posed and colored model into all camera views \\(c\\) that is then compared to the ground truth image \\(\\mathcal{I}_c\\). Importantly, this loss can supervise both the pose-dependent skinning field as well as the pose- and view-dependent appearance field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Laplacian Loss. Since the data terms alone can still lead to degenerate results, we also introduce some regularization on the geometry and the skinning weights. First, we regularize the posed geometry with a Laplacian loss" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.599, + 0.892, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l a p}} = \\sum_ {i = 1} ^ {N} \\| \\mathbf {v} _ {i} - \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {k \\in \\mathcal {N} _ {i}} \\mathbf {v} _ {k} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_i\\) is the indices of one-ring neighborhood of vertex \\(i\\). This regularizer ensures locally smooth geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.814 + ], + "angle": 0, + "content": "**Skinning Regularization.** We found that SkinNet can overfit the training poses and respective deformations by predicting non-local skinning weights, e.g., the foot joint has a non-zero skinning weight for vertices on the shoulder. This, however, harms generalization to novel poses at test time. To prevent this, we regularize the skinning weights by constraining their values based on the geodesic distance \\( d_{\\mathrm{geo}} \\) between the vertex and the initial body part it belongs to. This can be formalized with the following loss" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.824, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s k i n}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\mathbf {w} _ {i, j} \\left(\\frac {\\min _ {k \\in \\mathscr {A} _ {j}} d _ {\\text {g e o}} (\\bar {\\mathbf {v}} _ {i} , \\bar {\\mathbf {v}} _ {k})}{d _ {\\text {g e o m a x}}}\\right) ^ {r}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.88, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathscr {A} _ {j} = \\left\\{k \\mid j = \\underset {t} {\\operatorname {a r g m a x}} \\mathbf {w} _ {\\text {i n i t}, k, t} \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1381" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_j\\) is the set of vertices, which have the highest skinning weight with respect to joint \\(j\\). \\(r = 3\\) is a hyperparameter which controls how penalization increases as the geodesic distance increases. \\(d_{\\mathrm{geomax}}\\) is the maximum geodesic distance between any of the vertex pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Part-wise Regularization. In contrast to the clothed areas, the skin part has rather static skinning. Thus, we use the initial skinning to regularize the predicted skinning weights of the skin part, i.e., the set \\(\\mathcal{G}\\) of vertices, which have the parsing label skin. However, even for skin vertices, we found that the initial skinning weights around the joints are not accurate while the rigid parts (i.e., the vertices away from joints) usually have correct initial skinning. We identify the set of rigid vertices as \\(\\mathcal{R} = \\{k|\\max (\\mathbf{w}_{\\mathrm{init},k}) > u\\}\\) where \\(u = 0.95\\). Thus, our part loss" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.327, + 0.47, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathscr {L} _ {\\text {p a r t}} = \\sum_ {i \\in \\mathscr {R} \\cap \\mathscr {G}} \\| f _ {\\omega} (\\bar {\\mathbf {v}} _ {i}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {w} _ {\\text {i n i t}, i} \\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.41 + ], + "angle": 0, + "content": "enforces that the predicted skinning weights are close to the initial ones iff the human parsing label of a vertex is the skin \\((\\mathcal{G})\\) and the vertex belongs to the rigid part \\((\\mathcal{R})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.622 + ], + "angle": 0, + "content": "Training Strategy. Our training proceeds in four stages. First, we train SkinNet without using the rendering loss \\(\\mathcal{L}_{\\mathrm{rend}}\\). This step is required to ensure that the posed mesh roughly overlaps with the foreground masks. Next, the AlbedoNet is trained using only the rendering loss \\(\\mathcal{L}_{\\mathrm{rend}}\\) while the SkinNet weights \\(\\omega\\) are fixed. Importantly, the AlbedoNet is not pose-dependent, but we optimize the weights \\(\\omega'\\) across all frames. Afterward, the AlbedoNet and the SkinNet are fixed and we train the ShadowNet using the rendering loss \\(\\mathcal{L}_{\\mathrm{rend}}\\). Again we train on all training frames. Lastly, since we now have a pose- and view-dependent appearance model available, we refine the SkinNet weights including the rendering loss. We found this gives the best performance, as validated by our results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.165, + 0.649 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Dataset. We evaluate our method on 3 subjects wearing tight and loose clothing, i.e., pants or skirts. Two of them are from the DynaCap dataset [15] and they are referred as \\(D2\\) and \\(D5\\). \\(D2\\) is wearing short pants and a T-shirt while \\(D5\\) is in a skirt. The total numbers of cameras are 94 and 101 for those two sequences and our method can be trained using as few as 10 cameras. There are around 19000 training frames per subject depicting the actor in various articulated poses. For evaluation, there is a separate testing set of around 7000 frames per subject. For convenience, we keep 1 out of 10 frames for training and evaluation. Additionally, we captured a new subject, \\(V6\\), wearing a T-shirt and trousers using 116 cameras. We captured a separate training (17730 frames) and testing sequence (5000 frames). Importantly, for this sequence, we recover hand gestures and a respective 3D model including hands." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Metrics. For evaluation, we reconstruct the dense 3D geometry for the testing frames using multi-view stereo reconstruction [2]. We do not use NeuS here as it is very time-consuming. Those dense meshes serve as ground truth scans and we compare the posed meshes recovered by our method as well as competing methods in terms of the Chamfer distance and the Hausdorff distance going from the recovered mesh to the ground truth scan (M2S) and vice versa (S2M). We report the average results over all testing frames. All metrics are reported in centimeter scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.893, + 0.485 + ], + "angle": 0, + "content": "Competing Methods. We compare our work to our re-implementation of Pinocchio [7], which proposed a heat diffusion process to obtain a static set of skinning weights. Moreover, we compare to SCANimate [49] and SNARF [8], which, in stark contrast to our work, require dense point clouds for training. They learn static skinning weights and a pose-aware implicit shape. Since our approach mostly focuses on skinning, we compare to SCANimate with and without pose-aware shape, referred to as SCANimate and SCANimate* in the following. Finally, we compare to RigNet [59], which can jointly predict a rigging skeleton and static skinning weights given a template. Unfortunately, SkinningNet [41] and NeuroSkinning [34] do not provide their code and, thus, we cannot compare to them. For more details and comparisons to other (slightly less) related works [31, 56], we refer to the supplemental document." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.684, + 0.511 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.732 + ], + "angle": 0, + "content": "In Fig. 3, we qualitatively evaluate our approach. In the first column, we show the reconstructed static geometry for different subjects. In the following columns, we show the characters in novel poses from the test set overlayed onto the respective images. Note that our posed character nicely matches the reference image, which confirms that our pose-dependent skinning indeed generalizes to novel poses and results in accurately posed meshes. In Fig. 4, we further visualize the effect of the pose-dependent skinning weights obtained by our approach. The initial skinning weights are fixed across poses leading to the typical skinning artifacts and wrong deformation behavior. In contrast, our proposed pose-dependent skinning weights adapt according to the pose and as a result the deformations look more natural." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.633, + 0.758 + ], + "angle": 0, + "content": "4.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 1, we compare our results to previous works. Note that we achieve the most accurate results when compared to works only focusing on character skinning. This validates the necessity of pose-dependent skinning weights since other works rely on static ones. In this table, SCAN-imate and SNARF predict pose-aware shapes while assuming dense point clouds are given. Since it is not the scope of this work to predict pose-dependent non-rigid deformations and our method solely learns from video data, we placed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1382" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.835, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.321 + ], + "angle": 0, + "content": "Figure 3. Qualitative results. From left to right: Recovered static character template. Reference image showing a test pose. Our posed character overlayed onto the reference image. The precise overlay of the posed geometry confirms the effectiveness of our approach." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.333, + 0.872, + 0.459 + ], + "angle": 0, + "content": "
Quantitative Geometry Comparison
SubjectD2D5V6
MethodChamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓
Pinnochio [7]3.7602.1621.5995.0772.8112.2673.3581.8711.487
SCANimate* [49]3.7502.0991.6505.4532.9652.4883.5021.8901.612
RigNet [59]3.5992.0781.5214.9892.8362.1533.3691.8941.475
Ours3.0341.7461.2884.5122.4422.0702.9931.7191.274
SCANimate [49]2.8421.6461.1964.9822.2842.6983.1541.7141.440
SNARF [8]2.5911.4521.1394.3202.2882.0322.7601.4961.263
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.467, + 0.893, + 0.51 + ], + "angle": 0, + "content": "Table 1. We compare our method to previous works on 3 subjects. Our method obtains better results than other skinning-based approaches. It is even close to SCANimate and SNARF, which require dense pointclouds while we solely use multi-view video. Importantly, they focus on learning pose-aware shapes while our goal is to obtain a fully rigged and skinned character in a completely automated fashion." + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.52, + 0.43, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.785 + ], + "angle": 0, + "content": "Figure 4. Qualitative results showing the pose-dependent skinning weights. As a reference, we show the result using the initial static skinning weights. Notably, using static skinning weights leads to artifacts and worse deformation. When using our pose-dependent skinning, one can see a clear improvement in the final result and we also show how the weights change over different poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.856 + ], + "angle": 0, + "content": "them separately. Nonetheless, it is worth noting that our approach is close to SCANImate and SNARF in terms of accuracy and better than SCANImate for subject \\(D5\\) and \\(V6\\) even though our training does not require point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also visually compare to those works in Fig. 5. We can see qualitatively our method achieves a lower error especially in joint regions where most of the skinning artifacts" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.52, + 0.88, + 0.666 + ], + "angle": 0, + "content": "
Ablation Study on D2
MethodChamfer↓M2S↓S2M↓
Initial weights3.7612.1621.599
Static weights3.3541.9351.418
w/o Lrend3.1161.7831.333
w/ albedo-only3.0471.7501.297
w/ NeuS appearance3.1521.7861.365
w/ NeuS albedo3.1371.7951.342
w/ displacements3.1911.8051.386
Ours (10 views)3.1291.7911.337
Ours3.0341.7461.288
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.67, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Table 2. We compare our method against baselines. All designs improve the accuracy confirming the effectiveness of VINECS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "become visible. This can be nicely seen in the error maps, which visualize the per-vertex error. Although SCAnimate and SNARF perform well on naked body and tight clothes regions, they are worse than ours around loose clothing, as they rely on SMPL while our method does not have such a restriction. Further, the posed meshes clearly show fewer artifacts for our approach compared to others. Note that the entire character including the geometry, rigged skeleton, and the pose-dependent skinning can be obtained solely from multi-view video without any manual user input." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.089, + 0.842, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison. For each method, we visualize the recovered posed geometry as well as the per-vertex error map when comparing the ground truth in terms of M2S. Note that our method consistently shows the lowest error and also has the least visual artifacts." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.319, + 0.451, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.511 + ], + "angle": 0, + "content": "Figure 6. We compare the initial skinning, the static skinning optimized using our supervision scheme, ablating the rendering loss, and our pose-dependent skinning. We outperform all baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.182, + 0.54 + ], + "angle": 0, + "content": "4.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Next, we evaluate our design choices in Tab. 2. First, we evaluate the effect of pose-dependent skinning weights. Therefore, we compare our final result to the initial skinning, which we obtain by leveraging Pinocchio [7] and a version of our method where we use our supervision scheme, but optimize a static set of skinning weights, i.e., the weights are fixed across all poses. From the results, we can see that pose-dependent weights significantly improve the accuracy over the two baselines proofing the effectiveness of our formulation. This is also visualized in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Another key aspect when learning those weights solely from video data is our proposed supervision scheme. When we do not employ the dense rendering loss or soley use the albedo without shadow estimation, the performance drops. Moreover, we also evaluate the accuracy when using a dense rendering loss with the static appearance from NeuS and using NeuS appearance as albedo while the shadow network is learned. We found that our proposed rendering scheme performs best. To validate the effectiveness of training pose-dependent skinning, we replace the SkinNet by a network predicting per-vertex displacement, which generalizes worse than SkinNet. We train our model with sparse cameras (10 views), and the performance is still satisfying." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.748, + 0.336 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.893, + 0.627 + ], + "angle": 0, + "content": "Limitations. Although we propose the first method for learning pose-dependent skinning solely from multi-view video, our method still has some limitations, which opens up interesting directions for future work. Currently, we query the SkinNet for every vertex in canonical space, which may be inefficient when many points are sampled. Thus, future work could involve exploring more efficient architectures like hashgrids [43]. While hand and body can be faithfully animated, there is no control over the facial expression. Thus, extending our automated character creation pipeline with some parametric model for facial expressions might be a promising direction. Also, a joint consideration of rigging, skinning, and pose tracking may further improve the quality of the 3D animation-ready characters. Lastly, some artifacts still remain since skinning alone cannot account for high-frequency deformations. Combining VINECS with downstream tasks such as human surface tracking/modeling could alleviate such artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.893, + 0.818 + ], + "angle": 0, + "content": "Conclusion. We proposed the first method for creating a fully rigged and skinned character solely from multi-view video without any manual processing. To this end, we first acquire a canonical template using an implicit surface reconstruction, automated initial skinning and mesh parsing. Most importantly, we learn a pose-dependent skinning field from a multi-view video, which greatly reduces artifacts. We show how such a skinning field can be learned in a weakly supervised manner by an appearance model and differentiable rendering. Our work demonstrates that animatable character creation can be fully automated while maintaining high-quality geometry, rigging, and skinning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. We thank Heming Zhu and Diogo Luvizon for their help. This project was also supported by Saarbrücken Research Center for Visual Computing, Interaction, and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1384" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[2] Agisoft. PhotoScan. http://www.agisoft.com, 2016. 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.369 + ], + "angle": 0, + "content": "[3] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. ACM Trans. Graph., 21(3):612-619, 2002. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.468, + 0.441 + ], + "angle": 0, + "content": "[4] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a Correlated Model of Identity and Pose-Dependent Body Shape Variation for Real-Time Synthesis. In ACM SIGGRAPH / Eurographics Symposium on Computer Animation. The Eurographics Association, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.468, + 0.497 + ], + "angle": 0, + "content": "[5] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. ACM Transactions on Graphics, 24(3):408-416, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.5, + 0.468, + 0.54 + ], + "angle": 0, + "content": "[6] Seungbae Bang and Sung-Hee Lee. Spline interface for intuitive skinning weight editing. ACM Trans. Graph., 37(5), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[7] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Trans. Graph., 26(3), 2007, 2, 4, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[9] Olivier Dionne and Martin de Lasa. Geodesic voxel binding for production character meshes. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 173-180, New York, NY, USA, 2013. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.73, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[10] Olivier Dionne and Martin de Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE Transactions on Visualization and Computer Graphics, 20 (10):1367-1378, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.468, + 0.869 + ], + "angle": 0, + "content": "[11] Sven Forstmann, Jun Ohya, Artus Krohn-Grimberghe, and Ryan McDougall. Deformation styles for spline-based skeletal animation. In Proceedings of the 2007 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 141-150, Goslar, DEU, 2007. Eurographics Association. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[13] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[14] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 1:1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[15] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Trans. Graph., 40(4), 2021. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[16] Marc Habermann, Lingjie Liu, Weipeng Xu, Gerard Pons-Moll, Michael Zollhoefer, and Christian Theobalt. Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech., 6(3), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[17] Dae-Eun Hyun, Seung-Hyun Yoon, Jung-Woo Chang, Joon-Kyung Seong, Myung-Soo Kim, and Bert Jüttler. Sweep-based human deformation. The Visual Computer, 21:542-550, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[18] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine-Hornung. Bounded biharmonic weights for real-time deformation. Commun. ACM, 57(4):99-106, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Doug L. James and Christopher D. Twigg. Skinning mesh animations. ACM Trans. Graph., 24(3):399-407, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[20] Yue Jiang, Marc Habermann, Vladislav Golyanik, and Christian Theobalt. Hifecap: Monocular high-fidelity and expressive capture of human performances. In BMVC, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[21] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Trans. Graph., 31 (6), 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Ladislav Kavan and Jiří Žára. Spherical blend skinning: A real-time deformation of articulated models. In Proceedings of the 2005 Symposium on Interactive 3D Graphics and Games, page 9–16, New York, NY, USA, 2005. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46. ACM, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[26] Binh Huy Le and Zhigang Deng. Robust and accurate skeletal rigging from mesh sequences. ACM Trans. Graph., 33(4), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Binh Huy Le and Jessica K. Hodgins. Real-time skeletal skinning with optimized centers of rotation. ACM Trans. Graph., 35(4), 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1385" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] J. P. Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 165-172, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[29] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[30] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[31] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. 2022. 3, 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[32] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[33] Zhouyingcheng Liao, Jimei Yang, Jun Saito, Gerard Pons-Moll, and Yang Zhou. Skeleton-free pose transfer for stylized 3d characters. In European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[34] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Trans. Graph., 38(4), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[35] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[37] William Lorensen and Harvey Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21:163-, 1987. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[38] N. Magnenat-Thalmann, A. Laperri'ere, and D. Thalmann. Joint-dependent local deformations for hand animation and object grasping. In Proceedings of Graphics Interface '88, pages 26-33. Canadian Man-Computer Communications Society, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.706, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[40] Alex Mohr and Michael Gleicher. Building efficient, accurate character skins from examples. ACM Trans. Graph., 22 (3):562-568, 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] A. Mosella-Montoro and J. Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18572-18581, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[42] Tomohiko Mukai and Shigeru Kuriyama. Efficient dynamic skinning with low-rank helper bone controllers. ACM Trans. Graph., 35(4), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[44] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[46] Sang II Park and Jessica K. Hodgins. Capturing and animating skin deformation in human motion. ACM Trans. Graph., 25(3):881-889, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[47] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Animatable neural radiance fields for human body modeling. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[48] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 1(1): 9054-9063, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[49] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[50] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[51] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[52] Peter-Pike J. Sloan, Charles F. Rose, and Michael F. Cohen. Shape by example. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, pages 135-143, New York, NY, USA, 2001. Association for Computing Machinery. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.89, + 0.747 + ], + "angle": 0, + "content": "[53] TheCaptury. The Captury. http://www.thecapture.com/, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cirmrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1386" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "rithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[56] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, page 1–19, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[57] Xiaohuan Corina Wang and Cary Phillips. Multi-weight enveloping: Least-squares approximation techniques for skin animation. In Proceedings of the 2002 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 129–138, New York, NY, USA, 2002. Association for Computing Machinery. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[58] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects, pages 63-71, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[59] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. ACM Trans. on Graphics, 39, 2020. 1, 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[60] Ji Yang, Xinxin Zuo, Sen Wang, Zhenbo Yu, Xingyu Li, Bingbing Ni, Minglun Gong, and Li Cheng. Object wakeup: 3d object rigging from a single image. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 311-327. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[61] Xiaosong Yang, Arun Somasekharan, and Jian Jun Zhang. Curve skeleton skinning for human and creature characters. Computer Animation and Virtual Worlds, 17, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1387" + } + ] +] \ No newline at end of file diff --git a/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_origin.pdf b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7879bf802afdddfefc4a248dd9224c7224ceb9fd --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/14057e08-ebbd-4725-9942-632446ff87e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d75907a96fe1a0d148527b80f612d40cb70d182069576021fe6dfcd45fce249 +size 1986670 diff --git a/2024/VINECS_ Video-based Neural Character Skinning/full.md b/2024/VINECS_ Video-based Neural Character Skinning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..124682325c8193b0f9d7ea8220aefba7261f9b70 --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/full.md @@ -0,0 +1,291 @@ +# VINECS: Video-based Neural Character Skinning + +Zhouyingcheng Liao $^{1,2}$ + +Vladislav Golyanik1 + +Marc Habermann1,3 + +Christian Theobalt1,3 + +1Max Planck Institute for Informatics, Saarland Informatics Campus 2The University of Hong Kong + +$^{3}$ Saarbrücken Research Center for Visual Computing, Interaction and AI + +# Abstract + +Rigging and skinning clothed human avatars is a challenging task and traditionally requires a lot of manual work and expertise. Recent methods addressing it either generalize across different characters or focus on capturing the dynamics of a single character observed under different pose configurations. However, the former methods typically predict solely static skinning weights, which perform poorly for highly articulated poses, and the latter ones either require dense 3D character scans in different poses or cannot generate an explicit mesh with vertex correspondence over time. To address these challenges, we propose a fully automated approach for creating a fully rigged character with pose-dependent skinning weights, which can be solely learned from multi-view video. Therefore, we first acquire a rigged template, which is then statically skinned. Next, a coordinate-based MLP learns a skinning weights field parameterized over the position in a canonical pose space and the respective pose. Moreover, we introduce our pose- and view-dependent appearance field allowing us to differentiably render and supervise the posed mesh using multi-view imagery. We show that our approach outperforms state-of-the-art while not relying on dense 4D scans. More details can be found on our project page1. + +# 1. Introduction + +Animating 3D characters is a long-standing and challenging problem in computer graphics and vision. Traditionally, rigging, the process of positioning a sparse skeletal structure inside a dense 3D mesh, and skinning, the process of associating vertices with the skeleton, require large amounts of manual effort by experienced artists. This is not only expensive and time-consuming but also inherently difficult since one set of skinning weights is typically not ideal for all sorts of articulated character poses. + +There are also methods for automated rigging and skinning. Some recent works [33, 34, 41, 59] propose a learning-based solution for obtaining skinning and rigging + +![](images/6086a2dbb211e552a96dea27ad4f1f682c93d54696c16f3a7fdbb66d2cf5e476.jpg) +Figure 1. We propose the first end-to-end trainable method for generating a dense and rigged 3D character mesh including learned pose-dependent skinning weights solely from multiview videos. The results demonstrate that our model can generate visually-plausible geometry without requiring manual user input. + +given a single 3D character mesh. However, they do not consider pose-dependent skinning correctives and lead to the typical appearance artifacts. Therefore, other techniques focus on pose-dependent skinning formulations accounting for pose-dependent variations and deformations [3, 30, 40, 52, 57]. Noteworthy, none of these methods tried to learn pose-dependent skinning weight correctives purely from $2D$ image data. Instead, they typically assume a 3D mesh is given, and most of them solely work on a fixed mesh resolution, which cannot be easily adjusted. + +To overcome these limitations, we propose VINECS, i.e., Video-based NEural Character Skinning. Given solely a multi-view video of a human actor, our method creates a fully rigged, skinned, and textured 3D template (including hands) along with pose-dependent skinning weights while no manual editing is required. Moreover, VINECS enables multi-resolution character skinning since skinning weights can be sampled over a continuous 3D canonical space, which is not limited to any specific mesh resolution. + +For each character, we select one canonical pose frame, for which we extract a dense and textured 3D template mesh using implicit surface reconstruction techniques [55]. Given the template and the respective pose obtained from an off-the-shelf multi-view markerless capture system [53], + +we leverage an auto-rigging and skinning method, Pinocchio [7], for obtaining an initial set of skinning weights. Note that Pinocchio produces only a static set of skinning weights, which leads to the typical geometry artifacts when poses deviate too much from the canonical pose. Therefore, we propose a coordinate-based multi-layer perceptron (MLP)—which takes as input a 3D point on the surface of the template in canonical pose and the skeletal pose—and predicts the pose-dependent skinning weights at that point. Since we want to learn the skinning weights solely from multi-view imagery, we design dedicated losses consisting of a silhouette loss and a rendering loss. Specifically, we find that obtaining a well-behaved convergence is challenging only using a static texture for the rendering loss. Thus, we adopt the rendering formulation with an albedo network and a shadow network, which predict pose- and view-dependent appearance of our 3D character resulting in an improved convergence and skinning weights prediction. In summary, our technical contributions are as follows: + +- An end-to-end trainable method for animation-ready explicit character creation, involving static template generation, rigging, and pose-dependent skinning solely using 2D multi-view videos (as few as 10 views). +- A coordinate-based and pose-dependent skinning formulation, which enables multi-resolution skinning. +- A dedicated character appearance formulation and differentiable rendering enabling weakly supervised learning based on multi-view videos. + +We compare VINECS to the state-of-the-art and our results confirm that our method is a clear step toward accurate and automated creation of animatable 3D characters. + +Potential Impact. The focus of this work is not to model deformable geometry [13, 14, 20, 32] or photorealistic appearance [15, 16, 25, 35, 44, 47, 48] of humans but on learning 3D characters with pose-dependent skinning from 2D observations only. We believe our method can greatly benefit many downstream tasks as many existing works leverage naked body models [36] for space canonicalization and feature assignment, fundamentally limiting them to clothing types of the same topology as the human body. As our rigging and skinning are accurate, easy-to-use, and robust to any type of clothing, our approach can stimulate research on the modeling of geometry and appearance of loose apparel. + +# 2. Related Work + +Static Skinning. Linear Blend Skinning (LBS) [28, 38] linearly blends rigid transformations of each bone to obtain a posed geometry. However, this formulation can lead to a collapse of geometry near the skeleton joints, also called candy-wappers effect. To overcome this, several follow-up works have been proposed using spherical representations [22], sweep surfaces [17], curve-based skeleton parameterizations [61], Dual Quaternions [23], or by optimiz + +ing the centers of rotation [27]. Other works focused more on computing the skinning weights itself by modeling the weight distribution as a heat diffusion process [7], as illumination [58], or by voxelizing the mesh to increase robustness [9, 10]. Later on, deformation editing was also modeled using a Laplacian energy formulation [18], and user-defined splines on the mesh enable intuitive editing of skinning-based deformation behaviors [6]. They all only assume a single geometry and skeleton are given. + +In contrast, data-driven methods assume multiple instances of the same character under different poses or a large collection of rigged and skinned characters. Earlier works [19, 26] requires meshes under different poses and then automatically detect skeleton joints and compute the skinning weights. Recently, NeuroSkinning [34] was proposed as the first learning-based method predicting a static set of skinning weights given the mesh and a corresponding skeleton. RigNet [59] instead jointly predicts the skeleton as well as a set of skinning weights solely given a 3D mesh as input. Therefore, they collected a large dataset of 3D characters including artist-designed skeletons and skinning weights. SkinningNet [41] proposed a two-stream graph convolutional network architecture for improved skinning weight prediction given a mesh and the skeleton. Yang et al. [60] proposed to predict the 3D shape together with the rigging from a single image. Liao et al. [33] introduced a learning-based method which jointly learns to predict skinning weights and transfer poses. + +What all the above methods have in common is that they do not explicitly account for dynamic or pose-dependent effects, such as muscle bulging or coarse cloth deformations. In contrast, our proposed method learns pose-dependent skinning weights, which alleviate typical artifacts that arise when using a static set of skinning weights. Moreover, it can also represent coarse deformations induced by the skeletal pose, such as coarse cloth deformations. + +Dynamic Methods. More closely related to our method are approaches that account for dynamic or pose-dependent effects, which are discussed next. There are works [11] that allow an artist to add pose-dependent effects by painting onto specialized textures or by adding so-called helper bones [42]. Alternatively, Kavan et al. [21] proposed to find the optimal skinning weights by minimizing an elastic energy and adding joint-based deformers. Interestingly, those methods do not assume any dynamic observations of the character while still modeling dynamic effects. However, they either require manual editing [11, 42] or can mostly model elasticity-based dynamics [21]. + +Similar to the static methods, several dynamic methods assume a (large) dataset of character meshes is given. Earlier works [3, 40, 46, 57] account for dynamic effects by capturing a set of range scans depicting the subject in different poses. Given a new pose, a set of pose-dependent blend + +weights is estimated, which linearly combines the set of example poses in the dataset or alternatively uses linear and radial basis functions [52]. However, since the complexity of computation typically linearly scales with the number of example poses in the dataset, those methods cannot scale up to a very large number of examples (e.g., 20000 pose frames), which is in stark contrast to our approach that can deal with such an amount of data. There are also parametric naked human body models [4, 5, 36] learned by applying PCA over a large set of human scans, which can account for different shape- and pose-dependent deformations. In contrast, we are building a highly-detailed rigged and skinned character that can wear (loose) clothing in a completely automated fashion. Li et al. [30] predict static skinning weights and pose-dependent neural blend shapes. They assume a dataset of dense meshes is provided, whereas we purely learn pose-dependent skinning weights from videos. SCANimate [49] also requires dense point clouds, and learns a static set of skinning weights, while pose-dependent shapes are predicted by a pose-aware implicit field. SNARF [8] represents the character as a pose-dependent neural implicit surface in the canonical pose space while the learned skinning weights are pose-agnostic. Again, they require point cloud data while our method can be supervised directly on the video. In summary, none of the above methods that consider pose-dependent effects investigated learning solely from 2D imagery. Thus, we are the first method demonstrating that dynamic effects in the form of pose-dependent skinning weights can be learned only from videos. + +Recently, some works were proposed to skin and rig implicit surface representations learn an implicit human avatar. ARAH [56] learns an implicit human model using neural rendering. However, as it relies on a SMPL template and an inverse root finding to convert from view space to canonical space, it easily fails when the garment deviates much from the body. Importantly, it also does not have the vertex correspondence over time. TAVA [31] avoids these problems by getting rid of a template, which, however, leads to poor geometry recovery mainly due to their density-based geometry representation. While these methods can reconstruct an implicit avatar, they do not learn a high-quality skinning weight, which is important to traditional graphics pipeline. + +# 3. Method + +Given synchronized and calibrated multi-view RGB videos of a human character in motion, our goal is to learn an animatable 3D human model with dynamic skinning, without any manual process, such as meshing, rigging, and skinning. To this end, we first extract the static geometry of the actor from one canonical pose frame (typically depicting the actor in a T-pose) of the recording and automatically compute initial skinning weights (Sec. 3.1). However, the initial skinning weights are calculated only based on + +the static geometry of the human character, and, as studied in earlier works, one set of weights might be ideal for one pose while it can lead to strong artifacts in other poses. Thus, we propose a skinning network, called SkinNet, to predict the dynamic skinning weights as a function of the skeletal pose and 3D coordinates in the canonical space (Sec. 3.2). Given the skeletal pose, the static geometry, and the pose-dependent skinning weights, we leverage Linear Blend Skinning (LBS) to obtain the posed geometry. We then use a differentiable renderer to weakly supervise the skinning weights solely based on multi-view imagery. Thereafter, we propose a novel appearance field (Sec. 3.3), which is composed of an albedo component and a pose-and view-dependent shadow/shading component. In addition, we propose several regularization losses ensuring that the skinning weights are also robust to out-of-distribution poses and that the final animated geometry looks visually plausible (Sec. 3.4). An overview is shown in Fig. 2. + +Data Preparation. For a human subject, we record a multi-view video of the performance using $C$ calibrated and synchronized cameras with a total length of $F$ frames. For all frames $\mathcal{I}_{c,f}$ , we calculate the foreground segmentation using color keying or background matting [51] and compute respective distance transform images $\mathcal{D}_{c,f}$ . Here, $c$ and $f$ refer to the respective camera and frame index. We capture the size of the human skeleton and the motion using markerless motion capture [53]. The skeleton motion is parameterized by a root rotation $\alpha_{f} \in \mathbb{R}^{3}$ , a root translation $\mathbf{t}_f \in \mathbb{R}^3$ , and the joint angles $\rho_f \in \mathbb{R}^D$ . The full pose vector is defined as $\theta_f = [\alpha_f, \mathbf{t}_f, \rho_f]^T \in \mathbb{R}^{3 + 3 + D}$ . + +Notation. To represent a variable in the canonical space, we use a bar over the letter, i.e., $\bar{\mathbf{x}}$ . For any point in canonical and global space, we use $\bar{\mathbf{x}}$ and $\mathbf{x}$ , respectively, and mesh vertices are denoted as $\bar{\mathbf{v}}_i$ and $\mathbf{v}_i$ where $i$ is the $i$ th vertex. Without loss of generality, we assume the frame $f$ is fixed in the following (if not explicitly mentioned otherwise) and omit the subscript for better readability. + +# 3.1. Canonical Character Model + +As the first step, we acquire a canonical model of the actor. Recently, implicit surface and coordinate-based representations [45] have gained considerable attention due to their flexibility and comparably simple integration into learning frameworks, and they were also leveraged in the context of pose-dependent human deformation modeling [8]. Moreover, we compared the reconstruction accuracy on the first frame $f = 0$ between classical multi-view stereo and implicit reconstruction method called NeuS [55] in the supplemental material. Interestingly, NeuS has a much higher level of detail and comparably less noise on the surface. Thus, we use NeuS [55] to reconstruct the mesh in the first frame, also called the canonical frame/space. + +Implicit vs. Explicit Canonical Model. In stark con + +![](images/ca837b81452c319e118eb6c32059b87417436f8fb1c276cb015464647221369d.jpg) +Figure 2. Method overview. Our method, VINECS, is a fully automated solution to template generation, rigging, and pose-dependent skinning solely using multi-view videos. We first recover the skeletal poses using markerless motion capture [53] and a static template mesh using implicit surface reconstruction [55]. An initial skinning is obtained by a heat diffusion process [7]. Since one fixed set of skinning weights can lead to artifacts when significantly changing the character pose, we learn pose-dependent skinning weights using a coordinate-based MLP. An auxiliary appearance field is applied to better supervise the learning guide the training of pose-dependent skinning. Note that we do not intend to learn a detailed appearance and only the components in the blue box are used during inference. + +trast to Chen et al. [8], we run Marching Cubes [37] to convert the implicit surface into an explicit mesh and obtain per-vertex colors from NeuS by setting the viewing direction to the normal direction of the mesh. The reason for choosing an explicit mesh in canonical space is three-fold: 1.) It avoids running the rather slow Marching Cubes algorithm for every potentially posed mesh. 2.) We avoid the backward skinning [8], which boils down to a correspondence search that has to be solved using iterative optimization. 3.) We can maintain correspondence on the surface across poses, which is especially useful when texturing and other editing have to be applied. Thus, we aim for a hybrid strategy for template acquisition, i.e., using the state-of-the-art method for implicit surface reconstruction and then converting it to an explicit mesh. + +Meshing and Initial Skinning. During training, we downsample the template mesh to $N$ vertices where $\bar{\mathbf{v}}_i\in \mathbb{R}^3$ denotes the ith vertex. Given the template mesh and the corresponding posed skeleton, i.e. $\theta_{f}$ , we utilize Pinocchio [7] to obtain an initial set of skinning weights $\mathbf{w}_{\mathrm{init},i}\in \mathbb{R}^{J}$ for each vertex $i$ (see Fig. 4). Here, $J$ denotes the number of skinning joints. Given these weights and a skeletal pose $\theta$ , we can transform the vertex $\bar{\mathbf{v}}_i$ from the canonical space to the posed space using LBS [28]: + +$$ +\mathbf {v} _ {i} = L B S \left(\bar {\mathbf {v}} _ {i}, \mathbf {w} _ {\text {i n i t}, i}, \theta\right): \mathbb {R} ^ {3} \times \mathbb {R} ^ {J} \times \mathbb {R} ^ {D} \rightarrow \mathbb {R} ^ {3}. \tag {1} +$$ + +Mesh Parsing. Since different body parts and materials (skin vs. clothing) have different deformation behaviors, we compute per-vertex human parsing labels. We render the 3D model into all camera views and apply a 2D human parsing approach [29] on each view. We then perform max-voting to obtain the per-vertex human parsing labels. Note that so far, we obtained a segmented, rigged, and statically skinned character model without any manual user input. + +# 3.2. Pose-dependent Skinning Field + +So far, the initial skinning weights obtained in Sec. 3.1 have two main limitations: 1.) They are solely computed on a single static pose leading to artifacts such as candy wrappers and mesh distortions if the new pose significantly differs from the canonical one. 2.) The set of weights is computed per-vertex. Thus, the resolution of the mesh has to be fixed beforehand and any change in the mesh requires a new computation of skinning weights or some complicated weight transfer from one mesh to the other. + +To resolve this, we propose SkinNet $f_{\omega}$ , which is a coordinate-based MLP where $\omega$ denotes the trainable network weights. Given any 3D point $\bar{\mathbf{x}}$ in the canonical space, $f_{\omega}$ predicts its skinning weight $\mathbf{w}$ conditioned on the normalized human pose $\tilde{\theta}$ , which can be formulated as: + +$$ +\mathbf {w} _ {\tilde {\theta}} = f _ {\omega} (\bar {\mathbf {x}}, \tilde {\boldsymbol {\theta}}): \mathbb {R} ^ {3} \times \mathbb {R} ^ {D} \rightarrow \mathbb {R} ^ {J}, \tag {2} +$$ + +where $\bar{\mathbf{x}}$ and $\tilde{\theta}$ are concatenated before being fed into the MLP. Here, the normalized pose $\tilde{\theta}$ is obtained by neglecting the global translation as well as the yaw angle, i.e., the rotation around the spine. The intuition behind is that the skinning weights do not depend on where in global 3D space the person performs a respective pose, and, similarly, the skinning weight prediction should be agnostic to which direction the person is facing. Thus, we mask those degrees of freedom before feeding the pose vector into the network. To obtain the transformed point $\mathbf{x}$ in posed space, one can insert Eq. (2) into Eq. (1) resulting in: + +$$ +\mathbf {x} = L B S (\bar {\mathbf {x}}, f _ {\omega} (\bar {\mathbf {x}}, \tilde {\theta}), \theta). \tag {3} +$$ + +Revisiting the drawbacks of initial skinning discussed above, the SkinNet now supports pose-dependent skinning + +weights such that the artifacts of static skinning can be reduced. Moreover, since SkinNet is a coordinate-based MLP, it supports arbitrary mesh resolution since each vertex is queried independently, which is in stark contrast to an architecture that outputs all skinning weights at once and where the final tensor shape is fixed to a specific mesh resolution. + +# 3.3. Auxiliary Appearance Field + +We can now pose the geometry with our pose-dependent skinning weight formulation. However, since we want to supervise SkinNet solely on multi-view images, we also have to model the appearance of the actor. A naive choice would be to just leverage the static vertex colors acquired from NeuS. However, we found that this texture contains baked-in shadows and wrinkles, and cannot account for the pose- and view-dependent appearance changes harming the training of the SkinNet. We also refer to our ablation studies in Sec. 4.3. To address this issue, we propose an auxiliary appearance field, which consists of an albedo field predicting a pose- and view-agnostic color for a given point $\bar{\mathbf{x}}$ in the canonical space and a shadow field predicting the pose- and view-dependent shadow/shading properties of $\bar{\mathbf{x}}$ . Important to note here is that we are not interested in creating a highly photorealistic appearance model of the human, but we mainly use it as an auxiliary tool to better supervise the learning of the skinning weight network. + +More specifically, the albedo field, referred to as AlbedoNet, is an MLP that predicts a static albedo value $a_{\omega'}(\bar{\mathbf{x}}) = \mathbf{a} \in \mathbb{R}^3$ for a given 3D sample point in the canonical space. Here, $\omega'$ are the trainable weights of the albedo network. Moreover, the shadow field is also parameterized by an MLP $r_{\omega''}(\bar{\mathbf{x}}, \theta, \mathbf{n}, \mathbf{d}) = s \in \mathbb{R}^+$ , called ShadowNet, which predicts a scalar multiplier accounting for color changes due to shadows and shading. Note that we assume that such effects happen uniformly across all color channels and, thus, we are only predicting a single scalar instead of a scaling factor per color channel. In addition to taking the point in canonical space, the shadow field also takes the skeletal pose $\theta$ as input as well as the surface normal $\mathbf{n}$ of $\mathbf{x}$ in global space and the viewing direction $\mathbf{d}$ , which allows to potentially model pose- and view-dependent lighting effects such as shadows and shading. $\omega''$ denotes the respective network weights. Details about the architectures are provided in the supplemental document. The final color $\mathbf{c} \in \mathbb{R}^3$ for point $\mathbf{x}$ is computed as + +$$ +\mathbf {c} = a _ {\omega^ {\prime}} (\bar {\mathbf {x}}) r _ {\omega^ {\prime \prime}} (\bar {\mathbf {x}}, \theta , \mathbf {n}, \mathbf {d}). \tag {4} +$$ + +# 3.4. Multi-view Video-based Supervision + +Next, we describe our dedicated supervision strategy that solely requires multi-view imagery. We first introduce the individual loss terms and then describe our training strategy. Importantly, we can obtain a posed template mesh + +$\mathbf{V} \in \mathbb{R}^{N \times 3}$ and the pose- and view-dependent per-vertex colors $\mathbf{C}_c \in \mathbb{R}^{N \times 3}$ by evaluating Eq. (3) and Eq. (4) for all vertices $\bar{\mathbf{v}}_i$ and views $c$ . Moreover, the following losses that are applied to the deformed geometry and its appearance can directly backpropagate into the respective networks since our formulation is fully differentiable. + +Silhouette Loss. This loss [15] ensures that the projected posed geometry projects onto the foreground masks for all views (see supp. material). Since our pose-dependent skinning field Eq. (3) is differentiable with respect to the SkinNet weights $\omega$ , the silhouette loss can provide supervision and learn pose-dependent skinning by matching the posed model silhouette and the image silhouettes. + +Rendering Loss. Although the silhouette loss ensures that the posed geometry matches the multi-view silhouettes, it cannot supervise drifts on the visual hull. Supervision directly from the RGB images can help with resolving this issue since appearance features can constrain drifts in the image plane. Thus, we introduce the rendering loss + +$$ +\mathcal {L} _ {\text {r e n d}} = \sum_ {c = 1} ^ {C} \| \Pi_ {c} (\mathbf {V}, \mathbf {C} _ {c}) - \mathcal {I} _ {c} \| _ {1}, \tag {5} +$$ + +which leverages a differentiable renderer $\Pi_c$ based on the one of Habermann et al. [15] to render the posed and colored model into all camera views $c$ that is then compared to the ground truth image $\mathcal{I}_c$ . Importantly, this loss can supervise both the pose-dependent skinning field as well as the pose- and view-dependent appearance field. + +Laplacian Loss. Since the data terms alone can still lead to degenerate results, we also introduce some regularization on the geometry and the skinning weights. First, we regularize the posed geometry with a Laplacian loss + +$$ +\mathcal {L} _ {\mathrm {l a p}} = \sum_ {i = 1} ^ {N} \| \mathbf {v} _ {i} - \frac {1}{| \mathcal {N} _ {i} |} \sum_ {k \in \mathcal {N} _ {i}} \mathbf {v} _ {k} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\mathcal{N}_i$ is the indices of one-ring neighborhood of vertex $i$ . This regularizer ensures locally smooth geometry. + +**Skinning Regularization.** We found that SkinNet can overfit the training poses and respective deformations by predicting non-local skinning weights, e.g., the foot joint has a non-zero skinning weight for vertices on the shoulder. This, however, harms generalization to novel poses at test time. To prevent this, we regularize the skinning weights by constraining their values based on the geodesic distance $d_{\mathrm{geo}}$ between the vertex and the initial body part it belongs to. This can be formalized with the following loss + +$$ +\mathcal {L} _ {\text {s k i n}} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {J} \mathbf {w} _ {i, j} \left(\frac {\min _ {k \in \mathscr {A} _ {j}} d _ {\text {g e o}} (\bar {\mathbf {v}} _ {i} , \bar {\mathbf {v}} _ {k})}{d _ {\text {g e o m a x}}}\right) ^ {r}, \tag {7} +$$ + +$$ +\mathscr {A} _ {j} = \left\{k \mid j = \underset {t} {\operatorname {a r g m a x}} \mathbf {w} _ {\text {i n i t}, k, t} \right\}, \tag {8} +$$ + +where $\mathcal{A}_j$ is the set of vertices, which have the highest skinning weight with respect to joint $j$ . $r = 3$ is a hyperparameter which controls how penalization increases as the geodesic distance increases. $d_{\mathrm{geomax}}$ is the maximum geodesic distance between any of the vertex pairs. + +Part-wise Regularization. In contrast to the clothed areas, the skin part has rather static skinning. Thus, we use the initial skinning to regularize the predicted skinning weights of the skin part, i.e., the set $\mathcal{G}$ of vertices, which have the parsing label skin. However, even for skin vertices, we found that the initial skinning weights around the joints are not accurate while the rigid parts (i.e., the vertices away from joints) usually have correct initial skinning. We identify the set of rigid vertices as $\mathcal{R} = \{k|\max (\mathbf{w}_{\mathrm{init},k}) > u\}$ where $u = 0.95$ . Thus, our part loss + +$$ +\mathscr {L} _ {\text {p a r t}} = \sum_ {i \in \mathscr {R} \cap \mathscr {G}} \| f _ {\omega} (\bar {\mathbf {v}} _ {i}, \tilde {\boldsymbol {\theta}}) - \mathbf {w} _ {\text {i n i t}, i} \| _ {2} ^ {2}, \tag {9} +$$ + +enforces that the predicted skinning weights are close to the initial ones iff the human parsing label of a vertex is the skin $(\mathcal{G})$ and the vertex belongs to the rigid part $(\mathcal{R})$ . + +Training Strategy. Our training proceeds in four stages. First, we train SkinNet without using the rendering loss $\mathcal{L}_{\mathrm{rend}}$ . This step is required to ensure that the posed mesh roughly overlaps with the foreground masks. Next, the AlbedoNet is trained using only the rendering loss $\mathcal{L}_{\mathrm{rend}}$ while the SkinNet weights $\omega$ are fixed. Importantly, the AlbedoNet is not pose-dependent, but we optimize the weights $\omega'$ across all frames. Afterward, the AlbedoNet and the SkinNet are fixed and we train the ShadowNet using the rendering loss $\mathcal{L}_{\mathrm{rend}}$ . Again we train on all training frames. Lastly, since we now have a pose- and view-dependent appearance model available, we refine the SkinNet weights including the rendering loss. We found this gives the best performance, as validated by our results. + +# 4. Results + +Dataset. We evaluate our method on 3 subjects wearing tight and loose clothing, i.e., pants or skirts. Two of them are from the DynaCap dataset [15] and they are referred as $D2$ and $D5$ . $D2$ is wearing short pants and a T-shirt while $D5$ is in a skirt. The total numbers of cameras are 94 and 101 for those two sequences and our method can be trained using as few as 10 cameras. There are around 19000 training frames per subject depicting the actor in various articulated poses. For evaluation, there is a separate testing set of around 7000 frames per subject. For convenience, we keep 1 out of 10 frames for training and evaluation. Additionally, we captured a new subject, $V6$ , wearing a T-shirt and trousers using 116 cameras. We captured a separate training (17730 frames) and testing sequence (5000 frames). Importantly, for this sequence, we recover hand gestures and a respective 3D model including hands. + +Metrics. For evaluation, we reconstruct the dense 3D geometry for the testing frames using multi-view stereo reconstruction [2]. We do not use NeuS here as it is very time-consuming. Those dense meshes serve as ground truth scans and we compare the posed meshes recovered by our method as well as competing methods in terms of the Chamfer distance and the Hausdorff distance going from the recovered mesh to the ground truth scan (M2S) and vice versa (S2M). We report the average results over all testing frames. All metrics are reported in centimeter scale. + +Competing Methods. We compare our work to our re-implementation of Pinocchio [7], which proposed a heat diffusion process to obtain a static set of skinning weights. Moreover, we compare to SCANimate [49] and SNARF [8], which, in stark contrast to our work, require dense point clouds for training. They learn static skinning weights and a pose-aware implicit shape. Since our approach mostly focuses on skinning, we compare to SCANimate with and without pose-aware shape, referred to as SCANimate and SCANimate* in the following. Finally, we compare to RigNet [59], which can jointly predict a rigging skeleton and static skinning weights given a template. Unfortunately, SkinningNet [41] and NeuroSkinning [34] do not provide their code and, thus, we cannot compare to them. For more details and comparisons to other (slightly less) related works [31, 56], we refer to the supplemental document. + +# 4.1. Qualitative Results + +In Fig. 3, we qualitatively evaluate our approach. In the first column, we show the reconstructed static geometry for different subjects. In the following columns, we show the characters in novel poses from the test set overlayed onto the respective images. Note that our posed character nicely matches the reference image, which confirms that our pose-dependent skinning indeed generalizes to novel poses and results in accurately posed meshes. In Fig. 4, we further visualize the effect of the pose-dependent skinning weights obtained by our approach. The initial skinning weights are fixed across poses leading to the typical skinning artifacts and wrong deformation behavior. In contrast, our proposed pose-dependent skinning weights adapt according to the pose and as a result the deformations look more natural. + +# 4.2. Comparison + +In Tab. 1, we compare our results to previous works. Note that we achieve the most accurate results when compared to works only focusing on character skinning. This validates the necessity of pose-dependent skinning weights since other works rely on static ones. In this table, SCAN-imate and SNARF predict pose-aware shapes while assuming dense point clouds are given. Since it is not the scope of this work to predict pose-dependent non-rigid deformations and our method solely learns from video data, we placed + +![](images/719713ab2f42937e14871c69ffd912807906a8f24c90a9b142a6bc4127eab9a4.jpg) +Figure 3. Qualitative results. From left to right: Recovered static character template. Reference image showing a test pose. Our posed character overlayed onto the reference image. The precise overlay of the posed geometry confirms the effectiveness of our approach. + +
Quantitative Geometry Comparison
SubjectD2D5V6
MethodChamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓
Pinnochio [7]3.7602.1621.5995.0772.8112.2673.3581.8711.487
SCANimate* [49]3.7502.0991.6505.4532.9652.4883.5021.8901.612
RigNet [59]3.5992.0781.5214.9892.8362.1533.3691.8941.475
Ours3.0341.7461.2884.5122.4422.0702.9931.7191.274
SCANimate [49]2.8421.6461.1964.9822.2842.6983.1541.7141.440
SNARF [8]2.5911.4521.1394.3202.2882.0322.7601.4961.263
+ +![](images/4b7bfe1688536b0c2107190fe42e6468994f3c962a4c90860a9099dfaf15cc45.jpg) +Figure 4. Qualitative results showing the pose-dependent skinning weights. As a reference, we show the result using the initial static skinning weights. Notably, using static skinning weights leads to artifacts and worse deformation. When using our pose-dependent skinning, one can see a clear improvement in the final result and we also show how the weights change over different poses. + +them separately. Nonetheless, it is worth noting that our approach is close to SCANImate and SNARF in terms of accuracy and better than SCANImate for subject $D5$ and $V6$ even though our training does not require point clouds. + +We also visually compare to those works in Fig. 5. We can see qualitatively our method achieves a lower error especially in joint regions where most of the skinning artifacts + +Table 1. We compare our method to previous works on 3 subjects. Our method obtains better results than other skinning-based approaches. It is even close to SCANimate and SNARF, which require dense pointclouds while we solely use multi-view video. Importantly, they focus on learning pose-aware shapes while our goal is to obtain a fully rigged and skinned character in a completely automated fashion. + +
Ablation Study on D2
MethodChamfer↓M2S↓S2M↓
Initial weights3.7612.1621.599
Static weights3.3541.9351.418
w/o Lrend3.1161.7831.333
w/ albedo-only3.0471.7501.297
w/ NeuS appearance3.1521.7861.365
w/ NeuS albedo3.1371.7951.342
w/ displacements3.1911.8051.386
Ours (10 views)3.1291.7911.337
Ours3.0341.7461.288
+ +Table 2. We compare our method against baselines. All designs improve the accuracy confirming the effectiveness of VINECS. + +become visible. This can be nicely seen in the error maps, which visualize the per-vertex error. Although SCAnimate and SNARF perform well on naked body and tight clothes regions, they are worse than ours around loose clothing, as they rely on SMPL while our method does not have such a restriction. Further, the posed meshes clearly show fewer artifacts for our approach compared to others. Note that the entire character including the geometry, rigged skeleton, and the pose-dependent skinning can be obtained solely from multi-view video without any manual user input. + +![](images/7ebc3df62fe87c9fcb61b4eb00b466dfd019df2bef5405e00db6ebc93889705f.jpg) +Figure 5. Qualitative comparison. For each method, we visualize the recovered posed geometry as well as the per-vertex error map when comparing the ground truth in terms of M2S. Note that our method consistently shows the lowest error and also has the least visual artifacts. + +![](images/147be2e9b316f2464c834dd70fcdb6a9eb6118b251bc3dfa1049bb34874b0779.jpg) +Figure 6. We compare the initial skinning, the static skinning optimized using our supervision scheme, ablating the rendering loss, and our pose-dependent skinning. We outperform all baselines. + +# 4.3. Ablation + +Next, we evaluate our design choices in Tab. 2. First, we evaluate the effect of pose-dependent skinning weights. Therefore, we compare our final result to the initial skinning, which we obtain by leveraging Pinocchio [7] and a version of our method where we use our supervision scheme, but optimize a static set of skinning weights, i.e., the weights are fixed across all poses. From the results, we can see that pose-dependent weights significantly improve the accuracy over the two baselines proofing the effectiveness of our formulation. This is also visualized in Fig. 6. + +Another key aspect when learning those weights solely from video data is our proposed supervision scheme. When we do not employ the dense rendering loss or soley use the albedo without shadow estimation, the performance drops. Moreover, we also evaluate the accuracy when using a dense rendering loss with the static appearance from NeuS and using NeuS appearance as albedo while the shadow network is learned. We found that our proposed rendering scheme performs best. To validate the effectiveness of training pose-dependent skinning, we replace the SkinNet by a network predicting per-vertex displacement, which generalizes worse than SkinNet. We train our model with sparse cameras (10 views), and the performance is still satisfying. + +# 5. Discussion and Conclusion + +Limitations. Although we propose the first method for learning pose-dependent skinning solely from multi-view video, our method still has some limitations, which opens up interesting directions for future work. Currently, we query the SkinNet for every vertex in canonical space, which may be inefficient when many points are sampled. Thus, future work could involve exploring more efficient architectures like hashgrids [43]. While hand and body can be faithfully animated, there is no control over the facial expression. Thus, extending our automated character creation pipeline with some parametric model for facial expressions might be a promising direction. Also, a joint consideration of rigging, skinning, and pose tracking may further improve the quality of the 3D animation-ready characters. Lastly, some artifacts still remain since skinning alone cannot account for high-frequency deformations. Combining VINECS with downstream tasks such as human surface tracking/modeling could alleviate such artifacts. + +Conclusion. We proposed the first method for creating a fully rigged and skinned character solely from multi-view video without any manual processing. To this end, we first acquire a canonical template using an implicit surface reconstruction, automated initial skinning and mesh parsing. Most importantly, we learn a pose-dependent skinning field from a multi-view video, which greatly reduces artifacts. We show how such a skinning field can be learned in a weakly supervised manner by an appearance model and differentiable rendering. Our work demonstrates that animatable character creation can be fully automated while maintaining high-quality geometry, rigging, and skinning. + +Acknowledgements. We thank Heming Zhu and Diogo Luvizon for their help. This project was also supported by Saarbrücken Research Center for Visual Computing, Interaction, and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784). + +# References + +[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 2 +[2] Agisoft. PhotoScan. http://www.agisoft.com, 2016. 6, 1, 3 +[3] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. ACM Trans. Graph., 21(3):612-619, 2002. 1, 2 +[4] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a Correlated Model of Identity and Pose-Dependent Body Shape Variation for Real-Time Synthesis. In ACM SIGGRAPH / Eurographics Symposium on Computer Animation. The Eurographics Association, 2006. 3 +[5] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. ACM Transactions on Graphics, 24(3):408-416, 2005. 3 +[6] Seungbae Bang and Sung-Hee Lee. Spline interface for intuitive skinning weight editing. ACM Trans. Graph., 37(5), 2018. 2 +[7] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Trans. Graph., 26(3), 2007, 2, 4, 6, 7, 8, 3 +[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 3, 4, 6, 7 +[9] Olivier Dionne and Martin de Lasa. Geodesic voxel binding for production character meshes. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 173-180, New York, NY, USA, 2013. Association for Computing Machinery. 2 +[10] Olivier Dionne and Martin de Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE Transactions on Visualization and Computer Graphics, 20 (10):1367-1378, 2014. 2 +[11] Sven Forstmann, Jun Ohya, Artus Krohn-Grimberghe, and Ryan McDougall. Deformation styles for spline-based skeletal animation. In Proceedings of the 2007 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 141-150, Goslar, DEU, 2007. Eurographics Association. 2 +[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning + +shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 1 +[13] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2 +[14] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 1:1, 2020. 2 +[15] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Trans. Graph., 40(4), 2021. 2, 5, 6, 1 +[16] Marc Habermann, Lingjie Liu, Weipeng Xu, Gerard Pons-Moll, Michael Zollhoefer, and Christian Theobalt. Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech., 6(3), 2023. 2 +[17] Dae-Eun Hyun, Seung-Hyun Yoon, Jung-Woo Chang, Joon-Kyung Seong, Myung-Soo Kim, and Bert Jüttler. Sweep-based human deformation. The Visual Computer, 21:542-550, 2005. 2 +[18] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine-Hornung. Bounded biharmonic weights for real-time deformation. Commun. ACM, 57(4):99-106, 2014. 2 +[19] Doug L. James and Christopher D. Twigg. Skinning mesh animations. ACM Trans. Graph., 24(3):399-407, 2005. 2 +[20] Yue Jiang, Marc Habermann, Vladislav Golyanik, and Christian Theobalt. Hifecap: Monocular high-fidelity and expressive capture of human performances. In BMVC, 2022. 2 +[21] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Trans. Graph., 31 (6), 2012. 2 +[22] Ladislav Kavan and Jiří Žára. Spherical blend skinning: A real-time deformation of articulated models. In Proceedings of the 2005 Symposium on Interactive 3D Graphics and Games, page 9–16, New York, NY, USA, 2005. Association for Computing Machinery. 2 +[23] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46. ACM, 2007. 2 +[24] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2014. 3 +[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems, 2023. 2 +[26] Binh Huy Le and Zhigang Deng. Robust and accurate skeletal rigging from mesh sequences. ACM Trans. Graph., 33(4), 2014. 2 +[27] Binh Huy Le and Jessica K. Hodgins. Real-time skeletal skinning with optimized centers of rotation. ACM Trans. Graph., 35(4), 2016. 2 + +[28] J. P. Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 165-172, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2, 4 +[29] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4, 1 +[30] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 1, 3 +[31] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. 2022. 3, 6, 1, 2 +[32] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. 2020. 2 +[33] Zhouyingcheng Liao, Jimei Yang, Jun Saito, Gerard Pons-Moll, and Yang Zhou. Skeleton-free pose transfer for stylized 3d characters. In European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2 +[34] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Trans. Graph., 38(4), 2019. 1, 2, 6 +[35] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6), 2021. 2 +[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3 +[37] William Lorensen and Harvey Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21:163-, 1987. 4, 3 +[38] N. Magnenat-Thalmann, A. Laperri'ere, and D. Thalmann. Joint-dependent local deformations for hand animation and object grasping. In Proceedings of Graphics Interface '88, pages 26-33. Canadian Man-Computer Communications Society, 1988. 2 +[39] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 1 +[40] Alex Mohr and Michael Gleicher. Building efficient, accurate character skins from examples. ACM Trans. Graph., 22 (3):562-568, 2003. 1, 2 +[41] A. Mosella-Montoro and J. Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18572-18581, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1, 2, 6 + +[42] Tomohiko Mukai and Shigeru Kuriyama. Efficient dynamic skinning with low-rank helper bone controllers. ACM Trans. Graph., 35(4), 2016. 2 +[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 8 +[44] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. 2023. 2 +[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[46] Sang II Park and Jessica K. Hodgins. Capturing and animating skin deformation in human motion. ACM Trans. Graph., 25(3):881-889, 2006. 2 +[47] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Animatable neural radiance fields for human body modeling. ICCV, 2021. 2 +[48] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 1(1): 9054-9063, 2021. 2 +[49] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7 +[50] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. Advances in neural information processing systems, 29, 2016. 1 +[51] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[52] Peter-Pike J. Sloan, Charles F. Rose, and Michael F. Cohen. Shape by example. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, pages 135-143, New York, NY, USA, 2001. Association for Computing Machinery. 1, 3 +[53] TheCaptury. The Captury. http://www.thecapture.com/, 2020. 1, 3, 4 +[54] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cirmrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algo + +rithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. 3 +[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 1, 3, 4 +[56] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, page 1–19, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 6, 1, 2 +[57] Xiaohuan Corina Wang and Cary Phillips. Multi-weight enveloping: Least-squares approximation techniques for skin animation. In Proceedings of the 2002 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 129–138, New York, NY, USA, 2002. Association for Computing Machinery. 1, 2 +[58] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects, pages 63-71, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2 +[59] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. ACM Trans. on Graphics, 39, 2020. 1, 2, 6, 7, 3 +[60] Ji Yang, Xinxin Zuo, Sen Wang, Zhenbo Yu, Xingyu Li, Bingbing Ni, Minglun Gong, and Li Cheng. Object wakeup: 3d object rigging from a single image. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 311-327. Springer, 2022. 2 +[61] Xiaosong Yang, Arun Somasekharan, and Jian Jun Zhang. Curve skeleton skinning for human and creature characters. Computer Animation and Virtual Worlds, 17, 2006. 2 \ No newline at end of file diff --git a/2024/VINECS_ Video-based Neural Character Skinning/images.zip b/2024/VINECS_ Video-based Neural Character Skinning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..49db096076b31a9e63b6636d80d086d8405565e3 --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ec9b317ff6d418dfc083880bb2ffb3d7f9f65478aa71c2441bd12c4f100a3eb +size 466273 diff --git a/2024/VINECS_ Video-based Neural Character Skinning/layout.json b/2024/VINECS_ Video-based Neural Character Skinning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1be579915d0a0e54b81bdb02157075fb59986b0 --- /dev/null +++ b/2024/VINECS_ Video-based Neural Character Skinning/layout.json @@ -0,0 +1,8040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 452, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 452, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 452, + 121 + ], + "type": "text", + "content": "VINECS: Video-based Neural Character Skinning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 141, + 168, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 168, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 168, + 157 + ], + "type": "text", + "content": "Zhouyingcheng Liao" + }, + { + "bbox": [ + 53, + 141, + 168, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 143, + 291, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 143, + 291, + 157 + ], + "spans": [ + { + "bbox": [ + 193, + 143, + 291, + 157 + ], + "type": "text", + "content": "Vladislav Golyanik1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 142, + 415, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 142, + 415, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 142, + 415, + 156 + ], + "type": "text", + "content": "Marc Habermann1,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 440, + 142, + 543, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 142, + 543, + 156 + ], + "spans": [ + { + "bbox": [ + 440, + 142, + 543, + 156 + ], + "type": "text", + "content": "Christian Theobalt1,3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 159, + 496, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 159, + 496, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 159, + 496, + 171 + ], + "type": "text", + "content": "1Max Planck Institute for Informatics, Saarland Informatics Campus 2The University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 173, + 430, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 173, + 430, + 185 + ], + "spans": [ + { + "bbox": [ + 167, + 173, + 430, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 167, + 173, + 430, + 185 + ], + "type": "text", + "content": "Saarbrücken Research Center for Visual Computing, Interaction and AI" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "content": "Rigging and skinning clothed human avatars is a challenging task and traditionally requires a lot of manual work and expertise. Recent methods addressing it either generalize across different characters or focus on capturing the dynamics of a single character observed under different pose configurations. However, the former methods typically predict solely static skinning weights, which perform poorly for highly articulated poses, and the latter ones either require dense 3D character scans in different poses or cannot generate an explicit mesh with vertex correspondence over time. To address these challenges, we propose a fully automated approach for creating a fully rigged character with pose-dependent skinning weights, which can be solely learned from multi-view video. Therefore, we first acquire a rigged template, which is then statically skinned. Next, a coordinate-based MLP learns a skinning weights field parameterized over the position in a canonical pose space and the respective pose. Moreover, we introduce our pose- and view-dependent appearance field allowing us to differentiably render and supervise the posed mesh using multi-view imagery. We show that our approach outperforms state-of-the-art while not relying on dense 4D scans. More details can be found on our project page1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 543, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 650 + ], + "type": "text", + "content": "Animating 3D characters is a long-standing and challenging problem in computer graphics and vision. Traditionally, rigging, the process of positioning a sparse skeletal structure inside a dense 3D mesh, and skinning, the process of associating vertices with the skeleton, require large amounts of manual effort by experienced artists. This is not only expensive and time-consuming but also inherently difficult since one set of skinning weights is typically not ideal for all sorts of articulated character poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "content": "There are also methods for automated rigging and skinning. Some recent works [33, 34, 41, 59] propose a learning-based solution for obtaining skinning and rigging" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 212, + 527, + 363 + ], + "blocks": [ + { + "bbox": [ + 326, + 212, + 527, + 363 + ], + "lines": [ + { + "bbox": [ + 326, + 212, + 527, + 363 + ], + "spans": [ + { + "bbox": [ + 326, + 212, + 527, + 363 + ], + "type": "image", + "image_path": "6086a2dbb211e552a96dea27ad4f1f682c93d54696c16f3a7fdbb66d2cf5e476.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 364, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 304, + 364, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 547, + 419 + ], + "type": "text", + "content": "Figure 1. We propose the first end-to-end trainable method for generating a dense and rigged 3D character mesh including learned pose-dependent skinning weights solely from multiview videos. The results demonstrate that our model can generate visually-plausible geometry without requiring manual user input." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 422, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 542 + ], + "type": "text", + "content": "given a single 3D character mesh. However, they do not consider pose-dependent skinning correctives and lead to the typical appearance artifacts. Therefore, other techniques focus on pose-dependent skinning formulations accounting for pose-dependent variations and deformations [3, 30, 40, 52, 57]. Noteworthy, none of these methods tried to learn pose-dependent skinning weight correctives purely from " + }, + { + "bbox": [ + 304, + 422, + 547, + 542 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 304, + 422, + 547, + 542 + ], + "type": "text", + "content": " image data. Instead, they typically assume a 3D mesh is given, and most of them solely work on a fixed mesh resolution, which cannot be easily adjusted." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 544, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 652 + ], + "type": "text", + "content": "To overcome these limitations, we propose VINECS, i.e., Video-based NEural Character Skinning. Given solely a multi-view video of a human actor, our method creates a fully rigged, skinned, and textured 3D template (including hands) along with pose-dependent skinning weights while no manual editing is required. Moreover, VINECS enables multi-resolution character skinning since skinning weights can be sampled over a continuous 3D canonical space, which is not limited to any specific mesh resolution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "For each character, we select one canonical pose frame, for which we extract a dense and textured 3D template mesh using implicit surface reconstruction techniques [55]. Given the template and the respective pose obtained from an off-the-shelf multi-view markerless capture system [53]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "1https://people.mpi-inf.mpg.de/~mhaberma/projects/2023-Vinecs" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1377" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "we leverage an auto-rigging and skinning method, Pinocchio [7], for obtaining an initial set of skinning weights. Note that Pinocchio produces only a static set of skinning weights, which leads to the typical geometry artifacts when poses deviate too much from the canonical pose. Therefore, we propose a coordinate-based multi-layer perceptron (MLP)—which takes as input a 3D point on the surface of the template in canonical pose and the skeletal pose—and predicts the pose-dependent skinning weights at that point. Since we want to learn the skinning weights solely from multi-view imagery, we design dedicated losses consisting of a silhouette loss and a rendering loss. Specifically, we find that obtaining a well-behaved convergence is challenging only using a static texture for the rendering loss. Thus, we adopt the rendering formulation with an albedo network and a shadow network, which predict pose- and view-dependent appearance of our 3D character resulting in an improved convergence and skinning weights prediction. In summary, our technical contributions are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 287, + 408 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 300, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 346 + ], + "type": "text", + "content": "- An end-to-end trainable method for animation-ready explicit character creation, involving static template generation, rigging, and pose-dependent skinning solely using 2D multi-view videos (as few as 10 views)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 348, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 372 + ], + "type": "text", + "content": "- A coordinate-based and pose-dependent skinning formulation, which enables multi-resolution skinning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "type": "text", + "content": "- A dedicated character appearance formulation and differentiable rendering enabling weakly supervised learning based on multi-view videos." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 409, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 287, + 444 + ], + "type": "text", + "content": "We compare VINECS to the state-of-the-art and our results confirm that our method is a clear step toward accurate and automated creation of animatable 3D characters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 445, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 590 + ], + "type": "text", + "content": "Potential Impact. The focus of this work is not to model deformable geometry [13, 14, 20, 32] or photorealistic appearance [15, 16, 25, 35, 44, 47, 48] of humans but on learning 3D characters with pose-dependent skinning from 2D observations only. We believe our method can greatly benefit many downstream tasks as many existing works leverage naked body models [36] for space canonicalization and feature assignment, fundamentally limiting them to clothing types of the same topology as the human body. As our rigging and skinning are accurate, easy-to-use, and robust to any type of clothing, our approach can stimulate research on the modeling of geometry and appearance of loose apparel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Static Skinning. Linear Blend Skinning (LBS) [28, 38] linearly blends rigid transformations of each bone to obtain a posed geometry. However, this formulation can lead to a collapse of geometry near the skeleton joints, also called candy-wappers effect. To overcome this, several follow-up works have been proposed using spherical representations [22], sweep surfaces [17], curve-based skeleton parameterizations [61], Dual Quaternions [23], or by optimiz" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "ing the centers of rotation [27]. Other works focused more on computing the skinning weights itself by modeling the weight distribution as a heat diffusion process [7], as illumination [58], or by voxelizing the mesh to increase robustness [9, 10]. Later on, deformation editing was also modeled using a Laplacian energy formulation [18], and user-defined splines on the mesh enable intuitive editing of skinning-based deformation behaviors [6]. They all only assume a single geometry and skeleton are given." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 182, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 410 + ], + "type": "text", + "content": "In contrast, data-driven methods assume multiple instances of the same character under different poses or a large collection of rigged and skinned characters. Earlier works [19, 26] requires meshes under different poses and then automatically detect skeleton joints and compute the skinning weights. Recently, NeuroSkinning [34] was proposed as the first learning-based method predicting a static set of skinning weights given the mesh and a corresponding skeleton. RigNet [59] instead jointly predicts the skeleton as well as a set of skinning weights solely given a 3D mesh as input. Therefore, they collected a large dataset of 3D characters including artist-designed skeletons and skinning weights. SkinningNet [41] proposed a two-stream graph convolutional network architecture for improved skinning weight prediction given a mesh and the skeleton. Yang et al. [60] proposed to predict the 3D shape together with the rigging from a single image. Liao et al. [33] introduced a learning-based method which jointly learns to predict skinning weights and transfer poses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": "What all the above methods have in common is that they do not explicitly account for dynamic or pose-dependent effects, such as muscle bulging or coarse cloth deformations. In contrast, our proposed method learns pose-dependent skinning weights, which alleviate typical artifacts that arise when using a static set of skinning weights. Moreover, it can also represent coarse deformations induced by the skeletal pose, such as coarse cloth deformations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "type": "text", + "content": "Dynamic Methods. More closely related to our method are approaches that account for dynamic or pose-dependent effects, which are discussed next. There are works [11] that allow an artist to add pose-dependent effects by painting onto specialized textures or by adding so-called helper bones [42]. Alternatively, Kavan et al. [21] proposed to find the optimal skinning weights by minimizing an elastic energy and adding joint-based deformers. Interestingly, those methods do not assume any dynamic observations of the character while still modeling dynamic effects. However, they either require manual editing [11, 42] or can mostly model elasticity-based dynamics [21]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Similar to the static methods, several dynamic methods assume a (large) dataset of character meshes is given. Earlier works [3, 40, 46, 57] account for dynamic effects by capturing a set of range scans depicting the subject in different poses. Given a new pose, a set of pose-dependent blend" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "1378" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "content": "weights is estimated, which linearly combines the set of example poses in the dataset or alternatively uses linear and radial basis functions [52]. However, since the complexity of computation typically linearly scales with the number of example poses in the dataset, those methods cannot scale up to a very large number of examples (e.g., 20000 pose frames), which is in stark contrast to our approach that can deal with such an amount of data. There are also parametric naked human body models [4, 5, 36] learned by applying PCA over a large set of human scans, which can account for different shape- and pose-dependent deformations. In contrast, we are building a highly-detailed rigged and skinned character that can wear (loose) clothing in a completely automated fashion. Li et al. [30] predict static skinning weights and pose-dependent neural blend shapes. They assume a dataset of dense meshes is provided, whereas we purely learn pose-dependent skinning weights from videos. SCANimate [49] also requires dense point clouds, and learns a static set of skinning weights, while pose-dependent shapes are predicted by a pose-aware implicit field. SNARF [8] represents the character as a pose-dependent neural implicit surface in the canonical pose space while the learned skinning weights are pose-agnostic. Again, they require point cloud data while our method can be supervised directly on the video. In summary, none of the above methods that consider pose-dependent effects investigated learning solely from 2D imagery. Thus, we are the first method demonstrating that dynamic effects in the form of pose-dependent skinning weights can be learned only from videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 419, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 289, + 575 + ], + "type": "text", + "content": "Recently, some works were proposed to skin and rig implicit surface representations learn an implicit human avatar. ARAH [56] learns an implicit human model using neural rendering. However, as it relies on a SMPL template and an inverse root finding to convert from view space to canonical space, it easily fails when the garment deviates much from the body. Importantly, it also does not have the vertex correspondence over time. TAVA [31] avoids these problems by getting rid of a template, which, however, leads to poor geometry recovery mainly due to their density-based geometry representation. While these methods can reconstruct an implicit avatar, they do not learn a high-quality skinning weight, which is important to traditional graphics pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 586, + 104, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 104, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 104, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "content": "Given synchronized and calibrated multi-view RGB videos of a human character in motion, our goal is to learn an animatable 3D human model with dynamic skinning, without any manual process, such as meshing, rigging, and skinning. To this end, we first extract the static geometry of the actor from one canonical pose frame (typically depicting the actor in a T-pose) of the recording and automatically compute initial skinning weights (Sec. 3.1). However, the initial skinning weights are calculated only based on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "content": "the static geometry of the human character, and, as studied in earlier works, one set of weights might be ideal for one pose while it can lead to strong artifacts in other poses. Thus, we propose a skinning network, called SkinNet, to predict the dynamic skinning weights as a function of the skeletal pose and 3D coordinates in the canonical space (Sec. 3.2). Given the skeletal pose, the static geometry, and the pose-dependent skinning weights, we leverage Linear Blend Skinning (LBS) to obtain the posed geometry. We then use a differentiable renderer to weakly supervise the skinning weights solely based on multi-view imagery. Thereafter, we propose a novel appearance field (Sec. 3.3), which is composed of an albedo component and a pose-and view-dependent shadow/shading component. In addition, we propose several regularization losses ensuring that the skinning weights are also robust to out-of-distribution poses and that the final animated geometry looks visually plausible (Sec. 3.4). An overview is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": "Data Preparation. For a human subject, we record a multi-view video of the performance using " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": " calibrated and synchronized cameras with a total length of " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": " frames. For all frames " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{c,f}" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": ", we calculate the foreground segmentation using color keying or background matting [51] and compute respective distance transform images " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{c,f}" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": " refer to the respective camera and frame index. We capture the size of the human skeleton and the motion using markerless motion capture [53]. The skeleton motion is parameterized by a root rotation " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\alpha_{f} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": ", a root translation " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_f \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": ", and the joint angles " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\rho_f \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": ". The full pose vector is defined as " + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\theta_f = [\\alpha_f, \\mathbf{t}_f, \\rho_f]^T \\in \\mathbb{R}^{3 + 3 + D}" + }, + { + "bbox": [ + 304, + 289, + 546, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": "Notation. To represent a variable in the canonical space, we use a bar over the letter, i.e., " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": ". For any point in canonical and global space, we use " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": ", respectively, and mesh vertices are denoted as " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": "th vertex. Without loss of generality, we assume the frame " + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 434, + 547, + 518 + ], + "type": "text", + "content": " is fixed in the following (if not explicitly mentioned otherwise) and omit the subscript for better readability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 527, + 460, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 460, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 460, + 539 + ], + "type": "text", + "content": "3.1. Canonical Character Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": "As the first step, we acquire a canonical model of the actor. Recently, implicit surface and coordinate-based representations [45] have gained considerable attention due to their flexibility and comparably simple integration into learning frameworks, and they were also leveraged in the context of pose-dependent human deformation modeling [8]. Moreover, we compared the reconstruction accuracy on the first frame " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "f = 0" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " between classical multi-view stereo and implicit reconstruction method called NeuS [55] in the supplemental material. Interestingly, NeuS has a much higher level of detail and comparably less noise on the surface. Thus, we use NeuS [55] to reconstruct the mesh in the first frame, also called the canonical frame/space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Implicit vs. Explicit Canonical Model. In stark con" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1379" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 525, + 229 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 525, + 229 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 525, + 229 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 525, + 229 + ], + "type": "image", + "image_path": "ca837b81452c319e118eb6c32059b87417436f8fb1c276cb015464647221369d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. Method overview. Our method, VINECS, is a fully automated solution to template generation, rigging, and pose-dependent skinning solely using multi-view videos. We first recover the skeletal poses using markerless motion capture [53] and a static template mesh using implicit surface reconstruction [55]. An initial skinning is obtained by a heat diffusion process [7]. Since one fixed set of skinning weights can lead to artifacts when significantly changing the character pose, we learn pose-dependent skinning weights using a coordinate-based MLP. An auxiliary appearance field is applied to better supervise the learning guide the training of pose-dependent skinning. Note that we do not intend to learn a detailed appearance and only the components in the blue box are used during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 486 + ], + "type": "text", + "content": "trast to Chen et al. [8], we run Marching Cubes [37] to convert the implicit surface into an explicit mesh and obtain per-vertex colors from NeuS by setting the viewing direction to the normal direction of the mesh. The reason for choosing an explicit mesh in canonical space is three-fold: 1.) It avoids running the rather slow Marching Cubes algorithm for every potentially posed mesh. 2.) We avoid the backward skinning [8], which boils down to a correspondence search that has to be solved using iterative optimization. 3.) We can maintain correspondence on the surface across poses, which is especially useful when texturing and other editing have to be applied. Thus, we aim for a hybrid strategy for template acquisition, i.e., using the state-of-the-art method for implicit surface reconstruction and then converting it to an explicit mesh." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": "Meshing and Initial Skinning. During training, we downsample the template mesh to " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " vertices where " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " denotes the ith vertex. Given the template mesh and the corresponding posed skeleton, i.e. " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": ", we utilize Pinocchio [7] to obtain an initial set of skinning weights " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{init},i}\\in \\mathbb{R}^{J}" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " for each vertex " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " (see Fig. 4). Here, " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " denotes the number of skinning joints. Given these weights and a skeletal pose " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": ", we can transform the vertex " + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i" + }, + { + "bbox": [ + 46, + 487, + 289, + 594 + ], + "type": "text", + "content": " from the canonical space to the posed space using LBS [28]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 598, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 598, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 75, + 598, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} = L B S \\left(\\bar {\\mathbf {v}} _ {i}, \\mathbf {w} _ {\\text {i n i t}, i}, \\theta\\right): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {J} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {3}. \\tag {1}", + "image_path": "dda7849981050dc24eb0216417090744d573ed3ba16c8c7a0036f10c6a7369ed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Mesh Parsing. Since different body parts and materials (skin vs. clothing) have different deformation behaviors, we compute per-vertex human parsing labels. We render the 3D model into all camera views and apply a 2D human parsing approach [29] on each view. We then perform max-voting to obtain the per-vertex human parsing labels. Note that so far, we obtained a segmented, rigged, and statically skinned character model without any manual user input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 307, + 474, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 474, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 474, + 320 + ], + "type": "text", + "content": "3.2. Pose-dependent Skinning Field" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "type": "text", + "content": "So far, the initial skinning weights obtained in Sec. 3.1 have two main limitations: 1.) They are solely computed on a single static pose leading to artifacts such as candy wrappers and mesh distortions if the new pose significantly differs from the canonical one. 2.) The set of weights is computed per-vertex. Thus, the resolution of the mesh has to be fixed beforehand and any change in the mesh requires a new computation of skinning weights or some complicated weight transfer from one mesh to the other." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": "To resolve this, we propose SkinNet " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "f_{\\omega}" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": ", which is a coordinate-based MLP where " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": " denotes the trainable network weights. Given any 3D point " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": " in the canonical space, " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "f_{\\omega}" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": " predicts its skinning weight " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": " conditioned on the normalized human pose " + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 304, + 434, + 547, + 494 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 502, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 502, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 359, + 502, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {\\tilde {\\theta}} = f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\boldsymbol {\\theta}}): \\mathbb {R} ^ {3} \\times \\mathbb {R} ^ {D} \\rightarrow \\mathbb {R} ^ {J}, \\tag {2}", + "image_path": "ccca1839abd1bceb1d15ce2f14ce69b8bc45e1157c688826eedb35190ada3f31.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "content": " are concatenated before being fed into the MLP. Here, the normalized pose " + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}" + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "content": " is obtained by neglecting the global translation as well as the yaw angle, i.e., the rotation around the spine. The intuition behind is that the skinning weights do not depend on where in global 3D space the person performs a respective pose, and, similarly, the skinning weight prediction should be agnostic to which direction the person is facing. Thus, we mask those degrees of freedom before feeding the pose vector into the network. To obtain the transformed point " + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 525, + 547, + 658 + ], + "type": "text", + "content": " in posed space, one can insert Eq. (2) into Eq. (1) resulting in:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 375, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = L B S (\\bar {\\mathbf {x}}, f _ {\\omega} (\\bar {\\mathbf {x}}, \\tilde {\\theta}), \\theta). \\tag {3}", + "image_path": "7913d3e1dac99f735eefdb901c0025fbc3079a29e3191340a70e518187316177.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Revisiting the drawbacks of initial skinning discussed above, the SkinNet now supports pose-dependent skinning" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1380" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "weights such that the artifacts of static skinning can be reduced. Moreover, since SkinNet is a coordinate-based MLP, it supports arbitrary mesh resolution since each vertex is queried independently, which is in stark contrast to an architecture that outputs all skinning weights at once and where the final tensor shape is fixed to a specific mesh resolution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 200, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 200, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 200, + 165 + ], + "type": "text", + "content": "3.3. Auxiliary Appearance Field" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "text", + "content": "We can now pose the geometry with our pose-dependent skinning weight formulation. However, since we want to supervise SkinNet solely on multi-view images, we also have to model the appearance of the actor. A naive choice would be to just leverage the static vertex colors acquired from NeuS. However, we found that this texture contains baked-in shadows and wrinkles, and cannot account for the pose- and view-dependent appearance changes harming the training of the SkinNet. We also refer to our ablation studies in Sec. 4.3. To address this issue, we propose an auxiliary appearance field, which consists of an albedo field predicting a pose- and view-agnostic color for a given point " + }, + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "text", + "content": " in the canonical space and a shadow field predicting the pose- and view-dependent shadow/shading properties of " + }, + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 171, + 288, + 386 + ], + "type": "text", + "content": ". Important to note here is that we are not interested in creating a highly photorealistic appearance model of the human, but we mainly use it as an auxiliary tool to better supervise the learning of the skinning weight network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": "More specifically, the albedo field, referred to as AlbedoNet, is an MLP that predicts a static albedo value " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "a_{\\omega'}(\\bar{\\mathbf{x}}) = \\mathbf{a} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " for a given 3D sample point in the canonical space. Here, " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\omega'" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " are the trainable weights of the albedo network. Moreover, the shadow field is also parameterized by an MLP " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "r_{\\omega''}(\\bar{\\mathbf{x}}, \\theta, \\mathbf{n}, \\mathbf{d}) = s \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": ", called ShadowNet, which predicts a scalar multiplier accounting for color changes due to shadows and shading. Note that we assume that such effects happen uniformly across all color channels and, thus, we are only predicting a single scalar instead of a scaling factor per color channel. In addition to taking the point in canonical space, the shadow field also takes the skeletal pose " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " as input as well as the surface normal " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " in global space and the viewing direction " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": ", which allows to potentially model pose- and view-dependent lighting effects such as shadows and shading. " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\omega''" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " denotes the respective network weights. Details about the architectures are provided in the supplemental document. The final color " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " for point " + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 387, + 287, + 615 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 624, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 624, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 114, + 624, + 287, + 638 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = a _ {\\omega^ {\\prime}} (\\bar {\\mathbf {x}}) r _ {\\omega^ {\\prime \\prime}} (\\bar {\\mathbf {x}}, \\theta , \\mathbf {n}, \\mathbf {d}). \\tag {4}", + "image_path": "f95d3658442c3f98e79d750b9196f61c9fccadfbcdcb3412876c53ac370c8b6b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 241, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 241, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 241, + 659 + ], + "type": "text", + "content": "3.4. Multi-view Video-based Supervision" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Next, we describe our dedicated supervision strategy that solely requires multi-view imagery. We first introduce the individual loss terms and then describe our training strategy. Importantly, we can obtain a posed template mesh" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " and the pose- and view-dependent per-vertex colors " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_c \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " by evaluating Eq. (3) and Eq. (4) for all vertices " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " and views " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": ". Moreover, the following losses that are applied to the deformed geometry and its appearance can directly backpropagate into the respective networks since our formulation is fully differentiable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 227 + ], + "type": "text", + "content": "Silhouette Loss. This loss [15] ensures that the projected posed geometry projects onto the foreground masks for all views (see supp. material). Since our pose-dependent skinning field Eq. (3) is differentiable with respect to the SkinNet weights " + }, + { + "bbox": [ + 304, + 144, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 144, + 545, + 227 + ], + "type": "text", + "content": ", the silhouette loss can provide supervision and learn pose-dependent skinning by matching the posed model silhouette and the image silhouettes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "content": "Rendering Loss. Although the silhouette loss ensures that the posed geometry matches the multi-view silhouettes, it cannot supervise drifts on the visual hull. Supervision directly from the RGB images can help with resolving this issue since appearance features can constrain drifts in the image plane. Thus, we introduce the rendering loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 308, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 308, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 308, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d}} = \\sum_ {c = 1} ^ {C} \\| \\Pi_ {c} (\\mathbf {V}, \\mathbf {C} _ {c}) - \\mathcal {I} _ {c} \\| _ {1}, \\tag {5}", + "image_path": "4b27a46a163ec88fc1e5e812bd8cfdec635542053b9b31d0e2f28db5060bd745.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": "which leverages a differentiable renderer " + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\Pi_c" + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": " based on the one of Habermann et al. [15] to render the posed and colored model into all camera views " + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": " that is then compared to the ground truth image " + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_c" + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": ". Importantly, this loss can supervise both the pose-dependent skinning field as well as the pose- and view-dependent appearance field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 465 + ], + "type": "text", + "content": "Laplacian Loss. Since the data terms alone can still lead to degenerate results, we also introduce some regularization on the geometry and the skinning weights. First, we regularize the posed geometry with a Laplacian loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 474, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 358, + 474, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l a p}} = \\sum_ {i = 1} ^ {N} \\| \\mathbf {v} _ {i} - \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {k \\in \\mathcal {N} _ {i}} \\mathbf {v} _ {k} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "8f687a9698ef8c277d76dd33afc4d326ae6c629e7f134113936d02d6812c5a90.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "text", + "content": " is the indices of one-ring neighborhood of vertex " + }, + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 513, + 545, + 536 + ], + "type": "text", + "content": ". This regularizer ensures locally smooth geometry." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "text", + "content": "**Skinning Regularization.** We found that SkinNet can overfit the training poses and respective deformations by predicting non-local skinning weights, e.g., the foot joint has a non-zero skinning weight for vertices on the shoulder. This, however, harms generalization to novel poses at test time. To prevent this, we regularize the skinning weights by constraining their values based on the geodesic distance " + }, + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{geo}}" + }, + { + "bbox": [ + 304, + 537, + 545, + 644 + ], + "type": "text", + "content": " between the vertex and the initial body part it belongs to. This can be formalized with the following loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 334, + 652, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 652, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 334, + 652, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s k i n}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {J} \\mathbf {w} _ {i, j} \\left(\\frac {\\min _ {k \\in \\mathscr {A} _ {j}} d _ {\\text {g e o}} (\\bar {\\mathbf {v}} _ {i} , \\bar {\\mathbf {v}} _ {k})}{d _ {\\text {g e o m a x}}}\\right) ^ {r}, \\tag {7}", + "image_path": "2128fe32efb33cf8e3550c225ea809007428f37625b222e961dc04d523712202.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathscr {A} _ {j} = \\left\\{k \\mid j = \\underset {t} {\\operatorname {a r g m a x}} \\mathbf {w} _ {\\text {i n i t}, k, t} \\right\\}, \\tag {8}", + "image_path": "7387ba6b6e00806a7be6039e8624ad1533e3f340961fbe1c9c83a3997d78d7b3.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "1381" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " is the set of vertices, which have the highest skinning weight with respect to joint " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "r = 3" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " is a hyperparameter which controls how penalization increases as the geodesic distance increases. " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{geomax}}" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " is the maximum geodesic distance between any of the vertex pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "text", + "content": "Part-wise Regularization. In contrast to the clothed areas, the skin part has rather static skinning. Thus, we use the initial skinning to regularize the predicted skinning weights of the skin part, i.e., the set " + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "text", + "content": " of vertices, which have the parsing label skin. However, even for skin vertices, we found that the initial skinning weights around the joints are not accurate while the rigid parts (i.e., the vertices away from joints) usually have correct initial skinning. We identify the set of rigid vertices as " + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{k|\\max (\\mathbf{w}_{\\mathrm{init},k}) > u\\}" + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "inline_equation", + "content": "u = 0.95" + }, + { + "bbox": [ + 46, + 133, + 287, + 251 + ], + "type": "text", + "content": ". Thus, our part loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 258, + 287, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 258, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 90, + 258, + 287, + 281 + ], + "type": "interline_equation", + "content": "\\mathscr {L} _ {\\text {p a r t}} = \\sum_ {i \\in \\mathscr {R} \\cap \\mathscr {G}} \\| f _ {\\omega} (\\bar {\\mathbf {v}} _ {i}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {w} _ {\\text {i n i t}, i} \\| _ {2} ^ {2}, \\tag {9}", + "image_path": "414888ea2f00ec8cacb8626ba89b6473974361518bbe16ac6d4ad2d69ccd4266.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "text", + "content": "enforces that the predicted skinning weights are close to the initial ones iff the human parsing label of a vertex is the skin " + }, + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "inline_equation", + "content": "(\\mathcal{G})" + }, + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "text", + "content": " and the vertex belongs to the rigid part " + }, + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "inline_equation", + "content": "(\\mathcal{R})" + }, + { + "bbox": [ + 46, + 289, + 287, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": "Training Strategy. Our training proceeds in four stages. First, we train SkinNet without using the rendering loss " + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": ". This step is required to ensure that the posed mesh roughly overlaps with the foreground masks. Next, the AlbedoNet is trained using only the rendering loss " + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": " while the SkinNet weights " + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": " are fixed. Importantly, the AlbedoNet is not pose-dependent, but we optimize the weights " + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\omega'" + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": " across all frames. Afterward, the AlbedoNet and the SkinNet are fixed and we train the ShadowNet using the rendering loss " + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 325, + 287, + 492 + ], + "type": "text", + "content": ". Again we train on all training frames. Lastly, since we now have a pose- and view-dependent appearance model available, we refine the SkinNet weights including the rendering loss. We found this gives the best performance, as validated by our results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 100, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 100, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 100, + 514 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Dataset. We evaluate our method on 3 subjects wearing tight and loose clothing, i.e., pants or skirts. Two of them are from the DynaCap dataset [15] and they are referred as " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D2" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D2" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " is wearing short pants and a T-shirt while " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " is in a skirt. The total numbers of cameras are 94 and 101 for those two sequences and our method can be trained using as few as 10 cameras. There are around 19000 training frames per subject depicting the actor in various articulated poses. For evaluation, there is a separate testing set of around 7000 frames per subject. For convenience, we keep 1 out of 10 frames for training and evaluation. Additionally, we captured a new subject, " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "V6" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ", wearing a T-shirt and trousers using 116 cameras. We captured a separate training (17730 frames) and testing sequence (5000 frames). Importantly, for this sequence, we recover hand gestures and a respective 3D model including hands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "Metrics. For evaluation, we reconstruct the dense 3D geometry for the testing frames using multi-view stereo reconstruction [2]. We do not use NeuS here as it is very time-consuming. Those dense meshes serve as ground truth scans and we compare the posed meshes recovered by our method as well as competing methods in terms of the Chamfer distance and the Hausdorff distance going from the recovered mesh to the ground truth scan (M2S) and vice versa (S2M). We report the average results over all testing frames. All metrics are reported in centimeter scale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 193, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 546, + 384 + ], + "type": "text", + "content": "Competing Methods. We compare our work to our re-implementation of Pinocchio [7], which proposed a heat diffusion process to obtain a static set of skinning weights. Moreover, we compare to SCANimate [49] and SNARF [8], which, in stark contrast to our work, require dense point clouds for training. They learn static skinning weights and a pose-aware implicit shape. Since our approach mostly focuses on skinning, we compare to SCANimate with and without pose-aware shape, referred to as SCANimate and SCANimate* in the following. Finally, we compare to RigNet [59], which can jointly predict a rigging skeleton and static skinning weights given a template. Unfortunately, SkinningNet [41] and NeuroSkinning [34] do not provide their code and, thus, we cannot compare to them. For more details and comparisons to other (slightly less) related works [31, 56], we refer to the supplemental document." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 418, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 418, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 418, + 404 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "content": "In Fig. 3, we qualitatively evaluate our approach. In the first column, we show the reconstructed static geometry for different subjects. In the following columns, we show the characters in novel poses from the test set overlayed onto the respective images. Note that our posed character nicely matches the reference image, which confirms that our pose-dependent skinning indeed generalizes to novel poses and results in accurately posed meshes. In Fig. 4, we further visualize the effect of the pose-dependent skinning weights obtained by our approach. The initial skinning weights are fixed across poses leading to the typical skinning artifacts and wrong deformation behavior. In contrast, our proposed pose-dependent skinning weights adapt according to the pose and as a result the deformations look more natural." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "text", + "content": "4.2. Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 1, we compare our results to previous works. Note that we achieve the most accurate results when compared to works only focusing on character skinning. This validates the necessity of pose-dependent skinning weights since other works rely on static ones. In this table, SCAN-imate and SNARF predict pose-aware shapes while assuming dense point clouds are given. Since it is not the scope of this work to predict pose-dependent non-rigid deformations and our method solely learns from video data, we placed" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1382" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 511, + 230 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 511, + 230 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 511, + 230 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 511, + 230 + ], + "type": "image", + "image_path": "719713ab2f42937e14871c69ffd912807906a8f24c90a9b142a6bc4127eab9a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "type": "text", + "content": "Figure 3. Qualitative results. From left to right: Recovered static character template. Reference image showing a test pose. Our posed character overlayed onto the reference image. The precise overlay of the posed geometry confirms the effectiveness of our approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 263, + 533, + 363 + ], + "blocks": [ + { + "bbox": [ + 60, + 263, + 533, + 363 + ], + "lines": [ + { + "bbox": [ + 60, + 263, + 533, + 363 + ], + "spans": [ + { + "bbox": [ + 60, + 263, + 533, + 363 + ], + "type": "table", + "html": "
Quantitative Geometry Comparison
SubjectD2D5V6
MethodChamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓Chamfer↓M2S↓S2M↓
Pinnochio [7]3.7602.1621.5995.0772.8112.2673.3581.8711.487
SCANimate* [49]3.7502.0991.6505.4532.9652.4883.5021.8901.612
RigNet [59]3.5992.0781.5214.9892.8362.1533.3691.8941.475
Ours3.0341.7461.2884.5122.4422.0702.9931.7191.274
SCANimate [49]2.8421.6461.1964.9822.2842.6983.1541.7141.440
SNARF [8]2.5911.4521.1394.3202.2882.0322.7601.4961.263
", + "image_path": "62fa1d9ba254515436c5ca583ee449fa3449b29ed305ec001bc0318ca4e63ece.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 411, + 263, + 546 + ], + "blocks": [ + { + "bbox": [ + 47, + 411, + 263, + 546 + ], + "lines": [ + { + "bbox": [ + 47, + 411, + 263, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 263, + 546 + ], + "type": "image", + "image_path": "4b7bfe1688536b0c2107190fe42e6468994f3c962a4c90860a9099dfaf15cc45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 554, + 287, + 621 + ], + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 621 + ], + "type": "text", + "content": "Figure 4. Qualitative results showing the pose-dependent skinning weights. As a reference, we show the result using the initial static skinning weights. Notably, using static skinning weights leads to artifacts and worse deformation. When using our pose-dependent skinning, one can see a clear improvement in the final result and we also show how the weights change over different poses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "them separately. Nonetheless, it is worth noting that our approach is close to SCANImate and SNARF in terms of accuracy and better than SCANImate for subject " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "V6" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " even though our training does not require point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We also visually compare to those works in Fig. 5. We can see qualitatively our method achieves a lower error especially in joint regions where most of the skinning artifacts" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 411, + 538, + 527 + ], + "blocks": [ + { + "bbox": [ + 46, + 369, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 546, + 403 + ], + "type": "text", + "content": "Table 1. We compare our method to previous works on 3 subjects. Our method obtains better results than other skinning-based approaches. It is even close to SCANimate and SNARF, which require dense pointclouds while we solely use multi-view video. Importantly, they focus on learning pose-aware shapes while our goal is to obtain a fully rigged and skinned character in a completely automated fashion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 411, + 538, + 527 + ], + "lines": [ + { + "bbox": [ + 314, + 411, + 538, + 527 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 538, + 527 + ], + "type": "table", + "html": "
Ablation Study on D2
MethodChamfer↓M2S↓S2M↓
Initial weights3.7612.1621.599
Static weights3.3541.9351.418
w/o Lrend3.1161.7831.333
w/ albedo-only3.0471.7501.297
w/ NeuS appearance3.1521.7861.365
w/ NeuS albedo3.1371.7951.342
w/ displacements3.1911.8051.386
Ours (10 views)3.1291.7911.337
Ours3.0341.7461.288
", + "image_path": "9c2c5e053f2fd2086dc5513b3a9a6299cbc47c643bfff42d85ba7006ca99e658.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 530, + 545, + 553 + ], + "lines": [ + { + "bbox": [ + 306, + 530, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 545, + 553 + ], + "type": "text", + "content": "Table 2. We compare our method against baselines. All designs improve the accuracy confirming the effectiveness of VINECS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "become visible. This can be nicely seen in the error maps, which visualize the per-vertex error. Although SCAnimate and SNARF perform well on naked body and tight clothes regions, they are worse than ours around loose clothing, as they rely on SMPL while our method does not have such a restriction. Further, the posed meshes clearly show fewer artifacts for our approach compared to others. Note that the entire character including the geometry, rigged skeleton, and the pose-dependent skinning can be obtained solely from multi-view video without any manual user input." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1383" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 70, + 515, + 221 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 515, + 221 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 515, + 221 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 515, + 221 + ], + "type": "image", + "image_path": "7ebc3df62fe87c9fcb61b4eb00b466dfd019df2bef5405e00db6ebc93889705f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 546, + 246 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison. For each method, we visualize the recovered posed geometry as well as the per-vertex error map when comparing the ground truth in terms of M2S. Note that our method consistently shows the lowest error and also has the least visual artifacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 252, + 276, + 370 + ], + "blocks": [ + { + "bbox": [ + 50, + 252, + 276, + 370 + ], + "lines": [ + { + "bbox": [ + 50, + 252, + 276, + 370 + ], + "spans": [ + { + "bbox": [ + 50, + 252, + 276, + 370 + ], + "type": "image", + "image_path": "147be2e9b316f2464c834dd70fcdb6a9eb6118b251bc3dfa1049bb34874b0779.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 371, + 287, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 404 + ], + "type": "text", + "content": "Figure 6. We compare the initial skinning, the static skinning optimized using our supervision scheme, ablating the rendering loss, and our pose-dependent skinning. We outperform all baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 416, + 111, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 111, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 111, + 427 + ], + "type": "text", + "content": "4.3. Ablation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "type": "text", + "content": "Next, we evaluate our design choices in Tab. 2. First, we evaluate the effect of pose-dependent skinning weights. Therefore, we compare our final result to the initial skinning, which we obtain by leveraging Pinocchio [7] and a version of our method where we use our supervision scheme, but optimize a static set of skinning weights, i.e., the weights are fixed across all poses. From the results, we can see that pose-dependent weights significantly improve the accuracy over the two baselines proofing the effectiveness of our formulation. This is also visualized in Fig. 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Another key aspect when learning those weights solely from video data is our proposed supervision scheme. When we do not employ the dense rendering loss or soley use the albedo without shadow estimation, the performance drops. Moreover, we also evaluate the accuracy when using a dense rendering loss with the static appearance from NeuS and using NeuS appearance as albedo while the shadow network is learned. We found that our proposed rendering scheme performs best. To validate the effectiveness of training pose-dependent skinning, we replace the SkinNet by a network predicting per-vertex displacement, which generalizes worse than SkinNet. We train our model with sparse cameras (10 views), and the performance is still satisfying." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 253, + 457, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 457, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 457, + 266 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 280, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 546, + 496 + ], + "type": "text", + "content": "Limitations. Although we propose the first method for learning pose-dependent skinning solely from multi-view video, our method still has some limitations, which opens up interesting directions for future work. Currently, we query the SkinNet for every vertex in canonical space, which may be inefficient when many points are sampled. Thus, future work could involve exploring more efficient architectures like hashgrids [43]. While hand and body can be faithfully animated, there is no control over the facial expression. Thus, extending our automated character creation pipeline with some parametric model for facial expressions might be a promising direction. Also, a joint consideration of rigging, skinning, and pose tracking may further improve the quality of the 3D animation-ready characters. Lastly, some artifacts still remain since skinning alone cannot account for high-frequency deformations. Combining VINECS with downstream tasks such as human surface tracking/modeling could alleviate such artifacts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 503, + 546, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 546, + 647 + ], + "type": "text", + "content": "Conclusion. We proposed the first method for creating a fully rigged and skinned character solely from multi-view video without any manual processing. To this end, we first acquire a canonical template using an implicit surface reconstruction, automated initial skinning and mesh parsing. Most importantly, we learn a pose-dependent skinning field from a multi-view video, which greatly reduces artifacts. We show how such a skinning field can be learned in a weakly supervised manner by an appearance model and differentiable rendering. Our work demonstrates that animatable character creation can be fully automated while maintaining high-quality geometry, rigging, and skinning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgements. We thank Heming Zhu and Diogo Luvizon for their help. This project was also supported by Saarbrücken Research Center for Visual Computing, Interaction, and AI. Christian Theobalt was supported by ERC Consolidator Grant 4DReply (770784)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1384" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "type": "text", + "content": "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 235, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 287, + 257 + ], + "type": "text", + "content": "[2] Agisoft. PhotoScan. http://www.agisoft.com, 2016. 6, 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 287, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 292 + ], + "type": "text", + "content": "[3] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. ACM Trans. Graph., 21(3):612-619, 2002. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 293, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 286, + 349 + ], + "type": "text", + "content": "[4] Brett Allen, Brian Curless, Zoran Popovic, and Aaron Hertzmann. Learning a Correlated Model of Identity and Pose-Dependent Body Shape Variation for Real-Time Synthesis. In ACM SIGGRAPH / Eurographics Symposium on Computer Animation. The Eurographics Association, 2006. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 286, + 393 + ], + "type": "text", + "content": "[5] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. ACM Transactions on Graphics, 24(3):408-416, 2005. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 286, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 286, + 427 + ], + "type": "text", + "content": "[6] Seungbae Bang and Sung-Hee Lee. Spline interface for intuitive skinning weight editing. ACM Trans. Graph., 37(5), 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 286, + 463 + ], + "type": "text", + "content": "[7] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Trans. Graph., 26(3), 2007, 2, 4, 6, 7, 8, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 286, + 518 + ], + "type": "text", + "content": "[8] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 3, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 286, + 576 + ], + "type": "text", + "content": "[9] Olivier Dionne and Martin de Lasa. Geodesic voxel binding for production character meshes. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 173-180, New York, NY, USA, 2013. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "text", + "content": "[10] Olivier Dionne and Martin de Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE Transactions on Visualization and Computer Graphics, 20 (10):1367-1378, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 286, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 688 + ], + "type": "text", + "content": "[11] Sven Forstmann, Jun Ohya, Artus Krohn-Grimberghe, and Ryan McDougall. Deformation styles for spline-based skeletal animation. In Proceedings of the 2007 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 141-150, Goslar, DEU, 2007. Eurographics Association. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[13] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[14] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), 1:1, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[15] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Trans. Graph., 40(4), 2021. 2, 5, 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 296 + ], + "type": "text", + "content": "[16] Marc Habermann, Lingjie Liu, Weipeng Xu, Gerard Pons-Moll, Michael Zollhoefer, and Christian Theobalt. Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech., 6(3), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[17] Dae-Eun Hyun, Seung-Hyun Yoon, Jung-Woo Chang, Joon-Kyung Seong, Myung-Soo Kim, and Bert Jüttler. Sweep-based human deformation. The Visual Computer, 21:542-550, 2005. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "text", + "content": "[18] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine-Hornung. Bounded biharmonic weights for real-time deformation. Commun. ACM, 57(4):99-106, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "type": "text", + "content": "[19] Doug L. James and Christopher D. Twigg. Skinning mesh animations. ACM Trans. Graph., 24(3):399-407, 2005. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "text", + "content": "[20] Yue Jiang, Marc Habermann, Vladislav Golyanik, and Christian Theobalt. Hifecap: Monocular high-fidelity and expressive capture of human performances. In BMVC, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "text", + "content": "[21] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Trans. Graph., 31 (6), 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 521 + ], + "type": "text", + "content": "[22] Ladislav Kavan and Jiří Žára. Spherical blend skinning: A real-time deformation of articulated models. In Proceedings of the 2005 Symposium on Interactive 3D Graphics and Games, page 9–16, New York, NY, USA, 2005. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[23] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46. ACM, 2007. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "text", + "content": "[24] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "text", + "content": "[25] Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Habermann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "text", + "content": "[26] Binh Huy Le and Zhigang Deng. Robust and accurate skeletal rigging from mesh sequences. ACM Trans. Graph., 33(4), 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[27] Binh Huy Le and Jessica K. Hodgins. Real-time skeletal skinning with optimized centers of rotation. ACM Trans. Graph., 35(4), 2016. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1385" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[28] J. P. Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, page 165-172, USA, 2000. ACM Press/Addison-Wesley Publishing Co. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 172 + ], + "type": "text", + "content": "[29] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 287, + 216 + ], + "type": "text", + "content": "[30] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "text", + "content": "[31] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. 2022. 3, 6, 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "type": "text", + "content": "[32] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 294, + 287, + 338 + ], + "type": "text", + "content": "[33] Zhouyingcheng Liao, Jimei Yang, Jun Saito, Gerard Pons-Moll, and Yang Zhou. Skeleton-free pose transfer for stylized 3d characters. In European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 381 + ], + "type": "text", + "content": "[34] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Trans. Graph., 38(4), 2019. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 287, + 426 + ], + "type": "text", + "content": "[35] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 470 + ], + "type": "text", + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "type": "text", + "content": "[37] William Lorensen and Harvey Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21:163-, 1987. 4, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 504, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 287, + 558 + ], + "type": "text", + "content": "[38] N. Magnenat-Thalmann, A. Laperri'ere, and D. Thalmann. Joint-dependent local deformations for hand animation and object grasping. In Proceedings of Graphics Interface '88, pages 26-33. Canadian Man-Computer Communications Society, 1988. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 613 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 647 + ], + "type": "text", + "content": "[40] Alex Mohr and Michael Gleicher. Building efficient, accurate character skins from examples. ACM Trans. Graph., 22 (3):562-568, 2003. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "type": "text", + "content": "[41] A. Mosella-Montoro and J. Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18572-18581, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1, 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[42] Tomohiko Mukai and Shigeru Kuriyama. Efficient dynamic skinning with low-rank helper bone controllers. ACM Trans. Graph., 35(4), 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "type": "text", + "content": "[44] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "text", + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "type": "text", + "content": "[46] Sang II Park and Jessica K. Hodgins. Capturing and animating skin deformation in human motion. ACM Trans. Graph., 25(3):881-889, 2006. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "type": "text", + "content": "[47] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Animatable neural radiance fields for human body modeling. ICCV, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "type": "text", + "content": "[48] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 1(1): 9054-9063, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "text", + "content": "[49] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "type": "text", + "content": "[50] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 479, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 523 + ], + "type": "text", + "content": "[51] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "type": "text", + "content": "[52] Peter-Pike J. Sloan, Charles F. Rose, and Michael F. Cohen. Shape by example. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, pages 135-143, New York, NY, USA, 2001. Association for Computing Machinery. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 569, + 544, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 544, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 544, + 591 + ], + "type": "text", + "content": "[53] TheCaptury. The Captury. http://www.thecapture.com/, 2020. 1, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 713 + ], + "type": "text", + "content": "[54] Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stefan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cirmrman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antonio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algo" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1386" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "rithms for Scientific Computing in Python. Nature Methods, 17:261-272, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 1, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "text", + "content": "[56] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, page 1–19, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 6, 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "type": "text", + "content": "[57] Xiaohuan Corina Wang and Cary Phillips. Multi-weight enveloping: Least-squares approximation techniques for skin animation. In Proceedings of the 2002 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 129–138, New York, NY, USA, 2002. Association for Computing Machinery. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 318 + ], + "type": "text", + "content": "[58] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects, pages 63-71, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "text", + "content": "[59] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. ACM Trans. on Graphics, 39, 2020. 1, 2, 6, 7, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 430 + ], + "type": "text", + "content": "[60] Ji Yang, Xinxin Zuo, Sen Wang, Zhenbo Yu, Xingyu Li, Bingbing Ni, Minglun Gong, and Li Cheng. Object wakeup: 3d object rigging from a single image. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 311-327. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 286, + 463 + ], + "type": "text", + "content": "[61] Xiaosong Yang, Arun Somasekharan, and Jian Jun Zhang. Curve skeleton skinning for human and creature characters. Computer Animation and Virtual Worlds, 17, 2006. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1387" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_content_list.json b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..357cdc52bee0dd497f5ee09bdc6f9c867388c3e7 --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_content_list.json @@ -0,0 +1,1373 @@ +[ + { + "type": "text", + "text": "VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens", + "text_level": 1, + "bbox": [ + 88, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fan Ma $^{1}$ , Xiaojie Jin $^{2*}$ , Heng Wang $^{2}$ , Yuchen Xian $^{1}$ , Jiashi Feng $^{2}$ , Yi Yang $^{1*}$", + "bbox": [ + 179, + 202, + 785, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University 2ByteDance Inc.", + "bbox": [ + 328, + 220, + 643, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in large video-language models have displayed promising outcomes in video comprehension. Current approaches straightforwardly convert video into language tokens and employ large language models for multi-modal tasks. However, this method often leads to the generation of irrelevant content, commonly known as \"hallucination\", as the length of the text increases and the impact of the video diminishes. To address this problem, we propose VISTA-LLAMA, a novel framework that maintains the consistent distance between all visual tokens and any language tokens, irrespective of the generated text length. VISTA-LLAMA omits relative position encoding when determining attention weights between visual and text tokens, retaining the position encoding for text and text tokens. This amplifies the effect of visual tokens on text generation, especially when the relative distance is longer between visual and text tokens. The proposed attention mechanism significantly reduces the chance of producing irrelevant text related to the video content. Furthermore, we present a sequential visual projector that projects the current video frame into tokens of language space with the assistance of the previous frame. This approach not only captures the temporal relationship within the video, but also allows less visual tokens to encompass the entire video. Our approach significantly outperforms various previous methods (e.g., Video-ChatGPT, MovieChat) on four challenging open-ended video question answering benchmarks. We reach an accuracy of 60.7 on the zero-shot NExT-QA and 60.5 on the zero-shot MSRVTT-QA, setting a new state-of-the-art performance. This project is available at https://jinxxian.github.io/Vista-LLaMA.", + "bbox": [ + 76, + 320, + 473, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 804, + 209, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The surge in multi-modal vision-and-language models [1, 12, 14], capable of comprehending both visual (e.g., image/video) and language data, can be attributed to the recent", + "bbox": [ + 76, + 829, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d453bf5d37aa3c2ebd46200975efba1f530d25463bb66fd20963bba9ffca429.jpg", + "image_caption": [ + "Figure 1. Video language processing with LLaMA [26] and our VISTA-LLAMA. The vanilla LLaMA treats visual tokens (☑) the same as other language tokens (☑), weakening the impact for tokens in long distance. Our model retains the same mechanism for language tokens and strengthens the impact of the visual tokens. The intensity of the impact of each token is conveyed through the depth of the line color $(\\rightarrow)$ . Our model provides the accurate response for the presented scenario." + ], + "image_footnote": [], + "bbox": [ + 503, + 273, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "achieves of large language models (LLMs) such as GPT [8], GLM [38], and LLaMA [40]. Video-language models (video-LMs) pose greater challenges in scaling due to increased computational and annotation costs compared to image-language models (image-LMs). Recent research has therefore focused on effectively training video-LMs by utilizing pre-trained image-LMs [22, 40].", + "bbox": [ + 496, + 642, + 893, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-LMs benefit from this warm-start approach to enhance visual representation learning by projecting video frames into language space and treating videos as several prompt language tokens [18, 34]. However, this diminishes the visual impact of videos on text generation and lacks explicit temporal modeling in videos. The generated text is often not related to the video content, as depicted in Figure 1, a phenomenon known as hallucination in language processing [6]. The distance between the generated text token and the visual tokens in large language models may", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "be a contributing factor, especially when the visual tokens are distant from the generated text. Additionally, handling longer videos poses a challenge because of the constraints on context length in large language models, where numerous visual tokens take up a substantial portion of the context.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present VISTA-LLAMA, an innovative video language framework to tackle above issues. The primary concept is to preserve equal distance between all visual tokens and any language tokens, while also retaining the relative distance between any two language tokens, as depicted in Figure 1. The rotary position embedding is only applied to language tokens to capture relative distance when calculating similarity between language tokens. When computing similarity between visual and text tokens, the rotary position embedding is removed to reduce the impact of relative distance. With the equal distance to visual tokens (EDVT) attention, the impact of visual cues on text production is amplified without compromising the text production capability. The experiments also show that our design produces more precise text description for input videos and the phenomenon of visual hallucinations occurring in language models has been greatly reduced.", + "bbox": [ + 75, + 167, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further improve the temporal modeling of video, we introduce a sequential visual projector. Instead of mapping each frame of the video independently into fixed-length visual tokens, which ignores the temporal relationship between frames, we generate visual tokens for each frame using the previous projected visual tokens. It not only incorporates the temporal relationship between frames into the language model to improve video comprehension without any extra parameters, but also enables the language model to encode longer videos with fewer visual tokens by sampling fewer projected frames.", + "bbox": [ + 75, + 426, + 468, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of VISTA-LLAMA on four challenging video question answering (QA) benchmarks, where our method outperforms several previous works, and achieves the state-of-the-art in zero-shot NExTQA [30], and MSRVTT [32]. We also show that our attention design and the temporal modeling mechanism improves capacity of large language model on video-question answering by a large margin. Comprehensive experiments are conducted to demonstrate the effectiveness of our designs. We summarize our contributions as follows:", + "bbox": [ + 75, + 593, + 468, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel video-language model, dubbed as VISTA-LLAMA, to enhance the video understanding and facilitate temporal modeling within the language model.", + "- A novel multi-modal attention is proposed to enable reliable video text generation by maintaining equal distance to visual tokens. Additionally, temporal modeling is facilitated by employing a sequential visual projector.", + "- Our method exhibits superior empirical performance, establishing new benchmarks for zero-shot open-ended video question answering tasks." + ], + "bbox": [ + 76, + 750, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A detailed analysis further explains the design choices inherent in our proposed framework, contributing significantly to a better comprehension of multi-modal dynamics in large language models.", + "bbox": [ + 500, + 90, + 888, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 166, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Language Models", + "text_level": 1, + "bbox": [ + 500, + 191, + 722, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large language models (LLMs) [2, 19, 27, 28, 42] have demonstrated exceptional proficiency in understanding language and reasoning, resulting in the generation of high-quality natural language text in diverse domains. LLMs have already initiated a technological revolution and have found extensive application in various domains. Additionally, a series of open source large models, including LLaMA [26], and OPT [41], have significantly contributed to technological advancements and made remarkable contributions to the NLP community. Leveraging the foundation established by these impressive LLMs, researchers have further expanded their capabilities and developed exceptional models for diverse NLP tasks (e.g. Vicuna [4]). Our work is also built upon these remarkable LLMs, and we equip the language models with a novel attention mechanism and a sequential visual projector that enhances their abilities to comprehend visual content in videos.", + "bbox": [ + 496, + 215, + 890, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-modal Large Language Models", + "text_level": 1, + "bbox": [ + 500, + 484, + 823, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers have been diligently exploring the application of LLMs for processing multi-modal problems [9, 13, 15, 17]. The existing approaches can be classified into two primary categories. In the first category, LLMs are treated as controllers while the multi-modal models as tools [10]. When presented with the specific task, the LLMs first interact with user instructions and decide which tools to employ. Subsequently, it generates comprehensive responses by amalgamating the outcomes derived from these readily available multi-modal models. These approaches, such as Visual ChatGPT [29], HuggingGPT [21], and DoraemonGPT [35], have shown impressive results on various multi-modal tasks without training models. For the second category, large-scale multi-modal models are trained on the multi-modal data. The principle behind this line of researches is to align other modal pre-trained models with textual LLMs. Flamingo [1] incorporates a perceiver resampler and a gated cross-attention layer to connect a frozen image encoder with the LLM. BLIP2 [12] introduces a Q-Former to map each image into fixed-length tokens in the language embedding space via the learned queries. LLaVA [14], mPLUG-owl [36], and MiniGPT4 [42] develop the image-LLMs utilizing image-instruction training pairs. Video Chat [13] extends image encoders to enable large models to comprehend visual content in videos. Video-ChatGPT [18] is trained on video instructional data to give appropriate answers for multi-modal", + "bbox": [ + 496, + 508, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13152", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inputs. In this work, we improve the understanding of videos in a video-language model by incorporating temporal modeling and a novel multi-modal attention mechanism.", + "bbox": [ + 75, + 90, + 470, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Video Question Answering", + "text_level": 1, + "bbox": [ + 76, + 146, + 318, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Question Answering (VideoQA) is a task that involves answering language questions that are based on a given video [32]. It requires the ability to understand and reason across different semantic levels, which in turn demands a capacity for multi-modal understanding. Previous VideoQA benchmarks primarily concentrated on short videos, asking questions according to the specified visual facts (e.g., location and objects) [16, 33]. Recently, several new benchmarks have been proposed to focus on resolving temporal and causal questions in longer video clips [25]. NExT-QA [30] is an example of such a benchmark, aiming to uncover the causalities or intentions of particular events, and infer subsequent actions within the video. In this work, we investigate the video-language model on zero-shot open-ended VideoQA where no training question-answer pairs are provided in the training stage.", + "bbox": [ + 75, + 169, + 472, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 424, + 166, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 449, + 186, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VISTA-LLAMA comprises three fundamental components: a visual encoder, a visual projector, and a pre-trained large language model. Figure 2 shows an overview of VISTA-LLAMA. The components' design and implementation details are provided below:", + "bbox": [ + 75, + 472, + 470, + 547 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Visual Encoder. We utilize pre-trained EVA-CLIP-g [24] and ViT-L [20] as visual encoder. The last layer of ViT encoder is removed because it specializes in aggregating the features of the first token for contrastive learning.", + "- Visual Projector. The visual projector maps the output of the visual encoder into tokens that occupy the same space as the text features from word embedding. Various visual projectors are considered in Sec. 3.3. The visual projector is to be trained in our work.", + "- Pre-trained Large Language Model. In this study, we utilize LLaVa [14], which is fine-tuned on Vicuna-7B [4] using instructional image-text pairs, for processing videos and texts. The causal mask is employed in all attention processes, incorporating the attention interplay between visual and text tokens. The pre-trained language model is frozen in the present study, and a new attention mechanism is developed to maintain a consistent distance to visual tokens for all textual tokens." + ], + "bbox": [ + 76, + 550, + 468, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our framework, the video is first encoded into visual tokens using the frozen visual encoder and the trainable visual projector. These visual tokens are then combined with text tokens, which are projected with word embedding from the prompt and question sentences. The combined visual", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of EDVT attention in a PyTorch-like style." + ], + "code_body": "# x: hidden input in each attention layer\n# v_mask: indcite which input are from visual tokens\ndef edvtattention_layer(x, v_mask):\n # query, answer, value projection\n q, k, v = qkv_proj(x)\n # apply RoPE for query and key inputs\n r_q, r_k = rope(q, k)\n # attention weights without RoPE\n attention = bmm(q.T, k)\n # attention weights with RoPE\n rattention = bmm(r_q, r_k)\n # Merge attention weights based on visual token\n attention = v_mask * attention + (1 - v_mask) *\n rattention\n attention = Softmax(attention, dim=-1)\n # Update representation based on attention weights\n v = bmm(attention, v)\n out = linear_proj(v)\n return out", + "guess_lang": "python", + "bbox": [ + 500, + 128, + 880, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and text tokens are then input into the language model to generate answers.", + "bbox": [ + 498, + 421, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Equal Distance to Visual Tokens", + "text_level": 1, + "bbox": [ + 500, + 465, + 782, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. For the concatenate visual-text input, three linear projection layers are applied in every attention layer to produce the query $\\mathbf{Q}$ , key $\\mathbf{K}$ , and value $\\mathbf{V}$ . Let $\\mathbf{q}_j$ be $j^{th}$ query in $\\mathbf{Q}$ . The conventional attention mechanism updates the input by initially determining the similarity between the query and key, then applying the attention weights to the value state. Specifically, for the $j^{th}$ position, the update procedure can be articulated as follows:", + "bbox": [ + 496, + 489, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 628, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\text{sim}(\\mathbf{q}_j, \\mathbf{k}_i) = \\exp(\\mathbf{q}_j^T \\mathbf{k}_i / \\sqrt{d})$ . Here, a causal mask is utilized so that the $j^{th}$ query can only attend to the key positioned less than $j$ .", + "bbox": [ + 496, + 669, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "EDVT-Attention. The vanilla attention model lacks positional awareness, with no encoded relative distance for natural language processing. In contrast, Rotary Positional Embeddings (RoPE) [23] encodes the position data of tokens using a rotation matrix, which inherently includes an explicit relative position dependency. Within each attention layer, RoPE is implemented across all projected query and key inputs in order to compute the attention weights via leveraging relative distance between tokens. The query situated at the $j^{th}$ position incorporates rotary position embedding through $\\tilde{\\mathbf{q}}_j = \\mathbf{R}_j\\mathbf{q}_j$ , wherein $\\mathbf{R}_j \\in \\mathbb{R}^{d \\times d}$ represents the rotary matrix for the $j^{th}$ position. Consequently, the attention involving", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d09f2ea609b6164d00534d7ab7d44470a68833a75da8785b60ea4b2d101b7ccc.jpg", + "image_caption": [ + "Figure 2. The framework of VISTA-LLAMA. The visual encoder and large language model are both frozen ( ) during training, while the projector is trainable ( ) to map video into the language's space. The attention operation in each layer is present on the right part. Only the text tokens are applied with rotary position embedding to include relative distance information. The attention weights between visual and language tokens are calculated without the rotary position embedding. The casual mask is applied to the bottom-right attention weights." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relative position embedding is expressed as follows:", + "bbox": [ + 76, + 460, + 421, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} _ {r o p e} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i}) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i})}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 486, + 470, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RoPE inherently integrates relative position data via the multiplication of rotation matrices rather than appending it to the input as a positional embedding. In natural language processing, this relative proximity between two words is vital, given that remote words should have less influence than adjacent words when generating the current word. However, using the same attention mechanism for visual and text tokens may result in unintentional text generation, a phenomenon often referred to as hallucination in LLMs. With multi-modal input, the generated text should depend on the visual content, disregarding the influence of relative distance.", + "bbox": [ + 75, + 522, + 470, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this issue and enhance the video understanding with LLMs, we introduce the EDVT-Attention where the equal distance to visual tokens is maintained while the relative distance between text tokens is attained. As shown in Figure 2, the rotary position embedding is only applied on text tokens. Let $\\mathcal{V}$ and $\\mathcal{T}$ be the set of visual and text tokens. The EDVT-Attention is formulated as:", + "bbox": [ + 75, + 686, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} _ {e d v t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 800, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{q}}_j$ and $\\tilde{\\mathbf{k}}_j$ denote the query and key applied with rotary embedding, separately. The distance between language tokens is determined using the rotary matrix, while", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the correlation between visual and textual inputs remains unaffected by the rotary embedding. This improves the influence of visual information on long-term text generation and decreases the incidence of hallucinations wherein the fabricated content is absent in the videos.", + "bbox": [ + 496, + 460, + 893, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ease comprehension of how EDVT-Attention works, Algorithm 1 exhibits the pseudo-code in the decoder layer of LLMs. It involves the combinations of attention weights prior to and subsequent to the implementation of rotary embedding with the visual mask.", + "bbox": [ + 496, + 537, + 893, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sequential Visual Projector", + "text_level": 1, + "bbox": [ + 500, + 619, + 746, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual projector aims to map video features into the language embedding space, allowing for the fusion and processing of visual and textual inputs by the substantial language model. As shown in Figure 3, earlier visual projectors either employ the linear layer or the query transformer (Q-Former) [12] to directly convert frame features into language tokens. However, the lack of temporal relationship in these methods impedes thorough video understanding in LLMs. We introduce the sequential visual projector in Figure 3 to encode the temporal context into the visual tokens.", + "bbox": [ + 496, + 643, + 893, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{o} \\in \\mathbb{R}^{t \\times l \\times d}$ be the extracted video feature of length $t$ where $l$ denotes the length of visual tokens. We use a Q-Former $f_{Q}$ to map each frame into the fixed length of $k$ representations $\\mathbf{x}_i = f_{Q}(\\mathbf{o}_i, \\mathbf{p}) \\in \\mathbb{R}^{k \\times d}$ where $\\mathbf{p} \\in \\mathbb{R}^{k \\times d}$ is the learnable query embedding. In the prior approach, all projected frame features are merely combined with word tokens to serve as mixed input for LLMs. For encoding temporal", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13154", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/60b6b2d49ed7316370729aa4b1dd9dbf0caa6f7d7598e83eff2e24327aab1ee9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 129, + 88, + 416, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f33c44ada250c220cacf8291c1044bd26943cc5fa96d9e6c765ee4d3eb91d45.jpg", + "image_caption": [ + "Figure 3. Comparison of three visual projectors. (a): Each frame feature is projected into the visual tokens independently with the linear projection. (b): Q-Former uses shared learnable query tokens to separately map each frame into fixed-length tokens. (c): The sequential Q-Former with linear projection layer to enable temporal modeling." + ], + "image_footnote": [], + "bbox": [ + 127, + 222, + 370, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "context, we utilize the previously projected frame feature as the query to attend to the current frame feature. The current projected frame feature is then updated accordingly:", + "bbox": [ + 75, + 478, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = f _ {Q} \\left(\\mathbf {o} _ {i}, \\mathbf {x} _ {t - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 537, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{t - 1}\\in \\mathbb{R}^{k\\times d}$ represents the previous projected frame feature. This allows the visual tokens to encode the temporal relationship, as the current frame's visual token is generated using the previous feature.", + "bbox": [ + 75, + 564, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, we can represent the entire video with fewer tokens by merely sampling a small number of visual tokens. This method tackles the challenge of encoding lengthy videos by employing sequential encoding. By integrating prior visual context into subsequent visual tokens, we can accomplish an adequate representation of the entire video through sparse sampling of projected frame features. Besides, this technique permits larger language models to manage much longer videos without length limitation.", + "bbox": [ + 75, + 625, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 770, + 292, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We fine-tune the model based on LLaVA [14] and use $100K$ video instruction pairs provided by [18]. We add Q-Former initialized from BLIP2 [12] to project frame features into fixed-length tokens. We test our model with both ViT-L-14 [20] and EVA-CLIP-g [24] visual encoders in experiments. We only update the visual projector, which contains the Q-Former and linear projection layer to project the video", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features to the LLMs' input space. The visual backbone and the language model are frozen during training. The VISTA-LLAMA is fine-tuned for 3 epochs on 8 A100 80GB GPUs with a learning rate of $2e^{-5}$ and an overall batch size of 32. We run all the inference experiments with FP16 to save memory and faster testing.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 194, + 632, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 218, + 689, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our method is evaluated on four datasets:", + "bbox": [ + 500, + 242, + 841, + 256 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NExT-QA [30] is designed to advance video understanding from descrii ng to explaining the temporal actions. It comprises 5,440 videos and approximately $52K$ manually annotated QA pairs, which are categorized into temporal (Tem.), causal (Cau.), and descriptive (Des.) questions.", + "- MSVD-QA [32] is a dataset built upon Microsoft Research Video Description Corpus [3], commonly used in video caption tasks. The MSVD-QA dataset comprises a total of 1,970 video clips with 50,505 QA pairs.", + "- MSRVTT-QA [32] is based on MSR-VTT dataset [33], which includes $10K$ videos and $243K$ QA pairs with larger and has more complex scenes.", + "- ActiviytNet-QA [37] is a fully annotated and large-scale videoQA dataset. It contains 58K QA pairs derived from 5,800 complex web videos derived from the popular ActivityNet dataset [11]." + ], + "bbox": [ + 500, + 258, + 893, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. We adopt two metrics to evaluate the performance of video-language models.", + "bbox": [ + 500, + 502, + 893, + 532 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Open-Ended Zero-Shot Question-Answer Evaluation. As video-language models generate responses of varying lengths to open-ended questions, it is challenging to evaluate models with traditional word matching strategy. We employ LLM-Assisted evaluation, in line with [18], for fair comparison. Given the question, correct answer, and predicted answer, GPT3.5-turbo-0613 is used to return True or False judgement and relative score (0 to 5).", + "- Video-Based Text Generation Benchmarking. We evaluate the text generation performance following [18] from five aspects: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding. The test set for this evaluation is based on the ActivityNet-200 [7], featuring videos with rich, dense descriptive captions and associated question-answer pairs from human annotations. The evaluation pipeline is also built with the GPT-3.5 model and relative score (0 to 5) is generated." + ], + "bbox": [ + 500, + 535, + 893, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison to State-of-the-Arts", + "text_level": 1, + "bbox": [ + 500, + 816, + 784, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the zero-shot open-ended video question answering tasks, we compare our model with FrozenBiLM [34], VideoChat [13], LLaMA Adapter [40], VideoLLaMA [39], VideoChatGPT [18], and MovieChat [22] in Tab. 1. FrozenBiLM", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/53aa6ee4c8c512dde530fa115d5254afbc625ccffbdb947b8e8b039bd435b149.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNExT-QA [30]MSVD-QA [32]MSRVTT-QA [32]ActivityNet-QA [37]
AccuracyScoreAccuracyScoreAccuracyScoreAccuracyScore
FrozenBiLM [34]--32.2-16.8-24.7-
Video Chat [13]56.23.256.32.845.02.526.52.2
LLaMA Adapter [40]--54.93.143.82.734.22.7
Video LLaMA [39]--51.62.529.61.812.41.1
MovieChat [22]49.92.761.02.949.72.851.53.1
Video-ChatGPT [18]54.63.264.93.349.32.835.22.7
VISTA-LLAMA (Ours)60.73.465.33.660.53.348.33.3
", + "bbox": [ + 101, + 88, + 874, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42a793f29b9b338f826492d195ee87d20164f18bc926d3148ae8ba697aa891ca.jpg", + "table_caption": [ + "Table 1. Comparison with SoTA methods on zero-shot VideoQA. See §4.2 for more details." + ], + "table_footnote": [], + "table_body": "
MethodCr.Cs.De.Ct.Te.
Video Chat [13]2.232.242.502.531.94
LLaMA Adapter [40]2.032.152.322.301.98
Video LLaMA [39]1.961.792.182.161.82
Video-ChatGPT [18]2.402.372.522.621.98
VISTA-LLAMA (Ours)2.442.312.643.182.26
", + "bbox": [ + 81, + 279, + 470, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on video-based text generation with different video-language methods (§4.2). For clarity, five scores are reported (\"Cr\": Correctness of Information, \"Cs\": Consistency, \"De\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te\": Temporal Understanding).", + "bbox": [ + 75, + 383, + 470, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adapts frozen the bidirectional language model, showing promising results in zero-shot VideoQA settings. Other compared models are built on recent large auto-regressive language models. Despite pre-existing models have produced substantial results, VISTA-LLAMA consistently outperforms them, achieving state-of-the-art (SoTA) performance across three datasets: NExT-QA [30], MSVD-QA [32], and MSVTT-QA [32]. Our method obtains the highest results, with $60.5\\%$ accuracy on MSRvTT, elevating the performance of the second-best model by nearly $10\\%$ . Additionally, our method attains $60.7\\%$ accuracy on NExT-QA, markedly superior to Video-ChatGPT [18]. These results demonstrate VISTA-LLAMA's capability to comprehend video content and produce precise answers.", + "bbox": [ + 75, + 488, + 470, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the results of the evaluation of video-based text generation in Tab. 2. The results reveal its competent performance across all aspects when compared with the recent video-language models, Video Chat[13], VideoLLaMA [39], and Video-ChatGPT [18]. Despite being trained with identical datasets, our model outperforms Video-ChatGPT in four aspects. Our method offers a more comprehensive interpretation, and its responses are more in tune with the overarching context of the video content than comparable approaches. By utilizing the EDVT-Attention and sequential temporal modeling techniques, our model demonstrates a strong ability to generate text that is contextually appropriate, detailed, and includes precise timing for video inputs.", + "bbox": [ + 75, + 704, + 473, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ca9f919bd9e213840828c73cd7203dfe8f87ae302086bf8a131692ec539df82.jpg", + "image_caption": [ + "Figure 4. Comparison of different design choices on NExTQA [30]( $\\S$ 4.3). For clarity, accuracy of base model and two variants are given (\"Baseline\": the frozen LLaVA [14] with trainable Q-Former [12], \"w Seq\": base model with sequential visual projector, \"w Seq&EDVT\": base model with both sequential visual projector and EDVT-Attention)." + ], + "image_footnote": [], + "bbox": [ + 547, + 284, + 848, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Abalation Study", + "text_level": 1, + "bbox": [ + 500, + 573, + 661, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Design Choices. We investigate the effect of our two designs, including the equal-distance to visual-tokens (EDVT) attention and the sequential visual projector. As shown in Figure 4, our model is tested on NExT-QA [30], which comprises three different types of questions. The baseline model employs LLaVA [14] as the language model, and Q-Former in BLIP-2 [12] as the visual projector. We first incorporate temporal modeling into the baseline, utilizing the projected tokens from the previous frame as the query tokens to generate visual tokens for the current frame, a setup denoted as \"w Seq\" in Figure 4. Further enhancing this version with EDVT-Attention, we introduce \"w Seq&EDVT\", ultimately forming our final VISTA-LLAMA. The results affirm the efficacy of our design across all question types.", + "bbox": [ + 496, + 597, + 893, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EDVT-Attention. We here validate the effectiveness of our core EDVT-Attention. Table 3 reports the comparison results of different models in combination with our EDVT-Attention on NExT-QA [30]. Initially, we integrate the EDVT-Attention into Video-ChatGPT [18], where only the attention component is substituted with our EDVT-Attention", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06333a2bf1649cfdec5c68bb4e2f7398d8b3240179b5addc1fe34a84a1d4404e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNExT-QA [30]
Tem.Cau.Des.Avg.
Video-ChatGPT [18]37.665.154.954.6
+ EDVT-Attention39.5 (+1.9)72.8 (+7.7)54.859.3 (+4.7)
VISTA-LLAMA (ViT-L-14)34.069.142.253.6
+ EDVT-Attention36.8 (+2.8)72.2 (+3.1)47.7 (+5.5)56.5 (+2.9)
VISTA-LLAMA (EVA-CLIP-g)34.365.855.954.1
+ EDVT-Attention40.7 (+6.4)72.3 (+6.5)57.0 (+1.1)59.7 (+5.6)
", + "bbox": [ + 86, + 88, + 470, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e85abb7685829c2f4ccb8fb66e557e745fba1b4dc7145d952e2ef9d586dba89.jpg", + "image_caption": [ + "(a) Vanilla attention" + ], + "image_footnote": [], + "bbox": [ + 84, + 284, + 271, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1934748432e33a3fa4b48770222b00e3b7e8e0ac52d38cc123813ee71096d4d8.jpg", + "image_caption": [ + "(b) EDVT-Attention (Ours)" + ], + "image_footnote": [], + "bbox": [ + 274, + 284, + 460, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "during training. Further deployment of the EDVT-Attention into Video-ChatGPT [18] yields significant performance gains (e.g., $65.1\\%$ escalates to $72.8\\%$ for causal questions). The accuracy is enlarged across all three setting types, indicating that our EDVT-Attention considerably enhances the multi-modal understanding in large language models.", + "bbox": [ + 75, + 520, + 468, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further illustrate the attention weights of both EDVT-Attention and conventional attention in Figure 5. For this experiment, we employ 128 visual tokens and integrate the attention weights into the first four tokens to enhance visualization. The attention weights assigned to visual tokens are markedly greater in our EDVT-Attention compared to those in conventional attention. This indicates that the impact of visual tokens on language tokens is considerably more substantial under the proposed EDVT-Attention, shedding light on why the performance notably improves with our design.", + "bbox": [ + 75, + 611, + 470, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Projector. The visual projector is trainable within the model. Therefore, we have examined the impacts of different visual projector designs, as shown in Table 4. We evaluated three visual projector variants. The Q-Former, initialized using BLIP-2 [12], considerably outperforms the Q-Former that was initialized with BERT [5]. This indicates that visual projectors, when pre-trained on image-text pairs, can significantly enhance video comprehension. The model's performance peaks when integrating the sequential design,", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06eedae109a23c2a427e20f7d7e863fe47b2154233dbdd802a9f778f5939074e.jpg", + "table_caption": [ + "Table 3. Comparison of EDVT-Attention design with diffent visual encoders on NExT-QA [30] (§4.3). For clarity, accuracy of all questions and three types of questions are reported (\"Tem\": temporal, \"Cau\": causal, \"Des\": descriptive, \"Avg.\": average)." + ], + "table_footnote": [], + "table_body": "
Visual projectorNExT-QA [30]
Tem.Cau.Des.Avg.
Linear Projector37.665.154.954.6
Q-Former (BERT init.)35.262.749.251.8
Q-Former (BLIP-2 init.)34.365.855.954.1
SeqQ-Former (BLIP-2 init.)36.268.551.155.4
", + "bbox": [ + 509, + 88, + 893, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of different visual projectors on NExTQA [30]. The linear projector is initialized with pre-trained weights in LLaVa [14]. BERT init. and BLIP-2 init. indicate that the visual projector is initialized with weights from BERT [5] and BLIP-2 [12]. SeqQ-Former is the proposed sequential visual projector. See §4.3 for more details.", + "bbox": [ + 496, + 186, + 893, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18e3fbd0e90c2c5ebce735d5ba36164ab1d089c5857de038d52372f99d548245.jpg", + "image_caption": [ + "Figure 5. Comparison of attention weights for varying context lengths. Lighter colors represent higher weights. To improve clarity, we have combined visual token weights into the first four tokens. We recommend zooming in for optimal viewing.", + "Figure 6. The effect of training visual tokens on NExT-QA [30]. Accuracy of all questions and three types of questions, including temporal (Tem.), descriptive (Des.), and causal (Cau.), are presented with different colors." + ], + "image_footnote": [], + "bbox": [ + 501, + 286, + 890, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "referred to as \"SeqQ-Former\" in the figure, outperforming all other visual projectors in terms of accuracy.", + "bbox": [ + 496, + 611, + 890, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Training Visual Tokens. We assess the impact of visual tokens on NExT-QA as illustrated in Figure 6. In this study, a frame is converted into 32 visual tokens using Q-Former. Frames were sampled at various timestamps throughout the training process. The overall accuracy improves as more visual tokens are incorporated into the language model. However, for temporal queries, precision declined with an increased number of visual tokens. The accuracy of temporal reasoning questions is notably lower than that of other question types. Temporal reasoning presents more difficulties, and the language model may not excel at the temporal modeling of visual tokens. In terms of descriptive question types, the accuracy steadily increases as the model gains more visual information.", + "bbox": [ + 496, + 642, + 893, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We visualize the generated responses of Video-ChatGPT [18] and VISTA-LLAMA for different videos in Figure 7. The frames sampled at different times", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64cbfd04962d41864a9b94203e39d864c7b418d2f4495674c82cc46836b8a295.jpg", + "image_caption": [ + "Figure 7. Visualization results on different video questions. The questions and annotated answers are located on the left side. The generated text from Video-ChatGPT [31] and our model is presented in the green and orange boxes, respectively. See §4.3 for more details." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tamps is presented, with questions listed below the images. In the first video, the mobile phone is turned, sequentially exposing each side over time. The question asks for the side of the mobile phone at the end of the video. Video-ChatGPT gives the wrong response, as the camera is never present at the bottom left side in the video. The hallucination occurs in Video-ChatGPT, whereas our model predicts the answer correctly. The responses in three cases demonstrate that Video-ChatGPT provides unrelated answers that do not correspond to the video content. In the second case, the person in pink is always standing while the man is sitting on the ground, but Video-ChatGPT incorrectly states that the person in pink is also sitting on the ground. In the third video, the man is consistently walking up the mountain and never walking down. However, Video-ChatGPT falsely outputs that the man continues walking down the mountain. In contrast, our model provides the correct and reliable responses. This demonstrates that our model significantly reduces hallucination in video understanding and delivers more accurate responses.", + "bbox": [ + 75, + 584, + 472, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 583, + 620, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present VISTA-LLAMA to improve the video understanding in large language model. A new vision-aware attention is introduced to maintain same relative distance between all visual tokens and language tokens. In addition, we propose a sequential visual projector to map video into the language space to enable temporal modeling. Experiments on several video question answering task demonstrate that our designs significant improves the current SoTA. Current work is built on the pre-trained image-text architecture, the power of our designs could be enlarged when applied in pre-training stage. Furthermore, the performance of our sequential encoder on longer videos can also be assessed. We would consider evaluating the model in more tasks and applied the attention mechanism to more vision language models.", + "bbox": [ + 496, + 611, + 895, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554, U2336212), and China Postdoctoral Science Foundation (524000-X92302).", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13158", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 23716-23736, 2022. 1, 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2", + "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Annual Meeting of the Association for Computational Linguistics, 2011. 5", + "[4] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2, 3", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7", + "[6] Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495, 2023. 1", + "[7] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 5", + "[8] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30: 681–694, 2020. 1", + "[9] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yezhou Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14773-14783, 2022. 2", + "[10] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953–14962, 2023. 2", + "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 5", + "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 2, 4, 5, 6, 7", + "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao." + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 5, 6", + "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. ArXiv, abs/2304.08485, 2023. 1, 2, 3, 5, 6, 7", + "[15] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zeroshot video grounding with pseudo query lookup and verification. IEEE Transactions on Image Processing, 33:1643-1654, 2024. 2", + "[16] Fan Ma, Linchao Zhu, and Yi Yang. Weakly supervised moment localization with decoupled consistent concept prediction. International Journal of Computer Vision, 130(5): 1244-1258, 2022. 3", + "[17] Fan Ma, Xiaojie Jin, Heng Wang, Jingjia Huang, Linchao Zhu, Jiashi Feng, and Yi Yang. Temporal perceiving video-language pre-training. 2023. 2", + "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 2, 5, 6, 7", + "[19] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023.2", + "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5", + "[21] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2", + "[22] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 5, 6", + "[23] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3", + "[24] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3, 5", + "[25] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640, 2015. 3", + "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 1, 2", + "[27] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, Felix Hill, and Zacharias Janssen. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13159", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Luccioni, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2", + "[29] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 2", + "[30] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9772–9781, 2021. 2, 3, 5, 6, 7", + "[31] Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. ArXiv, abs/2304.01196, 2023. 8", + "[32] D. Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. Proceedings of the 25th ACM international conference on Multimedia, 2017. 2, 3, 5, 6", + "[33] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5288-5296, 2016. 3, 5", + "[34] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1, 5, 6", + "[35] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models, 2024. 2", + "[36] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 2", + "[37] Zhou Yu, D. Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 5, 6", + "[38] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414, 2022. 1", + "[39] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6", + "[40] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1, 5, 6", + "[41] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models. ArXiv, abs/2205.01068, 2022. 2", + "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 2" + ], + "bbox": [ + 501, + 90, + 893, + 217 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13160", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_model.json b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_model.json new file mode 100644 index 0000000000000000000000000000000000000000..03cfb8f5dbcef06ba6c19a8ecfefd5d0fb05610f --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.203, + 0.787, + 0.222 + ], + "angle": 0, + "content": "Fan Ma\\(^{1}\\), Xiaojie Jin\\(^{2*}\\), Heng Wang\\(^{2}\\), Yuchen Xian\\(^{1}\\), Jiashi Feng\\(^{2}\\), Yi Yang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.222, + 0.644, + 0.241 + ], + "angle": 0, + "content": "1Zhejiang University 2ByteDance Inc." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.289, + 0.314, + 0.306 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.474, + 0.775 + ], + "angle": 0, + "content": "Recent advances in large video-language models have displayed promising outcomes in video comprehension. Current approaches straightforwardly convert video into language tokens and employ large language models for multi-modal tasks. However, this method often leads to the generation of irrelevant content, commonly known as \"hallucination\", as the length of the text increases and the impact of the video diminishes. To address this problem, we propose VISTA-LLAMA, a novel framework that maintains the consistent distance between all visual tokens and any language tokens, irrespective of the generated text length. VISTA-LLAMA omits relative position encoding when determining attention weights between visual and text tokens, retaining the position encoding for text and text tokens. This amplifies the effect of visual tokens on text generation, especially when the relative distance is longer between visual and text tokens. The proposed attention mechanism significantly reduces the chance of producing irrelevant text related to the video content. Furthermore, we present a sequential visual projector that projects the current video frame into tokens of language space with the assistance of the previous frame. This approach not only captures the temporal relationship within the video, but also allows less visual tokens to encompass the entire video. Our approach significantly outperforms various previous methods (e.g., Video-ChatGPT, MovieChat) on four challenging open-ended video question answering benchmarks. We reach an accuracy of 60.7 on the zero-shot NExT-QA and 60.5 on the zero-shot MSRVTT-QA, setting a new state-of-the-art performance. This project is available at https://jinxxian.github.io/Vista-LLaMA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.805, + 0.21, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.471, + 0.877 + ], + "angle": 0, + "content": "The surge in multi-modal vision-and-language models [1, 12, 14], capable of comprehending both visual (e.g., image/video) and language data, can be attributed to the recent" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.274, + 0.891, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Video language processing with LLaMA [26] and our VISTA-LLAMA. The vanilla LLaMA treats visual tokens (☑) the same as other language tokens (☑), weakening the impact for tokens in long distance. Our model retains the same mechanism for language tokens and strengthens the impact of the visual tokens. The intensity of the impact of each token is conveyed through the depth of the line color \\((\\rightarrow)\\). Our model provides the accurate response for the presented scenario." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.749 + ], + "angle": 0, + "content": "achieves of large language models (LLMs) such as GPT [8], GLM [38], and LLaMA [40]. Video-language models (video-LMs) pose greater challenges in scaling due to increased computational and annotation costs compared to image-language models (image-LMs). Recent research has therefore focused on effectively training video-LMs by utilizing pre-trained image-LMs [22, 40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Video-LMs benefit from this warm-start approach to enhance visual representation learning by projecting video frames into language space and treating videos as several prompt language tokens [18, 34]. However, this diminishes the visual impact of videos on text generation and lacks explicit temporal modeling in videos. The generated text is often not related to the video content, as depicted in Figure 1, a phenomenon known as hallucination in language processing [6]. The distance between the generated text token and the visual tokens in large language models may" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "be a contributing factor, especially when the visual tokens are distant from the generated text. Additionally, handling longer videos poses a challenge because of the constraints on context length in large language models, where numerous visual tokens take up a substantial portion of the context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.426 + ], + "angle": 0, + "content": "We present VISTA-LLAMA, an innovative video language framework to tackle above issues. The primary concept is to preserve equal distance between all visual tokens and any language tokens, while also retaining the relative distance between any two language tokens, as depicted in Figure 1. The rotary position embedding is only applied to language tokens to capture relative distance when calculating similarity between language tokens. When computing similarity between visual and text tokens, the rotary position embedding is removed to reduce the impact of relative distance. With the equal distance to visual tokens (EDVT) attention, the impact of visual cues on text production is amplified without compromising the text production capability. The experiments also show that our design produces more precise text description for input videos and the phenomenon of visual hallucinations occurring in language models has been greatly reduced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.592 + ], + "angle": 0, + "content": "To further improve the temporal modeling of video, we introduce a sequential visual projector. Instead of mapping each frame of the video independently into fixed-length visual tokens, which ignores the temporal relationship between frames, we generate visual tokens for each frame using the previous projected visual tokens. It not only incorporates the temporal relationship between frames into the language model to improve video comprehension without any extra parameters, but also enables the language model to encode longer videos with fewer visual tokens by sampling fewer projected frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.744 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of VISTA-LLAMA on four challenging video question answering (QA) benchmarks, where our method outperforms several previous works, and achieves the state-of-the-art in zero-shot NExTQA [30], and MSRVTT [32]. We also show that our attention design and the temporal modeling mechanism improves capacity of large language model on video-question answering by a large margin. Comprehensive experiments are conducted to demonstrate the effectiveness of our designs. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.468, + 0.794 + ], + "angle": 0, + "content": "- We introduce a novel video-language model, dubbed as VISTA-LLAMA, to enhance the video understanding and facilitate temporal modeling within the language model." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.468, + 0.854 + ], + "angle": 0, + "content": "- A novel multi-modal attention is proposed to enable reliable video text generation by maintaining equal distance to visual tokens. Additionally, temporal modeling is facilitated by employing a sequential visual projector." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.857, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- Our method exhibits superior empirical performance, establishing new benchmarks for zero-shot open-ended video question answering tasks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.75, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.89, + 0.152 + ], + "angle": 0, + "content": "- A detailed analysis further explains the design choices inherent in our proposed framework, contributing significantly to a better comprehension of multi-modal dynamics in large language models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.167, + 0.642, + 0.183 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.723, + 0.209 + ], + "angle": 0, + "content": "2.1. Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Large language models (LLMs) [2, 19, 27, 28, 42] have demonstrated exceptional proficiency in understanding language and reasoning, resulting in the generation of high-quality natural language text in diverse domains. LLMs have already initiated a technological revolution and have found extensive application in various domains. Additionally, a series of open source large models, including LLaMA [26], and OPT [41], have significantly contributed to technological advancements and made remarkable contributions to the NLP community. Leveraging the foundation established by these impressive LLMs, researchers have further expanded their capabilities and developed exceptional models for diverse NLP tasks (e.g. Vicuna [4]). Our work is also built upon these remarkable LLMs, and we equip the language models with a novel attention mechanism and a sequential visual projector that enhances their abilities to comprehend visual content in videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.824, + 0.501 + ], + "angle": 0, + "content": "2.2. Multi-modal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Researchers have been diligently exploring the application of LLMs for processing multi-modal problems [9, 13, 15, 17]. The existing approaches can be classified into two primary categories. In the first category, LLMs are treated as controllers while the multi-modal models as tools [10]. When presented with the specific task, the LLMs first interact with user instructions and decide which tools to employ. Subsequently, it generates comprehensive responses by amalgamating the outcomes derived from these readily available multi-modal models. These approaches, such as Visual ChatGPT [29], HuggingGPT [21], and DoraemonGPT [35], have shown impressive results on various multi-modal tasks without training models. For the second category, large-scale multi-modal models are trained on the multi-modal data. The principle behind this line of researches is to align other modal pre-trained models with textual LLMs. Flamingo [1] incorporates a perceiver resampler and a gated cross-attention layer to connect a frozen image encoder with the LLM. BLIP2 [12] introduces a Q-Former to map each image into fixed-length tokens in the language embedding space via the learned queries. LLaVA [14], mPLUG-owl [36], and MiniGPT4 [42] develop the image-LLMs utilizing image-instruction training pairs. Video Chat [13] extends image encoders to enable large models to comprehend visual content in videos. Video-ChatGPT [18] is trained on video instructional data to give appropriate answers for multi-modal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.136 + ], + "angle": 0, + "content": "inputs. In this work, we improve the understanding of videos in a video-language model by incorporating temporal modeling and a novel multi-modal attention mechanism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.147, + 0.32, + 0.163 + ], + "angle": 0, + "content": "2.3. Video Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.473, + 0.412 + ], + "angle": 0, + "content": "Video Question Answering (VideoQA) is a task that involves answering language questions that are based on a given video [32]. It requires the ability to understand and reason across different semantic levels, which in turn demands a capacity for multi-modal understanding. Previous VideoQA benchmarks primarily concentrated on short videos, asking questions according to the specified visual facts (e.g., location and objects) [16, 33]. Recently, several new benchmarks have been proposed to focus on resolving temporal and causal questions in longer video clips [25]. NExT-QA [30] is an example of such a benchmark, aiming to uncover the causalities or intentions of particular events, and infer subsequent actions within the video. In this work, we investigate the video-language model on zero-shot open-ended VideoQA where no training question-answer pairs are provided in the training stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.168, + 0.44 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.45, + 0.187, + 0.464 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.548 + ], + "angle": 0, + "content": "VISTA-LLAMA comprises three fundamental components: a visual encoder, a visual projector, and a pre-trained large language model. Figure 2 shows an overview of VISTA-LLAMA. The components' design and implementation details are provided below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.551, + 0.469, + 0.611 + ], + "angle": 0, + "content": "- Visual Encoder. We utilize pre-trained EVA-CLIP-g [24] and ViT-L [20] as visual encoder. The last layer of ViT encoder is removed because it specializes in aggregating the features of the first token for contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.612, + 0.47, + 0.686 + ], + "angle": 0, + "content": "- Visual Projector. The visual projector maps the output of the visual encoder into tokens that occupy the same space as the text features from word embedding. Various visual projectors are considered in Sec. 3.3. The visual projector is to be trained in our work." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.687, + 0.469, + 0.821 + ], + "angle": 0, + "content": "- Pre-trained Large Language Model. In this study, we utilize LLaVa [14], which is fine-tuned on Vicuna-7B [4] using instructional image-text pairs, for processing videos and texts. The causal mask is employed in all attention processes, incorporating the attention interplay between visual and text tokens. The pre-trained language model is frozen in the present study, and a new attention mechanism is developed to maintain a consistent distance to visual tokens for all textual tokens." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.551, + 0.47, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In our framework, the video is first encoded into visual tokens using the frozen visual encoder and the trainable visual projector. These visual tokens are then combined with text tokens, which are projected with word embedding from the prompt and question sentences. The combined visual" + }, + { + "type": "code_caption", + "bbox": [ + 0.5, + 0.091, + 0.894, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of EDVT attention in a PyTorch-like style." + }, + { + "type": "code", + "bbox": [ + 0.501, + 0.13, + 0.882, + 0.362 + ], + "angle": 0, + "content": "# x: hidden input in each attention layer\n# v_mask: indcite which input are from visual tokens\ndef edvtattention_layer(x, v_mask):\n # query, answer, value projection\n q, k, v = qkv_proj(x)\n # apply RoPE for query and key inputs\n r_q, r_k = rope(q, k)\n # attention weights without RoPE\n attention = bmm(q.T, k)\n # attention weights with RoPE\n rattention = bmm(r_q, r_k)\n # Merge attention weights based on visual token\n attention = v_mask * attention + (1 - v_mask) *\n rattention\n attention = Softmax(attention, dim=-1)\n # Update representation based on attention weights\n v = bmm(attention, v)\n out = linear_proj(v)\n return out" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.453 + ], + "angle": 0, + "content": "and text tokens are then input into the language model to generate answers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.783, + 0.483 + ], + "angle": 0, + "content": "3.2. Equal Distance to Visual Tokens" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Preliminary. For the concatenate visual-text input, three linear projection layers are applied in every attention layer to produce the query \\(\\mathbf{Q}\\), key \\(\\mathbf{K}\\), and value \\(\\mathbf{V}\\). Let \\(\\mathbf{q}_j\\) be \\(j^{th}\\) query in \\(\\mathbf{Q}\\). The conventional attention mechanism updates the input by initially determining the similarity between the query and key, then applying the attention weights to the value state. Specifically, for the \\(j^{th}\\) position, the update procedure can be articulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.63, + 0.892, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.718 + ], + "angle": 0, + "content": "where \\( \\text{sim}(\\mathbf{q}_j, \\mathbf{k}_i) = \\exp(\\mathbf{q}_j^T \\mathbf{k}_i / \\sqrt{d}) \\). Here, a causal mask is utilized so that the \\( j^{th} \\) query can only attend to the key positioned less than \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "EDVT-Attention. The vanilla attention model lacks positional awareness, with no encoded relative distance for natural language processing. In contrast, Rotary Positional Embeddings (RoPE) [23] encodes the position data of tokens using a rotation matrix, which inherently includes an explicit relative position dependency. Within each attention layer, RoPE is implemented across all projected query and key inputs in order to compute the attention weights via leveraging relative distance between tokens. The query situated at the \\(j^{th}\\) position incorporates rotary position embedding through \\(\\tilde{\\mathbf{q}}_j = \\mathbf{R}_j\\mathbf{q}_j\\), wherein \\(\\mathbf{R}_j \\in \\mathbb{R}^{d \\times d}\\) represents the rotary matrix for the \\(j^{th}\\) position. Consequently, the attention involving" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.895, + 0.437 + ], + "angle": 0, + "content": "Figure 2. The framework of VISTA-LLAMA. The visual encoder and large language model are both frozen ( ) during training, while the projector is trainable ( ) to map video into the language's space. The attention operation in each layer is present on the right part. Only the text tokens are applied with rotary position embedding to include relative distance information. The attention weights between visual and language tokens are calculated without the rotary position embedding. The casual mask is applied to the bottom-right attention weights." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.423, + 0.479 + ], + "angle": 0, + "content": "relative position embedding is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.487, + 0.472, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} _ {r o p e} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i}) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i})}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.472, + 0.688 + ], + "angle": 0, + "content": "RoPE inherently integrates relative position data via the multiplication of rotation matrices rather than appending it to the input as a positional embedding. In natural language processing, this relative proximity between two words is vital, given that remote words should have less influence than adjacent words when generating the current word. However, using the same attention mechanism for visual and text tokens may result in unintentional text generation, a phenomenon often referred to as hallucination in LLMs. With multi-modal input, the generated text should depend on the visual content, disregarding the influence of relative distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.472, + 0.794 + ], + "angle": 0, + "content": "To alleviate this issue and enhance the video understanding with LLMs, we introduce the EDVT-Attention where the equal distance to visual tokens is maintained while the relative distance between text tokens is attained. As shown in Figure 2, the rotary position embedding is only applied on text tokens. Let \\(\\mathcal{V}\\) and \\(\\mathcal{T}\\) be the set of visual and text tokens. The EDVT-Attention is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.801, + 0.47, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} _ {e d v t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{q}}_j\\) and \\(\\tilde{\\mathbf{k}}_j\\) denote the query and key applied with rotary embedding, separately. The distance between language tokens is determined using the rotary matrix, while" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.537 + ], + "angle": 0, + "content": "the correlation between visual and textual inputs remains unaffected by the rotary embedding. This improves the influence of visual information on long-term text generation and decreases the incidence of hallucinations wherein the fabricated content is absent in the videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.613 + ], + "angle": 0, + "content": "For ease comprehension of how EDVT-Attention works, Algorithm 1 exhibits the pseudo-code in the decoder layer of LLMs. It involves the combinations of attention weights prior to and subsequent to the implementation of rotary embedding with the visual mask." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.748, + 0.637 + ], + "angle": 0, + "content": "3.3. Sequential Visual Projector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.795 + ], + "angle": 0, + "content": "The visual projector aims to map video features into the language embedding space, allowing for the fusion and processing of visual and textual inputs by the substantial language model. As shown in Figure 3, earlier visual projectors either employ the linear layer or the query transformer (Q-Former) [12] to directly convert frame features into language tokens. However, the lack of temporal relationship in these methods impedes thorough video understanding in LLMs. We introduce the sequential visual projector in Figure 3 to encode the temporal context into the visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Let \\(\\mathbf{o} \\in \\mathbb{R}^{t \\times l \\times d}\\) be the extracted video feature of length \\(t\\) where \\(l\\) denotes the length of visual tokens. We use a Q-Former \\(f_{Q}\\) to map each frame into the fixed length of \\(k\\) representations \\(\\mathbf{x}_i = f_{Q}(\\mathbf{o}_i, \\mathbf{p}) \\in \\mathbb{R}^{k \\times d}\\) where \\(\\mathbf{p} \\in \\mathbb{R}^{k \\times d}\\) is the learnable query embedding. In the prior approach, all projected frame features are merely combined with word tokens to serve as mixed input for LLMs. For encoding temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.089, + 0.418, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.223, + 0.372, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Figure 3. Comparison of three visual projectors. (a): Each frame feature is projected into the visual tokens independently with the linear projection. (b): Q-Former uses shared learnable query tokens to separately map each frame into fixed-length tokens. (c): The sequential Q-Former with linear projection layer to enable temporal modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.525 + ], + "angle": 0, + "content": "context, we utilize the previously projected frame feature as the query to attend to the current frame feature. The current projected frame feature is then updated accordingly:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.538, + 0.47, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = f _ {Q} \\left(\\mathbf {o} _ {i}, \\mathbf {x} _ {t - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.469, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{t - 1}\\in \\mathbb{R}^{k\\times d}\\) represents the previous projected frame feature. This allows the visual tokens to encode the temporal relationship, as the current frame's visual token is generated using the previous feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Moreover, we can represent the entire video with fewer tokens by merely sampling a small number of visual tokens. This method tackles the challenge of encoding lengthy videos by employing sequential encoding. By integrating prior visual context into subsequent visual tokens, we can accomplish an adequate representation of the entire video through sparse sampling of projected frame features. Besides, this technique permits larger language models to manage much longer videos without length limitation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.294, + 0.787 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We fine-tune the model based on LLaVA [14] and use \\(100K\\) video instruction pairs provided by [18]. We add Q-Former initialized from BLIP2 [12] to project frame features into fixed-length tokens. We test our model with both ViT-L-14 [20] and EVA-CLIP-g [24] visual encoders in experiments. We only update the visual projector, which contains the Q-Former and linear projection layer to project the video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.183 + ], + "angle": 0, + "content": "features to the LLMs' input space. The visual backbone and the language model are frozen during training. The VISTA-LLAMA is fine-tuned for 3 epochs on 8 A100 80GB GPUs with a learning rate of \\(2e^{-5}\\) and an overall batch size of 32. We run all the inference experiments with FP16 to save memory and faster testing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.633, + 0.212 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.219, + 0.691, + 0.236 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.842, + 0.257 + ], + "angle": 0, + "content": "Datasets. Our method is evaluated on four datasets:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.894, + 0.335 + ], + "angle": 0, + "content": "- NExT-QA [30] is designed to advance video understanding from descrii ng to explaining the temporal actions. It comprises 5,440 videos and approximately \\(52K\\) manually annotated QA pairs, which are categorized into temporal (Tem.), causal (Cau.), and descriptive (Des.) questions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.336, + 0.893, + 0.395 + ], + "angle": 0, + "content": "- MSVD-QA [32] is a dataset built upon Microsoft Research Video Description Corpus [3], commonly used in video caption tasks. The MSVD-QA dataset comprises a total of 1,970 video clips with 50,505 QA pairs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.396, + 0.893, + 0.44 + ], + "angle": 0, + "content": "- MSRVTT-QA [32] is based on MSR-VTT dataset [33], which includes \\(10K\\) videos and \\(243K\\) QA pairs with larger and has more complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.441, + 0.894, + 0.501 + ], + "angle": 0, + "content": "- ActiviytNet-QA [37] is a fully annotated and large-scale videoQA dataset. It contains 58K QA pairs derived from 5,800 complex web videos derived from the popular ActivityNet dataset [11]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.26, + 0.894, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.503, + 0.894, + 0.533 + ], + "angle": 0, + "content": "Evaluation. We adopt two metrics to evaluate the performance of video-language models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.536, + 0.894, + 0.655 + ], + "angle": 0, + "content": "- Open-Ended Zero-Shot Question-Answer Evaluation. As video-language models generate responses of varying lengths to open-ended questions, it is challenging to evaluate models with traditional word matching strategy. We employ LLM-Assisted evaluation, in line with [18], for fair comparison. Given the question, correct answer, and predicted answer, GPT3.5-turbo-0613 is used to return True or False judgement and relative score (0 to 5)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.657, + 0.894, + 0.807 + ], + "angle": 0, + "content": "- Video-Based Text Generation Benchmarking. We evaluate the text generation performance following [18] from five aspects: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding. The test set for this evaluation is based on the ActivityNet-200 [7], featuring videos with rich, dense descriptive captions and associated question-answer pairs from human annotations. The evaluation pipeline is also built with the GPT-3.5 model and relative score (0 to 5) is generated." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.536, + 0.894, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.785, + 0.833 + ], + "angle": 0, + "content": "4.2. Comparison to State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "For the zero-shot open-ended video question answering tasks, we compare our model with FrozenBiLM [34], VideoChat [13], LLaMA Adapter [40], VideoLLaMA [39], VideoChatGPT [18], and MovieChat [22] in Tab. 1. FrozenBiLM" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13155" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.875, + 0.231 + ], + "angle": 0, + "content": "
MethodNExT-QA [30]MSVD-QA [32]MSRVTT-QA [32]ActivityNet-QA [37]
AccuracyScoreAccuracyScoreAccuracyScoreAccuracyScore
FrozenBiLM [34]--32.2-16.8-24.7-
Video Chat [13]56.23.256.32.845.02.526.52.2
LLaMA Adapter [40]--54.93.143.82.734.22.7
Video LLaMA [39]--51.62.529.61.812.41.1
MovieChat [22]49.92.761.02.949.72.851.53.1
Video-ChatGPT [18]54.63.264.93.349.32.835.22.7
VISTA-LLAMA (Ours)60.73.465.33.660.53.348.33.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.207, + 0.242, + 0.761, + 0.256 + ], + "angle": 0, + "content": "Table 1. Comparison with SoTA methods on zero-shot VideoQA. See §4.2 for more details." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.28, + 0.472, + 0.376 + ], + "angle": 0, + "content": "
MethodCr.Cs.De.Ct.Te.
Video Chat [13]2.232.242.502.531.94
LLaMA Adapter [40]2.032.152.322.301.98
Video LLaMA [39]1.961.792.182.161.82
Video-ChatGPT [18]2.402.372.522.621.98
VISTA-LLAMA (Ours)2.442.312.643.182.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.385, + 0.472, + 0.456 + ], + "angle": 0, + "content": "Table 2. Quantitative results on video-based text generation with different video-language methods (§4.2). For clarity, five scores are reported (\"Cr\": Correctness of Information, \"Cs\": Consistency, \"De\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te\": Temporal Understanding)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.472, + 0.701 + ], + "angle": 0, + "content": "adapts frozen the bidirectional language model, showing promising results in zero-shot VideoQA settings. Other compared models are built on recent large auto-regressive language models. Despite pre-existing models have produced substantial results, VISTA-LLAMA consistently outperforms them, achieving state-of-the-art (SoTA) performance across three datasets: NExT-QA [30], MSVD-QA [32], and MSVTT-QA [32]. Our method obtains the highest results, with \\(60.5\\%\\) accuracy on MSRvTT, elevating the performance of the second-best model by nearly \\(10\\%\\). Additionally, our method attains \\(60.7\\%\\) accuracy on NExT-QA, markedly superior to Video-ChatGPT [18]. These results demonstrate VISTA-LLAMA's capability to comprehend video content and produce precise answers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We present the results of the evaluation of video-based text generation in Tab. 2. The results reveal its competent performance across all aspects when compared with the recent video-language models, Video Chat[13], VideoLLaMA [39], and Video-ChatGPT [18]. Despite being trained with identical datasets, our model outperforms Video-ChatGPT in four aspects. Our method offers a more comprehensive interpretation, and its responses are more in tune with the overarching context of the video content than comparable approaches. By utilizing the EDVT-Attention and sequential temporal modeling techniques, our model demonstrates a strong ability to generate text that is contextually appropriate, detailed, and includes precise timing for video inputs." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.285, + 0.849, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.461, + 0.896, + 0.545 + ], + "angle": 0, + "content": "Figure 4. Comparison of different design choices on NExTQA [30](\\(\\S\\)4.3). For clarity, accuracy of base model and two variants are given (\"Baseline\": the frozen LLaVA [14] with trainable Q-Former [12], \"w Seq\": base model with sequential visual projector, \"w Seq&EDVT\": base model with both sequential visual projector and EDVT-Attention)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.662, + 0.59 + ], + "angle": 0, + "content": "4.3. Abalation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Effect of Design Choices. We investigate the effect of our two designs, including the equal-distance to visual-tokens (EDVT) attention and the sequential visual projector. As shown in Figure 4, our model is tested on NExT-QA [30], which comprises three different types of questions. The baseline model employs LLaVA [14] as the language model, and Q-Former in BLIP-2 [12] as the visual projector. We first incorporate temporal modeling into the baseline, utilizing the projected tokens from the previous frame as the query tokens to generate visual tokens for the current frame, a setup denoted as \"w Seq\" in Figure 4. Further enhancing this version with EDVT-Attention, we introduce \"w Seq&EDVT\", ultimately forming our final VISTA-LLAMA. The results affirm the efficacy of our design across all question types." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "EDVT-Attention. We here validate the effectiveness of our core EDVT-Attention. Table 3 reports the comparison results of different models in combination with our EDVT-Attention on NExT-QA [30]. Initially, we integrate the EDVT-Attention into Video-ChatGPT [18], where only the attention component is substituted with our EDVT-Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.089, + 0.472, + 0.202 + ], + "angle": 0, + "content": "
MethodNExT-QA [30]
Tem.Cau.Des.Avg.
Video-ChatGPT [18]37.665.154.954.6
+ EDVT-Attention39.5 (+1.9)72.8 (+7.7)54.859.3 (+4.7)
VISTA-LLAMA (ViT-L-14)34.069.142.253.6
+ EDVT-Attention36.8 (+2.8)72.2 (+3.1)47.7 (+5.5)56.5 (+2.9)
VISTA-LLAMA (EVA-CLIP-g)34.365.855.954.1
+ EDVT-Attention40.7 (+6.4)72.3 (+6.5)57.0 (+1.1)59.7 (+5.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.268 + ], + "angle": 0, + "content": "Table 3. Comparison of EDVT-Attention design with diffent visual encoders on NExT-QA [30] (§4.3). For clarity, accuracy of all questions and three types of questions are reported (\"Tem\": temporal, \"Cau\": causal, \"Des\": descriptive, \"Avg.\": average)." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.285, + 0.272, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.412, + 0.232, + 0.422 + ], + "angle": 0, + "content": "(a) Vanilla attention" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.285, + 0.462, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.412, + 0.441, + 0.422 + ], + "angle": 0, + "content": "(b) EDVT-Attention (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.472, + 0.492 + ], + "angle": 0, + "content": "Figure 5. Comparison of attention weights for varying context lengths. Lighter colors represent higher weights. To improve clarity, we have combined visual token weights into the first four tokens. We recommend zooming in for optimal viewing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.612 + ], + "angle": 0, + "content": "during training. Further deployment of the EDVT-Attention into Video-ChatGPT [18] yields significant performance gains (e.g., \\(65.1\\%\\) escalates to \\(72.8\\%\\) for causal questions). The accuracy is enlarged across all three setting types, indicating that our EDVT-Attention considerably enhances the multi-modal understanding in large language models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.472, + 0.763 + ], + "angle": 0, + "content": "We further illustrate the attention weights of both EDVT-Attention and conventional attention in Figure 5. For this experiment, we employ 128 visual tokens and integrate the attention weights into the first four tokens to enhance visualization. The attention weights assigned to visual tokens are markedly greater in our EDVT-Attention compared to those in conventional attention. This indicates that the impact of visual tokens on language tokens is considerably more substantial under the proposed EDVT-Attention, shedding light on why the performance notably improves with our design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Visual Projector. The visual projector is trainable within the model. Therefore, we have examined the impacts of different visual projector designs, as shown in Table 4. We evaluated three visual projector variants. The Q-Former, initialized using BLIP-2 [12], considerably outperforms the Q-Former that was initialized with BERT [5]. This indicates that visual projectors, when pre-trained on image-text pairs, can significantly enhance video comprehension. The model's performance peaks when integrating the sequential design," + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.895, + 0.177 + ], + "angle": 0, + "content": "
Visual projectorNExT-QA [30]
Tem.Cau.Des.Avg.
Linear Projector37.665.154.954.6
Q-Former (BERT init.)35.262.749.251.8
Q-Former (BLIP-2 init.)34.365.855.954.1
SeqQ-Former (BLIP-2 init.)36.268.551.155.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.187, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Table 4. Comparison of different visual projectors on NExTQA [30]. The linear projector is initialized with pre-trained weights in LLaVa [14]. BERT init. and BLIP-2 init. indicate that the visual projector is initialized with weights from BERT [5] and BLIP-2 [12]. SeqQ-Former is the proposed sequential visual projector. See §4.3 for more details." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.287, + 0.892, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.526, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Figure 6. The effect of training visual tokens on NExT-QA [30]. Accuracy of all questions and three types of questions, including temporal (Tem.), descriptive (Des.), and causal (Cau.), are presented with different colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.642 + ], + "angle": 0, + "content": "referred to as \"SeqQ-Former\" in the figure, outperforming all other visual projectors in terms of accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Number of Training Visual Tokens. We assess the impact of visual tokens on NExT-QA as illustrated in Figure 6. In this study, a frame is converted into 32 visual tokens using Q-Former. Frames were sampled at various timestamps throughout the training process. The overall accuracy improves as more visual tokens are incorporated into the language model. However, for temporal queries, precision declined with an increased number of visual tokens. The accuracy of temporal reasoning questions is notably lower than that of other question types. Temporal reasoning presents more difficulties, and the language model may not excel at the temporal modeling of visual tokens. In terms of descriptive question types, the accuracy steadily increases as the model gains more visual information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We visualize the generated responses of Video-ChatGPT [18] and VISTA-LLAMA for different videos in Figure 7. The frames sampled at different times" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.53, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Figure 7. Visualization results on different video questions. The questions and annotated answers are located on the left side. The generated text from Video-ChatGPT [31] and our model is presented in the green and orange boxes, respectively. See §4.3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.473, + 0.889 + ], + "angle": 0, + "content": "tamps is presented, with questions listed below the images. In the first video, the mobile phone is turned, sequentially exposing each side over time. The question asks for the side of the mobile phone at the end of the video. Video-ChatGPT gives the wrong response, as the camera is never present at the bottom left side in the video. The hallucination occurs in Video-ChatGPT, whereas our model predicts the answer correctly. The responses in three cases demonstrate that Video-ChatGPT provides unrelated answers that do not correspond to the video content. In the second case, the person in pink is always standing while the man is sitting on the ground, but Video-ChatGPT incorrectly states that the person in pink is also sitting on the ground. In the third video, the man is consistently walking up the mountain and never walking down. However, Video-ChatGPT falsely outputs that the man continues walking down the mountain. In contrast, our model provides the correct and reliable responses. This demonstrates that our model significantly reduces hallucination in video understanding and delivers more accurate responses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.584, + 0.621, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.896, + 0.838 + ], + "angle": 0, + "content": "In this work, we present VISTA-LLAMA to improve the video understanding in large language model. A new vision-aware attention is introduced to maintain same relative distance between all visual tokens and language tokens. In addition, we propose a sequential visual projector to map video into the language space to enable temporal modeling. Experiments on several video question answering task demonstrate that our designs significant improves the current SoTA. Current work is built on the pre-trained image-text architecture, the power of our designs could be enlarged when applied in pre-training stage. Furthermore, the performance of our sequential encoder on longer videos can also be assessed. We would consider evaluating the model in more tasks and applied the attention mechanism to more vision language models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554, U2336212), and China Postdoctoral Science Foundation (524000-X92302)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 23716-23736, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Annual Meeting of the Association for Computational Linguistics, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[4] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.472, + 0.442 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.472, + 0.499 + ], + "angle": 0, + "content": "[6] Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.472, + 0.57 + ], + "angle": 0, + "content": "[7] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[8] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30: 681–694, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[9] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yezhou Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14773-14783, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[10] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953–14962, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. ArXiv, abs/2304.08485, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.895, + 0.217 + ], + "angle": 0, + "content": "[15] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zeroshot video grounding with pseudo query lookup and verification. IEEE Transactions on Image Processing, 33:1643-1654, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.895, + 0.273 + ], + "angle": 0, + "content": "[16] Fan Ma, Linchao Zhu, and Yi Yang. Weakly supervised moment localization with decoupled consistent concept prediction. International Journal of Computer Vision, 130(5): 1244-1258, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.895, + 0.316 + ], + "angle": 0, + "content": "[17] Fan Ma, Xiaojie Jin, Heng Wang, Jingjia Huang, Linchao Zhu, Jiashi Feng, and Yi Yang. Temporal perceiving video-language pre-training. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.895, + 0.372 + ], + "angle": 0, + "content": "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.895, + 0.398 + ], + "angle": 0, + "content": "[19] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.895, + 0.482 + ], + "angle": 0, + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.895, + 0.538 + ], + "angle": 0, + "content": "[21] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.895, + 0.607 + ], + "angle": 0, + "content": "[22] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.895, + 0.663 + ], + "angle": 0, + "content": "[23] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.895, + 0.706 + ], + "angle": 0, + "content": "[24] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.895, + 0.762 + ], + "angle": 0, + "content": "[25] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.895, + 0.845 + ], + "angle": 0, + "content": "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[27] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, Felix Hill, and Zacharias Janssen. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Luccioni, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[29] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[30] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9772–9781, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.333 + ], + "angle": 0, + "content": "[31] Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. ArXiv, abs/2304.01196, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[32] D. Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. Proceedings of the 25th ACM international conference on Multimedia, 2017. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.472, + 0.461 + ], + "angle": 0, + "content": "[33] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5288-5296, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[34] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[35] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[36] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[37] Zhou Yu, D. Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[38] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[39] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[40] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[41] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.894, + 0.161 + ], + "angle": 0, + "content": "Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models. ArXiv, abs/2205.01068, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13160" + } + ] +] \ No newline at end of file diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_origin.pdf b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05d153ffd9c3e5687f7449d6d899a9ae1625c843 --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/7b3a6ad4-efd5-421f-99e3-ea5d44dea958_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36d52da5ee47e6488c8a0a0f37df59cc81b820378a7209955ce9b084d5569351 +size 4410545 diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/full.md b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e9894c5ea8a3b9110e56ec5c5d2ef1cb1f229e49 --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/full.md @@ -0,0 +1,281 @@ +# VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens + +Fan Ma $^{1}$ , Xiaojie Jin $^{2*}$ , Heng Wang $^{2}$ , Yuchen Xian $^{1}$ , Jiashi Feng $^{2}$ , Yi Yang $^{1*}$ + +1Zhejiang University 2ByteDance Inc. + +# Abstract + +Recent advances in large video-language models have displayed promising outcomes in video comprehension. Current approaches straightforwardly convert video into language tokens and employ large language models for multi-modal tasks. However, this method often leads to the generation of irrelevant content, commonly known as "hallucination", as the length of the text increases and the impact of the video diminishes. To address this problem, we propose VISTA-LLAMA, a novel framework that maintains the consistent distance between all visual tokens and any language tokens, irrespective of the generated text length. VISTA-LLAMA omits relative position encoding when determining attention weights between visual and text tokens, retaining the position encoding for text and text tokens. This amplifies the effect of visual tokens on text generation, especially when the relative distance is longer between visual and text tokens. The proposed attention mechanism significantly reduces the chance of producing irrelevant text related to the video content. Furthermore, we present a sequential visual projector that projects the current video frame into tokens of language space with the assistance of the previous frame. This approach not only captures the temporal relationship within the video, but also allows less visual tokens to encompass the entire video. Our approach significantly outperforms various previous methods (e.g., Video-ChatGPT, MovieChat) on four challenging open-ended video question answering benchmarks. We reach an accuracy of 60.7 on the zero-shot NExT-QA and 60.5 on the zero-shot MSRVTT-QA, setting a new state-of-the-art performance. This project is available at https://jinxxian.github.io/Vista-LLaMA. + +# 1. Introduction + +The surge in multi-modal vision-and-language models [1, 12, 14], capable of comprehending both visual (e.g., image/video) and language data, can be attributed to the recent + +![](images/6d453bf5d37aa3c2ebd46200975efba1f530d25463bb66fd20963bba9ffca429.jpg) +Figure 1. Video language processing with LLaMA [26] and our VISTA-LLAMA. The vanilla LLaMA treats visual tokens (☑) the same as other language tokens (☑), weakening the impact for tokens in long distance. Our model retains the same mechanism for language tokens and strengthens the impact of the visual tokens. The intensity of the impact of each token is conveyed through the depth of the line color $(\rightarrow)$ . Our model provides the accurate response for the presented scenario. + +achieves of large language models (LLMs) such as GPT [8], GLM [38], and LLaMA [40]. Video-language models (video-LMs) pose greater challenges in scaling due to increased computational and annotation costs compared to image-language models (image-LMs). Recent research has therefore focused on effectively training video-LMs by utilizing pre-trained image-LMs [22, 40]. + +Video-LMs benefit from this warm-start approach to enhance visual representation learning by projecting video frames into language space and treating videos as several prompt language tokens [18, 34]. However, this diminishes the visual impact of videos on text generation and lacks explicit temporal modeling in videos. The generated text is often not related to the video content, as depicted in Figure 1, a phenomenon known as hallucination in language processing [6]. The distance between the generated text token and the visual tokens in large language models may + +be a contributing factor, especially when the visual tokens are distant from the generated text. Additionally, handling longer videos poses a challenge because of the constraints on context length in large language models, where numerous visual tokens take up a substantial portion of the context. + +We present VISTA-LLAMA, an innovative video language framework to tackle above issues. The primary concept is to preserve equal distance between all visual tokens and any language tokens, while also retaining the relative distance between any two language tokens, as depicted in Figure 1. The rotary position embedding is only applied to language tokens to capture relative distance when calculating similarity between language tokens. When computing similarity between visual and text tokens, the rotary position embedding is removed to reduce the impact of relative distance. With the equal distance to visual tokens (EDVT) attention, the impact of visual cues on text production is amplified without compromising the text production capability. The experiments also show that our design produces more precise text description for input videos and the phenomenon of visual hallucinations occurring in language models has been greatly reduced. + +To further improve the temporal modeling of video, we introduce a sequential visual projector. Instead of mapping each frame of the video independently into fixed-length visual tokens, which ignores the temporal relationship between frames, we generate visual tokens for each frame using the previous projected visual tokens. It not only incorporates the temporal relationship between frames into the language model to improve video comprehension without any extra parameters, but also enables the language model to encode longer videos with fewer visual tokens by sampling fewer projected frames. + +We demonstrate the effectiveness of VISTA-LLAMA on four challenging video question answering (QA) benchmarks, where our method outperforms several previous works, and achieves the state-of-the-art in zero-shot NExTQA [30], and MSRVTT [32]. We also show that our attention design and the temporal modeling mechanism improves capacity of large language model on video-question answering by a large margin. Comprehensive experiments are conducted to demonstrate the effectiveness of our designs. We summarize our contributions as follows: + +- We introduce a novel video-language model, dubbed as VISTA-LLAMA, to enhance the video understanding and facilitate temporal modeling within the language model. +- A novel multi-modal attention is proposed to enable reliable video text generation by maintaining equal distance to visual tokens. Additionally, temporal modeling is facilitated by employing a sequential visual projector. +- Our method exhibits superior empirical performance, establishing new benchmarks for zero-shot open-ended video question answering tasks. + +- A detailed analysis further explains the design choices inherent in our proposed framework, contributing significantly to a better comprehension of multi-modal dynamics in large language models. + +# 2. Related Work + +# 2.1. Large Language Models + +Large language models (LLMs) [2, 19, 27, 28, 42] have demonstrated exceptional proficiency in understanding language and reasoning, resulting in the generation of high-quality natural language text in diverse domains. LLMs have already initiated a technological revolution and have found extensive application in various domains. Additionally, a series of open source large models, including LLaMA [26], and OPT [41], have significantly contributed to technological advancements and made remarkable contributions to the NLP community. Leveraging the foundation established by these impressive LLMs, researchers have further expanded their capabilities and developed exceptional models for diverse NLP tasks (e.g. Vicuna [4]). Our work is also built upon these remarkable LLMs, and we equip the language models with a novel attention mechanism and a sequential visual projector that enhances their abilities to comprehend visual content in videos. + +# 2.2. Multi-modal Large Language Models + +Researchers have been diligently exploring the application of LLMs for processing multi-modal problems [9, 13, 15, 17]. The existing approaches can be classified into two primary categories. In the first category, LLMs are treated as controllers while the multi-modal models as tools [10]. When presented with the specific task, the LLMs first interact with user instructions and decide which tools to employ. Subsequently, it generates comprehensive responses by amalgamating the outcomes derived from these readily available multi-modal models. These approaches, such as Visual ChatGPT [29], HuggingGPT [21], and DoraemonGPT [35], have shown impressive results on various multi-modal tasks without training models. For the second category, large-scale multi-modal models are trained on the multi-modal data. The principle behind this line of researches is to align other modal pre-trained models with textual LLMs. Flamingo [1] incorporates a perceiver resampler and a gated cross-attention layer to connect a frozen image encoder with the LLM. BLIP2 [12] introduces a Q-Former to map each image into fixed-length tokens in the language embedding space via the learned queries. LLaVA [14], mPLUG-owl [36], and MiniGPT4 [42] develop the image-LLMs utilizing image-instruction training pairs. Video Chat [13] extends image encoders to enable large models to comprehend visual content in videos. Video-ChatGPT [18] is trained on video instructional data to give appropriate answers for multi-modal + +inputs. In this work, we improve the understanding of videos in a video-language model by incorporating temporal modeling and a novel multi-modal attention mechanism. + +# 2.3. Video Question Answering + +Video Question Answering (VideoQA) is a task that involves answering language questions that are based on a given video [32]. It requires the ability to understand and reason across different semantic levels, which in turn demands a capacity for multi-modal understanding. Previous VideoQA benchmarks primarily concentrated on short videos, asking questions according to the specified visual facts (e.g., location and objects) [16, 33]. Recently, several new benchmarks have been proposed to focus on resolving temporal and causal questions in longer video clips [25]. NExT-QA [30] is an example of such a benchmark, aiming to uncover the causalities or intentions of particular events, and infer subsequent actions within the video. In this work, we investigate the video-language model on zero-shot open-ended VideoQA where no training question-answer pairs are provided in the training stage. + +# 3. Method + +# 3.1. Overview + +VISTA-LLAMA comprises three fundamental components: a visual encoder, a visual projector, and a pre-trained large language model. Figure 2 shows an overview of VISTA-LLAMA. The components' design and implementation details are provided below: + +- Visual Encoder. We utilize pre-trained EVA-CLIP-g [24] and ViT-L [20] as visual encoder. The last layer of ViT encoder is removed because it specializes in aggregating the features of the first token for contrastive learning. +- Visual Projector. The visual projector maps the output of the visual encoder into tokens that occupy the same space as the text features from word embedding. Various visual projectors are considered in Sec. 3.3. The visual projector is to be trained in our work. +- Pre-trained Large Language Model. In this study, we utilize LLaVa [14], which is fine-tuned on Vicuna-7B [4] using instructional image-text pairs, for processing videos and texts. The causal mask is employed in all attention processes, incorporating the attention interplay between visual and text tokens. The pre-trained language model is frozen in the present study, and a new attention mechanism is developed to maintain a consistent distance to visual tokens for all textual tokens. + +In our framework, the video is first encoded into visual tokens using the frozen visual encoder and the trainable visual projector. These visual tokens are then combined with text tokens, which are projected with word embedding from the prompt and question sentences. The combined visual + +Algorithm 1 Pseudocode of EDVT attention in a PyTorch-like style. +```python +# x: hidden input in each attention layer +# v_mask: indcite which input are from visual tokens +def edvtattention_layer(x, v_mask): + # query, answer, value projection + q, k, v = qkv_proj(x) + # apply RoPE for query and key inputs + r_q, r_k = rope(q, k) + # attention weights without RoPE + attention = bmm(q.T, k) + # attention weights with RoPE + rattention = bmm(r_q, r_k) + # Merge attention weights based on visual token + attention = v_mask * attention + (1 - v_mask) * + rattention + attention = Softmax(attention, dim=-1) + # Update representation based on attention weights + v = bmm(attention, v) + out = linear_proj(v) + return out +``` + +and text tokens are then input into the language model to generate answers. + +# 3.2. Equal Distance to Visual Tokens + +Preliminary. For the concatenate visual-text input, three linear projection layers are applied in every attention layer to produce the query $\mathbf{Q}$ , key $\mathbf{K}$ , and value $\mathbf{V}$ . Let $\mathbf{q}_j$ be $j^{th}$ query in $\mathbf{Q}$ . The conventional attention mechanism updates the input by initially determining the similarity between the query and key, then applying the attention weights to the value state. Specifically, for the $j^{th}$ position, the update procedure can be articulated as follows: + +$$ +\operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) _ {j} = \frac {\sum_ {i = 1} ^ {j} \operatorname {s i m} \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right) \mathbf {v} _ {i}}{\sum_ {i = 1} ^ {j} \operatorname {s i m} \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right)}, \tag {1} +$$ + +where $\text{sim}(\mathbf{q}_j, \mathbf{k}_i) = \exp(\mathbf{q}_j^T \mathbf{k}_i / \sqrt{d})$ . Here, a causal mask is utilized so that the $j^{th}$ query can only attend to the key positioned less than $j$ . + +EDVT-Attention. The vanilla attention model lacks positional awareness, with no encoded relative distance for natural language processing. In contrast, Rotary Positional Embeddings (RoPE) [23] encodes the position data of tokens using a rotation matrix, which inherently includes an explicit relative position dependency. Within each attention layer, RoPE is implemented across all projected query and key inputs in order to compute the attention weights via leveraging relative distance between tokens. The query situated at the $j^{th}$ position incorporates rotary position embedding through $\tilde{\mathbf{q}}_j = \mathbf{R}_j\mathbf{q}_j$ , wherein $\mathbf{R}_j \in \mathbb{R}^{d \times d}$ represents the rotary matrix for the $j^{th}$ position. Consequently, the attention involving + +![](images/d09f2ea609b6164d00534d7ab7d44470a68833a75da8785b60ea4b2d101b7ccc.jpg) +Figure 2. The framework of VISTA-LLAMA. The visual encoder and large language model are both frozen ( ) during training, while the projector is trainable ( ) to map video into the language's space. The attention operation in each layer is present on the right part. Only the text tokens are applied with rotary position embedding to include relative distance information. The attention weights between visual and language tokens are calculated without the rotary position embedding. The casual mask is applied to the bottom-right attention weights. + +relative position embedding is expressed as follows: + +$$ +\operatorname {A t t e n t i o n} _ {r o p e} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) _ {j} = \frac {\sum_ {i = 1} ^ {j} s i m (\mathbf {R} _ {j} \mathbf {q} _ {j} , \mathbf {R} _ {i} \mathbf {k} _ {i}) \mathbf {v} _ {i}}{\sum_ {i = 1} ^ {j} s i m (\mathbf {R} _ {j} \mathbf {q} _ {j} , \mathbf {R} _ {i} \mathbf {k} _ {i})}. \tag {2} +$$ + +RoPE inherently integrates relative position data via the multiplication of rotation matrices rather than appending it to the input as a positional embedding. In natural language processing, this relative proximity between two words is vital, given that remote words should have less influence than adjacent words when generating the current word. However, using the same attention mechanism for visual and text tokens may result in unintentional text generation, a phenomenon often referred to as hallucination in LLMs. With multi-modal input, the generated text should depend on the visual content, disregarding the influence of relative distance. + +To alleviate this issue and enhance the video understanding with LLMs, we introduce the EDVT-Attention where the equal distance to visual tokens is maintained while the relative distance between text tokens is attained. As shown in Figure 2, the rotary position embedding is only applied on text tokens. Let $\mathcal{V}$ and $\mathcal{T}$ be the set of visual and text tokens. The EDVT-Attention is formulated as: + +$$ +\operatorname {A t t e n t i o n} _ {e d v t} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) _ {j} = \frac {\sum_ {k _ {i} \in \mathcal {T}} s i m \left(\tilde {\mathbf {q}} _ {j} , \tilde {\mathbf {k}} _ {i}\right) \mathbf {v} _ {i} + \sum_ {k _ {i} \in \mathcal {V}} s i m \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right) \mathbf {v} _ {i}}{\sum_ {k _ {i} \in \mathcal {T}} s i m \left(\tilde {\mathbf {q}} _ {j} , \tilde {\mathbf {k}} _ {i}\right) + \sum_ {k _ {i} \in \mathcal {V}} s i m \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right)}, \tag {3} +$$ + +where $\tilde{\mathbf{q}}_j$ and $\tilde{\mathbf{k}}_j$ denote the query and key applied with rotary embedding, separately. The distance between language tokens is determined using the rotary matrix, while + +the correlation between visual and textual inputs remains unaffected by the rotary embedding. This improves the influence of visual information on long-term text generation and decreases the incidence of hallucinations wherein the fabricated content is absent in the videos. + +For ease comprehension of how EDVT-Attention works, Algorithm 1 exhibits the pseudo-code in the decoder layer of LLMs. It involves the combinations of attention weights prior to and subsequent to the implementation of rotary embedding with the visual mask. + +# 3.3. Sequential Visual Projector + +The visual projector aims to map video features into the language embedding space, allowing for the fusion and processing of visual and textual inputs by the substantial language model. As shown in Figure 3, earlier visual projectors either employ the linear layer or the query transformer (Q-Former) [12] to directly convert frame features into language tokens. However, the lack of temporal relationship in these methods impedes thorough video understanding in LLMs. We introduce the sequential visual projector in Figure 3 to encode the temporal context into the visual tokens. + +Let $\mathbf{o} \in \mathbb{R}^{t \times l \times d}$ be the extracted video feature of length $t$ where $l$ denotes the length of visual tokens. We use a Q-Former $f_{Q}$ to map each frame into the fixed length of $k$ representations $\mathbf{x}_i = f_{Q}(\mathbf{o}_i, \mathbf{p}) \in \mathbb{R}^{k \times d}$ where $\mathbf{p} \in \mathbb{R}^{k \times d}$ is the learnable query embedding. In the prior approach, all projected frame features are merely combined with word tokens to serve as mixed input for LLMs. For encoding temporal + +![](images/60b6b2d49ed7316370729aa4b1dd9dbf0caa6f7d7598e83eff2e24327aab1ee9.jpg) + +![](images/0f33c44ada250c220cacf8291c1044bd26943cc5fa96d9e6c765ee4d3eb91d45.jpg) +Figure 3. Comparison of three visual projectors. (a): Each frame feature is projected into the visual tokens independently with the linear projection. (b): Q-Former uses shared learnable query tokens to separately map each frame into fixed-length tokens. (c): The sequential Q-Former with linear projection layer to enable temporal modeling. + +context, we utilize the previously projected frame feature as the query to attend to the current frame feature. The current projected frame feature is then updated accordingly: + +$$ +\mathbf {x} _ {t} = f _ {Q} \left(\mathbf {o} _ {i}, \mathbf {x} _ {t - 1}\right), \tag {4} +$$ + +where $\mathbf{x}_{t - 1}\in \mathbb{R}^{k\times d}$ represents the previous projected frame feature. This allows the visual tokens to encode the temporal relationship, as the current frame's visual token is generated using the previous feature. + +Moreover, we can represent the entire video with fewer tokens by merely sampling a small number of visual tokens. This method tackles the challenge of encoding lengthy videos by employing sequential encoding. By integrating prior visual context into subsequent visual tokens, we can accomplish an adequate representation of the entire video through sparse sampling of projected frame features. Besides, this technique permits larger language models to manage much longer videos without length limitation. + +# 3.4. Implementation Details + +We fine-tune the model based on LLaVA [14] and use $100K$ video instruction pairs provided by [18]. We add Q-Former initialized from BLIP2 [12] to project frame features into fixed-length tokens. We test our model with both ViT-L-14 [20] and EVA-CLIP-g [24] visual encoders in experiments. We only update the visual projector, which contains the Q-Former and linear projection layer to project the video + +features to the LLMs' input space. The visual backbone and the language model are frozen during training. The VISTA-LLAMA is fine-tuned for 3 epochs on 8 A100 80GB GPUs with a learning rate of $2e^{-5}$ and an overall batch size of 32. We run all the inference experiments with FP16 to save memory and faster testing. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. Our method is evaluated on four datasets: + +- NExT-QA [30] is designed to advance video understanding from descrii ng to explaining the temporal actions. It comprises 5,440 videos and approximately $52K$ manually annotated QA pairs, which are categorized into temporal (Tem.), causal (Cau.), and descriptive (Des.) questions. +- MSVD-QA [32] is a dataset built upon Microsoft Research Video Description Corpus [3], commonly used in video caption tasks. The MSVD-QA dataset comprises a total of 1,970 video clips with 50,505 QA pairs. +- MSRVTT-QA [32] is based on MSR-VTT dataset [33], which includes $10K$ videos and $243K$ QA pairs with larger and has more complex scenes. +- ActiviytNet-QA [37] is a fully annotated and large-scale videoQA dataset. It contains 58K QA pairs derived from 5,800 complex web videos derived from the popular ActivityNet dataset [11]. + +Evaluation. We adopt two metrics to evaluate the performance of video-language models. + +- Open-Ended Zero-Shot Question-Answer Evaluation. As video-language models generate responses of varying lengths to open-ended questions, it is challenging to evaluate models with traditional word matching strategy. We employ LLM-Assisted evaluation, in line with [18], for fair comparison. Given the question, correct answer, and predicted answer, GPT3.5-turbo-0613 is used to return True or False judgement and relative score (0 to 5). +- Video-Based Text Generation Benchmarking. We evaluate the text generation performance following [18] from five aspects: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding. The test set for this evaluation is based on the ActivityNet-200 [7], featuring videos with rich, dense descriptive captions and associated question-answer pairs from human annotations. The evaluation pipeline is also built with the GPT-3.5 model and relative score (0 to 5) is generated. + +# 4.2. Comparison to State-of-the-Arts + +For the zero-shot open-ended video question answering tasks, we compare our model with FrozenBiLM [34], VideoChat [13], LLaMA Adapter [40], VideoLLaMA [39], VideoChatGPT [18], and MovieChat [22] in Tab. 1. FrozenBiLM + +
MethodNExT-QA [30]MSVD-QA [32]MSRVTT-QA [32]ActivityNet-QA [37]
AccuracyScoreAccuracyScoreAccuracyScoreAccuracyScore
FrozenBiLM [34]--32.2-16.8-24.7-
Video Chat [13]56.23.256.32.845.02.526.52.2
LLaMA Adapter [40]--54.93.143.82.734.22.7
Video LLaMA [39]--51.62.529.61.812.41.1
MovieChat [22]49.92.761.02.949.72.851.53.1
Video-ChatGPT [18]54.63.264.93.349.32.835.22.7
VISTA-LLAMA (Ours)60.73.465.33.660.53.348.33.3
+ +Table 1. Comparison with SoTA methods on zero-shot VideoQA. See §4.2 for more details. + +
MethodCr.Cs.De.Ct.Te.
Video Chat [13]2.232.242.502.531.94
LLaMA Adapter [40]2.032.152.322.301.98
Video LLaMA [39]1.961.792.182.161.82
Video-ChatGPT [18]2.402.372.522.621.98
VISTA-LLAMA (Ours)2.442.312.643.182.26
+ +Table 2. Quantitative results on video-based text generation with different video-language methods (§4.2). For clarity, five scores are reported ("Cr": Correctness of Information, "Cs": Consistency, "De": Detail Orientation, "Ct": Contextual Understanding, "Te": Temporal Understanding). + +adapts frozen the bidirectional language model, showing promising results in zero-shot VideoQA settings. Other compared models are built on recent large auto-regressive language models. Despite pre-existing models have produced substantial results, VISTA-LLAMA consistently outperforms them, achieving state-of-the-art (SoTA) performance across three datasets: NExT-QA [30], MSVD-QA [32], and MSVTT-QA [32]. Our method obtains the highest results, with $60.5\%$ accuracy on MSRvTT, elevating the performance of the second-best model by nearly $10\%$ . Additionally, our method attains $60.7\%$ accuracy on NExT-QA, markedly superior to Video-ChatGPT [18]. These results demonstrate VISTA-LLAMA's capability to comprehend video content and produce precise answers. + +We present the results of the evaluation of video-based text generation in Tab. 2. The results reveal its competent performance across all aspects when compared with the recent video-language models, Video Chat[13], VideoLLaMA [39], and Video-ChatGPT [18]. Despite being trained with identical datasets, our model outperforms Video-ChatGPT in four aspects. Our method offers a more comprehensive interpretation, and its responses are more in tune with the overarching context of the video content than comparable approaches. By utilizing the EDVT-Attention and sequential temporal modeling techniques, our model demonstrates a strong ability to generate text that is contextually appropriate, detailed, and includes precise timing for video inputs. + +![](images/2ca9f919bd9e213840828c73cd7203dfe8f87ae302086bf8a131692ec539df82.jpg) +Figure 4. Comparison of different design choices on NExTQA [30]( $\S$ 4.3). For clarity, accuracy of base model and two variants are given ("Baseline": the frozen LLaVA [14] with trainable Q-Former [12], "w Seq": base model with sequential visual projector, "w Seq&EDVT": base model with both sequential visual projector and EDVT-Attention). + +# 4.3. Abalation Study + +Effect of Design Choices. We investigate the effect of our two designs, including the equal-distance to visual-tokens (EDVT) attention and the sequential visual projector. As shown in Figure 4, our model is tested on NExT-QA [30], which comprises three different types of questions. The baseline model employs LLaVA [14] as the language model, and Q-Former in BLIP-2 [12] as the visual projector. We first incorporate temporal modeling into the baseline, utilizing the projected tokens from the previous frame as the query tokens to generate visual tokens for the current frame, a setup denoted as "w Seq" in Figure 4. Further enhancing this version with EDVT-Attention, we introduce "w Seq&EDVT", ultimately forming our final VISTA-LLAMA. The results affirm the efficacy of our design across all question types. + +EDVT-Attention. We here validate the effectiveness of our core EDVT-Attention. Table 3 reports the comparison results of different models in combination with our EDVT-Attention on NExT-QA [30]. Initially, we integrate the EDVT-Attention into Video-ChatGPT [18], where only the attention component is substituted with our EDVT-Attention + +
MethodNExT-QA [30]
Tem.Cau.Des.Avg.
Video-ChatGPT [18]37.665.154.954.6
+ EDVT-Attention39.5 (+1.9)72.8 (+7.7)54.859.3 (+4.7)
VISTA-LLAMA (ViT-L-14)34.069.142.253.6
+ EDVT-Attention36.8 (+2.8)72.2 (+3.1)47.7 (+5.5)56.5 (+2.9)
VISTA-LLAMA (EVA-CLIP-g)34.365.855.954.1
+ EDVT-Attention40.7 (+6.4)72.3 (+6.5)57.0 (+1.1)59.7 (+5.6)
+ +![](images/7e85abb7685829c2f4ccb8fb66e557e745fba1b4dc7145d952e2ef9d586dba89.jpg) +(a) Vanilla attention + +![](images/1934748432e33a3fa4b48770222b00e3b7e8e0ac52d38cc123813ee71096d4d8.jpg) +(b) EDVT-Attention (Ours) + +during training. Further deployment of the EDVT-Attention into Video-ChatGPT [18] yields significant performance gains (e.g., $65.1\%$ escalates to $72.8\%$ for causal questions). The accuracy is enlarged across all three setting types, indicating that our EDVT-Attention considerably enhances the multi-modal understanding in large language models. + +We further illustrate the attention weights of both EDVT-Attention and conventional attention in Figure 5. For this experiment, we employ 128 visual tokens and integrate the attention weights into the first four tokens to enhance visualization. The attention weights assigned to visual tokens are markedly greater in our EDVT-Attention compared to those in conventional attention. This indicates that the impact of visual tokens on language tokens is considerably more substantial under the proposed EDVT-Attention, shedding light on why the performance notably improves with our design. + +Visual Projector. The visual projector is trainable within the model. Therefore, we have examined the impacts of different visual projector designs, as shown in Table 4. We evaluated three visual projector variants. The Q-Former, initialized using BLIP-2 [12], considerably outperforms the Q-Former that was initialized with BERT [5]. This indicates that visual projectors, when pre-trained on image-text pairs, can significantly enhance video comprehension. The model's performance peaks when integrating the sequential design, + +Table 3. Comparison of EDVT-Attention design with diffent visual encoders on NExT-QA [30] (§4.3). For clarity, accuracy of all questions and three types of questions are reported ("Tem": temporal, "Cau": causal, "Des": descriptive, "Avg.": average). + +
Visual projectorNExT-QA [30]
Tem.Cau.Des.Avg.
Linear Projector37.665.154.954.6
Q-Former (BERT init.)35.262.749.251.8
Q-Former (BLIP-2 init.)34.365.855.954.1
SeqQ-Former (BLIP-2 init.)36.268.551.155.4
+ +Table 4. Comparison of different visual projectors on NExTQA [30]. The linear projector is initialized with pre-trained weights in LLaVa [14]. BERT init. and BLIP-2 init. indicate that the visual projector is initialized with weights from BERT [5] and BLIP-2 [12]. SeqQ-Former is the proposed sequential visual projector. See §4.3 for more details. + +![](images/18e3fbd0e90c2c5ebce735d5ba36164ab1d089c5857de038d52372f99d548245.jpg) +Figure 5. Comparison of attention weights for varying context lengths. Lighter colors represent higher weights. To improve clarity, we have combined visual token weights into the first four tokens. We recommend zooming in for optimal viewing. +Figure 6. The effect of training visual tokens on NExT-QA [30]. Accuracy of all questions and three types of questions, including temporal (Tem.), descriptive (Des.), and causal (Cau.), are presented with different colors. + +referred to as "SeqQ-Former" in the figure, outperforming all other visual projectors in terms of accuracy. + +Number of Training Visual Tokens. We assess the impact of visual tokens on NExT-QA as illustrated in Figure 6. In this study, a frame is converted into 32 visual tokens using Q-Former. Frames were sampled at various timestamps throughout the training process. The overall accuracy improves as more visual tokens are incorporated into the language model. However, for temporal queries, precision declined with an increased number of visual tokens. The accuracy of temporal reasoning questions is notably lower than that of other question types. Temporal reasoning presents more difficulties, and the language model may not excel at the temporal modeling of visual tokens. In terms of descriptive question types, the accuracy steadily increases as the model gains more visual information. + +Quantitative Results. We visualize the generated responses of Video-ChatGPT [18] and VISTA-LLAMA for different videos in Figure 7. The frames sampled at different times + +![](images/64cbfd04962d41864a9b94203e39d864c7b418d2f4495674c82cc46836b8a295.jpg) +Figure 7. Visualization results on different video questions. The questions and annotated answers are located on the left side. The generated text from Video-ChatGPT [31] and our model is presented in the green and orange boxes, respectively. See §4.3 for more details. + +tamps is presented, with questions listed below the images. In the first video, the mobile phone is turned, sequentially exposing each side over time. The question asks for the side of the mobile phone at the end of the video. Video-ChatGPT gives the wrong response, as the camera is never present at the bottom left side in the video. The hallucination occurs in Video-ChatGPT, whereas our model predicts the answer correctly. The responses in three cases demonstrate that Video-ChatGPT provides unrelated answers that do not correspond to the video content. In the second case, the person in pink is always standing while the man is sitting on the ground, but Video-ChatGPT incorrectly states that the person in pink is also sitting on the ground. In the third video, the man is consistently walking up the mountain and never walking down. However, Video-ChatGPT falsely outputs that the man continues walking down the mountain. In contrast, our model provides the correct and reliable responses. This demonstrates that our model significantly reduces hallucination in video understanding and delivers more accurate responses. + +# 5. Conclusion + +In this work, we present VISTA-LLAMA to improve the video understanding in large language model. A new vision-aware attention is introduced to maintain same relative distance between all visual tokens and language tokens. In addition, we propose a sequential visual projector to map video into the language space to enable temporal modeling. Experiments on several video question answering task demonstrate that our designs significant improves the current SoTA. Current work is built on the pre-trained image-text architecture, the power of our designs could be enlarged when applied in pre-training stage. Furthermore, the performance of our sequential encoder on longer videos can also be assessed. We would consider evaluating the model in more tasks and applied the attention mechanism to more vision language models. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554, U2336212), and China Postdoctoral Science Foundation (524000-X92302). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 23716-23736, 2022. 1, 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2 +[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Annual Meeting of the Association for Computational Linguistics, 2011. 5 +[4] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2, 3 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7 +[6] Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495, 2023. 1 +[7] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 5 +[8] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30: 681–694, 2020. 1 +[9] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yezhou Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14773-14783, 2022. 2 +[10] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953–14962, 2023. 2 +[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 5 +[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 2, 4, 5, 6, 7 +[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. + +Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 5, 6 +[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. ArXiv, abs/2304.08485, 2023. 1, 2, 3, 5, 6, 7 +[15] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zeroshot video grounding with pseudo query lookup and verification. IEEE Transactions on Image Processing, 33:1643-1654, 2024. 2 +[16] Fan Ma, Linchao Zhu, and Yi Yang. Weakly supervised moment localization with decoupled consistent concept prediction. International Journal of Computer Vision, 130(5): 1244-1258, 2022. 3 +[17] Fan Ma, Xiaojie Jin, Heng Wang, Jingjia Huang, Linchao Zhu, Jiashi Feng, and Yi Yang. Temporal perceiving video-language pre-training. 2023. 2 +[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 2, 5, 6, 7 +[19] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023.2 +[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5 +[21] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2 +[22] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 5, 6 +[23] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3 +[24] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3, 5 +[25] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640, 2015. 3 +[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 1, 2 +[27] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, Felix Hill, and Zacharias Janssen. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 2 + +[28] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Luccioni, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2 +[29] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 2 +[30] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9772–9781, 2021. 2, 3, 5, 6, 7 +[31] Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. ArXiv, abs/2304.01196, 2023. 8 +[32] D. Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. Proceedings of the 25th ACM international conference on Multimedia, 2017. 2, 3, 5, 6 +[33] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5288-5296, 2016. 3, 5 +[34] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1, 5, 6 +[35] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models, 2024. 2 +[36] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 2 +[37] Zhou Yu, D. Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 5, 6 +[38] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414, 2022. 1 +[39] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6 +[40] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1, 5, 6 +[41] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. + +Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models. ArXiv, abs/2205.01068, 2022. 2 +[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 2 \ No newline at end of file diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/images.zip b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97fd2a27a76f02fa52d330bbe88882169100a2eb --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b007a0375936498e4593e86262832b76f75efed7095bc49fa974fd4adfabb73c +size 655880 diff --git a/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/layout.json b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..582472a86fe57b361d9dad2f09c837fa6e40c188 --- /dev/null +++ b/2024/VISTA-LLAMA_ Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens/layout.json @@ -0,0 +1,6939 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 54, + 102, + 541, + 138 + ], + "type": "text", + "content": "VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": "Fan Ma" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Xiaojie Jin" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Heng Wang" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Yuchen Xian" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Jiashi Feng" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "text", + "content": ", Yi Yang" + }, + { + "bbox": [ + 110, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 175, + 394, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 175, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 201, + 175, + 394, + 190 + ], + "type": "text", + "content": "1Zhejiang University 2ByteDance Inc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 254, + 290, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 290, + 613 + ], + "type": "text", + "content": "Recent advances in large video-language models have displayed promising outcomes in video comprehension. Current approaches straightforwardly convert video into language tokens and employ large language models for multi-modal tasks. However, this method often leads to the generation of irrelevant content, commonly known as \"hallucination\", as the length of the text increases and the impact of the video diminishes. To address this problem, we propose VISTA-LLAMA, a novel framework that maintains the consistent distance between all visual tokens and any language tokens, irrespective of the generated text length. VISTA-LLAMA omits relative position encoding when determining attention weights between visual and text tokens, retaining the position encoding for text and text tokens. This amplifies the effect of visual tokens on text generation, especially when the relative distance is longer between visual and text tokens. The proposed attention mechanism significantly reduces the chance of producing irrelevant text related to the video content. Furthermore, we present a sequential visual projector that projects the current video frame into tokens of language space with the assistance of the previous frame. This approach not only captures the temporal relationship within the video, but also allows less visual tokens to encompass the entire video. Our approach significantly outperforms various previous methods (e.g., Video-ChatGPT, MovieChat) on four challenging open-ended video question answering benchmarks. We reach an accuracy of 60.7 on the zero-shot NExT-QA and 60.5 on the zero-shot MSRVTT-QA, setting a new state-of-the-art performance. This project is available at https://jinxxian.github.io/Vista-LLaMA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 657, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 288, + 694 + ], + "type": "text", + "content": "The surge in multi-modal vision-and-language models [1, 12, 14], capable of comprehending both visual (e.g., image/video) and language data, can be attributed to the recent" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 217, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 308, + 217, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 403 + ], + "type": "image", + "image_path": "6d453bf5d37aa3c2ebd46200975efba1f530d25463bb66fd20963bba9ffca429.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 410, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 498 + ], + "type": "text", + "content": "Figure 1. Video language processing with LLaMA [26] and our VISTA-LLAMA. The vanilla LLaMA treats visual tokens (☑) the same as other language tokens (☑), weakening the impact for tokens in long distance. Our model retains the same mechanism for language tokens and strengthens the impact of the visual tokens. The intensity of the impact of each token is conveyed through the depth of the line color " + }, + { + "bbox": [ + 304, + 410, + 547, + 498 + ], + "type": "inline_equation", + "content": "(\\rightarrow)" + }, + { + "bbox": [ + 304, + 410, + 547, + 498 + ], + "type": "text", + "content": ". Our model provides the accurate response for the presented scenario." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 593 + ], + "type": "text", + "content": "achieves of large language models (LLMs) such as GPT [8], GLM [38], and LLaMA [40]. Video-language models (video-LMs) pose greater challenges in scaling due to increased computational and annotation costs compared to image-language models (image-LMs). Recent research has therefore focused on effectively training video-LMs by utilizing pre-trained image-LMs [22, 40]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Video-LMs benefit from this warm-start approach to enhance visual representation learning by projecting video frames into language space and treating videos as several prompt language tokens [18, 34]. However, this diminishes the visual impact of videos on text generation and lacks explicit temporal modeling in videos. The generated text is often not related to the video content, as depicted in Figure 1, a phenomenon known as hallucination in language processing [6]. The distance between the generated text token and the visual tokens in large language models may" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13151" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "be a contributing factor, especially when the visual tokens are distant from the generated text. Additionally, handling longer videos poses a challenge because of the constraints on context length in large language models, where numerous visual tokens take up a substantial portion of the context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 337 + ], + "type": "text", + "content": "We present VISTA-LLAMA, an innovative video language framework to tackle above issues. The primary concept is to preserve equal distance between all visual tokens and any language tokens, while also retaining the relative distance between any two language tokens, as depicted in Figure 1. The rotary position embedding is only applied to language tokens to capture relative distance when calculating similarity between language tokens. When computing similarity between visual and text tokens, the rotary position embedding is removed to reduce the impact of relative distance. With the equal distance to visual tokens (EDVT) attention, the impact of visual cues on text production is amplified without compromising the text production capability. The experiments also show that our design produces more precise text description for input videos and the phenomenon of visual hallucinations occurring in language models has been greatly reduced." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 468 + ], + "type": "text", + "content": "To further improve the temporal modeling of video, we introduce a sequential visual projector. Instead of mapping each frame of the video independently into fixed-length visual tokens, which ignores the temporal relationship between frames, we generate visual tokens for each frame using the previous projected visual tokens. It not only incorporates the temporal relationship between frames into the language model to improve video comprehension without any extra parameters, but also enables the language model to encode longer videos with fewer visual tokens by sampling fewer projected frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 470, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 589 + ], + "type": "text", + "content": "We demonstrate the effectiveness of VISTA-LLAMA on four challenging video question answering (QA) benchmarks, where our method outperforms several previous works, and achieves the state-of-the-art in zero-shot NExTQA [30], and MSRVTT [32]. We also show that our attention design and the temporal modeling mechanism improves capacity of large language model on video-question answering by a large margin. Comprehensive experiments are conducted to demonstrate the effectiveness of our designs. We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 594, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 594, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 628 + ], + "type": "text", + "content": "- We introduce a novel video-language model, dubbed as VISTA-LLAMA, to enhance the video understanding and facilitate temporal modeling within the language model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 630, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 286, + 676 + ], + "type": "text", + "content": "- A novel multi-modal attention is proposed to enable reliable video text generation by maintaining equal distance to visual tokens. Additionally, temporal modeling is facilitated by employing a sequential visual projector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 678, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 286, + 712 + ], + "type": "text", + "content": "- Our method exhibits superior empirical performance, establishing new benchmarks for zero-shot open-ended video question answering tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 544, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 544, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 544, + 120 + ], + "type": "text", + "content": "- A detailed analysis further explains the design choices inherent in our proposed framework, contributing significantly to a better comprehension of multi-modal dynamics in large language models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 152, + 442, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 442, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 442, + 165 + ], + "type": "text", + "content": "2.1. Large Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 171, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 373 + ], + "type": "text", + "content": "Large language models (LLMs) [2, 19, 27, 28, 42] have demonstrated exceptional proficiency in understanding language and reasoning, resulting in the generation of high-quality natural language text in diverse domains. LLMs have already initiated a technological revolution and have found extensive application in various domains. Additionally, a series of open source large models, including LLaMA [26], and OPT [41], have significantly contributed to technological advancements and made remarkable contributions to the NLP community. Leveraging the foundation established by these impressive LLMs, researchers have further expanded their capabilities and developed exceptional models for diverse NLP tasks (e.g. Vicuna [4]). Our work is also built upon these remarkable LLMs, and we equip the language models with a novel attention mechanism and a sequential visual projector that enhances their abilities to comprehend visual content in videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 384, + 504, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 504, + 396 + ], + "type": "text", + "content": "2.2. Multi-modal Large Language Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 403, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 713 + ], + "type": "text", + "content": "Researchers have been diligently exploring the application of LLMs for processing multi-modal problems [9, 13, 15, 17]. The existing approaches can be classified into two primary categories. In the first category, LLMs are treated as controllers while the multi-modal models as tools [10]. When presented with the specific task, the LLMs first interact with user instructions and decide which tools to employ. Subsequently, it generates comprehensive responses by amalgamating the outcomes derived from these readily available multi-modal models. These approaches, such as Visual ChatGPT [29], HuggingGPT [21], and DoraemonGPT [35], have shown impressive results on various multi-modal tasks without training models. For the second category, large-scale multi-modal models are trained on the multi-modal data. The principle behind this line of researches is to align other modal pre-trained models with textual LLMs. Flamingo [1] incorporates a perceiver resampler and a gated cross-attention layer to connect a frozen image encoder with the LLM. BLIP2 [12] introduces a Q-Former to map each image into fixed-length tokens in the language embedding space via the learned queries. LLaVA [14], mPLUG-owl [36], and MiniGPT4 [42] develop the image-LLMs utilizing image-instruction training pairs. Video Chat [13] extends image encoders to enable large models to comprehend visual content in videos. Video-ChatGPT [18] is trained on video instructional data to give appropriate answers for multi-modal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13152" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 107 + ], + "type": "text", + "content": "inputs. In this work, we improve the understanding of videos in a video-language model by incorporating temporal modeling and a novel multi-modal attention mechanism." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 195, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 195, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 195, + 129 + ], + "type": "text", + "content": "2.3. Video Question Answering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 134, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 289, + 326 + ], + "type": "text", + "content": "Video Question Answering (VideoQA) is a task that involves answering language questions that are based on a given video [32]. It requires the ability to understand and reason across different semantic levels, which in turn demands a capacity for multi-modal understanding. Previous VideoQA benchmarks primarily concentrated on short videos, asking questions according to the specified visual facts (e.g., location and objects) [16, 33]. Recently, several new benchmarks have been proposed to focus on resolving temporal and causal questions in longer video clips [25]. NExT-QA [30] is an example of such a benchmark, aiming to uncover the causalities or intentions of particular events, and infer subsequent actions within the video. In this work, we investigate the video-language model on zero-shot open-ended VideoQA where no training question-answer pairs are provided in the training stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 336, + 102, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 102, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 102, + 348 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 356, + 114, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 114, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 114, + 367 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 434 + ], + "type": "text", + "content": "VISTA-LLAMA comprises three fundamental components: a visual encoder, a visual projector, and a pre-trained large language model. Figure 2 shows an overview of VISTA-LLAMA. The components' design and implementation details are provided below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 436, + 287, + 650 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 436, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 287, + 483 + ], + "type": "text", + "content": "- Visual Encoder. We utilize pre-trained EVA-CLIP-g [24] and ViT-L [20] as visual encoder. The last layer of ViT encoder is removed because it specializes in aggregating the features of the first token for contrastive learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 484, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 543 + ], + "type": "text", + "content": "- Visual Projector. The visual projector maps the output of the visual encoder into tokens that occupy the same space as the text features from word embedding. Various visual projectors are considered in Sec. 3.3. The visual projector is to be trained in our work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 544, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 650 + ], + "type": "text", + "content": "- Pre-trained Large Language Model. In this study, we utilize LLaVa [14], which is fine-tuned on Vicuna-7B [4] using instructional image-text pairs, for processing videos and texts. The causal mask is employed in all attention processes, incorporating the attention interplay between visual and text tokens. The pre-trained language model is frozen in the present study, and a new attention mechanism is developed to maintain a consistent distance to visual tokens for all textual tokens." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "In our framework, the video is first encoded into visual tokens using the frozen visual encoder and the trainable visual projector. These visual tokens are then combined with text tokens, which are projected with word embedding from the prompt and question sentences. The combined visual" + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 306, + 102, + 539, + 286 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 95 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of EDVT attention in a PyTorch-like style." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 102, + 539, + 286 + ], + "lines": [ + { + "bbox": [ + 306, + 102, + 539, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 539, + 286 + ], + "type": "text", + "content": "# x: hidden input in each attention layer\n# v_mask: indcite which input are from visual tokens\ndef edvtattention_layer(x, v_mask):\n # query, answer, value projection\n q, k, v = qkv_proj(x)\n # apply RoPE for query and key inputs\n r_q, r_k = rope(q, k)\n # attention weights without RoPE\n attention = bmm(q.T, k)\n # attention weights with RoPE\n rattention = bmm(r_q, r_k)\n # Merge attention weights based on visual token\n attention = v_mask * attention + (1 - v_mask) *\n rattention\n attention = Softmax(attention, dim=-1)\n # Update representation based on attention weights\n v = bmm(attention, v)\n out = linear_proj(v)\n return out" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "type": "text", + "content": "and text tokens are then input into the language model to generate answers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "text", + "content": "3.2. Equal Distance to Visual Tokens" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": "Preliminary. For the concatenate visual-text input, three linear projection layers are applied in every attention layer to produce the query " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " query in " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": ". The conventional attention mechanism updates the input by initially determining the similarity between the query and key, then applying the attention weights to the value state. Specifically, for the " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " position, the update procedure can be articulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 498, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 498, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 349, + 498, + 545, + 521 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} \\operatorname {s i m} \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {1}", + "image_path": "3c3a9303d998e9486336ac992cbe85e4da28fd34cea8c3380aa9e90e2fa10c11.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\text{sim}(\\mathbf{q}_j, \\mathbf{k}_i) = \\exp(\\mathbf{q}_j^T \\mathbf{k}_i / \\sqrt{d})" + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "text", + "content": ". Here, a causal mask is utilized so that the " + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "text", + "content": " query can only attend to the key positioned less than " + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 530, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "EDVT-Attention. The vanilla attention model lacks positional awareness, with no encoded relative distance for natural language processing. In contrast, Rotary Positional Embeddings (RoPE) [23] encodes the position data of tokens using a rotation matrix, which inherently includes an explicit relative position dependency. Within each attention layer, RoPE is implemented across all projected query and key inputs in order to compute the attention weights via leveraging relative distance between tokens. The query situated at the " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " position incorporates rotary position embedding through " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}_j = \\mathbf{R}_j\\mathbf{q}_j" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ", wherein " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_j \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " represents the rotary matrix for the " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " position. Consequently, the attention involving" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13153" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 293 + ], + "type": "image", + "image_path": "d09f2ea609b6164d00534d7ab7d44470a68833a75da8785b60ea4b2d101b7ccc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "type": "text", + "content": "Figure 2. The framework of VISTA-LLAMA. The visual encoder and large language model are both frozen ( ) during training, while the projector is trainable ( ) to map video into the language's space. The attention operation in each layer is present on the right part. Only the text tokens are applied with rotary position embedding to include relative distance information. The attention weights between visual and language tokens are calculated without the rotary position embedding. The casual mask is applied to the bottom-right attention weights." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 365, + 258, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 258, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 258, + 379 + ], + "type": "text", + "content": "relative position embedding is expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 385, + 288, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 385, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 73, + 385, + 288, + 409 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} _ {r o p e} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i}) \\mathbf {v} _ {i}}{\\sum_ {i = 1} ^ {j} s i m (\\mathbf {R} _ {j} \\mathbf {q} _ {j} , \\mathbf {R} _ {i} \\mathbf {k} _ {i})}. \\tag {2}", + "image_path": "1db0950f136652adf2f3a054285122503737cd6733916c0eeebc76f9ab0dcd17.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 414, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 288, + 544 + ], + "type": "text", + "content": "RoPE inherently integrates relative position data via the multiplication of rotation matrices rather than appending it to the input as a positional embedding. In natural language processing, this relative proximity between two words is vital, given that remote words should have less influence than adjacent words when generating the current word. However, using the same attention mechanism for visual and text tokens may result in unintentional text generation, a phenomenon often referred to as hallucination in LLMs. With multi-modal input, the generated text should depend on the visual content, disregarding the influence of relative distance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "content": "To alleviate this issue and enhance the video understanding with LLMs, we introduce the EDVT-Attention where the equal distance to visual tokens is maintained while the relative distance between text tokens is attained. As shown in Figure 2, the rotary position embedding is only applied on text tokens. Let " + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "content": " be the set of visual and text tokens. The EDVT-Attention is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 287, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 287, + 676 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} _ {e d v t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {T}} s i m \\left(\\tilde {\\mathbf {q}} _ {j} , \\tilde {\\mathbf {k}} _ {i}\\right) + \\sum_ {k _ {i} \\in \\mathcal {V}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right)}, \\tag {3}", + "image_path": "214cd26a4e2caedbafb2565ddc2caac9142c0d6059f8e561b29d96e528647d9e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}_j" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{k}}_j" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " denote the query and key applied with rotary embedding, separately. The distance between language tokens is determined using the rotary matrix, while" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 425 + ], + "type": "text", + "content": "the correlation between visual and textual inputs remains unaffected by the rotary embedding. This improves the influence of visual information on long-term text generation and decreases the incidence of hallucinations wherein the fabricated content is absent in the videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 485 + ], + "type": "text", + "content": "For ease comprehension of how EDVT-Attention works, Algorithm 1 exhibits the pseudo-code in the decoder layer of LLMs. It involves the combinations of attention weights prior to and subsequent to the implementation of rotary embedding with the visual mask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 491, + 457, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 457, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 457, + 504 + ], + "type": "text", + "content": "3.3. Sequential Visual Projector" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "content": "The visual projector aims to map video features into the language embedding space, allowing for the fusion and processing of visual and textual inputs by the substantial language model. As shown in Figure 3, earlier visual projectors either employ the linear layer or the query transformer (Q-Former) [12] to directly convert frame features into language tokens. However, the lack of temporal relationship in these methods impedes thorough video understanding in LLMs. We introduce the sequential visual projector in Figure 3 to encode the temporal context into the visual tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^{t \\times l \\times d}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " be the extracted video feature of length " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " denotes the length of visual tokens. We use a Q-Former " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{Q}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " to map each frame into the fixed length of " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " representations " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f_{Q}(\\mathbf{o}_i, \\mathbf{p}) \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " is the learnable query embedding. In the prior approach, all projected frame features are merely combined with word tokens to serve as mixed input for LLMs. For encoding temporal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 70, + 255, + 171 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 255, + 171 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 255, + 171 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 255, + 171 + ], + "type": "image", + "image_path": "60b6b2d49ed7316370729aa4b1dd9dbf0caa6f7d7598e83eff2e24327aab1ee9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 176, + 227, + 281 + ], + "blocks": [ + { + "bbox": [ + 78, + 176, + 227, + 281 + ], + "lines": [ + { + "bbox": [ + 78, + 176, + 227, + 281 + ], + "spans": [ + { + "bbox": [ + 78, + 176, + 227, + 281 + ], + "type": "image", + "image_path": "0f33c44ada250c220cacf8291c1044bd26943cc5fa96d9e6c765ee4d3eb91d45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 357 + ], + "type": "text", + "content": "Figure 3. Comparison of three visual projectors. (a): Each frame feature is projected into the visual tokens independently with the linear projection. (b): Q-Former uses shared learnable query tokens to separately map each frame into fixed-length tokens. (c): The sequential Q-Former with linear projection layer to enable temporal modeling." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 379, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 415 + ], + "type": "text", + "content": "context, we utilize the previously projected frame feature as the query to attend to the current frame feature. The current projected frame feature is then updated accordingly:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 426, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 426, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 126, + 426, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = f _ {Q} \\left(\\mathbf {o} _ {i}, \\mathbf {x} _ {t - 1}\\right), \\tag {4}", + "image_path": "21a01896ed6a1a3fcb0e7b5e77d514d39bde4811b9a1025222a34ea373c63be6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}\\in \\mathbb{R}^{k\\times d}" + }, + { + "bbox": [ + 46, + 447, + 287, + 495 + ], + "type": "text", + "content": " represents the previous projected frame feature. This allows the visual tokens to encode the temporal relationship, as the current frame's visual token is generated using the previous feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": "Moreover, we can represent the entire video with fewer tokens by merely sampling a small number of visual tokens. This method tackles the challenge of encoding lengthy videos by employing sequential encoding. By integrating prior visual context into subsequent visual tokens, we can accomplish an adequate representation of the entire video through sparse sampling of projected frame features. Besides, this technique permits larger language models to manage much longer videos without length limitation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 610, + 179, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 179, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 179, + 623 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "We fine-tune the model based on LLaVA [14] and use " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " video instruction pairs provided by [18]. We add Q-Former initialized from BLIP2 [12] to project frame features into fixed-length tokens. We test our model with both ViT-L-14 [20] and EVA-CLIP-g [24] visual encoders in experiments. We only update the visual projector, which contains the Q-Former and linear projection layer to project the video" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "features to the LLMs' input space. The visual backbone and the language model are frozen during training. The VISTA-LLAMA is fine-tuned for 3 epochs on 8 A100 80GB GPUs with a learning rate of " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "2e^{-5}" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " and an overall batch size of 32. We run all the inference experiments with FP16 to save memory and faster testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 154, + 387, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 387, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 387, + 167 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 173, + 422, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 173, + 422, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 173, + 422, + 186 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 192, + 515, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 515, + 203 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 515, + 203 + ], + "type": "text", + "content": "Datasets. Our method is evaluated on four datasets:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 205, + 547, + 396 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 205, + 547, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 547, + 265 + ], + "type": "text", + "content": "- NExT-QA [30] is designed to advance video understanding from descrii ng to explaining the temporal actions. It comprises 5,440 videos and approximately " + }, + { + "bbox": [ + 306, + 205, + 547, + 265 + ], + "type": "inline_equation", + "content": "52K" + }, + { + "bbox": [ + 306, + 205, + 547, + 265 + ], + "type": "text", + "content": " manually annotated QA pairs, which are categorized into temporal (Tem.), causal (Cau.), and descriptive (Des.) questions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 266, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 546, + 312 + ], + "type": "text", + "content": "- MSVD-QA [32] is a dataset built upon Microsoft Research Video Description Corpus [3], commonly used in video caption tasks. The MSVD-QA dataset comprises a total of 1,970 video clips with 50,505 QA pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "text", + "content": "- MSRVTT-QA [32] is based on MSR-VTT dataset [33], which includes " + }, + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "text", + "content": " videos and " + }, + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "inline_equation", + "content": "243K" + }, + { + "bbox": [ + 306, + 313, + 546, + 348 + ], + "type": "text", + "content": " QA pairs with larger and has more complex scenes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 349, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 547, + 396 + ], + "type": "text", + "content": "- ActiviytNet-QA [37] is a fully annotated and large-scale videoQA dataset. It contains 58K QA pairs derived from 5,800 complex web videos derived from the popular ActivityNet dataset [11]." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 398, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 547, + 422 + ], + "type": "text", + "content": "Evaluation. We adopt two metrics to evaluate the performance of video-language models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 424, + 547, + 639 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 306, + 424, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 547, + 518 + ], + "type": "text", + "content": "- Open-Ended Zero-Shot Question-Answer Evaluation. As video-language models generate responses of varying lengths to open-ended questions, it is challenging to evaluate models with traditional word matching strategy. We employ LLM-Assisted evaluation, in line with [18], for fair comparison. Given the question, correct answer, and predicted answer, GPT3.5-turbo-0613 is used to return True or False judgement and relative score (0 to 5)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 520, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 547, + 639 + ], + "type": "text", + "content": "- Video-Based Text Generation Benchmarking. We evaluate the text generation performance following [18] from five aspects: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding. The test set for this evaluation is based on the ActivityNet-200 [7], featuring videos with rich, dense descriptive captions and associated question-answer pairs from human annotations. The evaluation pipeline is also built with the GPT-3.5 model and relative score (0 to 5) is generated." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 647, + 480, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 480, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 480, + 659 + ], + "type": "text", + "content": "4.2. Comparison to State-of-the-Arts" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "For the zero-shot open-ended video question answering tasks, we compare our model with FrozenBiLM [34], VideoChat [13], LLaMA Adapter [40], VideoLLaMA [39], VideoChatGPT [18], and MovieChat [22] in Tab. 1. FrozenBiLM" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13155" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 535, + 182 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 535, + 182 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 535, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 535, + 182 + ], + "type": "table", + "html": "
MethodNExT-QA [30]MSVD-QA [32]MSRVTT-QA [32]ActivityNet-QA [37]
AccuracyScoreAccuracyScoreAccuracyScoreAccuracyScore
FrozenBiLM [34]--32.2-16.8-24.7-
Video Chat [13]56.23.256.32.845.02.526.52.2
LLaMA Adapter [40]--54.93.143.82.734.22.7
Video LLaMA [39]--51.62.529.61.812.41.1
MovieChat [22]49.92.761.02.949.72.851.53.1
Video-ChatGPT [18]54.63.264.93.349.32.835.22.7
VISTA-LLAMA (Ours)60.73.465.33.660.53.348.33.3
", + "image_path": "53aa6ee4c8c512dde530fa115d5254afbc625ccffbdb947b8e8b039bd435b149.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 221, + 288, + 297 + ], + "blocks": [ + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "lines": [ + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "spans": [ + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "type": "text", + "content": "Table 1. Comparison with SoTA methods on zero-shot VideoQA. See §4.2 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 221, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 221, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 221, + 288, + 297 + ], + "type": "table", + "html": "
MethodCr.Cs.De.Ct.Te.
Video Chat [13]2.232.242.502.531.94
LLaMA Adapter [40]2.032.152.322.301.98
Video LLaMA [39]1.961.792.182.161.82
Video-ChatGPT [18]2.402.372.522.621.98
VISTA-LLAMA (Ours)2.442.312.643.182.26
", + "image_path": "42a793f29b9b338f826492d195ee87d20164f18bc926d3148ae8ba697aa891ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 288, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 361 + ], + "type": "text", + "content": "Table 2. Quantitative results on video-based text generation with different video-language methods (§4.2). For clarity, five scores are reported (\"Cr\": Correctness of Information, \"Cs\": Consistency, \"De\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te\": Temporal Understanding)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "text", + "content": "adapts frozen the bidirectional language model, showing promising results in zero-shot VideoQA settings. Other compared models are built on recent large auto-regressive language models. Despite pre-existing models have produced substantial results, VISTA-LLAMA consistently outperforms them, achieving state-of-the-art (SoTA) performance across three datasets: NExT-QA [30], MSVD-QA [32], and MSVTT-QA [32]. Our method obtains the highest results, with " + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "inline_equation", + "content": "60.5\\%" + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "text", + "content": " accuracy on MSRvTT, elevating the performance of the second-best model by nearly " + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "text", + "content": ". Additionally, our method attains " + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "inline_equation", + "content": "60.7\\%" + }, + { + "bbox": [ + 46, + 387, + 288, + 555 + ], + "type": "text", + "content": " accuracy on NExT-QA, markedly superior to Video-ChatGPT [18]. These results demonstrate VISTA-LLAMA's capability to comprehend video content and produce precise answers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 290, + 715 + ], + "type": "text", + "content": "We present the results of the evaluation of video-based text generation in Tab. 2. The results reveal its competent performance across all aspects when compared with the recent video-language models, Video Chat[13], VideoLLaMA [39], and Video-ChatGPT [18]. Despite being trained with identical datasets, our model outperforms Video-ChatGPT in four aspects. Our method offers a more comprehensive interpretation, and its responses are more in tune with the overarching context of the video content than comparable approaches. By utilizing the EDVT-Attention and sequential temporal modeling techniques, our model demonstrates a strong ability to generate text that is contextually appropriate, detailed, and includes precise timing for video inputs." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 335, + 225, + 519, + 353 + ], + "blocks": [ + { + "bbox": [ + 335, + 225, + 519, + 353 + ], + "lines": [ + { + "bbox": [ + 335, + 225, + 519, + 353 + ], + "spans": [ + { + "bbox": [ + 335, + 225, + 519, + 353 + ], + "type": "image", + "image_path": "2ca9f919bd9e213840828c73cd7203dfe8f87ae302086bf8a131692ec539df82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 365, + 548, + 431 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 548, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 548, + 431 + ], + "type": "text", + "content": "Figure 4. Comparison of different design choices on NExTQA [30](" + }, + { + "bbox": [ + 304, + 365, + 548, + 431 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 304, + 365, + 548, + 431 + ], + "type": "text", + "content": "4.3). For clarity, accuracy of base model and two variants are given (\"Baseline\": the frozen LLaVA [14] with trainable Q-Former [12], \"w Seq\": base model with sequential visual projector, \"w Seq&EDVT\": base model with both sequential visual projector and EDVT-Attention)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 454, + 405, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 405, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 405, + 467 + ], + "type": "text", + "content": "4.3. Abalation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "type": "text", + "content": "Effect of Design Choices. We investigate the effect of our two designs, including the equal-distance to visual-tokens (EDVT) attention and the sequential visual projector. As shown in Figure 4, our model is tested on NExT-QA [30], which comprises three different types of questions. The baseline model employs LLaVA [14] as the language model, and Q-Former in BLIP-2 [12] as the visual projector. We first incorporate temporal modeling into the baseline, utilizing the projected tokens from the previous frame as the query tokens to generate visual tokens for the current frame, a setup denoted as \"w Seq\" in Figure 4. Further enhancing this version with EDVT-Attention, we introduce \"w Seq&EDVT\", ultimately forming our final VISTA-LLAMA. The results affirm the efficacy of our design across all question types." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "EDVT-Attention. We here validate the effectiveness of our core EDVT-Attention. Table 3 reports the comparison results of different models in combination with our EDVT-Attention on NExT-QA [30]. Initially, we integrate the EDVT-Attention into Video-ChatGPT [18], where only the attention component is substituted with our EDVT-Attention" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13156" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 288, + 159 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 288, + 159 + ], + "type": "table", + "html": "
MethodNExT-QA [30]
Tem.Cau.Des.Avg.
Video-ChatGPT [18]37.665.154.954.6
+ EDVT-Attention39.5 (+1.9)72.8 (+7.7)54.859.3 (+4.7)
VISTA-LLAMA (ViT-L-14)34.069.142.253.6
+ EDVT-Attention36.8 (+2.8)72.2 (+3.1)47.7 (+5.5)56.5 (+2.9)
VISTA-LLAMA (EVA-CLIP-g)34.365.855.954.1
+ EDVT-Attention40.7 (+6.4)72.3 (+6.5)57.0 (+1.1)59.7 (+5.6)
", + "image_path": "06333a2bf1649cfdec5c68bb4e2f7398d8b3240179b5addc1fe34a84a1d4404e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 225, + 166, + 323 + ], + "blocks": [ + { + "bbox": [ + 52, + 225, + 166, + 323 + ], + "lines": [ + { + "bbox": [ + 52, + 225, + 166, + 323 + ], + "spans": [ + { + "bbox": [ + 52, + 225, + 166, + 323 + ], + "type": "image", + "image_path": "7e85abb7685829c2f4ccb8fb66e557e745fba1b4dc7145d952e2ef9d586dba89.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 326, + 141, + 334 + ], + "lines": [ + { + "bbox": [ + 78, + 326, + 141, + 334 + ], + "spans": [ + { + "bbox": [ + 78, + 326, + 141, + 334 + ], + "type": "text", + "content": "(a) Vanilla attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 225, + 282, + 323 + ], + "blocks": [ + { + "bbox": [ + 168, + 225, + 282, + 323 + ], + "lines": [ + { + "bbox": [ + 168, + 225, + 282, + 323 + ], + "spans": [ + { + "bbox": [ + 168, + 225, + 282, + 323 + ], + "type": "image", + "image_path": "1934748432e33a3fa4b48770222b00e3b7e8e0ac52d38cc123813ee71096d4d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 326, + 269, + 334 + ], + "lines": [ + { + "bbox": [ + 181, + 326, + 269, + 334 + ], + "spans": [ + { + "bbox": [ + 181, + 326, + 269, + 334 + ], + "type": "text", + "content": "(b) EDVT-Attention (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "content": "during training. Further deployment of the EDVT-Attention into Video-ChatGPT [18] yields significant performance gains (e.g., " + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "inline_equation", + "content": "65.1\\%" + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "content": " escalates to " + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "inline_equation", + "content": "72.8\\%" + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "content": " for causal questions). The accuracy is enlarged across all three setting types, indicating that our EDVT-Attention considerably enhances the multi-modal understanding in large language models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 484, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 604 + ], + "type": "text", + "content": "We further illustrate the attention weights of both EDVT-Attention and conventional attention in Figure 5. For this experiment, we employ 128 visual tokens and integrate the attention weights into the first four tokens to enhance visualization. The attention weights assigned to visual tokens are markedly greater in our EDVT-Attention compared to those in conventional attention. This indicates that the impact of visual tokens on language tokens is considerably more substantial under the proposed EDVT-Attention, shedding light on why the performance notably improves with our design." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Visual Projector. The visual projector is trainable within the model. Therefore, we have examined the impacts of different visual projector designs, as shown in Table 4. We evaluated three visual projector variants. The Q-Former, initialized using BLIP-2 [12], considerably outperforms the Q-Former that was initialized with BERT [5]. This indicates that visual projectors, when pre-trained on image-text pairs, can significantly enhance video comprehension. The model's performance peaks when integrating the sequential design," + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 547, + 140 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": "Table 3. Comparison of EDVT-Attention design with diffent visual encoders on NExT-QA [30] (§4.3). For clarity, accuracy of all questions and three types of questions are reported (\"Tem\": temporal, \"Cau\": causal, \"Des\": descriptive, \"Avg.\": average)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 70, + 547, + 140 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 547, + 140 + ], + "type": "table", + "html": "
Visual projectorNExT-QA [30]
Tem.Cau.Des.Avg.
Linear Projector37.665.154.954.6
Q-Former (BERT init.)35.262.749.251.8
Q-Former (BLIP-2 init.)34.365.855.954.1
SeqQ-Former (BLIP-2 init.)36.268.551.155.4
", + "image_path": "06eedae109a23c2a427e20f7d7e863fe47b2154233dbdd802a9f778f5939074e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 148, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 547, + 214 + ], + "type": "text", + "content": "Table 4. Comparison of different visual projectors on NExTQA [30]. The linear projector is initialized with pre-trained weights in LLaVa [14]. BERT init. and BLIP-2 init. indicate that the visual projector is initialized with weights from BERT [5] and BLIP-2 [12]. SeqQ-Former is the proposed sequential visual projector. See §4.3 for more details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 227, + 545, + 410 + ], + "blocks": [ + { + "bbox": [ + 46, + 344, + 288, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 288, + 389 + ], + "type": "text", + "content": "Figure 5. Comparison of attention weights for varying context lengths. Lighter colors represent higher weights. To improve clarity, we have combined visual token weights into the first four tokens. We recommend zooming in for optimal viewing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 227, + 545, + 410 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 410 + ], + "type": "image", + "image_path": "18e3fbd0e90c2c5ebce735d5ba36164ab1d089c5857de038d52372f99d548245.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 416, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 304, + 416, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 547, + 460 + ], + "type": "text", + "content": "Figure 6. The effect of training visual tokens on NExT-QA [30]. Accuracy of all questions and three types of questions, including temporal (Tem.), descriptive (Des.), and causal (Cau.), are presented with different colors." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 484, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 508 + ], + "type": "text", + "content": "referred to as \"SeqQ-Former\" in the figure, outperforming all other visual projectors in terms of accuracy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "type": "text", + "content": "Number of Training Visual Tokens. We assess the impact of visual tokens on NExT-QA as illustrated in Figure 6. In this study, a frame is converted into 32 visual tokens using Q-Former. Frames were sampled at various timestamps throughout the training process. The overall accuracy improves as more visual tokens are incorporated into the language model. However, for temporal queries, precision declined with an increased number of visual tokens. The accuracy of temporal reasoning questions is notably lower than that of other question types. Temporal reasoning presents more difficulties, and the language model may not excel at the temporal modeling of visual tokens. In terms of descriptive question types, the accuracy steadily increases as the model gains more visual information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Quantitative Results. We visualize the generated responses of Video-ChatGPT [18] and VISTA-LLAMA for different videos in Figure 7. The frames sampled at different times" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13157" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 411 + ], + "type": "image", + "image_path": "64cbfd04962d41864a9b94203e39d864c7b418d2f4495674c82cc46836b8a295.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 419, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 419, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 546, + 444 + ], + "type": "text", + "content": "Figure 7. Visualization results on different video questions. The questions and annotated answers are located on the left side. The generated text from Video-ChatGPT [31] and our model is presented in the green and orange boxes, respectively. See §4.3 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 463, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 289, + 704 + ], + "type": "text", + "content": "tamps is presented, with questions listed below the images. In the first video, the mobile phone is turned, sequentially exposing each side over time. The question asks for the side of the mobile phone at the end of the video. Video-ChatGPT gives the wrong response, as the camera is never present at the bottom left side in the video. The hallucination occurs in Video-ChatGPT, whereas our model predicts the answer correctly. The responses in three cases demonstrate that Video-ChatGPT provides unrelated answers that do not correspond to the video content. In the second case, the person in pink is always standing while the man is sitting on the ground, but Video-ChatGPT incorrectly states that the person in pink is also sitting on the ground. In the third video, the man is consistently walking up the mountain and never walking down. However, Video-ChatGPT falsely outputs that the man continues walking down the mountain. In contrast, our model provides the correct and reliable responses. This demonstrates that our model significantly reduces hallucination in video understanding and delivers more accurate responses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 462, + 380, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 380, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 380, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 484, + 548, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 548, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 548, + 663 + ], + "type": "text", + "content": "In this work, we present VISTA-LLAMA to improve the video understanding in large language model. A new vision-aware attention is introduced to maintain same relative distance between all visual tokens and language tokens. In addition, we propose a sequential visual projector to map video into the language space to enable temporal modeling. Experiments on several video question answering task demonstrate that our designs significant improves the current SoTA. Current work is built on the pre-trained image-text architecture, the power of our designs could be enlarged when applied in pre-training stage. Furthermore, the performance of our sequential encoder on longer videos can also be assessed. We would consider evaluating the model in more tasks and applied the attention mechanism to more vision language models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554, U2336212), and China Postdoctoral Science Foundation (524000-X92302)." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13158" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 23716-23736, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "text", + "content": "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Annual Meeting of the Association for Computational Linguistics, 2011. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "text", + "content": "[4] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "text", + "content": "[6] Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "text", + "content": "[7] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 485 + ], + "type": "text", + "content": "[8] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30: 681–694, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "text", + "content": "[9] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yezhou Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14773-14783, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "text", + "content": "[10] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953–14962, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "text", + "content": "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. ArXiv, abs/2304.08485, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 547, + 171 + ], + "type": "text", + "content": "[15] Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zeroshot video grounding with pseudo query lookup and verification. IEEE Transactions on Image Processing, 33:1643-1654, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 172, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 547, + 216 + ], + "type": "text", + "content": "[16] Fan Ma, Linchao Zhu, and Yi Yang. Weakly supervised moment localization with decoupled consistent concept prediction. International Journal of Computer Vision, 130(5): 1244-1258, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 547, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 250 + ], + "type": "text", + "content": "[17] Fan Ma, Xiaojie Jin, Heng Wang, Jingjia Huang, Linchao Zhu, Jiashi Feng, and Yi Yang. Temporal perceiving video-language pre-training. 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 547, + 294 + ], + "type": "text", + "content": "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 315 + ], + "type": "text", + "content": "[19] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 547, + 381 + ], + "type": "text", + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 547, + 426 + ], + "type": "text", + "content": "[21] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "type": "text", + "content": "[22] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "type": "text", + "content": "[23] Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "type": "text", + "content": "[24] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "text", + "content": "[25] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640, 2015. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 604, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 547, + 669 + ], + "type": "text", + "content": "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[27] Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, Felix Hill, and Zacharias Janssen. Multimodal few-shot learning with frozen language models. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13159" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Luccioni, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[29] Chenfei Wu, Sheng-Kai Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. ArXiv, abs/2303.04671, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "text", + "content": "[30] Junbin Xiao, Xindi Shang, Angela Yao, and Tat seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9772–9781, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "type": "text", + "content": "[31] Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. Baize: An open-source chat model with parameter-efficient tuning on self-chat data. ArXiv, abs/2304.01196, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[32] D. Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. Proceedings of the 25th ACM international conference on Multimedia, 2017. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 365 + ], + "type": "text", + "content": "[33] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5288-5296, 2016. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "type": "text", + "content": "[34] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 443 + ], + "type": "text", + "content": "[35] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "text", + "content": "[36] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "type": "text", + "content": "[37] Zhou Yu, D. Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "type": "text", + "content": "[38] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "type": "text", + "content": "[39] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[40] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[41] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 172 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 72, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 547, + 127 + ], + "type": "text", + "content": "Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models. ArXiv, abs/2205.01068, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv, abs/2304.10592, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13160" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_content_list.json b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1cd06702e3f6912b0b3086eb140214c394a50c2b --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "VLP: Vision Language Planning for Autonomous Driving", + "text_level": 1, + "bbox": [ + 194, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenbin Pan $^{1*}$ Burhaneddin Yaman $^{2\\text{圆}}$ Tommaso Nesti $^{2}$ Abhirup Mallik $^{2}$ Alessandro G Allievi $^{2}$ Senem Velipasalar $^{1}$ Liu Ren $^{2}$ $^{1}$ Syracuse University", + "bbox": [ + 189, + 179, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)", + "bbox": [ + 156, + 234, + 812, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cpan14, svelipas}@syr.edu,", + "bbox": [ + 359, + 253, + 594, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{burhaneddin.yaman,tommaso.nesti,abhirup.mallik,alessandro.allievi,liu. ren}@us.bosch.com", + "bbox": [ + 93, + 271, + 870, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving is a complex and challenging task that aims at safe motion planning through scene understanding and reasoning. While vision-only autonomous driving methods have recently achieved notable performance, through enhanced scene understanding, several key issues, including lack of reasoning, low generalization performance and long-tail scenarios, still need to be addressed. In this paper, we present VLP, a novel Vision-Language-Planning framework that exploits language models to bridge the gap between linguistic understanding and autonomous driving. VLP enhances autonomous driving systems by strengthening both the source memory foundation and the self-driving car's contextual understanding. VLP achieves state-of-the-art end-to-end planning performance on the challenging NuScenes dataset by achieving $35.9\\%$ and $60.5\\%$ reduction in terms of average L2 error and collision rates, respectively, compared to the previous best method. Moreover, VLP shows improved performance in challenging long-tail scenarios and strong generalization capabilities when faced with new urban environments.", + "bbox": [ + 75, + 353, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving is a complex problem requiring scene understanding and reasoning to ensure safe motion planning. This sophisticated challenge can be broadly divided into three main tasks, namely perception, prediction and planning (also known as P3). Conventional methods adopt a modular approach by developing and optimizing each task in a disjoint manner without a holistic view, leading to compounding errors and safety concerns [15, 16, 18]. End-to-end autonomous driving systems (ADS), unifying all P3 tasks, have garnered attention for their potential to enhance", + "bbox": [ + 75, + 712, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6850a411b255ab31d7194d8ce722ca14b816f6fdc78873a73b44366a523b7f43.jpg", + "image_caption": [ + "Figure 1. New-city generalization ability of ADS for planning is evaluated by training on Boston city and testing on Singapore, and vice versa. Our proposed VLP shows strong generalization ability by significantly outperforming state-of-the-art vision-only method, VAD[18], in terms of both L2 error and collision rate." + ], + "image_footnote": [], + "bbox": [ + 506, + 335, + 689, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e27faeb36b6d1b5f939040fb2888fe7125b35b4eda94e08cd1f762375b89245.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 335, + 880, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "safe planning. Existing vision-based ADS typically follow a two-stage process: bird's eye view (BEV) feature extraction and downstream tasks [16, 18]. BEV feature extraction transforms multi-view camera data into a structured top-view representation embedding spatial information around the ego-car. BEV features are further utilized as the information pool by the downstream P3 tasks. While effective, these vision-only methods struggle with out-of-domain generalization, such as maintaining performance in new cities and long-tail scenarios, hindering real-world deployment.", + "bbox": [ + 496, + 566, + 892, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The way humans process and interpret visual driving scenes naturally involve a coherent cognitive framework that maintains a consistent logical flow. This enables humans effortlessly make correct decisions even when faced with previously unseen data or scenarios, eliminating generalization issues in driving tasks. Then, this raises a pertinent question: How can ADS achieve human-like driving to navigate diverse and dynamic road environments? Recent advancements in language models (LMs) have led to unprecedented common-sense ability and generalization performance across unseen data and tasks for the natural language processing domain [5, 8, 25]. The superior common-", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done while interned at Bosch Research North America.", + "bbox": [ + 91, + 875, + 426, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 91, + 887, + 228, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sense capability of LMs have led to the emergence of multimodal LMs for diverse applications ranging from medical imaging to robotics [4, 9, 33]. However, incorporation of the reasoning ability of LMs into real-world autonomous driving tasks, to address generalization and long-tail scenarios, is yet to be fully-explored.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we propose a Vision Language Planning (VLP) framework, which integrates the commonsense capability of LMs into vision-based ADS for safe self-driving. Our VLP consists of two key components: Agent-centric Learning Paradigm (ALP) and Self-driving-car-centric Learning Paradigm (SLP), leveraging LMs to enhance the ADS from reasoning and decision-making aspects, respectively.", + "bbox": [ + 75, + 183, + 467, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The BEV feature map serves as the source memory pool in ADS for downstream decoding tasks. It summarizes and encodes the driving environment surrounding the self-driving car, including vehicles, pedestrians, lanes, and more, into a unified feature map. Hence, capturing comprehensive and necessary details in each local position of BEV is critical for safe and precise self-driving performance. To enhance the local semantic representation and reasoning capabilities of BEV, we introduce an innovative Agent-centric Learning Paradigm (ALP) module. ALP integrates the consistent feature space of a pretrained LM to revamp the agent features on the BEV, actively shaping semantics, and guiding the BEV reasoning process. Leveraging on the common sense and logic flow embedded in the LM, our ALP equips the ADS with robustness and consistent BEV feature space, enhancing its effectiveness in diverse driving scenarios.", + "bbox": [ + 75, + 306, + 467, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In ADS, the planning module aggregates information from the preceding perception and prediction phases to make the final decisions for self-driving. This global perspective culminates in the formation of a planning query, directly influencing the safety and accuracy of the self-driving navigation. Considering the critical role of the planning module within ADS, we also present a novel Self-driving-car-centric Learning Paradigm (SLP) to elevate the decoding and acquiring information ability of the planning query. In the SLP, we align the planning query with intended goals and the ego-vehicle driving status by leveraging the knowledge encoded in the pretrained LM. The language model's comprehension capabilities contribute to more informed decision-making during the planning phase as well as enabling a more robust planning query formation process.", + "bbox": [ + 75, + 549, + 467, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through VLP, we bridge the gap between human-like reasoning and autonomous driving, enhancing the model's contextual awareness and its ability to generalize effectively in complex, ever-changing real-world scenarios, as shown in Fig.1. The main contributions of this work are summarized as follows:", + "bbox": [ + 75, + 792, + 467, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose VLP, a Vision Language Planning model,", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which incorporates reasoning capability of LMs into vision-based autonomous driving systems as an enhancement of motion planning and self-driving safety.", + "bbox": [ + 511, + 90, + 888, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VLP is composed of novel components ALP and SLP, aiming to improve the ADS from self-driving BEV reasoning and self-driving decision-making aspects, respectively.", + "- Through extensive experiments in real-world driving scenarios, we show that VLP significantly and consistently outperforms the state-of-the-art vision-based approaches across a spectrum of driving tasks, including open-loop planning, multi-object tracking, motion forecasting, and more.", + "- We conduct the first new-city generalization study on the nuScenes dataset [6] by training and testing on distinct cities, demonstrating the remarkable zero-shot generalization ability of our VLP approach over vision-only methods.", + "- To the best of our knowledge, this is the first work introducing LM into multiple stages of ADS to address the generalization ability in new cities and long-tail cases." + ], + "bbox": [ + 500, + 136, + 888, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 420, + 638, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 500, + 445, + 790, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early end-to-end approaches have adopted a single neural network to perform motion planning without explicitly designing intermediate tasks, such as prediction [3]. Thus, such approaches not only suffer from the sub-optimal performance but also lack interpretability. Modular end-to-end frameworks introduce intermediate tasks for interpretability and provide improved performance [7, 10, 15, 16, 18, 23, 28]. These interpretable end-to-end frameworks explicitly model perception, prediction and planning (P3) components and train them together with a joint optimization strategy to enhance planning. UniAD [16] and VAD [18] are two prominent state-of-the-art end-to-end frameworks. UniAD leverages rasterized scene representations and explicitly identifies crucial components within the P3 framework. It coordinates all tasks through a unified query design for safe planning. On the other hand, VAD uses vectorized representations for efficient planning and safety improvement. While these approaches generally achieve good performance, their generalization performance have been limited due to lack of common sense and reasoning process.", + "bbox": [ + 496, + 464, + 890, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision-Language Models", + "text_level": 1, + "bbox": [ + 500, + 775, + 725, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large language models (LLMs), which are trained on massive amounts of text data, have shown unprecedented success in language-related tasks [5, 8, 29]. The capability of LLMs for common-sense understanding and versatility in tackling diverse language tasks has driven the exploration of multi-modal LLMs [20, 21, 27]. Vision-language models (VLM) integrate visual and textual information to en", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "able common sense understanding of the content by training on large scale open-world vision and text data [2, 17, 20, 27]. CLIP [27] uses a contrastive learning objective to learn a joint embedding space for vision and text data, and has shown immense zero-shot generalization performance, showcasing the importance of VLMs compared to vision-only approaches.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Autonomous Driving with Language Models", + "text_level": 1, + "bbox": [ + 76, + 207, + 454, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, numerous works have been proposed for extending LMS to the embodied AI domain to improve zero-shot generalization performance [4, 9, 30, 35]. These works incorporate sensory inputs from embodied agents in the form of language with vision and language data for decision making process [4, 9]. While these methods have mainly focused on the robotics domain, few works have been proposed for leveraging embodied LMs for autonomous driving tasks [11, 24, 31, 32, 34]. In particular, DiLu [32] and GPT-Driver [24] propose GPT-based driver agents for closed-loop simulation tasks. In [31], an open loop driving commentator, which combines vision and low-level driving actions with language, is proposed to interpret driving actions and for reasoning. Our work significantly differs from these approaches by leveraging LMs for end-to-end motion planning task for real-world driving applications.", + "bbox": [ + 75, + 228, + 472, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 484, + 212, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a Vision Language Planning (VLP) model, which equips the ADS with the capacity to emulate human common sense and engage in contextual reasoning for safe motion planning. Our proposed VLP model, illustrated in Fig. 2a, comprises two innovative components that leverage LMs in both local and global contexts. The Agent-centric Learning Paradigm (ALP) concentrates on refining local details to enhance source memory reasoning, while the Self-driving-car-centric Learning Paradigm (SLP) focuses on guiding the planning process for the self-driving-car (SDC). Below, we first provide an overview of how general ADS work in Sec.3.1. Subsequently, we describe our proposed ALP and SLP in detail in Sec. 3.2 and Sec. 3.3, respectively.", + "bbox": [ + 75, + 507, + 472, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 729, + 207, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In vision-based ADS, a sequence of multi-view camera images serves as the source input, offering abundant visual data for downstream tasks. Initially, an image backbone extracts fundamental features for perception. These features are then processed by a BEV encoder to transform multi-view features into a unified 2D representation from a top-down view. This BEV feature map encapsulates crucial perception details like object positions, lane markings, and road boundaries [22]. Acting as an information hub, this BEV feature map supports diverse downstream tasks such", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as 3D object tracking, mapping, motion prediction, occupancy prediction, and path planning [16, 18]. Specific task-related information is decoded using a query-based transformer decoder designed for each task to efficiently access the source memory [16]. In the final planning stage, an ego-query models the ego-vehicle's status and interacts with environmental features to determine the optimal path for autonomous driving.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Agent-centric Learning Paradigm: Enhancing BEV Source Memory", + "text_level": 1, + "bbox": [ + 498, + 215, + 890, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The BEV feature map serves as the foundational source memory pool and holds a pivotal role in ADS. Ensuring that a BEV map provides comprehensive information with intricate details for driving is essential for enabling downstream decoders to make safe, precise, and human-like decisions. However, the BEV in ADS is derived from multiview camera images rather than a true bird's-eye-view image, which can introduce discrepancies between the produced BEV and the expected BEV representation. Therefore, to mitigate the discrepancies, we introduce the Agent-centric Learning Paradigm (ALP) to align the produced BEV with a true bird's-eye-view map. Through direct supervision on the BEV map with LM, our ALP enables the refinement of local details and alignment with the desired human perspective.", + "bbox": [ + 496, + 250, + 890, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agent BEV features. In our proposed ALP, three kinds of BEV agents are considered: ego-vehicle (self-driving-car), foreground (FG) objects, and lane elements. We first align the ground-truth area of each agent with the produced BEV map, and crop the regions of interest. We utilize the 3D bounding box to crop the ego-car and FG object area, and panoptic scene mask to segment the lane area. Subsequently, we perform a pooling operation on the obtained local BEV region, to generate a single feature representation for the corresponding agent. After pooling, the local agent features in each sample along the batch are concatenated to formulate an Agent-BEV tensor denoted as $\\mathcal{A}_{bev} \\in \\mathbb{R}^{N_B \\times C}$ , where $N_B$ and $C$ denote total number of agents in the batch and feature dimension, respectively. The process can be formulated as:", + "bbox": [ + 496, + 477, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {b e v} ^ {E g o} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {E g o})) \\text {,} \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {F G} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {F G})) \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {\\text {L a n e}} = \\operatorname {P o o l} \\left(\\operatorname {S e g} \\left(B E V, \\operatorname {S e g M a s k} ^ {\\text {L a n e}}\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {1} \\\\ \\mathcal {A} _ {b e v} = B a t c h \\left(\\left[ \\mathcal {A} _ {b e v} ^ {E g o}; \\mathcal {A} _ {b e v} ^ {F G}; \\mathcal {A} _ {b e v} ^ {L a n e} \\right]\\right), \\in \\mathbb {R} ^ {N _ {b a t c h} \\times C}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 705, + 890, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Batch[;]$ denotes the concatenation operation, $\\mathcal{A}_{bev}^{(\\cdot)}$ represents the single agent feature of the corresponding type, $bbox_{3D}^{(\\cdot)}$ and $SegMask^{Lane}$ indicate the ground-truth 3D bounding box and the lane segmentation mask, respectively.", + "bbox": [ + 496, + 789, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agent expectation features. To ensure that local BEV features express the desired information, we conduct a BEV-", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14e20e0a0aab98af833b8caf17461440292d8bddcff7879081c620ab73773bf0.jpg", + "image_caption": [ + "a)", + "b)", + "Figure 2. a) The overview of proposed vision language planning (VLP) framework. VLP enhances ADS from self-driving BEV-reasoning and self-driving decision-making aspects, through two innovative modules, ALP and SLP, respectively. Leveraging LM and contrastive learning, ALP conducts agent-wise learning for refining local details on BEV, while SLP engages sample-wise learning for advancing global context understanding ability of the ADS. VLP is only active during training, ensuring no additional parameters or computations are introduced during inference. b) Prompt formats used in VLP." + ], + "image_footnote": [ + "Ego-car prompt: This is the self-driving car. It is {class}, and its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. It is currently {command}. Its future trajectory will be {{x1, y1}...{xn, yn}]. Foreground prompt: This object is {class}. Its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. Its future trajectory will be {{x1, y1}...{xt, yt}]. Lane prompt: This lane belongs to a {lane-label}. The label belongs to {thing/stuff} class. Its 2D bounding box on the BEV map is {cx, cy, w, h}. Planning prompt: The self-driving car is driving in an urban area. It is currently {command}. The future trajectory of the car for the next P timestamps will be {{[x1, y1]...{xp, yp}]}" + ], + "bbox": [ + 86, + 92, + 665, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "expectation alignment process by leveraging LM and contrastive learning. We precisely define the perceptual information expected from the corresponding agent, such as agent labels, bounding boxes, and future trajectories. These driving-related ground-truth information, which should also be embedded in the local BEV feature is formulated into a prompt as illustrated in Fig. 2b. The description is then passed to the language encoder, $LM$ , to generate the corresponding agent expectation feature. We apply an MLP layer $\\mathcal{F}_{bev}$ to adapt the expectation feature to the BEV feature space. Then, the agent expectation features are concatenated along the batch to generate an Agent-Expectation tensor denoted as $\\mathcal{A}_{exp} \\in \\mathbb{R}^{N_B \\times C}$ . The procedure can be formulated as in Eq. (2):", + "bbox": [ + 75, + 388, + 472, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {e x p} ^ {E g o} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {E g o} [ y ^ {E g o} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {F G} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {F G} [ y ^ {F G} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {L a n e} = \\mathcal {F} _ {b e v} \\left(L M \\left(\\boldsymbol {T} ^ {L a n e} \\left[ y ^ {L a n e} \\right]\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {2} \\\\ \\mathcal {A} _ {e x p} = B a t c h \\left(\\left[ \\mathcal {A} _ {e x p} ^ {e g o}; \\mathcal {A} _ {e x p} ^ {F G}; \\mathcal {A} _ {e x p} ^ {l a n e} \\right]\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 604, + 467, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_{exp}^{(\\cdot)}, T^{(\\cdot)}$ , and $y^{(\\cdot)}$ represent the single agent expectation feature, the description template, and the ground-truth in ADS for the respective agent. Note that during training, we freeze the $LM$ and only train the adaptation layer $\\mathcal{F}_{bev}$ to save memory and retain the pre-trained knowledge in the $LM$ .", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning Formulation in ALP. Given the agent BEV feature $\\mathcal{A}_{bev}$ and the agent expectation feature $\\mathcal{A}_{exp}$ , the alignment between the produced BEV and the expected BEV is facilitated through a contrastive learning process [27]. In particular, L2 normalization is applied to both $\\mathcal{A}_{bev}$ and $\\mathcal{A}_{exp}$ to standardize the feature vectors and ensure training stability. Then, matrix multiplication is performed between the normalized two-modality tensors with", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a learnable logit scale $\\alpha_{alp}$ to produce a similarity matrix $S_{pred} \\in \\mathbb{R}^{N_B \\times N_B}$ in agent-wise. The ground truth for the similarity matrix, $S_{gt}$ , is a unit matrix, signifying that each agent's BEV feature should be closest to its corresponding expected feature in the shared space. To optimize this alignment, we apply a cross-entropy loss, $\\mathcal{L}_{CE}$ , along both the Agent-BEV mode axis and Agent-Expectation mode axis of the $S_{pred}$ , and compute an average value as the final ALP loss, $\\mathcal{L}_{alp}$ , for the BEV source reasoning part. The entire process can be formally expressed as:", + "bbox": [ + 496, + 388, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\mathcal {S} _ {p r e d} = \\alpha_ {a l p} \\times \\left(\\frac {\\mathcal {A} _ {b e v}}{\\| \\mathcal {A} _ {b e v} \\| _ {2}} \\otimes \\frac {\\mathcal {A} _ {e x p}}{\\| \\mathcal {A} _ {e x p} \\| _ {2}}\\right), \\\\ \\mathcal {L} _ {a l p} = \\left(\\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {q t}, d i m = 0\\right) \\right. \\end{array} \\tag {3} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {g t}, d i m = 1\\right)) / 2. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 550, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. SDC-centric Learning Paradigm: Enhancing Ego-vehicle Query", + "text_level": 1, + "bbox": [ + 500, + 619, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ego-vehicle query feature. The ego-vehicle dynamic modeling is the core function of the ADS. In previous ADS, a trainable ego-query $\\mathcal{E}_{query}$ is applied to interact with other agents, $\\mathcal{A}_{query}^{FG}$ and $\\mathcal{A}_{query}^{Lane}$ , on BEV map, to gather the self-driving perception/prediction information. The produced ego-car query feature $\\mathcal{E}_{qfeat}$ is further processed to predict the future waypoints of the self-driving car $y_{pred}^{plan}$ . The process is formulated as in Eq. (4):", + "bbox": [ + 496, + 657, + 890, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {q f e a t} = \\mathcal {M} ^ {i n t e r} \\left(\\mathcal {E} _ {q u e r y}, \\mathcal {A} _ {q u e r y} ^ {F G}, \\mathcal {A} _ {q u e r y} ^ {l a n e}\\right) \\in \\mathbb {R} ^ {B \\times C}, \\tag {4} \\\\ y _ {p r e d} ^ {\\text {p l a n}} = \\mathcal {M} ^ {\\text {p l a n}} \\left(\\mathcal {E} _ {q f e a t}\\right) \\in \\mathbb {R} ^ {B \\times P \\times 2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 789, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}^{inter}$ and $\\mathcal{M}^{plan}$ denote the intermediate ego information gathering module and the planning head in ADS, respectively. $P \\times 2$ indicate the planned waypoints in the next P timesteps, and $B$ is the batch size during training.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although such mechanism can achieve good planning results, solely relying on numeric ground-truth can make it challenging to understand the rationale behind the ADS' decisions, which can result in inconsistent feature learning for the planning and lead to limited adaptability to new environments. To address these limitations, we propose a SDCcentric Learning Paradigm (SLP) to equip the ADS with the capability of making decisions from continuous and robust feature space.", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ego-vehicle planning feature. We template a language description of the ego-vehicle status with the planning ground truth (GT) including high-level driving command and the future trajectory of the ego-vehicle, as illustrated in Fig. 2b. The prompt is sent to the $LM$ to obtain the ground truth planning features of the ego-car embedding driving scenario information and human driving logic. An MLP layer $\\mathcal{F}_{ego}$ is applied to adapt the textual planning features to the ego-query feature space. The process can be expressed as in Eq. (5):", + "bbox": [ + 75, + 227, + 472, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {p r o m p t} = \\mathcal {F} _ {e g o} (L M (\\boldsymbol {T} _ {s l p} ^ {E g o} [ y ^ {p l a n} ])) \\in \\mathbb {R} ^ {B \\times C}, \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 398, + 468, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{slp}^{Ego}$ indicates the prompt for ego-car description used in SLP, and $y^{plan}$ denotes the planning ground truth. Note that the $LM$ is a shared off-the-shelf model for both ALP and SLP. As in ALP, only the adaptation layer $\\mathcal{F}_{ego}$ is trainable during training to save memory and retain the pre-trained knowledge in $LM$ .", + "bbox": [ + 75, + 431, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive Learning Formulation in SLP. Similar to ALP, we employ a sample-wise contrastive learning approach for SLP to align the ego-vehicle query feature with the ego-vehicle textual planning feature as follows:", + "bbox": [ + 75, + 527, + 468, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {S} _ {p r e d} ^ {s l p} = \\alpha_ {s l p} \\times (\\frac {\\mathcal {E} _ {q f e a t}}{\\| \\mathcal {E} _ {q f e a t} \\| _ {2}} \\otimes \\frac {\\mathcal {E} _ {p r o m p t}}{\\| \\mathcal {E} _ {p r o m p t} \\| _ {2}}), \\\\ \\mathcal {L} _ {s l p} = \\left(\\mathcal {L} _ {C E} \\left(S _ {p r e d} ^ {s l p}, S _ {g t} ^ {s l p}, d i m = 0\\right) \\right. \\tag {6} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d} ^ {s l p}, \\mathcal {S} _ {g t} ^ {s l p}, d i m = 1\\right)) / 2, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 592, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{slp}$ is a learnable logit scale, and $S_{pred}^{slp}$ and $S_{gt}^{slp}$ are the predicted and ground truth similarity matrices, respectively. Aligning the two modes closely in the feature space, the contrastive process refines the produced ego-vehicle query using the ground truth-embedded linguistic feature. The sample-wise contrastive loss rectifies the relationships between individual samples from a human (linguistic) perspective, establishing a connection for the current data point with the world-wise common sense embedded in the pretrained language model.", + "bbox": [ + 75, + 667, + 472, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 76, + 830, + 220, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss, $\\mathcal{L}_{vlp}$ , of our VLP training is composed of two parts: the BEV encoder reasoning loss $\\mathcal{L}_{enc}$ and the decoder decision-making loss $\\mathcal{L}_{dec}$ such that:", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v l p} = \\omega_ {e n c} \\mathcal {L} _ {e n c} + \\omega_ {d e c} \\mathcal {L} _ {d e c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 108, + 890, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{enc}$ and $\\omega_{dec}$ represent the weights for $\\mathcal{L}_{enc}$ and $\\mathcal{L}_{dec}$ , respectively.", + "bbox": [ + 500, + 132, + 890, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoder reasoning loss, $\\mathcal{L}_{enc}$ , is equal to the loss produced in ALP module, i.e. $\\mathcal{L}_{enc} = \\mathcal{L}_{alp}$ . The decision-making loss includes the losses in all downstream tasks (perception/prediction/planning) in ADS and the SLP loss:", + "bbox": [ + 498, + 164, + 890, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e c} = \\mathcal {L} _ {p e r c} + \\mathcal {L} _ {p r e d} + \\mathcal {L} _ {p l a n} + \\mathcal {L} _ {s l p}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 232, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the off-the-shelf LM head is discarded during inference, and thus, our method introduces no additional parameters and computations to the original ADS.", + "bbox": [ + 498, + 256, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 316, + 633, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 342, + 718, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments on the challenging nuScenes dataset [6], which is the first large-scale public dataset to provide data from the entire sensor suite of an autonomous vehicle (6 cameras, 1 LIDAR, 5 RADAR, GPS and IMU sensors). The nuScenes contains 1000 driving scenes from Boston and Singapore, two cities that are known for their dense traffic and highly challenging driving conditions. In our experiments, we utilize 6 camera images as our vision inputs.", + "bbox": [ + 498, + 366, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We adopt two recent SOTA models in autonomous driving, namely UniAD [16] and VAD [18], as our baselines to evaluate the effectiveness of our approach.", + "bbox": [ + 498, + 503, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We incorporate our proposed VLP training strategy into vision-only approaches of UniAD [16] and VAD [18], and refer to this enriched versions as VLP-UniAD and VLP-VAD, respectively, for clarity when reporting the results. All methods are trained with the same hyper-parameters reported in the respective baselines [16, 18] for commensurate comparison. For language model, we use the pretrained LM in CLIP [27]. In VLP study, we set $T = 6$ (3 seconds) for planning, which is same as baselines [16] [18]. Experiments are conducted with 8 NVIDIA Tesla A100 GPUs. More details regarding experimental setup can be found in supplementary materials.", + "bbox": [ + 498, + 549, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Open-loop Planning", + "text_level": 1, + "bbox": [ + 500, + 739, + 691, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Planning constitutes the cornerstone of any ADS, playing a pivotal role in ensuring safety and devising efficient routes for the ego-car. In Tab. 1, we present a series of comparative experiments that showcase the performance of our open-loop planning in comparison to the baseline models. As can be seen in rows 4-6 of the table, the integration of just SLP leads to noticeable reductions in both the L2 error and collision rates for all the baseline models. Moving down the table, rows 7 to 9 demonstrate that the inclusion", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of both VLP components (SLP and ALP together) consistently yields further improvements in these planning metrics. In particular, VLP-UniAD shows a $28.1\\%$ and $48.4\\%$ reduction in terms of average L2 error and collision rate, respectively, compared to baseline UniAD. Similarly, in comparison with VAD, VLP-VAD achieves $35.9\\%$ and $60.5\\%$ reduction for average L2 error and collision rate, respectively. These significant results underscore the effectiveness of both SLP and ALP, as well as their adaptability across various ADS configurations. The reduced L2 error and collision rate achieved through VLP integration contribute to safer road planning in the realm of self-driving.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Perception and Prediction", + "text_level": 1, + "bbox": [ + 76, + 287, + 313, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we showcase the consistent effectiveness of our proposed VLP across a spectrum of essential driving tasks. Our proposed approach excels in various perception and prediction tasks, including multi-object tracking, mapping, motion forecasting, occupancy prediction, 3D object detection, and vectorized scene segmentation. These results underscore the versatility and reliability of our VLP framework across a wide range of critical driving tasks.", + "bbox": [ + 75, + 306, + 468, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Object Tracking. The results for multi-object tracking (MOT) are presented in Tab. 2. We apply the standard evaluation protocols of AMOTA (Average Multi-Object Tracking Accuracy), AMOTP (Average Multi-Object Tracking Precision), Recall, and IDS (Identity Switches) to evaluate the MOT performance. As can be seen, with the incorporation of SLP, AMOTA, Recall, and IDS values all surpass those of UniAD, while AMOTP experiences a slight degradation. The combined integration of both SLP and ALP results in improvements across all metrics when compared to the UniAD baseline. The enhanced performance in MOT proves that our VLP can help the system better predict and respond to the movements of other objects on the road, reducing the risk of accidents.", + "bbox": [ + 75, + 430, + 472, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Online Mapping. In Tab. 3, we present the results of online mapping, encompassing four key mapping elements: lanes, drivable areas, dividers, and pedestrian crossings. The evaluation employs the Intersection-over-Union (IoU) metric to measure the overlap between the predicted and ground-truth maps. The results reveal that the inclusion of the SLP leads to improvements in mapping accuracy for all four elements. Specifically, it enhances the IoU for drivable areas and crossings by $5.5\\%$ and $6.5\\%$ , respectively. Furthermore, the integration of both SLP and ALP yields even further improvement, with IoU achieving by $6.6\\%$ and $10.2\\%$ for drivable areas and crossings, respectively. These enhancements represent substantial gains compared to the SLP-only configuration. These results underscore the valuable contributions of each paradigm in online mapping. They also demonstrate the effectiveness of VLP in bridging the gap between vision-based and language-based information, ul", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "timely enhancing the system's comprehension of various road elements.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Forecasting. We present the motion forecasting results in Tab. 4, applying the same evaluation metrics as those in UniAD [16], namely minADE (minimum Average Displacement Error), minFDE (minimum Final Displacement Error), MR (Miss Rate), and EPA (End-to-end Prediction Accuracy). The results demonstrate that the inclusion of the SLP contributes to a reduction in minADE, minFDE, and MR, while simultaneously increasing EPA for the ADS. This outcome signifies that the supervision and refinement applied to the final ego-car feature also has a positive influence on the motion prediction of other vehicles around the ego-car. The improvements are more obvious with the EPA metrics. Equipped with full components of VLP, the performance of the ADS, in motion prediction, experiences further enhancements. This observation underscores the capability of ALP to empower the system in the identification and accurate prediction of the movements of various agents on the road.", + "bbox": [ + 496, + 123, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Occupancy Prediction. We present the occupancy prediction results in Tab. 5, which have been obtained using the IoU and Video Panoptic Quality (VPQ) within two distance ranges around the ego-car, namely near (\"-n.\", covering a $30 \\times 30m$ area, and far (\"-f.\", spanning a $50 \\times 50m$ area. The results highlight the effectiveness of the SLP in consistently improving all four metrics. SLP, primarily focusing on enhancing the ego-car feature, demonstrates consistent performance gains in occupancy prediction. The inclusion of ALP with SLP shows similar improvement over the baseline as well. Since both configurations of the VLP prioritize attention to the ego-car during training, more improvements are observed within the nearby areas.", + "bbox": [ + 496, + 397, + 890, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Object Detection. We show the 3D object detection results in Tab. 6 using nuScenes detection metrics of mean Average Precision (mAP), Average Translation Error (ATE), Average Scale Error (ASE), and nuScenes detection score (NDS) as indicated in [6]. As observed in the table, the incorporation of the VLP consistently leads to improvements in NDS. Notably, models incorporating both the SLP and ALP generally perform better than those with SLP alone, providing empirical evidence of the ALP's effectiveness in enhancing the local BEV representation ability.", + "bbox": [ + 496, + 595, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vectorized Scene Segmentation in VAD. The VAD [18] framework represents the driving scene as a fully vectorized structure, categorizing map elements into road boundaries, dividers, and pedestrian crossings. Thus, vectorized scene segmentation experiments are conducted within the VAD framework, with IoU serving as the evaluation metric. As depicted in Tab. 7, the introduction of SLP yields substantial improvements. More specifically, the segmentation for boundaries, dividers, and crossings sees improvements of $14.9\\%$ , $19.2\\%$ , and $23.1\\%$ , respectively, contributing to an", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88bbbeb6ad61d7c2571916bcc2d037aa2c1027cc937a755496d7815af89111e3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDModelSLPALPL2 (m) ↓Col. Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
0NMP [36]--2.31---1.92
1SA-NMP [36]--2.05---1.59
2FF [14]0.551.202.541.430.060.171.070.43
3EO [19]0.671.362.781.600.040.090.880.33
4ST-P3 [15]1.332.112.902.110.230.621.270.71
5UniAD [16]0.480.961.651.030.050.170.710.31
6VLP-UniAD✓0.430.861.470.920.030.150.480.22
7VLP-UniAD✓✓0.360.681.190.740.030.120.320.16
8VAD [18]0.460.761.120.780.210.350.580.38
9VLP-VAD✓0.260.470.780.500.120.170.420.23
10VLP-VAD✓✓0.300.530.840.550.010.070.380.15
", + "bbox": [ + 200, + 88, + 766, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74f712f142e8e404cb9d6636225c4ce8f1cf0f3056ce75edcff575169f18d74c.jpg", + "table_caption": [ + "Table 1. Open-loop planning performance. VLP achieves significant end-to-end planning performance improvement over counterpart vision only UniAD and VAD methods on the nuScenes validation dataset [6]. Based on the planning results, we can conclude that both SLP and ALP components plays a vital role to ensure safe motion planning." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPAMOTA↑AMOTP↓Recall↑IDS↓
ViP3D [10]21.71.62536.3-
QD3DT [13]24.21.51839.9-
MUTR3D [37]29.41.49842.73822
UniAD [16]35.91.32046.7906
VLP-UniAD✓36.61.33246.8820
VLP-UniAD✓✓36.81.31547.3678
", + "bbox": [ + 80, + 349, + 465, + 446 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5145210261e75ee1b8749ab7c4d1f3aa6bf25450842312d0621e66a8ab97d210.jpg", + "table_caption": [ + "Table 2. Multi-object tracking. VLP shows improved performance over vision-only MOT techniques." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPLanes↑Drivable↑Divider↑Crossing↑
UniAD [16]31.369.125.713.8
VLP-UniAD✓32.069.527.114.7
VLP-UniAD✓✓32.370.227.415.2
", + "bbox": [ + 80, + 488, + 465, + 550 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/941274d4d3fc374bab96b2d4c13365cba6fd055da843c358d79ad0d5d94573fb.jpg", + "table_caption": [ + "Table 3. Online mapping. VLP demonstrates improved segmentation IoU $(\\%)$ performance over the UniAD." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPminADE(m)↓minFDE(m)↓MR↓EPA↑
PnPNet [23]1.151.950.2260.222
ViP3D [10]2.052.840.2460.226
UniAD [16]0.711.020.1510.456
VLP-UniAD✓0.721.040.1540.459
VLP-UniAD✓✓0.680.980.1330.460
VAD [18]0.781.070.1210.598
VLP-VAD✓0.771.030.1100.621
VLP-VAD✓✓0.771.050.1150.621
", + "bbox": [ + 80, + 590, + 465, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "overall increase of $18.6\\%$ in the mean IoU (mIoU). VLP with SLP and ALP also achieves similar improvement over the VAD baseline. This increased accuracy is crucial for developing a safe and efficient self-driving system.", + "bbox": [ + 75, + 763, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization", + "text_level": 1, + "bbox": [ + 76, + 834, + 227, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Autonomous vehicles are intended to operate in diverse environments. Generalization allows a system to be deployed in various urban landscapes, suburban areas, or even rural", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0fe3c04aa5af30b81cb22779cbe8aabd545155d6b9b3711436715704afc9e295.jpg", + "table_caption": [ + "Table 4. Motion forecasting. VLP achieves better motion forecasting over counterpart vision based methods." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPIoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑
FIERY [12]59.436.750.229.9
StretchBEV [1]55.537.146.029.0
ST-P3 [15]-38.9-32.1
UniAD [16]63.440.254.733.5
VLP-UniAD✓64.240.755.834.5
VLP-UniAD✓✓64.140.255.934.1
", + "bbox": [ + 504, + 349, + 888, + 445 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f86c999808b76a6485a313632bb51e0bfe7166f4d3744e7a197763fb2afc0b1.jpg", + "table_caption": [ + "Table 5. Occupancy prediction. VLP shows improvement in both near $(30\\times 30m)$ and far $(50\\times 50m)$ ranges, denoted as \"n.\" and \"f.\", respectively." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPmAP↑mATE↓mASE↓NDS↑
VAD [18]0.270.700.300.389
VLP-VAD✓0.270.670.300.394
VLP-VAD✓✓0.280.670.300.406
", + "bbox": [ + 503, + 502, + 888, + 573 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24e8c7a9fe665d0f62705f29ab23d2aa6afbdba54651cf82048324c10665c3de.jpg", + "table_caption": [ + "Table 6. 3D object detection. VLP achieves improved object detection performance over the baseline." + ], + "table_footnote": [], + "table_body": "
ModelSLPALPBoundary↑Divider↑Crossing↑mIoU↑
VAD [18]45.642.231.639.8
VLP-VAD✓52.450.338.947.2
VLP-VAD✓✓49.448.439.245.7
", + "bbox": [ + 503, + 614, + 888, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Vectorized scene segmentation. VLP shows improved segmentation IoU (%) performance over the VAD.", + "bbox": [ + 500, + 683, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "settings, making the technology applicable and accessible on a broader scale. Long-tail cases, which are scenarios that occur infrequently, are often underrepresented in training data. Generalization helps the system cope with these rare but critical situations, reducing the risk of biased decision-making based on inadequate exposure during training. In this section, we evaluate the VLP in terms of generalization ability to new cities and long-tail cases.", + "bbox": [ + 496, + 732, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "New-city Generalization. To keep the same number of multi-view camera inputs, we construct a multi-city dataset exclusively from nuScenes, by encompassing data from two", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14766", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b53566a6685299f38a16a40f3e9b594bd4c0d76f73de44d5deefaabee3e7c5e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVLPBostonSingapore
Avg.L2↓Avg.Col↓Avg.L2↓Avg.Col↓
UniAD [16]1.240.321.050.37
VLP-UniAD✓1.140.260.870.34
VAD [18]0.860.270.780.39
VLP-VAD✓0.730.220.630.20
", + "bbox": [ + 81, + 89, + 464, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distinct urban environments, Boston and Singapore. To comprehensively assess generalizability, we conduct two sets of experiments: (i) training on Boston and testing on Singapore; and (ii) conversely, training on Singapore and testing on Boston. The results, presented in Tab. 8, reveal a noteworthy reduction in planning L2 error and collision rates for the baselines in both scenarios with the integration of VLP. In particular, VLP-VAD achieves strong generalization over counterpart vision-only VAD approach (in Boston: $15.1\\%$ and $18.5\\%$ , in Singapore: $19.2\\%$ and $48.7\\%$ reduction in terms of average L2 error and collision rates, respectively). This signifies VLP's prowess in enhancing the safety and reliability of the ADS as well as its capacity to transcend training confines and excel in diverse real-world conditions.", + "bbox": [ + 75, + 260, + 468, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Long-tail Generalization. We assess the performance of UniAD in Multi-Object Tracking (MOT) and VAD in 3D detection concerning long-tail scenarios, as detailed in Tab. 9a and Tab. 9b, respectively. Utilizing the long-tail split methodology from [26], derived from per-class object counts, we observe that VLP consistently enhances the generalization capabilities of both frameworks. Specifically, VLP provides substantial improvements, increasing AMOTA and Recall by $3.7\\%$ and $7.4\\%$ , respectively, for UniAD. In the case of VAD, VLP improves the mean Average Precision (mAP) by $15.9\\%$ . These results highlight the robust feature space cultivated by our proposed VLP, emphasizing its efficacy in handling long-tail scenarios.", + "bbox": [ + 75, + 488, + 468, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 695, + 240, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of SLP and ALP. VLP comprises ALP and SLP components. Extensive experiments presented in prior tables show the importance and necessity of both SLP and ALP for enhanced planning and safety.", + "bbox": [ + 76, + 718, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Format. The design of the prompt format is pivotal for the success of our VLP, serving as the structured input that guides the learning process. We have performed ablation studies to systematically investigate the impact of different components of the ground truth information included in the prompt. The driving ground truth encompasses labels, bounding boxes, trajectories, and high-level commands. To demonstrate the indispensability of including all aspects of", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15f6d463baf6d9b93dee76e7bf9a77c003b941fc74eadbb7631824db70164c3f.jpg", + "table_caption": [ + "Table 8. New-city generalization. To evaluate the generalization performance, we train the model on Boston and test it on Singapore, and vice versa. VLP shows remarkable zero-shot generalization performance improvement over the vision-only methods." + ], + "table_footnote": [], + "table_body": "
ModelVLPAMOTA↑AMOTP↓Recall↑IDS↓
UniAD[16]29.61.44639.268
VLP-UniAD✓30.71.43542.166
", + "bbox": [ + 504, + 89, + 905, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/678de87bf61eb1464135f19e66b1fa69f98195a94b9c56555df2751e11dca488.jpg", + "table_caption": [ + "(a) Multi-object tracking on long-tail cases." + ], + "table_footnote": [], + "table_body": "
ModelVLPmAP↑mATE↓mASE↓mAOE↓
VAD [18]17.60.790.330.95
VLP-VAD✓20.40.750.330.83
", + "bbox": [ + 504, + 171, + 903, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4076e661ea3e61ca583947930aa8a49ecdaa5f1faac35c43403cb30d9e0b61cd.jpg", + "table_caption": [ + "(b) Object detection on long-tail cases.", + "Table 9. Long-tail generalization. VLP shows consistent improvement for long-tail cases across both a) multi-object tracking and b) object detection tasks." + ], + "table_footnote": [], + "table_body": "
label-gtbbox-gttraj-gtcommd-gtAvg.L2↓Avg.Col↓
X0.640.23
X0.560.30
X0.590.36
X0.590.26
✓✓✓✓0.520.17
", + "bbox": [ + 521, + 306, + 867, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Ablation for prompt information. All ground-truth information components contributes to improved planning performance. We provide full task ablation results in the supplementary.", + "bbox": [ + 498, + 398, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ground truth in the prompt, we conducted experiments where each element was selectively removed in individual settings. The results, presented in Tab. 10, clearly indicate that the exclusion of any ground truth component leads to a degradation in planning ability. This underscores the necessity of incorporating the entire spectrum of ground truth information in the prompt for optimal performance.", + "bbox": [ + 496, + 462, + 890, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a novel Vision-Language-Planning (VLP) approach to enhance the capabilities of Autonomous Driving Systems (ADS). Our approach leverages both self-driving-car-centric learning paradigm (SLP) and agent-wise learning paradigm (ALP) guided by language prompts to create a comprehensive understanding of the environment. Through a series of experiments on various driving tasks, we have demonstrated the effectiveness of our VLP approach in improving perception, prediction, and planning aspects of ADS. The generalization experiments showcased the robustness of our VLP approach, proving its adaptability to new cities and long-tail cases. By extending the capabilities of ADS beyond the training environment, our VLP approach paves the way for safer and more reliable autonomous driving in real-world conditions.", + "bbox": [ + 496, + 612, + 890, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our experiments are currently confined to the nuScenes dataset and camera modality as baseline vision-based approaches. We will assess VLP on a broader range of datasets and sensor modalities in our future work.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 7", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3", + "[3] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 2", + "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2, 3", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6, 7", + "[7] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2", + "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2", + "[9] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3", + "[10] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 2, 7", + "[11] Md Zahid Hasan, Jiajing Chen, Jiyang Wang, Ameya Joshi, Senem Velipasalar, Chinmay Hegde, Anuj Sharma, and Soumik Sarkar. Vision-language models can identify distracted driver behavior from naturalistic videos. arXiv preprint arXiv:2306.10159, 2023. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 7", + "[13] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1992-2008, 2022. 7", + "[14] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 7", + "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 7", + "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3", + "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1, 2, 3, 5, 6, 7, 8", + "[19] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. In European Conference on Computer Vision, pages 353-369. Springer, 2022. 7", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3", + "[21] Liunian Harold Li, Mark Yatskar, D Yin, CJ Hsieh, and KW Chang. Visualbert: A simple and performant baseline for vision and language. arxiv 2019. arXiv preprint arXiv:1908.03557. 2", + "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 3", + "[23] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end per" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14768", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 2, 7", + "[24] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3", + "[25] OpenAI. Introducing chatgpt, 2023. 1", + "[26] Neehar Peri, Achal Dave, Deva Ramanan, and Shu Kong. Towards long-tailed 3d detection. In Conference on Robot Learning, pages 1904–1915. PMLR, 2023. 8", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5", + "[28] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 2", + "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[30] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 3", + "[31] Wayve. Lingo-1: Exploring natural language for autonomous driving. https://wayve.ai/thinking/lingo - natural - language - autonomous - driving/, 2023.3", + "[32] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. arXiv preprint arXiv:2309.16292, 2023. 3", + "[33] Shawn Xu, Lin Yang, Christopher Kelly, Marcin Sieniek, Timo Kohlberger, Martin Ma, Wei-Hung Weng, Attila Kiraly, Sahar Kazemzadeh, Zakkai Melamed, et al. Elixir: Towards a general purpose x-ray artificial intelligence system through alignment of large language models and radiology vision encoders. arXiv preprint arXiv:2308.01317, 2023. 2", + "[34] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 3", + "[35] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 7", + "[37] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. Mutr3d: A multi-camera tracking framework via 3d-to-2d queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4537–4546, 2022. 7" + ], + "bbox": [ + 501, + 90, + 890, + 231 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14769", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_model.json b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1eac41952fcd19acb499bb6d9183296fb7cb5dda --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.776, + 0.154 + ], + "angle": 0, + "content": "VLP: Vision Language Planning for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.18, + 0.777, + 0.234 + ], + "angle": 0, + "content": "Chenbin Pan\\(^{1*}\\) Burhaneddin Yaman\\(^{2\\text{圆}}\\) Tommaso Nesti\\(^{2}\\) Abhirup Mallik\\(^{2}\\) Alessandro G Allievi\\(^{2}\\) Senem Velipasalar\\(^{1}\\) Liu Ren\\(^{2}\\) \\(^{1}\\)Syracuse University" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.235, + 0.813, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.255, + 0.595, + 0.27 + ], + "angle": 0, + "content": "{cpan14, svelipas}@syr.edu," + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.272, + 0.871, + 0.288 + ], + "angle": 0, + "content": "{burhaneddin.yaman,tommaso.nesti,abhirup.mallik,alessandro.allievi,liu. ren}@us.bosch.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.658 + ], + "angle": 0, + "content": "Autonomous driving is a complex and challenging task that aims at safe motion planning through scene understanding and reasoning. While vision-only autonomous driving methods have recently achieved notable performance, through enhanced scene understanding, several key issues, including lack of reasoning, low generalization performance and long-tail scenarios, still need to be addressed. In this paper, we present VLP, a novel Vision-Language-Planning framework that exploits language models to bridge the gap between linguistic understanding and autonomous driving. VLP enhances autonomous driving systems by strengthening both the source memory foundation and the self-driving car's contextual understanding. VLP achieves state-of-the-art end-to-end planning performance on the challenging NuScenes dataset by achieving \\(35.9\\%\\) and \\(60.5\\%\\) reduction in terms of average L2 error and collision rates, respectively, compared to the previous best method. Moreover, VLP shows improved performance in challenging long-tail scenarios and strong generalization capabilities when faced with new urban environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.687, + 0.21, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.865 + ], + "angle": 0, + "content": "Autonomous driving is a complex problem requiring scene understanding and reasoning to ensure safe motion planning. This sophisticated challenge can be broadly divided into three main tasks, namely perception, prediction and planning (also known as P3). Conventional methods adopt a modular approach by developing and optimizing each task in a disjoint manner without a holistic view, leading to compounding errors and safety concerns [15, 16, 18]. End-to-end autonomous driving systems (ADS), unifying all P3 tasks, have garnered attention for their potential to enhance" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.337, + 0.691, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.337, + 0.882, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.481, + 0.893, + 0.552 + ], + "angle": 0, + "content": "Figure 1. New-city generalization ability of ADS for planning is evaluated by training on Boston city and testing on Singapore, and vice versa. Our proposed VLP shows strong generalization ability by significantly outperforming state-of-the-art vision-only method, VAD[18], in terms of both L2 error and collision rate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.718 + ], + "angle": 0, + "content": "safe planning. Existing vision-based ADS typically follow a two-stage process: bird's eye view (BEV) feature extraction and downstream tasks [16, 18]. BEV feature extraction transforms multi-view camera data into a structured top-view representation embedding spatial information around the ego-car. BEV features are further utilized as the information pool by the downstream P3 tasks. While effective, these vision-only methods struggle with out-of-domain generalization, such as maintaining performance in new cities and long-tail scenarios, hindering real-world deployment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The way humans process and interpret visual driving scenes naturally involve a coherent cognitive framework that maintains a consistent logical flow. This enables humans effortlessly make correct decisions even when faced with previously unseen data or scenarios, eliminating generalization issues in driving tasks. Then, this raises a pertinent question: How can ADS achieve human-like driving to navigate diverse and dynamic road environments? Recent advancements in language models (LMs) have led to unprecedented common-sense ability and generalization performance across unseen data and tasks for the natural language processing domain [5, 8, 25]. The superior common-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.875, + 0.428, + 0.888 + ], + "angle": 0, + "content": "*Work done while interned at Bosch Research North America." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.229, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.875, + 0.428, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "sense capability of LMs have led to the emergence of multimodal LMs for diverse applications ranging from medical imaging to robotics [4, 9, 33]. However, incorporation of the reasoning ability of LMs into real-world autonomous driving tasks, to address generalization and long-tail scenarios, is yet to be fully-explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.305 + ], + "angle": 0, + "content": "To bridge this gap, we propose a Vision Language Planning (VLP) framework, which integrates the commonsense capability of LMs into vision-based ADS for safe self-driving. Our VLP consists of two key components: Agent-centric Learning Paradigm (ALP) and Self-driving-car-centric Learning Paradigm (SLP), leveraging LMs to enhance the ADS from reasoning and decision-making aspects, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.468, + 0.548 + ], + "angle": 0, + "content": "The BEV feature map serves as the source memory pool in ADS for downstream decoding tasks. It summarizes and encodes the driving environment surrounding the self-driving car, including vehicles, pedestrians, lanes, and more, into a unified feature map. Hence, capturing comprehensive and necessary details in each local position of BEV is critical for safe and precise self-driving performance. To enhance the local semantic representation and reasoning capabilities of BEV, we introduce an innovative Agent-centric Learning Paradigm (ALP) module. ALP integrates the consistent feature space of a pretrained LM to revamp the agent features on the BEV, actively shaping semantics, and guiding the BEV reasoning process. Leveraging on the common sense and logic flow embedded in the LM, our ALP equips the ADS with robustness and consistent BEV feature space, enhancing its effectiveness in diverse driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.468, + 0.791 + ], + "angle": 0, + "content": "In ADS, the planning module aggregates information from the preceding perception and prediction phases to make the final decisions for self-driving. This global perspective culminates in the formation of a planning query, directly influencing the safety and accuracy of the self-driving navigation. Considering the critical role of the planning module within ADS, we also present a novel Self-driving-car-centric Learning Paradigm (SLP) to elevate the decoding and acquiring information ability of the planning query. In the SLP, we align the planning query with intended goals and the ego-vehicle driving status by leveraging the knowledge encoded in the pretrained LM. The language model's comprehension capabilities contribute to more informed decision-making during the planning phase as well as enabling a more robust planning query formation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.468, + 0.882 + ], + "angle": 0, + "content": "Through VLP, we bridge the gap between human-like reasoning and autonomous driving, enhancing the model's contextual awareness and its ability to generalize effectively in complex, ever-changing real-world scenarios, as shown in Fig.1. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We propose VLP, a Vision Language Planning model," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.89, + 0.136 + ], + "angle": 0, + "content": "which incorporates reasoning capability of LMs into vision-based autonomous driving systems as an enhancement of motion planning and self-driving safety." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.89, + 0.195 + ], + "angle": 0, + "content": "- VLP is composed of novel components ALP and SLP, aiming to improve the ADS from self-driving BEV reasoning and self-driving decision-making aspects, respectively." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.198, + 0.89, + 0.285 + ], + "angle": 0, + "content": "- Through extensive experiments in real-world driving scenarios, we show that VLP significantly and consistently outperforms the state-of-the-art vision-based approaches across a spectrum of driving tasks, including open-loop planning, multi-object tracking, motion forecasting, and more." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.288, + 0.89, + 0.36 + ], + "angle": 0, + "content": "- We conduct the first new-city generalization study on the nuScenes dataset [6] by training and testing on distinct cities, demonstrating the remarkable zero-shot generalization ability of our VLP approach over vision-only methods." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.363, + 0.89, + 0.407 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work introducing LM into multiple stages of ADS to address the generalization ability in new cities and long-tail cases." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.137, + 0.89, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.64, + 0.436 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.446, + 0.792, + 0.461 + ], + "angle": 0, + "content": "2.1. End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.768 + ], + "angle": 0, + "content": "Early end-to-end approaches have adopted a single neural network to perform motion planning without explicitly designing intermediate tasks, such as prediction [3]. Thus, such approaches not only suffer from the sub-optimal performance but also lack interpretability. Modular end-to-end frameworks introduce intermediate tasks for interpretability and provide improved performance [7, 10, 15, 16, 18, 23, 28]. These interpretable end-to-end frameworks explicitly model perception, prediction and planning (P3) components and train them together with a joint optimization strategy to enhance planning. UniAD [16] and VAD [18] are two prominent state-of-the-art end-to-end frameworks. UniAD leverages rasterized scene representations and explicitly identifies crucial components within the P3 framework. It coordinates all tasks through a unified query design for safe planning. On the other hand, VAD uses vectorized representations for efficient planning and safety improvement. While these approaches generally achieve good performance, their generalization performance have been limited due to lack of common sense and reasoning process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.727, + 0.791 + ], + "angle": 0, + "content": "2.2. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Large language models (LLMs), which are trained on massive amounts of text data, have shown unprecedented success in language-related tasks [5, 8, 29]. The capability of LLMs for common-sense understanding and versatility in tackling diverse language tasks has driven the exploration of multi-modal LLMs [20, 21, 27]. Vision-language models (VLM) integrate visual and textual information to en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14761" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "able common sense understanding of the content by training on large scale open-world vision and text data [2, 17, 20, 27]. CLIP [27] uses a contrastive learning objective to learn a joint embedding space for vision and text data, and has shown immense zero-shot generalization performance, showcasing the importance of VLMs compared to vision-only approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.208, + 0.455, + 0.226 + ], + "angle": 0, + "content": "2.3. Autonomous Driving with Language Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.471 + ], + "angle": 0, + "content": "In recent years, numerous works have been proposed for extending LMS to the embodied AI domain to improve zero-shot generalization performance [4, 9, 30, 35]. These works incorporate sensory inputs from embodied agents in the form of language with vision and language data for decision making process [4, 9]. While these methods have mainly focused on the robotics domain, few works have been proposed for leveraging embodied LMs for autonomous driving tasks [11, 24, 31, 32, 34]. In particular, DiLu [32] and GPT-Driver [24] propose GPT-based driver agents for closed-loop simulation tasks. In [31], an open loop driving commentator, which combines vision and low-level driving actions with language, is proposed to interpret driving actions and for reasoning. Our work significantly differs from these approaches by leveraging LMs for end-to-end motion planning task for real-world driving applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.214, + 0.503 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.473, + 0.72 + ], + "angle": 0, + "content": "We present a Vision Language Planning (VLP) model, which equips the ADS with the capacity to emulate human common sense and engage in contextual reasoning for safe motion planning. Our proposed VLP model, illustrated in Fig. 2a, comprises two innovative components that leverage LMs in both local and global contexts. The Agent-centric Learning Paradigm (ALP) concentrates on refining local details to enhance source memory reasoning, while the Self-driving-car-centric Learning Paradigm (SLP) focuses on guiding the planning process for the self-driving-car (SDC). Below, we first provide an overview of how general ADS work in Sec.3.1. Subsequently, we describe our proposed ALP and SLP in detail in Sec. 3.2 and Sec. 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.208, + 0.746 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In vision-based ADS, a sequence of multi-view camera images serves as the source input, offering abundant visual data for downstream tasks. Initially, an image backbone extracts fundamental features for perception. These features are then processed by a BEV encoder to transform multi-view features into a unified 2D representation from a top-down view. This BEV feature map encapsulates crucial perception details like object positions, lane markings, and road boundaries [22]. Acting as an information hub, this BEV feature map supports diverse downstream tasks such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "as 3D object tracking, mapping, motion prediction, occupancy prediction, and path planning [16, 18]. Specific task-related information is decoded using a query-based transformer decoder designed for each task to efficiently access the source memory [16]. In the final planning stage, an ego-query models the ego-vehicle's status and interacts with environmental features to determine the optimal path for autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.217, + 0.892, + 0.248 + ], + "angle": 0, + "content": "3.2. Agent-centric Learning Paradigm: Enhancing BEV Source Memory" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.478 + ], + "angle": 0, + "content": "The BEV feature map serves as the foundational source memory pool and holds a pivotal role in ADS. Ensuring that a BEV map provides comprehensive information with intricate details for driving is essential for enabling downstream decoders to make safe, precise, and human-like decisions. However, the BEV in ADS is derived from multiview camera images rather than a true bird's-eye-view image, which can introduce discrepancies between the produced BEV and the expected BEV representation. Therefore, to mitigate the discrepancies, we introduce the Agent-centric Learning Paradigm (ALP) to align the produced BEV with a true bird's-eye-view map. Through direct supervision on the BEV map with LM, our ALP enables the refinement of local details and alignment with the desired human perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Agent BEV features. In our proposed ALP, three kinds of BEV agents are considered: ego-vehicle (self-driving-car), foreground (FG) objects, and lane elements. We first align the ground-truth area of each agent with the produced BEV map, and crop the regions of interest. We utilize the 3D bounding box to crop the ego-car and FG object area, and panoptic scene mask to segment the lane area. Subsequently, we perform a pooling operation on the obtained local BEV region, to generate a single feature representation for the corresponding agent. After pooling, the local agent features in each sample along the batch are concatenated to formulate an Agent-BEV tensor denoted as \\(\\mathcal{A}_{bev} \\in \\mathbb{R}^{N_B \\times C}\\), where \\(N_B\\) and \\(C\\) denote total number of agents in the batch and feature dimension, respectively. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.707, + 0.892, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {b e v} ^ {E g o} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {E g o})) \\text {,} \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {F G} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {F G})) \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {\\text {L a n e}} = \\operatorname {P o o l} \\left(\\operatorname {S e g} \\left(B E V, \\operatorname {S e g M a s k} ^ {\\text {L a n e}}\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {1} \\\\ \\mathcal {A} _ {b e v} = B a t c h \\left(\\left[ \\mathcal {A} _ {b e v} ^ {E g o}; \\mathcal {A} _ {b e v} ^ {F G}; \\mathcal {A} _ {b e v} ^ {L a n e} \\right]\\right), \\in \\mathbb {R} ^ {N _ {b a t c h} \\times C}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.871 + ], + "angle": 0, + "content": "where \\( Batch[;] \\) denotes the concatenation operation, \\( \\mathcal{A}_{bev}^{(\\cdot)} \\) represents the single agent feature of the corresponding type, \\( bbox_{3D}^{(\\cdot)} \\) and \\( SegMask^{Lane} \\) indicate the ground-truth 3D bounding box and the lane segmentation mask, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Agent expectation features. To ensure that local BEV features express the desired information, we conduct a BEV-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14762" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.093, + 0.1, + 0.105 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.666, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.094, + 0.678, + 0.105 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.683, + 0.097, + 0.884, + 0.278 + ], + "angle": 0, + "content": "Ego-car prompt: This is the self-driving car. It is {class}, and its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. It is currently {command}. Its future trajectory will be {{x1, y1}...{xn, yn}]. Foreground prompt: This object is {class}. Its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. Its future trajectory will be {{x1, y1}...{xt, yt}]. Lane prompt: This lane belongs to a {lane-label}. The label belongs to {thing/stuff} class. Its 2D bounding box on the BEV map is {cx, cy, w, h}. Planning prompt: The self-driving car is driving in an urban area. It is currently {command}. The future trajectory of the car for the next P timestamps will be {{[x1, y1]...{xp, yp}]}" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Figure 2. a) The overview of proposed vision language planning (VLP) framework. VLP enhances ADS from self-driving BEV-reasoning and self-driving decision-making aspects, through two innovative modules, ALP and SLP, respectively. Leveraging LM and contrastive learning, ALP conducts agent-wise learning for refining local details on BEV, while SLP engages sample-wise learning for advancing global context understanding ability of the ADS. VLP is only active during training, ensuring no additional parameters or computations are introduced during inference. b) Prompt formats used in VLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.603 + ], + "angle": 0, + "content": "expectation alignment process by leveraging LM and contrastive learning. We precisely define the perceptual information expected from the corresponding agent, such as agent labels, bounding boxes, and future trajectories. These driving-related ground-truth information, which should also be embedded in the local BEV feature is formulated into a prompt as illustrated in Fig. 2b. The description is then passed to the language encoder, \\(LM\\), to generate the corresponding agent expectation feature. We apply an MLP layer \\(\\mathcal{F}_{bev}\\) to adapt the expectation feature to the BEV feature space. Then, the agent expectation features are concatenated along the batch to generate an Agent-Expectation tensor denoted as \\(\\mathcal{A}_{exp} \\in \\mathbb{R}^{N_B \\times C}\\). The procedure can be formulated as in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.606, + 0.468, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {e x p} ^ {E g o} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {E g o} [ y ^ {E g o} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {F G} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {F G} [ y ^ {F G} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {L a n e} = \\mathcal {F} _ {b e v} \\left(L M \\left(\\boldsymbol {T} ^ {L a n e} \\left[ y ^ {L a n e} \\right]\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {2} \\\\ \\mathcal {A} _ {e x p} = B a t c h \\left(\\left[ \\mathcal {A} _ {e x p} ^ {e g o}; \\mathcal {A} _ {e x p} ^ {F G}; \\mathcal {A} _ {e x p} ^ {l a n e} \\right]\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.778 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_{exp}^{(\\cdot)}, T^{(\\cdot)}\\), and \\(y^{(\\cdot)}\\) represent the single agent expectation feature, the description template, and the ground-truth in ADS for the respective agent. Note that during training, we freeze the \\(LM\\) and only train the adaptation layer \\(\\mathcal{F}_{bev}\\) to save memory and retain the pre-trained knowledge in the \\(LM\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Contrastive Learning Formulation in ALP. Given the agent BEV feature \\(\\mathcal{A}_{bev}\\) and the agent expectation feature \\(\\mathcal{A}_{exp}\\), the alignment between the produced BEV and the expected BEV is facilitated through a contrastive learning process [27]. In particular, L2 normalization is applied to both \\(\\mathcal{A}_{bev}\\) and \\(\\mathcal{A}_{exp}\\) to standardize the feature vectors and ensure training stability. Then, matrix multiplication is performed between the normalized two-modality tensors with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.542 + ], + "angle": 0, + "content": "a learnable logit scale \\(\\alpha_{alp}\\) to produce a similarity matrix \\(S_{pred} \\in \\mathbb{R}^{N_B \\times N_B}\\) in agent-wise. The ground truth for the similarity matrix, \\(S_{gt}\\), is a unit matrix, signifying that each agent's BEV feature should be closest to its corresponding expected feature in the shared space. To optimize this alignment, we apply a cross-entropy loss, \\(\\mathcal{L}_{CE}\\), along both the Agent-BEV mode axis and Agent-Expectation mode axis of the \\(S_{pred}\\), and compute an average value as the final ALP loss, \\(\\mathcal{L}_{alp}\\), for the BEV source reasoning part. The entire process can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.551, + 0.892, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\mathcal {S} _ {p r e d} = \\alpha_ {a l p} \\times \\left(\\frac {\\mathcal {A} _ {b e v}}{\\| \\mathcal {A} _ {b e v} \\| _ {2}} \\otimes \\frac {\\mathcal {A} _ {e x p}}{\\| \\mathcal {A} _ {e x p} \\| _ {2}}\\right), \\\\ \\mathcal {L} _ {a l p} = \\left(\\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {q t}, d i m = 0\\right) \\right. \\end{array} \\tag {3} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {g t}, d i m = 1\\right)) / 2. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.891, + 0.651 + ], + "angle": 0, + "content": "3.3. SDC-centric Learning Paradigm: Enhancing Ego-vehicle Query" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Ego-vehicle query feature. The ego-vehicle dynamic modeling is the core function of the ADS. In previous ADS, a trainable ego-query \\(\\mathcal{E}_{query}\\) is applied to interact with other agents, \\(\\mathcal{A}_{query}^{FG}\\) and \\(\\mathcal{A}_{query}^{Lane}\\), on BEV map, to gather the self-driving perception/prediction information. The produced ego-car query feature \\(\\mathcal{E}_{qfeat}\\) is further processed to predict the future waypoints of the self-driving car \\(y_{pred}^{plan}\\). The process is formulated as in Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {q f e a t} = \\mathcal {M} ^ {i n t e r} \\left(\\mathcal {E} _ {q u e r y}, \\mathcal {A} _ {q u e r y} ^ {F G}, \\mathcal {A} _ {q u e r y} ^ {l a n e}\\right) \\in \\mathbb {R} ^ {B \\times C}, \\tag {4} \\\\ y _ {p r e d} ^ {\\text {p l a n}} = \\mathcal {M} ^ {\\text {p l a n}} \\left(\\mathcal {E} _ {q f e a t}\\right) \\in \\mathbb {R} ^ {B \\times P \\times 2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}^{inter}\\) and \\(\\mathcal{M}^{plan}\\) denote the intermediate ego information gathering module and the planning head in ADS, respectively. \\(P \\times 2\\) indicate the planned waypoints in the next P timesteps, and \\(B\\) is the batch size during training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14763" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "Although such mechanism can achieve good planning results, solely relying on numeric ground-truth can make it challenging to understand the rationale behind the ADS' decisions, which can result in inconsistent feature learning for the planning and lead to limited adaptability to new environments. To address these limitations, we propose a SDCcentric Learning Paradigm (SLP) to equip the ADS with the capability of making decisions from continuous and robust feature space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.381 + ], + "angle": 0, + "content": "Ego-vehicle planning feature. We template a language description of the ego-vehicle status with the planning ground truth (GT) including high-level driving command and the future trajectory of the ego-vehicle, as illustrated in Fig. 2b. The prompt is sent to the \\(LM\\) to obtain the ground truth planning features of the ego-car embedding driving scenario information and human driving logic. An MLP layer \\(\\mathcal{F}_{ego}\\) is applied to adapt the textual planning features to the ego-query feature space. The process can be expressed as in Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.399, + 0.47, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {p r o m p t} = \\mathcal {F} _ {e g o} (L M (\\boldsymbol {T} _ {s l p} ^ {E g o} [ y ^ {p l a n} ])) \\in \\mathbb {R} ^ {B \\times C}, \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.527 + ], + "angle": 0, + "content": "where \\( T_{slp}^{Ego} \\) indicates the prompt for ego-car description used in SLP, and \\( y^{plan} \\) denotes the planning ground truth. Note that the \\( LM \\) is a shared off-the-shelf model for both ALP and SLP. As in ALP, only the adaptation layer \\( \\mathcal{F}_{ego} \\) is trainable during training to save memory and retain the pre-trained knowledge in \\( LM \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Contrastive Learning Formulation in SLP. Similar to ALP, we employ a sample-wise contrastive learning approach for SLP to align the ego-vehicle query feature with the ego-vehicle textual planning feature as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.593, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {S} _ {p r e d} ^ {s l p} = \\alpha_ {s l p} \\times (\\frac {\\mathcal {E} _ {q f e a t}}{\\| \\mathcal {E} _ {q f e a t} \\| _ {2}} \\otimes \\frac {\\mathcal {E} _ {p r o m p t}}{\\| \\mathcal {E} _ {p r o m p t} \\| _ {2}}), \\\\ \\mathcal {L} _ {s l p} = \\left(\\mathcal {L} _ {C E} \\left(S _ {p r e d} ^ {s l p}, S _ {g t} ^ {s l p}, d i m = 0\\right) \\right. \\tag {6} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d} ^ {s l p}, \\mathcal {S} _ {g t} ^ {s l p}, d i m = 1\\right)) / 2, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.473, + 0.822 + ], + "angle": 0, + "content": "where \\(\\alpha_{slp}\\) is a learnable logit scale, and \\(S_{pred}^{slp}\\) and \\(S_{gt}^{slp}\\) are the predicted and ground truth similarity matrices, respectively. Aligning the two modes closely in the feature space, the contrastive process refines the produced ego-vehicle query using the ground truth-embedded linguistic feature. The sample-wise contrastive loss rectifies the relationships between individual samples from a human (linguistic) perspective, establishing a connection for the current data point with the world-wise common sense embedded in the pretrained language model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.221, + 0.848 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The overall loss, \\(\\mathcal{L}_{vlp}\\), of our VLP training is composed of two parts: the BEV encoder reasoning loss \\(\\mathcal{L}_{enc}\\) and the decoder decision-making loss \\(\\mathcal{L}_{dec}\\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.109, + 0.892, + 0.125 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v l p} = \\omega_ {e n c} \\mathcal {L} _ {e n c} + \\omega_ {d e c} \\mathcal {L} _ {d e c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.133, + 0.892, + 0.163 + ], + "angle": 0, + "content": "where \\(\\omega_{enc}\\) and \\(\\omega_{dec}\\) represent the weights for \\(\\mathcal{L}_{enc}\\) and \\(\\mathcal{L}_{dec}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.165, + 0.892, + 0.225 + ], + "angle": 0, + "content": "The encoder reasoning loss, \\(\\mathcal{L}_{enc}\\), is equal to the loss produced in ALP module, i.e. \\(\\mathcal{L}_{enc} = \\mathcal{L}_{alp}\\). The decision-making loss includes the losses in all downstream tasks (perception/prediction/planning) in ADS and the SLP loss:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.233, + 0.892, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e c} = \\mathcal {L} _ {p e r c} + \\mathcal {L} _ {p r e d} + \\mathcal {L} _ {p l a n} + \\mathcal {L} _ {s l p}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Note that the off-the-shelf LM head is discarded during inference, and thus, our method introduces no additional parameters and computations to the original ADS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.634, + 0.335 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.343, + 0.719, + 0.359 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Dataset. We conduct experiments on the challenging nuScenes dataset [6], which is the first large-scale public dataset to provide data from the entire sensor suite of an autonomous vehicle (6 cameras, 1 LIDAR, 5 RADAR, GPS and IMU sensors). The nuScenes contains 1000 driving scenes from Boston and Singapore, two cities that are known for their dense traffic and highly challenging driving conditions. In our experiments, we utilize 6 camera images as our vision inputs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Baselines. We adopt two recent SOTA models in autonomous driving, namely UniAD [16] and VAD [18], as our baselines to evaluate the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Training. We incorporate our proposed VLP training strategy into vision-only approaches of UniAD [16] and VAD [18], and refer to this enriched versions as VLP-UniAD and VLP-VAD, respectively, for clarity when reporting the results. All methods are trained with the same hyper-parameters reported in the respective baselines [16, 18] for commensurate comparison. For language model, we use the pretrained LM in CLIP [27]. In VLP study, we set \\( T = 6 \\) (3 seconds) for planning, which is same as baselines [16] [18]. Experiments are conducted with 8 NVIDIA Tesla A100 GPUs. More details regarding experimental setup can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.692, + 0.758 + ], + "angle": 0, + "content": "4.2. Open-loop Planning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Planning constitutes the cornerstone of any ADS, playing a pivotal role in ensuring safety and devising efficient routes for the ego-car. In Tab. 1, we present a series of comparative experiments that showcase the performance of our open-loop planning in comparison to the baseline models. As can be seen in rows 4-6 of the table, the integration of just SLP leads to noticeable reductions in both the L2 error and collision rates for all the baseline models. Moving down the table, rows 7 to 9 demonstrate that the inclusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14764" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "of both VLP components (SLP and ALP together) consistently yields further improvements in these planning metrics. In particular, VLP-UniAD shows a \\(28.1\\%\\) and \\(48.4\\%\\) reduction in terms of average L2 error and collision rate, respectively, compared to baseline UniAD. Similarly, in comparison with VAD, VLP-VAD achieves \\(35.9\\%\\) and \\(60.5\\%\\) reduction for average L2 error and collision rate, respectively. These significant results underscore the effectiveness of both SLP and ALP, as well as their adaptability across various ADS configurations. The reduced L2 error and collision rate achieved through VLP integration contribute to safer road planning in the realm of self-driving." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.288, + 0.315, + 0.304 + ], + "angle": 0, + "content": "4.3. Perception and Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.43 + ], + "angle": 0, + "content": "In this section, we showcase the consistent effectiveness of our proposed VLP across a spectrum of essential driving tasks. Our proposed approach excels in various perception and prediction tasks, including multi-object tracking, mapping, motion forecasting, occupancy prediction, 3D object detection, and vectorized scene segmentation. These results underscore the versatility and reliability of our VLP framework across a wide range of critical driving tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.473, + 0.644 + ], + "angle": 0, + "content": "Multi-Object Tracking. The results for multi-object tracking (MOT) are presented in Tab. 2. We apply the standard evaluation protocols of AMOTA (Average Multi-Object Tracking Accuracy), AMOTP (Average Multi-Object Tracking Precision), Recall, and IDS (Identity Switches) to evaluate the MOT performance. As can be seen, with the incorporation of SLP, AMOTA, Recall, and IDS values all surpass those of UniAD, while AMOTP experiences a slight degradation. The combined integration of both SLP and ALP results in improvements across all metrics when compared to the UniAD baseline. The enhanced performance in MOT proves that our VLP can help the system better predict and respond to the movements of other objects on the road, reducing the risk of accidents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Online Mapping. In Tab. 3, we present the results of online mapping, encompassing four key mapping elements: lanes, drivable areas, dividers, and pedestrian crossings. The evaluation employs the Intersection-over-Union (IoU) metric to measure the overlap between the predicted and ground-truth maps. The results reveal that the inclusion of the SLP leads to improvements in mapping accuracy for all four elements. Specifically, it enhances the IoU for drivable areas and crossings by \\(5.5\\%\\) and \\(6.5\\%\\), respectively. Furthermore, the integration of both SLP and ALP yields even further improvement, with IoU achieving by \\(6.6\\%\\) and \\(10.2\\%\\) for drivable areas and crossings, respectively. These enhancements represent substantial gains compared to the SLP-only configuration. These results underscore the valuable contributions of each paradigm in online mapping. They also demonstrate the effectiveness of VLP in bridging the gap between vision-based and language-based information, ul" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "timely enhancing the system's comprehension of various road elements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Motion Forecasting. We present the motion forecasting results in Tab. 4, applying the same evaluation metrics as those in UniAD [16], namely minADE (minimum Average Displacement Error), minFDE (minimum Final Displacement Error), MR (Miss Rate), and EPA (End-to-end Prediction Accuracy). The results demonstrate that the inclusion of the SLP contributes to a reduction in minADE, minFDE, and MR, while simultaneously increasing EPA for the ADS. This outcome signifies that the supervision and refinement applied to the final ego-car feature also has a positive influence on the motion prediction of other vehicles around the ego-car. The improvements are more obvious with the EPA metrics. Equipped with full components of VLP, the performance of the ADS, in motion prediction, experiences further enhancements. This observation underscores the capability of ALP to empower the system in the identification and accurate prediction of the movements of various agents on the road." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Occupancy Prediction. We present the occupancy prediction results in Tab. 5, which have been obtained using the IoU and Video Panoptic Quality (VPQ) within two distance ranges around the ego-car, namely near (\"-n.\", covering a \\(30 \\times 30m\\) area, and far (\"-f.\", spanning a \\(50 \\times 50m\\) area. The results highlight the effectiveness of the SLP in consistently improving all four metrics. SLP, primarily focusing on enhancing the ego-car feature, demonstrates consistent performance gains in occupancy prediction. The inclusion of ALP with SLP shows similar improvement over the baseline as well. Since both configurations of the VLP prioritize attention to the ego-car during training, more improvements are observed within the nearby areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.748 + ], + "angle": 0, + "content": "3D Object Detection. We show the 3D object detection results in Tab. 6 using nuScenes detection metrics of mean Average Precision (mAP), Average Translation Error (ATE), Average Scale Error (ASE), and nuScenes detection score (NDS) as indicated in [6]. As observed in the table, the incorporation of the VLP consistently leads to improvements in NDS. Notably, models incorporating both the SLP and ALP generally perform better than those with SLP alone, providing empirical evidence of the ALP's effectiveness in enhancing the local BEV representation ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Vectorized Scene Segmentation in VAD. The VAD [18] framework represents the driving scene as a fully vectorized structure, categorizing map elements into road boundaries, dividers, and pedestrian crossings. Thus, vectorized scene segmentation experiments are conducted within the VAD framework, with IoU serving as the evaluation metric. As depicted in Tab. 7, the introduction of SLP yields substantial improvements. More specifically, the segmentation for boundaries, dividers, and crossings sees improvements of \\(14.9\\%\\), \\(19.2\\%\\), and \\(23.1\\%\\), respectively, contributing to an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14765" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.089, + 0.767, + 0.288 + ], + "angle": 0, + "content": "
IDModelSLPALPL2 (m) ↓Col. Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
0NMP [36]--2.31---1.92
1SA-NMP [36]--2.05---1.59
2FF [14]0.551.202.541.430.060.171.070.43
3EO [19]0.671.362.781.600.040.090.880.33
4ST-P3 [15]1.332.112.902.110.230.621.270.71
5UniAD [16]0.480.961.651.030.050.170.710.31
6VLP-UniAD✓0.430.861.470.920.030.150.480.22
7VLP-UniAD✓✓0.360.681.190.740.030.120.320.16
8VAD [18]0.460.761.120.780.210.350.580.38
9VLP-VAD✓0.260.470.780.500.120.170.420.23
10VLP-VAD✓✓0.300.530.840.550.010.070.380.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.293, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Table 1. Open-loop planning performance. VLP achieves significant end-to-end planning performance improvement over counterpart vision only UniAD and VAD methods on the nuScenes validation dataset [6]. Based on the planning results, we can conclude that both SLP and ALP components plays a vital role to ensure safe motion planning." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.351, + 0.466, + 0.447 + ], + "angle": 0, + "content": "
ModelSLPALPAMOTA↑AMOTP↓Recall↑IDS↓
ViP3D [10]21.71.62536.3-
QD3DT [13]24.21.51839.9-
MUTR3D [37]29.41.49842.73822
UniAD [16]35.91.32046.7906
VLP-UniAD✓36.61.33246.8820
VLP-UniAD✓✓36.81.31547.3678
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.453, + 0.469, + 0.482 + ], + "angle": 0, + "content": "Table 2. Multi-object tracking. VLP shows improved performance over vision-only MOT techniques." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.489, + 0.466, + 0.551 + ], + "angle": 0, + "content": "
ModelSLPALPLanes↑Drivable↑Divider↑Crossing↑
UniAD [16]31.369.125.713.8
VLP-UniAD✓32.069.527.114.7
VLP-UniAD✓✓32.370.227.415.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.556, + 0.469, + 0.584 + ], + "angle": 0, + "content": "Table 3. Online mapping. VLP demonstrates improved segmentation IoU \\((\\%)\\) performance over the UniAD." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.591, + 0.466, + 0.711 + ], + "angle": 0, + "content": "
ModelSLPALPminADE(m)↓minFDE(m)↓MR↓EPA↑
PnPNet [23]1.151.950.2260.222
ViP3D [10]2.052.840.2460.226
UniAD [16]0.711.020.1510.456
VLP-UniAD✓0.721.040.1540.459
VLP-UniAD✓✓0.680.980.1330.460
VAD [18]0.781.070.1210.598
VLP-VAD✓0.771.030.1100.621
VLP-VAD✓✓0.771.050.1150.621
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.715, + 0.469, + 0.743 + ], + "angle": 0, + "content": "Table 4. Motion forecasting. VLP achieves better motion forecasting over counterpart vision based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.825 + ], + "angle": 0, + "content": "overall increase of \\(18.6\\%\\) in the mean IoU (mIoU). VLP with SLP and ALP also achieves similar improvement over the VAD baseline. This increased accuracy is crucial for developing a safe and efficient self-driving system." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.835, + 0.228, + 0.85 + ], + "angle": 0, + "content": "4.4. Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Autonomous vehicles are intended to operate in diverse environments. Generalization allows a system to be deployed in various urban landscapes, suburban areas, or even rural" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.351, + 0.889, + 0.446 + ], + "angle": 0, + "content": "
ModelSLPALPIoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑
FIERY [12]59.436.750.229.9
StretchBEV [1]55.537.146.029.0
ST-P3 [15]-38.9-32.1
UniAD [16]63.440.254.733.5
VLP-UniAD✓64.240.755.834.5
VLP-UniAD✓✓64.140.255.934.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.452, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Table 5. Occupancy prediction. VLP shows improvement in both near \\((30\\times 30m)\\) and far \\((50\\times 50m)\\) ranges, denoted as \"n.\" and \"f.\", respectively." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.503, + 0.889, + 0.574 + ], + "angle": 0, + "content": "
ModelSLPALPmAP↑mATE↓mASE↓NDS↑
VAD [18]0.270.700.300.389
VLP-VAD✓0.270.670.300.394
VLP-VAD✓✓0.280.670.300.406
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.578, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Table 6. 3D object detection. VLP achieves improved object detection performance over the baseline." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.616, + 0.889, + 0.68 + ], + "angle": 0, + "content": "
ModelSLPALPBoundary↑Divider↑Crossing↑mIoU↑
VAD [18]45.642.231.639.8
VLP-VAD✓52.450.338.947.2
VLP-VAD✓✓49.448.439.245.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.684, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Table 7. Vectorized scene segmentation. VLP shows improved segmentation IoU (%) performance over the VAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.854 + ], + "angle": 0, + "content": "settings, making the technology applicable and accessible on a broader scale. Long-tail cases, which are scenarios that occur infrequently, are often underrepresented in training data. Generalization helps the system cope with these rare but critical situations, reducing the risk of biased decision-making based on inadequate exposure during training. In this section, we evaluate the VLP in terms of generalization ability to new cities and long-tail cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "New-city Generalization. To keep the same number of multi-view camera inputs, we construct a multi-city dataset exclusively from nuScenes, by encompassing data from two" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14766" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.465, + 0.18 + ], + "angle": 0, + "content": "
ModelVLPBostonSingapore
Avg.L2↓Avg.Col↓Avg.L2↓Avg.Col↓
UniAD [16]1.240.321.050.37
VLP-UniAD✓1.140.260.870.34
VAD [18]0.860.270.780.39
VLP-VAD✓0.730.220.630.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.185, + 0.469, + 0.242 + ], + "angle": 0, + "content": "Table 8. New-city generalization. To evaluate the generalization performance, we train the model on Boston and test it on Singapore, and vice versa. VLP shows remarkable zero-shot generalization performance improvement over the vision-only methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.487 + ], + "angle": 0, + "content": "distinct urban environments, Boston and Singapore. To comprehensively assess generalizability, we conduct two sets of experiments: (i) training on Boston and testing on Singapore; and (ii) conversely, training on Singapore and testing on Boston. The results, presented in Tab. 8, reveal a noteworthy reduction in planning L2 error and collision rates for the baselines in both scenarios with the integration of VLP. In particular, VLP-VAD achieves strong generalization over counterpart vision-only VAD approach (in Boston: \\(15.1\\%\\) and \\(18.5\\%\\), in Singapore: \\(19.2\\%\\) and \\(48.7\\%\\) reduction in terms of average L2 error and collision rates, respectively). This signifies VLP's prowess in enhancing the safety and reliability of the ADS as well as its capacity to transcend training confines and excel in diverse real-world conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.685 + ], + "angle": 0, + "content": "Long-tail Generalization. We assess the performance of UniAD in Multi-Object Tracking (MOT) and VAD in 3D detection concerning long-tail scenarios, as detailed in Tab. 9a and Tab. 9b, respectively. Utilizing the long-tail split methodology from [26], derived from per-class object counts, we observe that VLP consistently enhances the generalization capabilities of both frameworks. Specifically, VLP provides substantial improvements, increasing AMOTA and Recall by \\(3.7\\%\\) and \\(7.4\\%\\), respectively, for UniAD. In the case of VAD, VLP improves the mean Average Precision (mAP) by \\(15.9\\%\\). These results highlight the robust feature space cultivated by our proposed VLP, emphasizing its efficacy in handling long-tail scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.241, + 0.71 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.719, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Effectiveness of SLP and ALP. VLP comprises ALP and SLP components. Extensive experiments presented in prior tables show the importance and necessity of both SLP and ALP for enhanced planning and safety." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Prompt Format. The design of the prompt format is pivotal for the success of our VLP, serving as the structured input that guides the learning process. We have performed ablation studies to systematically investigate the impact of different components of the ground truth information included in the prompt. The driving ground truth encompasses labels, bounding boxes, trajectories, and high-level commands. To demonstrate the indispensability of including all aspects of" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.906, + 0.149 + ], + "angle": 0, + "content": "
ModelVLPAMOTA↑AMOTP↓Recall↑IDS↓
UniAD[16]29.61.44639.268
VLP-UniAD✓30.71.43542.166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.591, + 0.152, + 0.82, + 0.164 + ], + "angle": 0, + "content": "(a) Multi-object tracking on long-tail cases." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.172, + 0.905, + 0.236 + ], + "angle": 0, + "content": "
ModelVLPmAP↑mATE↓mASE↓mAOE↓
VAD [18]17.60.790.330.95
VLP-VAD✓20.40.750.330.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.602, + 0.239, + 0.808, + 0.251 + ], + "angle": 0, + "content": "(b) Object detection on long-tail cases." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.26, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Table 9. Long-tail generalization. VLP shows consistent improvement for long-tail cases across both a) multi-object tracking and b) object detection tasks." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.308, + 0.868, + 0.395 + ], + "angle": 0, + "content": "
label-gtbbox-gttraj-gtcommd-gtAvg.L2↓Avg.Col↓
X0.640.23
X0.560.30
X0.590.36
X0.590.26
✓✓✓✓0.520.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Table 10. Ablation for prompt information. All ground-truth information components contributes to improved planning performance. We provide full task ablation results in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.57 + ], + "angle": 0, + "content": "the ground truth in the prompt, we conducted experiments where each element was selectively removed in individual settings. The results, presented in Tab. 10, clearly indicate that the exclusion of any ground truth component leads to a degradation in planning ability. This underscores the necessity of incorporating the entire spectrum of ground truth information in the prompt for optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.619, + 0.602 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We have introduced a novel Vision-Language-Planning (VLP) approach to enhance the capabilities of Autonomous Driving Systems (ADS). Our approach leverages both self-driving-car-centric learning paradigm (SLP) and agent-wise learning paradigm (ALP) guided by language prompts to create a comprehensive understanding of the environment. Through a series of experiments on various driving tasks, we have demonstrated the effectiveness of our VLP approach in improving perception, prediction, and planning aspects of ADS. The generalization experiments showcased the robustness of our VLP approach, proving its adaptability to new cities and long-tail cases. By extending the capabilities of ADS beyond the training environment, our VLP approach paves the way for safer and more reliable autonomous driving in real-world conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Limitations. Our experiments are currently confined to the nuScenes dataset and camera modality as baseline vision-based approaches. We will assess VLP on a broader range of datasets and sensor modalities in our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14767" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.407 + ], + "angle": 0, + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.472, + 0.48 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[7] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.691, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[9] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[10] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Md Zahid Hasan, Jiajing Chen, Jiyang Wang, Ameya Joshi, Senem Velipasalar, Chinmay Hegde, Anuj Sharma, and Soumik Sarkar. Vision-language models can identify distracted driver behavior from naturalistic videos. arXiv preprint arXiv:2306.10159, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[12] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[13] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1992-2008, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[14] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.894, + 0.372 + ], + "angle": 0, + "content": "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.894, + 0.537 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.894, + 0.677 + ], + "angle": 0, + "content": "[19] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. In European Conference on Computer Vision, pages 353-369. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.894, + 0.747 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[21] Liunian Harold Li, Mark Yatskar, D Yin, CJ Hsieh, and KW Chang. Visualbert: A simple and performant baseline for vision and language. arxiv 2019. arXiv preprint arXiv:1908.03557. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.894, + 0.874 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[23] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end per" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "ception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[24] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.339, + 0.195 + ], + "angle": 0, + "content": "[25] OpenAI. Introducing chatgpt, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.198, + 0.469, + 0.239 + ], + "angle": 0, + "content": "[26] Neehar Peri, Achal Dave, Deva Ramanan, and Shu Kong. Towards long-tailed 3d detection. In Conference on Robot Learning, pages 1904–1915. PMLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.469, + 0.423 + ], + "angle": 0, + "content": "[28] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.469, + 0.555 + ], + "angle": 0, + "content": "[30] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[31] Wayve. Lingo-1: Exploring natural language for autonomous driving. https://wayve.ai/thinking/lingo - natural - language - autonomous - driving/, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[32] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. arXiv preprint arXiv:2309.16292, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[33] Shawn Xu, Lin Yang, Christopher Kelly, Marcin Sieniek, Timo Kohlberger, Martin Ma, Wei-Hung Weng, Attila Kiraly, Sahar Kazemzadeh, Zakkai Melamed, et al. Elixir: Towards a general purpose x-ray artificial intelligence system through alignment of large language models and radiology vision encoders. arXiv preprint arXiv:2308.01317, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[34] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[36] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[37] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. Mutr3d: A multi-camera tracking framework via 3d-to-2d queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4537–4546, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14769" + } + ] +] \ No newline at end of file diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_origin.pdf b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccaecb0453a8da3e47109ef68f6711dc33d79d3e --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/16cda83b-9dea-4aab-a622-217c5537ce0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7965216d7d54c0e31205f6355cb4bd4b98a40c4a0d854acbc985d77791747b98 +size 921760 diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/full.md b/2024/VLP_ Vision Language Planning for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa4b4001b469a936b45b4af9016929ddb1e201ad --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/full.md @@ -0,0 +1,304 @@ +# VLP: Vision Language Planning for Autonomous Driving + +Chenbin Pan $^{1*}$ Burhaneddin Yaman $^{2\text{圆}}$ Tommaso Nesti $^{2}$ Abhirup Mallik $^{2}$ Alessandro G Allievi $^{2}$ Senem Velipasalar $^{1}$ Liu Ren $^{2}$ $^{1}$ Syracuse University + +$^{2}$ Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI) + +{cpan14, svelipas}@syr.edu, + +{burhaneddin.yaman,tommaso.nesti,abhirup.mallik,alessandro.allievi,liu. ren}@us.bosch.com + +# Abstract + +Autonomous driving is a complex and challenging task that aims at safe motion planning through scene understanding and reasoning. While vision-only autonomous driving methods have recently achieved notable performance, through enhanced scene understanding, several key issues, including lack of reasoning, low generalization performance and long-tail scenarios, still need to be addressed. In this paper, we present VLP, a novel Vision-Language-Planning framework that exploits language models to bridge the gap between linguistic understanding and autonomous driving. VLP enhances autonomous driving systems by strengthening both the source memory foundation and the self-driving car's contextual understanding. VLP achieves state-of-the-art end-to-end planning performance on the challenging NuScenes dataset by achieving $35.9\%$ and $60.5\%$ reduction in terms of average L2 error and collision rates, respectively, compared to the previous best method. Moreover, VLP shows improved performance in challenging long-tail scenarios and strong generalization capabilities when faced with new urban environments. + +# 1. Introduction + +Autonomous driving is a complex problem requiring scene understanding and reasoning to ensure safe motion planning. This sophisticated challenge can be broadly divided into three main tasks, namely perception, prediction and planning (also known as P3). Conventional methods adopt a modular approach by developing and optimizing each task in a disjoint manner without a holistic view, leading to compounding errors and safety concerns [15, 16, 18]. End-to-end autonomous driving systems (ADS), unifying all P3 tasks, have garnered attention for their potential to enhance + +![](images/6850a411b255ab31d7194d8ce722ca14b816f6fdc78873a73b44366a523b7f43.jpg) +Figure 1. New-city generalization ability of ADS for planning is evaluated by training on Boston city and testing on Singapore, and vice versa. Our proposed VLP shows strong generalization ability by significantly outperforming state-of-the-art vision-only method, VAD[18], in terms of both L2 error and collision rate. + +![](images/3e27faeb36b6d1b5f939040fb2888fe7125b35b4eda94e08cd1f762375b89245.jpg) + +safe planning. Existing vision-based ADS typically follow a two-stage process: bird's eye view (BEV) feature extraction and downstream tasks [16, 18]. BEV feature extraction transforms multi-view camera data into a structured top-view representation embedding spatial information around the ego-car. BEV features are further utilized as the information pool by the downstream P3 tasks. While effective, these vision-only methods struggle with out-of-domain generalization, such as maintaining performance in new cities and long-tail scenarios, hindering real-world deployment. + +The way humans process and interpret visual driving scenes naturally involve a coherent cognitive framework that maintains a consistent logical flow. This enables humans effortlessly make correct decisions even when faced with previously unseen data or scenarios, eliminating generalization issues in driving tasks. Then, this raises a pertinent question: How can ADS achieve human-like driving to navigate diverse and dynamic road environments? Recent advancements in language models (LMs) have led to unprecedented common-sense ability and generalization performance across unseen data and tasks for the natural language processing domain [5, 8, 25]. The superior common- + +sense capability of LMs have led to the emergence of multimodal LMs for diverse applications ranging from medical imaging to robotics [4, 9, 33]. However, incorporation of the reasoning ability of LMs into real-world autonomous driving tasks, to address generalization and long-tail scenarios, is yet to be fully-explored. + +To bridge this gap, we propose a Vision Language Planning (VLP) framework, which integrates the commonsense capability of LMs into vision-based ADS for safe self-driving. Our VLP consists of two key components: Agent-centric Learning Paradigm (ALP) and Self-driving-car-centric Learning Paradigm (SLP), leveraging LMs to enhance the ADS from reasoning and decision-making aspects, respectively. + +The BEV feature map serves as the source memory pool in ADS for downstream decoding tasks. It summarizes and encodes the driving environment surrounding the self-driving car, including vehicles, pedestrians, lanes, and more, into a unified feature map. Hence, capturing comprehensive and necessary details in each local position of BEV is critical for safe and precise self-driving performance. To enhance the local semantic representation and reasoning capabilities of BEV, we introduce an innovative Agent-centric Learning Paradigm (ALP) module. ALP integrates the consistent feature space of a pretrained LM to revamp the agent features on the BEV, actively shaping semantics, and guiding the BEV reasoning process. Leveraging on the common sense and logic flow embedded in the LM, our ALP equips the ADS with robustness and consistent BEV feature space, enhancing its effectiveness in diverse driving scenarios. + +In ADS, the planning module aggregates information from the preceding perception and prediction phases to make the final decisions for self-driving. This global perspective culminates in the formation of a planning query, directly influencing the safety and accuracy of the self-driving navigation. Considering the critical role of the planning module within ADS, we also present a novel Self-driving-car-centric Learning Paradigm (SLP) to elevate the decoding and acquiring information ability of the planning query. In the SLP, we align the planning query with intended goals and the ego-vehicle driving status by leveraging the knowledge encoded in the pretrained LM. The language model's comprehension capabilities contribute to more informed decision-making during the planning phase as well as enabling a more robust planning query formation process. + +Through VLP, we bridge the gap between human-like reasoning and autonomous driving, enhancing the model's contextual awareness and its ability to generalize effectively in complex, ever-changing real-world scenarios, as shown in Fig.1. The main contributions of this work are summarized as follows: + +- We propose VLP, a Vision Language Planning model, + +which incorporates reasoning capability of LMs into vision-based autonomous driving systems as an enhancement of motion planning and self-driving safety. + +- VLP is composed of novel components ALP and SLP, aiming to improve the ADS from self-driving BEV reasoning and self-driving decision-making aspects, respectively. +- Through extensive experiments in real-world driving scenarios, we show that VLP significantly and consistently outperforms the state-of-the-art vision-based approaches across a spectrum of driving tasks, including open-loop planning, multi-object tracking, motion forecasting, and more. +- We conduct the first new-city generalization study on the nuScenes dataset [6] by training and testing on distinct cities, demonstrating the remarkable zero-shot generalization ability of our VLP approach over vision-only methods. +- To the best of our knowledge, this is the first work introducing LM into multiple stages of ADS to address the generalization ability in new cities and long-tail cases. + +# 2. Related Work + +# 2.1. End-to-End Autonomous Driving + +Early end-to-end approaches have adopted a single neural network to perform motion planning without explicitly designing intermediate tasks, such as prediction [3]. Thus, such approaches not only suffer from the sub-optimal performance but also lack interpretability. Modular end-to-end frameworks introduce intermediate tasks for interpretability and provide improved performance [7, 10, 15, 16, 18, 23, 28]. These interpretable end-to-end frameworks explicitly model perception, prediction and planning (P3) components and train them together with a joint optimization strategy to enhance planning. UniAD [16] and VAD [18] are two prominent state-of-the-art end-to-end frameworks. UniAD leverages rasterized scene representations and explicitly identifies crucial components within the P3 framework. It coordinates all tasks through a unified query design for safe planning. On the other hand, VAD uses vectorized representations for efficient planning and safety improvement. While these approaches generally achieve good performance, their generalization performance have been limited due to lack of common sense and reasoning process. + +# 2.2. Vision-Language Models + +Large language models (LLMs), which are trained on massive amounts of text data, have shown unprecedented success in language-related tasks [5, 8, 29]. The capability of LLMs for common-sense understanding and versatility in tackling diverse language tasks has driven the exploration of multi-modal LLMs [20, 21, 27]. Vision-language models (VLM) integrate visual and textual information to en + +able common sense understanding of the content by training on large scale open-world vision and text data [2, 17, 20, 27]. CLIP [27] uses a contrastive learning objective to learn a joint embedding space for vision and text data, and has shown immense zero-shot generalization performance, showcasing the importance of VLMs compared to vision-only approaches. + +# 2.3. Autonomous Driving with Language Models + +In recent years, numerous works have been proposed for extending LMS to the embodied AI domain to improve zero-shot generalization performance [4, 9, 30, 35]. These works incorporate sensory inputs from embodied agents in the form of language with vision and language data for decision making process [4, 9]. While these methods have mainly focused on the robotics domain, few works have been proposed for leveraging embodied LMs for autonomous driving tasks [11, 24, 31, 32, 34]. In particular, DiLu [32] and GPT-Driver [24] propose GPT-based driver agents for closed-loop simulation tasks. In [31], an open loop driving commentator, which combines vision and low-level driving actions with language, is proposed to interpret driving actions and for reasoning. Our work significantly differs from these approaches by leveraging LMs for end-to-end motion planning task for real-world driving applications. + +# 3. Methodology + +We present a Vision Language Planning (VLP) model, which equips the ADS with the capacity to emulate human common sense and engage in contextual reasoning for safe motion planning. Our proposed VLP model, illustrated in Fig. 2a, comprises two innovative components that leverage LMs in both local and global contexts. The Agent-centric Learning Paradigm (ALP) concentrates on refining local details to enhance source memory reasoning, while the Self-driving-car-centric Learning Paradigm (SLP) focuses on guiding the planning process for the self-driving-car (SDC). Below, we first provide an overview of how general ADS work in Sec.3.1. Subsequently, we describe our proposed ALP and SLP in detail in Sec. 3.2 and Sec. 3.3, respectively. + +# 3.1. Preliminary + +In vision-based ADS, a sequence of multi-view camera images serves as the source input, offering abundant visual data for downstream tasks. Initially, an image backbone extracts fundamental features for perception. These features are then processed by a BEV encoder to transform multi-view features into a unified 2D representation from a top-down view. This BEV feature map encapsulates crucial perception details like object positions, lane markings, and road boundaries [22]. Acting as an information hub, this BEV feature map supports diverse downstream tasks such + +as 3D object tracking, mapping, motion prediction, occupancy prediction, and path planning [16, 18]. Specific task-related information is decoded using a query-based transformer decoder designed for each task to efficiently access the source memory [16]. In the final planning stage, an ego-query models the ego-vehicle's status and interacts with environmental features to determine the optimal path for autonomous driving. + +# 3.2. Agent-centric Learning Paradigm: Enhancing BEV Source Memory + +The BEV feature map serves as the foundational source memory pool and holds a pivotal role in ADS. Ensuring that a BEV map provides comprehensive information with intricate details for driving is essential for enabling downstream decoders to make safe, precise, and human-like decisions. However, the BEV in ADS is derived from multiview camera images rather than a true bird's-eye-view image, which can introduce discrepancies between the produced BEV and the expected BEV representation. Therefore, to mitigate the discrepancies, we introduce the Agent-centric Learning Paradigm (ALP) to align the produced BEV with a true bird's-eye-view map. Through direct supervision on the BEV map with LM, our ALP enables the refinement of local details and alignment with the desired human perspective. + +Agent BEV features. In our proposed ALP, three kinds of BEV agents are considered: ego-vehicle (self-driving-car), foreground (FG) objects, and lane elements. We first align the ground-truth area of each agent with the produced BEV map, and crop the regions of interest. We utilize the 3D bounding box to crop the ego-car and FG object area, and panoptic scene mask to segment the lane area. Subsequently, we perform a pooling operation on the obtained local BEV region, to generate a single feature representation for the corresponding agent. After pooling, the local agent features in each sample along the batch are concatenated to formulate an Agent-BEV tensor denoted as $\mathcal{A}_{bev} \in \mathbb{R}^{N_B \times C}$ , where $N_B$ and $C$ denote total number of agents in the batch and feature dimension, respectively. The process can be formulated as: + +$$ +\begin{array}{l} \mathcal {A} _ {b e v} ^ {E g o} = \operatorname {P o o l} (\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {E g o})) \text {,} \in \mathbb {R} ^ {C}, \\ \mathcal {A} _ {b e v} ^ {F G} = \operatorname {P o o l} (\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {F G})) \in \mathbb {R} ^ {C}, \\ \mathcal {A} _ {b e v} ^ {\text {L a n e}} = \operatorname {P o o l} \left(\operatorname {S e g} \left(B E V, \operatorname {S e g M a s k} ^ {\text {L a n e}}\right)\right), \in \mathbb {R} ^ {C}, \tag {1} \\ \mathcal {A} _ {b e v} = B a t c h \left(\left[ \mathcal {A} _ {b e v} ^ {E g o}; \mathcal {A} _ {b e v} ^ {F G}; \mathcal {A} _ {b e v} ^ {L a n e} \right]\right), \in \mathbb {R} ^ {N _ {b a t c h} \times C}, \\ \end{array} +$$ + +where $Batch[;]$ denotes the concatenation operation, $\mathcal{A}_{bev}^{(\cdot)}$ represents the single agent feature of the corresponding type, $bbox_{3D}^{(\cdot)}$ and $SegMask^{Lane}$ indicate the ground-truth 3D bounding box and the lane segmentation mask, respectively. + +Agent expectation features. To ensure that local BEV features express the desired information, we conduct a BEV- + +a) +b) +Figure 2. a) The overview of proposed vision language planning (VLP) framework. VLP enhances ADS from self-driving BEV-reasoning and self-driving decision-making aspects, through two innovative modules, ALP and SLP, respectively. Leveraging LM and contrastive learning, ALP conducts agent-wise learning for refining local details on BEV, while SLP engages sample-wise learning for advancing global context understanding ability of the ADS. VLP is only active during training, ensuring no additional parameters or computations are introduced during inference. b) Prompt formats used in VLP. +![](images/14e20e0a0aab98af833b8caf17461440292d8bddcff7879081c620ab73773bf0.jpg) +Ego-car prompt: This is the self-driving car. It is {class}, and its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. It is currently {command}. Its future trajectory will be {{x1, y1}...{xn, yn}]. Foreground prompt: This object is {class}. Its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. Its future trajectory will be {{x1, y1}...{xt, yt}]. Lane prompt: This lane belongs to a {lane-label}. The label belongs to {thing/stuff} class. Its 2D bounding box on the BEV map is {cx, cy, w, h}. Planning prompt: The self-driving car is driving in an urban area. It is currently {command}. The future trajectory of the car for the next P timestamps will be {{[x1, y1]...{xp, yp}]} + +expectation alignment process by leveraging LM and contrastive learning. We precisely define the perceptual information expected from the corresponding agent, such as agent labels, bounding boxes, and future trajectories. These driving-related ground-truth information, which should also be embedded in the local BEV feature is formulated into a prompt as illustrated in Fig. 2b. The description is then passed to the language encoder, $LM$ , to generate the corresponding agent expectation feature. We apply an MLP layer $\mathcal{F}_{bev}$ to adapt the expectation feature to the BEV feature space. Then, the agent expectation features are concatenated along the batch to generate an Agent-Expectation tensor denoted as $\mathcal{A}_{exp} \in \mathbb{R}^{N_B \times C}$ . The procedure can be formulated as in Eq. (2): + +$$ +\begin{array}{l} \mathcal {A} _ {e x p} ^ {E g o} = \mathcal {F} _ {b e v} (L M (\boldsymbol {T} ^ {E g o} [ y ^ {E g o} ])), \in \mathbb {R} ^ {C}, \\ \mathcal {A} _ {e x p} ^ {F G} = \mathcal {F} _ {b e v} (L M (\boldsymbol {T} ^ {F G} [ y ^ {F G} ])), \in \mathbb {R} ^ {C}, \\ \mathcal {A} _ {e x p} ^ {L a n e} = \mathcal {F} _ {b e v} \left(L M \left(\boldsymbol {T} ^ {L a n e} \left[ y ^ {L a n e} \right]\right)\right), \in \mathbb {R} ^ {C}, \tag {2} \\ \mathcal {A} _ {e x p} = B a t c h \left(\left[ \mathcal {A} _ {e x p} ^ {e g o}; \mathcal {A} _ {e x p} ^ {F G}; \mathcal {A} _ {e x p} ^ {l a n e} \right]\right), \\ \end{array} +$$ + +where $\mathcal{A}_{exp}^{(\cdot)}, T^{(\cdot)}$ , and $y^{(\cdot)}$ represent the single agent expectation feature, the description template, and the ground-truth in ADS for the respective agent. Note that during training, we freeze the $LM$ and only train the adaptation layer $\mathcal{F}_{bev}$ to save memory and retain the pre-trained knowledge in the $LM$ . + +Contrastive Learning Formulation in ALP. Given the agent BEV feature $\mathcal{A}_{bev}$ and the agent expectation feature $\mathcal{A}_{exp}$ , the alignment between the produced BEV and the expected BEV is facilitated through a contrastive learning process [27]. In particular, L2 normalization is applied to both $\mathcal{A}_{bev}$ and $\mathcal{A}_{exp}$ to standardize the feature vectors and ensure training stability. Then, matrix multiplication is performed between the normalized two-modality tensors with + +a learnable logit scale $\alpha_{alp}$ to produce a similarity matrix $S_{pred} \in \mathbb{R}^{N_B \times N_B}$ in agent-wise. The ground truth for the similarity matrix, $S_{gt}$ , is a unit matrix, signifying that each agent's BEV feature should be closest to its corresponding expected feature in the shared space. To optimize this alignment, we apply a cross-entropy loss, $\mathcal{L}_{CE}$ , along both the Agent-BEV mode axis and Agent-Expectation mode axis of the $S_{pred}$ , and compute an average value as the final ALP loss, $\mathcal{L}_{alp}$ , for the BEV source reasoning part. The entire process can be formally expressed as: + +$$ +\begin{array}{l} \begin{array}{l} \mathcal {S} _ {p r e d} = \alpha_ {a l p} \times \left(\frac {\mathcal {A} _ {b e v}}{\| \mathcal {A} _ {b e v} \| _ {2}} \otimes \frac {\mathcal {A} _ {e x p}}{\| \mathcal {A} _ {e x p} \| _ {2}}\right), \\ \mathcal {L} _ {a l p} = \left(\mathcal {L} _ {C E} \left(\mathcal {S} _ {p r e d}, \mathcal {S} _ {q t}, d i m = 0\right) \right. \end{array} \tag {3} \\ + \mathcal {L} _ {C E} \left(\mathcal {S} _ {p r e d}, \mathcal {S} _ {g t}, d i m = 1\right)) / 2. \\ \end{array} +$$ + +# 3.3. SDC-centric Learning Paradigm: Enhancing Ego-vehicle Query + +Ego-vehicle query feature. The ego-vehicle dynamic modeling is the core function of the ADS. In previous ADS, a trainable ego-query $\mathcal{E}_{query}$ is applied to interact with other agents, $\mathcal{A}_{query}^{FG}$ and $\mathcal{A}_{query}^{Lane}$ , on BEV map, to gather the self-driving perception/prediction information. The produced ego-car query feature $\mathcal{E}_{qfeat}$ is further processed to predict the future waypoints of the self-driving car $y_{pred}^{plan}$ . The process is formulated as in Eq. (4): + +$$ +\begin{array}{l} \mathcal {E} _ {q f e a t} = \mathcal {M} ^ {i n t e r} \left(\mathcal {E} _ {q u e r y}, \mathcal {A} _ {q u e r y} ^ {F G}, \mathcal {A} _ {q u e r y} ^ {l a n e}\right) \in \mathbb {R} ^ {B \times C}, \tag {4} \\ y _ {p r e d} ^ {\text {p l a n}} = \mathcal {M} ^ {\text {p l a n}} \left(\mathcal {E} _ {q f e a t}\right) \in \mathbb {R} ^ {B \times P \times 2}, \\ \end{array} +$$ + +where $\mathcal{M}^{inter}$ and $\mathcal{M}^{plan}$ denote the intermediate ego information gathering module and the planning head in ADS, respectively. $P \times 2$ indicate the planned waypoints in the next P timesteps, and $B$ is the batch size during training. + +Although such mechanism can achieve good planning results, solely relying on numeric ground-truth can make it challenging to understand the rationale behind the ADS' decisions, which can result in inconsistent feature learning for the planning and lead to limited adaptability to new environments. To address these limitations, we propose a SDCcentric Learning Paradigm (SLP) to equip the ADS with the capability of making decisions from continuous and robust feature space. + +Ego-vehicle planning feature. We template a language description of the ego-vehicle status with the planning ground truth (GT) including high-level driving command and the future trajectory of the ego-vehicle, as illustrated in Fig. 2b. The prompt is sent to the $LM$ to obtain the ground truth planning features of the ego-car embedding driving scenario information and human driving logic. An MLP layer $\mathcal{F}_{ego}$ is applied to adapt the textual planning features to the ego-query feature space. The process can be expressed as in Eq. (5): + +$$ +\mathcal {E} _ {p r o m p t} = \mathcal {F} _ {e g o} (L M (\boldsymbol {T} _ {s l p} ^ {E g o} [ y ^ {p l a n} ])) \in \mathbb {R} ^ {B \times C}, \quad (5) +$$ + +where $T_{slp}^{Ego}$ indicates the prompt for ego-car description used in SLP, and $y^{plan}$ denotes the planning ground truth. Note that the $LM$ is a shared off-the-shelf model for both ALP and SLP. As in ALP, only the adaptation layer $\mathcal{F}_{ego}$ is trainable during training to save memory and retain the pre-trained knowledge in $LM$ . + +Contrastive Learning Formulation in SLP. Similar to ALP, we employ a sample-wise contrastive learning approach for SLP to align the ego-vehicle query feature with the ego-vehicle textual planning feature as follows: + +$$ +\begin{array}{l} \mathcal {S} _ {p r e d} ^ {s l p} = \alpha_ {s l p} \times (\frac {\mathcal {E} _ {q f e a t}}{\| \mathcal {E} _ {q f e a t} \| _ {2}} \otimes \frac {\mathcal {E} _ {p r o m p t}}{\| \mathcal {E} _ {p r o m p t} \| _ {2}}), \\ \mathcal {L} _ {s l p} = \left(\mathcal {L} _ {C E} \left(S _ {p r e d} ^ {s l p}, S _ {g t} ^ {s l p}, d i m = 0\right) \right. \tag {6} \\ + \mathcal {L} _ {C E} \left(\mathcal {S} _ {p r e d} ^ {s l p}, \mathcal {S} _ {g t} ^ {s l p}, d i m = 1\right)) / 2, \\ \end{array} +$$ + +where $\alpha_{slp}$ is a learnable logit scale, and $S_{pred}^{slp}$ and $S_{gt}^{slp}$ are the predicted and ground truth similarity matrices, respectively. Aligning the two modes closely in the feature space, the contrastive process refines the produced ego-vehicle query using the ground truth-embedded linguistic feature. The sample-wise contrastive loss rectifies the relationships between individual samples from a human (linguistic) perspective, establishing a connection for the current data point with the world-wise common sense embedded in the pretrained language model. + +# 3.4. Training Loss + +The overall loss, $\mathcal{L}_{vlp}$ , of our VLP training is composed of two parts: the BEV encoder reasoning loss $\mathcal{L}_{enc}$ and the decoder decision-making loss $\mathcal{L}_{dec}$ such that: + +$$ +\mathcal {L} _ {v l p} = \omega_ {e n c} \mathcal {L} _ {e n c} + \omega_ {d e c} \mathcal {L} _ {d e c}, \tag {7} +$$ + +where $\omega_{enc}$ and $\omega_{dec}$ represent the weights for $\mathcal{L}_{enc}$ and $\mathcal{L}_{dec}$ , respectively. + +The encoder reasoning loss, $\mathcal{L}_{enc}$ , is equal to the loss produced in ALP module, i.e. $\mathcal{L}_{enc} = \mathcal{L}_{alp}$ . The decision-making loss includes the losses in all downstream tasks (perception/prediction/planning) in ADS and the SLP loss: + +$$ +\mathcal {L} _ {d e c} = \mathcal {L} _ {p e r c} + \mathcal {L} _ {p r e d} + \mathcal {L} _ {p l a n} + \mathcal {L} _ {s l p}. \tag {8} +$$ + +Note that the off-the-shelf LM head is discarded during inference, and thus, our method introduces no additional parameters and computations to the original ADS. + +# 4. Experiments + +# 4.1. Implementation Details + +Dataset. We conduct experiments on the challenging nuScenes dataset [6], which is the first large-scale public dataset to provide data from the entire sensor suite of an autonomous vehicle (6 cameras, 1 LIDAR, 5 RADAR, GPS and IMU sensors). The nuScenes contains 1000 driving scenes from Boston and Singapore, two cities that are known for their dense traffic and highly challenging driving conditions. In our experiments, we utilize 6 camera images as our vision inputs. + +Baselines. We adopt two recent SOTA models in autonomous driving, namely UniAD [16] and VAD [18], as our baselines to evaluate the effectiveness of our approach. + +Training. We incorporate our proposed VLP training strategy into vision-only approaches of UniAD [16] and VAD [18], and refer to this enriched versions as VLP-UniAD and VLP-VAD, respectively, for clarity when reporting the results. All methods are trained with the same hyper-parameters reported in the respective baselines [16, 18] for commensurate comparison. For language model, we use the pretrained LM in CLIP [27]. In VLP study, we set $T = 6$ (3 seconds) for planning, which is same as baselines [16] [18]. Experiments are conducted with 8 NVIDIA Tesla A100 GPUs. More details regarding experimental setup can be found in supplementary materials. + +# 4.2. Open-loop Planning + +Planning constitutes the cornerstone of any ADS, playing a pivotal role in ensuring safety and devising efficient routes for the ego-car. In Tab. 1, we present a series of comparative experiments that showcase the performance of our open-loop planning in comparison to the baseline models. As can be seen in rows 4-6 of the table, the integration of just SLP leads to noticeable reductions in both the L2 error and collision rates for all the baseline models. Moving down the table, rows 7 to 9 demonstrate that the inclusion + +of both VLP components (SLP and ALP together) consistently yields further improvements in these planning metrics. In particular, VLP-UniAD shows a $28.1\%$ and $48.4\%$ reduction in terms of average L2 error and collision rate, respectively, compared to baseline UniAD. Similarly, in comparison with VAD, VLP-VAD achieves $35.9\%$ and $60.5\%$ reduction for average L2 error and collision rate, respectively. These significant results underscore the effectiveness of both SLP and ALP, as well as their adaptability across various ADS configurations. The reduced L2 error and collision rate achieved through VLP integration contribute to safer road planning in the realm of self-driving. + +# 4.3. Perception and Prediction + +In this section, we showcase the consistent effectiveness of our proposed VLP across a spectrum of essential driving tasks. Our proposed approach excels in various perception and prediction tasks, including multi-object tracking, mapping, motion forecasting, occupancy prediction, 3D object detection, and vectorized scene segmentation. These results underscore the versatility and reliability of our VLP framework across a wide range of critical driving tasks. + +Multi-Object Tracking. The results for multi-object tracking (MOT) are presented in Tab. 2. We apply the standard evaluation protocols of AMOTA (Average Multi-Object Tracking Accuracy), AMOTP (Average Multi-Object Tracking Precision), Recall, and IDS (Identity Switches) to evaluate the MOT performance. As can be seen, with the incorporation of SLP, AMOTA, Recall, and IDS values all surpass those of UniAD, while AMOTP experiences a slight degradation. The combined integration of both SLP and ALP results in improvements across all metrics when compared to the UniAD baseline. The enhanced performance in MOT proves that our VLP can help the system better predict and respond to the movements of other objects on the road, reducing the risk of accidents. + +Online Mapping. In Tab. 3, we present the results of online mapping, encompassing four key mapping elements: lanes, drivable areas, dividers, and pedestrian crossings. The evaluation employs the Intersection-over-Union (IoU) metric to measure the overlap between the predicted and ground-truth maps. The results reveal that the inclusion of the SLP leads to improvements in mapping accuracy for all four elements. Specifically, it enhances the IoU for drivable areas and crossings by $5.5\%$ and $6.5\%$ , respectively. Furthermore, the integration of both SLP and ALP yields even further improvement, with IoU achieving by $6.6\%$ and $10.2\%$ for drivable areas and crossings, respectively. These enhancements represent substantial gains compared to the SLP-only configuration. These results underscore the valuable contributions of each paradigm in online mapping. They also demonstrate the effectiveness of VLP in bridging the gap between vision-based and language-based information, ul + +timely enhancing the system's comprehension of various road elements. + +Motion Forecasting. We present the motion forecasting results in Tab. 4, applying the same evaluation metrics as those in UniAD [16], namely minADE (minimum Average Displacement Error), minFDE (minimum Final Displacement Error), MR (Miss Rate), and EPA (End-to-end Prediction Accuracy). The results demonstrate that the inclusion of the SLP contributes to a reduction in minADE, minFDE, and MR, while simultaneously increasing EPA for the ADS. This outcome signifies that the supervision and refinement applied to the final ego-car feature also has a positive influence on the motion prediction of other vehicles around the ego-car. The improvements are more obvious with the EPA metrics. Equipped with full components of VLP, the performance of the ADS, in motion prediction, experiences further enhancements. This observation underscores the capability of ALP to empower the system in the identification and accurate prediction of the movements of various agents on the road. + +Occupancy Prediction. We present the occupancy prediction results in Tab. 5, which have been obtained using the IoU and Video Panoptic Quality (VPQ) within two distance ranges around the ego-car, namely near ("-n.", covering a $30 \times 30m$ area, and far ("-f.", spanning a $50 \times 50m$ area. The results highlight the effectiveness of the SLP in consistently improving all four metrics. SLP, primarily focusing on enhancing the ego-car feature, demonstrates consistent performance gains in occupancy prediction. The inclusion of ALP with SLP shows similar improvement over the baseline as well. Since both configurations of the VLP prioritize attention to the ego-car during training, more improvements are observed within the nearby areas. + +3D Object Detection. We show the 3D object detection results in Tab. 6 using nuScenes detection metrics of mean Average Precision (mAP), Average Translation Error (ATE), Average Scale Error (ASE), and nuScenes detection score (NDS) as indicated in [6]. As observed in the table, the incorporation of the VLP consistently leads to improvements in NDS. Notably, models incorporating both the SLP and ALP generally perform better than those with SLP alone, providing empirical evidence of the ALP's effectiveness in enhancing the local BEV representation ability. + +Vectorized Scene Segmentation in VAD. The VAD [18] framework represents the driving scene as a fully vectorized structure, categorizing map elements into road boundaries, dividers, and pedestrian crossings. Thus, vectorized scene segmentation experiments are conducted within the VAD framework, with IoU serving as the evaluation metric. As depicted in Tab. 7, the introduction of SLP yields substantial improvements. More specifically, the segmentation for boundaries, dividers, and crossings sees improvements of $14.9\%$ , $19.2\%$ , and $23.1\%$ , respectively, contributing to an + +
IDModelSLPALPL2 (m) ↓Col. Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
0NMP [36]--2.31---1.92
1SA-NMP [36]--2.05---1.59
2FF [14]0.551.202.541.430.060.171.070.43
3EO [19]0.671.362.781.600.040.090.880.33
4ST-P3 [15]1.332.112.902.110.230.621.270.71
5UniAD [16]0.480.961.651.030.050.170.710.31
6VLP-UniAD✓0.430.861.470.920.030.150.480.22
7VLP-UniAD✓✓0.360.681.190.740.030.120.320.16
8VAD [18]0.460.761.120.780.210.350.580.38
9VLP-VAD✓0.260.470.780.500.120.170.420.23
10VLP-VAD✓✓0.300.530.840.550.010.070.380.15
+ +Table 1. Open-loop planning performance. VLP achieves significant end-to-end planning performance improvement over counterpart vision only UniAD and VAD methods on the nuScenes validation dataset [6]. Based on the planning results, we can conclude that both SLP and ALP components plays a vital role to ensure safe motion planning. + +
ModelSLPALPAMOTA↑AMOTP↓Recall↑IDS↓
ViP3D [10]21.71.62536.3-
QD3DT [13]24.21.51839.9-
MUTR3D [37]29.41.49842.73822
UniAD [16]35.91.32046.7906
VLP-UniAD✓36.61.33246.8820
VLP-UniAD✓✓36.81.31547.3678
+ +Table 2. Multi-object tracking. VLP shows improved performance over vision-only MOT techniques. + +
ModelSLPALPLanes↑Drivable↑Divider↑Crossing↑
UniAD [16]31.369.125.713.8
VLP-UniAD✓32.069.527.114.7
VLP-UniAD✓✓32.370.227.415.2
+ +Table 3. Online mapping. VLP demonstrates improved segmentation IoU $(\%)$ performance over the UniAD. + +
ModelSLPALPminADE(m)↓minFDE(m)↓MR↓EPA↑
PnPNet [23]1.151.950.2260.222
ViP3D [10]2.052.840.2460.226
UniAD [16]0.711.020.1510.456
VLP-UniAD✓0.721.040.1540.459
VLP-UniAD✓✓0.680.980.1330.460
VAD [18]0.781.070.1210.598
VLP-VAD✓0.771.030.1100.621
VLP-VAD✓✓0.771.050.1150.621
+ +overall increase of $18.6\%$ in the mean IoU (mIoU). VLP with SLP and ALP also achieves similar improvement over the VAD baseline. This increased accuracy is crucial for developing a safe and efficient self-driving system. + +# 4.4. Generalization + +Autonomous vehicles are intended to operate in diverse environments. Generalization allows a system to be deployed in various urban landscapes, suburban areas, or even rural + +Table 4. Motion forecasting. VLP achieves better motion forecasting over counterpart vision based methods. + +
ModelSLPALPIoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑
FIERY [12]59.436.750.229.9
StretchBEV [1]55.537.146.029.0
ST-P3 [15]-38.9-32.1
UniAD [16]63.440.254.733.5
VLP-UniAD✓64.240.755.834.5
VLP-UniAD✓✓64.140.255.934.1
+ +Table 5. Occupancy prediction. VLP shows improvement in both near $(30\times 30m)$ and far $(50\times 50m)$ ranges, denoted as "n." and "f.", respectively. + +
ModelSLPALPmAP↑mATE↓mASE↓NDS↑
VAD [18]0.270.700.300.389
VLP-VAD✓0.270.670.300.394
VLP-VAD✓✓0.280.670.300.406
+ +Table 6. 3D object detection. VLP achieves improved object detection performance over the baseline. + +
ModelSLPALPBoundary↑Divider↑Crossing↑mIoU↑
VAD [18]45.642.231.639.8
VLP-VAD✓52.450.338.947.2
VLP-VAD✓✓49.448.439.245.7
+ +Table 7. Vectorized scene segmentation. VLP shows improved segmentation IoU (%) performance over the VAD. + +settings, making the technology applicable and accessible on a broader scale. Long-tail cases, which are scenarios that occur infrequently, are often underrepresented in training data. Generalization helps the system cope with these rare but critical situations, reducing the risk of biased decision-making based on inadequate exposure during training. In this section, we evaluate the VLP in terms of generalization ability to new cities and long-tail cases. + +New-city Generalization. To keep the same number of multi-view camera inputs, we construct a multi-city dataset exclusively from nuScenes, by encompassing data from two + +
ModelVLPBostonSingapore
Avg.L2↓Avg.Col↓Avg.L2↓Avg.Col↓
UniAD [16]1.240.321.050.37
VLP-UniAD✓1.140.260.870.34
VAD [18]0.860.270.780.39
VLP-VAD✓0.730.220.630.20
+ +distinct urban environments, Boston and Singapore. To comprehensively assess generalizability, we conduct two sets of experiments: (i) training on Boston and testing on Singapore; and (ii) conversely, training on Singapore and testing on Boston. The results, presented in Tab. 8, reveal a noteworthy reduction in planning L2 error and collision rates for the baselines in both scenarios with the integration of VLP. In particular, VLP-VAD achieves strong generalization over counterpart vision-only VAD approach (in Boston: $15.1\%$ and $18.5\%$ , in Singapore: $19.2\%$ and $48.7\%$ reduction in terms of average L2 error and collision rates, respectively). This signifies VLP's prowess in enhancing the safety and reliability of the ADS as well as its capacity to transcend training confines and excel in diverse real-world conditions. + +Long-tail Generalization. We assess the performance of UniAD in Multi-Object Tracking (MOT) and VAD in 3D detection concerning long-tail scenarios, as detailed in Tab. 9a and Tab. 9b, respectively. Utilizing the long-tail split methodology from [26], derived from per-class object counts, we observe that VLP consistently enhances the generalization capabilities of both frameworks. Specifically, VLP provides substantial improvements, increasing AMOTA and Recall by $3.7\%$ and $7.4\%$ , respectively, for UniAD. In the case of VAD, VLP improves the mean Average Precision (mAP) by $15.9\%$ . These results highlight the robust feature space cultivated by our proposed VLP, emphasizing its efficacy in handling long-tail scenarios. + +# 4.5. Ablation Studies + +Effectiveness of SLP and ALP. VLP comprises ALP and SLP components. Extensive experiments presented in prior tables show the importance and necessity of both SLP and ALP for enhanced planning and safety. + +Prompt Format. The design of the prompt format is pivotal for the success of our VLP, serving as the structured input that guides the learning process. We have performed ablation studies to systematically investigate the impact of different components of the ground truth information included in the prompt. The driving ground truth encompasses labels, bounding boxes, trajectories, and high-level commands. To demonstrate the indispensability of including all aspects of + +Table 8. New-city generalization. To evaluate the generalization performance, we train the model on Boston and test it on Singapore, and vice versa. VLP shows remarkable zero-shot generalization performance improvement over the vision-only methods. + +
ModelVLPAMOTA↑AMOTP↓Recall↑IDS↓
UniAD[16]29.61.44639.268
VLP-UniAD✓30.71.43542.166
+ +(a) Multi-object tracking on long-tail cases. + +
ModelVLPmAP↑mATE↓mASE↓mAOE↓
VAD [18]17.60.790.330.95
VLP-VAD✓20.40.750.330.83
+ +(b) Object detection on long-tail cases. +Table 9. Long-tail generalization. VLP shows consistent improvement for long-tail cases across both a) multi-object tracking and b) object detection tasks. + +
label-gtbbox-gttraj-gtcommd-gtAvg.L2↓Avg.Col↓
X0.640.23
X0.560.30
X0.590.36
X0.590.26
✓✓✓✓0.520.17
+ +Table 10. Ablation for prompt information. All ground-truth information components contributes to improved planning performance. We provide full task ablation results in the supplementary. + +the ground truth in the prompt, we conducted experiments where each element was selectively removed in individual settings. The results, presented in Tab. 10, clearly indicate that the exclusion of any ground truth component leads to a degradation in planning ability. This underscores the necessity of incorporating the entire spectrum of ground truth information in the prompt for optimal performance. + +# 5. Conclusion + +We have introduced a novel Vision-Language-Planning (VLP) approach to enhance the capabilities of Autonomous Driving Systems (ADS). Our approach leverages both self-driving-car-centric learning paradigm (SLP) and agent-wise learning paradigm (ALP) guided by language prompts to create a comprehensive understanding of the environment. Through a series of experiments on various driving tasks, we have demonstrated the effectiveness of our VLP approach in improving perception, prediction, and planning aspects of ADS. The generalization experiments showcased the robustness of our VLP approach, proving its adaptability to new cities and long-tail cases. By extending the capabilities of ADS beyond the training environment, our VLP approach paves the way for safer and more reliable autonomous driving in real-world conditions. + +Limitations. Our experiments are currently confined to the nuScenes dataset and camera modality as baseline vision-based approaches. We will assess VLP on a broader range of datasets and sensor modalities in our future work. + +# References + +[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 7 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3 +[3] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 2 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2, 3 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6, 7 +[7] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2 +[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2 +[9] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3 +[10] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 2, 7 +[11] Md Zahid Hasan, Jiajing Chen, Jiyang Wang, Ameya Joshi, Senem Velipasalar, Chinmay Hegde, Anuj Sharma, and Soumik Sarkar. Vision-language models can identify distracted driver behavior from naturalistic videos. arXiv preprint arXiv:2306.10159, 2023. 3 + +[12] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 7 +[13] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1992-2008, 2022. 7 +[14] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 7 +[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 7 +[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3 +[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1, 2, 3, 5, 6, 7, 8 +[19] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. In European Conference on Computer Vision, pages 353-369. Springer, 2022. 7 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3 +[21] Liunian Harold Li, Mark Yatskar, D Yin, CJ Hsieh, and KW Chang. Visualbert: A simple and performant baseline for vision and language. arxiv 2019. arXiv preprint arXiv:1908.03557. 2 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 3 +[23] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end per + +ception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 2, 7 +[24] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3 +[25] OpenAI. Introducing chatgpt, 2023. 1 +[26] Neehar Peri, Achal Dave, Deva Ramanan, and Shu Kong. Towards long-tailed 3d detection. In Conference on Robot Learning, pages 1904–1915. PMLR, 2023. 8 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5 +[28] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 2 +[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[30] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 3 +[31] Wayve. Lingo-1: Exploring natural language for autonomous driving. https://wayve.ai/thinking/lingo - natural - language - autonomous - driving/, 2023.3 +[32] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. arXiv preprint arXiv:2309.16292, 2023. 3 +[33] Shawn Xu, Lin Yang, Christopher Kelly, Marcin Sieniek, Timo Kohlberger, Martin Ma, Wei-Hung Weng, Attila Kiraly, Sahar Kazemzadeh, Zakkai Melamed, et al. Elixir: Towards a general purpose x-ray artificial intelligence system through alignment of large language models and radiology vision encoders. arXiv preprint arXiv:2308.01317, 2023. 2 +[34] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 3 +[35] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 3 + +[36] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 7 +[37] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. Mutr3d: A multi-camera tracking framework via 3d-to-2d queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4537–4546, 2022. 7 \ No newline at end of file diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/images.zip b/2024/VLP_ Vision Language Planning for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..128457651335037e6109f8f3508b53e394a37b74 --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33b7df5426d777b46b8642b303add8c942416adc813e65e3dc159791ae8338fa +size 401642 diff --git a/2024/VLP_ Vision Language Planning for Autonomous Driving/layout.json b/2024/VLP_ Vision Language Planning for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19fae84097935607004d02dc5056678cd29d5b2a --- /dev/null +++ b/2024/VLP_ Vision Language Planning for Autonomous Driving/layout.json @@ -0,0 +1,8073 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "text", + "content": "VLP: Vision Language Planning for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": "Chenbin Pan" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Burhaneddin Yaman" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2\\text{圆}}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Tommaso Nesti" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Abhirup Mallik" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Alessandro G Allievi" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Senem Velipasalar" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": " Liu Ren" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 142, + 475, + 185 + ], + "type": "text", + "content": "Syracuse University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 186, + 497, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 186, + 497, + 200 + ], + "spans": [ + { + "bbox": [ + 96, + 186, + 497, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 186, + 497, + 200 + ], + "type": "text", + "content": "Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 201, + 364, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 201, + 364, + 213 + ], + "spans": [ + { + "bbox": [ + 220, + 201, + 364, + 213 + ], + "type": "text", + "content": "{cpan14, svelipas}@syr.edu," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 215, + 533, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 215, + 533, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 215, + 533, + 228 + ], + "type": "text", + "content": "{burhaneddin.yaman,tommaso.nesti,abhirup.mallik,alessandro.allievi,liu. ren}@us.bosch.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "text", + "content": "Autonomous driving is a complex and challenging task that aims at safe motion planning through scene understanding and reasoning. While vision-only autonomous driving methods have recently achieved notable performance, through enhanced scene understanding, several key issues, including lack of reasoning, low generalization performance and long-tail scenarios, still need to be addressed. In this paper, we present VLP, a novel Vision-Language-Planning framework that exploits language models to bridge the gap between linguistic understanding and autonomous driving. VLP enhances autonomous driving systems by strengthening both the source memory foundation and the self-driving car's contextual understanding. VLP achieves state-of-the-art end-to-end planning performance on the challenging NuScenes dataset by achieving " + }, + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "inline_equation", + "content": "35.9\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "inline_equation", + "content": "60.5\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 521 + ], + "type": "text", + "content": " reduction in terms of average L2 error and collision rates, respectively, compared to the previous best method. Moreover, VLP shows improved performance in challenging long-tail scenarios and strong generalization capabilities when faced with new urban environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 685 + ], + "type": "text", + "content": "Autonomous driving is a complex problem requiring scene understanding and reasoning to ensure safe motion planning. This sophisticated challenge can be broadly divided into three main tasks, namely perception, prediction and planning (also known as P3). Conventional methods adopt a modular approach by developing and optimizing each task in a disjoint manner without a holistic view, leading to compounding errors and safety concerns [15, 16, 18]. End-to-end autonomous driving systems (ADS), unifying all P3 tasks, have garnered attention for their potential to enhance" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 266, + 422, + 376 + ], + "blocks": [ + { + "bbox": [ + 310, + 266, + 422, + 376 + ], + "lines": [ + { + "bbox": [ + 310, + 266, + 422, + 376 + ], + "spans": [ + { + "bbox": [ + 310, + 266, + 422, + 376 + ], + "type": "image", + "image_path": "6850a411b255ab31d7194d8ce722ca14b816f6fdc78873a73b44366a523b7f43.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 380, + 546, + 437 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 437 + ], + "type": "text", + "content": "Figure 1. New-city generalization ability of ADS for planning is evaluated by training on Boston city and testing on Singapore, and vice versa. Our proposed VLP shows strong generalization ability by significantly outperforming state-of-the-art vision-only method, VAD[18], in terms of both L2 error and collision rate." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 266, + 539, + 376 + ], + "blocks": [ + { + "bbox": [ + 427, + 266, + 539, + 376 + ], + "lines": [ + { + "bbox": [ + 427, + 266, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 427, + 266, + 539, + 376 + ], + "type": "image", + "image_path": "3e27faeb36b6d1b5f939040fb2888fe7125b35b4eda94e08cd1f762375b89245.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 568 + ], + "type": "text", + "content": "safe planning. Existing vision-based ADS typically follow a two-stage process: bird's eye view (BEV) feature extraction and downstream tasks [16, 18]. BEV feature extraction transforms multi-view camera data into a structured top-view representation embedding spatial information around the ego-car. BEV features are further utilized as the information pool by the downstream P3 tasks. While effective, these vision-only methods struggle with out-of-domain generalization, such as maintaining performance in new cities and long-tail scenarios, hindering real-world deployment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "The way humans process and interpret visual driving scenes naturally involve a coherent cognitive framework that maintains a consistent logical flow. This enables humans effortlessly make correct decisions even when faced with previously unseen data or scenarios, eliminating generalization issues in driving tasks. Then, this raises a pertinent question: How can ADS achieve human-like driving to navigate diverse and dynamic road environments? Recent advancements in language models (LMs) have led to unprecedented common-sense ability and generalization performance across unseen data and tasks for the natural language processing domain [5, 8, 25]. The superior common-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 261, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 261, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 261, + 703 + ], + "type": "text", + "content": "*Work done while interned at Bosch Research North America." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 703, + 140, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 140, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 140, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14760" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "sense capability of LMs have led to the emergence of multimodal LMs for diverse applications ranging from medical imaging to robotics [4, 9, 33]. However, incorporation of the reasoning ability of LMs into real-world autonomous driving tasks, to address generalization and long-tail scenarios, is yet to be fully-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 286, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 241 + ], + "type": "text", + "content": "To bridge this gap, we propose a Vision Language Planning (VLP) framework, which integrates the commonsense capability of LMs into vision-based ADS for safe self-driving. Our VLP consists of two key components: Agent-centric Learning Paradigm (ALP) and Self-driving-car-centric Learning Paradigm (SLP), leveraging LMs to enhance the ADS from reasoning and decision-making aspects, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "type": "text", + "content": "The BEV feature map serves as the source memory pool in ADS for downstream decoding tasks. It summarizes and encodes the driving environment surrounding the self-driving car, including vehicles, pedestrians, lanes, and more, into a unified feature map. Hence, capturing comprehensive and necessary details in each local position of BEV is critical for safe and precise self-driving performance. To enhance the local semantic representation and reasoning capabilities of BEV, we introduce an innovative Agent-centric Learning Paradigm (ALP) module. ALP integrates the consistent feature space of a pretrained LM to revamp the agent features on the BEV, actively shaping semantics, and guiding the BEV reasoning process. Leveraging on the common sense and logic flow embedded in the LM, our ALP equips the ADS with robustness and consistent BEV feature space, enhancing its effectiveness in diverse driving scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "type": "text", + "content": "In ADS, the planning module aggregates information from the preceding perception and prediction phases to make the final decisions for self-driving. This global perspective culminates in the formation of a planning query, directly influencing the safety and accuracy of the self-driving navigation. Considering the critical role of the planning module within ADS, we also present a novel Self-driving-car-centric Learning Paradigm (SLP) to elevate the decoding and acquiring information ability of the planning query. In the SLP, we align the planning query with intended goals and the ego-vehicle driving status by leveraging the knowledge encoded in the pretrained LM. The language model's comprehension capabilities contribute to more informed decision-making during the planning phase as well as enabling a more robust planning query formation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 628, + 286, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 698 + ], + "type": "text", + "content": "Through VLP, we bridge the gap between human-like reasoning and autonomous driving, enhancing the model's contextual awareness and its ability to generalize effectively in complex, ever-changing real-world scenarios, as shown in Fig.1. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "- We propose VLP, a Vision Language Planning model," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 544, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 107 + ], + "type": "text", + "content": "which incorporates reasoning capability of LMs into vision-based autonomous driving systems as an enhancement of motion planning and self-driving safety." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 108, + 544, + 322 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 108, + 544, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 544, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 544, + 154 + ], + "type": "text", + "content": "- VLP is composed of novel components ALP and SLP, aiming to improve the ADS from self-driving BEV reasoning and self-driving decision-making aspects, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 156, + 544, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 544, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 544, + 225 + ], + "type": "text", + "content": "- Through extensive experiments in real-world driving scenarios, we show that VLP significantly and consistently outperforms the state-of-the-art vision-based approaches across a spectrum of driving tasks, including open-loop planning, multi-object tracking, motion forecasting, and more." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 544, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 544, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 544, + 285 + ], + "type": "text", + "content": "- We conduct the first new-city generalization study on the nuScenes dataset [6] by training and testing on distinct cities, demonstrating the remarkable zero-shot generalization ability of our VLP approach over vision-only methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 287, + 544, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 544, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 544, + 322 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work introducing LM into multiple stages of ADS to address the generalization ability in new cities and long-tail cases." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 333, + 391, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 391, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 391, + 345 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "type": "text", + "content": "2.1. End-to-End Autonomous Driving" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 368, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 608 + ], + "type": "text", + "content": "Early end-to-end approaches have adopted a single neural network to perform motion planning without explicitly designing intermediate tasks, such as prediction [3]. Thus, such approaches not only suffer from the sub-optimal performance but also lack interpretability. Modular end-to-end frameworks introduce intermediate tasks for interpretability and provide improved performance [7, 10, 15, 16, 18, 23, 28]. These interpretable end-to-end frameworks explicitly model perception, prediction and planning (P3) components and train them together with a joint optimization strategy to enhance planning. UniAD [16] and VAD [18] are two prominent state-of-the-art end-to-end frameworks. UniAD leverages rasterized scene representations and explicitly identifies crucial components within the P3 framework. It coordinates all tasks through a unified query design for safe planning. On the other hand, VAD uses vectorized representations for efficient planning and safety improvement. While these approaches generally achieve good performance, their generalization performance have been limited due to lack of common sense and reasoning process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 614, + 444, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 444, + 626 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 444, + 626 + ], + "type": "text", + "content": "2.2. Vision-Language Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Large language models (LLMs), which are trained on massive amounts of text data, have shown unprecedented success in language-related tasks [5, 8, 29]. The capability of LLMs for common-sense understanding and versatility in tackling diverse language tasks has driven the exploration of multi-modal LLMs [20, 21, 27]. Vision-language models (VLM) integrate visual and textual information to en" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14761" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "able common sense understanding of the content by training on large scale open-world vision and text data [2, 17, 20, 27]. CLIP [27] uses a contrastive learning objective to learn a joint embedding space for vision and text data, and has shown immense zero-shot generalization performance, showcasing the importance of VLMs compared to vision-only approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 164, + 278, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 278, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 278, + 178 + ], + "type": "text", + "content": "2.3. Autonomous Driving with Language Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 373 + ], + "type": "text", + "content": "In recent years, numerous works have been proposed for extending LMS to the embodied AI domain to improve zero-shot generalization performance [4, 9, 30, 35]. These works incorporate sensory inputs from embodied agents in the form of language with vision and language data for decision making process [4, 9]. While these methods have mainly focused on the robotics domain, few works have been proposed for leveraging embodied LMs for autonomous driving tasks [11, 24, 31, 32, 34]. In particular, DiLu [32] and GPT-Driver [24] propose GPT-based driver agents for closed-loop simulation tasks. In [31], an open loop driving commentator, which combines vision and low-level driving actions with language, is proposed to interpret driving actions and for reasoning. Our work significantly differs from these approaches by leveraging LMs for end-to-end motion planning task for real-world driving applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 384, + 130, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 130, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 130, + 398 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 289, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 289, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 289, + 570 + ], + "type": "text", + "content": "We present a Vision Language Planning (VLP) model, which equips the ADS with the capacity to emulate human common sense and engage in contextual reasoning for safe motion planning. Our proposed VLP model, illustrated in Fig. 2a, comprises two innovative components that leverage LMs in both local and global contexts. The Agent-centric Learning Paradigm (ALP) concentrates on refining local details to enhance source memory reasoning, while the Self-driving-car-centric Learning Paradigm (SLP) focuses on guiding the planning process for the self-driving-car (SDC). Below, we first provide an overview of how general ADS work in Sec.3.1. Subsequently, we describe our proposed ALP and SLP in detail in Sec. 3.2 and Sec. 3.3, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In vision-based ADS, a sequence of multi-view camera images serves as the source input, offering abundant visual data for downstream tasks. Initially, an image backbone extracts fundamental features for perception. These features are then processed by a BEV encoder to transform multi-view features into a unified 2D representation from a top-down view. This BEV feature map encapsulates crucial perception details like object positions, lane markings, and road boundaries [22]. Acting as an information hub, this BEV feature map supports diverse downstream tasks such" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "as 3D object tracking, mapping, motion prediction, occupancy prediction, and path planning [16, 18]. Specific task-related information is decoded using a query-based transformer decoder designed for each task to efficiently access the source memory [16]. In the final planning stage, an ego-query models the ego-vehicle's status and interacts with environmental features to determine the optimal path for autonomous driving." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 171, + 545, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 196 + ], + "type": "text", + "content": "3.2. Agent-centric Learning Paradigm: Enhancing BEV Source Memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 378 + ], + "type": "text", + "content": "The BEV feature map serves as the foundational source memory pool and holds a pivotal role in ADS. Ensuring that a BEV map provides comprehensive information with intricate details for driving is essential for enabling downstream decoders to make safe, precise, and human-like decisions. However, the BEV in ADS is derived from multiview camera images rather than a true bird's-eye-view image, which can introduce discrepancies between the produced BEV and the expected BEV representation. Therefore, to mitigate the discrepancies, we introduce the Agent-centric Learning Paradigm (ALP) to align the produced BEV with a true bird's-eye-view map. Through direct supervision on the BEV map with LM, our ALP enables the refinement of local details and alignment with the desired human perspective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "text", + "content": "Agent BEV features. In our proposed ALP, three kinds of BEV agents are considered: ego-vehicle (self-driving-car), foreground (FG) objects, and lane elements. We first align the ground-truth area of each agent with the produced BEV map, and crop the regions of interest. We utilize the 3D bounding box to crop the ego-car and FG object area, and panoptic scene mask to segment the lane area. Subsequently, we perform a pooling operation on the obtained local BEV region, to generate a single feature representation for the corresponding agent. After pooling, the local agent features in each sample along the batch are concatenated to formulate an Agent-BEV tensor denoted as " + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{bev} \\in \\mathbb{R}^{N_B \\times C}" + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "inline_equation", + "content": "N_B" + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 378, + 545, + 557 + ], + "type": "text", + "content": " denote total number of agents in the batch and feature dimension, respectively. The process can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 559, + 545, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 545, + 621 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {b e v} ^ {E g o} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {E g o})) \\text {,} \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {F G} = \\operatorname {P o o l} (\\operatorname {C r o p} (B E V, b b o x _ {3 D} ^ {F G})) \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {b e v} ^ {\\text {L a n e}} = \\operatorname {P o o l} \\left(\\operatorname {S e g} \\left(B E V, \\operatorname {S e g M a s k} ^ {\\text {L a n e}}\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {1} \\\\ \\mathcal {A} _ {b e v} = B a t c h \\left(\\left[ \\mathcal {A} _ {b e v} ^ {E g o}; \\mathcal {A} _ {b e v} ^ {F G}; \\mathcal {A} _ {b e v} ^ {L a n e} \\right]\\right), \\in \\mathbb {R} ^ {N _ {b a t c h} \\times C}, \\\\ \\end{array}", + "image_path": "f355b9d51ca00648aad38888c1ae06d9d80f42636e70b817ef6674a010bef767.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "inline_equation", + "content": "Batch[;]" + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{bev}^{(\\cdot)}" + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "content": " represents the single agent feature of the corresponding type, " + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "inline_equation", + "content": "bbox_{3D}^{(\\cdot)}" + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "inline_equation", + "content": "SegMask^{Lane}" + }, + { + "bbox": [ + 304, + 625, + 545, + 689 + ], + "type": "text", + "content": " indicate the ground-truth 3D bounding box and the lane segmentation mask, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Agent expectation features. To ensure that local BEV features express the desired information, we conduct a BEV-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14762" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 407, + 223 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 61, + 83 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 61, + 83 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 61, + 83 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 73, + 407, + 223 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 407, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 407, + 223 + ], + "type": "image", + "image_path": "14e20e0a0aab98af833b8caf17461440292d8bddcff7879081c620ab73773bf0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 74, + 414, + 83 + ], + "lines": [ + { + "bbox": [ + 406, + 74, + 414, + 83 + ], + "spans": [ + { + "bbox": [ + 406, + 74, + 414, + 83 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 76, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 417, + 76, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 417, + 76, + 541, + 220 + ], + "type": "text", + "content": "Ego-car prompt: This is the self-driving car. It is {class}, and its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. It is currently {command}. Its future trajectory will be {{x1, y1}...{xn, yn}]. Foreground prompt: This object is {class}. Its 3D bounding box is {cx, cy, w, l, cz, h, sin, cos, vx, vy}. Its future trajectory will be {{x1, y1}...{xt, yt}]. Lane prompt: This lane belongs to a {lane-label}. The label belongs to {thing/stuff} class. Its 2D bounding box on the BEV map is {cx, cy, w, h}. Planning prompt: The self-driving car is driving in an urban area. It is currently {command}. The future trajectory of the car for the next P timestamps will be {{[x1, y1]...{xp, yp}]}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 238, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 545, + 293 + ], + "type": "text", + "content": "Figure 2. a) The overview of proposed vision language planning (VLP) framework. VLP enhances ADS from self-driving BEV-reasoning and self-driving decision-making aspects, through two innovative modules, ALP and SLP, respectively. Leveraging LM and contrastive learning, ALP conducts agent-wise learning for refining local details on BEV, while SLP engages sample-wise learning for advancing global context understanding ability of the ADS. VLP is only active during training, ensuring no additional parameters or computations are introduced during inference. b) Prompt formats used in VLP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "text", + "content": "expectation alignment process by leveraging LM and contrastive learning. We precisely define the perceptual information expected from the corresponding agent, such as agent labels, bounding boxes, and future trajectories. These driving-related ground-truth information, which should also be embedded in the local BEV feature is formulated into a prompt as illustrated in Fig. 2b. The description is then passed to the language encoder, " + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "text", + "content": ", to generate the corresponding agent expectation feature. We apply an MLP layer " + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{bev}" + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "text", + "content": " to adapt the expectation feature to the BEV feature space. Then, the agent expectation features are concatenated along the batch to generate an Agent-Expectation tensor denoted as " + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{exp} \\in \\mathbb{R}^{N_B \\times C}" + }, + { + "bbox": [ + 46, + 308, + 289, + 477 + ], + "type": "text", + "content": ". The procedure can be formulated as in Eq. (2):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 479, + 286, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 479, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 84, + 479, + 286, + 541 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {e x p} ^ {E g o} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {E g o} [ y ^ {E g o} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {F G} = \\mathcal {F} _ {b e v} (L M (\\boldsymbol {T} ^ {F G} [ y ^ {F G} ])), \\in \\mathbb {R} ^ {C}, \\\\ \\mathcal {A} _ {e x p} ^ {L a n e} = \\mathcal {F} _ {b e v} \\left(L M \\left(\\boldsymbol {T} ^ {L a n e} \\left[ y ^ {L a n e} \\right]\\right)\\right), \\in \\mathbb {R} ^ {C}, \\tag {2} \\\\ \\mathcal {A} _ {e x p} = B a t c h \\left(\\left[ \\mathcal {A} _ {e x p} ^ {e g o}; \\mathcal {A} _ {e x p} ^ {F G}; \\mathcal {A} _ {e x p} ^ {l a n e} \\right]\\right), \\\\ \\end{array}", + "image_path": "aebcdcd0298f8184f5e05903eefad6dfd7a518c7e82cd9ba576208efc651115c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{exp}^{(\\cdot)}, T^{(\\cdot)}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "y^{(\\cdot)}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " represent the single agent expectation feature, the description template, and the ground-truth in ADS for the respective agent. Note that during training, we freeze the " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " and only train the adaptation layer " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{bev}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " to save memory and retain the pre-trained knowledge in the " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Contrastive Learning Formulation in ALP. Given the agent BEV feature " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{bev}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and the agent expectation feature " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{exp}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", the alignment between the produced BEV and the expected BEV is facilitated through a contrastive learning process [27]. In particular, L2 normalization is applied to both " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{bev}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{exp}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " to standardize the feature vectors and ensure training stability. Then, matrix multiplication is performed between the normalized two-modality tensors with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": "a learnable logit scale " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\alpha_{alp}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": " to produce a similarity matrix " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "S_{pred} \\in \\mathbb{R}^{N_B \\times N_B}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": " in agent-wise. The ground truth for the similarity matrix, " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "S_{gt}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": ", is a unit matrix, signifying that each agent's BEV feature should be closest to its corresponding expected feature in the shared space. To optimize this alignment, we apply a cross-entropy loss, " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": ", along both the Agent-BEV mode axis and Agent-Expectation mode axis of the " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "S_{pred}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": ", and compute an average value as the final ALP loss, " + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{alp}" + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": ", for the BEV source reasoning part. The entire process can be formally expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 436, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 436, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 348, + 436, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\mathcal {S} _ {p r e d} = \\alpha_ {a l p} \\times \\left(\\frac {\\mathcal {A} _ {b e v}}{\\| \\mathcal {A} _ {b e v} \\| _ {2}} \\otimes \\frac {\\mathcal {A} _ {e x p}}{\\| \\mathcal {A} _ {e x p} \\| _ {2}}\\right), \\\\ \\mathcal {L} _ {a l p} = \\left(\\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {q t}, d i m = 0\\right) \\right. \\end{array} \\tag {3} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d}, \\mathcal {S} _ {g t}, d i m = 1\\right)) / 2. \\\\ \\end{array}", + "image_path": "1351eb7f606ea3ee3c698c53b27ffad1e9444f77d9bccdf67f89eca378e1f8f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 491, + 545, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 545, + 515 + ], + "type": "text", + "content": "3.3. SDC-centric Learning Paradigm: Enhancing Ego-vehicle Query" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": "Ego-vehicle query feature. The ego-vehicle dynamic modeling is the core function of the ADS. In previous ADS, a trainable ego-query " + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{query}" + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": " is applied to interact with other agents, " + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{query}^{FG}" + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{query}^{Lane}" + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": ", on BEV map, to gather the self-driving perception/prediction information. The produced ego-car query feature " + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{qfeat}" + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": " is further processed to predict the future waypoints of the self-driving car " + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "inline_equation", + "content": "y_{pred}^{plan}" + }, + { + "bbox": [ + 304, + 521, + 545, + 618 + ], + "type": "text", + "content": ". The process is formulated as in Eq. (4):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 625, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 625, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {q f e a t} = \\mathcal {M} ^ {i n t e r} \\left(\\mathcal {E} _ {q u e r y}, \\mathcal {A} _ {q u e r y} ^ {F G}, \\mathcal {A} _ {q u e r y} ^ {l a n e}\\right) \\in \\mathbb {R} ^ {B \\times C}, \\tag {4} \\\\ y _ {p r e d} ^ {\\text {p l a n}} = \\mathcal {M} ^ {\\text {p l a n}} \\left(\\mathcal {E} _ {q f e a t}\\right) \\in \\mathbb {R} ^ {B \\times P \\times 2}, \\\\ \\end{array}", + "image_path": "342bd76f5324fd149a945556fcf58937d00c872139dd11644d3dace1a5a6b0de.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{inter}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{plan}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " denote the intermediate ego information gathering module and the planning head in ADS, respectively. " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "P \\times 2" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " indicate the planned waypoints in the next P timesteps, and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the batch size during training." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14763" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "Although such mechanism can achieve good planning results, solely relying on numeric ground-truth can make it challenging to understand the rationale behind the ADS' decisions, which can result in inconsistent feature learning for the planning and lead to limited adaptability to new environments. To address these limitations, we propose a SDCcentric Learning Paradigm (SLP) to equip the ADS with the capability of making decisions from continuous and robust feature space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "text", + "content": "Ego-vehicle planning feature. We template a language description of the ego-vehicle status with the planning ground truth (GT) including high-level driving command and the future trajectory of the ego-vehicle, as illustrated in Fig. 2b. The prompt is sent to the " + }, + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "text", + "content": " to obtain the ground truth planning features of the ego-car embedding driving scenario information and human driving logic. An MLP layer " + }, + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ego}" + }, + { + "bbox": [ + 46, + 180, + 289, + 301 + ], + "type": "text", + "content": " is applied to adapt the textual planning features to the ego-query feature space. The process can be expressed as in Eq. (5):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 316, + 287, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 316, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 70, + 316, + 287, + 332 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {p r o m p t} = \\mathcal {F} _ {e g o} (L M (\\boldsymbol {T} _ {s l p} ^ {E g o} [ y ^ {p l a n} ])) \\in \\mathbb {R} ^ {B \\times C}, \\quad (5)", + "image_path": "c430bf4490aaad4d7a1a7735e78d34095cead2f208d62dc2a266fdfca55787e3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "inline_equation", + "content": "T_{slp}^{Ego}" + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": " indicates the prompt for ego-car description used in SLP, and " + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "inline_equation", + "content": "y^{plan}" + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": " denotes the planning ground truth. Note that the " + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": " is a shared off-the-shelf model for both ALP and SLP. As in ALP, only the adaptation layer " + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ego}" + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": " is trainable during training to save memory and retain the pre-trained knowledge in " + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 342, + 287, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 465 + ], + "type": "text", + "content": "Contrastive Learning Formulation in SLP. Similar to ALP, we employ a sample-wise contrastive learning approach for SLP to align the ego-vehicle query feature with the ego-vehicle textual planning feature as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 469, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 469, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 81, + 469, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {S} _ {p r e d} ^ {s l p} = \\alpha_ {s l p} \\times (\\frac {\\mathcal {E} _ {q f e a t}}{\\| \\mathcal {E} _ {q f e a t} \\| _ {2}} \\otimes \\frac {\\mathcal {E} _ {p r o m p t}}{\\| \\mathcal {E} _ {p r o m p t} \\| _ {2}}), \\\\ \\mathcal {L} _ {s l p} = \\left(\\mathcal {L} _ {C E} \\left(S _ {p r e d} ^ {s l p}, S _ {g t} ^ {s l p}, d i m = 0\\right) \\right. \\tag {6} \\\\ + \\mathcal {L} _ {C E} \\left(\\mathcal {S} _ {p r e d} ^ {s l p}, \\mathcal {S} _ {g t} ^ {s l p}, d i m = 1\\right)) / 2, \\\\ \\end{array}", + "image_path": "9b52dc194d6223bbac2e39fbff197fbd42d501e318827aeaa5b012a5429e5183.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "inline_equation", + "content": "\\alpha_{slp}" + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "text", + "content": " is a learnable logit scale, and " + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "inline_equation", + "content": "S_{pred}^{slp}" + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "inline_equation", + "content": "S_{gt}^{slp}" + }, + { + "bbox": [ + 46, + 529, + 289, + 651 + ], + "type": "text", + "content": " are the predicted and ground truth similarity matrices, respectively. Aligning the two modes closely in the feature space, the contrastive process refines the produced ego-vehicle query using the ground truth-embedded linguistic feature. The sample-wise contrastive loss rectifies the relationships between individual samples from a human (linguistic) perspective, establishing a connection for the current data point with the world-wise common sense embedded in the pretrained language model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 135, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 135, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 135, + 671 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "The overall loss, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vlp}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", of our VLP training is composed of two parts: the BEV encoder reasoning loss " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enc}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " and the decoder decision-making loss " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 363, + 86, + 545, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 86, + 545, + 99 + ], + "spans": [ + { + "bbox": [ + 363, + 86, + 545, + 99 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v l p} = \\omega_ {e n c} \\mathcal {L} _ {e n c} + \\omega_ {d e c} \\mathcal {L} _ {d e c}, \\tag {7}", + "image_path": "e86ffaff0872f38d5679244728086c2b4d1ed4c3084d7c96b04b79de9e3f4de8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\omega_{enc}" + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\omega_{dec}" + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "content": " represent the weights for " + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enc}" + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 306, + 105, + 545, + 129 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "text", + "content": "The encoder reasoning loss, " + }, + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enc}" + }, + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "text", + "content": ", is equal to the loss produced in ALP module, i.e. " + }, + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enc} = \\mathcal{L}_{alp}" + }, + { + "bbox": [ + 305, + 130, + 545, + 178 + ], + "type": "text", + "content": ". The decision-making loss includes the losses in all downstream tasks (perception/prediction/planning) in ADS and the SLP loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 184, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 184, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 345, + 184, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e c} = \\mathcal {L} _ {p e r c} + \\mathcal {L} _ {p r e d} + \\mathcal {L} _ {p l a n} + \\mathcal {L} _ {s l p}. \\tag {8}", + "image_path": "64225bad3d036bf802a8c203d9e66a5f5beb29041022ead880d0ce07eaee21b2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 239 + ], + "type": "text", + "content": "Note that the off-the-shelf LM head is discarded during inference, and thus, our method introduces no additional parameters and computations to the original ADS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 251, + 388, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 388, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 388, + 265 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 271, + 440, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 271, + 440, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 440, + 284 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 290, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 398 + ], + "type": "text", + "content": "Dataset. We conduct experiments on the challenging nuScenes dataset [6], which is the first large-scale public dataset to provide data from the entire sensor suite of an autonomous vehicle (6 cameras, 1 LIDAR, 5 RADAR, GPS and IMU sensors). The nuScenes contains 1000 driving scenes from Boston and Singapore, two cities that are known for their dense traffic and highly challenging driving conditions. In our experiments, we utilize 6 camera images as our vision inputs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "content": "Baselines. We adopt two recent SOTA models in autonomous driving, namely UniAD [16] and VAD [18], as our baselines to evaluate the effectiveness of our approach." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 435, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 545, + 578 + ], + "type": "text", + "content": "Training. We incorporate our proposed VLP training strategy into vision-only approaches of UniAD [16] and VAD [18], and refer to this enriched versions as VLP-UniAD and VLP-VAD, respectively, for clarity when reporting the results. All methods are trained with the same hyper-parameters reported in the respective baselines [16, 18] for commensurate comparison. For language model, we use the pretrained LM in CLIP [27]. In VLP study, we set " + }, + { + "bbox": [ + 305, + 435, + 545, + 578 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 305, + 435, + 545, + 578 + ], + "type": "text", + "content": " (3 seconds) for planning, which is same as baselines [16] [18]. Experiments are conducted with 8 NVIDIA Tesla A100 GPUs. More details regarding experimental setup can be found in supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 586, + 423, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 423, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 423, + 600 + ], + "type": "text", + "content": "4.2. Open-loop Planning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "Planning constitutes the cornerstone of any ADS, playing a pivotal role in ensuring safety and devising efficient routes for the ego-car. In Tab. 1, we present a series of comparative experiments that showcase the performance of our open-loop planning in comparison to the baseline models. As can be seen in rows 4-6 of the table, the integration of just SLP leads to noticeable reductions in both the L2 error and collision rates for all the baseline models. Moving down the table, rows 7 to 9 demonstrate that the inclusion" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14764" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "of both VLP components (SLP and ALP together) consistently yields further improvements in these planning metrics. In particular, VLP-UniAD shows a " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "28.1\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "48.4\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " reduction in terms of average L2 error and collision rate, respectively, compared to baseline UniAD. Similarly, in comparison with VAD, VLP-VAD achieves " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "35.9\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "60.5\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " reduction for average L2 error and collision rate, respectively. These significant results underscore the effectiveness of both SLP and ALP, as well as their adaptability across various ADS configurations. The reduced L2 error and collision rate achieved through VLP integration contribute to safer road planning in the realm of self-driving." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 228, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 192, + 240 + ], + "type": "text", + "content": "4.3. Perception and Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 340 + ], + "type": "text", + "content": "In this section, we showcase the consistent effectiveness of our proposed VLP across a spectrum of essential driving tasks. Our proposed approach excels in various perception and prediction tasks, including multi-object tracking, mapping, motion forecasting, occupancy prediction, 3D object detection, and vectorized scene segmentation. These results underscore the versatility and reliability of our VLP framework across a wide range of critical driving tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 289, + 510 + ], + "type": "text", + "content": "Multi-Object Tracking. The results for multi-object tracking (MOT) are presented in Tab. 2. We apply the standard evaluation protocols of AMOTA (Average Multi-Object Tracking Accuracy), AMOTP (Average Multi-Object Tracking Precision), Recall, and IDS (Identity Switches) to evaluate the MOT performance. As can be seen, with the incorporation of SLP, AMOTA, Recall, and IDS values all surpass those of UniAD, while AMOTP experiences a slight degradation. The combined integration of both SLP and ALP results in improvements across all metrics when compared to the UniAD baseline. The enhanced performance in MOT proves that our VLP can help the system better predict and respond to the movements of other objects on the road, reducing the risk of accidents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "Online Mapping. In Tab. 3, we present the results of online mapping, encompassing four key mapping elements: lanes, drivable areas, dividers, and pedestrian crossings. The evaluation employs the Intersection-over-Union (IoU) metric to measure the overlap between the predicted and ground-truth maps. The results reveal that the inclusion of the SLP leads to improvements in mapping accuracy for all four elements. Specifically, it enhances the IoU for drivable areas and crossings by " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", respectively. Furthermore, the integration of both SLP and ALP yields even further improvement, with IoU achieving by " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " for drivable areas and crossings, respectively. These enhancements represent substantial gains compared to the SLP-only configuration. These results underscore the valuable contributions of each paradigm in online mapping. They also demonstrate the effectiveness of VLP in bridging the gap between vision-based and language-based information, ul" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "timely enhancing the system's comprehension of various road elements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 98, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 312 + ], + "type": "text", + "content": "Motion Forecasting. We present the motion forecasting results in Tab. 4, applying the same evaluation metrics as those in UniAD [16], namely minADE (minimum Average Displacement Error), minFDE (minimum Final Displacement Error), MR (Miss Rate), and EPA (End-to-end Prediction Accuracy). The results demonstrate that the inclusion of the SLP contributes to a reduction in minADE, minFDE, and MR, while simultaneously increasing EPA for the ADS. This outcome signifies that the supervision and refinement applied to the final ego-car feature also has a positive influence on the motion prediction of other vehicles around the ego-car. The improvements are more obvious with the EPA metrics. Equipped with full components of VLP, the performance of the ADS, in motion prediction, experiences further enhancements. This observation underscores the capability of ALP to empower the system in the identification and accurate prediction of the movements of various agents on the road." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "content": "Occupancy Prediction. We present the occupancy prediction results in Tab. 5, which have been obtained using the IoU and Video Panoptic Quality (VPQ) within two distance ranges around the ego-car, namely near (\"-n.\", covering a " + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "inline_equation", + "content": "30 \\times 30m" + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "content": " area, and far (\"-f.\", spanning a " + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "inline_equation", + "content": "50 \\times 50m" + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "content": " area. The results highlight the effectiveness of the SLP in consistently improving all four metrics. SLP, primarily focusing on enhancing the ego-car feature, demonstrates consistent performance gains in occupancy prediction. The inclusion of ALP with SLP shows similar improvement over the baseline as well. Since both configurations of the VLP prioritize attention to the ego-car during training, more improvements are observed within the nearby areas." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": "3D Object Detection. We show the 3D object detection results in Tab. 6 using nuScenes detection metrics of mean Average Precision (mAP), Average Translation Error (ATE), Average Scale Error (ASE), and nuScenes detection score (NDS) as indicated in [6]. As observed in the table, the incorporation of the VLP consistently leads to improvements in NDS. Notably, models incorporating both the SLP and ALP generally perform better than those with SLP alone, providing empirical evidence of the ALP's effectiveness in enhancing the local BEV representation ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Vectorized Scene Segmentation in VAD. The VAD [18] framework represents the driving scene as a fully vectorized structure, categorizing map elements into road boundaries, dividers, and pedestrian crossings. Thus, vectorized scene segmentation experiments are conducted within the VAD framework, with IoU serving as the evaluation metric. As depicted in Tab. 7, the introduction of SLP yields substantial improvements. More specifically, the segmentation for boundaries, dividers, and crossings sees improvements of " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "14.9\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "19.2\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "23.1\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", respectively, contributing to an" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14765" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 70, + 469, + 228 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 469, + 228 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 469, + 228 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 469, + 228 + ], + "type": "table", + "html": "
IDModelSLPALPL2 (m) ↓Col. Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
0NMP [36]--2.31---1.92
1SA-NMP [36]--2.05---1.59
2FF [14]0.551.202.541.430.060.171.070.43
3EO [19]0.671.362.781.600.040.090.880.33
4ST-P3 [15]1.332.112.902.110.230.621.270.71
5UniAD [16]0.480.961.651.030.050.170.710.31
6VLP-UniAD✓0.430.861.470.920.030.150.480.22
7VLP-UniAD✓✓0.360.681.190.740.030.120.320.16
8VAD [18]0.460.761.120.780.210.350.580.38
9VLP-VAD✓0.260.470.780.500.120.170.420.23
10VLP-VAD✓✓0.300.530.840.550.010.070.380.15
", + "image_path": "88bbbeb6ad61d7c2571916bcc2d037aa2c1027cc937a755496d7815af89111e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 277, + 285, + 354 + ], + "blocks": [ + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "type": "text", + "content": "Table 1. Open-loop planning performance. VLP achieves significant end-to-end planning performance improvement over counterpart vision only UniAD and VAD methods on the nuScenes validation dataset [6]. Based on the planning results, we can conclude that both SLP and ALP components plays a vital role to ensure safe motion planning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 277, + 285, + 354 + ], + "lines": [ + { + "bbox": [ + 49, + 277, + 285, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 285, + 354 + ], + "type": "table", + "html": "
ModelSLPALPAMOTA↑AMOTP↓Recall↑IDS↓
ViP3D [10]21.71.62536.3-
QD3DT [13]24.21.51839.9-
MUTR3D [37]29.41.49842.73822
UniAD [16]35.91.32046.7906
VLP-UniAD✓36.61.33246.8820
VLP-UniAD✓✓36.81.31547.3678
", + "image_path": "74f712f142e8e404cb9d6636225c4ce8f1cf0f3056ce75edcff575169f18d74c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 387, + 285, + 436 + ], + "blocks": [ + { + "bbox": [ + 47, + 358, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 358, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 287, + 381 + ], + "type": "text", + "content": "Table 2. Multi-object tracking. VLP shows improved performance over vision-only MOT techniques." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 387, + 285, + 436 + ], + "lines": [ + { + "bbox": [ + 49, + 387, + 285, + 436 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 285, + 436 + ], + "type": "table", + "html": "
ModelSLPALPLanes↑Drivable↑Divider↑Crossing↑
UniAD [16]31.369.125.713.8
VLP-UniAD✓32.069.527.114.7
VLP-UniAD✓✓32.370.227.415.2
", + "image_path": "5145210261e75ee1b8749ab7c4d1f3aa6bf25450842312d0621e66a8ab97d210.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 49, + 468, + 285, + 563 + ], + "blocks": [ + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "type": "text", + "content": "Table 3. Online mapping. VLP demonstrates improved segmentation IoU " + }, + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "type": "text", + "content": " performance over the UniAD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 468, + 285, + 563 + ], + "lines": [ + { + "bbox": [ + 49, + 468, + 285, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 285, + 563 + ], + "type": "table", + "html": "
ModelSLPALPminADE(m)↓minFDE(m)↓MR↓EPA↑
PnPNet [23]1.151.950.2260.222
ViP3D [10]2.052.840.2460.226
UniAD [16]0.711.020.1510.456
VLP-UniAD✓0.721.040.1540.459
VLP-UniAD✓✓0.680.980.1330.460
VAD [18]0.781.070.1210.598
VLP-VAD✓0.771.030.1100.621
VLP-VAD✓✓0.771.050.1150.621
", + "image_path": "941274d4d3fc374bab96b2d4c13365cba6fd055da843c358d79ad0d5d94573fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": "overall increase of " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " in the mean IoU (mIoU). VLP with SLP and ALP also achieves similar improvement over the VAD baseline. This increased accuracy is crucial for developing a safe and efficient self-driving system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 661, + 139, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 139, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 139, + 673 + ], + "type": "text", + "content": "4.4. Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Autonomous vehicles are intended to operate in diverse environments. Generalization allows a system to be deployed in various urban landscapes, suburban areas, or even rural" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 277, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 47, + 566, + 287, + 588 + ], + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 588 + ], + "type": "text", + "content": "Table 4. Motion forecasting. VLP achieves better motion forecasting over counterpart vision based methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 277, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 309, + 277, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 309, + 277, + 544, + 353 + ], + "type": "table", + "html": "
ModelSLPALPIoU-n.↑IoU-f.↑VPQ-n.↑VPQ-f.↑
FIERY [12]59.436.750.229.9
StretchBEV [1]55.537.146.029.0
ST-P3 [15]-38.9-32.1
UniAD [16]63.440.254.733.5
VLP-UniAD✓64.240.755.834.5
VLP-UniAD✓✓64.140.255.934.1
", + "image_path": "0fe3c04aa5af30b81cb22779cbe8aabd545155d6b9b3711436715704afc9e295.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 398, + 544, + 454 + ], + "blocks": [ + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "type": "text", + "content": "Table 5. Occupancy prediction. VLP shows improvement in both near " + }, + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "(30\\times 30m)" + }, + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "type": "text", + "content": " and far " + }, + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "(50\\times 50m)" + }, + { + "bbox": [ + 306, + 357, + 545, + 392 + ], + "type": "text", + "content": " ranges, denoted as \"n.\" and \"f.\", respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 398, + 544, + 454 + ], + "lines": [ + { + "bbox": [ + 308, + 398, + 544, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 544, + 454 + ], + "type": "table", + "html": "
ModelSLPALPmAP↑mATE↓mASE↓NDS↑
VAD [18]0.270.700.300.389
VLP-VAD✓0.270.670.300.394
VLP-VAD✓✓0.280.670.300.406
", + "image_path": "0f86c999808b76a6485a313632bb51e0bfe7166f4d3744e7a197763fb2afc0b1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 487, + 544, + 538 + ], + "blocks": [ + { + "bbox": [ + 306, + 457, + 545, + 480 + ], + "lines": [ + { + "bbox": [ + 306, + 457, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 545, + 480 + ], + "type": "text", + "content": "Table 6. 3D object detection. VLP achieves improved object detection performance over the baseline." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 487, + 544, + 538 + ], + "lines": [ + { + "bbox": [ + 308, + 487, + 544, + 538 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 544, + 538 + ], + "type": "table", + "html": "
ModelSLPALPBoundary↑Divider↑Crossing↑mIoU↑
VAD [18]45.642.231.639.8
VLP-VAD✓52.450.338.947.2
VLP-VAD✓✓49.448.439.245.7
", + "image_path": "24e8c7a9fe665d0f62705f29ab23d2aa6afbdba54651cf82048324c10665c3de.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 541, + 545, + 563 + ], + "lines": [ + { + "bbox": [ + 306, + 541, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 545, + 563 + ], + "type": "text", + "content": "Table 7. Vectorized scene segmentation. VLP shows improved segmentation IoU (%) performance over the VAD." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "type": "text", + "content": "settings, making the technology applicable and accessible on a broader scale. Long-tail cases, which are scenarios that occur infrequently, are often underrepresented in training data. Generalization helps the system cope with these rare but critical situations, reducing the risk of biased decision-making based on inadequate exposure during training. In this section, we evaluate the VLP in terms of generalization ability to new cities and long-tail cases." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "New-city Generalization. To keep the same number of multi-view camera inputs, we construct a multi-city dataset exclusively from nuScenes, by encompassing data from two" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14766" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 284, + 142 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 284, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 284, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 284, + 142 + ], + "type": "table", + "html": "
ModelVLPBostonSingapore
Avg.L2↓Avg.Col↓Avg.L2↓Avg.Col↓
UniAD [16]1.240.321.050.37
VLP-UniAD✓1.140.260.870.34
VAD [18]0.860.270.780.39
VLP-VAD✓0.730.220.630.20
", + "image_path": "0b53566a6685299f38a16a40f3e9b594bd4c0d76f73de44d5deefaabee3e7c5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "content": "distinct urban environments, Boston and Singapore. To comprehensively assess generalizability, we conduct two sets of experiments: (i) training on Boston and testing on Singapore; and (ii) conversely, training on Singapore and testing on Boston. The results, presented in Tab. 8, reveal a noteworthy reduction in planning L2 error and collision rates for the baselines in both scenarios with the integration of VLP. In particular, VLP-VAD achieves strong generalization over counterpart vision-only VAD approach (in Boston: " + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "inline_equation", + "content": "18.5\\%" + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "content": ", in Singapore: " + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "inline_equation", + "content": "19.2\\%" + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 46, + 206, + 287, + 385 + ], + "type": "text", + "content": " reduction in terms of average L2 error and collision rates, respectively). This signifies VLP's prowess in enhancing the safety and reliability of the ADS as well as its capacity to transcend training confines and excel in diverse real-world conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "text", + "content": "Long-tail Generalization. We assess the performance of UniAD in Multi-Object Tracking (MOT) and VAD in 3D detection concerning long-tail scenarios, as detailed in Tab. 9a and Tab. 9b, respectively. Utilizing the long-tail split methodology from [26], derived from per-class object counts, we observe that VLP consistently enhances the generalization capabilities of both frameworks. Specifically, VLP provides substantial improvements, increasing AMOTA and Recall by " + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "text", + "content": ", respectively, for UniAD. In the case of VAD, VLP improves the mean Average Precision (mAP) by " + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 46, + 387, + 287, + 542 + ], + "type": "text", + "content": ". These results highlight the robust feature space cultivated by our proposed VLP, emphasizing its efficacy in handling long-tail scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 147, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 147, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 147, + 562 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 569, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 287, + 617 + ], + "type": "text", + "content": "Effectiveness of SLP and ALP. VLP comprises ALP and SLP components. Extensive experiments presented in prior tables show the importance and necessity of both SLP and ALP for enhanced planning and safety." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Prompt Format. The design of the prompt format is pivotal for the success of our VLP, serving as the structured input that guides the learning process. We have performed ablation studies to systematically investigate the impact of different components of the ground truth information included in the prompt. The driving ground truth encompasses labels, bounding boxes, trajectories, and high-level commands. To demonstrate the indispensability of including all aspects of" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 554, + 118 + ], + "blocks": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "type": "text", + "content": "Table 8. New-city generalization. To evaluate the generalization performance, we train the model on Boston and test it on Singapore, and vice versa. VLP shows remarkable zero-shot generalization performance improvement over the vision-only methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 554, + 118 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 554, + 118 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 554, + 118 + ], + "type": "table", + "html": "
ModelVLPAMOTA↑AMOTP↓Recall↑IDS↓
UniAD[16]29.61.44639.268
VLP-UniAD✓30.71.43542.166
", + "image_path": "15f6d463baf6d9b93dee76e7bf9a77c003b941fc74eadbb7631824db70164c3f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 136, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 361, + 120, + 501, + 129 + ], + "lines": [ + { + "bbox": [ + 361, + 120, + 501, + 129 + ], + "spans": [ + { + "bbox": [ + 361, + 120, + 501, + 129 + ], + "type": "text", + "content": "(a) Multi-object tracking on long-tail cases." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 136, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 136, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 136, + 553, + 186 + ], + "type": "table", + "html": "
ModelVLPmAP↑mATE↓mASE↓mAOE↓
VAD [18]17.60.790.330.95
VLP-VAD✓20.40.750.330.83
", + "image_path": "678de87bf61eb1464135f19e66b1fa69f98195a94b9c56555df2751e11dca488.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 243, + 531, + 312 + ], + "blocks": [ + { + "bbox": [ + 368, + 189, + 494, + 198 + ], + "lines": [ + { + "bbox": [ + 368, + 189, + 494, + 198 + ], + "spans": [ + { + "bbox": [ + 368, + 189, + 494, + 198 + ], + "type": "text", + "content": "(b) Object detection on long-tail cases." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 205, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 239 + ], + "type": "text", + "content": "Table 9. Long-tail generalization. VLP shows consistent improvement for long-tail cases across both a) multi-object tracking and b) object detection tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 243, + 531, + 312 + ], + "lines": [ + { + "bbox": [ + 319, + 243, + 531, + 312 + ], + "spans": [ + { + "bbox": [ + 319, + 243, + 531, + 312 + ], + "type": "table", + "html": "
label-gtbbox-gttraj-gtcommd-gtAvg.L2↓Avg.Col↓
X0.640.23
X0.560.30
X0.590.36
X0.590.26
✓✓✓✓0.520.17
", + "image_path": "4076e661ea3e61ca583947930aa8a49ecdaa5f1faac35c43403cb30d9e0b61cd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 350 + ], + "type": "text", + "content": "Table 10. Ablation for prompt information. All ground-truth information components contributes to improved planning performance. We provide full task ablation results in the supplementary." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 366, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 451 + ], + "type": "text", + "content": "the ground truth in the prompt, we conducted experiments where each element was selectively removed in individual settings. The results, presented in Tab. 10, clearly indicate that the exclusion of any ground truth component leads to a degradation in planning ability. This underscores the necessity of incorporating the entire spectrum of ground truth information in the prompt for optimal performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 464, + 378, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 476 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": "We have introduced a novel Vision-Language-Planning (VLP) approach to enhance the capabilities of Autonomous Driving Systems (ADS). Our approach leverages both self-driving-car-centric learning paradigm (SLP) and agent-wise learning paradigm (ALP) guided by language prompts to create a comprehensive understanding of the environment. Through a series of experiments on various driving tasks, we have demonstrated the effectiveness of our VLP approach in improving perception, prediction, and planning aspects of ADS. The generalization experiments showcased the robustness of our VLP approach, proving its adaptability to new cities and long-tail cases. By extending the capabilities of ADS beyond the training environment, our VLP approach paves the way for safer and more reliable autonomous driving in real-world conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Limitations. Our experiments are currently confined to the nuScenes dataset and camera modality as baseline vision-based approaches. We will assess VLP on a broader range of datasets and sensor modalities in our future work." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14767" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 322 + ], + "type": "text", + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 446 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[7] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "text", + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 547, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 288, + 602 + ], + "type": "text", + "content": "[9] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "text", + "content": "[10] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[11] Md Zahid Hasan, Jiajing Chen, Jiyang Wang, Ameya Joshi, Senem Velipasalar, Chinmay Hegde, Anuj Sharma, and Soumik Sarkar. Vision-language models can identify distracted driver behavior from naturalistic videos. arXiv preprint arXiv:2306.10159, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[12] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "text", + "content": "[13] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1992-2008, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "text", + "content": "[14] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 239, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 547, + 294 + ], + "type": "text", + "content": "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 294, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 360 + ], + "type": "text", + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 361, + 547, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 425 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 547, + 480 + ], + "type": "text", + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 481, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 547, + 536 + ], + "type": "text", + "content": "[19] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. In European Conference on Computer Vision, pages 353-369. Springer, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "type": "text", + "content": "[21] Liunian Harold Li, Mark Yatskar, D Yin, CJ Hsieh, and KW Chang. Visualbert: A simple and performant baseline for vision and language. arxiv 2019. arXiv preprint arXiv:1908.03557. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 636, + 547, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 547, + 692 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 547, + 692 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 692, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 547, + 715 + ], + "type": "text", + "content": "[23] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end per" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14768" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "ception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 140 + ], + "type": "text", + "content": "[24] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 207, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 207, + 154 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 207, + 154 + ], + "type": "text", + "content": "[25] OpenAI. Introducing chatgpt, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 156, + 287, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 156, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 48, + 156, + 287, + 189 + ], + "type": "text", + "content": "[26] Neehar Peri, Achal Dave, Deva Ramanan, and Shu Kong. Towards long-tailed 3d detection. In Conference on Robot Learning, pages 1904–1915. PMLR, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 191, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 257 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 259, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 287, + 335 + ], + "type": "text", + "content": "[28] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 395, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 287, + 439 + ], + "type": "text", + "content": "[30] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "text", + "content": "[31] Wayve. Lingo-1: Exploring natural language for autonomous driving. https://wayve.ai/thinking/lingo - natural - language - autonomous - driving/, 2023.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "text", + "content": "[32] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. arXiv preprint arXiv:2309.16292, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "text", + "content": "[33] Shawn Xu, Lin Yang, Christopher Kelly, Marcin Sieniek, Timo Kohlberger, Martin Ma, Wei-Hung Weng, Attila Kiraly, Sahar Kazemzadeh, Zakkai Melamed, et al. Elixir: Towards a general purpose x-ray artificial intelligence system through alignment of large language models and radiology vision encoders. arXiv preprint arXiv:2308.01317, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[34] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[35] Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schuurmans. Foundation models for decision making: Problems, methods, and opportunities. arXiv preprint arXiv:2303.04129, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 183 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "text", + "content": "[36] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[37] Tianyuan Zhang, Xuanyao Chen, Yue Wang, Yilun Wang, and Hang Zhao. Mutr3d: A multi-camera tracking framework via 3d-to-2d queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4537–4546, 2022. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14769" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_content_list.json b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77be33db6b478b157f420386f1db397a33c789e1 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_content_list.json @@ -0,0 +1,1518 @@ +[ + { + "type": "text", + "text": "VMC: Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models", + "text_level": 1, + "bbox": [ + 104, + 128, + 866, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeonho Jeong $^{1*}$ Geon Yeong Park $^{2*}$ Jong Chul Ye $^{1,2}$ $^{1}$ Kim Jaechul Graduate School of AI, $^{2}$ Bio and Brain Engineering Korea Advanced Institute of Science and Technology (KAIST) * Indicates co-first authors {hyeonho.jeong, pky3436,jong.ye}@kaist.ac.kr", + "bbox": [ + 217, + 202, + 750, + 294 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0361a88fc4a77228b5db573a17f446329bd9e6e33d3a7cdaf70c88dcc415c254.jpg", + "image_caption": [ + "Figure 1. Using only a single video portraying any type of motion, our Video Motion Customization framework allows for generating a wide variety of videos characterized by the same motion but in entirely distinct contexts and better spatial/temporal resolution. 8-frame input videos are translated to 29-frame videos in different contexts while closely following the target motion. The visualized frames for the first video are at indexes 1, 9, and 17. A comprehensive view of these motions in the form of videos can be explored at our project page." + ], + "image_footnote": [], + "bbox": [ + 96, + 335, + 305, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31a55f9578dab9c43f41e6dda0a2f3cbbb42773e6d1f8ee82849c0c543e21c0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 337, + 493, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad270e66ab0145940372d36ee4cc5e4ef1c3557496971ef197312449cd91e26e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 337, + 681, + 681 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85d9be07120ca207ac10214a67a82eb6b5f906f1249ef04e6df3ed43f0608468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 337, + 872, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 750, + 313, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video diffusion models have advanced video generation significantly. However, customizing these models to generate videos with tailored motions presents a substantial challenge. In specific, they encounter hurdles in $(\\mathbf{a})$ accurately reproducing motion from a target video, and $(\\mathbf{b})$ creating diverse visual variations. For example, straightforward extensions of static image customization methods", + "bbox": [ + 73, + 794, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to video often lead to intricate entanglements of appearance and motion data. To tackle this, here we present the Video Motion Customization (VMC) framework, a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high", + "bbox": [ + 496, + 752, + 893, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9212", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our code and data can be found at https://video-motioncustomization.github.io/.", + "bbox": [ + 73, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 195, + 207, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The evolution of diffusion models [12, 26, 29] has significantly advanced Text-to-Image (T2I) generation, notably when paired with extensive text-image datasets [3, 23]. While cascaded diffusion pipelines [2, 9, 13, 25, 31, 34, 36] have extended this success to Text-to-Video (T2V) generation, current models lack the ability to replicate specific motions or generate diverse variations of the same motion with distinct visual attributes and backgrounds. Addressing this, we tackle the challenge of Motion Customization [35]—adapting pre-trained Video Diffusion Models (VDM) to produce motion-specific videos in different contexts, while maintaining the same motion patterns of target subjects.", + "bbox": [ + 75, + 220, + 468, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a few subject images for reference, appearance customization [8, 17, 21, 22, 24, 32] in generative models aims to fine-tune models to generate subject images in diverse contexts. However, these approaches, despite varying optimization objectives, commonly strive for faithful image (frame) reconstruction by minimizing the $\\ell_2$ -distance between predicted and ground-truth noise. This may lead to the entangled learning of appearance and motion.", + "bbox": [ + 75, + 417, + 468, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this, we present VMC, a new framework aimed at adapting pre-trained VDM's temporal attention layers via our proposed Motion Distillation objective. This approach utilizes residual vectors between consecutive (latent) frames to obtain the motion vectors that trace motion trajectories in the target video. Consequently, we fine-tune VDM's temporal attention layers to align the ground-truth image-space residuals with their denoised estimates, which equivalently aligns predicted and ground-truth source noise differences and motion vectors within VDM. This enables lightweight and fast one-shot training. To further facilitate the appearance-invariant motion distillation, we transform faithful text prompts into appearance-invariant prompts, e.g. \"A bird is flying above a lake in the forest\" $\\rightarrow$ \"A bird is flying\" in Fig. 1. This encourages the modules to focus on the motion information and ignore others, such as appearance, structure, background, etc. During inference, our procedure initiates by sampling key-frames using the adapted key-frame generation U-Net, followed by temporal interpolation and spatial super-resolution. To summarize, VMC makes the following key contributions:", + "bbox": [ + 75, + 537, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel fine-tuning strategy which focuses solely on temporal attention layers in the key-frame gen", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration module. This enables lightweight training (15GB vRAM) and fast training ( $< 5$ minutes).", + "bbox": [ + 509, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our knowledge, we mark a pioneering case of finetuning only the temporal attention layers in video diffusion models, without optimizing spatial self or crossattention layers, while achieving successful motion customization.", + "- We introduce a novel motion distillation objective that leverages the residual vectors between consecutive (latent) frames as motion vectors.", + "- We present the concept of appearance-invariant prompts, which further facilitates the process of motion learning when combined with our motion distillation loss." + ], + "bbox": [ + 500, + 121, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 297, + 637, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models aim to generate samples from the Gaussian noise through iterative denoising processes. Given a clean sample $\\pmb{x}_0 \\sim p_{\\mathrm{data}}(\\pmb{x})$ , the forward process is defined as a Markov chain with forward conditional densities", + "bbox": [ + 498, + 323, + 890, + 398 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\beta_ {t} \\boldsymbol {x} _ {t - 1}, (1 - \\beta_ {t}) I\\right) \\tag {1} \\\\ p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha}) I\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 402, + 890, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t\\in \\mathbb{R}^d$ is a noisy latent variable at a timestep $t$ that has the same dimension as $\\pmb{x}_0$ , and $\\beta_{t}$ denotes an increasing sequence of noise schedule where $\\alpha_{t}\\coloneqq 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t}\\coloneqq \\Pi_{i = 1}^{t}\\alpha_{i}$ . Then, the goal of diffusion model training is to obtain a residual denoiser $\\epsilon_{\\theta}$ :", + "bbox": [ + 498, + 446, + 890, + 521 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right), \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} \\left(\\boldsymbol {x} _ {0}\\right), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 526, + 895, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It can be shown that this epsilon matching in (2) is equivalent to the Denoising Score Matching (DSM [14, 28]) with different parameterization:", + "bbox": [ + 498, + 563, + 890, + 607 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon} \\left[ \\left\\| \\boldsymbol {s} _ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}) \\right\\| \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 612, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $s_{\\theta *}(\\pmb{x}_t,t)\\simeq -\\frac{\\pmb{x}_t - \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0}{1 - \\bar{\\alpha}} = -\\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t)$ . The reverse sampling from $q(\\pmb{x}_{t - 1}|\\pmb{x}_t,\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t))$ is then achieved by", + "bbox": [ + 498, + 642, + 888, + 691 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} * \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 696, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,I)$ and $\\tilde{\\beta}_t\\coloneqq \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t$ .To accelerate sampling, DDIM [27] further proposes another sampling method as follows:", + "bbox": [ + 498, + 734, + 888, + 779 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\hat {\\boldsymbol {x}} _ {0} (t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\eta^ {2} \\tilde {\\beta} _ {t} ^ {2}} \\boldsymbol {\\epsilon} _ {\\theta *} (\\boldsymbol {x} _ {t}, t) + \\eta \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 784, + 903, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\eta \\in [0,1]$ is a stochasticity parameter, and $\\hat{\\pmb{x}}_0(t)$ is the denoised estimate which can be equivalently derived using Tweedie's formula [6]:", + "bbox": [ + 498, + 821, + 888, + 867 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {0} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta *} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 872, + 890, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9213", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/31f12a97d47b50cea52bb90ce97db8d052d51b23c0a2b518d7d5316b74b8e19f.jpg", + "image_caption": [ + "(a) Training" + ], + "image_footnote": [], + "bbox": [ + 84, + 109, + 433, + 281 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/044dd9c26bb049203cec23828269b96567d442914c2b3015211ef46ce89762aa.jpg", + "image_caption": [ + "(b) Inference", + "Figure 2. Overview. The proposed Video Motion Customization (VMC) framework distills the motion trajectories from the residual between noisy latent frames, namely motion vector $\\delta \\pmb{v}_t^n$ . Specifically, we fine-tune only the temporal attention layers of the key-frame generation model by aligning the ground-truth and predicted motion vectors. After training, the customized key-frame generator is leveraged for target motion-driven video generation with new appearances context, e.g. \"A chicken is walking in a city\"." + ], + "image_footnote": [], + "bbox": [ + 468, + 109, + 893, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a text-guided Diffusion Model, the training objective is often given by:", + "bbox": [ + 75, + 371, + 468, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon , c} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) - \\boldsymbol {\\epsilon} \\| \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 411, + 468, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ represents the textual embedding. Throughout this paper, we will often omit $c$ from $\\epsilon_{\\theta}(\\pmb{x}_t,t,c)$ if it does not lead to notational ambiguity.", + "bbox": [ + 75, + 444, + 468, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Diffusion Models. Video diffusion models [11, 13, 34] further attempt to model the video data distribution. Specifically, Let $(\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}$ represents the $N$ -frame input video sequence. Then, for a given $n$ -th frame $\\pmb{v}^n\\in \\mathbb{R}^d$ , let $\\pmb{v}^{1:N}\\in \\mathbb{R}^{N\\times d}$ represents a whole video vector. Let $\\pmb{v}_t^n = \\sqrt{\\bar{\\alpha}_t}\\pmb{v}^n +\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}_t^n$ represents the $n$ -th noisy frame latent sampled from $p_t(\\pmb{v}_t^n |\\pmb{v}^n)$ , where $\\pmb{\\epsilon}_t^n\\sim \\mathcal{N}(0,I)$ . We similarly define $(\\pmb{v}_t^n)_{n\\in \\{1,\\dots ,N}$ , $\\pmb{v}_t^{1:N}$ , and $\\pmb{\\epsilon}_t^{1:N}$ . The goal of video diffusion model training is then to obtain a residual denoiser $\\pmb{\\epsilon}_{\\theta}$ with textual condition $c$ and video input that satisfies:", + "bbox": [ + 75, + 507, + 472, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {v} _ {t} ^ {1: N}, \\boldsymbol {v} ^ {1: N}, \\boldsymbol {\\epsilon} _ {t} ^ {1: N}, c} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {\\epsilon} _ {t} ^ {1: N} \\right\\| \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 684, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(\\pmb{v}_{t}^{1:N}, t, c), \\epsilon_{t}^{1:N} \\in \\mathbb{R}^{N \\times d}$ . In this work, we denote the predicted noise of $n$ -th frame as $\\epsilon_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t, c) \\in \\mathbb{R}^{d}$ .", + "bbox": [ + 75, + 717, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, contemporary video diffusion models often employ cascaded inference pipelines for high-resolution outputs. For instance, [34] initially generates a low-resolution video with strong text-video correlation, further enhancing its resolution via temporal interpolation and spatial super-resolution modules.", + "bbox": [ + 75, + 750, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In exploring video generative tasks through diffusion models, two primary approaches have emerged: leveraging foundational Video Diffusion Models (VDMs) [7, 16, 30, 35] or pre-trained Text-to-Image (T2I) models [4, 10,", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "15, 32]. To extend image diffusion models to videos, several architectural modifications are made. Typically, U-Net generative modules integrate temporal attention blocks after spatial attentions [11]. Moreover, 2D convolution layers are inflated to 3D convolution layers by altering kernels [11].", + "bbox": [ + 498, + 371, + 893, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Video Motion Customization", + "text_level": 1, + "bbox": [ + 498, + 462, + 766, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input video, our main goal is to (a) distill the motion patterns $M_*$ of target subjects, and (b) customize the input video in different contexts while maintaining the same motion patterns $M_*$ , e.g. Sharks w/ motion $M_* \\rightarrow$ Airplanes w/ motion $M_*$ , with minimal computational costs.", + "bbox": [ + 496, + 487, + 890, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose a novel video motion customization framework, namely VMC, which leverages cascaded video diffusion models with robust temporal priors. One notable aspect of the proposed framework is that we perform fine-tuning only on the key-frame generation module, also referred to as the T2V base model, within the cascaded VDMs, which guarantees computational and memory efficiency. Specifically, within the key-frame generation model, our fine-tuning process only targets the temporal attention layers. This facilitates adaptation while preserving the model's inherent capacity for generic synthesis. Notably, we freeze the subsequent frame interpolation and spatial super-resolution modules as-is (Fig. 2).", + "bbox": [ + 496, + 579, + 892, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Attention Adaptation", + "text_level": 1, + "bbox": [ + 500, + 785, + 779, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to distill the motion $M_*$ , we first propose a new objective function for temporal attention adaptation using residual cosine similarity. Our intuition is that residual vectors between consecutive frames may include information about the motion trajectories.", + "bbox": [ + 496, + 809, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $(\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}$ represents the $N$ -frame input video", + "bbox": [ + 517, + 885, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9214", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6f7da401475cafc1fcb67d25265a6ef82a6e92f4c9c1e7a3607bc2bac368ff42.jpg", + "image_caption": [ + "Figure 3. Training. The proposed framework aims to learn motion by $\\delta \\epsilon_t^n$ -alignment using (16) or (17). Note that we only fine-tune the temporal attention layers in the key-frame generation U-Net. The blue circle represents the diffusion forward process." + ], + "image_footnote": [], + "bbox": [ + 78, + 80, + 472, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sequence. As defined in Section 2, for a given noisy video latent vector $\\pmb{v}_t^{1:N}$ with $\\epsilon_t^{1:N}$ , let $\\pmb{v}_t^n$ represents the $n$ -th noisy frame latent sampled from $p_t(\\pmb{v}_t^n|\\pmb{v}^n)$ with $\\epsilon_t^n$ . We will interchangeably use $\\pmb{v}^n$ and $\\pmb{v}_0^n$ for notational simplicity. Likewise, $\\pmb{v}_t^{n + c}$ is defined as $\\pmb{v}_t^n$ , with $c > 0$ representing the fixed frame stride. Then, we define the frame residual vector at time $t \\geqslant 0$ as", + "bbox": [ + 75, + 272, + 470, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\boldsymbol {v} _ {t} ^ {n} := \\boldsymbol {v} _ {t} ^ {n + c} - \\boldsymbol {v} _ {t} ^ {n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 387, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we similarly define the epsilon residual vector $\\delta \\pmb{\\epsilon}_t^n$ . In the rest of the paper, we interchangeably use frame residual vector and motion vector.", + "bbox": [ + 75, + 414, + 468, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We expect that these motion vectors may encode information about motion patterns, where such information may vary depending on the time $t$ and its corresponding noise level. The difference vector $\\delta \\pmb{v}_t^n$ can be delineated as:", + "bbox": [ + 75, + 460, + 468, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\delta \\boldsymbol {v} _ {t} ^ {n} = \\sqrt {\\bar {\\alpha} _ {t}} \\left(\\boldsymbol {v} _ {0} ^ {n + c} - \\boldsymbol {v} _ {0} ^ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\left(\\boldsymbol {\\epsilon} _ {t} ^ {n + c} - \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right) \\tag {10} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 529, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta \\pmb{\\epsilon}_t^n$ is normally distributed with zero mean and $2I$ variance. In essence, $\\delta \\pmb{v}_t^n$ can be acquired through the following diffusion kernel:", + "bbox": [ + 75, + 577, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\delta \\boldsymbol {v} _ {0} ^ {n}\\right) = \\mathcal {N} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n}, 2 \\left(1 - \\bar {\\alpha} _ {t}\\right) I\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 632, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of this, our goal is to transfer motion information to the temporal attention layers by leveraging the motion vectors. For this, we first simulate the motion vectors using video diffusion models. Specifically, as similarly done in (6), the denoised video vector estimates $\\hat{v_0}^{1:N}(t)$ can be derived by applying Tweedie's formula:", + "bbox": [ + 75, + 659, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {v}} _ {0} ^ {1: N} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {v} _ {t} ^ {1: N} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 758, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{v}}_0^{1:N}(t)$ is an empirical Bayes optimal posterior expectation $\\mathbb{E}[\\pmb{v}_0^{1:N}|\\pmb{v}_t^{1:N}]$ . Then, the denoised motion vector estimate $\\delta \\hat{\\pmb{v}}_0^n$ can be defined in terms of $\\delta \\pmb{v}_t^n$ and $\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N},t)$ by using (12):", + "bbox": [ + 76, + 800, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\boldsymbol {v} _ {0} ^ {n} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 872, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t) \\coloneqq \\pmb{\\epsilon}_{\\theta}^{n + c}(\\pmb{v}_{t}^{1:N}, t) - \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t)$ is abbreviated as $\\delta \\pmb{\\epsilon}_{\\theta,t}^{n}$ for notational simplicity. Similarly, one can obtain ground-truth motion vector $\\delta \\pmb{v}_{0}^{n}$ by using (10):", + "bbox": [ + 498, + 89, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\boldsymbol {v} _ {0} ^ {n} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 143, + 890, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, our objective is to finetune $\\theta$ by aligning the motion vector $\\delta \\pmb{v}_0^n$ and its denoised estimate $\\delta \\hat{\\pmb{v}}_0^n (t)$ :", + "bbox": [ + 498, + 181, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) \\right], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 219, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with a loss function $\\ell_{\\mathrm{align}}:\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}$ . By using $\\ell_{2}$ -distance for $\\ell_{\\mathrm{align}}$ , this is equivalent to matching $\\delta \\pmb{\\epsilon}_{\\theta ,t}^{n}$ and $\\delta \\pmb{\\epsilon}_t^n$", + "bbox": [ + 498, + 253, + 890, + 300 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\left\\| \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n} - \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n} \\right\\| ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 306, + 890, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, aligning the ground-truth and predicted motion vectors translates into aligning epsilon residuals.", + "bbox": [ + 498, + 344, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While this objective demonstrates effective empirical performance, our additional observations indicate that using $\\ell_{\\cos}(\\delta \\pmb{\\epsilon}_t^n,\\delta \\pmb{\\epsilon}_{\\theta ,t}^n)$ may further improve the distillation, where $\\ell_{\\cos}(\\pmb {x},\\pmb {y}) = 1 - \\frac{\\langle\\pmb{x},\\pmb{y}\\rangle}{\\|\\pmb{x}\\|\\|\\pmb{y}\\|}$ for $\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{d}$ (more analysis in section 4.3). Accordingly, our optimization framework is finally defined as follows:", + "bbox": [ + 498, + 375, + 890, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\cos} \\left(\\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right) \\right]. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 478, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, the proposed optimization framework aims to maximize the residual cosine similarity between $\\delta \\epsilon_{t}^{n}$ and $\\delta \\epsilon_{\\theta ,t}^{n}$ . Hence, this optimization approach can be seamlessly applied to video diffusion models trained using epsilon-matching. Practically, we exclusively fine-tune the temporal attention layers $\\theta_{\\mathrm{TA}}\\subset \\theta$ originally designed for dynamic temporal data assimilation [32]. The frame stride remains fixed at $c = 1$ across all experiments.", + "bbox": [ + 498, + 507, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Appearance-invariant Prompts", + "text_level": 1, + "bbox": [ + 500, + 635, + 774, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In motion distillation, it is crucial to filter out disruptive variations that are unrelated to motion. These variations may include changes in appearance and background, distortions, consecutive frame inconsistencies, etc. To achieve this, we further utilize appearance-invariant prompts. Diverging from traditional generative customization frameworks [21, 22, 32, 35] that rely on text prompts that \"faithfully\" describe the input image or video during model finetuning, our framework purposely employs \"unfaithful\" text prompts during the training phase. Specifically, our approach involves the removal of background information. For instance, the text prompt 'a cat is roaring on the grass under the tree' is simplified to 'a cat is roaring' as presented in Fig. 4. This reduces background complexity as in Fig. 4a compared to Fig. 4b, facilitating the application of new appearance in motion distillation.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9215", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8b241d6b7e01941bd15eaae26211ac51d8fc4be9868648d3827b48987c32a6c.jpg", + "image_caption": [ + "Input Video Frame", + "Figure 4. Appearance-invariant Prompt. Comparison of input reconstruction with and without appearance-invariant prompt: (a) and (b) depict sampled low-resolution (64x40) keyframes. For (a), the training prompt used was \"A cat is roaring,\" while for (b), the training prompt was \"A cat is roaring on the grass under the tree.\" Our appearance-invariant prompt enables the removal of background information that can disturb motion distillation." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 218, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f5ce7399e27c6b934c263a6e0894c7c6e57d7648ab058027c7107c844d5fdd7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 220, + 92, + 341, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3999abfa2c6faff59f8015cec203b9c978af6f9c24cfd698d7c786f8e888729.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 344, + 92, + 467, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Inference Pipeline", + "text_level": 1, + "bbox": [ + 76, + 275, + 251, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once trained, in the inference phase, our process begins by computing inverted latents from the input video through DDIM inversion. Subsequently, the inverted latents are fed into the temporally fine-tuned keyframe generation model, yielding short and low-resolution keyframes. These keyframes then undergo temporal extension using the unaltered frame interpolation model. Lastly, the interpolated frames are subjected to spatial enlargement through the spatial super-resolution model. Refer to Fig. 2 for overview.", + "bbox": [ + 75, + 299, + 467, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 446, + 207, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 470, + 292, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we choose Show-1 [34] as our VDM backbone and its publicly available pre-trained weights. All experiments were conducted using a single NVIDIA RTX 6000 GPU. VMC with Show-1 demonstrates efficient resource usage, requiring only 15GB of vRAM during mixed-precision training [18], which is completed within 5 minutes. During inference, generating a single video comprising 29 frames at a resolution of $576 \\times 320$ consumes 18GB of vRAM and takes approximately 12 minutes.", + "bbox": [ + 75, + 494, + 467, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 76, + 638, + 282, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Selection. In our experiments, we draw upon a dataset that comprises 24 videos. These videos encompass a broad spectrum of motion types occurring in various contexts, encompassing vehicles, humans, birds, plants, diffusion processes, mammals, sea creatures, and more. This diversity provides a comprehensive range of motion scenarios for our assessment. Out of these 24 videos, 13 are sourced from the DAVIS dataset [19], 10 from the WebVid dataset [1], and 1 video is obtained from LAMP [33].", + "bbox": [ + 75, + 662, + 467, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Our method is compared against four contemporary baselines that integrate depth map signals into the diffusion denoising process to assimilate motion information. Notably, our approach operates without the necessity of depth maps during both training and inference, in contrast to these baseline methods.", + "bbox": [ + 75, + 809, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, VideoComposer (VC) [30] is an open-source latent-based video diffusion model tailored for compositional video generation tasks. Gen-1 [7] introduces a video diffusion architecture incorporating additional structure and content guidance for video-to-video translation. In contrast to our targeted fine-tuning of temporal attention, Tune-A-Video (TAV) [32] fine-tunes self, cross, and temporal attention layers within a pre-trained, but inflated T2I model on input videos. Control-A-Video (CAV) [5] introduces a controllable T2V diffusion model utilizing control signals and a first-frame conditioning strategy. Notably, while closely aligned with our framework, Motion Director [35] lacks available code at the time of our research.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. We offer visual comparisons of our method against four baselines in Fig. 5. The compared baselines face challenges in adapting the motion of the input video to new contexts. They exhibit difficulties in applying the overall motion, neglecting the specific background indicated in the target text (e.g., \"underwater\" or \"on the sand\"). Additionally, they face difficulties in deviating from the original shape of the subject in the input video, leading to issues like a shark-shaped airplane, an owl-shaped seagull, or preservation of the shape of the ground where a seagull is taking off. In contrast, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization, e.g. Two sharks are moving. $\\rightarrow$ Two airplanes are moving in the sky.", + "bbox": [ + 496, + 306, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. We further quantitatively demonstrate the effectiveness of our method against the baselines through automatic metrics and user study.", + "bbox": [ + 496, + 537, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Automatic Metrics. We use CLIP [20] for automatic metrics. For textual alignment, we compute the average cosine similarity between the target prompt and the generated frames. In terms of frame consistency, we obtain CLIP image features within the output video and then calculate the average cosine similarity among all pairs of video frames. For methods that generate temporally interpolated frames, we utilized the keyframe indexes to calculate the metric for a fair evaluation. To illustrate, in the case of VMC, which takes an 8-frame input and produces a 29-frame output, we considered the frames at the following indexes: 1, 5, 9, 13, 17, 21, 25, 29. As shown in Table 1, VMC outperforms baselines in both text alignment and temporal consistency. User Study. We conducted a survey involving a total of 27 participants to assess four key aspects: the preservation of motion between the input video and the generated output video, appearance diversity in the output video compared to the input video, the text alignment with the target prompt, and the overall consistency of the generated frames. The survey utilized a rating scale ranging from 1 to 5. For assessing motion preservation, we employed the question:", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9216", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f15857e49dbd158cdbd77b61822c6b0b8f6a777eadef8c6164321adc53fa0144.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 452 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b60dba8fc6b8f85670faeabc53fde866bbf28286f439420a5fffb79674dd238e.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison against state-of-the-art baselines. In contrast to other baselines, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization." + ], + "image_footnote": [], + "bbox": [ + 76, + 468, + 890, + 834 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10144d2a9770d5c3e13a84ce8b1cfd4315954555bb4effc20d947209615c3fe7.jpg", + "image_caption": [ + "Figure 6. Comparative analysis of the proposed frameworks with fine-tuning (a) temporal attention and (b) self- and cross-attention layers." + ], + "image_footnote": [], + "bbox": [ + 80, + 85, + 883, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bb1e48579f3623fc6d20ac193ecfbdf9cd855bf479e2c39445f73735fd4b1c5.jpg", + "image_caption": [ + "Figure 7. Comparative analysis of the proposed frameworks with (a) $\\ell_{cos}$ and (b) $\\ell_2$ loss functions." + ], + "image_footnote": [], + "bbox": [ + 78, + 305, + 890, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"To what extent is the motion of the input video retained in the output video?\" To evaluate appearance diversity, participants were asked: \"To what extent does the appearance of the output video avoid being restricted on the input video's appearance?\" Tab. 1 shows that our method surpasses the baselines in all four aspects.", + "bbox": [ + 75, + 515, + 468, + 606 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e81146b2bab543c598e1b3543a0a724d8b2b637fa09d93a3ac3b18872c24b939.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Text AlignmentTemporal ConsistencyMotion PreservationAppearance DiversityText AlignmentTemporal Consistency
VC0.7980.9583.453.432.963.03
Gen-10.7800.9573.463.172.872.73
TAV0.7580.9473.502.882.672.80
CAV0.7640.9522.752.452.072.00
Ours0.8010.9594.424.544.564.57
", + "bbox": [ + 78, + 611, + 465, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation using CLIP (left) and user study (right). VMC significantly outperforms the other baselines.", + "bbox": [ + 76, + 681, + 468, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 726, + 240, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on attention layers. We conducted a comparative study evaluating the performance of fine-tuning: (a) temporal attention layers and (b) self- and cross-attention layers. Illustrated in Fig. 6, both frameworks exhibit proficient motion learning capabilities. Notably, the utilization of customized temporal attention layers (a) yields smoother frame transitions, indicating the effectiveness of the optimization framework (17) in encouraging motion distillation, with a slight preference observed for customized temporal attention layers.", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This observation stems from the premise that integrating the proposed motion distillation objective (17) may autonomously and accurately embed motion information within temporal attention layers [11, 13]. This suggests a potential application of the motion distillation objective for training large-scale video diffusion models, warranting further exploration in future research endeavors.", + "bbox": [ + 496, + 515, + 890, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Choice of loss functions. In addition, we conducted a comparative analysis on distinct training loss functions in (17): the $\\ell_{2}$ -distance and $\\ell_{\\mathrm{cos}}$ as delineated in (17). As depicted in Fig. 7, the $\\delta \\epsilon$ -matching process in (15) and (17) demonstrates compatibility with generic loss functions. While both $\\ell_{2}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})$ and $\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})$ are promising objectives, the marginal superiority of $\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})$ led to its adoption for visualizations in this study.", + "bbox": [ + 496, + 640, + 892, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of adaptation. To assess the importance of temporal attention adaptation, we conducted a visualization of customized generations without temporal attention adaptation, as detailed in Section 3.1. Specifically, from our original architecture in Fig. 2, we omitted attention adaptation and performed inference by maintaining the U-Net modules in a frozen state. The outcomes depicted in Fig. 9 indicate that while DDIM inversion guides the generations", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9218", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8faab0ded7e24ddeea311f461b75fe1caccf811075bb8e63d69932a4bcaafe9.jpg", + "image_caption": [ + "Figure 8. Left: Style transfer on two videos. Right: Motion customization results on the video of \"A seagull is walking backward.\"" + ], + "image_footnote": [], + "bbox": [ + 80, + 82, + 893, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6bc297438c3589159b9cd65aaabcf9bebb62eb90e5ff771ed49148e5505c50a.jpg", + "image_caption": [ + "Figure 9. Ablation study on temporal attention adaptation. Without temporal attention adaptation, motion distillation fails." + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 464, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to mimic the motion of the input video, it alone does not ensure successful motion distillation. The observed changes in appearance and motion exhibit an entangled relationship. Consequently, this underlines the necessity of an explicit motion distillation objective to achieve consistent motion transfer, independent of any alterations in appearance.", + "bbox": [ + 75, + 657, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Additional results", + "text_level": 1, + "bbox": [ + 76, + 756, + 250, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Style Transfer. We illustrate video style transfer applications in Fig. 8-Left. We incorporate style prompts at the end of the text after applying appearance-invariant prompting (see Section 3.2). Target styles are fluidly injected while preserving the distilled motion of an input video.", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning Backward Motion. To further verify our video motion customization capabilities, we present a challenging scenario: extracting backward motion from a reversed", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "video sequence where frames are arranged in reverse order. This scenario, an exceedingly rare event in real-world videos, is highly improbable within standard training video datasets [1]. Illustrated in Fig. 8, our VMC framework showcases proficiency in learning \"a bird walking backward\" motion and generating diverse videos with distinct subjects and backgrounds. This capability not only enables leveraging the distilled motion but also offers prospects for further contextual editing.", + "bbox": [ + 496, + 306, + 892, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 460, + 617, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Video Motion Customization (VMC), addressing challenges in adapting Text-to-Video (T2V) models to generate motion-driven diverse visual customizations. Existing models struggle with accurately replicating motion from a target video and creating varied visual outputs, leading to entanglements of appearance and motion data. To overcome this, our VMC framework presents a novel one-shot tuning approach, focusing on adapting temporal attention layers within video diffusion models. This framework stands out for its efficiency in time and memory, ease of implementation, and minimal hyperparameters.", + "bbox": [ + 496, + 479, + 890, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethics Statement. Our work employs a generative model that necessitates caution due to its potential for misuse in creating deceptive content with negative societal impacts.", + "bbox": [ + 496, + 647, + 890, + 694 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "This research was supported by Field-oriented Technology Development Project for Customs Administration through National Research Foundation of Korea (NRF) funded by the Ministry of Science & ICT and Korea Customs Service (NRF-2021M3I1A1097938), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)).", + "bbox": [ + 496, + 708, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9219", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 8", + "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2", + "[4] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3", + "[5] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 5", + "[6] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2", + "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 3, 5", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2", + "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3", + "[11] J Ho, T Salimans, A Gritsenko, W Chan, M Norouzi, and DJ Fleet. Video diffusion models. arxiv 2022. arXiv preprint arXiv:2204.03458. 3, 7", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 2", + "[15] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3", + "[16] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 3", + "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 2", + "[18] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 5", + "[19] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5", + "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4", + "[22] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 4", + "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2", + "[25] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International confer" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[28] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2", + "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[30] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3, 5", + "[31] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2", + "[32] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2, 3, 4, 5", + "[33] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 5", + "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3, 5", + "[35] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2, 3, 4, 5", + "[36] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 739 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9221", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_model.json b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5dcc91577904bcb235f539b285f4306f3bd3c7a2 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.867, + 0.175 + ], + "angle": 0, + "content": "VMC: Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.751, + 0.295 + ], + "angle": 0, + "content": "Hyeonho Jeong\\(^{1*}\\) Geon Yeong Park\\(^{2*}\\) Jong Chul Ye\\(^{1,2}\\) \n\\(^{1}\\)Kim Jaechul Graduate School of AI, \\(^{2}\\)Bio and Brain Engineering Korea Advanced Institute of Science and Technology (KAIST) * Indicates co-first authors {hyeonho.jeong, pky3436,jong.ye}@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.336, + 0.307, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.338, + 0.495, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.338, + 0.682, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.338, + 0.874, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.683, + 0.895, + 0.741 + ], + "angle": 0, + "content": "Figure 1. Using only a single video portraying any type of motion, our Video Motion Customization framework allows for generating a wide variety of videos characterized by the same motion but in entirely distinct contexts and better spatial/temporal resolution. 8-frame input videos are translated to 29-frame videos in different contexts while closely following the target motion. The visualized frames for the first video are at indexes 1, 9, and 17. A comprehensive view of these motions in the form of videos can be explored at our project page." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.751, + 0.314, + 0.768 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Text-to-video diffusion models have advanced video generation significantly. However, customizing these models to generate videos with tailored motions presents a substantial challenge. In specific, they encounter hurdles in \\((\\mathbf{a})\\) accurately reproducing motion from a target video, and \\((\\mathbf{b})\\) creating diverse visual variations. For example, straightforward extensions of static image customization methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.895, + 0.891 + ], + "angle": 0, + "content": "to video often lead to intricate entanglements of appearance and motion data. To tackle this, here we present the Video Motion Customization (VMC) framework, a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our code and data can be found at https://video-motioncustomization.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.208, + 0.211 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.469, + 0.417 + ], + "angle": 0, + "content": "The evolution of diffusion models [12, 26, 29] has significantly advanced Text-to-Image (T2I) generation, notably when paired with extensive text-image datasets [3, 23]. While cascaded diffusion pipelines [2, 9, 13, 25, 31, 34, 36] have extended this success to Text-to-Video (T2V) generation, current models lack the ability to replicate specific motions or generate diverse variations of the same motion with distinct visual attributes and backgrounds. Addressing this, we tackle the challenge of Motion Customization [35]—adapting pre-trained Video Diffusion Models (VDM) to produce motion-specific videos in different contexts, while maintaining the same motion patterns of target subjects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.538 + ], + "angle": 0, + "content": "Given a few subject images for reference, appearance customization [8, 17, 21, 22, 24, 32] in generative models aims to fine-tune models to generate subject images in diverse contexts. However, these approaches, despite varying optimization objectives, commonly strive for faithful image (frame) reconstruction by minimizing the \\(\\ell_2\\)-distance between predicted and ground-truth noise. This may lead to the entangled learning of appearance and motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.869 + ], + "angle": 0, + "content": "To tackle this, we present VMC, a new framework aimed at adapting pre-trained VDM's temporal attention layers via our proposed Motion Distillation objective. This approach utilizes residual vectors between consecutive (latent) frames to obtain the motion vectors that trace motion trajectories in the target video. Consequently, we fine-tune VDM's temporal attention layers to align the ground-truth image-space residuals with their denoised estimates, which equivalently aligns predicted and ground-truth source noise differences and motion vectors within VDM. This enables lightweight and fast one-shot training. To further facilitate the appearance-invariant motion distillation, we transform faithful text prompts into appearance-invariant prompts, e.g. \"A bird is flying above a lake in the forest\" \\(\\rightarrow\\) \"A bird is flying\" in Fig. 1. This encourages the modules to focus on the motion information and ignore others, such as appearance, structure, background, etc. During inference, our procedure initiates by sampling key-frames using the adapted key-frame generation U-Net, followed by temporal interpolation and spatial super-resolution. To summarize, VMC makes the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We introduce a novel fine-tuning strategy which focuses solely on temporal attention layers in the key-frame gen" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "eration module. This enables lightweight training (15GB vRAM) and fast training (\\(< 5\\) minutes)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.891, + 0.196 + ], + "angle": 0, + "content": "- To our knowledge, we mark a pioneering case of finetuning only the temporal attention layers in video diffusion models, without optimizing spatial self or crossattention layers, while achieving successful motion customization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.197, + 0.891, + 0.242 + ], + "angle": 0, + "content": "- We introduce a novel motion distillation objective that leverages the residual vectors between consecutive (latent) frames as motion vectors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.891, + 0.287 + ], + "angle": 0, + "content": "- We present the concept of appearance-invariant prompts, which further facilitates the process of motion learning when combined with our motion distillation loss." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.891, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.638, + 0.314 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models aim to generate samples from the Gaussian noise through iterative denoising processes. Given a clean sample \\( \\pmb{x}_0 \\sim p_{\\mathrm{data}}(\\pmb{x}) \\), the forward process is defined as a Markov chain with forward conditional densities" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.403, + 0.892, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\beta_ {t} \\boldsymbol {x} _ {t - 1}, (1 - \\beta_ {t}) I\\right) \\tag {1} \\\\ p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha}) I\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.522 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_t\\in \\mathbb{R}^d\\) is a noisy latent variable at a timestep \\(t\\) that has the same dimension as \\(\\pmb{x}_0\\), and \\(\\beta_{t}\\) denotes an increasing sequence of noise schedule where \\(\\alpha_{t}\\coloneqq 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t}\\coloneqq \\Pi_{i = 1}^{t}\\alpha_{i}\\). Then, the goal of diffusion model training is to obtain a residual denoiser \\(\\epsilon_{\\theta}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.527, + 0.896, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right), \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} \\left(\\boldsymbol {x} _ {0}\\right), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.891, + 0.608 + ], + "angle": 0, + "content": "It can be shown that this epsilon matching in (2) is equivalent to the Denoising Score Matching (DSM [14, 28]) with different parameterization:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.613, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon} \\left[ \\left\\| \\boldsymbol {s} _ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}) \\right\\| \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.89, + 0.692 + ], + "angle": 0, + "content": "where \\(s_{\\theta *}(\\pmb{x}_t,t)\\simeq -\\frac{\\pmb{x}_t - \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0}{1 - \\bar{\\alpha}} = -\\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t)\\). The reverse sampling from \\(q(\\pmb{x}_{t - 1}|\\pmb{x}_t,\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t))\\) is then achieved by" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.697, + 0.891, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} * \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.89, + 0.78 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) and \\(\\tilde{\\beta}_t\\coloneqq \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t\\) .To accelerate sampling, DDIM [27] further proposes another sampling method as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.785, + 0.905, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\hat {\\boldsymbol {x}} _ {0} (t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\eta^ {2} \\tilde {\\beta} _ {t} ^ {2}} \\boldsymbol {\\epsilon} _ {\\theta *} (\\boldsymbol {x} _ {t}, t) + \\eta \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.823, + 0.89, + 0.868 + ], + "angle": 0, + "content": "where \\(\\eta \\in [0,1]\\) is a stochasticity parameter, and \\(\\hat{\\pmb{x}}_0(t)\\) is the denoised estimate which can be equivalently derived using Tweedie's formula [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.873, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {0} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta *} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9213" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.094, + 0.162, + 0.108 + ], + "angle": 0, + "content": "(a) Training" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.11, + 0.434, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.094, + 0.558, + 0.107 + ], + "angle": 0, + "content": "(b) Inference" + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.111, + 0.895, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Overview. The proposed Video Motion Customization (VMC) framework distills the motion trajectories from the residual between noisy latent frames, namely motion vector \\(\\delta \\pmb{v}_t^n\\). Specifically, we fine-tune only the temporal attention layers of the key-frame generation model by aligning the ground-truth and predicted motion vectors. After training, the customized key-frame generator is leveraged for target motion-driven video generation with new appearances context, e.g. \"A chicken is walking in a city\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.403 + ], + "angle": 0, + "content": "For a text-guided Diffusion Model, the training objective is often given by:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.412, + 0.47, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon , c} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) - \\boldsymbol {\\epsilon} \\| \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.492 + ], + "angle": 0, + "content": "where \\(c\\) represents the textual embedding. Throughout this paper, we will often omit \\(c\\) from \\(\\epsilon_{\\theta}(\\pmb{x}_t,t,c)\\) if it does not lead to notational ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.473, + 0.676 + ], + "angle": 0, + "content": "Video Diffusion Models. Video diffusion models [11, 13, 34] further attempt to model the video data distribution. Specifically, Let \\((\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}\\) represents the \\(N\\)-frame input video sequence. Then, for a given \\(n\\)-th frame \\(\\pmb{v}^n\\in \\mathbb{R}^d\\), let \\(\\pmb{v}^{1:N}\\in \\mathbb{R}^{N\\times d}\\) represents a whole video vector. Let \\(\\pmb{v}_t^n = \\sqrt{\\bar{\\alpha}_t}\\pmb{v}^n +\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}_t^n\\) represents the \\(n\\)-th noisy frame latent sampled from \\(p_t(\\pmb{v}_t^n |\\pmb{v}^n)\\), where \\(\\pmb{\\epsilon}_t^n\\sim \\mathcal{N}(0,I)\\). We similarly define \\((\\pmb{v}_t^n)_{n\\in \\{1,\\dots ,N}\\), \\(\\pmb{v}_t^{1:N}\\), and \\(\\pmb{\\epsilon}_t^{1:N}\\). The goal of video diffusion model training is then to obtain a residual denoiser \\(\\pmb{\\epsilon}_{\\theta}\\) with textual condition \\(c\\) and video input that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.685, + 0.47, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {v} _ {t} ^ {1: N}, \\boldsymbol {v} ^ {1: N}, \\boldsymbol {\\epsilon} _ {t} ^ {1: N}, c} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {\\epsilon} _ {t} ^ {1: N} \\right\\| \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.751 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(\\pmb{v}_{t}^{1:N}, t, c), \\epsilon_{t}^{1:N} \\in \\mathbb{R}^{N \\times d}\\). In this work, we denote the predicted noise of \\(n\\)-th frame as \\(\\epsilon_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t, c) \\in \\mathbb{R}^{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.84 + ], + "angle": 0, + "content": "In practice, contemporary video diffusion models often employ cascaded inference pipelines for high-resolution outputs. For instance, [34] initially generates a low-resolution video with strong text-video correlation, further enhancing its resolution via temporal interpolation and spatial super-resolution modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In exploring video generative tasks through diffusion models, two primary approaches have emerged: leveraging foundational Video Diffusion Models (VDMs) [7, 16, 30, 35] or pre-trained Text-to-Image (T2I) models [4, 10," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.372, + 0.895, + 0.449 + ], + "angle": 0, + "content": "15, 32]. To extend image diffusion models to videos, several architectural modifications are made. Typically, U-Net generative modules integrate temporal attention blocks after spatial attentions [11]. Moreover, 2D convolution layers are inflated to 3D convolution layers by altering kernels [11]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.463, + 0.767, + 0.479 + ], + "angle": 0, + "content": "3. Video Motion Customization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.578 + ], + "angle": 0, + "content": "Given an input video, our main goal is to (a) distill the motion patterns \\( M_* \\) of target subjects, and (b) customize the input video in different contexts while maintaining the same motion patterns \\( M_* \\), e.g. Sharks w/ motion \\( M_* \\rightarrow \\) Airplanes w/ motion \\( M_* \\), with minimal computational costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.777 + ], + "angle": 0, + "content": "To this end, we propose a novel video motion customization framework, namely VMC, which leverages cascaded video diffusion models with robust temporal priors. One notable aspect of the proposed framework is that we perform fine-tuning only on the key-frame generation module, also referred to as the T2V base model, within the cascaded VDMs, which guarantees computational and memory efficiency. Specifically, within the key-frame generation model, our fine-tuning process only targets the temporal attention layers. This facilitates adaptation while preserving the model's inherent capacity for generic synthesis. Notably, we freeze the subsequent frame interpolation and spatial super-resolution modules as-is (Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.78, + 0.803 + ], + "angle": 0, + "content": "3.1. Temporal Attention Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.885 + ], + "angle": 0, + "content": "In order to distill the motion \\( M_* \\), we first propose a new objective function for temporal attention adaptation using residual cosine similarity. Our intuition is that residual vectors between consecutive frames may include information about the motion trajectories." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.904 + ], + "angle": 0, + "content": "Let \\((\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}\\) represents the \\(N\\) -frame input video" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.082, + 0.473, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.192, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Figure 3. Training. The proposed framework aims to learn motion by \\(\\delta \\epsilon_t^n\\)-alignment using (16) or (17). Note that we only fine-tune the temporal attention layers in the key-frame generation U-Net. The blue circle represents the diffusion forward process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.471, + 0.378 + ], + "angle": 0, + "content": "sequence. As defined in Section 2, for a given noisy video latent vector \\(\\pmb{v}_t^{1:N}\\) with \\(\\epsilon_t^{1:N}\\), let \\(\\pmb{v}_t^n\\) represents the \\(n\\)-th noisy frame latent sampled from \\(p_t(\\pmb{v}_t^n|\\pmb{v}^n)\\) with \\(\\epsilon_t^n\\). We will interchangeably use \\(\\pmb{v}^n\\) and \\(\\pmb{v}_0^n\\) for notational simplicity. Likewise, \\(\\pmb{v}_t^{n + c}\\) is defined as \\(\\pmb{v}_t^n\\), with \\(c > 0\\) representing the fixed frame stride. Then, we define the frame residual vector at time \\(t \\geqslant 0\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.388, + 0.469, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\delta \\boldsymbol {v} _ {t} ^ {n} := \\boldsymbol {v} _ {t} ^ {n + c} - \\boldsymbol {v} _ {t} ^ {n}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.459 + ], + "angle": 0, + "content": "where we similarly define the epsilon residual vector \\(\\delta \\pmb{\\epsilon}_t^n\\). In the rest of the paper, we interchangeably use frame residual vector and motion vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.469, + 0.521 + ], + "angle": 0, + "content": "We expect that these motion vectors may encode information about motion patterns, where such information may vary depending on the time \\( t \\) and its corresponding noise level. The difference vector \\( \\delta \\pmb{v}_t^n \\) can be delineated as:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.53, + 0.469, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\delta \\boldsymbol {v} _ {t} ^ {n} = \\sqrt {\\bar {\\alpha} _ {t}} \\left(\\boldsymbol {v} _ {0} ^ {n + c} - \\boldsymbol {v} _ {0} ^ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\left(\\boldsymbol {\\epsilon} _ {t} ^ {n + c} - \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right) \\tag {10} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.623 + ], + "angle": 0, + "content": "where \\(\\delta \\pmb{\\epsilon}_t^n\\) is normally distributed with zero mean and \\(2I\\) variance. In essence, \\(\\delta \\pmb{v}_t^n\\) can be acquired through the following diffusion kernel:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.633, + 0.469, + 0.651 + ], + "angle": 0, + "content": "\\[\np \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\delta \\boldsymbol {v} _ {0} ^ {n}\\right) = \\mathcal {N} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n}, 2 \\left(1 - \\bar {\\alpha} _ {t}\\right) I\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.75 + ], + "angle": 0, + "content": "In light of this, our goal is to transfer motion information to the temporal attention layers by leveraging the motion vectors. For this, we first simulate the motion vectors using video diffusion models. Specifically, as similarly done in (6), the denoised video vector estimates \\(\\hat{v_0}^{1:N}(t)\\) can be derived by applying Tweedie's formula:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.759, + 0.469, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {v}} _ {0} ^ {1: N} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {v} _ {t} ^ {1: N} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.801, + 0.469, + 0.864 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{v}}_0^{1:N}(t)\\) is an empirical Bayes optimal posterior expectation \\(\\mathbb{E}[\\pmb{v}_0^{1:N}|\\pmb{v}_t^{1:N}]\\). Then, the denoised motion vector estimate \\(\\delta \\hat{\\pmb{v}}_0^n\\) can be defined in terms of \\(\\delta \\pmb{v}_t^n\\) and \\(\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N},t)\\) by using (12):" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.873, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\delta \\boldsymbol {v} _ {0} ^ {n} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.138 + ], + "angle": 0, + "content": "where \\(\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t) \\coloneqq \\pmb{\\epsilon}_{\\theta}^{n + c}(\\pmb{v}_{t}^{1:N}, t) - \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t)\\) is abbreviated as \\(\\delta \\pmb{\\epsilon}_{\\theta,t}^{n}\\) for notational simplicity. Similarly, one can obtain ground-truth motion vector \\(\\delta \\pmb{v}_{0}^{n}\\) by using (10):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.144, + 0.892, + 0.176 + ], + "angle": 0, + "content": "\\[\n\\delta \\boldsymbol {v} _ {0} ^ {n} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.182, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Then, our objective is to finetune \\(\\theta\\) by aligning the motion vector \\(\\delta \\pmb{v}_0^n\\) and its denoised estimate \\(\\delta \\hat{\\pmb{v}}_0^n (t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.22, + 0.892, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) \\right], \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.301 + ], + "angle": 0, + "content": "with a loss function \\(\\ell_{\\mathrm{align}}:\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}\\). By using \\(\\ell_{2}\\)-distance for \\(\\ell_{\\mathrm{align}}\\), this is equivalent to matching \\(\\delta \\pmb{\\epsilon}_{\\theta ,t}^{n}\\) and \\(\\delta \\pmb{\\epsilon}_t^n\\)" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.308, + 0.892, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\left\\| \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n} - \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n} \\right\\| ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.345, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Notably, aligning the ground-truth and predicted motion vectors translates into aligning epsilon residuals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.47 + ], + "angle": 0, + "content": "While this objective demonstrates effective empirical performance, our additional observations indicate that using \\(\\ell_{\\cos}(\\delta \\pmb{\\epsilon}_t^n,\\delta \\pmb{\\epsilon}_{\\theta ,t}^n)\\) may further improve the distillation, where \\(\\ell_{\\cos}(\\pmb {x},\\pmb {y}) = 1 - \\frac{\\langle\\pmb{x},\\pmb{y}\\rangle}{\\|\\pmb{x}\\|\\|\\pmb{y}\\|}\\) for \\(\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{d}\\) (more analysis in section 4.3). Accordingly, our optimization framework is finally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.479, + 0.892, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\cos} \\left(\\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right) \\right]. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.628 + ], + "angle": 0, + "content": "In other words, the proposed optimization framework aims to maximize the residual cosine similarity between \\(\\delta \\epsilon_{t}^{n}\\) and \\(\\delta \\epsilon_{\\theta ,t}^{n}\\). Hence, this optimization approach can be seamlessly applied to video diffusion models trained using epsilon-matching. Practically, we exclusively fine-tune the temporal attention layers \\(\\theta_{\\mathrm{TA}}\\subset \\theta\\) originally designed for dynamic temporal data assimilation [32]. The frame stride remains fixed at \\(c = 1\\) across all experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.776, + 0.653 + ], + "angle": 0, + "content": "3.2. Appearance-invariant Prompts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In motion distillation, it is crucial to filter out disruptive variations that are unrelated to motion. These variations may include changes in appearance and background, distortions, consecutive frame inconsistencies, etc. To achieve this, we further utilize appearance-invariant prompts. Diverging from traditional generative customization frameworks [21, 22, 32, 35] that rely on text prompts that \"faithfully\" describe the input image or video during model finetuning, our framework purposely employs \"unfaithful\" text prompts during the training phase. Specifically, our approach involves the removal of background information. For instance, the text prompt 'a cat is roaring on the grass under the tree' is simplified to 'a cat is roaring' as presented in Fig. 4. This reduces background complexity as in Fig. 4a compared to Fig. 4b, facilitating the application of new appearance in motion distillation." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9215" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.083, + 0.191, + 0.093 + ], + "angle": 0, + "content": "Input Video Frame" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.219, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.084, + 0.288, + 0.092 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.093, + 0.343, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.084, + 0.413, + 0.092 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.093, + 0.468, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.156, + 0.47, + 0.252 + ], + "angle": 0, + "content": "Figure 4. Appearance-invariant Prompt. Comparison of input reconstruction with and without appearance-invariant prompt: (a) and (b) depict sampled low-resolution (64x40) keyframes. For (a), the training prompt used was \"A cat is roaring,\" while for (b), the training prompt was \"A cat is roaring on the grass under the tree.\" Our appearance-invariant prompt enables the removal of background information that can disturb motion distillation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.276, + 0.252, + 0.292 + ], + "angle": 0, + "content": "3.3. Inference Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.468, + 0.435 + ], + "angle": 0, + "content": "Once trained, in the inference phase, our process begins by computing inverted latents from the input video through DDIM inversion. Subsequently, the inverted latents are fed into the temporally fine-tuned keyframe generation model, yielding short and low-resolution keyframes. These keyframes then undergo temporal extension using the unaltered frame interpolation model. Lastly, the interpolated frames are subjected to spatial enlargement through the spatial super-resolution model. Refer to Fig. 2 for overview." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.208, + 0.464 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.293, + 0.488 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.468, + 0.631 + ], + "angle": 0, + "content": "In our experiments, we choose Show-1 [34] as our VDM backbone and its publicly available pre-trained weights. All experiments were conducted using a single NVIDIA RTX 6000 GPU. VMC with Show-1 demonstrates efficient resource usage, requiring only 15GB of vRAM during mixed-precision training [18], which is completed within 5 minutes. During inference, generating a single video comprising 29 frames at a resolution of \\(576 \\times 320\\) consumes 18GB of vRAM and takes approximately 12 minutes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.284, + 0.655 + ], + "angle": 0, + "content": "4.2. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.468, + 0.799 + ], + "angle": 0, + "content": "Dataset Selection. In our experiments, we draw upon a dataset that comprises 24 videos. These videos encompass a broad spectrum of motion types occurring in various contexts, encompassing vehicles, humans, birds, plants, diffusion processes, mammals, sea creatures, and more. This diversity provides a comprehensive range of motion scenarios for our assessment. Out of these 24 videos, 13 are sourced from the DAVIS dataset [19], 10 from the WebVid dataset [1], and 1 video is obtained from LAMP [33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Baselines. Our method is compared against four contemporary baselines that integrate depth map signals into the diffusion denoising process to assimilate motion information. Notably, our approach operates without the necessity of depth maps during both training and inference, in contrast to these baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Specifically, VideoComposer (VC) [30] is an open-source latent-based video diffusion model tailored for compositional video generation tasks. Gen-1 [7] introduces a video diffusion architecture incorporating additional structure and content guidance for video-to-video translation. In contrast to our targeted fine-tuning of temporal attention, Tune-A-Video (TAV) [32] fine-tunes self, cross, and temporal attention layers within a pre-trained, but inflated T2I model on input videos. Control-A-Video (CAV) [5] introduces a controllable T2V diffusion model utilizing control signals and a first-frame conditioning strategy. Notably, while closely aligned with our framework, Motion Director [35] lacks available code at the time of our research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Qualitative Results. We offer visual comparisons of our method against four baselines in Fig. 5. The compared baselines face challenges in adapting the motion of the input video to new contexts. They exhibit difficulties in applying the overall motion, neglecting the specific background indicated in the target text (e.g., \"underwater\" or \"on the sand\"). Additionally, they face difficulties in deviating from the original shape of the subject in the input video, leading to issues like a shark-shaped airplane, an owl-shaped seagull, or preservation of the shape of the ground where a seagull is taking off. In contrast, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization, e.g. Two sharks are moving. \\(\\rightarrow\\) Two airplanes are moving in the sky." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.891, + 0.583 + ], + "angle": 0, + "content": "Quantitative Results. We further quantitatively demonstrate the effectiveness of our method against the baselines through automatic metrics and user study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Automatic Metrics. We use CLIP [20] for automatic metrics. For textual alignment, we compute the average cosine similarity between the target prompt and the generated frames. In terms of frame consistency, we obtain CLIP image features within the output video and then calculate the average cosine similarity among all pairs of video frames. For methods that generate temporally interpolated frames, we utilized the keyframe indexes to calculate the metric for a fair evaluation. To illustrate, in the case of VMC, which takes an 8-frame input and produces a 29-frame output, we considered the frames at the following indexes: 1, 5, 9, 13, 17, 21, 25, 29. As shown in Table 1, VMC outperforms baselines in both text alignment and temporal consistency. User Study. We conducted a survey involving a total of 27 participants to assess four key aspects: the preservation of motion between the input video and the generated output video, appearance diversity in the output video compared to the input video, the text alignment with the target prompt, and the overall consistency of the generated frames. The survey utilized a rating scale ranging from 1 to 5. For assessing motion preservation, we employed the question:" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.469, + 0.891, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.837, + 0.895, + 0.865 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison against state-of-the-art baselines. In contrast to other baselines, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.087, + 0.885, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 6. Comparative analysis of the proposed frameworks with fine-tuning (a) temporal attention and (b) self- and cross-attention layers." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.306, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.475, + 0.778, + 0.49 + ], + "angle": 0, + "content": "Figure 7. Comparative analysis of the proposed frameworks with (a) \\(\\ell_{cos}\\) and (b) \\(\\ell_2\\) loss functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\"To what extent is the motion of the input video retained in the output video?\" To evaluate appearance diversity, participants were asked: \"To what extent does the appearance of the output video avoid being restricted on the input video's appearance?\" Tab. 1 shows that our method surpasses the baselines in all four aspects." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.612, + 0.467, + 0.681 + ], + "angle": 0, + "content": "
Text AlignmentTemporal ConsistencyMotion PreservationAppearance DiversityText AlignmentTemporal Consistency
VC0.7980.9583.453.432.963.03
Gen-10.7800.9573.463.172.872.73
TAV0.7580.9473.502.882.672.80
CAV0.7640.9522.752.452.072.00
Ours0.8010.9594.424.544.564.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.682, + 0.469, + 0.71 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation using CLIP (left) and user study (right). VMC significantly outperforms the other baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.241, + 0.741 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparisons on attention layers. We conducted a comparative study evaluating the performance of fine-tuning: (a) temporal attention layers and (b) self- and cross-attention layers. Illustrated in Fig. 6, both frameworks exhibit proficient motion learning capabilities. Notably, the utilization of customized temporal attention layers (a) yields smoother frame transitions, indicating the effectiveness of the optimization framework (17) in encouraging motion distillation, with a slight preference observed for customized temporal attention layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.622 + ], + "angle": 0, + "content": "This observation stems from the premise that integrating the proposed motion distillation objective (17) may autonomously and accurately embed motion information within temporal attention layers [11, 13]. This suggests a potential application of the motion distillation objective for training large-scale video diffusion models, warranting further exploration in future research endeavors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.762 + ], + "angle": 0, + "content": "Choice of loss functions. In addition, we conducted a comparative analysis on distinct training loss functions in (17): the \\(\\ell_{2}\\)-distance and \\(\\ell_{\\mathrm{cos}}\\) as delineated in (17). As depicted in Fig. 7, the \\(\\delta \\epsilon\\)-matching process in (15) and (17) demonstrates compatibility with generic loss functions. While both \\(\\ell_{2}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})\\) and \\(\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})\\) are promising objectives, the marginal superiority of \\(\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})\\) led to its adoption for visualizations in this study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Importance of adaptation. To assess the importance of temporal attention adaptation, we conducted a visualization of customized generations without temporal attention adaptation, as detailed in Section 3.1. Specifically, from our original architecture in Fig. 2, we omitted attention adaptation and performed inference by maintaining the U-Net modules in a frozen state. The outcomes depicted in Fig. 9 indicate that while DDIM inversion guides the generations" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.083, + 0.894, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.267, + 0.871, + 0.281 + ], + "angle": 0, + "content": "Figure 8. Left: Style transfer on two videos. Right: Motion customization results on the video of \"A seagull is walking backward.\"" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.303, + 0.465, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.603, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Figure 9. Ablation study on temporal attention adaptation. Without temporal attention adaptation, motion distillation fails." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.749 + ], + "angle": 0, + "content": "to mimic the motion of the input video, it alone does not ensure successful motion distillation. The observed changes in appearance and motion exhibit an entangled relationship. Consequently, this underlines the necessity of an explicit motion distillation objective to achieve consistent motion transfer, independent of any alterations in appearance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.251, + 0.772 + ], + "angle": 0, + "content": "4.4. Additional results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Video Style Transfer. We illustrate video style transfer applications in Fig. 8-Left. We incorporate style prompts at the end of the text after applying appearance-invariant prompting (see Section 3.2). Target styles are fluidly injected while preserving the distilled motion of an input video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Learning Backward Motion. To further verify our video motion customization capabilities, we present a challenging scenario: extracting backward motion from a reversed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.445 + ], + "angle": 0, + "content": "video sequence where frames are arranged in reverse order. This scenario, an exceedingly rare event in real-world videos, is highly improbable within standard training video datasets [1]. Illustrated in Fig. 8, our VMC framework showcases proficiency in learning \"a bird walking backward\" motion and generating diverse videos with distinct subjects and backgrounds. This capability not only enables leveraging the distilled motion but also offers prospects for further contextual editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.619, + 0.477 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.648 + ], + "angle": 0, + "content": "This paper introduces Video Motion Customization (VMC), addressing challenges in adapting Text-to-Video (T2V) models to generate motion-driven diverse visual customizations. Existing models struggle with accurately replicating motion from a target video and creating varied visual outputs, leading to entanglements of appearance and motion data. To overcome this, our VMC framework presents a novel one-shot tuning approach, focusing on adapting temporal attention layers within video diffusion models. This framework stands out for its efficiency in time and memory, ease of implementation, and minimal hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Ethics Statement. Our work employs a generative model that necessitates caution due to its potential for misuse in creating deceptive content with negative societal impacts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.709, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This research was supported by Field-oriented Technology Development Project for Customs Administration through National Research Foundation of Korea (NRF) funded by the Ministry of Science & ICT and Korea Customs Service (NRF-2021M3I1A1097938), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST))." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9219" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[4] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[5] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[6] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[11] J Ho, T Salimans, A Gritsenko, W Chan, M Norouzi, and DJ Fleet. Video diffusion models. arxiv 2022. arXiv preprint arXiv:2204.03458. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[22] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[25] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International confer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ence on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[28] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[30] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[31] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[32] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[33] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[35] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[36] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9221" + } + ] +] \ No newline at end of file diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_origin.pdf b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ecc22a00ff24f514191bc4e6765496f2ce7f4ac8 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/ffbb1e4a-8f04-44a4-a263-284d8b8f46c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fe6a9e74f3893cb99321911c7ad5e1a24dcdbad65b2e7dce7f7260c8fa064c4 +size 5234303 diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/full.md b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bbc0b887b9b4b98a3b597d254a05e89f9a883090 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/full.md @@ -0,0 +1,307 @@ +# VMC: Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models + +Hyeonho Jeong $^{1*}$ Geon Yeong Park $^{2*}$ Jong Chul Ye $^{1,2}$ $^{1}$ Kim Jaechul Graduate School of AI, $^{2}$ Bio and Brain Engineering Korea Advanced Institute of Science and Technology (KAIST) * Indicates co-first authors {hyeonho.jeong, pky3436,jong.ye}@kaist.ac.kr + +![](images/0361a88fc4a77228b5db573a17f446329bd9e6e33d3a7cdaf70c88dcc415c254.jpg) +Figure 1. Using only a single video portraying any type of motion, our Video Motion Customization framework allows for generating a wide variety of videos characterized by the same motion but in entirely distinct contexts and better spatial/temporal resolution. 8-frame input videos are translated to 29-frame videos in different contexts while closely following the target motion. The visualized frames for the first video are at indexes 1, 9, and 17. A comprehensive view of these motions in the form of videos can be explored at our project page. + +![](images/31a55f9578dab9c43f41e6dda0a2f3cbbb42773e6d1f8ee82849c0c543e21c0a.jpg) + +![](images/ad270e66ab0145940372d36ee4cc5e4ef1c3557496971ef197312449cd91e26e.jpg) + +![](images/85d9be07120ca207ac10214a67a82eb6b5f906f1249ef04e6df3ed43f0608468.jpg) + +# Abstract + +Text-to-video diffusion models have advanced video generation significantly. However, customizing these models to generate videos with tailored motions presents a substantial challenge. In specific, they encounter hurdles in $(\mathbf{a})$ accurately reproducing motion from a target video, and $(\mathbf{b})$ creating diverse visual variations. For example, straightforward extensions of static image customization methods + +to video often lead to intricate entanglements of appearance and motion data. To tackle this, here we present the Video Motion Customization (VMC) framework, a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high + +frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our code and data can be found at https://video-motioncustomization.github.io/. + +# 1. Introduction + +The evolution of diffusion models [12, 26, 29] has significantly advanced Text-to-Image (T2I) generation, notably when paired with extensive text-image datasets [3, 23]. While cascaded diffusion pipelines [2, 9, 13, 25, 31, 34, 36] have extended this success to Text-to-Video (T2V) generation, current models lack the ability to replicate specific motions or generate diverse variations of the same motion with distinct visual attributes and backgrounds. Addressing this, we tackle the challenge of Motion Customization [35]—adapting pre-trained Video Diffusion Models (VDM) to produce motion-specific videos in different contexts, while maintaining the same motion patterns of target subjects. + +Given a few subject images for reference, appearance customization [8, 17, 21, 22, 24, 32] in generative models aims to fine-tune models to generate subject images in diverse contexts. However, these approaches, despite varying optimization objectives, commonly strive for faithful image (frame) reconstruction by minimizing the $\ell_2$ -distance between predicted and ground-truth noise. This may lead to the entangled learning of appearance and motion. + +To tackle this, we present VMC, a new framework aimed at adapting pre-trained VDM's temporal attention layers via our proposed Motion Distillation objective. This approach utilizes residual vectors between consecutive (latent) frames to obtain the motion vectors that trace motion trajectories in the target video. Consequently, we fine-tune VDM's temporal attention layers to align the ground-truth image-space residuals with their denoised estimates, which equivalently aligns predicted and ground-truth source noise differences and motion vectors within VDM. This enables lightweight and fast one-shot training. To further facilitate the appearance-invariant motion distillation, we transform faithful text prompts into appearance-invariant prompts, e.g. "A bird is flying above a lake in the forest" $\rightarrow$ "A bird is flying" in Fig. 1. This encourages the modules to focus on the motion information and ignore others, such as appearance, structure, background, etc. During inference, our procedure initiates by sampling key-frames using the adapted key-frame generation U-Net, followed by temporal interpolation and spatial super-resolution. To summarize, VMC makes the following key contributions: + +- We introduce a novel fine-tuning strategy which focuses solely on temporal attention layers in the key-frame gen + +eration module. This enables lightweight training (15GB vRAM) and fast training ( $< 5$ minutes). + +- To our knowledge, we mark a pioneering case of finetuning only the temporal attention layers in video diffusion models, without optimizing spatial self or crossattention layers, while achieving successful motion customization. +- We introduce a novel motion distillation objective that leverages the residual vectors between consecutive (latent) frames as motion vectors. +- We present the concept of appearance-invariant prompts, which further facilitates the process of motion learning when combined with our motion distillation loss. + +# 2. Preliminaries + +Diffusion Models. Diffusion models aim to generate samples from the Gaussian noise through iterative denoising processes. Given a clean sample $\pmb{x}_0 \sim p_{\mathrm{data}}(\pmb{x})$ , the forward process is defined as a Markov chain with forward conditional densities + +$$ +\begin{array}{l} p \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t} \mid \beta_ {t} \boldsymbol {x} _ {t - 1}, (1 - \beta_ {t}) I\right) \tag {1} \\ p _ {t} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t} \mid \sqrt {\bar {\alpha}} \boldsymbol {x} _ {0}, (1 - \bar {\alpha}) I\right), \\ \end{array} +$$ + +where $\pmb{x}_t\in \mathbb{R}^d$ is a noisy latent variable at a timestep $t$ that has the same dimension as $\pmb{x}_0$ , and $\beta_{t}$ denotes an increasing sequence of noise schedule where $\alpha_{t}\coloneqq 1 - \beta_{t}$ and $\bar{\alpha}_{t}\coloneqq \Pi_{i = 1}^{t}\alpha_{i}$ . Then, the goal of diffusion model training is to obtain a residual denoiser $\epsilon_{\theta}$ : + +$$ +\min _ {\theta} \mathbb {E} _ {\boldsymbol {x} _ {t} \sim p _ {t} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right), \boldsymbol {x} _ {0} \sim p _ {\text {d a t a}} \left(\boldsymbol {x} _ {0}\right), \epsilon \sim \mathcal {N} (0, I)} \left[ \| \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t) - \boldsymbol {\epsilon} \| \right]. \tag {2} +$$ + +It can be shown that this epsilon matching in (2) is equivalent to the Denoising Score Matching (DSM [14, 28]) with different parameterization: + +$$ +\min _ {\theta} \mathbb {E} _ {\boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}, \epsilon} \left[ \left\| \boldsymbol {s} _ {\theta} ^ {t} (\boldsymbol {x} _ {t}) - \nabla_ {\boldsymbol {x} _ {t}} \log p _ {t} (\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}) \right\| \right], \tag {3} +$$ + +where $s_{\theta *}(\pmb{x}_t,t)\simeq -\frac{\pmb{x}_t - \sqrt{\bar{\alpha}_t}\pmb{x}_0}{1 - \bar{\alpha}} = -\frac{1}{\sqrt{1 - \bar{\alpha}_t}}\pmb{\epsilon}_{\theta *}(\pmb{x}_t,t)$ . The reverse sampling from $q(\pmb{x}_{t - 1}|\pmb{x}_t,\pmb{\epsilon}_{\theta *}(\pmb{x}_t,t))$ is then achieved by + +$$ +\boldsymbol {x} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} * \left(\boldsymbol {x} _ {t}, t\right)\right) + \tilde {\beta} _ {t} \boldsymbol {\epsilon}, \tag {4} +$$ + +where $\epsilon \sim \mathcal{N}(0,I)$ and $\tilde{\beta}_t\coloneqq \frac{1 - \bar{\alpha}_{t - 1}}{1 - \bar{\alpha}_t}\beta_t$ .To accelerate sampling, DDIM [27] further proposes another sampling method as follows: + +$$ +\boldsymbol {x} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \hat {\boldsymbol {x}} _ {0} (t) + \sqrt {1 - \bar {\alpha} _ {t - 1} - \eta^ {2} \tilde {\beta} _ {t} ^ {2}} \boldsymbol {\epsilon} _ {\theta *} (\boldsymbol {x} _ {t}, t) + \eta \tilde {\beta} _ {t} \boldsymbol {\epsilon}, \tag {5} +$$ + +where $\eta \in [0,1]$ is a stochasticity parameter, and $\hat{\pmb{x}}_0(t)$ is the denoised estimate which can be equivalently derived using Tweedie's formula [6]: + +$$ +\hat {\boldsymbol {x}} _ {0} (t) := \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\boldsymbol {x} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {\theta *} \left(\boldsymbol {x} _ {t}, t\right)\right). \tag {6} +$$ + +![](images/31f12a97d47b50cea52bb90ce97db8d052d51b23c0a2b518d7d5316b74b8e19f.jpg) +(a) Training + +![](images/044dd9c26bb049203cec23828269b96567d442914c2b3015211ef46ce89762aa.jpg) +(b) Inference +Figure 2. Overview. The proposed Video Motion Customization (VMC) framework distills the motion trajectories from the residual between noisy latent frames, namely motion vector $\delta \pmb{v}_t^n$ . Specifically, we fine-tune only the temporal attention layers of the key-frame generation model by aligning the ground-truth and predicted motion vectors. After training, the customized key-frame generator is leveraged for target motion-driven video generation with new appearances context, e.g. "A chicken is walking in a city". + +For a text-guided Diffusion Model, the training objective is often given by: + +$$ +\min _ {\theta} \mathbb {E} _ {\boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}, \epsilon , c} \left[ \| \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t, c) - \boldsymbol {\epsilon} \| \right], \tag {7} +$$ + +where $c$ represents the textual embedding. Throughout this paper, we will often omit $c$ from $\epsilon_{\theta}(\pmb{x}_t,t,c)$ if it does not lead to notational ambiguity. + +Video Diffusion Models. Video diffusion models [11, 13, 34] further attempt to model the video data distribution. Specifically, Let $(\pmb{v}^n)_{n\in \{1,\dots ,N\}}$ represents the $N$ -frame input video sequence. Then, for a given $n$ -th frame $\pmb{v}^n\in \mathbb{R}^d$ , let $\pmb{v}^{1:N}\in \mathbb{R}^{N\times d}$ represents a whole video vector. Let $\pmb{v}_t^n = \sqrt{\bar{\alpha}_t}\pmb{v}^n +\sqrt{1 - \bar{\alpha}_t}\pmb{\epsilon}_t^n$ represents the $n$ -th noisy frame latent sampled from $p_t(\pmb{v}_t^n |\pmb{v}^n)$ , where $\pmb{\epsilon}_t^n\sim \mathcal{N}(0,I)$ . We similarly define $(\pmb{v}_t^n)_{n\in \{1,\dots ,N}$ , $\pmb{v}_t^{1:N}$ , and $\pmb{\epsilon}_t^{1:N}$ . The goal of video diffusion model training is then to obtain a residual denoiser $\pmb{\epsilon}_{\theta}$ with textual condition $c$ and video input that satisfies: + +$$ +\min _ {\theta} \mathbb {E} _ {\boldsymbol {v} _ {t} ^ {1: N}, \boldsymbol {v} ^ {1: N}, \boldsymbol {\epsilon} _ {t} ^ {1: N}, c} \left[ \left\| \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {v} _ {t} ^ {1: N}, t, c\right) - \boldsymbol {\epsilon} _ {t} ^ {1: N} \right\| \right], \tag {8} +$$ + +where $\epsilon_{\theta}(\pmb{v}_{t}^{1:N}, t, c), \epsilon_{t}^{1:N} \in \mathbb{R}^{N \times d}$ . In this work, we denote the predicted noise of $n$ -th frame as $\epsilon_{\theta}^{n}(\pmb{v}_{t}^{1:N}, t, c) \in \mathbb{R}^{d}$ . + +In practice, contemporary video diffusion models often employ cascaded inference pipelines for high-resolution outputs. For instance, [34] initially generates a low-resolution video with strong text-video correlation, further enhancing its resolution via temporal interpolation and spatial super-resolution modules. + +In exploring video generative tasks through diffusion models, two primary approaches have emerged: leveraging foundational Video Diffusion Models (VDMs) [7, 16, 30, 35] or pre-trained Text-to-Image (T2I) models [4, 10, + +15, 32]. To extend image diffusion models to videos, several architectural modifications are made. Typically, U-Net generative modules integrate temporal attention blocks after spatial attentions [11]. Moreover, 2D convolution layers are inflated to 3D convolution layers by altering kernels [11]. + +# 3. Video Motion Customization + +Given an input video, our main goal is to (a) distill the motion patterns $M_*$ of target subjects, and (b) customize the input video in different contexts while maintaining the same motion patterns $M_*$ , e.g. Sharks w/ motion $M_* \rightarrow$ Airplanes w/ motion $M_*$ , with minimal computational costs. + +To this end, we propose a novel video motion customization framework, namely VMC, which leverages cascaded video diffusion models with robust temporal priors. One notable aspect of the proposed framework is that we perform fine-tuning only on the key-frame generation module, also referred to as the T2V base model, within the cascaded VDMs, which guarantees computational and memory efficiency. Specifically, within the key-frame generation model, our fine-tuning process only targets the temporal attention layers. This facilitates adaptation while preserving the model's inherent capacity for generic synthesis. Notably, we freeze the subsequent frame interpolation and spatial super-resolution modules as-is (Fig. 2). + +# 3.1. Temporal Attention Adaptation + +In order to distill the motion $M_*$ , we first propose a new objective function for temporal attention adaptation using residual cosine similarity. Our intuition is that residual vectors between consecutive frames may include information about the motion trajectories. + +Let $(\pmb{v}^n)_{n\in \{1,\dots ,N\}}$ represents the $N$ -frame input video + +![](images/6f7da401475cafc1fcb67d25265a6ef82a6e92f4c9c1e7a3607bc2bac368ff42.jpg) +Figure 3. Training. The proposed framework aims to learn motion by $\delta \epsilon_t^n$ -alignment using (16) or (17). Note that we only fine-tune the temporal attention layers in the key-frame generation U-Net. The blue circle represents the diffusion forward process. + +sequence. As defined in Section 2, for a given noisy video latent vector $\pmb{v}_t^{1:N}$ with $\epsilon_t^{1:N}$ , let $\pmb{v}_t^n$ represents the $n$ -th noisy frame latent sampled from $p_t(\pmb{v}_t^n|\pmb{v}^n)$ with $\epsilon_t^n$ . We will interchangeably use $\pmb{v}^n$ and $\pmb{v}_0^n$ for notational simplicity. Likewise, $\pmb{v}_t^{n + c}$ is defined as $\pmb{v}_t^n$ , with $c > 0$ representing the fixed frame stride. Then, we define the frame residual vector at time $t \geqslant 0$ as + +$$ +\delta \boldsymbol {v} _ {t} ^ {n} := \boldsymbol {v} _ {t} ^ {n + c} - \boldsymbol {v} _ {t} ^ {n}, \tag {9} +$$ + +where we similarly define the epsilon residual vector $\delta \pmb{\epsilon}_t^n$ . In the rest of the paper, we interchangeably use frame residual vector and motion vector. + +We expect that these motion vectors may encode information about motion patterns, where such information may vary depending on the time $t$ and its corresponding noise level. The difference vector $\delta \pmb{v}_t^n$ can be delineated as: + +$$ +\begin{array}{l} \delta \boldsymbol {v} _ {t} ^ {n} = \sqrt {\bar {\alpha} _ {t}} \left(\boldsymbol {v} _ {0} ^ {n + c} - \boldsymbol {v} _ {0} ^ {n}\right) + \sqrt {1 - \bar {\alpha} _ {t}} \left(\boldsymbol {\epsilon} _ {t} ^ {n + c} - \boldsymbol {\epsilon} _ {t} ^ {n}\right) \tag {10} \\ = \sqrt {\bar {\alpha} _ {t}} \delta \boldsymbol {v} _ {0} ^ {n} + \sqrt {1 - \bar {\alpha} _ {t}} \delta \boldsymbol {\epsilon} _ {t} ^ {n}, \\ \end{array} +$$ + +where $\delta \pmb{\epsilon}_t^n$ is normally distributed with zero mean and $2I$ variance. In essence, $\delta \pmb{v}_t^n$ can be acquired through the following diffusion kernel: + +$$ +p \left(\delta \boldsymbol {v} _ {t} ^ {n} \mid \delta \boldsymbol {v} _ {0} ^ {n}\right) = \mathcal {N} \left(\delta \boldsymbol {v} _ {t} ^ {n} \mid \sqrt {\bar {\alpha} _ {t}} \delta \boldsymbol {v} _ {0} ^ {n}, 2 \left(1 - \bar {\alpha} _ {t}\right) I\right). \tag {11} +$$ + +In light of this, our goal is to transfer motion information to the temporal attention layers by leveraging the motion vectors. For this, we first simulate the motion vectors using video diffusion models. Specifically, as similarly done in (6), the denoised video vector estimates $\hat{v_0}^{1:N}(t)$ can be derived by applying Tweedie's formula: + +$$ +\hat {\boldsymbol {v}} _ {0} ^ {1: N} (t) := \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\boldsymbol {v} _ {t} ^ {1: N} - \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {v} _ {t} ^ {1: N}, t\right)\right), \tag {12} +$$ + +where $\hat{\pmb{v}}_0^{1:N}(t)$ is an empirical Bayes optimal posterior expectation $\mathbb{E}[\pmb{v}_0^{1:N}|\pmb{v}_t^{1:N}]$ . Then, the denoised motion vector estimate $\delta \hat{\pmb{v}}_0^n$ can be defined in terms of $\delta \pmb{v}_t^n$ and $\delta \pmb{\epsilon}_{\theta}^{n}(\pmb{v}_{t}^{1:N},t)$ by using (12): + +$$ +\delta \boldsymbol {v} _ {0} ^ {n} (t) := \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\delta \boldsymbol {v} _ {t} ^ {n} - \sqrt {1 - \bar {\alpha} _ {t}} \delta \boldsymbol {\epsilon} _ {\theta , t} ^ {n}\right), \tag {13} +$$ + +where $\delta \pmb{\epsilon}_{\theta}^{n}(\pmb{v}_{t}^{1:N}, t) \coloneqq \pmb{\epsilon}_{\theta}^{n + c}(\pmb{v}_{t}^{1:N}, t) - \pmb{\epsilon}_{\theta}^{n}(\pmb{v}_{t}^{1:N}, t)$ is abbreviated as $\delta \pmb{\epsilon}_{\theta,t}^{n}$ for notational simplicity. Similarly, one can obtain ground-truth motion vector $\delta \pmb{v}_{0}^{n}$ by using (10): + +$$ +\delta \boldsymbol {v} _ {0} ^ {n} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\delta \boldsymbol {v} _ {t} ^ {n} - \sqrt {1 - \bar {\alpha} _ {t}} \delta \boldsymbol {\epsilon} _ {t} ^ {n}\right). \tag {14} +$$ + +Then, our objective is to finetune $\theta$ by aligning the motion vector $\delta \pmb{v}_0^n$ and its denoised estimate $\delta \hat{\pmb{v}}_0^n (t)$ : + +$$ +\min _ {\theta} \mathbb {E} _ {t, n, \boldsymbol {\epsilon} ^ {t, n}, \boldsymbol {\epsilon} ^ {t, n + c}} \left[ \ell_ {\text {a l i g n}} \left(\delta \boldsymbol {v} _ {0} ^ {n}, \delta \hat {\boldsymbol {v}} _ {0} ^ {n} (t)\right) \right], \tag {15} +$$ + +with a loss function $\ell_{\mathrm{align}}:\mathbb{R}^d\times \mathbb{R}^d\to \mathbb{R}$ . By using $\ell_{2}$ -distance for $\ell_{\mathrm{align}}$ , this is equivalent to matching $\delta \pmb{\epsilon}_{\theta ,t}^{n}$ and $\delta \pmb{\epsilon}_t^n$ + +$$ +\ell_ {\text {a l i g n}} \left(\delta \boldsymbol {v} _ {0} ^ {n}, \delta \hat {\boldsymbol {v}} _ {0} ^ {n} (t)\right) = \frac {1 - \bar {\alpha} _ {t}}{\bar {\alpha} _ {t}} \left\| \delta \boldsymbol {\epsilon} _ {t} ^ {n} - \delta \boldsymbol {\epsilon} _ {\theta , t} ^ {n} \right\| ^ {2}. \tag {16} +$$ + +Notably, aligning the ground-truth and predicted motion vectors translates into aligning epsilon residuals. + +While this objective demonstrates effective empirical performance, our additional observations indicate that using $\ell_{\cos}(\delta \pmb{\epsilon}_t^n,\delta \pmb{\epsilon}_{\theta ,t}^n)$ may further improve the distillation, where $\ell_{\cos}(\pmb {x},\pmb {y}) = 1 - \frac{\langle\pmb{x},\pmb{y}\rangle}{\|\pmb{x}\|\|\pmb{y}\|}$ for $\pmb {x},\pmb {y}\in \mathbb{R}^{d}$ (more analysis in section 4.3). Accordingly, our optimization framework is finally defined as follows: + +$$ +\min _ {\theta} \mathbb {E} _ {t, n, \boldsymbol {\epsilon} ^ {t, n}, \boldsymbol {\epsilon} ^ {t, n + c}} \left[ \ell_ {\cos} \left(\delta \boldsymbol {\epsilon} _ {t} ^ {n}, \delta \boldsymbol {\epsilon} _ {\theta , t} ^ {n}\right) \right]. \tag {17} +$$ + +In other words, the proposed optimization framework aims to maximize the residual cosine similarity between $\delta \epsilon_{t}^{n}$ and $\delta \epsilon_{\theta ,t}^{n}$ . Hence, this optimization approach can be seamlessly applied to video diffusion models trained using epsilon-matching. Practically, we exclusively fine-tune the temporal attention layers $\theta_{\mathrm{TA}}\subset \theta$ originally designed for dynamic temporal data assimilation [32]. The frame stride remains fixed at $c = 1$ across all experiments. + +# 3.2. Appearance-invariant Prompts + +In motion distillation, it is crucial to filter out disruptive variations that are unrelated to motion. These variations may include changes in appearance and background, distortions, consecutive frame inconsistencies, etc. To achieve this, we further utilize appearance-invariant prompts. Diverging from traditional generative customization frameworks [21, 22, 32, 35] that rely on text prompts that "faithfully" describe the input image or video during model finetuning, our framework purposely employs "unfaithful" text prompts during the training phase. Specifically, our approach involves the removal of background information. For instance, the text prompt 'a cat is roaring on the grass under the tree' is simplified to 'a cat is roaring' as presented in Fig. 4. This reduces background complexity as in Fig. 4a compared to Fig. 4b, facilitating the application of new appearance in motion distillation. + +![](images/a8b241d6b7e01941bd15eaae26211ac51d8fc4be9868648d3827b48987c32a6c.jpg) +Input Video Frame +Figure 4. Appearance-invariant Prompt. Comparison of input reconstruction with and without appearance-invariant prompt: (a) and (b) depict sampled low-resolution (64x40) keyframes. For (a), the training prompt used was "A cat is roaring," while for (b), the training prompt was "A cat is roaring on the grass under the tree." Our appearance-invariant prompt enables the removal of background information that can disturb motion distillation. + +![](images/1f5ce7399e27c6b934c263a6e0894c7c6e57d7648ab058027c7107c844d5fdd7.jpg) +(a) + +![](images/a3999abfa2c6faff59f8015cec203b9c978af6f9c24cfd698d7c786f8e888729.jpg) +(b) + +# 3.3. Inference Pipeline + +Once trained, in the inference phase, our process begins by computing inverted latents from the input video through DDIM inversion. Subsequently, the inverted latents are fed into the temporally fine-tuned keyframe generation model, yielding short and low-resolution keyframes. These keyframes then undergo temporal extension using the unaltered frame interpolation model. Lastly, the interpolated frames are subjected to spatial enlargement through the spatial super-resolution model. Refer to Fig. 2 for overview. + +# 4. Experiments + +# 4.1. Implementation Details + +In our experiments, we choose Show-1 [34] as our VDM backbone and its publicly available pre-trained weights. All experiments were conducted using a single NVIDIA RTX 6000 GPU. VMC with Show-1 demonstrates efficient resource usage, requiring only 15GB of vRAM during mixed-precision training [18], which is completed within 5 minutes. During inference, generating a single video comprising 29 frames at a resolution of $576 \times 320$ consumes 18GB of vRAM and takes approximately 12 minutes. + +# 4.2. Baseline Comparisons + +Dataset Selection. In our experiments, we draw upon a dataset that comprises 24 videos. These videos encompass a broad spectrum of motion types occurring in various contexts, encompassing vehicles, humans, birds, plants, diffusion processes, mammals, sea creatures, and more. This diversity provides a comprehensive range of motion scenarios for our assessment. Out of these 24 videos, 13 are sourced from the DAVIS dataset [19], 10 from the WebVid dataset [1], and 1 video is obtained from LAMP [33]. + +Baselines. Our method is compared against four contemporary baselines that integrate depth map signals into the diffusion denoising process to assimilate motion information. Notably, our approach operates without the necessity of depth maps during both training and inference, in contrast to these baseline methods. + +Specifically, VideoComposer (VC) [30] is an open-source latent-based video diffusion model tailored for compositional video generation tasks. Gen-1 [7] introduces a video diffusion architecture incorporating additional structure and content guidance for video-to-video translation. In contrast to our targeted fine-tuning of temporal attention, Tune-A-Video (TAV) [32] fine-tunes self, cross, and temporal attention layers within a pre-trained, but inflated T2I model on input videos. Control-A-Video (CAV) [5] introduces a controllable T2V diffusion model utilizing control signals and a first-frame conditioning strategy. Notably, while closely aligned with our framework, Motion Director [35] lacks available code at the time of our research. + +Qualitative Results. We offer visual comparisons of our method against four baselines in Fig. 5. The compared baselines face challenges in adapting the motion of the input video to new contexts. They exhibit difficulties in applying the overall motion, neglecting the specific background indicated in the target text (e.g., "underwater" or "on the sand"). Additionally, they face difficulties in deviating from the original shape of the subject in the input video, leading to issues like a shark-shaped airplane, an owl-shaped seagull, or preservation of the shape of the ground where a seagull is taking off. In contrast, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization, e.g. Two sharks are moving. $\rightarrow$ Two airplanes are moving in the sky. + +Quantitative Results. We further quantitatively demonstrate the effectiveness of our method against the baselines through automatic metrics and user study. + +Automatic Metrics. We use CLIP [20] for automatic metrics. For textual alignment, we compute the average cosine similarity between the target prompt and the generated frames. In terms of frame consistency, we obtain CLIP image features within the output video and then calculate the average cosine similarity among all pairs of video frames. For methods that generate temporally interpolated frames, we utilized the keyframe indexes to calculate the metric for a fair evaluation. To illustrate, in the case of VMC, which takes an 8-frame input and produces a 29-frame output, we considered the frames at the following indexes: 1, 5, 9, 13, 17, 21, 25, 29. As shown in Table 1, VMC outperforms baselines in both text alignment and temporal consistency. User Study. We conducted a survey involving a total of 27 participants to assess four key aspects: the preservation of motion between the input video and the generated output video, appearance diversity in the output video compared to the input video, the text alignment with the target prompt, and the overall consistency of the generated frames. The survey utilized a rating scale ranging from 1 to 5. For assessing motion preservation, we employed the question: + +![](images/f15857e49dbd158cdbd77b61822c6b0b8f6a777eadef8c6164321adc53fa0144.jpg) + +![](images/b60dba8fc6b8f85670faeabc53fde866bbf28286f439420a5fffb79674dd238e.jpg) +Figure 5. Qualitative comparison against state-of-the-art baselines. In contrast to other baselines, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization. + +![](images/10144d2a9770d5c3e13a84ce8b1cfd4315954555bb4effc20d947209615c3fe7.jpg) +Figure 6. Comparative analysis of the proposed frameworks with fine-tuning (a) temporal attention and (b) self- and cross-attention layers. + +![](images/0bb1e48579f3623fc6d20ac193ecfbdf9cd855bf479e2c39445f73735fd4b1c5.jpg) +Figure 7. Comparative analysis of the proposed frameworks with (a) $\ell_{cos}$ and (b) $\ell_2$ loss functions. + +"To what extent is the motion of the input video retained in the output video?" To evaluate appearance diversity, participants were asked: "To what extent does the appearance of the output video avoid being restricted on the input video's appearance?" Tab. 1 shows that our method surpasses the baselines in all four aspects. + +
Text AlignmentTemporal ConsistencyMotion PreservationAppearance DiversityText AlignmentTemporal Consistency
VC0.7980.9583.453.432.963.03
Gen-10.7800.9573.463.172.872.73
TAV0.7580.9473.502.882.672.80
CAV0.7640.9522.752.452.072.00
Ours0.8010.9594.424.544.564.57
+ +Table 1. Quantitative evaluation using CLIP (left) and user study (right). VMC significantly outperforms the other baselines. + +# 4.3. Ablation Studies + +Comparisons on attention layers. We conducted a comparative study evaluating the performance of fine-tuning: (a) temporal attention layers and (b) self- and cross-attention layers. Illustrated in Fig. 6, both frameworks exhibit proficient motion learning capabilities. Notably, the utilization of customized temporal attention layers (a) yields smoother frame transitions, indicating the effectiveness of the optimization framework (17) in encouraging motion distillation, with a slight preference observed for customized temporal attention layers. + +This observation stems from the premise that integrating the proposed motion distillation objective (17) may autonomously and accurately embed motion information within temporal attention layers [11, 13]. This suggests a potential application of the motion distillation objective for training large-scale video diffusion models, warranting further exploration in future research endeavors. + +Choice of loss functions. In addition, we conducted a comparative analysis on distinct training loss functions in (17): the $\ell_{2}$ -distance and $\ell_{\mathrm{cos}}$ as delineated in (17). As depicted in Fig. 7, the $\delta \epsilon$ -matching process in (15) and (17) demonstrates compatibility with generic loss functions. While both $\ell_{2}(\delta \epsilon_{t}^{n}, \delta \epsilon_{\theta, t}^{n})$ and $\ell_{\mathrm{cos}}(\delta \epsilon_{t}^{n}, \delta \epsilon_{\theta, t}^{n})$ are promising objectives, the marginal superiority of $\ell_{\mathrm{cos}}(\delta \epsilon_{t}^{n}, \delta \epsilon_{\theta, t}^{n})$ led to its adoption for visualizations in this study. + +Importance of adaptation. To assess the importance of temporal attention adaptation, we conducted a visualization of customized generations without temporal attention adaptation, as detailed in Section 3.1. Specifically, from our original architecture in Fig. 2, we omitted attention adaptation and performed inference by maintaining the U-Net modules in a frozen state. The outcomes depicted in Fig. 9 indicate that while DDIM inversion guides the generations + +![](images/f8faab0ded7e24ddeea311f461b75fe1caccf811075bb8e63d69932a4bcaafe9.jpg) +Figure 8. Left: Style transfer on two videos. Right: Motion customization results on the video of "A seagull is walking backward." + +![](images/a6bc297438c3589159b9cd65aaabcf9bebb62eb90e5ff771ed49148e5505c50a.jpg) +Figure 9. Ablation study on temporal attention adaptation. Without temporal attention adaptation, motion distillation fails. + +to mimic the motion of the input video, it alone does not ensure successful motion distillation. The observed changes in appearance and motion exhibit an entangled relationship. Consequently, this underlines the necessity of an explicit motion distillation objective to achieve consistent motion transfer, independent of any alterations in appearance. + +# 4.4. Additional results + +Video Style Transfer. We illustrate video style transfer applications in Fig. 8-Left. We incorporate style prompts at the end of the text after applying appearance-invariant prompting (see Section 3.2). Target styles are fluidly injected while preserving the distilled motion of an input video. + +Learning Backward Motion. To further verify our video motion customization capabilities, we present a challenging scenario: extracting backward motion from a reversed + +video sequence where frames are arranged in reverse order. This scenario, an exceedingly rare event in real-world videos, is highly improbable within standard training video datasets [1]. Illustrated in Fig. 8, our VMC framework showcases proficiency in learning "a bird walking backward" motion and generating diverse videos with distinct subjects and backgrounds. This capability not only enables leveraging the distilled motion but also offers prospects for further contextual editing. + +# 5. Conclusion + +This paper introduces Video Motion Customization (VMC), addressing challenges in adapting Text-to-Video (T2V) models to generate motion-driven diverse visual customizations. Existing models struggle with accurately replicating motion from a target video and creating varied visual outputs, leading to entanglements of appearance and motion data. To overcome this, our VMC framework presents a novel one-shot tuning approach, focusing on adapting temporal attention layers within video diffusion models. This framework stands out for its efficiency in time and memory, ease of implementation, and minimal hyperparameters. + +Ethics Statement. Our work employs a generative model that necessitates caution due to its potential for misuse in creating deceptive content with negative societal impacts. + +# References + +[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 8 +[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2 +[4] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3 +[5] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 5 +[6] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2 +[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 3, 5 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2 +[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3 +[11] J Ho, T Salimans, A Gritsenko, W Chan, M Norouzi, and DJ Fleet. Video diffusion models. arxiv 2022. arXiv preprint arXiv:2204.03458. 3, 7 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 7 + +[14] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 2 +[15] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3 +[16] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 3 +[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 2 +[18] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 5 +[19] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5 +[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4 +[22] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 4 +[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2 +[25] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International confer + +ence on machine learning, pages 2256-2265. PMLR, 2015. 2 +[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[28] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[30] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3, 5 +[31] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2 +[32] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2, 3, 4, 5 +[33] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 5 +[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3, 5 +[35] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2, 3, 4, 5 +[36] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 \ No newline at end of file diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/images.zip b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8c5a8c5c0a94834d919061f25383f56471f121f2 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6abef4acb929e9fa01a2fab07f34c5a832b0320038c5a8420534b65e3dce95a2 +size 1157097 diff --git a/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/layout.json b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ce6e1f1c7b1298a1c76713966ec1858a5813b0 --- /dev/null +++ b/2024/VMC_ Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models/layout.json @@ -0,0 +1,8249 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 530, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 530, + 138 + ], + "type": "text", + "content": "VMC: Video Motion Customization using Temporal Attention Adaption for Text-to-Video Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "content": "Hyeonho Jeong" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "content": " Geon Yeong Park" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "content": " Jong Chul Ye" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "content": "Kim Jaechul Graduate School of AI, " + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 459, + 233 + ], + "type": "text", + "content": "Bio and Brain Engineering Korea Advanced Institute of Science and Technology (KAIST) * Indicates co-first authors {hyeonho.jeong, pky3436,jong.ye}@kaist.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 266, + 187, + 540 + ], + "blocks": [ + { + "bbox": [ + 59, + 266, + 187, + 540 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 187, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 187, + 540 + ], + "type": "image", + "image_path": "0361a88fc4a77228b5db573a17f446329bd9e6e33d3a7cdaf70c88dcc415c254.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 540, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 540, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 547, + 586 + ], + "type": "text", + "content": "Figure 1. Using only a single video portraying any type of motion, our Video Motion Customization framework allows for generating a wide variety of videos characterized by the same motion but in entirely distinct contexts and better spatial/temporal resolution. 8-frame input videos are translated to 29-frame videos in different contexts while closely following the target motion. The visualized frames for the first video are at indexes 1, 9, and 17. A comprehensive view of these motions in the form of videos can be explored at our project page." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 190, + 267, + 302, + 540 + ], + "blocks": [ + { + "bbox": [ + 190, + 267, + 302, + 540 + ], + "lines": [ + { + "bbox": [ + 190, + 267, + 302, + 540 + ], + "spans": [ + { + "bbox": [ + 190, + 267, + 302, + 540 + ], + "type": "image", + "image_path": "31a55f9578dab9c43f41e6dda0a2f3cbbb42773e6d1f8ee82849c0c543e21c0a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 267, + 417, + 540 + ], + "blocks": [ + { + "bbox": [ + 306, + 267, + 417, + 540 + ], + "lines": [ + { + "bbox": [ + 306, + 267, + 417, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 417, + 540 + ], + "type": "image", + "image_path": "ad270e66ab0145940372d36ee4cc5e4ef1c3557496971ef197312449cd91e26e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 422, + 267, + 534, + 540 + ], + "blocks": [ + { + "bbox": [ + 422, + 267, + 534, + 540 + ], + "lines": [ + { + "bbox": [ + 422, + 267, + 534, + 540 + ], + "spans": [ + { + "bbox": [ + 422, + 267, + 534, + 540 + ], + "type": "image", + "image_path": "85d9be07120ca207ac10214a67a82eb6b5f906f1249ef04e6df3ed43f0608468.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 594, + 192, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 594, + 192, + 608 + ], + "spans": [ + { + "bbox": [ + 143, + 594, + 192, + 608 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "text", + "content": "Text-to-video diffusion models have advanced video generation significantly. However, customizing these models to generate videos with tailored motions presents a substantial challenge. In specific, they encounter hurdles in " + }, + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{a})" + }, + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "text", + "content": " accurately reproducing motion from a target video, and " + }, + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{b})" + }, + { + "bbox": [ + 45, + 629, + 288, + 715 + ], + "type": "text", + "content": " creating diverse visual variations. For example, straightforward extensions of static image customization methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 596, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 705 + ], + "type": "text", + "content": "to video often lead to intricate entanglements of appearance and motion data. To tackle this, here we present the Video Motion Customization (VMC) framework, a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "9212" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 287, + 131 + ], + "type": "text", + "content": "frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our code and data can be found at https://video-motioncustomization.github.io/." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 330 + ], + "type": "text", + "content": "The evolution of diffusion models [12, 26, 29] has significantly advanced Text-to-Image (T2I) generation, notably when paired with extensive text-image datasets [3, 23]. While cascaded diffusion pipelines [2, 9, 13, 25, 31, 34, 36] have extended this success to Text-to-Video (T2V) generation, current models lack the ability to replicate specific motions or generate diverse variations of the same motion with distinct visual attributes and backgrounds. Addressing this, we tackle the challenge of Motion Customization [35]—adapting pre-trained Video Diffusion Models (VDM) to produce motion-specific videos in different contexts, while maintaining the same motion patterns of target subjects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "text", + "content": "Given a few subject images for reference, appearance customization [8, 17, 21, 22, 24, 32] in generative models aims to fine-tune models to generate subject images in diverse contexts. However, these approaches, despite varying optimization objectives, commonly strive for faithful image (frame) reconstruction by minimizing the " + }, + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "text", + "content": "-distance between predicted and ground-truth noise. This may lead to the entangled learning of appearance and motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "text", + "content": "To tackle this, we present VMC, a new framework aimed at adapting pre-trained VDM's temporal attention layers via our proposed Motion Distillation objective. This approach utilizes residual vectors between consecutive (latent) frames to obtain the motion vectors that trace motion trajectories in the target video. Consequently, we fine-tune VDM's temporal attention layers to align the ground-truth image-space residuals with their denoised estimates, which equivalently aligns predicted and ground-truth source noise differences and motion vectors within VDM. This enables lightweight and fast one-shot training. To further facilitate the appearance-invariant motion distillation, we transform faithful text prompts into appearance-invariant prompts, e.g. \"A bird is flying above a lake in the forest\" " + }, + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "text", + "content": " \"A bird is flying\" in Fig. 1. This encourages the modules to focus on the motion information and ignore others, such as appearance, structure, background, etc. During inference, our procedure initiates by sampling key-frames using the adapted key-frame generation U-Net, followed by temporal interpolation and spatial super-resolution. To summarize, VMC makes the following key contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "- We introduce a novel fine-tuning strategy which focuses solely on temporal attention layers in the key-frame gen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "text", + "content": "eration module. This enables lightweight training (15GB vRAM) and fast training (" + }, + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "< 5" + }, + { + "bbox": [ + 312, + 72, + 544, + 95 + ], + "type": "text", + "content": " minutes)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 96, + 545, + 227 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 545, + 155 + ], + "type": "text", + "content": "- To our knowledge, we mark a pioneering case of finetuning only the temporal attention layers in video diffusion models, without optimizing spatial self or crossattention layers, while achieving successful motion customization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "type": "text", + "content": "- We introduce a novel motion distillation objective that leverages the residual vectors between consecutive (latent) frames as motion vectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "type": "text", + "content": "- We present the concept of appearance-invariant prompts, which further facilitates the process of motion learning when combined with our motion distillation loss." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 236, + 390, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 390, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 390, + 248 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 316 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models aim to generate samples from the Gaussian noise through iterative denoising processes. Given a clean sample " + }, + { + "bbox": [ + 305, + 256, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\sim p_{\\mathrm{data}}(\\pmb{x})" + }, + { + "bbox": [ + 305, + 256, + 545, + 316 + ], + "type": "text", + "content": ", the forward process is defined as a Markov chain with forward conditional densities" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 319, + 545, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 319, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 337, + 319, + 545, + 348 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\beta_ {t} \\boldsymbol {x} _ {t - 1}, (1 - \\beta_ {t}) I\\right) \\tag {1} \\\\ p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha}) I\\right), \\\\ \\end{array}", + "image_path": "023f626a697460d6303dd836198757402b0cac265c1d623a047ab083ae34b329.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": " is a noisy latent variable at a timestep " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": " that has the same dimension as " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": " denotes an increasing sequence of noise schedule where " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\alpha_{t}\\coloneqq 1 - \\beta_{t}" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}\\coloneqq \\Pi_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": ". Then, the goal of diffusion model training is to obtain a residual denoiser " + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 354, + 545, + 413 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 417, + 548, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 548, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 548, + 446 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right), \\boldsymbol {x} _ {0} \\sim p _ {\\text {d a t a}} \\left(\\boldsymbol {x} _ {0}\\right), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {\\epsilon} \\| \\right]. \\tag {2}", + "image_path": "02b823df4d3dd732cdf0edcd8d9b971d6e2842d181c8eb25f40e72fc11ebfd55.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 446, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 481 + ], + "type": "text", + "content": "It can be shown that this epsilon matching in (2) is equivalent to the Denoising Score Matching (DSM [14, 28]) with different parameterization:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 485, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 485, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 320, + 485, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon} \\left[ \\left\\| \\boldsymbol {s} _ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} (\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}) \\right\\| \\right], \\tag {3}", + "image_path": "c1e27989b6687768c5cf09ed8d53c800dc222412632cabc2645738d857345b5c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "inline_equation", + "content": "s_{\\theta *}(\\pmb{x}_t,t)\\simeq -\\frac{\\pmb{x}_t - \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0}{1 - \\bar{\\alpha}} = -\\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "text", + "content": ". The reverse sampling from " + }, + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{t - 1}|\\pmb{x}_t,\\pmb{\\epsilon}_{\\theta *}(\\pmb{x}_t,t))" + }, + { + "bbox": [ + 305, + 509, + 544, + 548 + ], + "type": "text", + "content": " is then achieved by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 552, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 552, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 552, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} * \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {4}", + "image_path": "dd6fe9441a4dbf6b9dee5d155fdc5b1bd14510d8e14c65b15e84a05ebaaaec64.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_t\\coloneqq \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t" + }, + { + "bbox": [ + 305, + 582, + 544, + 617 + ], + "type": "text", + "content": " .To accelerate sampling, DDIM [27] further proposes another sampling method as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 621, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\hat {\\boldsymbol {x}} _ {0} (t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\eta^ {2} \\tilde {\\beta} _ {t} ^ {2}} \\boldsymbol {\\epsilon} _ {\\theta *} (\\boldsymbol {x} _ {t}, t) + \\eta \\tilde {\\beta} _ {t} \\boldsymbol {\\epsilon}, \\tag {5}", + "image_path": "b77b7f05335a9b994d289672d23f74d604a94cd50f02cf5e40e54b9ce75fe521.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "text", + "content": " is a stochasticity parameter, and " + }, + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0(t)" + }, + { + "bbox": [ + 305, + 651, + 544, + 687 + ], + "type": "text", + "content": " is the denoised estimate which can be equivalently derived using Tweedie's formula [6]:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 340, + 691, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 691, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 340, + 691, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {0} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta *} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right). \\tag {6}", + "image_path": "1a5235dd9f54c07ec28c85fba80355bef42315ab815ae84096bb7d2afefe32e0.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9213" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 87, + 265, + 223 + ], + "blocks": [ + { + "bbox": [ + 52, + 74, + 99, + 85 + ], + "lines": [ + { + "bbox": [ + 52, + 74, + 99, + 85 + ], + "spans": [ + { + "bbox": [ + 52, + 74, + 99, + 85 + ], + "type": "text", + "content": "(a) Training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 87, + 265, + 223 + ], + "lines": [ + { + "bbox": [ + 52, + 87, + 265, + 223 + ], + "spans": [ + { + "bbox": [ + 52, + 87, + 265, + 223 + ], + "type": "image", + "image_path": "31f12a97d47b50cea52bb90ce97db8d052d51b23c0a2b518d7d5316b74b8e19f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 287, + 87, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 290, + 74, + 341, + 84 + ], + "lines": [ + { + "bbox": [ + 290, + 74, + 341, + 84 + ], + "spans": [ + { + "bbox": [ + 290, + 74, + 341, + 84 + ], + "type": "text", + "content": "(b) Inference" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 87, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 287, + 87, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 287, + 87, + 547, + 215 + ], + "type": "image", + "image_path": "044dd9c26bb049203cec23828269b96567d442914c2b3015211ef46ce89762aa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 275 + ], + "type": "text", + "content": "Figure 2. Overview. The proposed Video Motion Customization (VMC) framework distills the motion trajectories from the residual between noisy latent frames, namely motion vector " + }, + { + "bbox": [ + 46, + 228, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_t^n" + }, + { + "bbox": [ + 46, + 228, + 547, + 275 + ], + "type": "text", + "content": ". Specifically, we fine-tune only the temporal attention layers of the key-frame generation model by aligning the ground-truth and predicted motion vectors. After training, the customized key-frame generator is leveraged for target motion-driven video generation with new appearances context, e.g. \"A chicken is walking in a city\"." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 319 + ], + "type": "text", + "content": "For a text-guided Diffusion Model, the training objective is often given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 326, + 287, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 326, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 93, + 326, + 287, + 344 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\epsilon , c} \\left[ \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) - \\boldsymbol {\\epsilon} \\| \\right], \\tag {7}", + "image_path": "7123f240088d7420f75b6790ed58e19931cd9a4df7a7c0c135826ec1f9cd2299.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": " represents the textual embedding. Throughout this paper, we will often omit " + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t,t,c)" + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": " if it does not lead to notational ambiguity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": "Video Diffusion Models. Video diffusion models [11, 13, 34] further attempt to model the video data distribution. Specifically, Let " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "(\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": "-frame input video sequence. Then, for a given " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{v}^n\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{v}^{1:N}\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": " represents a whole video vector. Let " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^n = \\sqrt{\\bar{\\alpha}_t}\\pmb{v}^n +\\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}_t^n" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": "-th noisy frame latent sampled from " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "p_t(\\pmb{v}_t^n |\\pmb{v}^n)" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_t^n\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ". We similarly define " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "(\\pmb{v}_t^n)_{n\\in \\{1,\\dots ,N}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^{1:N}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_t^{1:N}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": ". The goal of video diffusion model training is then to obtain a residual denoiser " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": " with textual condition " + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 402, + 289, + 535 + ], + "type": "text", + "content": " and video input that satisfies:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 542, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 542, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 62, + 542, + 287, + 561 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {v} _ {t} ^ {1: N}, \\boldsymbol {v} ^ {1: N}, \\boldsymbol {\\epsilon} _ {t} ^ {1: N}, c} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {\\epsilon} _ {t} ^ {1: N} \\right\\| \\right], \\tag {8}", + "image_path": "371ccc3ee126ecd5f73d84389a931984c07213864773e36bb084df5443e1e5f9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{v}_{t}^{1:N}, t, c), \\epsilon_{t}^{1:N} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "text", + "content": ". In this work, we denote the predicted noise of " + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "text", + "content": "-th frame as " + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t, c) \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 568, + 287, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": "In practice, contemporary video diffusion models often employ cascaded inference pipelines for high-resolution outputs. For instance, [34] initially generates a low-resolution video with strong text-video correlation, further enhancing its resolution via temporal interpolation and spatial super-resolution modules." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In exploring video generative tasks through diffusion models, two primary approaches have emerged: leveraging foundational Video Diffusion Models (VDMs) [7, 16, 30, 35] or pre-trained Text-to-Image (T2I) models [4, 10," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 294, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 547, + 355 + ], + "type": "text", + "content": "15, 32]. To extend image diffusion models to videos, several architectural modifications are made. Typically, U-Net generative modules integrate temporal attention blocks after spatial attentions [11]. Moreover, 2D convolution layers are inflated to 3D convolution layers by altering kernels [11]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 366, + 469, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 469, + 379 + ], + "type": "text", + "content": "3. Video Motion Customization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": "Given an input video, our main goal is to (a) distill the motion patterns " + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "inline_equation", + "content": "M_*" + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": " of target subjects, and (b) customize the input video in different contexts while maintaining the same motion patterns " + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "inline_equation", + "content": "M_*" + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": ", e.g. Sharks w/ motion " + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "inline_equation", + "content": "M_* \\rightarrow" + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": " Airplanes w/ motion " + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "inline_equation", + "content": "M_*" + }, + { + "bbox": [ + 304, + 386, + 545, + 457 + ], + "type": "text", + "content": ", with minimal computational costs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 459, + 546, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 615 + ], + "type": "text", + "content": "To this end, we propose a novel video motion customization framework, namely VMC, which leverages cascaded video diffusion models with robust temporal priors. One notable aspect of the proposed framework is that we perform fine-tuning only on the key-frame generation module, also referred to as the T2V base model, within the cascaded VDMs, which guarantees computational and memory efficiency. Specifically, within the key-frame generation model, our fine-tuning process only targets the temporal attention layers. This facilitates adaptation while preserving the model's inherent capacity for generic synthesis. Notably, we freeze the subsequent frame interpolation and spatial super-resolution modules as-is (Fig. 2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 622, + 477, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 477, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 477, + 635 + ], + "type": "text", + "content": "3.1. Temporal Attention Adaptation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "content": "In order to distill the motion " + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "inline_equation", + "content": "M_*" + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "content": ", we first propose a new objective function for temporal attention adaptation using residual cosine similarity. Our intuition is that residual vectors between consecutive frames may include information about the motion trajectories." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{v}^n)_{n\\in \\{1,\\dots ,N\\}}" + }, + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 317, + 701, + 545, + 715 + ], + "type": "text", + "content": " -frame input video" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9214" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 64, + 289, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 64, + 289, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 64, + 289, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 64, + 289, + 149 + ], + "type": "image", + "image_path": "6f7da401475cafc1fcb67d25265a6ef82a6e92f4c9c1e7a3607bc2bac368ff42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 152, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 196 + ], + "type": "text", + "content": "Figure 3. Training. The proposed framework aims to learn motion by " + }, + { + "bbox": [ + 46, + 152, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\delta \\epsilon_t^n" + }, + { + "bbox": [ + 46, + 152, + 287, + 196 + ], + "type": "text", + "content": "-alignment using (16) or (17). Note that we only fine-tune the temporal attention layers in the key-frame generation U-Net. The blue circle represents the diffusion forward process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": "sequence. As defined in Section 2, for a given noisy video latent vector " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^{1:N}" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{1:N}" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": "-th noisy frame latent sampled from " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "p_t(\\pmb{v}_t^n|\\pmb{v}^n)" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\epsilon_t^n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": ". We will interchangeably use " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}^n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}_0^n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " for notational simplicity. Likewise, " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^{n + c}" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^n" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "c > 0" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " representing the fixed frame stride. Then, we define the frame residual vector at time " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "t \\geqslant 0" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 307, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 307, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 125, + 307, + 287, + 321 + ], + "type": "interline_equation", + "content": "\\delta \\boldsymbol {v} _ {t} ^ {n} := \\boldsymbol {v} _ {t} ^ {n + c} - \\boldsymbol {v} _ {t} ^ {n}, \\tag {9}", + "image_path": "9775d08d73e57944e0c5276eed34f2bb613c1bb1492beb85a748b1cf406846b0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 328, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 363 + ], + "type": "text", + "content": "where we similarly define the epsilon residual vector " + }, + { + "bbox": [ + 46, + 328, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_t^n" + }, + { + "bbox": [ + 46, + 328, + 287, + 363 + ], + "type": "text", + "content": ". In the rest of the paper, we interchangeably use frame residual vector and motion vector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "text", + "content": "We expect that these motion vectors may encode information about motion patterns, where such information may vary depending on the time " + }, + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "text", + "content": " and its corresponding noise level. The difference vector " + }, + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_t^n" + }, + { + "bbox": [ + 46, + 365, + 287, + 412 + ], + "type": "text", + "content": " can be delineated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 419, + 287, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 419, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 419, + 287, + 450 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\delta \\boldsymbol {v} _ {t} ^ {n} = \\sqrt {\\bar {\\alpha} _ {t}} \\left(\\boldsymbol {v} _ {0} ^ {n + c} - \\boldsymbol {v} _ {0} ^ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\left(\\boldsymbol {\\epsilon} _ {t} ^ {n + c} - \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right) \\tag {10} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\\\ \\end{array}", + "image_path": "28c6ee320a709c4d1b646887526f4f1265f130562307be89cfc9396798fa5155.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_t^n" + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "text", + "content": " is normally distributed with zero mean and " + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "inline_equation", + "content": "2I" + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "text", + "content": " variance. In essence, " + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_t^n" + }, + { + "bbox": [ + 46, + 457, + 287, + 493 + ], + "type": "text", + "content": " can be acquired through the following diffusion kernel:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 501, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 501, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 59, + 501, + 287, + 515 + ], + "type": "interline_equation", + "content": "p \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\delta \\boldsymbol {v} _ {0} ^ {n}\\right) = \\mathcal {N} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} \\mid \\sqrt {\\bar {\\alpha} _ {t}} \\delta \\boldsymbol {v} _ {0} ^ {n}, 2 \\left(1 - \\bar {\\alpha} _ {t}\\right) I\\right). \\tag {11}", + "image_path": "511309c37c77163d9478637056c58276cf96dfa7fc477fdaed861f7919fe400d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": "In light of this, our goal is to transfer motion information to the temporal attention layers by leveraging the motion vectors. For this, we first simulate the motion vectors using video diffusion models. Specifically, as similarly done in (6), the denoised video vector estimates " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\hat{v_0}^{1:N}(t)" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " can be derived by applying Tweedie's formula:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 601, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 601, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 61, + 601, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {v}} _ {0} ^ {1: N} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {v} _ {t} ^ {1: N} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {v} _ {t} ^ {1: N}, t\\right)\\right), \\tag {12}", + "image_path": "2bfd7433524d8f2aff83f813c3312597b46518da4092bc8c0d6c9bcae2484147.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_0^{1:N}(t)" + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": " is an empirical Bayes optimal posterior expectation " + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\pmb{v}_0^{1:N}|\\pmb{v}_t^{1:N}]" + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": ". Then, the denoised motion vector estimate " + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\delta \\hat{\\pmb{v}}_0^n" + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": " can be defined in terms of " + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_t^n" + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N},t)" + }, + { + "bbox": [ + 47, + 634, + 287, + 684 + ], + "type": "text", + "content": " by using (12):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 691, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 691, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 84, + 691, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\delta \\boldsymbol {v} _ {0} ^ {n} (t) := \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right), \\tag {13}", + "image_path": "926ee4cada511d7824663a5264053942b840b6e986c3321b278039e9707c47c7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t) \\coloneqq \\pmb{\\epsilon}_{\\theta}^{n + c}(\\pmb{v}_{t}^{1:N}, t) - \\pmb{\\epsilon}_{\\theta}^{n}(\\pmb{v}_{t}^{1:N}, t)" + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "text", + "content": " is abbreviated as " + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_{\\theta,t}^{n}" + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "text", + "content": " for notational simplicity. Similarly, one can obtain ground-truth motion vector " + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_{0}^{n}" + }, + { + "bbox": [ + 305, + 71, + 545, + 109 + ], + "type": "text", + "content": " by using (10):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 114, + 545, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 114, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 545, + 139 + ], + "type": "interline_equation", + "content": "\\delta \\boldsymbol {v} _ {0} ^ {n} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\delta \\boldsymbol {v} _ {t} ^ {n} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}\\right). \\tag {14}", + "image_path": "4c2587902b87ad2333678e851ff5209a154f1e6575c3ea7d1ab4b6458248c117.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "content": "Then, our objective is to finetune " + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "content": " by aligning the motion vector " + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{v}_0^n" + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "content": " and its denoised estimate " + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\delta \\hat{\\pmb{v}}_0^n (t)" + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 174, + 545, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 174, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 341, + 174, + 545, + 194 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) \\right], \\tag {15}", + "image_path": "bf1f09842edc5072f7ac457b06a209dcac0f78a72f0e1fafba53c41c0bef7c0a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "content": "with a loss function " + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{align}}:\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}" + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "content": ". By using " + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "content": "-distance for " + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{align}}" + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "content": ", this is equivalent to matching " + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_{\\theta ,t}^{n}" + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 201, + 545, + 238 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{\\epsilon}_t^n" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 322, + 243, + 545, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 243, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 322, + 243, + 545, + 268 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {a l i g n}} \\left(\\delta \\boldsymbol {v} _ {0} ^ {n}, \\delta \\hat {\\boldsymbol {v}} _ {0} ^ {n} (t)\\right) = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\left\\| \\delta \\boldsymbol {\\epsilon} _ {t} ^ {n} - \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n} \\right\\| ^ {2}. \\tag {16}", + "image_path": "d1996f1885e0a6be6a725b2d4b00fff6deb94834e41c825091ed208392b3af7f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 273, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 273, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 545, + 296 + ], + "type": "text", + "content": "Notably, aligning the ground-truth and predicted motion vectors translates into aligning epsilon residuals." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "text", + "content": "While this objective demonstrates effective empirical performance, our additional observations indicate that using " + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\ell_{\\cos}(\\delta \\pmb{\\epsilon}_t^n,\\delta \\pmb{\\epsilon}_{\\theta ,t}^n)" + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "text", + "content": " may further improve the distillation, where " + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\ell_{\\cos}(\\pmb {x},\\pmb {y}) = 1 - \\frac{\\langle\\pmb{x},\\pmb{y}\\rangle}{\\|\\pmb{x}\\|\\|\\pmb{y}\\|}" + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 305, + 297, + 545, + 372 + ], + "type": "text", + "content": " (more analysis in section 4.3). Accordingly, our optimization framework is finally defined as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 379, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 379, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 350, + 379, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {t, n, \\boldsymbol {\\epsilon} ^ {t, n}, \\boldsymbol {\\epsilon} ^ {t, n + c}} \\left[ \\ell_ {\\cos} \\left(\\delta \\boldsymbol {\\epsilon} _ {t} ^ {n}, \\delta \\boldsymbol {\\epsilon} _ {\\theta , t} ^ {n}\\right) \\right]. \\tag {17}", + "image_path": "7f247a0907b532c8e96646886e0578bc53959231e2d6655152429b605e265ce0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "content": "In other words, the proposed optimization framework aims to maximize the residual cosine similarity between " + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\delta \\epsilon_{t}^{n}" + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\delta \\epsilon_{\\theta ,t}^{n}" + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "content": ". Hence, this optimization approach can be seamlessly applied to video diffusion models trained using epsilon-matching. Practically, we exclusively fine-tune the temporal attention layers " + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{TA}}\\subset \\theta" + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "content": " originally designed for dynamic temporal data assimilation [32]. The frame stride remains fixed at " + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 305, + 402, + 545, + 497 + ], + "type": "text", + "content": " across all experiments." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 503, + 474, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 474, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 474, + 517 + ], + "type": "text", + "content": "3.2. Appearance-invariant Prompts" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "In motion distillation, it is crucial to filter out disruptive variations that are unrelated to motion. These variations may include changes in appearance and background, distortions, consecutive frame inconsistencies, etc. To achieve this, we further utilize appearance-invariant prompts. Diverging from traditional generative customization frameworks [21, 22, 32, 35] that rely on text prompts that \"faithfully\" describe the input image or video during model finetuning, our framework purposely employs \"unfaithful\" text prompts during the training phase. Specifically, our approach involves the removal of background information. For instance, the text prompt 'a cat is roaring on the grass under the tree' is simplified to 'a cat is roaring' as presented in Fig. 4. This reduces background complexity as in Fig. 4a compared to Fig. 4b, facilitating the application of new appearance in motion distillation." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9215" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 134, + 121 + ], + "blocks": [ + { + "bbox": [ + 66, + 65, + 116, + 73 + ], + "lines": [ + { + "bbox": [ + 66, + 65, + 116, + 73 + ], + "spans": [ + { + "bbox": [ + 66, + 65, + 116, + 73 + ], + "type": "text", + "content": "Input Video Frame" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 73, + 134, + 121 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 134, + 121 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 134, + 121 + ], + "type": "image", + "image_path": "a8b241d6b7e01941bd15eaae26211ac51d8fc4be9868648d3827b48987c32a6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 123, + 287, + 199 + ], + "lines": [ + { + "bbox": [ + 47, + 123, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 287, + 199 + ], + "type": "text", + "content": "Figure 4. Appearance-invariant Prompt. Comparison of input reconstruction with and without appearance-invariant prompt: (a) and (b) depict sampled low-resolution (64x40) keyframes. For (a), the training prompt used was \"A cat is roaring,\" while for (b), the training prompt was \"A cat is roaring on the grass under the tree.\" Our appearance-invariant prompt enables the removal of background information that can disturb motion distillation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 135, + 73, + 209, + 121 + ], + "blocks": [ + { + "bbox": [ + 168, + 66, + 176, + 72 + ], + "lines": [ + { + "bbox": [ + 168, + 66, + 176, + 72 + ], + "spans": [ + { + "bbox": [ + 168, + 66, + 176, + 72 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 73, + 209, + 121 + ], + "lines": [ + { + "bbox": [ + 135, + 73, + 209, + 121 + ], + "spans": [ + { + "bbox": [ + 135, + 73, + 209, + 121 + ], + "type": "image", + "image_path": "1f5ce7399e27c6b934c263a6e0894c7c6e57d7648ab058027c7107c844d5fdd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 211, + 73, + 286, + 121 + ], + "blocks": [ + { + "bbox": [ + 244, + 66, + 252, + 72 + ], + "lines": [ + { + "bbox": [ + 244, + 66, + 252, + 72 + ], + "spans": [ + { + "bbox": [ + 244, + 66, + 252, + 72 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 73, + 286, + 121 + ], + "lines": [ + { + "bbox": [ + 211, + 73, + 286, + 121 + ], + "spans": [ + { + "bbox": [ + 211, + 73, + 286, + 121 + ], + "type": "image", + "image_path": "a3999abfa2c6faff59f8015cec203b9c978af6f9c24cfd698d7c786f8e888729.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 218, + 154, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 154, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 154, + 231 + ], + "type": "text", + "content": "3.3. Inference Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 237, + 286, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 286, + 344 + ], + "type": "text", + "content": "Once trained, in the inference phase, our process begins by computing inverted latents from the input video through DDIM inversion. Subsequently, the inverted latents are fed into the temporally fine-tuned keyframe generation model, yielding short and low-resolution keyframes. These keyframes then undergo temporal extension using the unaltered frame interpolation model. Lastly, the interpolated frames are subjected to spatial enlargement through the spatial super-resolution model. Refer to Fig. 2 for overview." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 354, + 127, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 127, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 127, + 367 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 373, + 179, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 179, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 179, + 386 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 392, + 286, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 286, + 499 + ], + "type": "text", + "content": "In our experiments, we choose Show-1 [34] as our VDM backbone and its publicly available pre-trained weights. All experiments were conducted using a single NVIDIA RTX 6000 GPU. VMC with Show-1 demonstrates efficient resource usage, requiring only 15GB of vRAM during mixed-precision training [18], which is completed within 5 minutes. During inference, generating a single video comprising 29 frames at a resolution of " + }, + { + "bbox": [ + 46, + 392, + 286, + 499 + ], + "type": "inline_equation", + "content": "576 \\times 320" + }, + { + "bbox": [ + 46, + 392, + 286, + 499 + ], + "type": "text", + "content": " consumes 18GB of vRAM and takes approximately 12 minutes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 506, + 173, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 173, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 173, + 518 + ], + "type": "text", + "content": "4.2. Baseline Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 525, + 286, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 286, + 632 + ], + "type": "text", + "content": "Dataset Selection. In our experiments, we draw upon a dataset that comprises 24 videos. These videos encompass a broad spectrum of motion types occurring in various contexts, encompassing vehicles, humans, birds, plants, diffusion processes, mammals, sea creatures, and more. This diversity provides a comprehensive range of motion scenarios for our assessment. Out of these 24 videos, 13 are sourced from the DAVIS dataset [19], 10 from the WebVid dataset [1], and 1 video is obtained from LAMP [33]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": "Baselines. Our method is compared against four contemporary baselines that integrate depth map signals into the diffusion denoising process to assimilate motion information. Notably, our approach operates without the necessity of depth maps during both training and inference, in contrast to these baseline methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "Specifically, VideoComposer (VC) [30] is an open-source latent-based video diffusion model tailored for compositional video generation tasks. Gen-1 [7] introduces a video diffusion architecture incorporating additional structure and content guidance for video-to-video translation. In contrast to our targeted fine-tuning of temporal attention, Tune-A-Video (TAV) [32] fine-tunes self, cross, and temporal attention layers within a pre-trained, but inflated T2I model on input videos. Control-A-Video (CAV) [5] introduces a controllable T2V diffusion model utilizing control signals and a first-frame conditioning strategy. Notably, while closely aligned with our framework, Motion Director [35] lacks available code at the time of our research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 243, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 411 + ], + "type": "text", + "content": "Qualitative Results. We offer visual comparisons of our method against four baselines in Fig. 5. The compared baselines face challenges in adapting the motion of the input video to new contexts. They exhibit difficulties in applying the overall motion, neglecting the specific background indicated in the target text (e.g., \"underwater\" or \"on the sand\"). Additionally, they face difficulties in deviating from the original shape of the subject in the input video, leading to issues like a shark-shaped airplane, an owl-shaped seagull, or preservation of the shape of the ground where a seagull is taking off. In contrast, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization, e.g. Two sharks are moving. " + }, + { + "bbox": [ + 304, + 243, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 243, + 545, + 411 + ], + "type": "text", + "content": " Two airplanes are moving in the sky." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 461 + ], + "type": "text", + "content": "Quantitative Results. We further quantitatively demonstrate the effectiveness of our method against the baselines through automatic metrics and user study." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Automatic Metrics. We use CLIP [20] for automatic metrics. For textual alignment, we compute the average cosine similarity between the target prompt and the generated frames. In terms of frame consistency, we obtain CLIP image features within the output video and then calculate the average cosine similarity among all pairs of video frames. For methods that generate temporally interpolated frames, we utilized the keyframe indexes to calculate the metric for a fair evaluation. To illustrate, in the case of VMC, which takes an 8-frame input and produces a 29-frame output, we considered the frames at the following indexes: 1, 5, 9, 13, 17, 21, 25, 29. As shown in Table 1, VMC outperforms baselines in both text alignment and temporal consistency. User Study. We conducted a survey involving a total of 27 participants to assess four key aspects: the preservation of motion between the input video and the generated output video, appearance diversity in the output video compared to the input video, the text alignment with the target prompt, and the overall consistency of the generated frames. The survey utilized a rating scale ranging from 1 to 5. For assessing motion preservation, we employed the question:" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9216" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 358 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 358 + ], + "type": "image", + "image_path": "f15857e49dbd158cdbd77b61822c6b0b8f6a777eadef8c6164321adc53fa0144.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 371, + 545, + 661 + ], + "blocks": [ + { + "bbox": [ + 47, + 371, + 545, + 661 + ], + "lines": [ + { + "bbox": [ + 47, + 371, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 545, + 661 + ], + "type": "image", + "image_path": "b60dba8fc6b8f85670faeabc53fde866bbf28286f439420a5fffb79674dd238e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 547, + 685 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 547, + 685 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison against state-of-the-art baselines. In contrast to other baselines, the proposed framework succeeds in motion-driven customization, even for difficult compositional customization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9217" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 541, + 215 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 541, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 541, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 541, + 215 + ], + "type": "image", + "image_path": "10144d2a9770d5c3e13a84ce8b1cfd4315954555bb4effc20d947209615c3fe7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 545, + 232 + ], + "type": "text", + "content": "Figure 6. Comparative analysis of the proposed frameworks with fine-tuning (a) temporal attention and (b) self- and cross-attention layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 242, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 48, + 242, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 48, + 242, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 545, + 371 + ], + "type": "image", + "image_path": "0bb1e48579f3623fc6d20ac193ecfbdf9cd855bf479e2c39445f73735fd4b1c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "lines": [ + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "type": "text", + "content": "Figure 7. Comparative analysis of the proposed frameworks with (a) " + }, + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "type": "inline_equation", + "content": "\\ell_{cos}" + }, + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "type": "text", + "content": " and (b) " + }, + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 116, + 376, + 476, + 388 + ], + "type": "text", + "content": " loss functions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "type": "text", + "content": "\"To what extent is the motion of the input video retained in the output video?\" To evaluate appearance diversity, participants were asked: \"To what extent does the appearance of the output video avoid being restricted on the input video's appearance?\" Tab. 1 shows that our method surpasses the baselines in all four aspects." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 484, + 285, + 539 + ], + "blocks": [ + { + "bbox": [ + 48, + 484, + 285, + 539 + ], + "lines": [ + { + "bbox": [ + 48, + 484, + 285, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 285, + 539 + ], + "type": "table", + "html": "
Text AlignmentTemporal ConsistencyMotion PreservationAppearance DiversityText AlignmentTemporal Consistency
VC0.7980.9583.453.432.963.03
Gen-10.7800.9573.463.172.872.73
TAV0.7580.9473.502.882.672.80
CAV0.7640.9522.752.452.072.00
Ours0.8010.9594.424.544.564.57
", + "image_path": "e81146b2bab543c598e1b3543a0a724d8b2b637fa09d93a3ac3b18872c24b939.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 287, + 562 + ], + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 562 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation using CLIP (left) and user study (right). VMC significantly outperforms the other baselines." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 575, + 147, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 147, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 147, + 586 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Comparisons on attention layers. We conducted a comparative study evaluating the performance of fine-tuning: (a) temporal attention layers and (b) self- and cross-attention layers. Illustrated in Fig. 6, both frameworks exhibit proficient motion learning capabilities. Notably, the utilization of customized temporal attention layers (a) yields smoother frame transitions, indicating the effectiveness of the optimization framework (17) in encouraging motion distillation, with a slight preference observed for customized temporal attention layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 492 + ], + "type": "text", + "content": "This observation stems from the premise that integrating the proposed motion distillation objective (17) may autonomously and accurately embed motion information within temporal attention layers [11, 13]. This suggests a potential application of the motion distillation objective for training large-scale video diffusion models, warranting further exploration in future research endeavors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": "Choice of loss functions. In addition, we conducted a comparative analysis on distinct training loss functions in (17): the " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": "-distance and " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{cos}}" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": " as delineated in (17). As depicted in Fig. 7, the " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\delta \\epsilon" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": "-matching process in (15) and (17) demonstrates compatibility with generic loss functions. While both " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\ell_{2}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": " are promising objectives, the marginal superiority of " + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{cos}}(\\delta \\epsilon_{t}^{n}, \\delta \\epsilon_{\\theta, t}^{n})" + }, + { + "bbox": [ + 304, + 507, + 546, + 603 + ], + "type": "text", + "content": " led to its adoption for visualizations in this study." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Importance of adaptation. To assess the importance of temporal attention adaptation, we conducted a visualization of customized generations without temporal attention adaptation, as detailed in Section 3.1. Specifically, from our original architecture in Fig. 2, we omitted attention adaptation and performed inference by maintaining the U-Net modules in a frozen state. The outcomes depicted in Fig. 9 indicate that while DDIM inversion guides the generations" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9218" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 65, + 547, + 209 + ], + "blocks": [ + { + "bbox": [ + 49, + 65, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 49, + 65, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 65, + 547, + 209 + ], + "type": "image", + "image_path": "f8faab0ded7e24ddeea311f461b75fe1caccf811075bb8e63d69932a4bcaafe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 211, + 533, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 211, + 533, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 211, + 533, + 222 + ], + "type": "text", + "content": "Figure 8. Left: Style transfer on two videos. Right: Motion customization results on the video of \"A seagull is walking backward.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 284, + 473 + ], + "blocks": [ + { + "bbox": [ + 50, + 239, + 284, + 473 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 284, + 473 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 284, + 473 + ], + "type": "image", + "image_path": "a6bc297438c3589159b9cd65aaabcf9bebb62eb90e5ff771ed49148e5505c50a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "lines": [ + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 287, + 501 + ], + "type": "text", + "content": "Figure 9. Ablation study on temporal attention adaptation. Without temporal attention adaptation, motion distillation fails." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 593 + ], + "type": "text", + "content": "to mimic the motion of the input video, it alone does not ensure successful motion distillation. The observed changes in appearance and motion exhibit an entangled relationship. Consequently, this underlines the necessity of an explicit motion distillation objective to achieve consistent motion transfer, independent of any alterations in appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 153, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 153, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 153, + 611 + ], + "type": "text", + "content": "4.4. Additional results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "Video Style Transfer. We illustrate video style transfer applications in Fig. 8-Left. We incorporate style prompts at the end of the text after applying appearance-invariant prompting (see Section 3.2). Target styles are fluidly injected while preserving the distilled motion of an input video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Learning Backward Motion. To further verify our video motion customization capabilities, we present a challenging scenario: extracting backward motion from a reversed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 243, + 546, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 352 + ], + "type": "text", + "content": "video sequence where frames are arranged in reverse order. This scenario, an exceedingly rare event in real-world videos, is highly improbable within standard training video datasets [1]. Illustrated in Fig. 8, our VMC framework showcases proficiency in learning \"a bird walking backward\" motion and generating diverse videos with distinct subjects and backgrounds. This capability not only enables leveraging the distilled motion but also offers prospects for further contextual editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "type": "text", + "content": "This paper introduces Video Motion Customization (VMC), addressing challenges in adapting Text-to-Video (T2V) models to generate motion-driven diverse visual customizations. Existing models struggle with accurately replicating motion from a target video and creating varied visual outputs, leading to entanglements of appearance and motion data. To overcome this, our VMC framework presents a novel one-shot tuning approach, focusing on adapting temporal attention layers within video diffusion models. This framework stands out for its efficiency in time and memory, ease of implementation, and minimal hyperparameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 513, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 550 + ], + "type": "text", + "content": "Ethics Statement. Our work employs a generative model that necessitates caution due to its potential for misuse in creating deceptive content with negative societal impacts." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 561, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 546, + 713 + ], + "type": "text", + "content": "This research was supported by Field-oriented Technology Development Project for Customs Administration through National Research Foundation of Korea (NRF) funded by the Ministry of Science & ICT and Korea Customs Service (NRF-2021M3I1A1097938), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST))." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9219" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "text", + "content": "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[4] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "text", + "content": "[5] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "text", + "content": "[6] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "text", + "content": "[7] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 555 + ], + "type": "text", + "content": "[9] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[10] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[11] J Ho, T Salimans, A Gritsenko, W Chan, M Norouzi, and DJ Fleet. Video diffusion models. arxiv 2022. arXiv preprint arXiv:2204.03458. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[15] Hyeonho Jeong and Jong Chul Ye. Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[16] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "text", + "content": "[18] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[19] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 333, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 398 + ], + "type": "text", + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 465 + ], + "type": "text", + "content": "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "text", + "content": "[22] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "text", + "content": "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[25] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International confer" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9220" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "ence on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[28] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "text", + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[30] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[31] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 386 + ], + "type": "text", + "content": "[32] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "type": "text", + "content": "[33] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 485 + ], + "type": "text", + "content": "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "text", + "content": "[35] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "text", + "content": "[36] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9221" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_content_list.json b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47f2920878e24a72583bba1d445b84feba952a00 --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources", + "text_level": 1, + "bbox": [ + 236, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fan Fei $^{1,2,4}$", + "bbox": [ + 246, + 199, + 339, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiajun Tang $^{1,2,4}$", + "bbox": [ + 359, + 200, + 480, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ping Tan3,4", + "bbox": [ + 501, + 200, + 591, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boxin Shi $^{1,2\\#}$", + "bbox": [ + 612, + 200, + 718, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{3}$ Hong Kong University of Science and Technology $^{4}$ Light Illusions", + "bbox": [ + 107, + 223, + 861, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 300, + 313, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a versatile multi-view inverse rendering framework with near- and far-field light sources. Tackling the fundamental challenge of inherent ambiguity in inverse rendering, our framework adopts a lightweight yet inclusive lighting model for different near- and far-field lights, thus is able to make use of input images under varied lighting conditions available during capture. It leverages observations under each lighting to disentangle the intrinsic geometry and material from the external lighting, using both neural radiance field rendering and physically-based surface rendering on the 3D implicit fields. After training, the reconstructed scene is extracted to a textured triangle mesh for seamless integration into industrial rendering software for various applications. Quantitatively and qualitatively tested on synthetic and real-world scenes, our method shows superiority to state-of-the-art multi-view inverse rendering methods in both speed and quality.", + "bbox": [ + 75, + 325, + 473, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 597, + 209, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The reconstruction of 3D scenes from multi-view RGB imagery has experienced significant advancements following the development of Neural Radiance Fields (NeRF) [19], and both the speed and quality of the reconstruction have reached an unprecedented level [18, 20]. Despite these improvements, a common limitation is to represent only the radiance field, which is a complex product of the external lighting interacting with the intrinsic geometry and material of the scene [11]. The entangled nature of this representation generally hampers the ability to accurately render the scene under unseen lighting conditions, because the influence of the original lighting is embedded within the newly rendered scene. Recent approaches [24, 41, 46] have incorporated inverse rendering [22] to separate material properties and lighting effects, extending the application of the reconstruction beyond novel view synthesis to novel scenarios involving relighting and material editing [2]. Some", + "bbox": [ + 75, + 622, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/666655ee759469bc512e25e6e361d78e0c88657a3c0f2c78b5e8fb4eba83864d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 300, + 669, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f96c9997a11d572c792b71a5aa2263c68a0ac967236f2c8372f60973b522a1b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 300, + 890, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ed3a4b2f4591e97d11b8176d393cd24d837d0542efced48509631783bdb842e.jpg", + "image_caption": [ + "Figure 1. Given multi-view images under possibly varied lighting, our method leverages all present lighting conditions to reconstruct scene geometry and material disentangled with lighting. The trained fields are extracted to textured meshes for seamless integration into industrial renderers for various applications." + ], + "image_footnote": [], + "bbox": [ + 503, + 404, + 658, + 517 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb050fcc3eab0ef1ec35bac7fe3512835a41341a3386d274f77df58c883014ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 404, + 890, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods [2, 6, 9, 21, 32] have advanced to the extraction of detailed triangle meshes with material UV textures, which can serve as economical and lifelike 3D models for gaming and cinematography industries [9, 21], thereby marking a transformative step in digital asset creation.", + "bbox": [ + 496, + 598, + 892, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering presents several fundamental challenges, one of them being its severe inherent ambiguity [41]. To combat this, multi-view inverse rendering methodologies usually reduce the ambiguity by imposing various constraints on each scene component. These methods can be categorized based on the assumption on the amount and types of lighting conditions present in the input images (Tab. 1). The majority of methods [9, 21, 24, 35, 36, 39, 41, 42] assumes the imagery to be captured under one fixed lighting condition (rows 1, 2). However, in such scenarios, only one case of scene appearances among all possible cases under different lighting is observed, causing severe lighting-material ambiguity, posing considerable difficulties to material estimation. In a bid to alleviate this ambiguity, recent techniques [2, 6, 10, 12, 23, 32] utilize input", + "bbox": [ + 496, + 674, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author. E-mail: shiboxin@pku.edu.cn.", + "bbox": [ + 76, + 886, + 375, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f287bb1c871bd177c92373877b5d111ff1d3c3698d7c1f85887194aa065931a0.jpg", + "table_caption": [ + "Table 1. A summary of multi-view inverse rendering methods based on their assumptions and usages of input lighting conditions." + ], + "table_footnote": [ + "${}^{a}$ These methods also support input images under single far-field lighting. In that case, they belong to single far-field lighting methods (row 1)." + ], + "table_body": "
Methods# of far-field light# of near-field lightMaterial optimizationLighting-material ambiguityCapture workload
(1) [24, 39, 41, 42]SingleNoneOnly from singleStrongLow
(2) NeILF [35]Single 5D incident light field
(3) [2, 10, 12, 32]aMultipleNoneOnly from far-fieldModerateHigh
(4) WildLight [6]SingleSingleOnly from near-fieldModerateLow
(5) OursSingleSingleFrom near- and far-fieldWeakLow
MultipleSingle or multipleAlmost noneHigh
", + "bbox": [ + 83, + 107, + 885, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images under varied lighting to discern materials from lighting (rows 3, 4). This demands more complex lighting models. Existing varied-lighting methods [2, 10, 12, 23, 32] typically model multiple, but only far-field, lighting conditions to maintain manageable unknown lighting parameters. This means throughout the capture, the lighting setups should be largely changed by either moving the objects to reconstruct or adapting far-field light sources, causing increased workload to the data collection process. Alternatively, near-field light sources, such as flashlights, can function as easily controllable light sources for disambiguation. WildLight [6] suggests the use of a camera-collocated flashlight in addition to the ambient lighting for inverse rendering. Nevertheless, it limits itself by using only the isolated appearance under the flashlight for material estimation, not harnessing available observations under the ambient lighting and degrading the quality of its estimated material (row 4).", + "bbox": [ + 75, + 260, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing the challenges and restrictions previously outlined, this paper introduces VMINer, a Versatile Multiview Inverse rendering framework with Near- and far-field light sources (row 5). The distinctive feature of our framework is that it uses a lightweight yet inclusive lighting model for different far- and near-field light sources and efficiently leverages available observations under each lighting to disentangle lighting from other components, making the most of whatever lighting conditions are at hand to boost the practicality and quality of the reconstruction, as shown in Fig. 1. Although our method accepts the simplest setting of inputs under single far-field lighting, a more effective compromise is to employ an additional flashlight to enhance quality without overly complicating the capture process. The ideal scenario for the highest quality reconstruction would involve capturing the scene under varied far-field lighting and using a flashlight.", + "bbox": [ + 75, + 517, + 468, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nonetheless, implementing such a versatile framework is far from straightforward. The method has to render the scene under different lighting conditions, including spatially-varying and changing ones to represent possibly moving near-field lights. We propose to solve the problem by modeling near-field lights as point sources in 3D space with adjustable positions and intensities, which could be anisotropic and not necessarily camera-aligned (e.g.,", + "bbox": [ + 75, + 773, + 470, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stationary desk lamps). This representation effectively approximates a broad range of common near-field lights and is easy for spatially-consistent editing and optimization. We use 3D implicit fields to model the scenes' radiance, shape, and material and minimize the error of the neural or physically-based re-rendered scene appearance under each input lighting. To account for multiple lighting situations and moving near-field lights, we integrate specially tailored multi-layer perceptrons (MLPs) that additionally process light directions and embeddings. Post-training, the fields are converted into a textured triangle mesh, ready for seamless integration into industrial rendering software like Blender [1], facilitating a broad spectrum of applications. Our extensive experiments, covering both synthetic and real-world scenes, demonstrate that VMINer surpasses prior methods quantitatively and qualitatively.", + "bbox": [ + 496, + 260, + 893, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 517, + 640, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Preliminary. The rendering equation [11] models the reflected radiance from any point on a surface as a result of an intricate surface integral. This integral comprises contributions from three fundamental scene components: lighting, material, and geometry. The equation is expressed as:", + "bbox": [ + 496, + 542, + 893, + 619 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) = \\int_ {\\Omega} \\underbrace {L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}})} _ {\\text {l i g h t i n g}} \\underbrace {f (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}})} _ {\\text {m a t e r i a l}} \\underbrace {(\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) ^ {+}} _ {\\text {g e o m e t r y}} \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 625, + 890, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this equation, $L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}})$ represents the radiance reflected in the outgoing direction $\\omega_{\\mathrm{o}}$ from a surface point $\\mathbf{x}$ in 3D space. $\\Omega$ denotes the unit hemisphere encompassing all incident directions $\\omega_{\\mathrm{i}}$ with $\\omega_{\\mathrm{i}} \\cdot \\mathbf{n} > 0$ . $L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})$ is the incident radiance at $\\mathbf{x}$ from direction $\\omega_{\\mathrm{i}}$ , and $f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}})$ is the bidirectional reflectance distribution function (BRDF), denoting the ratio of light reflected along $\\omega_{\\mathrm{o}}$ at $\\mathbf{x}$ from $\\omega_{\\mathrm{i}}$ . The term $(\\omega_{\\mathrm{i}} \\cdot \\mathbf{n})^{+}$ , where $x^{+} \\triangleq \\max(x, 0)$ , signifies the cosine weakening factor. The incident lighting $L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})$ may originate from other points within the scene, necessitating recursive evaluation of this integral.", + "bbox": [ + 496, + 671, + 893, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view Inverse Rendering. Multi-view inverse rendering techniques [2-4, 6, 9, 10, 12, 21, 23, 24, 34-36, 38-43] generally employ implicit or explicit 3D fields to model the scene's geometry and material, while also estimating", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ca25f6f05cfdb6c0ded64883c3a2eba6824f20bd9a754e282e7ed753d092aa1.jpg", + "image_caption": [ + "Figure 2. Bottom-left: VMINer takes as input multi-view RGB images with foreground masks under possibly varied far- and near-field lighting. The total number of lighting conditions and their types should be given. Each image is illuminated by one far-field lighting whose index is \"far index\", and near-field lights whose on/off states are \"near on/off\". Top-left: VMINer models each far-field lighting $(\\mathrm{far}_i)$ and near-field lighting $(\\mathrm{near}_i)$ separately as parametric models, from which incoming radiance is queried at any position $\\mathbf{x}$ and direction $\\omega$ . Bottom-mid: It utilizes 3D implicit representations for scene geometry, material, and radiance. Four MLPs process $\\mathbf{x}$ and view direction $\\mathbf{v}$ encoded by different encoders $\\mathrm{enc}_{\\{\\mathrm{geo},\\mathrm{mat},\\mathrm{dir}\\}}$ with light embeddings $\\mathbf{F}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}$ to get the SDF $s$ , the neural appearance descriptor $\\mathbf{z}$ , the BRDF parameters $\\beta$ , and the radiance $\\mathbf{c}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}$ under each lighting. Bottom-right: It then uses radiance field rendering to re-render the appearance $\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}$ separately under each lighting condition, added up to $\\mathbf{C}_{\\mathrm{rf}}$ according to the per-image lighting condition. Top-right: Physically-based surface rendering uses Monte Carlo integration to evaluate both direct illumination $L_{\\mathrm{d}\\{\\mathrm{near},\\mathrm{far}\\}_i}$ with secondary visibility and indirect illumination $L_{\\mathrm{ind}\\{\\mathrm{near},\\mathrm{far}\\}_i}$ . The rendered appearances are aggregated to $\\mathbf{C}_{\\mathrm{pb}}$ . $\\mathbf{C}_{\\{\\mathrm{rf},\\mathrm{pb}\\}}$ are compared with the observation $\\mathbf{C}_{\\mathrm{gt}}$ to train the scene model. $\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}$ are also used as additional supervision signals for $\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{near},\\mathrm{far}\\}_i}$ ." + ], + "image_footnote": [ + ": input : variable : refer to another entry : untrainable operator : trainable operator / MLP $\\bigotimes \\bigoplus$ element-wise multiplication / add" + ], + "bbox": [ + 80, + 89, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lighting to recreate the scene's appearance. They optimize each scene component from scratch by aligning the rendered with the observed appearance, while avoiding potential local minima, which, although do not faithfully represent the scene, can replicate the observed input. Addressing this inherently ambiguous nature of inverse rendering, these methods typically impose constraints on each scene component. For geometry and material, assumptions of known geometry [35, 36, 46] are made, along with leveraging the recovered shape from radiance fields [6, 10, 41], adopting Lambertian or uniform materials [23, 39], controlling BRDF parameter smoothness [10], or integrating a learned BRDF latent space [41]. Regarding lighting, many approaches presuppose known lighting conditions [24, 40] or fixed lighting shared across input images [35, 38, 41, 42]. More recent techniques [2, 6, 10, 12, 23, 32] have turned to using images with varied lighting conditions to mitigate the lighting-material ambiguity. They typically accept input images that can be grouped into sets, each illuminated by one distinct and fixed far-field lighting. WildLight [6] exceptionally employs a camera-collocated flashlight along with fixed ambient lighting. However, it does not model the ambient lighting and thus its material property estimation relies exclusively on flashlight observations, where the light and view direction always coincide [5], limiting its BRDF estimation ability. Our VMINer, in contrast, accommodates", + "bbox": [ + 75, + 503, + 472, + 896 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "input images under single or multiple far-field lighting conditions with near-field light sources and leverages each of them to enhance the quality of the estimated material.", + "bbox": [ + 496, + 503, + 892, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lighting Models. Spatially-uniform (SU) lighting models rely on the assumption that the scene's lighting originates from a distant source, modeled as spherical Gaussians (SGs) [2, 10], spherical harmonics (SH) [12, 23], or MLPs [32]. In contrast, spatially-varying (SV) lighting models acknowledge the presence of near-field light sources, causing different locations to receive different incident lighting. They include parametric 3D lights [7, 17, 36, 37], outgoing light fields such as volumetric SGs [31] and neural out-of-view lighting volumes [45], and incident light fields like SV environment maps [47], SVSGs [16], and neural incident light fields (NeILF) [35, 38]. Our VMINer integrates distant lights as SGs and near-field lights as point lights with adjustable position and intensity. It is less complex compared to other SV lighting models, facilitating spatially consistent editing and optimization, while being able to approximate common lighting setups in real life.", + "bbox": [ + 496, + 549, + 893, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 816, + 671, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, VMINer reconstructs the scene lighting (Sec. 3.1) and 3D fields (Sec. 3.2) from input images using differentiable rendering (Sec. 3.3). The training scheme and loss functions are described in Sec. 3.4.", + "bbox": [ + 496, + 839, + 892, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Versatile Lighting Model", + "text_level": 1, + "bbox": [ + 76, + 90, + 305, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection models the lighting term in Eq. (1) and corresponds to the top-left part of Fig. 2.", + "bbox": [ + 75, + 113, + 468, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VMINer harnesses the diversity of lighting conditions present in the input RGB images to disentangle lighting effects. The lighting model plays a crucial role here: 1) Since lighting conditions are unknown initially, they must be optimized alongside other scene parameters using physically-based rendering during training. 2) Different lighting setups, each compatible with the model, are necessary during the capture phase to achieve this goal. Given these prerequisites, it is essential to design a lighting model that accommodates common far-field and near-field light sources, while being lightweight enough for efficient and robust optimization and physically-based rendering. With these considerations, we propose a versatile lighting model that is both inclusive and lightweight. This representation models far-field and near-field lighting separately as below.", + "bbox": [ + 75, + 143, + 470, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Far-field Lighting. In our model, each far-field lighting condition $i$ , where $i$ belongs to the set $\\mathbb{N}_{N_{\\mathrm{far}}} \\triangleq \\{1, \\dots, N_{\\mathrm{far}}\\}$ and $N_{\\mathrm{far}}$ denotes the number of far-field lighting, is represented using 128-lobe SGs [28] commonly adopted in existing far-field lighting techniques [33, 39, 42]. For each lobe $j$ of lighting $i$ , there are six parameters: lobe axis $\\pmb{\\xi}_{ij} \\in \\mathbb{S}^2$ , lobe sharpness $\\lambda_{ij} \\in \\mathbb{R}_+$ , and lobe RGB amplitude $\\pmb{\\mu}_{ij} \\in \\mathbb{R}_+^3$ . The incident radiance from far-field lighting $i$ along direction $\\omega$ at any 3D position $\\mathbf{x}$ is calculated using the formula (ignoring visibility for now):", + "bbox": [ + 76, + 373, + 470, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\operatorname {f a r} _ {i}} (\\boldsymbol {\\omega}) = \\sum_ {j = 1} ^ {1 2 8} c _ {i j} \\boldsymbol {\\mu} _ {i j} e ^ {\\lambda_ {i} (\\boldsymbol {\\omega} \\cdot \\boldsymbol {\\xi} _ {i j} - 1)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 529, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{ij} = \\lambda_{ij} / (2\\pi (1 - e^{-2\\lambda_{ij}}))$ acts as the normalization factor related to roughness. Notably, $\\mathbf{x}$ does not appear as an input to $L_{\\mathrm{far}_i}$ , reflecting the nature of distant lighting.", + "bbox": [ + 76, + 578, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Near-field Lighting. Each near-field lighting $i$ , included in the set $\\mathbb{N}_{N_{\\mathrm{near}}}$ with $N_{\\mathrm{near}}$ representing the number of near-field light sources, is modeled as a point light. These point lights can have moving positions $\\mathbf{p}_i \\in \\mathbb{R}^3$ and exhibit anisotropic radiation characterized by $l^{\\mathrm{th}}$ -order ( $l$ can be 0, 1, or 2) SH coefficients $\\mathbf{h}_i \\in \\mathbb{R}^{3 \\times (l + 1)^2}$ . Our method accommodates two types of near-field lighting: camera-collocated lights, positioned at the camera ray origin for each image, and stationary lights, which remain fixed across all images. We observe that in neural radiance field rendering, the radiance under a stationary near-field light, especially when the light is active in all images, can be challenging to distinguish from radiance under far-field lighting. Also, the radiance from these two sources aids in material estimation in a similar way. Therefore, in practical applications, we favor using a moving flashlight as the near-field light to provide unique information for material estimation. The incident radiance from near-field lighting $i$ at a 3D po", + "bbox": [ + 75, + 627, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sition $\\mathbf{x}$ is computed as (also ignoring visibility for now):", + "bbox": [ + 500, + 90, + 890, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n e a r} i} (\\mathbf {x}, \\boldsymbol {\\omega}, \\mathbf {p} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathrm {S H} (\\boldsymbol {\\omega} ; \\mathbf {h} _ {i})}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2} ^ {2}} & \\text {i f} \\boldsymbol {\\omega} = \\frac {\\mathbf {p} _ {i} - \\mathbf {x}}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 111, + 890, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{SH}(\\omega ;\\mathbf{h})$ calculates the SH at direction $\\omega$ with coefficients $\\mathbf{h}$ , and $1 / \\| \\mathbf{p}_i - \\mathbf{x}\\|_2^2$ signifies the inverse-square lighting attenuation for point lights. Here we include $\\mathbf{p}_i$ as an input because it may be set differently across images. It is important to note that with respect to the incident direction $\\omega$ , $L_{\\mathrm{far}}$ is a continuous function, while $L_{\\mathrm{near}}$ is a discrete function, being non-zero only in a single direction. As a result, to render appearances under direct lighting, Monte Carlo integration is essential for far-field lighting, while a simple multiplication suffices for near-field lighting. Further details about this process are provided in Sec. 3.3.", + "bbox": [ + 496, + 162, + 893, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Geometry, Material, and Radiance Fields", + "text_level": 1, + "bbox": [ + 498, + 335, + 852, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection models the material and geometry terms in Eq. (1) and corresponds to the bottom-mid part of Fig. 2.", + "bbox": [ + 496, + 359, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry. Multi-view reconstruction methodologies generally hinge on two geometry representations: volume density [2, 10, 19, 41] and the signed distance function (SDF) [33, 39, 42, 44] fields. We choose the SDF field for our geometry representation due to its clearly defined surface at the zero-level isosurface, which simplifies and enhances the post-training mesh extraction process. Our approach utilizes implicit fields as MLPs over explicit structures like voxel grids for their compactness and flexibility. For each 3D position $\\mathbf{x}$ , a multi-resolution hash grid [20] is first employed for positional encoding, yielding a feature vector $\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x}) \\in \\mathbb{R}^{16}$ . The geometry MLP $\\mathrm{M}_{\\mathrm{geo}}$ then predicts the SDF value at $\\mathbf{x}$ :", + "bbox": [ + 496, + 390, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (\\mathbf {x}), \\mathbf {z} (\\mathbf {x}) = \\mathrm {M} _ {\\mathrm {g e o}} (\\mathbf {x}, \\operatorname {e n c} _ {\\mathrm {g e o}} (\\mathbf {x})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 592, + 890, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s \\in \\mathbb{R}$ represents the signed distance (positive outside the surface, negative inside), and $\\mathbf{z} \\in \\mathbb{R}^{13}$ is a descriptor of local appearance. To maintain differentiability in rendering and to aid in geometry optimization, during training we apply NeuS [29] techniques to transform SDF values along rays into volume densities. The surface normal $\\mathbf{n} \\in \\mathbb{S}^2$ is derived as the gradient of the SDF $s$ with respect to $\\mathbf{x}$ :", + "bbox": [ + 496, + 614, + 893, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\partial s}{\\partial \\mathbf {x}} / \\left\\| \\frac {\\partial s}{\\partial \\mathbf {x}} \\right\\| _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 723, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathrm{M}_{\\mathrm{geo}}$ only approximates a strict SDF field that has $\\left\\| \\frac{\\partial s}{\\partial \\mathbf{x}} \\right\\|_2 = 1$ , so we normalize $\\mathbf{n}$ to ensure it is a unit vector.", + "bbox": [ + 498, + 763, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Material. For accurate scene material recovery using physically-based re-rendering, it is imperative to model the material in such a way that the renderer can query the BRDF at any point on the surface. Considering that material properties are independent of lighting and view directions, we represent the scene's material through an implicit 3D field, much like the SDF field utilized for geometry. To achieve", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this, we employ another multi-resolution hash grid $\\mathrm{enc}_{\\mathrm{mat}}$ for positional encoding. We adopt the simplified GGX BRDF model [27] with a fixed fresnel parameter to help alleviate ambiguity. The material MLP, $\\mathbf{M}_{\\mathrm{mat}}$ , is tasked with predicting the SVBRDF parameters $\\beta (\\mathbf{x})$ :", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\beta} (\\mathbf {x}) = \\operatorname {M} _ {\\text {m a t}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {m a t}} (\\mathbf {x})). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 172, + 468, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Radiance. Although our method does not strictly require a radiance field for rendering scene appearance - given that appearance could be rendered solely using physically-based surface rendering (PBR) - we find that incorporating a radiance field at the start of training significantly eases the optimization of scene geometry and appearance. PBR, while physically accurate, tends to introduce instability and slow down the training process due to its inherent ambiguity and computational intensity. Moreover, the radiance field-rendered results can serve as supplementary supervision, aiding PBR in better separating scene appearance under each lighting condition and thus further diminishing lighting-material ambiguity. Hence, we deploy radiance MLPs to predict view-dependent and lighting-dependent radiance at each position. For each lighting condition $i$ , a light embedding $\\mathbf{F}_{\\{\\mathrm{far,near}\\}_i} \\in \\mathbb{R}^{16}$ is utilized. The view direction $\\mathbf{v} \\in \\mathbb{S}^2$ is processed through directional encoding that projects it onto the coefficients of the $3^{\\mathrm{rd}}$ -order SH basis, yielding $\\mathrm{enc}_{\\mathrm{dir}}(\\mathbf{v}) \\in \\mathbb{R}^{4^2 = 16}$ [20]. The far-field radiance MLP $\\mathrm{M}_{\\mathrm{far}}$ computes the outgoing radiance $\\mathbf{c}_{\\mathrm{far}_i}$ from position $\\mathbf{x}$ along view direction $\\mathbf{v}$ under far-field lighting $i$ :", + "bbox": [ + 75, + 194, + 472, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\text {f a r} _ {i}} (\\mathbf {x}, \\mathbf {v}) = \\mathrm {M} _ {\\text {f a r}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\text {f a r} _ {i}}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 520, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we incorporate the surface normal $\\mathbf{n}$ as an additional MLP input, following Instant-NSR [44] and WildLight [6], as this has shown to aid shape recovery.", + "bbox": [ + 75, + 542, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The near-field radiance MLP differs slightly, as nearfield lighting can move in different images, thus the current light position $\\mathbf{p}_i$ is a necessary input. Additionally, the radiance under near-field point light takes the form of the rendering equation Eq. (1) without the integral, allowing the incoming radiance and cosine term to be explicitly included. We follow ReNe [25] to input the relative direction of the point light $\\omega_{i} = \\frac{\\mathbf{p}_{i} - \\mathbf{x}}{\\|\\mathbf{p}_{i} - \\mathbf{x}\\|_{2}}$ instead of the absolute position $\\mathbf{p}_i$ into the MLP. Therefore, the outgoing neural radiance $\\mathbf{c}_{\\mathrm{near}_i}$ under near-field lighting $i$ is calculated as:", + "bbox": [ + 75, + 587, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {c} _ {\\mathrm {n e a r} _ {i}} (\\mathbf {x}, \\mathbf {v}, \\mathbf {p} _ {i}) = \\mathrm {M} _ {\\mathrm {n e a r}} \\left(\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\mathrm {e n c} _ {\\mathrm {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\mathrm {n e a r} _ {i}}, \\boldsymbol {\\omega} _ {i}\\right) \\\\ \\otimes \\frac {L _ {\\text {n e a r} _ {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {i})}{\\left\\| \\mathbf {p} _ {i} - \\mathbf {x} \\right\\| _ {2} ^ {2}} \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 744, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes element-wise multiplication. On the right side, the second term denotes the incident radiance attenuated by square distance, the third term denotes the cosine weakening factor, and $\\mathrm{M}_{\\mathrm{near}}$ accounts for other factors like reflectance and visibility. Owing to the instability of these terms at the training's outset, we use an annealing strategy,", + "bbox": [ + 75, + 804, + 468, + 893 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as recommended in prior works [44], gradually replacing default values (e.g., 1 for incident radiance) with those optimized during training, to ease through the process.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Differentiable Rendering", + "text_level": 1, + "bbox": [ + 498, + 155, + 728, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To simplify the notation, we omit $\\mathbf{p}_i$ and the lighting index $i$ in the equations within this subsection. Our approach optimizes the lighting alongside the implicit fields primarily by reducing the discrepancy between the observed and differentiably re-rendered appearances. The re-rendering is accomplished using two types of renderer: a neural radiance field renderer (Fig. 2 bottom-right) and a physically-based surface renderer (Fig. 2 top-right; Eq. (1)). We handle each lighting condition separately in rendering, and then aggregate the results based on per-image lighting condition.", + "bbox": [ + 496, + 176, + 890, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume Rendering for SDF Field. Consider a camera ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ ( $t > 0$ ), with $\\mathbf{o} \\in \\mathbb{R}^3$ as its origin and $\\mathbf{d} \\in \\mathbb{S}^2$ as its direction. We incorporate importance sampling using an occupancy grid [14] to get $N \\leq 1024$ points along this ray, denoted as $\\mathbf{r}(t_i)$ , $i \\in \\mathbb{N}_N$ . Utilizing techniques from NeuS [29], we transform the signed distances $s(\\mathbf{r}(t_i))$ into discrete opacity values as follows:", + "bbox": [ + 496, + 335, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\left(\\frac {\\Phi_ {b} (s (\\mathbf {r} (t _ {i}))) - \\Phi_ {b} (s (\\mathbf {r} (t _ {i + 1})))}{\\Phi_ {b} (s (\\mathbf {r} (t _ {i})))}\\right) ^ {+}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 446, + 890, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_b(s) = 1 / (1 + e^{-bs})$ is the cumulative opacity distribution function. The parameter $b$ is trainable and tends to increase during training, focusing opacity more narrowly around the surface where $s = 0$ . The accumulated transmittance is then calculated as $T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})$ .", + "bbox": [ + 496, + 488, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural Radiance Field Rendering. The color $\\mathbf{C}_{\\mathrm{rf}}$ of a camera ray under either far- or near-field lighting conditions, as rendered through the neural radiance field, is determined by accumulating the radiance along the ray's path:", + "bbox": [ + 496, + 571, + 890, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r , n e a r} \\}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {\\{\\text {f a r , n e a r} \\}} (\\mathbf {r} (t _ {i}), - \\mathbf {d}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 637, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formula sums the radiance from each sampled point to yield the color of the ray under the specified lighting.", + "bbox": [ + 496, + 681, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Physically-based Rendering. Our method employs differentiable surface rendering, modeling the observed radiance as reflected from a single surface point. We compute the depth $t$ of this surface point similarly to Eq. (10): $t = \\sum_{i=1}^{N} T_i \\alpha_i t_i$ . Accumulation along the ray also applies to the BRDF parameters $\\beta$ and the surface normal $n$ . For each input lighting, we render the appearance that models secondary visibility and indirect illumination. The rendering method under direct illumination differs for far- and near-field lighting. For direct far-field lighting $L_{\\mathrm{dfar}}$ , we employ Monte Carlo integration to evaluate the rendering equation (Eq. (1)) as the incident light comes from all", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11804", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "directions:", + "bbox": [ + 76, + 90, + 150, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {C} _ {\\mathrm {p b , d f a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {o}} (\\mathbf {r} (t), - \\mathbf {d}) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\frac {L _ {\\mathrm {d f a r}} (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s}) f (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s} , - \\mathbf {d}) \\left(\\boldsymbol {\\omega} _ {s} \\cdot \\mathbf {n}\\right) ^ {+}}{p \\left(\\boldsymbol {\\omega} _ {s}\\right)}, \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 108, + 468, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S$ is the number of sampled directions, $\\omega_{s}$ a sampled incident direction, and $p(\\omega_{s})$ the probability density function (PDF) for the sampled direction. We use multiple importance sampling [26] to combine evaluations from different strategies: BRDF importance sampling [13], SG lighting importance sampling [33], and cosine importance sampling. we follow differentiable rendering works [15] to use a small $S = 20$ , as stochastic gradient descent handles noisy gradients. The incident radiance, now taking the visibility of light sources into account, is given by:", + "bbox": [ + 75, + 172, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d f a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}) = L _ {\\mathrm {f a r}} (\\boldsymbol {\\omega} _ {s}) V (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 330, + 468, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $V(\\mathbf{r}(t), \\boldsymbol{\\omega}_s) \\in [0,1]$ representing visibility, or the inverse of the accumulated opacity along a secondary ray $\\mathbf{r}_{\\mathrm{sec}}(t') = \\mathbf{r}(t) + t' \\boldsymbol{\\omega}_s$ , ( $t' > 0$ ), computed in a manner akin to Eq. (10). For direct near-field lighting $L_{\\mathrm{dnear}}$ , the integral in Eq. (1) simplifies to a multiplication, as the light originates from a single direction:", + "bbox": [ + 76, + 352, + 468, + 443 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d n e a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {d n e a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega}) f (\\mathbf {r} (t), \\boldsymbol {\\omega}, - \\mathbf {d}) \\left(\\boldsymbol {\\omega} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 448, + 468, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega = (\\mathbf{p} - \\mathbf{r}(t)) / \\| \\mathbf{p} - \\mathbf{r}(t)\\|_{2}$ is the direction of incident light, and $L_{\\mathrm{dnear}}(\\mathbf{r}(t),\\omega)$ is the potentially occluded incident radiance, computed similarly to Eq. (12).", + "bbox": [ + 75, + 479, + 468, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Indirect illumination $L_{\\mathrm{ind}}$ relevant to both far-field and near-field lighting, considers radiance reflected from the scene itself, potentially from all directions, thus necessitating Monte Carlo integration. The neural radiance field $\\mathbf{C}_{\\mathrm{rf}}$ substitutes multi-bounce path tracing for indirect illumination, given its encapsulation of the scene radiance under infinite lighting bounces. The radiance $\\mathbf{C}_{\\mathrm{pb,ind}\\{\\mathrm{far,near}\\}}$ from indirect illumination is evaluated similarly to Eq. (12), but with the $L_{\\mathrm{dfar}}$ replaced by the indirect lighting $L_{\\mathrm{ind}\\{\\mathrm{far,near}\\}}$ :", + "bbox": [ + 75, + 525, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i n d} *} (\\mathbf {r} (t), \\omega_ {s}) = \\mathbf {c} _ {*} \\left(\\mathbf {x} ^ {\\prime}, - \\omega_ {s}\\right) \\left(1 - V (\\mathbf {r} (t), \\omega_ {s})\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 681, + 468, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{x}'$ is the intersection of the secondary ray $\\mathbf{r}_{\\mathrm{sec}}(t')$ with the scene geometry. The radiance from indirect illumination is also computed for each lighting condition separately, then combined with direct illumination to yield the complete PBR radiance for the camera ray under a specific lighting condition:", + "bbox": [ + 75, + 703, + 468, + 794 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r , n e a r} \\}} = \\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d} \\{\\text {f a r , n e a r} \\}} + \\mathbf {C} _ {\\mathrm {p b}, \\text {i n d} \\{\\text {f a r , n e a r} \\}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 800, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 76, + 830, + 181, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Schemes. Our training process is divided into two sequential stages. In the first stage, the focus is on training the geometry and radiance field without employ", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing PBR. The objectives are: 1) to recover the scene's geometry, 2) to distinguish the appearance under each lighting condition, a necessity due to the potential presence of multiple light sources in one image, and 3) to utilize the reconstructed radiance field for indirect illumination. In the second stage, we train the material field and scene lighting using PBR, aiming to estimate material properties.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Functions. We average the loss functions across batches of camera rays. During the first stage, our loss function compares the total re-rendered radiance under all lighting conditions present in an image – including a far-field light $i$ and any active near-field lights – with the ground-truth (GT) color $\\mathbf{C}_{\\mathrm{gt}}$ from the input observation. This comparison trains the geometry and radiance field:", + "bbox": [ + 496, + 200, + 890, + 306 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r f}} = \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\text {f a r} _ {i}} + \\sum_ {\\text {n e a r} i \\text {i s o n}} \\mathbf {C} _ {\\mathrm {r f}, \\text {n e a r} _ {i}} - \\mathbf {C} _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 311, + 890, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Eikonal loss [8] is applied to the gradients of the SDF values $s$ for geometric regularization:", + "bbox": [ + 496, + 347, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e i k}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\left\\| \\frac {\\partial s (\\mathbf {r} (t _ {j}))}{\\partial \\mathbf {r} (t _ {j})} \\right\\| _ {2} - 1\\right) ^ {2}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 383, + 890, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A silhouette loss using GT alpha $\\alpha_{\\mathrm{gt}}$ from the foreground mask and a normal smoothness loss [41] aid shape recovery:", + "bbox": [ + 496, + 430, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s i l}} = \\left(\\left(1 - T _ {N + 1}\\right) - \\alpha_ {\\mathrm {g t}}\\right) ^ {2}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 467, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n s}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right)\\right) - \\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right) + \\boldsymbol {\\epsilon}\\right)\\right) ^ {2}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 488, + 890, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon \\in \\mathbb{R}^3$ is a small random perturbation. The total loss for stage one is $\\mathcal{L}_{\\mathrm{rf}} + \\lambda_{\\mathrm{eik}}\\mathcal{L}_{\\mathrm{eik}} + \\lambda_{\\mathrm{sil}}\\mathcal{L}_{\\mathrm{sil}} + \\lambda_{\\mathrm{ns}}\\mathcal{L}_{\\mathrm{ns}}$ , with $\\lambda_*$ representing the loss weights. In stage two, PBR colors replace radiance field-rendered colors for comparison with $\\mathbf{C}_{\\mathrm{gt}}$ , and a novel self-consistency loss is introduced between neural and PBR radiance under each lighting for additional supervision. The direct supervision loss $\\mathcal{L}_{\\mathrm{pb}}$ mirrors $\\mathcal{L}_{\\mathrm{rf}}$ with $\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{far,near}\\}_i}$ replaced by $\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{far,near}\\}_i}$ . The perlighting self-consistency loss $\\mathcal{L}_{\\mathrm{con}}$ is defined as:", + "bbox": [ + 496, + 537, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n}} = \\sum_ {i = 1} ^ {N _ {\\{\\text {f a r}, \\text {n e a r} \\}}} \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} - \\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 672, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This loss is crucial for discerning contributions from each lighting in images with multiple active light sources, thereby reducing ambiguity and improving material estimation. A material smoothness loss $\\mathcal{L}_{\\mathrm{ms}}$ is used similar to $\\mathcal{L}_{\\mathrm{ns}}$ . The total loss for stage two is $\\mathcal{L}_{\\mathrm{pb}} + \\lambda_{\\mathrm{con}}\\mathcal{L}_{\\mathrm{con}} + \\lambda_{\\mathrm{ms}}\\mathcal{L}_{\\mathrm{ms}}$ .", + "bbox": [ + 496, + 719, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is trained on an NVIDIA RTX 3090 GPU for a total of 40,000 steps. The first 30,000 steps of stage one take about 20 minutes, while the subsequent 10,000 steps of stage two require about 40 minutes. Post-training, we follow a procedure similar to WildLight [6] to extract the fields into textured meshes, which can be easily integrated into industry-standard rendering software like Blender [1],", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac4b766d47d1ca8f325230f955f1f0236da4f91fda954960aa459b98c77b5209.jpg", + "table_caption": [ + "Table 2. Quantitative comparison results with state-of-the-art methods averaged on 6 synthetic scenes. We show results of surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, the specular reflection part of FV relit RGB, and training time on a single RTX 3090 GPU. We mark the best and the second best results in each column. $\\uparrow (\\downarrow)$ means bigger (smaller) is better." + ], + "table_footnote": [], + "table_body": "
MethodInput lighting conditionsNormal MAngE↓AlbedoView synthesisFV relitFV relit (spec)Time
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
(1) TensoIR [10]17.6626.480.92129.480.91228.180.90128.300.861300 mins
(2) NVDiffRecMC [9]Single far-field16.2426.520.91527.130.91326.610.90125.570.836150 mins
(3) Ours12.3924.500.88228.200.93427.460.92127.560.87145 mins
(4) WildLight [6]Single far-field + Flashlight11.4928.860.94029.870.92930.440.93027.710.8631440 mins
(5) Ours10.8931.620.95332.090.95332.000.95330.750.90660 mins
(6) TensoIR [10]Two far-field16.2427.180.92929.680.91228.660.90228.460.863300 mins
(7) Ours11.7026.070.90229.590.94229.270.93429.090.89045 mins
(8) OursTwo far + Single near10.7932.040.95732.100.95032.380.95431.400.91060 mins
", + "bbox": [ + 80, + 133, + 888, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "enabling fast and high-quality rendering suitable for various applications. Check the supplement for more details.", + "bbox": [ + 76, + 316, + 470, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 359, + 209, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments involve a comprehensive comparison with state-of-the-art methods (Sec. 4.2), ablation studies (Sec. 4.3), and evaluations on both synthetic datasets (Sec. 4.1) and real-world images (Sec. 4.4). The supplement shows further data creation details and more experiments.", + "bbox": [ + 76, + 383, + 468, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Synthetic Datasets", + "text_level": 1, + "bbox": [ + 76, + 468, + 254, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We collect 6 synthetic scenes comprising a variety of scene geometries and materials. For training, each scene is rendered under four distinct lighting setups: 1) single far-field light, 2) single far-field light with a camera-collocated flashlight, 3) two far-field lights, and 4) two far-field lights with a near-field light source. We render 100 training images per setting using random viewpoints and per-image lighting conditions. The far-field lighting from the first setting is used in subsequent settings, with the last two sharing their far-field lighting. For testing, 200 images under one common far-field lighting are rendered to assess novel view synthesis under seen lighting, and another 200 under unseen far-field lighting to evaluate free-viewpoint (FV) relighting, which largely depends on the geometry and material estimation quality. This methodology provides a relatively fair comparison across methods with different abilities regarding supported input lighting conditions, trained under similar conditions and tested all using the same sets.", + "bbox": [ + 75, + 491, + 470, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 771, + 441, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with leading multi-view inverse rendering methods under each input lighting scenario. For single far-field lighting inputs (Tab. 1 row 1), we compare with NVDiffRecMC [9] and TensoIR [10] (single-lighting input). For multiple far-field lighting scenarios (Tab. 1 row 3), we compare with TensoIR [10]. For setups involving a far-field light and a flashlight (Tab. 1 row 4), we compare", + "bbox": [ + 76, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with WildLight [6]. We use Blender [1] for high-quality relit results for our method, WildLight [6], and NVDiffRecMC [9], importing trained SDF and material fields as textured meshes. Since TensoIR [10] employs density-based geometry and cannot extract high-quality meshes, we utilize its differentiable renderer for relighting.", + "bbox": [ + 496, + 316, + 890, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our quantitative comparison employs mean angular error (MAngE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) [30] metrics for surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, and specular reflection of FV relit RGB. We follow NeRFactor [41] in assuming albedo and lighting brightness as scale-invariant, necessitating metric comparisons using corresponding scales. For visualization and metrics, we scale each RGB channel of albedo and relit images from all methods by a global scale to minimize mean squared error against ground truth.", + "bbox": [ + 496, + 407, + 892, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results are detailed in Tab. 2. We observe that when only input images under single far-field lighting are available (rows 1-3), our method gives comparable results as TensoIR [10] and NVDiffRecMC [9]. Adding a flashlight (rows 4, 5) significantly helps in geometry and material estimation (comparing rows 3, 5), where the greatest improvement is made regarding the diffuse albedo. Also, our method clearly surpasses WildLight [6] in both quality and speed in this setting. Leveraging another far-field lighting condition (rows 6, 7) brings similar effects as using a flashlight, but generally with a lower degree of improvement. Further adding a flashlight (row 8) gives the best results among all settings and methods, but the gain is marginal compared to the setting of single far-field lighting with a flashlight (comparing rows 5, 8).", + "bbox": [ + 496, + 571, + 893, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 shows the qualitative comparison on two synthetic scenes: LEGO (left) with detailed geometry, and TROOPER (right) with highly-reflective SV materials. The results show that our method predicts more detailed and faithful shape, diffuse albedo, and specular parameters that could produce accurate and realistic relit results under unseen", + "bbox": [ + 496, + 799, + 893, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11806", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2fb9ddc485f21d9ae4368406011da2716409663ac6d6e3046cbad3fe42965c6.jpg", + "image_caption": [ + "Figure 3. Comparison with state-of-the-art methods. We test NVDiffRecMC [9] (single far-field lighting in input), TensoIR [10] (two far-field lighting in input), and WildLight [6] and our method (single far-field lighting with a flashlight in input) on two synthetic scenes. The intensity of the specular reflection is multiplied by $3 \\times$ for the visualization purpose. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 493, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee09e28916cb666d5eed9b22d6bf9faf6bea8a608629c3c7f21f6da5d7825cd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 885, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/417858a2c896cfcfa89dcd3d9f8c12898f6afdae47db8df7a6c7db8a9ba88a1b.jpg", + "table_caption": [ + "Table 3. Ablation study on our method using inputs under single far-field lighting with a flashlight." + ], + "table_footnote": [], + "table_body": "
ModelNormalAlbedoFV RelitRelit (spec)
MAngE↓PSNR↑PSNR↑PSNR↑
(1) w/o Lcon8.45227.7029.8529.08
(2) Mod. cnear8.57227.7930.0328.87
(3) w/o encmat8.26327.7930.0728.92
(4) Full model8.29828.0530.1929.36
", + "bbox": [ + 84, + 430, + 464, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lighting. In contrast, WildLight [6] tends to overly smooth geometry and material. TensoIR [10] can not handle highly-reflective surfaces, and lighting is baked into its predicted diffuse albedo for TROOPER.", + "bbox": [ + 75, + 532, + 468, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 599, + 240, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 3, we quantitatively evaluate ablation models on a subset of our synthetic scenes: 1) the model excluding the self-consistency loss $\\mathcal{L}_{\\mathrm{con}}$ , 2) the model where the output of the near-field radiance MLP is not explicitly multiplied by the cosine weakening factor and incident radiance (Eq. (8)), 3) the model with $\\mathrm{enc}_{\\mathrm{mat}}(\\mathbf{x})$ replaced by $\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x})$ . The results show that all the above techniques enhance the geometry and materials estimated by our method.", + "bbox": [ + 75, + 623, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Real-world Results", + "text_level": 1, + "bbox": [ + 76, + 750, + 261, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 shows the results of our method on two real-world scenes captured under single far-field lighting and a flashlight. Our method faithfully recovers the geometry and material of the objects and produces realistic relit results.", + "bbox": [ + 75, + 773, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a versatile multi-view inverse rendering framework, distinguished by its ability to leverage input im", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbf612ad1b7d69f94b9168c7654cd04c1bf2429980067f248be7b32e4ec3c024.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 398, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2170a37d6d11c10633e96c8b249bafbbbe6b0a96d45ffda7305aa22d64cfe820.jpg", + "image_caption": [ + "Figure 4. The results of our method on two real-world scenes, using multi-view images under one ambient lighting and a flashlight. The specular reflection intensity is increased for visualization." + ], + "image_footnote": [], + "bbox": [ + 501, + 515, + 890, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages under varied far- and near-field light sources available in capture for better geometry and material estimation.", + "bbox": [ + 498, + 684, + 890, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. VMINer does not model the background, necessitating a foreground mask for each image. It also does not consider unknown tone-mapping curves applied during the image signal processing (ISP) stage, using a fixed curve with $\\gamma = 2.2$ to the computed linear radiance. It depends on distinguishing the contribution from each lighting, which implies that, for optimal results, near-field light sources should be switched on/off during capture.", + "bbox": [ + 496, + 714, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 851, + 658, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grand No. 62136001, 62088102).", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Blender Foundation. The Blender project - free and open 3D creation software. https://www.blender.org. Accessed: 2023-11-07. 2, 6, 7", + "[2] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4", + "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. NeuralPIL: Neural pre-integrated lighting for reflectance decomposition. In Proc. of Neural Information Processing Systems (NeurIPS), 2021.", + "[4] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P. A. Lensch, and Varun Jampani. SAMURAI: shape and material from unconstrained real-world arbitrary image collections. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 2", + "[5] Brent Burley. Physically-based shading at disney. In SIGGRAPH 2012 Courses, 2012. 3", + "[6] Ziang Cheng, Junxuan Li, and Hongdong Li. WildLight: In-the-wild inverse rendering with a flashlight. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7, 8", + "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3", + "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proc. of International Conference on Machine Learning (ICML), 2020. 6", + "[9] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using Monte Carlo rendering and denoising. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 7, 8", + "[10] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 4, 7, 8", + "[11] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 1, 2", + "[12] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG), 41(4):56:1-56:12, 2022. 1, 2, 3", + "[13] Eric P Lafortune and Yves D Willems. Using the modified phong reflectance model for physically based rendering. Katholieke Universiteit Leuven. Departement Computerwetenschappen, 1994. 6" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates NeRFs. arXiv preprint arXiv:2305.04966, 2023. 5", + "[15] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable Monte Carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6): 222, 2018. 6", + "[16] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[17] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Milos Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In Proc. of European Conference on Computer Vision (ECCV), 2022. 3", + "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1, 4", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 1, 4, 5", + "[21] Jacob Munkberg, Wenzheng Chen, Jon Hasselgren, Alex Evans, Tianchang Shen, Thomas Müller, Jun Gao, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[22] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 1", + "[23] Viktor Rudnev, Mohamed Elgharib, William A. P. Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3", + "[24] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3", + "[25] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. ReLight My NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11808", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 6", + "[27] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. In Proc. of the Eurographics Symposium on Rendering Techniques, 2007. 5", + "[28] Jiaping Wang, Peiran Ren, Minmin Gong, John M. Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM Transactions on Graphics (TOG), 28(5):133, 2009. 4", + "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proc. of Neural Information Processing Systems (NeurIPS), 2021. 4, 5", + "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7", + "[31] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[32] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3", + "[33] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 6", + "[34] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 2", + "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3", + "[36] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(8):10129-10142, 2023. 1, 2, 3", + "[37] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 3", + "[38] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++:" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Inter-reflectable light fields for geometry and material estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3", + "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4", + "[40] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: inverse rendering by optimizing neural SDFs and materials from photometric images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[41] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 2021. 1, 2, 3, 4, 6, 7", + "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4", + "[43] Youjia Zhang, Teng Xu, Junqing Yu, Yuteng Ye, Yanqing Jing, Junle Wang, Jingyi Yu, and Wei Yang. NeMF: Inverse volume rendering with neural microflake field. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[44] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG), 41(6):235:1-235:17, 2022. 4, 5", + "[45] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable Monte Carlo raytracing. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2022. 3", + "[46] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Di-anbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I $^2$ -SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3", + "[47] Yongjie Zhu, Yinda Zhang, Si Li, and Boxin Shi. Spatially-varying outdoor lighting estimation from intrinsics. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11809", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_model.json b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c0bdbfa398d50f7d48f701efbd0f38a90b0c5fb --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.131, + 0.737, + 0.177 + ], + "angle": 0, + "content": "VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.2, + 0.341, + 0.218 + ], + "angle": 0, + "content": "Fan Fei\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.201, + 0.482, + 0.22 + ], + "angle": 0, + "content": "Jiajun Tang\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.201, + 0.593, + 0.22 + ], + "angle": 0, + "content": "Ping Tan3,4" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.201, + 0.72, + 0.218 + ], + "angle": 0, + "content": "Boxin Shi\\(^{1,2\\#}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.224, + 0.862, + 0.28 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University \\(^{2}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University \\(^{3}\\)Hong Kong University of Science and Technology \\(^{4}\\)Light Illusions" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.301, + 0.314, + 0.317 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.474, + 0.584 + ], + "angle": 0, + "content": "This paper introduces a versatile multi-view inverse rendering framework with near- and far-field light sources. Tackling the fundamental challenge of inherent ambiguity in inverse rendering, our framework adopts a lightweight yet inclusive lighting model for different near- and far-field lights, thus is able to make use of input images under varied lighting conditions available during capture. It leverages observations under each lighting to disentangle the intrinsic geometry and material from the external lighting, using both neural radiance field rendering and physically-based surface rendering on the 3D implicit fields. After training, the reconstructed scene is extracted to a textured triangle mesh for seamless integration into industrial rendering software for various applications. Quantitatively and qualitatively tested on synthetic and real-world scenes, our method shows superiority to state-of-the-art multi-view inverse rendering methods in both speed and quality." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.598, + 0.21, + 0.614 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.88 + ], + "angle": 0, + "content": "The reconstruction of 3D scenes from multi-view RGB imagery has experienced significant advancements following the development of Neural Radiance Fields (NeRF) [19], and both the speed and quality of the reconstruction have reached an unprecedented level [18, 20]. Despite these improvements, a common limitation is to represent only the radiance field, which is a complex product of the external lighting interacting with the intrinsic geometry and material of the scene [11]. The entangled nature of this representation generally hampers the ability to accurately render the scene under unseen lighting conditions, because the influence of the original lighting is embedded within the newly rendered scene. Recent approaches [24, 41, 46] have incorporated inverse rendering [22] to separate material properties and lighting effects, extending the application of the reconstruction beyond novel view synthesis to novel scenarios involving relighting and material editing [2]. Some" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.301, + 0.67, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.301, + 0.891, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.405, + 0.66, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.405, + 0.891, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.521, + 0.893, + 0.592 + ], + "angle": 0, + "content": "Figure 1. Given multi-view images under possibly varied lighting, our method leverages all present lighting conditions to reconstruct scene geometry and material disentangled with lighting. The trained fields are extracted to textured meshes for seamless integration into industrial renderers for various applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.674 + ], + "angle": 0, + "content": "methods [2, 6, 9, 21, 32] have advanced to the extraction of detailed triangle meshes with material UV textures, which can serve as economical and lifelike 3D models for gaming and cinematography industries [9, 21], thereby marking a transformative step in digital asset creation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Inverse rendering presents several fundamental challenges, one of them being its severe inherent ambiguity [41]. To combat this, multi-view inverse rendering methodologies usually reduce the ambiguity by imposing various constraints on each scene component. These methods can be categorized based on the assumption on the amount and types of lighting conditions present in the input images (Tab. 1). The majority of methods [9, 21, 24, 35, 36, 39, 41, 42] assumes the imagery to be captured under one fixed lighting condition (rows 1, 2). However, in such scenarios, only one case of scene appearances among all possible cases under different lighting is observed, causing severe lighting-material ambiguity, posing considerable difficulties to material estimation. In a bid to alleviate this ambiguity, recent techniques [2, 6, 10, 12, 23, 32] utilize input" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.376, + 0.901 + ], + "angle": 0, + "content": "Corresponding author. E-mail: shiboxin@pku.edu.cn." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11800" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.091, + 0.867, + 0.105 + ], + "angle": 0, + "content": "Table 1. A summary of multi-view inverse rendering methods based on their assumptions and usages of input lighting conditions." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.108, + 0.887, + 0.242 + ], + "angle": 0, + "content": "
Methods# of far-field light# of near-field lightMaterial optimizationLighting-material ambiguityCapture workload
(1) [24, 39, 41, 42]SingleNoneOnly from singleStrongLow
(2) NeILF [35]Single 5D incident light field
(3) [2, 10, 12, 32]aMultipleNoneOnly from far-fieldModerateHigh
(4) WildLight [6]SingleSingleOnly from near-fieldModerateLow
(5) OursSingleSingleFrom near- and far-fieldWeakLow
MultipleSingle or multipleAlmost noneHigh
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.096, + 0.243, + 0.847, + 0.256 + ], + "angle": 0, + "content": "\\( {}^{a} \\) These methods also support input images under single far-field lighting. In that case, they belong to single far-field lighting methods (row 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.518 + ], + "angle": 0, + "content": "images under varied lighting to discern materials from lighting (rows 3, 4). This demands more complex lighting models. Existing varied-lighting methods [2, 10, 12, 23, 32] typically model multiple, but only far-field, lighting conditions to maintain manageable unknown lighting parameters. This means throughout the capture, the lighting setups should be largely changed by either moving the objects to reconstruct or adapting far-field light sources, causing increased workload to the data collection process. Alternatively, near-field light sources, such as flashlights, can function as easily controllable light sources for disambiguation. WildLight [6] suggests the use of a camera-collocated flashlight in addition to the ambient lighting for inverse rendering. Nevertheless, it limits itself by using only the isolated appearance under the flashlight for material estimation, not harnessing available observations under the ambient lighting and degrading the quality of its estimated material (row 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Addressing the challenges and restrictions previously outlined, this paper introduces VMINer, a Versatile Multiview Inverse rendering framework with Near- and far-field light sources (row 5). The distinctive feature of our framework is that it uses a lightweight yet inclusive lighting model for different far- and near-field light sources and efficiently leverages available observations under each lighting to disentangle lighting from other components, making the most of whatever lighting conditions are at hand to boost the practicality and quality of the reconstruction, as shown in Fig. 1. Although our method accepts the simplest setting of inputs under single far-field lighting, a more effective compromise is to employ an additional flashlight to enhance quality without overly complicating the capture process. The ideal scenario for the highest quality reconstruction would involve capturing the scene under varied far-field lighting and using a flashlight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.895 + ], + "angle": 0, + "content": "Nonetheless, implementing such a versatile framework is far from straightforward. The method has to render the scene under different lighting conditions, including spatially-varying and changing ones to represent possibly moving near-field lights. We propose to solve the problem by modeling near-field lights as point sources in 3D space with adjustable positions and intensities, which could be anisotropic and not necessarily camera-aligned (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.895, + 0.504 + ], + "angle": 0, + "content": "stationary desk lamps). This representation effectively approximates a broad range of common near-field lights and is easy for spatially-consistent editing and optimization. We use 3D implicit fields to model the scenes' radiance, shape, and material and minimize the error of the neural or physically-based re-rendered scene appearance under each input lighting. To account for multiple lighting situations and moving near-field lights, we integrate specially tailored multi-layer perceptrons (MLPs) that additionally process light directions and embeddings. Post-training, the fields are converted into a textured triangle mesh, ready for seamless integration into industrial rendering software like Blender [1], facilitating a broad spectrum of applications. Our extensive experiments, covering both synthetic and real-world scenes, demonstrate that VMINer surpasses prior methods quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.642, + 0.535 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.894, + 0.621 + ], + "angle": 0, + "content": "Preliminary. The rendering equation [11] models the reflected radiance from any point on a surface as a result of an intricate surface integral. This integral comprises contributions from three fundamental scene components: lighting, material, and geometry. The equation is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.625, + 0.892, + 0.667 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) = \\int_ {\\Omega} \\underbrace {L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}})} _ {\\text {l i g h t i n g}} \\underbrace {f (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}})} _ {\\text {m a t e r i a l}} \\underbrace {(\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) ^ {+}} _ {\\text {g e o m e t r y}} \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.838 + ], + "angle": 0, + "content": "In this equation, \\( L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}}) \\) represents the radiance reflected in the outgoing direction \\( \\omega_{\\mathrm{o}} \\) from a surface point \\( \\mathbf{x} \\) in 3D space. \\( \\Omega \\) denotes the unit hemisphere encompassing all incident directions \\( \\omega_{\\mathrm{i}} \\) with \\( \\omega_{\\mathrm{i}} \\cdot \\mathbf{n} > 0 \\). \\( L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}}) \\) is the incident radiance at \\( \\mathbf{x} \\) from direction \\( \\omega_{\\mathrm{i}} \\), and \\( f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}}) \\) is the bidirectional reflectance distribution function (BRDF), denoting the ratio of light reflected along \\( \\omega_{\\mathrm{o}} \\) at \\( \\mathbf{x} \\) from \\( \\omega_{\\mathrm{i}} \\). The term \\( (\\omega_{\\mathrm{i}} \\cdot \\mathbf{n})^{+} \\), where \\( x^{+} \\triangleq \\max(x, 0) \\), signifies the cosine weakening factor. The incident lighting \\( L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}}) \\) may originate from other points within the scene, necessitating recursive evaluation of this integral." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Multi-view Inverse Rendering. Multi-view inverse rendering techniques [2-4, 6, 9, 10, 12, 21, 23, 24, 34-36, 38-43] generally employ implicit or explicit 3D fields to model the scene's geometry and material, while also estimating" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.08, + 0.325, + 0.891, + 0.341 + ], + "angle": 0, + "content": ": input : variable : refer to another entry : untrainable operator : trainable operator / MLP \\(\\bigotimes \\bigoplus\\) element-wise multiplication / add" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Figure 2. Bottom-left: VMINer takes as input multi-view RGB images with foreground masks under possibly varied far- and near-field lighting. The total number of lighting conditions and their types should be given. Each image is illuminated by one far-field lighting whose index is \"far index\", and near-field lights whose on/off states are \"near on/off\". Top-left: VMINer models each far-field lighting \\((\\mathrm{far}_i)\\) and near-field lighting \\((\\mathrm{near}_i)\\) separately as parametric models, from which incoming radiance is queried at any position \\(\\mathbf{x}\\) and direction \\(\\omega\\). Bottom-mid: It utilizes 3D implicit representations for scene geometry, material, and radiance. Four MLPs process \\(\\mathbf{x}\\) and view direction \\(\\mathbf{v}\\) encoded by different encoders \\(\\mathrm{enc}_{\\{\\mathrm{geo},\\mathrm{mat},\\mathrm{dir}\\}}\\) with light embeddings \\(\\mathbf{F}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}\\) to get the SDF \\(s\\), the neural appearance descriptor \\(\\mathbf{z}\\), the BRDF parameters \\(\\beta\\), and the radiance \\(\\mathbf{c}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}\\) under each lighting. Bottom-right: It then uses radiance field rendering to re-render the appearance \\(\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}\\) separately under each lighting condition, added up to \\(\\mathbf{C}_{\\mathrm{rf}}\\) according to the per-image lighting condition. Top-right: Physically-based surface rendering uses Monte Carlo integration to evaluate both direct illumination \\(L_{\\mathrm{d}\\{\\mathrm{near},\\mathrm{far}\\}_i}\\) with secondary visibility and indirect illumination \\(L_{\\mathrm{ind}\\{\\mathrm{near},\\mathrm{far}\\}_i}\\). The rendered appearances are aggregated to \\(\\mathbf{C}_{\\mathrm{pb}}\\). \\(\\mathbf{C}_{\\{\\mathrm{rf},\\mathrm{pb}\\}}\\) are compared with the observation \\(\\mathbf{C}_{\\mathrm{gt}}\\) to train the scene model. \\(\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}\\) are also used as additional supervision signals for \\(\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{near},\\mathrm{far}\\}_i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.473, + 0.897 + ], + "angle": 0, + "content": "lighting to recreate the scene's appearance. They optimize each scene component from scratch by aligning the rendered with the observed appearance, while avoiding potential local minima, which, although do not faithfully represent the scene, can replicate the observed input. Addressing this inherently ambiguous nature of inverse rendering, these methods typically impose constraints on each scene component. For geometry and material, assumptions of known geometry [35, 36, 46] are made, along with leveraging the recovered shape from radiance fields [6, 10, 41], adopting Lambertian or uniform materials [23, 39], controlling BRDF parameter smoothness [10], or integrating a learned BRDF latent space [41]. Regarding lighting, many approaches presuppose known lighting conditions [24, 40] or fixed lighting shared across input images [35, 38, 41, 42]. More recent techniques [2, 6, 10, 12, 23, 32] have turned to using images with varied lighting conditions to mitigate the lighting-material ambiguity. They typically accept input images that can be grouped into sets, each illuminated by one distinct and fixed far-field lighting. WildLight [6] exceptionally employs a camera-collocated flashlight along with fixed ambient lighting. However, it does not model the ambient lighting and thus its material property estimation relies exclusively on flashlight observations, where the light and view direction always coincide [5], limiting its BRDF estimation ability. Our VMINer, in contrast, accommodates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.549 + ], + "angle": 0, + "content": "input images under single or multiple far-field lighting conditions with near-field light sources and leverages each of them to enhance the quality of the estimated material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.807 + ], + "angle": 0, + "content": "Lighting Models. Spatially-uniform (SU) lighting models rely on the assumption that the scene's lighting originates from a distant source, modeled as spherical Gaussians (SGs) [2, 10], spherical harmonics (SH) [12, 23], or MLPs [32]. In contrast, spatially-varying (SV) lighting models acknowledge the presence of near-field light sources, causing different locations to receive different incident lighting. They include parametric 3D lights [7, 17, 36, 37], outgoing light fields such as volumetric SGs [31] and neural out-of-view lighting volumes [45], and incident light fields like SV environment maps [47], SVSGs [16], and neural incident light fields (NeILF) [35, 38]. Our VMINer integrates distant lights as SGs and near-field lights as point lights with adjustable position and intensity. It is less complex compared to other SV lighting models, facilitating spatially consistent editing and optimization, while being able to approximate common lighting setups in real life." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.673, + 0.835 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.9 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, VMINer reconstructs the scene lighting (Sec. 3.1) and 3D fields (Sec. 3.2) from input images using differentiable rendering (Sec. 3.3). The training scheme and loss functions are described in Sec. 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11802" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.307, + 0.108 + ], + "angle": 0, + "content": "3.1. Versatile Lighting Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.145 + ], + "angle": 0, + "content": "This subsection models the lighting term in Eq. (1) and corresponds to the top-left part of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.144, + 0.471, + 0.372 + ], + "angle": 0, + "content": "VMINer harnesses the diversity of lighting conditions present in the input RGB images to disentangle lighting effects. The lighting model plays a crucial role here: 1) Since lighting conditions are unknown initially, they must be optimized alongside other scene parameters using physically-based rendering during training. 2) Different lighting setups, each compatible with the model, are necessary during the capture phase to achieve this goal. Given these prerequisites, it is essential to design a lighting model that accommodates common far-field and near-field light sources, while being lightweight enough for efficient and robust optimization and physically-based rendering. With these considerations, we propose a versatile lighting model that is both inclusive and lightweight. This representation models far-field and near-field lighting separately as below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.374, + 0.471, + 0.526 + ], + "angle": 0, + "content": "Far-field Lighting. In our model, each far-field lighting condition \\(i\\), where \\(i\\) belongs to the set \\(\\mathbb{N}_{N_{\\mathrm{far}}} \\triangleq \\{1, \\dots, N_{\\mathrm{far}}\\}\\) and \\(N_{\\mathrm{far}}\\) denotes the number of far-field lighting, is represented using 128-lobe SGs [28] commonly adopted in existing far-field lighting techniques [33, 39, 42]. For each lobe \\(j\\) of lighting \\(i\\), there are six parameters: lobe axis \\(\\pmb{\\xi}_{ij} \\in \\mathbb{S}^2\\), lobe sharpness \\(\\lambda_{ij} \\in \\mathbb{R}_+\\), and lobe RGB amplitude \\(\\pmb{\\mu}_{ij} \\in \\mathbb{R}_+^3\\). The incident radiance from far-field lighting \\(i\\) along direction \\(\\omega\\) at any 3D position \\(\\mathbf{x}\\) is calculated using the formula (ignoring visibility for now):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.53, + 0.47, + 0.573 + ], + "angle": 0, + "content": "\\[\nL _ {\\operatorname {f a r} _ {i}} (\\boldsymbol {\\omega}) = \\sum_ {j = 1} ^ {1 2 8} c _ {i j} \\boldsymbol {\\mu} _ {i j} e ^ {\\lambda_ {i} (\\boldsymbol {\\omega} \\cdot \\boldsymbol {\\xi} _ {i j} - 1)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.579, + 0.47, + 0.626 + ], + "angle": 0, + "content": "where \\( c_{ij} = \\lambda_{ij} / (2\\pi (1 - e^{-2\\lambda_{ij}})) \\) acts as the normalization factor related to roughness. Notably, \\( \\mathbf{x} \\) does not appear as an input to \\( L_{\\mathrm{far}_i} \\), reflecting the nature of distant lighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Near-field Lighting. Each near-field lighting \\(i\\), included in the set \\(\\mathbb{N}_{N_{\\mathrm{near}}}\\) with \\(N_{\\mathrm{near}}\\) representing the number of near-field light sources, is modeled as a point light. These point lights can have moving positions \\(\\mathbf{p}_i \\in \\mathbb{R}^3\\) and exhibit anisotropic radiation characterized by \\(l^{\\mathrm{th}}\\)-order (\\(l\\) can be 0, 1, or 2) SH coefficients \\(\\mathbf{h}_i \\in \\mathbb{R}^{3 \\times (l + 1)^2}\\). Our method accommodates two types of near-field lighting: camera-collocated lights, positioned at the camera ray origin for each image, and stationary lights, which remain fixed across all images. We observe that in neural radiance field rendering, the radiance under a stationary near-field light, especially when the light is active in all images, can be challenging to distinguish from radiance under far-field lighting. Also, the radiance from these two sources aids in material estimation in a similar way. Therefore, in practical applications, we favor using a moving flashlight as the near-field light to provide unique information for material estimation. The incident radiance from near-field lighting \\(i\\) at a 3D po" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.107 + ], + "angle": 0, + "content": "sition \\(\\mathbf{x}\\) is computed as (also ignoring visibility for now):" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.112, + 0.892, + 0.16 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n e a r} i} (\\mathbf {x}, \\boldsymbol {\\omega}, \\mathbf {p} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathrm {S H} (\\boldsymbol {\\omega} ; \\mathbf {h} _ {i})}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2} ^ {2}} & \\text {i f} \\boldsymbol {\\omega} = \\frac {\\mathbf {p} _ {i} - \\mathbf {x}}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.895, + 0.332 + ], + "angle": 0, + "content": "where \\(\\mathrm{SH}(\\omega ;\\mathbf{h})\\) calculates the SH at direction \\(\\omega\\) with coefficients \\(\\mathbf{h}\\), and \\(1 / \\| \\mathbf{p}_i - \\mathbf{x}\\|_2^2\\) signifies the inverse-square lighting attenuation for point lights. Here we include \\(\\mathbf{p}_i\\) as an input because it may be set differently across images. It is important to note that with respect to the incident direction \\(\\omega\\), \\(L_{\\mathrm{far}}\\) is a continuous function, while \\(L_{\\mathrm{near}}\\) is a discrete function, being non-zero only in a single direction. As a result, to render appearances under direct lighting, Monte Carlo integration is essential for far-field lighting, while a simple multiplication suffices for near-field lighting. Further details about this process are provided in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.854, + 0.353 + ], + "angle": 0, + "content": "3.2. Geometry, Material, and Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.391 + ], + "angle": 0, + "content": "This subsection models the material and geometry terms in Eq. (1) and corresponds to the bottom-mid part of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Geometry. Multi-view reconstruction methodologies generally hinge on two geometry representations: volume density [2, 10, 19, 41] and the signed distance function (SDF) [33, 39, 42, 44] fields. We choose the SDF field for our geometry representation due to its clearly defined surface at the zero-level isosurface, which simplifies and enhances the post-training mesh extraction process. Our approach utilizes implicit fields as MLPs over explicit structures like voxel grids for their compactness and flexibility. For each 3D position \\(\\mathbf{x}\\), a multi-resolution hash grid [20] is first employed for positional encoding, yielding a feature vector \\(\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x}) \\in \\mathbb{R}^{16}\\). The geometry MLP \\(\\mathrm{M}_{\\mathrm{geo}}\\) then predicts the SDF value at \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.593, + 0.891, + 0.61 + ], + "angle": 0, + "content": "\\[\ns (\\mathbf {x}), \\mathbf {z} (\\mathbf {x}) = \\mathrm {M} _ {\\mathrm {g e o}} (\\mathbf {x}, \\operatorname {e n c} _ {\\mathrm {g e o}} (\\mathbf {x})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.721 + ], + "angle": 0, + "content": "where \\( s \\in \\mathbb{R} \\) represents the signed distance (positive outside the surface, negative inside), and \\( \\mathbf{z} \\in \\mathbb{R}^{13} \\) is a descriptor of local appearance. To maintain differentiability in rendering and to aid in geometry optimization, during training we apply NeuS [29] techniques to transform SDF values along rays into volume densities. The surface normal \\( \\mathbf{n} \\in \\mathbb{S}^2 \\) is derived as the gradient of the SDF \\( s \\) with respect to \\( \\mathbf{x} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.724, + 0.892, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\partial s}{\\partial \\mathbf {x}} / \\left\\| \\frac {\\partial s}{\\partial \\mathbf {x}} \\right\\| _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.764, + 0.892, + 0.794 + ], + "angle": 0, + "content": "\\(\\mathrm{M}_{\\mathrm{geo}}\\) only approximates a strict SDF field that has \\(\\left\\| \\frac{\\partial s}{\\partial \\mathbf{x}} \\right\\|_2 = 1\\), so we normalize \\(\\mathbf{n}\\) to ensure it is a unit vector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Material. For accurate scene material recovery using physically-based re-rendering, it is imperative to model the material in such a way that the renderer can query the BRDF at any point on the surface. Considering that material properties are independent of lighting and view directions, we represent the scene's material through an implicit 3D field, much like the SDF field utilized for geometry. To achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11803" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "this, we employ another multi-resolution hash grid \\(\\mathrm{enc}_{\\mathrm{mat}}\\) for positional encoding. We adopt the simplified GGX BRDF model [27] with a fixed fresnel parameter to help alleviate ambiguity. The material MLP, \\(\\mathbf{M}_{\\mathrm{mat}}\\), is tasked with predicting the SVBRDF parameters \\(\\beta (\\mathbf{x})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.174, + 0.469, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\beta} (\\mathbf {x}) = \\operatorname {M} _ {\\text {m a t}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {m a t}} (\\mathbf {x})). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.195, + 0.473, + 0.515 + ], + "angle": 0, + "content": "Radiance. Although our method does not strictly require a radiance field for rendering scene appearance - given that appearance could be rendered solely using physically-based surface rendering (PBR) - we find that incorporating a radiance field at the start of training significantly eases the optimization of scene geometry and appearance. PBR, while physically accurate, tends to introduce instability and slow down the training process due to its inherent ambiguity and computational intensity. Moreover, the radiance field-rendered results can serve as supplementary supervision, aiding PBR in better separating scene appearance under each lighting condition and thus further diminishing lighting-material ambiguity. Hence, we deploy radiance MLPs to predict view-dependent and lighting-dependent radiance at each position. For each lighting condition \\(i\\), a light embedding \\(\\mathbf{F}_{\\{\\mathrm{far,near}\\}_i} \\in \\mathbb{R}^{16}\\) is utilized. The view direction \\(\\mathbf{v} \\in \\mathbb{S}^2\\) is processed through directional encoding that projects it onto the coefficients of the \\(3^{\\mathrm{rd}}\\)-order SH basis, yielding \\(\\mathrm{enc}_{\\mathrm{dir}}(\\mathbf{v}) \\in \\mathbb{R}^{4^2 = 16}\\) [20]. The far-field radiance MLP \\(\\mathrm{M}_{\\mathrm{far}}\\) computes the outgoing radiance \\(\\mathbf{c}_{\\mathrm{far}_i}\\) from position \\(\\mathbf{x}\\) along view direction \\(\\mathbf{v}\\) under far-field lighting \\(i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.521, + 0.469, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\text {f a r} _ {i}} (\\mathbf {x}, \\mathbf {v}) = \\mathrm {M} _ {\\text {f a r}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\text {f a r} _ {i}}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.469, + 0.587 + ], + "angle": 0, + "content": "where we incorporate the surface normal \\(\\mathbf{n}\\) as an additional MLP input, following Instant-NSR [44] and WildLight [6], as this has shown to aid shape recovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.47, + 0.741 + ], + "angle": 0, + "content": "The near-field radiance MLP differs slightly, as nearfield lighting can move in different images, thus the current light position \\(\\mathbf{p}_i\\) is a necessary input. Additionally, the radiance under near-field point light takes the form of the rendering equation Eq. (1) without the integral, allowing the incoming radiance and cosine term to be explicitly included. We follow ReNe [25] to input the relative direction of the point light \\(\\omega_{i} = \\frac{\\mathbf{p}_{i} - \\mathbf{x}}{\\|\\mathbf{p}_{i} - \\mathbf{x}\\|_{2}}\\) instead of the absolute position \\(\\mathbf{p}_i\\) into the MLP. Therefore, the outgoing neural radiance \\(\\mathbf{c}_{\\mathrm{near}_i}\\) under near-field lighting \\(i\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {c} _ {\\mathrm {n e a r} _ {i}} (\\mathbf {x}, \\mathbf {v}, \\mathbf {p} _ {i}) = \\mathrm {M} _ {\\mathrm {n e a r}} \\left(\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\mathrm {e n c} _ {\\mathrm {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\mathrm {n e a r} _ {i}}, \\boldsymbol {\\omega} _ {i}\\right) \\\\ \\otimes \\frac {L _ {\\text {n e a r} _ {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {i})}{\\left\\| \\mathbf {p} _ {i} - \\mathbf {x} \\right\\| _ {2} ^ {2}} \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.47, + 0.895 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes element-wise multiplication. On the right side, the second term denotes the incident radiance attenuated by square distance, the third term denotes the cosine weakening factor, and \\(\\mathrm{M}_{\\mathrm{near}}\\) accounts for other factors like reflectance and visibility. Owing to the instability of these terms at the training's outset, we use an annealing strategy," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "as recommended in prior works [44], gradually replacing default values (e.g., 1 for incident radiance) with those optimized during training, to ease through the process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.156, + 0.729, + 0.172 + ], + "angle": 0, + "content": "3.3. Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.178, + 0.892, + 0.33 + ], + "angle": 0, + "content": "To simplify the notation, we omit \\(\\mathbf{p}_i\\) and the lighting index \\(i\\) in the equations within this subsection. Our approach optimizes the lighting alongside the implicit fields primarily by reducing the discrepancy between the observed and differentiably re-rendered appearances. The re-rendering is accomplished using two types of renderer: a neural radiance field renderer (Fig. 2 bottom-right) and a physically-based surface renderer (Fig. 2 top-right; Eq. (1)). We handle each lighting condition separately in rendering, and then aggregate the results based on per-image lighting condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Volume Rendering for SDF Field. Consider a camera ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) (\\(t > 0\\)), with \\(\\mathbf{o} \\in \\mathbb{R}^3\\) as its origin and \\(\\mathbf{d} \\in \\mathbb{S}^2\\) as its direction. We incorporate importance sampling using an occupancy grid [14] to get \\(N \\leq 1024\\) points along this ray, denoted as \\(\\mathbf{r}(t_i)\\), \\(i \\in \\mathbb{N}_N\\). Utilizing techniques from NeuS [29], we transform the signed distances \\(s(\\mathbf{r}(t_i))\\) into discrete opacity values as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.448, + 0.892, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\left(\\frac {\\Phi_ {b} (s (\\mathbf {r} (t _ {i}))) - \\Phi_ {b} (s (\\mathbf {r} (t _ {i + 1})))}{\\Phi_ {b} (s (\\mathbf {r} (t _ {i})))}\\right) ^ {+}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\(\\Phi_b(s) = 1 / (1 + e^{-bs})\\) is the cumulative opacity distribution function. The parameter \\(b\\) is trainable and tends to increase during training, focusing opacity more narrowly around the surface where \\(s = 0\\). The accumulated transmittance is then calculated as \\(T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Neural Radiance Field Rendering. The color \\(\\mathbf{C}_{\\mathrm{rf}}\\) of a camera ray under either far- or near-field lighting conditions, as rendered through the neural radiance field, is determined by accumulating the radiance along the ray's path:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.638, + 0.892, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r , n e a r} \\}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {\\{\\text {f a r , n e a r} \\}} (\\mathbf {r} (t _ {i}), - \\mathbf {d}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.713 + ], + "angle": 0, + "content": "This formula sums the radiance from each sampled point to yield the color of the ray under the specified lighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Physically-based Rendering. Our method employs differentiable surface rendering, modeling the observed radiance as reflected from a single surface point. We compute the depth \\( t \\) of this surface point similarly to Eq. (10): \\( t = \\sum_{i=1}^{N} T_i \\alpha_i t_i \\). Accumulation along the ray also applies to the BRDF parameters \\( \\beta \\) and the surface normal \\( n \\). For each input lighting, we render the appearance that models secondary visibility and indirect illumination. The rendering method under direct illumination differs for far- and near-field lighting. For direct far-field lighting \\( L_{\\mathrm{dfar}} \\), we employ Monte Carlo integration to evaluate the rendering equation (Eq. (1)) as the incident light comes from all" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11804" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.151, + 0.105 + ], + "angle": 0, + "content": "directions:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.109, + 0.469, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {C} _ {\\mathrm {p b , d f a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {o}} (\\mathbf {r} (t), - \\mathbf {d}) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\frac {L _ {\\mathrm {d f a r}} (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s}) f (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s} , - \\mathbf {d}) \\left(\\boldsymbol {\\omega} _ {s} \\cdot \\mathbf {n}\\right) ^ {+}}{p \\left(\\boldsymbol {\\omega} _ {s}\\right)}, \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.174, + 0.47, + 0.326 + ], + "angle": 0, + "content": "where \\( S \\) is the number of sampled directions, \\( \\omega_{s} \\) a sampled incident direction, and \\( p(\\omega_{s}) \\) the probability density function (PDF) for the sampled direction. We use multiple importance sampling [26] to combine evaluations from different strategies: BRDF importance sampling [13], SG lighting importance sampling [33], and cosine importance sampling. we follow differentiable rendering works [15] to use a small \\( S = 20 \\), as stochastic gradient descent handles noisy gradients. The incident radiance, now taking the visibility of light sources into account, is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.331, + 0.469, + 0.348 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d f a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}) = L _ {\\mathrm {f a r}} (\\boldsymbol {\\omega} _ {s}) V (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.47, + 0.444 + ], + "angle": 0, + "content": "with \\( V(\\mathbf{r}(t), \\boldsymbol{\\omega}_s) \\in [0,1] \\) representing visibility, or the inverse of the accumulated opacity along a secondary ray \\( \\mathbf{r}_{\\mathrm{sec}}(t') = \\mathbf{r}(t) + t' \\boldsymbol{\\omega}_s \\), (\\( t' > 0 \\)), computed in a manner akin to Eq. (10). For direct near-field lighting \\( L_{\\mathrm{dnear}} \\), the integral in Eq. (1) simplifies to a multiplication, as the light originates from a single direction:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.449, + 0.469, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d n e a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {d n e a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega}) f (\\mathbf {r} (t), \\boldsymbol {\\omega}, - \\mathbf {d}) \\left(\\boldsymbol {\\omega} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.469, + 0.526 + ], + "angle": 0, + "content": "where \\(\\omega = (\\mathbf{p} - \\mathbf{r}(t)) / \\| \\mathbf{p} - \\mathbf{r}(t)\\|_{2}\\) is the direction of incident light, and \\(L_{\\mathrm{dnear}}(\\mathbf{r}(t),\\omega)\\) is the potentially occluded incident radiance, computed similarly to Eq. (12)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.469, + 0.679 + ], + "angle": 0, + "content": "Indirect illumination \\(L_{\\mathrm{ind}}\\) relevant to both far-field and near-field lighting, considers radiance reflected from the scene itself, potentially from all directions, thus necessitating Monte Carlo integration. The neural radiance field \\(\\mathbf{C}_{\\mathrm{rf}}\\) substitutes multi-bounce path tracing for indirect illumination, given its encapsulation of the scene radiance under infinite lighting bounces. The radiance \\(\\mathbf{C}_{\\mathrm{pb,ind}\\{\\mathrm{far,near}\\}}\\) from indirect illumination is evaluated similarly to Eq. (12), but with the \\(L_{\\mathrm{dfar}}\\) replaced by the indirect lighting \\(L_{\\mathrm{ind}\\{\\mathrm{far,near}\\}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.682, + 0.469, + 0.7 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i n d} *} (\\mathbf {r} (t), \\omega_ {s}) = \\mathbf {c} _ {*} \\left(\\mathbf {x} ^ {\\prime}, - \\omega_ {s}\\right) \\left(1 - V (\\mathbf {r} (t), \\omega_ {s})\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}'\\) is the intersection of the secondary ray \\(\\mathbf{r}_{\\mathrm{sec}}(t')\\) with the scene geometry. The radiance from indirect illumination is also computed for each lighting condition separately, then combined with direct illumination to yield the complete PBR radiance for the camera ray under a specific lighting condition:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.801, + 0.469, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r , n e a r} \\}} = \\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d} \\{\\text {f a r , n e a r} \\}} + \\mathbf {C} _ {\\mathrm {p b}, \\text {i n d} \\{\\text {f a r , n e a r} \\}}. \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.182, + 0.849 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Training Schemes. Our training process is divided into two sequential stages. In the first stage, the focus is on training the geometry and radiance field without employ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "ing PBR. The objectives are: 1) to recover the scene's geometry, 2) to distinguish the appearance under each lighting condition, a necessity due to the potential presence of multiple light sources in one image, and 3) to utilize the reconstructed radiance field for indirect illumination. In the second stage, we train the material field and scene lighting using PBR, aiming to estimate material properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Loss Functions. We average the loss functions across batches of camera rays. During the first stage, our loss function compares the total re-rendered radiance under all lighting conditions present in an image – including a far-field light \\(i\\) and any active near-field lights – with the ground-truth (GT) color \\(\\mathbf{C}_{\\mathrm{gt}}\\) from the input observation. This comparison trains the geometry and radiance field:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.313, + 0.892, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r f}} = \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\text {f a r} _ {i}} + \\sum_ {\\text {n e a r} i \\text {i s o n}} \\mathbf {C} _ {\\mathrm {r f}, \\text {n e a r} _ {i}} - \\mathbf {C} _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.379 + ], + "angle": 0, + "content": "The Eikonal loss [8] is applied to the gradients of the SDF values \\( s \\) for geometric regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.384, + 0.891, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e i k}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\left\\| \\frac {\\partial s (\\mathbf {r} (t _ {j}))}{\\partial \\mathbf {r} (t _ {j})} \\right\\| _ {2} - 1\\right) ^ {2}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.462 + ], + "angle": 0, + "content": "A silhouette loss using GT alpha \\(\\alpha_{\\mathrm{gt}}\\) from the foreground mask and a normal smoothness loss [41] aid shape recovery:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.468, + 0.891, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s i l}} = \\left(\\left(1 - T _ {N + 1}\\right) - \\alpha_ {\\mathrm {g t}}\\right) ^ {2}. \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.489, + 0.891, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n s}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right)\\right) - \\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right) + \\boldsymbol {\\epsilon}\\right)\\right) ^ {2}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.674 + ], + "angle": 0, + "content": "where \\(\\epsilon \\in \\mathbb{R}^3\\) is a small random perturbation. The total loss for stage one is \\(\\mathcal{L}_{\\mathrm{rf}} + \\lambda_{\\mathrm{eik}}\\mathcal{L}_{\\mathrm{eik}} + \\lambda_{\\mathrm{sil}}\\mathcal{L}_{\\mathrm{sil}} + \\lambda_{\\mathrm{ns}}\\mathcal{L}_{\\mathrm{ns}}\\), with \\(\\lambda_*\\) representing the loss weights. In stage two, PBR colors replace radiance field-rendered colors for comparison with \\(\\mathbf{C}_{\\mathrm{gt}}\\), and a novel self-consistency loss is introduced between neural and PBR radiance under each lighting for additional supervision. The direct supervision loss \\(\\mathcal{L}_{\\mathrm{pb}}\\) mirrors \\(\\mathcal{L}_{\\mathrm{rf}}\\) with \\(\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{far,near}\\}_i}\\) replaced by \\(\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{far,near}\\}_i}\\). The perlighting self-consistency loss \\(\\mathcal{L}_{\\mathrm{con}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.674, + 0.892, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n}} = \\sum_ {i = 1} ^ {N _ {\\{\\text {f a r}, \\text {n e a r} \\}}} \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} - \\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.796 + ], + "angle": 0, + "content": "This loss is crucial for discerning contributions from each lighting in images with multiple active light sources, thereby reducing ambiguity and improving material estimation. A material smoothness loss \\(\\mathcal{L}_{\\mathrm{ms}}\\) is used similar to \\(\\mathcal{L}_{\\mathrm{ns}}\\). The total loss for stage two is \\(\\mathcal{L}_{\\mathrm{pb}} + \\lambda_{\\mathrm{con}}\\mathcal{L}_{\\mathrm{con}} + \\lambda_{\\mathrm{ms}}\\mathcal{L}_{\\mathrm{ms}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our model is trained on an NVIDIA RTX 3090 GPU for a total of 40,000 steps. The first 30,000 steps of stage one take about 20 minutes, while the subsequent 10,000 steps of stage two require about 40 minutes. Post-training, we follow a procedure similar to WildLight [6] to extract the fields into textured meshes, which can be easily integrated into industry-standard rendering software like Blender [1]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11805" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.091, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison results with state-of-the-art methods averaged on 6 synthetic scenes. We show results of surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, the specular reflection part of FV relit RGB, and training time on a single RTX 3090 GPU. We mark the best and the second best results in each column. \\(\\uparrow (\\downarrow)\\) means bigger (smaller) is better." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.135, + 0.89, + 0.309 + ], + "angle": 0, + "content": "
MethodInput lighting conditionsNormal MAngE↓AlbedoView synthesisFV relitFV relit (spec)Time
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
(1) TensoIR [10]17.6626.480.92129.480.91228.180.90128.300.861300 mins
(2) NVDiffRecMC [9]Single far-field16.2426.520.91527.130.91326.610.90125.570.836150 mins
(3) Ours12.3924.500.88228.200.93427.460.92127.560.87145 mins
(4) WildLight [6]Single far-field + Flashlight11.4928.860.94029.870.92930.440.93027.710.8631440 mins
(5) Ours10.8931.620.95332.090.95332.000.95330.750.90660 mins
(6) TensoIR [10]Two far-field16.2427.180.92929.680.91228.660.90228.460.863300 mins
(7) Ours11.7026.070.90229.590.94229.270.93429.090.89045 mins
(8) OursTwo far + Single near10.7932.040.95732.100.95032.380.95431.400.91060 mins
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.317, + 0.471, + 0.348 + ], + "angle": 0, + "content": "enabling fast and high-quality rendering suitable for various applications. Check the supplement for more details." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.36, + 0.21, + 0.377 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Our experiments involve a comprehensive comparison with state-of-the-art methods (Sec. 4.2), ablation studies (Sec. 4.3), and evaluations on both synthetic datasets (Sec. 4.1) and real-world images (Sec. 4.4). The supplement shows further data creation details and more experiments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.469, + 0.255, + 0.485 + ], + "angle": 0, + "content": "4.1. Synthetic Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.765 + ], + "angle": 0, + "content": "We collect 6 synthetic scenes comprising a variety of scene geometries and materials. For training, each scene is rendered under four distinct lighting setups: 1) single far-field light, 2) single far-field light with a camera-collocated flashlight, 3) two far-field lights, and 4) two far-field lights with a near-field light source. We render 100 training images per setting using random viewpoints and per-image lighting conditions. The far-field lighting from the first setting is used in subsequent settings, with the last two sharing their far-field lighting. For testing, 200 images under one common far-field lighting are rendered to assess novel view synthesis under seen lighting, and another 200 under unseen far-field lighting to evaluate free-viewpoint (FV) relighting, which largely depends on the geometry and material estimation quality. This methodology provides a relatively fair comparison across methods with different abilities regarding supported input lighting conditions, trained under similar conditions and tested all using the same sets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.442, + 0.789 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We compare our method with leading multi-view inverse rendering methods under each input lighting scenario. For single far-field lighting inputs (Tab. 1 row 1), we compare with NVDiffRecMC [9] and TensoIR [10] (single-lighting input). For multiple far-field lighting scenarios (Tab. 1 row 3), we compare with TensoIR [10]. For setups involving a far-field light and a flashlight (Tab. 1 row 4), we compare" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.408 + ], + "angle": 0, + "content": "with WildLight [6]. We use Blender [1] for high-quality relit results for our method, WildLight [6], and NVDiffRecMC [9], importing trained SDF and material fields as textured meshes. Since TensoIR [10] employs density-based geometry and cannot extract high-quality meshes, we utilize its differentiable renderer for relighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.573 + ], + "angle": 0, + "content": "Our quantitative comparison employs mean angular error (MAngE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) [30] metrics for surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, and specular reflection of FV relit RGB. We follow NeRFactor [41] in assuming albedo and lighting brightness as scale-invariant, necessitating metric comparisons using corresponding scales. For visualization and metrics, we scale each RGB channel of albedo and relit images from all methods by a global scale to minimize mean squared error against ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.894, + 0.799 + ], + "angle": 0, + "content": "The quantitative results are detailed in Tab. 2. We observe that when only input images under single far-field lighting are available (rows 1-3), our method gives comparable results as TensoIR [10] and NVDiffRecMC [9]. Adding a flashlight (rows 4, 5) significantly helps in geometry and material estimation (comparing rows 3, 5), where the greatest improvement is made regarding the diffuse albedo. Also, our method clearly surpasses WildLight [6] in both quality and speed in this setting. Leveraging another far-field lighting condition (rows 6, 7) brings similar effects as using a flashlight, but generally with a lower degree of improvement. Further adding a flashlight (row 8) gives the best results among all settings and methods, but the gain is marginal compared to the setting of single far-field lighting with a flashlight (comparing rows 5, 8)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.894, + 0.892 + ], + "angle": 0, + "content": "Fig. 3 shows the qualitative comparison on two synthetic scenes: LEGO (left) with detailed geometry, and TROOPER (right) with highly-reflective SV materials. The results show that our method predicts more detailed and faithful shape, diffuse albedo, and specular parameters that could produce accurate and realistic relit results under unseen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11806" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.495, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.887, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Figure 3. Comparison with state-of-the-art methods. We test NVDiffRecMC [9] (single far-field lighting in input), TensoIR [10] (two far-field lighting in input), and WildLight [6] and our method (single far-field lighting with a flashlight in input) on two synthetic scenes. The intensity of the specular reflection is multiplied by \\(3 \\times\\) for the visualization purpose. Please zoom in for details." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.4, + 0.47, + 0.428 + ], + "angle": 0, + "content": "Table 3. Ablation study on our method using inputs under single far-field lighting with a flashlight." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.431, + 0.465, + 0.53 + ], + "angle": 0, + "content": "
ModelNormalAlbedoFV RelitRelit (spec)
MAngE↓PSNR↑PSNR↑PSNR↑
(1) w/o Lcon8.45227.7029.8529.08
(2) Mod. cnear8.57227.7930.0328.87
(3) w/o encmat8.26327.7930.0728.92
(4) Full model8.29828.0530.1929.36
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.594 + ], + "angle": 0, + "content": "lighting. In contrast, WildLight [6] tends to overly smooth geometry and material. TensoIR [10] can not handle highly-reflective surfaces, and lighting is baked into its predicted diffuse albedo for TROOPER." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.241, + 0.616 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.745 + ], + "angle": 0, + "content": "In Tab. 3, we quantitatively evaluate ablation models on a subset of our synthetic scenes: 1) the model excluding the self-consistency loss \\(\\mathcal{L}_{\\mathrm{con}}\\), 2) the model where the output of the near-field radiance MLP is not explicitly multiplied by the cosine weakening factor and incident radiance (Eq. (8)), 3) the model with \\(\\mathrm{enc}_{\\mathrm{mat}}(\\mathbf{x})\\) replaced by \\(\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x})\\). The results show that all the above techniques enhance the geometry and materials estimated by our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.262, + 0.767 + ], + "angle": 0, + "content": "4.4. Real-world Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Fig. 4 shows the results of our method on two real-world scenes captured under single far-field lighting and a flashlight. Our method faithfully recovers the geometry and material of the objects and produces realistic relit results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We introduce a versatile multi-view inverse rendering framework, distinguished by its ability to leverage input im" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.399, + 0.891, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.516, + 0.891, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Figure 4. The results of our method on two real-world scenes, using multi-view images under one ambient lighting and a flashlight. The specular reflection intensity is increased for visualization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.715 + ], + "angle": 0, + "content": "ages under varied far- and near-field light sources available in capture for better geometry and material estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Limitations. VMINer does not model the background, necessitating a foreground mask for each image. It also does not consider unknown tone-mapping curves applied during the image signal processing (ISP) stage, using a fixed curve with \\(\\gamma = 2.2\\) to the computed linear radiance. It depends on distinguishing the contribution from each lighting, which implies that, for optimal results, near-field light sources should be switched on/off during capture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.852, + 0.66, + 0.869 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grand No. 62136001, 62088102)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11807" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Blender Foundation. The Blender project - free and open 3D creation software. https://www.blender.org. Accessed: 2023-11-07. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. NeuralPIL: Neural pre-integrated lighting for reflectance decomposition. In Proc. of Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[4] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P. A. Lensch, and Varun Jampani. SAMURAI: shape and material from unconstrained real-world arbitrary image collections. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[5] Brent Burley. Physically-based shading at disney. In SIGGRAPH 2012 Courses, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] Ziang Cheng, Junxuan Li, and Hongdong Li. WildLight: In-the-wild inverse rendering with a flashlight. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proc. of International Conference on Machine Learning (ICML), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[9] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using Monte Carlo rendering and denoising. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[11] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[12] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG), 41(4):56:1-56:12, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Eric P Lafortune and Yves D Willems. Using the modified phong reflectance model for physically based rendering. Katholieke Universiteit Leuven. Departement Computerwetenschappen, 1994. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates NeRFs. arXiv preprint arXiv:2305.04966, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[15] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable Monte Carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6): 222, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[16] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Milos Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In Proc. of European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Jacob Munkberg, Wenzheng Chen, Jon Hasselgren, Alex Evans, Tianchang Shen, Thomas Müller, Jun Gao, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[22] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Viktor Rudnev, Mohamed Elgharib, William A. P. Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. ReLight My NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11808" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[26] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[27] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. In Proc. of the Eurographics Symposium on Rendering Techniques, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[28] Jiaping Wang, Peiran Ren, Minmin Gong, John M. Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM Transactions on Graphics (TOG), 28(5):133, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proc. of Neural Information Processing Systems (NeurIPS), 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[31] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[32] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[34] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[36] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(8):10129-10142, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[37] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Inter-reflectable light fields for geometry and material estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[40] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: inverse rendering by optimizing neural SDFs and materials from photometric images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[41] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[43] Youjia Zhang, Teng Xu, Junqing Yu, Yuteng Ye, Yanqing Jing, Junle Wang, Jingyi Yu, and Wei Yang. NeMF: Inverse volume rendering with neural microflake field. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[44] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG), 41(6):235:1-235:17, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[45] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable Monte Carlo raytracing. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[46] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Di-anbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I\\(^2\\)-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[47] Yongjie Zhu, Yinda Zhang, Si Li, and Boxin Shi. Spatially-varying outdoor lighting estimation from intrinsics. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11809" + } + ] +] \ No newline at end of file diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_origin.pdf b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71b4683665977c6e9103cfab1e429c9620c5e5fe --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/e2a0f876-3b69-4ca2-9f34-bbb92417c831_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a82a6b0f7349c0831f30656cdace39042df0492080cb128a362f9ddd16348c51 +size 9031617 diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/full.md b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd7b1fa2214a00cd914ecce27dd0c06c75e53483 --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/full.md @@ -0,0 +1,348 @@ +# VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources + +Fan Fei $^{1,2,4}$ + +Jiajun Tang $^{1,2,4}$ + +Ping Tan3,4 + +Boxin Shi $^{1,2\#}$ + +$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{3}$ Hong Kong University of Science and Technology $^{4}$ Light Illusions + +# Abstract + +This paper introduces a versatile multi-view inverse rendering framework with near- and far-field light sources. Tackling the fundamental challenge of inherent ambiguity in inverse rendering, our framework adopts a lightweight yet inclusive lighting model for different near- and far-field lights, thus is able to make use of input images under varied lighting conditions available during capture. It leverages observations under each lighting to disentangle the intrinsic geometry and material from the external lighting, using both neural radiance field rendering and physically-based surface rendering on the 3D implicit fields. After training, the reconstructed scene is extracted to a textured triangle mesh for seamless integration into industrial rendering software for various applications. Quantitatively and qualitatively tested on synthetic and real-world scenes, our method shows superiority to state-of-the-art multi-view inverse rendering methods in both speed and quality. + +# 1. Introduction + +The reconstruction of 3D scenes from multi-view RGB imagery has experienced significant advancements following the development of Neural Radiance Fields (NeRF) [19], and both the speed and quality of the reconstruction have reached an unprecedented level [18, 20]. Despite these improvements, a common limitation is to represent only the radiance field, which is a complex product of the external lighting interacting with the intrinsic geometry and material of the scene [11]. The entangled nature of this representation generally hampers the ability to accurately render the scene under unseen lighting conditions, because the influence of the original lighting is embedded within the newly rendered scene. Recent approaches [24, 41, 46] have incorporated inverse rendering [22] to separate material properties and lighting effects, extending the application of the reconstruction beyond novel view synthesis to novel scenarios involving relighting and material editing [2]. Some + +![](images/666655ee759469bc512e25e6e361d78e0c88657a3c0f2c78b5e8fb4eba83864d.jpg) + +![](images/f96c9997a11d572c792b71a5aa2263c68a0ac967236f2c8372f60973b522a1b9.jpg) + +![](images/4ed3a4b2f4591e97d11b8176d393cd24d837d0542efced48509631783bdb842e.jpg) +Figure 1. Given multi-view images under possibly varied lighting, our method leverages all present lighting conditions to reconstruct scene geometry and material disentangled with lighting. The trained fields are extracted to textured meshes for seamless integration into industrial renderers for various applications. + +![](images/cb050fcc3eab0ef1ec35bac7fe3512835a41341a3386d274f77df58c883014ea.jpg) + +methods [2, 6, 9, 21, 32] have advanced to the extraction of detailed triangle meshes with material UV textures, which can serve as economical and lifelike 3D models for gaming and cinematography industries [9, 21], thereby marking a transformative step in digital asset creation. + +Inverse rendering presents several fundamental challenges, one of them being its severe inherent ambiguity [41]. To combat this, multi-view inverse rendering methodologies usually reduce the ambiguity by imposing various constraints on each scene component. These methods can be categorized based on the assumption on the amount and types of lighting conditions present in the input images (Tab. 1). The majority of methods [9, 21, 24, 35, 36, 39, 41, 42] assumes the imagery to be captured under one fixed lighting condition (rows 1, 2). However, in such scenarios, only one case of scene appearances among all possible cases under different lighting is observed, causing severe lighting-material ambiguity, posing considerable difficulties to material estimation. In a bid to alleviate this ambiguity, recent techniques [2, 6, 10, 12, 23, 32] utilize input + +Table 1. A summary of multi-view inverse rendering methods based on their assumptions and usages of input lighting conditions. + +
Methods# of far-field light# of near-field lightMaterial optimizationLighting-material ambiguityCapture workload
(1) [24, 39, 41, 42]SingleNoneOnly from singleStrongLow
(2) NeILF [35]Single 5D incident light field
(3) [2, 10, 12, 32]aMultipleNoneOnly from far-fieldModerateHigh
(4) WildLight [6]SingleSingleOnly from near-fieldModerateLow
(5) OursSingleSingleFrom near- and far-fieldWeakLow
MultipleSingle or multipleAlmost noneHigh
+ +${}^{a}$ These methods also support input images under single far-field lighting. In that case, they belong to single far-field lighting methods (row 1). + +images under varied lighting to discern materials from lighting (rows 3, 4). This demands more complex lighting models. Existing varied-lighting methods [2, 10, 12, 23, 32] typically model multiple, but only far-field, lighting conditions to maintain manageable unknown lighting parameters. This means throughout the capture, the lighting setups should be largely changed by either moving the objects to reconstruct or adapting far-field light sources, causing increased workload to the data collection process. Alternatively, near-field light sources, such as flashlights, can function as easily controllable light sources for disambiguation. WildLight [6] suggests the use of a camera-collocated flashlight in addition to the ambient lighting for inverse rendering. Nevertheless, it limits itself by using only the isolated appearance under the flashlight for material estimation, not harnessing available observations under the ambient lighting and degrading the quality of its estimated material (row 4). + +Addressing the challenges and restrictions previously outlined, this paper introduces VMINer, a Versatile Multiview Inverse rendering framework with Near- and far-field light sources (row 5). The distinctive feature of our framework is that it uses a lightweight yet inclusive lighting model for different far- and near-field light sources and efficiently leverages available observations under each lighting to disentangle lighting from other components, making the most of whatever lighting conditions are at hand to boost the practicality and quality of the reconstruction, as shown in Fig. 1. Although our method accepts the simplest setting of inputs under single far-field lighting, a more effective compromise is to employ an additional flashlight to enhance quality without overly complicating the capture process. The ideal scenario for the highest quality reconstruction would involve capturing the scene under varied far-field lighting and using a flashlight. + +Nonetheless, implementing such a versatile framework is far from straightforward. The method has to render the scene under different lighting conditions, including spatially-varying and changing ones to represent possibly moving near-field lights. We propose to solve the problem by modeling near-field lights as point sources in 3D space with adjustable positions and intensities, which could be anisotropic and not necessarily camera-aligned (e.g., + +stationary desk lamps). This representation effectively approximates a broad range of common near-field lights and is easy for spatially-consistent editing and optimization. We use 3D implicit fields to model the scenes' radiance, shape, and material and minimize the error of the neural or physically-based re-rendered scene appearance under each input lighting. To account for multiple lighting situations and moving near-field lights, we integrate specially tailored multi-layer perceptrons (MLPs) that additionally process light directions and embeddings. Post-training, the fields are converted into a textured triangle mesh, ready for seamless integration into industrial rendering software like Blender [1], facilitating a broad spectrum of applications. Our extensive experiments, covering both synthetic and real-world scenes, demonstrate that VMINer surpasses prior methods quantitatively and qualitatively. + +# 2. Related Work + +Preliminary. The rendering equation [11] models the reflected radiance from any point on a surface as a result of an intricate surface integral. This integral comprises contributions from three fundamental scene components: lighting, material, and geometry. The equation is expressed as: + +$$ +L _ {\mathrm {o}} (\mathbf {x}, \boldsymbol {\omega} _ {\mathrm {o}}) = \int_ {\Omega} \underbrace {L _ {\mathrm {i}} (\mathbf {x} , \boldsymbol {\omega} _ {\mathrm {i}})} _ {\text {l i g h t i n g}} \underbrace {f (\mathbf {x} , \boldsymbol {\omega} _ {\mathrm {i}} , \boldsymbol {\omega} _ {\mathrm {o}})} _ {\text {m a t e r i a l}} \underbrace {(\boldsymbol {\omega} _ {\mathrm {i}} \cdot \mathbf {n}) ^ {+}} _ {\text {g e o m e t r y}} \mathrm {d} \boldsymbol {\omega} _ {\mathrm {i}}. \tag {1} +$$ + +In this equation, $L_{\mathrm{o}}(\mathbf{x}, \omega_{\mathrm{o}})$ represents the radiance reflected in the outgoing direction $\omega_{\mathrm{o}}$ from a surface point $\mathbf{x}$ in 3D space. $\Omega$ denotes the unit hemisphere encompassing all incident directions $\omega_{\mathrm{i}}$ with $\omega_{\mathrm{i}} \cdot \mathbf{n} > 0$ . $L_{\mathrm{i}}(\mathbf{x}, \omega_{\mathrm{i}})$ is the incident radiance at $\mathbf{x}$ from direction $\omega_{\mathrm{i}}$ , and $f(\mathbf{x}, \omega_{\mathrm{i}}, \omega_{\mathrm{o}})$ is the bidirectional reflectance distribution function (BRDF), denoting the ratio of light reflected along $\omega_{\mathrm{o}}$ at $\mathbf{x}$ from $\omega_{\mathrm{i}}$ . The term $(\omega_{\mathrm{i}} \cdot \mathbf{n})^{+}$ , where $x^{+} \triangleq \max(x, 0)$ , signifies the cosine weakening factor. The incident lighting $L_{\mathrm{i}}(\mathbf{x}, \omega_{\mathrm{i}})$ may originate from other points within the scene, necessitating recursive evaluation of this integral. + +Multi-view Inverse Rendering. Multi-view inverse rendering techniques [2-4, 6, 9, 10, 12, 21, 23, 24, 34-36, 38-43] generally employ implicit or explicit 3D fields to model the scene's geometry and material, while also estimating + +Figure 2. Bottom-left: VMINer takes as input multi-view RGB images with foreground masks under possibly varied far- and near-field lighting. The total number of lighting conditions and their types should be given. Each image is illuminated by one far-field lighting whose index is "far index", and near-field lights whose on/off states are "near on/off". Top-left: VMINer models each far-field lighting $(\mathrm{far}_i)$ and near-field lighting $(\mathrm{near}_i)$ separately as parametric models, from which incoming radiance is queried at any position $\mathbf{x}$ and direction $\omega$ . Bottom-mid: It utilizes 3D implicit representations for scene geometry, material, and radiance. Four MLPs process $\mathbf{x}$ and view direction $\mathbf{v}$ encoded by different encoders $\mathrm{enc}_{\{\mathrm{geo},\mathrm{mat},\mathrm{dir}\}}$ with light embeddings $\mathbf{F}_{\{\mathrm{far},\mathrm{near}\}_i}$ to get the SDF $s$ , the neural appearance descriptor $\mathbf{z}$ , the BRDF parameters $\beta$ , and the radiance $\mathbf{c}_{\{\mathrm{far},\mathrm{near}\}_i}$ under each lighting. Bottom-right: It then uses radiance field rendering to re-render the appearance $\mathbf{C}_{\mathrm{rf},\{\mathrm{near},\mathrm{far}\}_i}$ separately under each lighting condition, added up to $\mathbf{C}_{\mathrm{rf}}$ according to the per-image lighting condition. Top-right: Physically-based surface rendering uses Monte Carlo integration to evaluate both direct illumination $L_{\mathrm{d}\{\mathrm{near},\mathrm{far}\}_i}$ with secondary visibility and indirect illumination $L_{\mathrm{ind}\{\mathrm{near},\mathrm{far}\}_i}$ . The rendered appearances are aggregated to $\mathbf{C}_{\mathrm{pb}}$ . $\mathbf{C}_{\{\mathrm{rf},\mathrm{pb}\}}$ are compared with the observation $\mathbf{C}_{\mathrm{gt}}$ to train the scene model. $\mathbf{C}_{\mathrm{rf},\{\mathrm{near},\mathrm{far}\}_i}$ are also used as additional supervision signals for $\mathbf{C}_{\mathrm{pb},\{\mathrm{near},\mathrm{far}\}_i}$ . +![](images/0ca25f6f05cfdb6c0ded64883c3a2eba6824f20bd9a754e282e7ed753d092aa1.jpg) +: input : variable : refer to another entry : untrainable operator : trainable operator / MLP $\bigotimes \bigoplus$ element-wise multiplication / add + +lighting to recreate the scene's appearance. They optimize each scene component from scratch by aligning the rendered with the observed appearance, while avoiding potential local minima, which, although do not faithfully represent the scene, can replicate the observed input. Addressing this inherently ambiguous nature of inverse rendering, these methods typically impose constraints on each scene component. For geometry and material, assumptions of known geometry [35, 36, 46] are made, along with leveraging the recovered shape from radiance fields [6, 10, 41], adopting Lambertian or uniform materials [23, 39], controlling BRDF parameter smoothness [10], or integrating a learned BRDF latent space [41]. Regarding lighting, many approaches presuppose known lighting conditions [24, 40] or fixed lighting shared across input images [35, 38, 41, 42]. More recent techniques [2, 6, 10, 12, 23, 32] have turned to using images with varied lighting conditions to mitigate the lighting-material ambiguity. They typically accept input images that can be grouped into sets, each illuminated by one distinct and fixed far-field lighting. WildLight [6] exceptionally employs a camera-collocated flashlight along with fixed ambient lighting. However, it does not model the ambient lighting and thus its material property estimation relies exclusively on flashlight observations, where the light and view direction always coincide [5], limiting its BRDF estimation ability. Our VMINer, in contrast, accommodates + +input images under single or multiple far-field lighting conditions with near-field light sources and leverages each of them to enhance the quality of the estimated material. + +Lighting Models. Spatially-uniform (SU) lighting models rely on the assumption that the scene's lighting originates from a distant source, modeled as spherical Gaussians (SGs) [2, 10], spherical harmonics (SH) [12, 23], or MLPs [32]. In contrast, spatially-varying (SV) lighting models acknowledge the presence of near-field light sources, causing different locations to receive different incident lighting. They include parametric 3D lights [7, 17, 36, 37], outgoing light fields such as volumetric SGs [31] and neural out-of-view lighting volumes [45], and incident light fields like SV environment maps [47], SVSGs [16], and neural incident light fields (NeILF) [35, 38]. Our VMINer integrates distant lights as SGs and near-field lights as point lights with adjustable position and intensity. It is less complex compared to other SV lighting models, facilitating spatially consistent editing and optimization, while being able to approximate common lighting setups in real life. + +# 3. Proposed Method + +As illustrated in Fig. 2, VMINer reconstructs the scene lighting (Sec. 3.1) and 3D fields (Sec. 3.2) from input images using differentiable rendering (Sec. 3.3). The training scheme and loss functions are described in Sec. 3.4. + +# 3.1. Versatile Lighting Model + +This subsection models the lighting term in Eq. (1) and corresponds to the top-left part of Fig. 2. + +VMINer harnesses the diversity of lighting conditions present in the input RGB images to disentangle lighting effects. The lighting model plays a crucial role here: 1) Since lighting conditions are unknown initially, they must be optimized alongside other scene parameters using physically-based rendering during training. 2) Different lighting setups, each compatible with the model, are necessary during the capture phase to achieve this goal. Given these prerequisites, it is essential to design a lighting model that accommodates common far-field and near-field light sources, while being lightweight enough for efficient and robust optimization and physically-based rendering. With these considerations, we propose a versatile lighting model that is both inclusive and lightweight. This representation models far-field and near-field lighting separately as below. + +Far-field Lighting. In our model, each far-field lighting condition $i$ , where $i$ belongs to the set $\mathbb{N}_{N_{\mathrm{far}}} \triangleq \{1, \dots, N_{\mathrm{far}}\}$ and $N_{\mathrm{far}}$ denotes the number of far-field lighting, is represented using 128-lobe SGs [28] commonly adopted in existing far-field lighting techniques [33, 39, 42]. For each lobe $j$ of lighting $i$ , there are six parameters: lobe axis $\pmb{\xi}_{ij} \in \mathbb{S}^2$ , lobe sharpness $\lambda_{ij} \in \mathbb{R}_+$ , and lobe RGB amplitude $\pmb{\mu}_{ij} \in \mathbb{R}_+^3$ . The incident radiance from far-field lighting $i$ along direction $\omega$ at any 3D position $\mathbf{x}$ is calculated using the formula (ignoring visibility for now): + +$$ +L _ {\operatorname {f a r} _ {i}} (\boldsymbol {\omega}) = \sum_ {j = 1} ^ {1 2 8} c _ {i j} \boldsymbol {\mu} _ {i j} e ^ {\lambda_ {i} (\boldsymbol {\omega} \cdot \boldsymbol {\xi} _ {i j} - 1)}, \tag {2} +$$ + +where $c_{ij} = \lambda_{ij} / (2\pi (1 - e^{-2\lambda_{ij}}))$ acts as the normalization factor related to roughness. Notably, $\mathbf{x}$ does not appear as an input to $L_{\mathrm{far}_i}$ , reflecting the nature of distant lighting. + +Near-field Lighting. Each near-field lighting $i$ , included in the set $\mathbb{N}_{N_{\mathrm{near}}}$ with $N_{\mathrm{near}}$ representing the number of near-field light sources, is modeled as a point light. These point lights can have moving positions $\mathbf{p}_i \in \mathbb{R}^3$ and exhibit anisotropic radiation characterized by $l^{\mathrm{th}}$ -order ( $l$ can be 0, 1, or 2) SH coefficients $\mathbf{h}_i \in \mathbb{R}^{3 \times (l + 1)^2}$ . Our method accommodates two types of near-field lighting: camera-collocated lights, positioned at the camera ray origin for each image, and stationary lights, which remain fixed across all images. We observe that in neural radiance field rendering, the radiance under a stationary near-field light, especially when the light is active in all images, can be challenging to distinguish from radiance under far-field lighting. Also, the radiance from these two sources aids in material estimation in a similar way. Therefore, in practical applications, we favor using a moving flashlight as the near-field light to provide unique information for material estimation. The incident radiance from near-field lighting $i$ at a 3D po + +sition $\mathbf{x}$ is computed as (also ignoring visibility for now): + +$$ +L _ {\text {n e a r} i} (\mathbf {x}, \boldsymbol {\omega}, \mathbf {p} _ {i}) = \left\{ \begin{array}{l l} \frac {\mathrm {S H} (\boldsymbol {\omega} ; \mathbf {h} _ {i})}{\| \mathbf {p} _ {i} - \mathbf {x} \| _ {2} ^ {2}} & \text {i f} \boldsymbol {\omega} = \frac {\mathbf {p} _ {i} - \mathbf {x}}{\| \mathbf {p} _ {i} - \mathbf {x} \| _ {2}}, \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {3} +$$ + +where $\mathrm{SH}(\omega ;\mathbf{h})$ calculates the SH at direction $\omega$ with coefficients $\mathbf{h}$ , and $1 / \| \mathbf{p}_i - \mathbf{x}\|_2^2$ signifies the inverse-square lighting attenuation for point lights. Here we include $\mathbf{p}_i$ as an input because it may be set differently across images. It is important to note that with respect to the incident direction $\omega$ , $L_{\mathrm{far}}$ is a continuous function, while $L_{\mathrm{near}}$ is a discrete function, being non-zero only in a single direction. As a result, to render appearances under direct lighting, Monte Carlo integration is essential for far-field lighting, while a simple multiplication suffices for near-field lighting. Further details about this process are provided in Sec. 3.3. + +# 3.2. Geometry, Material, and Radiance Fields + +This subsection models the material and geometry terms in Eq. (1) and corresponds to the bottom-mid part of Fig. 2. + +Geometry. Multi-view reconstruction methodologies generally hinge on two geometry representations: volume density [2, 10, 19, 41] and the signed distance function (SDF) [33, 39, 42, 44] fields. We choose the SDF field for our geometry representation due to its clearly defined surface at the zero-level isosurface, which simplifies and enhances the post-training mesh extraction process. Our approach utilizes implicit fields as MLPs over explicit structures like voxel grids for their compactness and flexibility. For each 3D position $\mathbf{x}$ , a multi-resolution hash grid [20] is first employed for positional encoding, yielding a feature vector $\mathrm{enc}_{\mathrm{geo}}(\mathbf{x}) \in \mathbb{R}^{16}$ . The geometry MLP $\mathrm{M}_{\mathrm{geo}}$ then predicts the SDF value at $\mathbf{x}$ : + +$$ +s (\mathbf {x}), \mathbf {z} (\mathbf {x}) = \mathrm {M} _ {\mathrm {g e o}} (\mathbf {x}, \operatorname {e n c} _ {\mathrm {g e o}} (\mathbf {x})), \tag {4} +$$ + +where $s \in \mathbb{R}$ represents the signed distance (positive outside the surface, negative inside), and $\mathbf{z} \in \mathbb{R}^{13}$ is a descriptor of local appearance. To maintain differentiability in rendering and to aid in geometry optimization, during training we apply NeuS [29] techniques to transform SDF values along rays into volume densities. The surface normal $\mathbf{n} \in \mathbb{S}^2$ is derived as the gradient of the SDF $s$ with respect to $\mathbf{x}$ : + +$$ +\mathbf {n} (\mathbf {x}) = \frac {\partial s}{\partial \mathbf {x}} / \left\| \frac {\partial s}{\partial \mathbf {x}} \right\| _ {2}. \tag {5} +$$ + +$\mathrm{M}_{\mathrm{geo}}$ only approximates a strict SDF field that has $\left\| \frac{\partial s}{\partial \mathbf{x}} \right\|_2 = 1$ , so we normalize $\mathbf{n}$ to ensure it is a unit vector. + +Material. For accurate scene material recovery using physically-based re-rendering, it is imperative to model the material in such a way that the renderer can query the BRDF at any point on the surface. Considering that material properties are independent of lighting and view directions, we represent the scene's material through an implicit 3D field, much like the SDF field utilized for geometry. To achieve + +this, we employ another multi-resolution hash grid $\mathrm{enc}_{\mathrm{mat}}$ for positional encoding. We adopt the simplified GGX BRDF model [27] with a fixed fresnel parameter to help alleviate ambiguity. The material MLP, $\mathbf{M}_{\mathrm{mat}}$ , is tasked with predicting the SVBRDF parameters $\beta (\mathbf{x})$ : + +$$ +\boldsymbol {\beta} (\mathbf {x}) = \operatorname {M} _ {\text {m a t}} (\mathbf {x}, \mathbf {z}, \mathbf {n}, \operatorname {e n c} _ {\text {m a t}} (\mathbf {x})). \tag {6} +$$ + +Radiance. Although our method does not strictly require a radiance field for rendering scene appearance - given that appearance could be rendered solely using physically-based surface rendering (PBR) - we find that incorporating a radiance field at the start of training significantly eases the optimization of scene geometry and appearance. PBR, while physically accurate, tends to introduce instability and slow down the training process due to its inherent ambiguity and computational intensity. Moreover, the radiance field-rendered results can serve as supplementary supervision, aiding PBR in better separating scene appearance under each lighting condition and thus further diminishing lighting-material ambiguity. Hence, we deploy radiance MLPs to predict view-dependent and lighting-dependent radiance at each position. For each lighting condition $i$ , a light embedding $\mathbf{F}_{\{\mathrm{far,near}\}_i} \in \mathbb{R}^{16}$ is utilized. The view direction $\mathbf{v} \in \mathbb{S}^2$ is processed through directional encoding that projects it onto the coefficients of the $3^{\mathrm{rd}}$ -order SH basis, yielding $\mathrm{enc}_{\mathrm{dir}}(\mathbf{v}) \in \mathbb{R}^{4^2 = 16}$ [20]. The far-field radiance MLP $\mathrm{M}_{\mathrm{far}}$ computes the outgoing radiance $\mathbf{c}_{\mathrm{far}_i}$ from position $\mathbf{x}$ along view direction $\mathbf{v}$ under far-field lighting $i$ : + +$$ +\mathbf {c} _ {\text {f a r} _ {i}} (\mathbf {x}, \mathbf {v}) = \mathrm {M} _ {\text {f a r}} (\mathbf {x}, \mathbf {z}, \mathbf {n}, \operatorname {e n c} _ {\text {d i r}} (\mathbf {v}), \mathbf {F} _ {\text {f a r} _ {i}}), \tag {7} +$$ + +where we incorporate the surface normal $\mathbf{n}$ as an additional MLP input, following Instant-NSR [44] and WildLight [6], as this has shown to aid shape recovery. + +The near-field radiance MLP differs slightly, as nearfield lighting can move in different images, thus the current light position $\mathbf{p}_i$ is a necessary input. Additionally, the radiance under near-field point light takes the form of the rendering equation Eq. (1) without the integral, allowing the incoming radiance and cosine term to be explicitly included. We follow ReNe [25] to input the relative direction of the point light $\omega_{i} = \frac{\mathbf{p}_{i} - \mathbf{x}}{\|\mathbf{p}_{i} - \mathbf{x}\|_{2}}$ instead of the absolute position $\mathbf{p}_i$ into the MLP. Therefore, the outgoing neural radiance $\mathbf{c}_{\mathrm{near}_i}$ under near-field lighting $i$ is calculated as: + +$$ +\begin{array}{l} \mathbf {c} _ {\mathrm {n e a r} _ {i}} (\mathbf {x}, \mathbf {v}, \mathbf {p} _ {i}) = \mathrm {M} _ {\mathrm {n e a r}} \left(\mathbf {x}, \mathbf {z}, \mathbf {n}, \mathrm {e n c} _ {\mathrm {d i r}} (\mathbf {v}), \mathbf {F} _ {\mathrm {n e a r} _ {i}}, \boldsymbol {\omega} _ {i}\right) \\ \otimes \frac {L _ {\text {n e a r} _ {i}} (\mathbf {x} , \boldsymbol {\omega} _ {i})}{\left\| \mathbf {p} _ {i} - \mathbf {x} \right\| _ {2} ^ {2}} \left(\boldsymbol {\omega} _ {i} \cdot \mathbf {n}\right) ^ {+}, \tag {8} \\ \end{array} +$$ + +where $\otimes$ denotes element-wise multiplication. On the right side, the second term denotes the incident radiance attenuated by square distance, the third term denotes the cosine weakening factor, and $\mathrm{M}_{\mathrm{near}}$ accounts for other factors like reflectance and visibility. Owing to the instability of these terms at the training's outset, we use an annealing strategy, + +as recommended in prior works [44], gradually replacing default values (e.g., 1 for incident radiance) with those optimized during training, to ease through the process. + +# 3.3. Differentiable Rendering + +To simplify the notation, we omit $\mathbf{p}_i$ and the lighting index $i$ in the equations within this subsection. Our approach optimizes the lighting alongside the implicit fields primarily by reducing the discrepancy between the observed and differentiably re-rendered appearances. The re-rendering is accomplished using two types of renderer: a neural radiance field renderer (Fig. 2 bottom-right) and a physically-based surface renderer (Fig. 2 top-right; Eq. (1)). We handle each lighting condition separately in rendering, and then aggregate the results based on per-image lighting condition. + +Volume Rendering for SDF Field. Consider a camera ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ ( $t > 0$ ), with $\mathbf{o} \in \mathbb{R}^3$ as its origin and $\mathbf{d} \in \mathbb{S}^2$ as its direction. We incorporate importance sampling using an occupancy grid [14] to get $N \leq 1024$ points along this ray, denoted as $\mathbf{r}(t_i)$ , $i \in \mathbb{N}_N$ . Utilizing techniques from NeuS [29], we transform the signed distances $s(\mathbf{r}(t_i))$ into discrete opacity values as follows: + +$$ +\alpha_ {i} = \left(\frac {\Phi_ {b} (s (\mathbf {r} (t _ {i}))) - \Phi_ {b} (s (\mathbf {r} (t _ {i + 1})))}{\Phi_ {b} (s (\mathbf {r} (t _ {i})))}\right) ^ {+}, \tag {9} +$$ + +where $\Phi_b(s) = 1 / (1 + e^{-bs})$ is the cumulative opacity distribution function. The parameter $b$ is trainable and tends to increase during training, focusing opacity more narrowly around the surface where $s = 0$ . The accumulated transmittance is then calculated as $T_{i} = \prod_{j = 1}^{i - 1}(1 - \alpha_{j})$ . + +Neural Radiance Field Rendering. The color $\mathbf{C}_{\mathrm{rf}}$ of a camera ray under either far- or near-field lighting conditions, as rendered through the neural radiance field, is determined by accumulating the radiance along the ray's path: + +$$ +\mathbf {C} _ {\mathrm {r f}, \{\text {f a r , n e a r} \}} (\mathbf {o}, \mathbf {d}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} _ {\{\text {f a r , n e a r} \}} (\mathbf {r} (t _ {i}), - \mathbf {d}). \tag {10} +$$ + +This formula sums the radiance from each sampled point to yield the color of the ray under the specified lighting. + +Physically-based Rendering. Our method employs differentiable surface rendering, modeling the observed radiance as reflected from a single surface point. We compute the depth $t$ of this surface point similarly to Eq. (10): $t = \sum_{i=1}^{N} T_i \alpha_i t_i$ . Accumulation along the ray also applies to the BRDF parameters $\beta$ and the surface normal $n$ . For each input lighting, we render the appearance that models secondary visibility and indirect illumination. The rendering method under direct illumination differs for far- and near-field lighting. For direct far-field lighting $L_{\mathrm{dfar}}$ , we employ Monte Carlo integration to evaluate the rendering equation (Eq. (1)) as the incident light comes from all + +directions: + +$$ +\begin{array}{l} \mathbf {C} _ {\mathrm {p b , d f a r}} (\mathbf {o}, \mathbf {d}) = L _ {\mathrm {o}} (\mathbf {r} (t), - \mathbf {d}) \\ = \frac {1}{S} \sum_ {s = 1} ^ {S} \frac {L _ {\mathrm {d f a r}} (\mathbf {r} (t) , \boldsymbol {\omega} _ {s}) f (\mathbf {r} (t) , \boldsymbol {\omega} _ {s} , - \mathbf {d}) \left(\boldsymbol {\omega} _ {s} \cdot \mathbf {n}\right) ^ {+}}{p \left(\boldsymbol {\omega} _ {s}\right)}, \tag {11} \\ \end{array} +$$ + +where $S$ is the number of sampled directions, $\omega_{s}$ a sampled incident direction, and $p(\omega_{s})$ the probability density function (PDF) for the sampled direction. We use multiple importance sampling [26] to combine evaluations from different strategies: BRDF importance sampling [13], SG lighting importance sampling [33], and cosine importance sampling. we follow differentiable rendering works [15] to use a small $S = 20$ , as stochastic gradient descent handles noisy gradients. The incident radiance, now taking the visibility of light sources into account, is given by: + +$$ +L _ {\mathrm {d f a r}} (\mathbf {r} (t), \boldsymbol {\omega} _ {s}) = L _ {\mathrm {f a r}} (\boldsymbol {\omega} _ {s}) V (\mathbf {r} (t), \boldsymbol {\omega} _ {s}), \tag {12} +$$ + +with $V(\mathbf{r}(t), \boldsymbol{\omega}_s) \in [0,1]$ representing visibility, or the inverse of the accumulated opacity along a secondary ray $\mathbf{r}_{\mathrm{sec}}(t') = \mathbf{r}(t) + t' \boldsymbol{\omega}_s$ , ( $t' > 0$ ), computed in a manner akin to Eq. (10). For direct near-field lighting $L_{\mathrm{dnear}}$ , the integral in Eq. (1) simplifies to a multiplication, as the light originates from a single direction: + +$$ +\mathbf {C} _ {\mathrm {p b}, \mathrm {d n e a r}} (\mathbf {o}, \mathbf {d}) = L _ {\mathrm {d n e a r}} (\mathbf {r} (t), \boldsymbol {\omega}) f (\mathbf {r} (t), \boldsymbol {\omega}, - \mathbf {d}) \left(\boldsymbol {\omega} \cdot \mathbf {n}\right) ^ {+}, \tag {13} +$$ + +where $\omega = (\mathbf{p} - \mathbf{r}(t)) / \| \mathbf{p} - \mathbf{r}(t)\|_{2}$ is the direction of incident light, and $L_{\mathrm{dnear}}(\mathbf{r}(t),\omega)$ is the potentially occluded incident radiance, computed similarly to Eq. (12). + +Indirect illumination $L_{\mathrm{ind}}$ relevant to both far-field and near-field lighting, considers radiance reflected from the scene itself, potentially from all directions, thus necessitating Monte Carlo integration. The neural radiance field $\mathbf{C}_{\mathrm{rf}}$ substitutes multi-bounce path tracing for indirect illumination, given its encapsulation of the scene radiance under infinite lighting bounces. The radiance $\mathbf{C}_{\mathrm{pb,ind}\{\mathrm{far,near}\}}$ from indirect illumination is evaluated similarly to Eq. (12), but with the $L_{\mathrm{dfar}}$ replaced by the indirect lighting $L_{\mathrm{ind}\{\mathrm{far,near}\}}$ : + +$$ +L _ {\text {i n d} *} (\mathbf {r} (t), \omega_ {s}) = \mathbf {c} _ {*} \left(\mathbf {x} ^ {\prime}, - \omega_ {s}\right) \left(1 - V (\mathbf {r} (t), \omega_ {s})\right), \tag {14} +$$ + +where $\mathbf{x}'$ is the intersection of the secondary ray $\mathbf{r}_{\mathrm{sec}}(t')$ with the scene geometry. The radiance from indirect illumination is also computed for each lighting condition separately, then combined with direct illumination to yield the complete PBR radiance for the camera ray under a specific lighting condition: + +$$ +\mathbf {C} _ {\mathrm {p b}, \{\text {f a r , n e a r} \}} = \mathbf {C} _ {\mathrm {p b}, \mathrm {d} \{\text {f a r , n e a r} \}} + \mathbf {C} _ {\mathrm {p b}, \text {i n d} \{\text {f a r , n e a r} \}}. \tag {15} +$$ + +# 3.4. Training + +Training Schemes. Our training process is divided into two sequential stages. In the first stage, the focus is on training the geometry and radiance field without employ + +ing PBR. The objectives are: 1) to recover the scene's geometry, 2) to distinguish the appearance under each lighting condition, a necessity due to the potential presence of multiple light sources in one image, and 3) to utilize the reconstructed radiance field for indirect illumination. In the second stage, we train the material field and scene lighting using PBR, aiming to estimate material properties. + +Loss Functions. We average the loss functions across batches of camera rays. During the first stage, our loss function compares the total re-rendered radiance under all lighting conditions present in an image – including a far-field light $i$ and any active near-field lights – with the ground-truth (GT) color $\mathbf{C}_{\mathrm{gt}}$ from the input observation. This comparison trains the geometry and radiance field: + +$$ +\mathcal {L} _ {\mathrm {r f}} = \left\| \mathbf {C} _ {\mathrm {r f}, \text {f a r} _ {i}} + \sum_ {\text {n e a r} i \text {i s o n}} \mathbf {C} _ {\mathrm {r f}, \text {n e a r} _ {i}} - \mathbf {C} _ {\mathrm {g t}} \right\| _ {2} ^ {2}. \tag {16} +$$ + +The Eikonal loss [8] is applied to the gradients of the SDF values $s$ for geometric regularization: + +$$ +\mathcal {L} _ {\mathrm {e i k}} = \sum_ {j = 1} ^ {N} T _ {j} \alpha_ {j} \left(\left\| \frac {\partial s (\mathbf {r} (t _ {j}))}{\partial \mathbf {r} (t _ {j})} \right\| _ {2} - 1\right) ^ {2}, \tag {17} +$$ + +A silhouette loss using GT alpha $\alpha_{\mathrm{gt}}$ from the foreground mask and a normal smoothness loss [41] aid shape recovery: + +$$ +\mathcal {L} _ {\mathrm {s i l}} = \left(\left(1 - T _ {N + 1}\right) - \alpha_ {\mathrm {g t}}\right) ^ {2}. \tag {18} +$$ + +$$ +\mathcal {L} _ {\mathrm {n s}} = \sum_ {j = 1} ^ {N} T _ {j} \alpha_ {j} \left(\mathbf {n} \left(\mathbf {r} \left(t _ {j}\right)\right) - \mathbf {n} \left(\mathbf {r} \left(t _ {j}\right) + \boldsymbol {\epsilon}\right)\right) ^ {2}, \tag {19} +$$ + +where $\epsilon \in \mathbb{R}^3$ is a small random perturbation. The total loss for stage one is $\mathcal{L}_{\mathrm{rf}} + \lambda_{\mathrm{eik}}\mathcal{L}_{\mathrm{eik}} + \lambda_{\mathrm{sil}}\mathcal{L}_{\mathrm{sil}} + \lambda_{\mathrm{ns}}\mathcal{L}_{\mathrm{ns}}$ , with $\lambda_*$ representing the loss weights. In stage two, PBR colors replace radiance field-rendered colors for comparison with $\mathbf{C}_{\mathrm{gt}}$ , and a novel self-consistency loss is introduced between neural and PBR radiance under each lighting for additional supervision. The direct supervision loss $\mathcal{L}_{\mathrm{pb}}$ mirrors $\mathcal{L}_{\mathrm{rf}}$ with $\mathbf{C}_{\mathrm{rf},\{\mathrm{far,near}\}_i}$ replaced by $\mathbf{C}_{\mathrm{pb},\{\mathrm{far,near}\}_i}$ . The perlighting self-consistency loss $\mathcal{L}_{\mathrm{con}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {c o n}} = \sum_ {i = 1} ^ {N _ {\{\text {f a r}, \text {n e a r} \}}} \left\| \mathbf {C} _ {\mathrm {r f}, \{\text {f a r}, \text {n e a r} \} _ {i}} - \mathbf {C} _ {\mathrm {p b}, \{\text {f a r}, \text {n e a r} \} _ {i}} \right\| _ {2} ^ {2}. \tag {20} +$$ + +This loss is crucial for discerning contributions from each lighting in images with multiple active light sources, thereby reducing ambiguity and improving material estimation. A material smoothness loss $\mathcal{L}_{\mathrm{ms}}$ is used similar to $\mathcal{L}_{\mathrm{ns}}$ . The total loss for stage two is $\mathcal{L}_{\mathrm{pb}} + \lambda_{\mathrm{con}}\mathcal{L}_{\mathrm{con}} + \lambda_{\mathrm{ms}}\mathcal{L}_{\mathrm{ms}}$ . + +Our model is trained on an NVIDIA RTX 3090 GPU for a total of 40,000 steps. The first 30,000 steps of stage one take about 20 minutes, while the subsequent 10,000 steps of stage two require about 40 minutes. Post-training, we follow a procedure similar to WildLight [6] to extract the fields into textured meshes, which can be easily integrated into industry-standard rendering software like Blender [1], + +Table 2. Quantitative comparison results with state-of-the-art methods averaged on 6 synthetic scenes. We show results of surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, the specular reflection part of FV relit RGB, and training time on a single RTX 3090 GPU. We mark the best and the second best results in each column. $\uparrow (\downarrow)$ means bigger (smaller) is better. + +
MethodInput lighting conditionsNormal MAngE↓AlbedoView synthesisFV relitFV relit (spec)Time
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
(1) TensoIR [10]17.6626.480.92129.480.91228.180.90128.300.861300 mins
(2) NVDiffRecMC [9]Single far-field16.2426.520.91527.130.91326.610.90125.570.836150 mins
(3) Ours12.3924.500.88228.200.93427.460.92127.560.87145 mins
(4) WildLight [6]Single far-field + Flashlight11.4928.860.94029.870.92930.440.93027.710.8631440 mins
(5) Ours10.8931.620.95332.090.95332.000.95330.750.90660 mins
(6) TensoIR [10]Two far-field16.2427.180.92929.680.91228.660.90228.460.863300 mins
(7) Ours11.7026.070.90229.590.94229.270.93429.090.89045 mins
(8) OursTwo far + Single near10.7932.040.95732.100.95032.380.95431.400.91060 mins
+ +enabling fast and high-quality rendering suitable for various applications. Check the supplement for more details. + +# 4. Experiments + +Our experiments involve a comprehensive comparison with state-of-the-art methods (Sec. 4.2), ablation studies (Sec. 4.3), and evaluations on both synthetic datasets (Sec. 4.1) and real-world images (Sec. 4.4). The supplement shows further data creation details and more experiments. + +# 4.1. Synthetic Datasets + +We collect 6 synthetic scenes comprising a variety of scene geometries and materials. For training, each scene is rendered under four distinct lighting setups: 1) single far-field light, 2) single far-field light with a camera-collocated flashlight, 3) two far-field lights, and 4) two far-field lights with a near-field light source. We render 100 training images per setting using random viewpoints and per-image lighting conditions. The far-field lighting from the first setting is used in subsequent settings, with the last two sharing their far-field lighting. For testing, 200 images under one common far-field lighting are rendered to assess novel view synthesis under seen lighting, and another 200 under unseen far-field lighting to evaluate free-viewpoint (FV) relighting, which largely depends on the geometry and material estimation quality. This methodology provides a relatively fair comparison across methods with different abilities regarding supported input lighting conditions, trained under similar conditions and tested all using the same sets. + +# 4.2. Comparison with State-of-the-art Methods + +We compare our method with leading multi-view inverse rendering methods under each input lighting scenario. For single far-field lighting inputs (Tab. 1 row 1), we compare with NVDiffRecMC [9] and TensoIR [10] (single-lighting input). For multiple far-field lighting scenarios (Tab. 1 row 3), we compare with TensoIR [10]. For setups involving a far-field light and a flashlight (Tab. 1 row 4), we compare + +with WildLight [6]. We use Blender [1] for high-quality relit results for our method, WildLight [6], and NVDiffRecMC [9], importing trained SDF and material fields as textured meshes. Since TensoIR [10] employs density-based geometry and cannot extract high-quality meshes, we utilize its differentiable renderer for relighting. + +Our quantitative comparison employs mean angular error (MAngE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) [30] metrics for surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, and specular reflection of FV relit RGB. We follow NeRFactor [41] in assuming albedo and lighting brightness as scale-invariant, necessitating metric comparisons using corresponding scales. For visualization and metrics, we scale each RGB channel of albedo and relit images from all methods by a global scale to minimize mean squared error against ground truth. + +The quantitative results are detailed in Tab. 2. We observe that when only input images under single far-field lighting are available (rows 1-3), our method gives comparable results as TensoIR [10] and NVDiffRecMC [9]. Adding a flashlight (rows 4, 5) significantly helps in geometry and material estimation (comparing rows 3, 5), where the greatest improvement is made regarding the diffuse albedo. Also, our method clearly surpasses WildLight [6] in both quality and speed in this setting. Leveraging another far-field lighting condition (rows 6, 7) brings similar effects as using a flashlight, but generally with a lower degree of improvement. Further adding a flashlight (row 8) gives the best results among all settings and methods, but the gain is marginal compared to the setting of single far-field lighting with a flashlight (comparing rows 5, 8). + +Fig. 3 shows the qualitative comparison on two synthetic scenes: LEGO (left) with detailed geometry, and TROOPER (right) with highly-reflective SV materials. The results show that our method predicts more detailed and faithful shape, diffuse albedo, and specular parameters that could produce accurate and realistic relit results under unseen + +![](images/b2fb9ddc485f21d9ae4368406011da2716409663ac6d6e3046cbad3fe42965c6.jpg) +Figure 3. Comparison with state-of-the-art methods. We test NVDiffRecMC [9] (single far-field lighting in input), TensoIR [10] (two far-field lighting in input), and WildLight [6] and our method (single far-field lighting with a flashlight in input) on two synthetic scenes. The intensity of the specular reflection is multiplied by $3 \times$ for the visualization purpose. Please zoom in for details. + +![](images/ee09e28916cb666d5eed9b22d6bf9faf6bea8a608629c3c7f21f6da5d7825cd0.jpg) + +Table 3. Ablation study on our method using inputs under single far-field lighting with a flashlight. + +
ModelNormalAlbedoFV RelitRelit (spec)
MAngE↓PSNR↑PSNR↑PSNR↑
(1) w/o Lcon8.45227.7029.8529.08
(2) Mod. cnear8.57227.7930.0328.87
(3) w/o encmat8.26327.7930.0728.92
(4) Full model8.29828.0530.1929.36
+ +lighting. In contrast, WildLight [6] tends to overly smooth geometry and material. TensoIR [10] can not handle highly-reflective surfaces, and lighting is baked into its predicted diffuse albedo for TROOPER. + +# 4.3. Ablation Studies + +In Tab. 3, we quantitatively evaluate ablation models on a subset of our synthetic scenes: 1) the model excluding the self-consistency loss $\mathcal{L}_{\mathrm{con}}$ , 2) the model where the output of the near-field radiance MLP is not explicitly multiplied by the cosine weakening factor and incident radiance (Eq. (8)), 3) the model with $\mathrm{enc}_{\mathrm{mat}}(\mathbf{x})$ replaced by $\mathrm{enc}_{\mathrm{geo}}(\mathbf{x})$ . The results show that all the above techniques enhance the geometry and materials estimated by our method. + +# 4.4. Real-world Results + +Fig. 4 shows the results of our method on two real-world scenes captured under single far-field lighting and a flashlight. Our method faithfully recovers the geometry and material of the objects and produces realistic relit results. + +# 5. Conclusion + +We introduce a versatile multi-view inverse rendering framework, distinguished by its ability to leverage input im + +![](images/cbf612ad1b7d69f94b9168c7654cd04c1bf2429980067f248be7b32e4ec3c024.jpg) + +![](images/2170a37d6d11c10633e96c8b249bafbbbe6b0a96d45ffda7305aa22d64cfe820.jpg) +Figure 4. The results of our method on two real-world scenes, using multi-view images under one ambient lighting and a flashlight. The specular reflection intensity is increased for visualization. + +ages under varied far- and near-field light sources available in capture for better geometry and material estimation. + +Limitations. VMINer does not model the background, necessitating a foreground mask for each image. It also does not consider unknown tone-mapping curves applied during the image signal processing (ISP) stage, using a fixed curve with $\gamma = 2.2$ to the computed linear radiance. It depends on distinguishing the contribution from each lighting, which implies that, for optimal results, near-field light sources should be switched on/off during capture. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China (Grand No. 62136001, 62088102). + +# References + +[1] Blender Foundation. The Blender project - free and open 3D creation software. https://www.blender.org. Accessed: 2023-11-07. 2, 6, 7 +[2] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4 +[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. NeuralPIL: Neural pre-integrated lighting for reflectance decomposition. In Proc. of Neural Information Processing Systems (NeurIPS), 2021. +[4] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P. A. Lensch, and Varun Jampani. SAMURAI: shape and material from unconstrained real-world arbitrary image collections. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 2 +[5] Brent Burley. Physically-based shading at disney. In SIGGRAPH 2012 Courses, 2012. 3 +[6] Ziang Cheng, Junxuan Li, and Hongdong Li. WildLight: In-the-wild inverse rendering with a flashlight. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7, 8 +[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3 +[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proc. of International Conference on Machine Learning (ICML), 2020. 6 +[9] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using Monte Carlo rendering and denoising. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 7, 8 +[10] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 4, 7, 8 +[11] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 1, 2 +[12] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG), 41(4):56:1-56:12, 2022. 1, 2, 3 +[13] Eric P Lafortune and Yves D Willems. Using the modified phong reflectance model for physically based rendering. Katholieke Universiteit Leuven. Departement Computerwetenschappen, 1994. 6 + +[14] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates NeRFs. arXiv preprint arXiv:2305.04966, 2023. 5 +[15] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable Monte Carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6): 222, 2018. 6 +[16] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[17] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Milos Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In Proc. of European Conference on Computer Vision (ECCV), 2022. 3 +[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1, 4 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 1, 4, 5 +[21] Jacob Munkberg, Wenzheng Chen, Jon Hasselgren, Alex Evans, Tianchang Shen, Thomas Müller, Jun Gao, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[22] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 1 +[23] Viktor Rudnev, Mohamed Elgharib, William A. P. Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3 +[24] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3 +[25] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. ReLight My NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 + +[26] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 6 +[27] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. In Proc. of the Eurographics Symposium on Rendering Techniques, 2007. 5 +[28] Jiaping Wang, Peiran Ren, Minmin Gong, John M. Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM Transactions on Graphics (TOG), 28(5):133, 2009. 4 +[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proc. of Neural Information Processing Systems (NeurIPS), 2021. 4, 5 +[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7 +[31] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[32] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3 +[33] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 6 +[34] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 2 +[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3 +[36] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(8):10129-10142, 2023. 1, 2, 3 +[37] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 3 +[38] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: + +Inter-reflectable light fields for geometry and material estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3 +[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4 +[40] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: inverse rendering by optimizing neural SDFs and materials from photometric images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[41] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 2021. 1, 2, 3, 4, 6, 7 +[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4 +[43] Youjia Zhang, Teng Xu, Junqing Yu, Yuteng Ye, Yanqing Jing, Junle Wang, Jingyi Yu, and Wei Yang. NeMF: Inverse volume rendering with neural microflake field. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[44] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG), 41(6):235:1-235:17, 2022. 4, 5 +[45] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable Monte Carlo raytracing. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2022. 3 +[46] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Di-anbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I $^2$ -SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3 +[47] Yongjie Zhu, Yinda Zhang, Si Li, and Boxin Shi. Spatially-varying outdoor lighting estimation from intrinsics. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 \ No newline at end of file diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/images.zip b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4af6fc3fb0a467e38ebc3bdb17f521c30ff031ba --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00f0e33838c5c2dad6041a06e96e782833789e7d50b22c324267edc2c7f789c9 +size 688645 diff --git a/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/layout.json b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab62893eb81f15a52f45d74c5184965e3b09abe2 --- /dev/null +++ b/2024/VMINer_ Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources/layout.json @@ -0,0 +1,10035 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 451, + 140 + ], + "type": "text", + "content": "VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 158, + 208, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 158, + 208, + 172 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 208, + 172 + ], + "type": "text", + "content": "Fan Fei" + }, + { + "bbox": [ + 151, + 158, + 208, + 172 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 159, + 294, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 159, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 220, + 159, + 294, + 174 + ], + "type": "text", + "content": "Jiajun Tang" + }, + { + "bbox": [ + 220, + 159, + 294, + 174 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 159, + 362, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 362, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 362, + 174 + ], + "type": "text", + "content": "Ping Tan3,4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 375, + 159, + 440, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 159, + 440, + 172 + ], + "spans": [ + { + "bbox": [ + 375, + 159, + 440, + 172 + ], + "type": "text", + "content": "Boxin Shi" + }, + { + "bbox": [ + 375, + 159, + 440, + 172 + ], + "type": "inline_equation", + "content": "^{1,2\\#}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "spans": [ + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University " + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University " + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 177, + 527, + 221 + ], + "type": "text", + "content": "Light Illusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 238, + 192, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 238, + 192, + 251 + ], + "spans": [ + { + "bbox": [ + 143, + 238, + 192, + 251 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 258, + 290, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 462 + ], + "type": "text", + "content": "This paper introduces a versatile multi-view inverse rendering framework with near- and far-field light sources. Tackling the fundamental challenge of inherent ambiguity in inverse rendering, our framework adopts a lightweight yet inclusive lighting model for different near- and far-field lights, thus is able to make use of input images under varied lighting conditions available during capture. It leverages observations under each lighting to disentangle the intrinsic geometry and material from the external lighting, using both neural radiance field rendering and physically-based surface rendering on the 3D implicit fields. After training, the reconstructed scene is extracted to a textured triangle mesh for seamless integration into industrial rendering software for various applications. Quantitatively and qualitatively tested on synthetic and real-world scenes, our method shows superiority to state-of-the-art multi-view inverse rendering methods in both speed and quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 473, + 128, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 128, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 128, + 486 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 493, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 696 + ], + "type": "text", + "content": "The reconstruction of 3D scenes from multi-view RGB imagery has experienced significant advancements following the development of Neural Radiance Fields (NeRF) [19], and both the speed and quality of the reconstruction have reached an unprecedented level [18, 20]. Despite these improvements, a common limitation is to represent only the radiance field, which is a complex product of the external lighting interacting with the intrinsic geometry and material of the scene [11]. The entangled nature of this representation generally hampers the ability to accurately render the scene under unseen lighting conditions, because the influence of the original lighting is embedded within the newly rendered scene. Recent approaches [24, 41, 46] have incorporated inverse rendering [22] to separate material properties and lighting effects, extending the application of the reconstruction beyond novel view synthesis to novel scenarios involving relighting and material editing [2]. Some" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 238, + 410, + 318 + ], + "blocks": [ + { + "bbox": [ + 308, + 238, + 410, + 318 + ], + "lines": [ + { + "bbox": [ + 308, + 238, + 410, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 410, + 318 + ], + "type": "image", + "image_path": "666655ee759469bc512e25e6e361d78e0c88657a3c0f2c78b5e8fb4eba83864d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 416, + 238, + 545, + 318 + ], + "blocks": [ + { + "bbox": [ + 416, + 238, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 416, + 238, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 416, + 238, + 545, + 318 + ], + "type": "image", + "image_path": "f96c9997a11d572c792b71a5aa2263c68a0ac967236f2c8372f60973b522a1b9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 320, + 403, + 410 + ], + "blocks": [ + { + "bbox": [ + 308, + 320, + 403, + 410 + ], + "lines": [ + { + "bbox": [ + 308, + 320, + 403, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 403, + 410 + ], + "type": "image", + "image_path": "4ed3a4b2f4591e97d11b8176d393cd24d837d0542efced48509631783bdb842e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 412, + 546, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 412, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 546, + 468 + ], + "type": "text", + "content": "Figure 1. Given multi-view images under possibly varied lighting, our method leverages all present lighting conditions to reconstruct scene geometry and material disentangled with lighting. The trained fields are extracted to textured meshes for seamless integration into industrial renderers for various applications." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 406, + 320, + 545, + 410 + ], + "blocks": [ + { + "bbox": [ + 406, + 320, + 545, + 410 + ], + "lines": [ + { + "bbox": [ + 406, + 320, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 406, + 320, + 545, + 410 + ], + "type": "image", + "image_path": "cb050fcc3eab0ef1ec35bac7fe3512835a41341a3386d274f77df58c883014ea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 533 + ], + "type": "text", + "content": "methods [2, 6, 9, 21, 32] have advanced to the extraction of detailed triangle meshes with material UV textures, which can serve as economical and lifelike 3D models for gaming and cinematography industries [9, 21], thereby marking a transformative step in digital asset creation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": "Inverse rendering presents several fundamental challenges, one of them being its severe inherent ambiguity [41]. To combat this, multi-view inverse rendering methodologies usually reduce the ambiguity by imposing various constraints on each scene component. These methods can be categorized based on the assumption on the amount and types of lighting conditions present in the input images (Tab. 1). The majority of methods [9, 21, 24, 35, 36, 39, 41, 42] assumes the imagery to be captured under one fixed lighting condition (rows 1, 2). However, in such scenarios, only one case of scene appearances among all possible cases under different lighting is observed, causing severe lighting-material ambiguity, posing considerable difficulties to material estimation. In a bid to alleviate this ambiguity, recent techniques [2, 6, 10, 12, 23, 32] utilize input" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 230, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 230, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 230, + 713 + ], + "type": "text", + "content": "Corresponding author. E-mail: shiboxin@pku.edu.cn." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11800" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 85, + 542, + 191 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 530, + 83 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 530, + 83 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 530, + 83 + ], + "type": "text", + "content": "Table 1. A summary of multi-view inverse rendering methods based on their assumptions and usages of input lighting conditions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 85, + 542, + 191 + ], + "lines": [ + { + "bbox": [ + 51, + 85, + 542, + 191 + ], + "spans": [ + { + "bbox": [ + 51, + 85, + 542, + 191 + ], + "type": "table", + "html": "
Methods# of far-field light# of near-field lightMaterial optimizationLighting-material ambiguityCapture workload
(1) [24, 39, 41, 42]SingleNoneOnly from singleStrongLow
(2) NeILF [35]Single 5D incident light field
(3) [2, 10, 12, 32]aMultipleNoneOnly from far-fieldModerateHigh
(4) WildLight [6]SingleSingleOnly from near-fieldModerateLow
(5) OursSingleSingleFrom near- and far-fieldWeakLow
MultipleSingle or multipleAlmost noneHigh
", + "image_path": "f287bb1c871bd177c92373877b5d111ff1d3c3698d7c1f85887194aa065931a0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 58, + 192, + 518, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 192, + 518, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 192, + 518, + 202 + ], + "type": "inline_equation", + "content": "{}^{a}" + }, + { + "bbox": [ + 58, + 192, + 518, + 202 + ], + "type": "text", + "content": " These methods also support input images under single far-field lighting. In that case, they belong to single far-field lighting methods (row 1)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 410 + ], + "type": "text", + "content": "images under varied lighting to discern materials from lighting (rows 3, 4). This demands more complex lighting models. Existing varied-lighting methods [2, 10, 12, 23, 32] typically model multiple, but only far-field, lighting conditions to maintain manageable unknown lighting parameters. This means throughout the capture, the lighting setups should be largely changed by either moving the objects to reconstruct or adapting far-field light sources, causing increased workload to the data collection process. Alternatively, near-field light sources, such as flashlights, can function as easily controllable light sources for disambiguation. WildLight [6] suggests the use of a camera-collocated flashlight in addition to the ambient lighting for inverse rendering. Nevertheless, it limits itself by using only the isolated appearance under the flashlight for material estimation, not harnessing available observations under the ambient lighting and degrading the quality of its estimated material (row 4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 613 + ], + "type": "text", + "content": "Addressing the challenges and restrictions previously outlined, this paper introduces VMINer, a Versatile Multiview Inverse rendering framework with Near- and far-field light sources (row 5). The distinctive feature of our framework is that it uses a lightweight yet inclusive lighting model for different far- and near-field light sources and efficiently leverages available observations under each lighting to disentangle lighting from other components, making the most of whatever lighting conditions are at hand to boost the practicality and quality of the reconstruction, as shown in Fig. 1. Although our method accepts the simplest setting of inputs under single far-field lighting, a more effective compromise is to employ an additional flashlight to enhance quality without overly complicating the capture process. The ideal scenario for the highest quality reconstruction would involve capturing the scene under varied far-field lighting and using a flashlight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 613, + 288, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 708 + ], + "type": "text", + "content": "Nonetheless, implementing such a versatile framework is far from straightforward. The method has to render the scene under different lighting conditions, including spatially-varying and changing ones to represent possibly moving near-field lights. We propose to solve the problem by modeling near-field lights as point sources in 3D space with adjustable positions and intensities, which could be anisotropic and not necessarily camera-aligned (e.g.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 206, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 547, + 399 + ], + "type": "text", + "content": "stationary desk lamps). This representation effectively approximates a broad range of common near-field lights and is easy for spatially-consistent editing and optimization. We use 3D implicit fields to model the scenes' radiance, shape, and material and minimize the error of the neural or physically-based re-rendered scene appearance under each input lighting. To account for multiple lighting situations and moving near-field lights, we integrate specially tailored multi-layer perceptrons (MLPs) that additionally process light directions and embeddings. Post-training, the fields are converted into a textured triangle mesh, ready for seamless integration into industrial rendering software like Blender [1], facilitating a broad spectrum of applications. Our extensive experiments, covering both synthetic and real-world scenes, demonstrate that VMINer surpasses prior methods quantitatively and qualitatively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 410, + 392, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 392, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 392, + 423 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 430, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 491 + ], + "type": "text", + "content": "Preliminary. The rendering equation [11] models the reflected radiance from any point on a surface as a result of an intricate surface integral. This integral comprises contributions from three fundamental scene components: lighting, material, and geometry. The equation is expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 495, + 545, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 545, + 528 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) = \\int_ {\\Omega} \\underbrace {L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}})} _ {\\text {l i g h t i n g}} \\underbrace {f (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}})} _ {\\text {m a t e r i a l}} \\underbrace {(\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) ^ {+}} _ {\\text {g e o m e t r y}} \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}. \\tag {1}", + "image_path": "6971f605b2f344d7f1f2578d1c2abb93668c5ab02f32efc44b8a732f1909ce85.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}})" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " represents the radiance reflected in the outgoing direction " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " from a surface point " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " in 3D space. " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " denotes the unit hemisphere encompassing all incident directions " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}} \\cdot \\mathbf{n} > 0" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " is the incident radiance at " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " from direction " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}})" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " is the bidirectional reflectance distribution function (BRDF), denoting the ratio of light reflected along " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "(\\omega_{\\mathrm{i}} \\cdot \\mathbf{n})^{+}" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "x^{+} \\triangleq \\max(x, 0)" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": ", signifies the cosine weakening factor. The incident lighting " + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})" + }, + { + "bbox": [ + 304, + 532, + 547, + 663 + ], + "type": "text", + "content": " may originate from other points within the scene, necessitating recursive evaluation of this integral." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Multi-view Inverse Rendering. Multi-view inverse rendering techniques [2-4, 6, 9, 10, 12, 21, 23, 24, 34-36, 38-43] generally employ implicit or explicit 3D fields to model the scene's geometry and material, while also estimating" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11801" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 255 + ], + "type": "image", + "image_path": "0ca25f6f05cfdb6c0ded64883c3a2eba6824f20bd9a754e282e7ed753d092aa1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 257, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 48, + 257, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 545, + 270 + ], + "type": "text", + "content": ": input : variable : refer to another entry : untrainable operator : trainable operator / MLP " + }, + { + "bbox": [ + 48, + 257, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\bigotimes \\bigoplus" + }, + { + "bbox": [ + 48, + 257, + 545, + 270 + ], + "type": "text", + "content": " element-wise multiplication / add" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": "Figure 2. Bottom-left: VMINer takes as input multi-view RGB images with foreground masks under possibly varied far- and near-field lighting. The total number of lighting conditions and their types should be given. Each image is illuminated by one far-field lighting whose index is \"far index\", and near-field lights whose on/off states are \"near on/off\". Top-left: VMINer models each far-field lighting " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "(\\mathrm{far}_i)" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " and near-field lighting " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "(\\mathrm{near}_i)" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " separately as parametric models, from which incoming radiance is queried at any position " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ". Bottom-mid: It utilizes 3D implicit representations for scene geometry, material, and radiance. Four MLPs process " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " encoded by different encoders " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\{\\mathrm{geo},\\mathrm{mat},\\mathrm{dir}\\}}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " with light embeddings " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " to get the SDF " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ", the neural appearance descriptor " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ", the BRDF parameters " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ", and the radiance " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\{\\mathrm{far},\\mathrm{near}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " under each lighting. Bottom-right: It then uses radiance field rendering to re-render the appearance " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " separately under each lighting condition, added up to " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf}}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " according to the per-image lighting condition. Top-right: Physically-based surface rendering uses Monte Carlo integration to evaluate both direct illumination " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{d}\\{\\mathrm{near},\\mathrm{far}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " with secondary visibility and indirect illumination " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}\\{\\mathrm{near},\\mathrm{far}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ". The rendered appearances are aggregated to " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{pb}}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\{\\mathrm{rf},\\mathrm{pb}\\}}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " are compared with the observation " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " to train the scene model. " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{near},\\mathrm{far}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": " are also used as additional supervision signals for " + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{near},\\mathrm{far}\\}_i}" + }, + { + "bbox": [ + 46, + 272, + 546, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 399, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 710 + ], + "type": "text", + "content": "lighting to recreate the scene's appearance. They optimize each scene component from scratch by aligning the rendered with the observed appearance, while avoiding potential local minima, which, although do not faithfully represent the scene, can replicate the observed input. Addressing this inherently ambiguous nature of inverse rendering, these methods typically impose constraints on each scene component. For geometry and material, assumptions of known geometry [35, 36, 46] are made, along with leveraging the recovered shape from radiance fields [6, 10, 41], adopting Lambertian or uniform materials [23, 39], controlling BRDF parameter smoothness [10], or integrating a learned BRDF latent space [41]. Regarding lighting, many approaches presuppose known lighting conditions [24, 40] or fixed lighting shared across input images [35, 38, 41, 42]. More recent techniques [2, 6, 10, 12, 23, 32] have turned to using images with varied lighting conditions to mitigate the lighting-material ambiguity. They typically accept input images that can be grouped into sets, each illuminated by one distinct and fixed far-field lighting. WildLight [6] exceptionally employs a camera-collocated flashlight along with fixed ambient lighting. However, it does not model the ambient lighting and thus its material property estimation relies exclusively on flashlight observations, where the light and view direction always coincide [5], limiting its BRDF estimation ability. Our VMINer, in contrast, accommodates" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 399, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 434 + ], + "type": "text", + "content": "input images under single or multiple far-field lighting conditions with near-field light sources and leverages each of them to enhance the quality of the estimated material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 435, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 639 + ], + "type": "text", + "content": "Lighting Models. Spatially-uniform (SU) lighting models rely on the assumption that the scene's lighting originates from a distant source, modeled as spherical Gaussians (SGs) [2, 10], spherical harmonics (SH) [12, 23], or MLPs [32]. In contrast, spatially-varying (SV) lighting models acknowledge the presence of near-field light sources, causing different locations to receive different incident lighting. They include parametric 3D lights [7, 17, 36, 37], outgoing light fields such as volumetric SGs [31] and neural out-of-view lighting volumes [45], and incident light fields like SV environment maps [47], SVSGs [16], and neural incident light fields (NeILF) [35, 38]. Our VMINer integrates distant lights as SGs and near-field lights as point lights with adjustable position and intensity. It is less complex compared to other SV lighting models, facilitating spatially consistent editing and optimization, while being able to approximate common lighting setups in real life." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 647, + 411, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 411, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 411, + 661 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "content": "As illustrated in Fig. 2, VMINer reconstructs the scene lighting (Sec. 3.1) and 3D fields (Sec. 3.2) from input images using differentiable rendering (Sec. 3.3). The training scheme and loss functions are described in Sec. 3.4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11802" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "type": "text", + "content": "3.1. Versatile Lighting Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 114 + ], + "type": "text", + "content": "This subsection models the lighting term in Eq. (1) and corresponds to the top-left part of Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 114, + 288, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 288, + 294 + ], + "type": "text", + "content": "VMINer harnesses the diversity of lighting conditions present in the input RGB images to disentangle lighting effects. The lighting model plays a crucial role here: 1) Since lighting conditions are unknown initially, they must be optimized alongside other scene parameters using physically-based rendering during training. 2) Different lighting setups, each compatible with the model, are necessary during the capture phase to achieve this goal. Given these prerequisites, it is essential to design a lighting model that accommodates common far-field and near-field light sources, while being lightweight enough for efficient and robust optimization and physically-based rendering. With these considerations, we propose a versatile lighting model that is both inclusive and lightweight. This representation models far-field and near-field lighting separately as below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": "Far-field Lighting. In our model, each far-field lighting condition " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " belongs to the set " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_{N_{\\mathrm{far}}} \\triangleq \\{1, \\dots, N_{\\mathrm{far}}\\}" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{far}}" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " denotes the number of far-field lighting, is represented using 128-lobe SGs [28] commonly adopted in existing far-field lighting techniques [33, 39, 42]. For each lobe " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " of lighting " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": ", there are six parameters: lobe axis " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}_{ij} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": ", lobe sharpness " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\lambda_{ij} \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": ", and lobe RGB amplitude " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{ij} \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": ". The incident radiance from far-field lighting " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " along direction " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " at any 3D position " + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 296, + 288, + 416 + ], + "type": "text", + "content": " is calculated using the formula (ignoring visibility for now):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 419, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 419, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 94, + 419, + 287, + 453 + ], + "type": "interline_equation", + "content": "L _ {\\operatorname {f a r} _ {i}} (\\boldsymbol {\\omega}) = \\sum_ {j = 1} ^ {1 2 8} c _ {i j} \\boldsymbol {\\mu} _ {i j} e ^ {\\lambda_ {i} (\\boldsymbol {\\omega} \\cdot \\boldsymbol {\\xi} _ {i j} - 1)}, \\tag {2}", + "image_path": "0169b96e5d3b5d585923692667d3f880b1227d86b7ea0fc45b09dcff0c156c61.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "inline_equation", + "content": "c_{ij} = \\lambda_{ij} / (2\\pi (1 - e^{-2\\lambda_{ij}}))" + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "text", + "content": " acts as the normalization factor related to roughness. Notably, " + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "text", + "content": " does not appear as an input to " + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{far}_i}" + }, + { + "bbox": [ + 47, + 458, + 287, + 495 + ], + "type": "text", + "content": ", reflecting the nature of distant lighting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": "Near-field Lighting. Each near-field lighting " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": ", included in the set " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{N}_{N_{\\mathrm{near}}}" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{near}}" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": " representing the number of near-field light sources, is modeled as a point light. These point lights can have moving positions " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": " and exhibit anisotropic radiation characterized by " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": "-order (" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": " can be 0, 1, or 2) SH coefficients " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i \\in \\mathbb{R}^{3 \\times (l + 1)^2}" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": ". Our method accommodates two types of near-field lighting: camera-collocated lights, positioned at the camera ray origin for each image, and stationary lights, which remain fixed across all images. We observe that in neural radiance field rendering, the radiance under a stationary near-field light, especially when the light is active in all images, can be challenging to distinguish from radiance under far-field lighting. Also, the radiance from these two sources aids in material estimation in a similar way. Therefore, in practical applications, we favor using a moving flashlight as the near-field light to provide unique information for material estimation. The incident radiance from near-field lighting " + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 497, + 288, + 714 + ], + "type": "text", + "content": " at a 3D po" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "content": "sition " + }, + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "content": " is computed as (also ignoring visibility for now):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 88, + 545, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 88, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 88, + 545, + 126 + ], + "type": "interline_equation", + "content": "L _ {\\text {n e a r} i} (\\mathbf {x}, \\boldsymbol {\\omega}, \\mathbf {p} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathrm {S H} (\\boldsymbol {\\omega} ; \\mathbf {h} _ {i})}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2} ^ {2}} & \\text {i f} \\boldsymbol {\\omega} = \\frac {\\mathbf {p} _ {i} - \\mathbf {x}}{\\| \\mathbf {p} _ {i} - \\mathbf {x} \\| _ {2}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}", + "image_path": "be46b358c933b808590712b6971ce1ee116c66c0af39549d5ae9f3d444f0e2a1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{SH}(\\omega ;\\mathbf{h})" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " calculates the SH at direction " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " with coefficients " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "1 / \\| \\mathbf{p}_i - \\mathbf{x}\\|_2^2" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " signifies the inverse-square lighting attenuation for point lights. Here we include " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " as an input because it may be set differently across images. It is important to note that with respect to the incident direction " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{far}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " is a continuous function, while " + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{near}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 262 + ], + "type": "text", + "content": " is a discrete function, being non-zero only in a single direction. As a result, to render appearances under direct lighting, Monte Carlo integration is essential for far-field lighting, while a simple multiplication suffices for near-field lighting. Further details about this process are provided in Sec. 3.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 266, + 522, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 522, + 279 + ], + "type": "text", + "content": "3.2. Geometry, Material, and Radiance Fields" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 309 + ], + "type": "text", + "content": "This subsection models the material and geometry terms in Eq. (1) and corresponds to the bottom-mid part of Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": "Geometry. Multi-view reconstruction methodologies generally hinge on two geometry representations: volume density [2, 10, 19, 41] and the signed distance function (SDF) [33, 39, 42, 44] fields. We choose the SDF field for our geometry representation due to its clearly defined surface at the zero-level isosurface, which simplifies and enhances the post-training mesh extraction process. Our approach utilizes implicit fields as MLPs over explicit structures like voxel grids for their compactness and flexibility. For each 3D position " + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": ", a multi-resolution hash grid [20] is first employed for positional encoding, yielding a feature vector " + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x}) \\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": ". The geometry MLP " + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": " then predicts the SDF value at " + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 355, + 469, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 469, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 355, + 469, + 545, + 483 + ], + "type": "interline_equation", + "content": "s (\\mathbf {x}), \\mathbf {z} (\\mathbf {x}) = \\mathrm {M} _ {\\mathrm {g e o}} (\\mathbf {x}, \\operatorname {e n c} _ {\\mathrm {g e o}} (\\mathbf {x})), \\tag {4}", + "image_path": "9e0f95db09ee0f28e9661ef6dd950b1dc136d34b2f164712e36bed0d12db4435.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": " represents the signed distance (positive outside the surface, negative inside), and " + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{13}" + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": " is a descriptor of local appearance. To maintain differentiability in rendering and to aid in geometry optimization, during training we apply NeuS [29] techniques to transform SDF values along rays into volume densities. The surface normal " + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": " is derived as the gradient of the SDF " + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 487, + 547, + 571 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 573, + 545, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 573, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 376, + 573, + 545, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) = \\frac {\\partial s}{\\partial \\mathbf {x}} / \\left\\| \\frac {\\partial s}{\\partial \\mathbf {x}} \\right\\| _ {2}. \\tag {5}", + "image_path": "fa1ccf1bcf54a2cedbc2360ddaa12f2f886fbd7c99c131bebc5232755db1e87b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "text", + "content": " only approximates a strict SDF field that has " + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\left\\| \\frac{\\partial s}{\\partial \\mathbf{x}} \\right\\|_2 = 1" + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "text", + "content": ", so we normalize " + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 305, + 605, + 545, + 628 + ], + "type": "text", + "content": " to ensure it is a unit vector." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Material. For accurate scene material recovery using physically-based re-rendering, it is imperative to model the material in such a way that the renderer can query the BRDF at any point on the surface. Considering that material properties are independent of lighting and view directions, we represent the scene's material through an implicit 3D field, much like the SDF field utilized for geometry. To achieve" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11803" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "this, we employ another multi-resolution hash grid " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\mathrm{mat}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " for positional encoding. We adopt the simplified GGX BRDF model [27] with a fixed fresnel parameter to help alleviate ambiguity. The material MLP, " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{mat}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ", is tasked with predicting the SVBRDF parameters " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\beta (\\mathbf{x})" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 137, + 287, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 137, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 94, + 137, + 287, + 150 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\beta} (\\mathbf {x}) = \\operatorname {M} _ {\\text {m a t}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {m a t}} (\\mathbf {x})). \\tag {6}", + "image_path": "1f9b2b2c1766d57cd4270364e33b70a5286e6d104e21afd977054a611a41b342.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": "Radiance. Although our method does not strictly require a radiance field for rendering scene appearance - given that appearance could be rendered solely using physically-based surface rendering (PBR) - we find that incorporating a radiance field at the start of training significantly eases the optimization of scene geometry and appearance. PBR, while physically accurate, tends to introduce instability and slow down the training process due to its inherent ambiguity and computational intensity. Moreover, the radiance field-rendered results can serve as supplementary supervision, aiding PBR in better separating scene appearance under each lighting condition and thus further diminishing lighting-material ambiguity. Hence, we deploy radiance MLPs to predict view-dependent and lighting-dependent radiance at each position. For each lighting condition " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": ", a light embedding " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\{\\mathrm{far,near}\\}_i} \\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " is utilized. The view direction " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " is processed through directional encoding that projects it onto the coefficients of the " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": "-order SH basis, yielding " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\mathrm{dir}}(\\mathbf{v}) \\in \\mathbb{R}^{4^2 = 16}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " [20]. The far-field radiance MLP " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{far}}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " computes the outgoing radiance " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{far}_i}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " from position " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " along view direction " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": " under far-field lighting " + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 154, + 289, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 412, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 412, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 75, + 412, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\text {f a r} _ {i}} (\\mathbf {x}, \\mathbf {v}) = \\mathrm {M} _ {\\text {f a r}} (\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\operatorname {e n c} _ {\\text {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\text {f a r} _ {i}}), \\tag {7}", + "image_path": "3d9ecab91717b57f5fbed0326ff454932faf517189e0133660d16c7cd41867e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 464 + ], + "type": "text", + "content": "where we incorporate the surface normal " + }, + { + "bbox": [ + 46, + 430, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 430, + 287, + 464 + ], + "type": "text", + "content": " as an additional MLP input, following Instant-NSR [44] and WildLight [6], as this has shown to aid shape recovery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": "The near-field radiance MLP differs slightly, as nearfield lighting can move in different images, thus the current light position " + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": " is a necessary input. Additionally, the radiance under near-field point light takes the form of the rendering equation Eq. (1) without the integral, allowing the incoming radiance and cosine term to be explicitly included. We follow ReNe [25] to input the relative direction of the point light " + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\omega_{i} = \\frac{\\mathbf{p}_{i} - \\mathbf{x}}{\\|\\mathbf{p}_{i} - \\mathbf{x}\\|_{2}}" + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": " instead of the absolute position " + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": " into the MLP. Therefore, the outgoing neural radiance " + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{near}_i}" + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": " under near-field lighting " + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 465, + 287, + 586 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 590, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 52, + 590, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {c} _ {\\mathrm {n e a r} _ {i}} (\\mathbf {x}, \\mathbf {v}, \\mathbf {p} _ {i}) = \\mathrm {M} _ {\\mathrm {n e a r}} \\left(\\mathbf {x}, \\mathbf {z}, \\mathbf {n}, \\mathrm {e n c} _ {\\mathrm {d i r}} (\\mathbf {v}), \\mathbf {F} _ {\\mathrm {n e a r} _ {i}}, \\boldsymbol {\\omega} _ {i}\\right) \\\\ \\otimes \\frac {L _ {\\text {n e a r} _ {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {i})}{\\left\\| \\mathbf {p} _ {i} - \\mathbf {x} \\right\\| _ {2} ^ {2}} \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {8} \\\\ \\end{array}", + "image_path": "02c09aedf104dc80194ccf30085f7a1fc0a7f5f1f9fb5030efa2cde17f775cbe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "text", + "content": " denotes element-wise multiplication. On the right side, the second term denotes the incident radiance attenuated by square distance, the third term denotes the cosine weakening factor, and " + }, + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{near}}" + }, + { + "bbox": [ + 46, + 637, + 287, + 708 + ], + "type": "text", + "content": " accounts for other factors like reflectance and visibility. Owing to the instability of these terms at the training's outset, we use an annealing strategy," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "as recommended in prior works [44], gradually replacing default values (e.g., 1 for incident radiance) with those optimized during training, to ease through the process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 123, + 446, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 123, + 446, + 136 + ], + "spans": [ + { + "bbox": [ + 305, + 123, + 446, + 136 + ], + "type": "text", + "content": "3.3. Differentiable Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "text", + "content": "To simplify the notation, we omit " + }, + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "text", + "content": " and the lighting index " + }, + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 140, + 545, + 261 + ], + "type": "text", + "content": " in the equations within this subsection. Our approach optimizes the lighting alongside the implicit fields primarily by reducing the discrepancy between the observed and differentiably re-rendered appearances. The re-rendering is accomplished using two types of renderer: a neural radiance field renderer (Fig. 2 bottom-right) and a physically-based surface renderer (Fig. 2 top-right; Eq. (1)). We handle each lighting condition separately in rendering, and then aggregate the results based on per-image lighting condition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": "Volume Rendering for SDF Field. Consider a camera ray " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": "), with " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": " as its origin and " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": " as its direction. We incorporate importance sampling using an occupancy grid [14] to get " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "N \\leq 1024" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": " points along this ray, denoted as " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t_i)" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "i \\in \\mathbb{N}_N" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": ". Utilizing techniques from NeuS [29], we transform the signed distances " + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "inline_equation", + "content": "s(\\mathbf{r}(t_i))" + }, + { + "bbox": [ + 304, + 266, + 545, + 350 + ], + "type": "text", + "content": " into discrete opacity values as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 354, + 545, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 354, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 336, + 354, + 545, + 383 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\left(\\frac {\\Phi_ {b} (s (\\mathbf {r} (t _ {i}))) - \\Phi_ {b} (s (\\mathbf {r} (t _ {i + 1})))}{\\Phi_ {b} (s (\\mathbf {r} (t _ {i})))}\\right) ^ {+}, \\tag {9}", + "image_path": "f5aefd1f761e75f05f92ef62a583bdb26c90a9a32011e8f15f90fbea08d896c1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\Phi_b(s) = 1 / (1 + e^{-bs})" + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "content": " is the cumulative opacity distribution function. The parameter " + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "content": " is trainable and tends to increase during training, focusing opacity more narrowly around the surface where " + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "content": ". The accumulated transmittance is then calculated as " + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 304, + 387, + 545, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 453, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 501 + ], + "type": "text", + "content": "Neural Radiance Field Rendering. The color " + }, + { + "bbox": [ + 304, + 453, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf}}" + }, + { + "bbox": [ + 304, + 453, + 545, + 501 + ], + "type": "text", + "content": " of a camera ray under either far- or near-field lighting conditions, as rendered through the neural radiance field, is determined by accumulating the radiance along the ray's path:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 505, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 505, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r , n e a r} \\}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {\\{\\text {f a r , n e a r} \\}} (\\mathbf {r} (t _ {i}), - \\mathbf {d}). \\tag {10}", + "image_path": "a09b7397acee2a81973cf6fd7877aec067dc4570120f7a731b2a88c6ced8a1c0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 564 + ], + "type": "text", + "content": "This formula sums the radiance from each sampled point to yield the color of the ray under the specified lighting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Physically-based Rendering. Our method employs differentiable surface rendering, modeling the observed radiance as reflected from a single surface point. We compute the depth " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " of this surface point similarly to Eq. (10): " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = \\sum_{i=1}^{N} T_i \\alpha_i t_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Accumulation along the ray also applies to the BRDF parameters " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". For each input lighting, we render the appearance that models secondary visibility and indirect illumination. The rendering method under direct illumination differs for far- and near-field lighting. For direct far-field lighting " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dfar}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", we employ Monte Carlo integration to evaluate the rendering equation (Eq. (1)) as the incident light comes from all" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11804" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 92, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 92, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 92, + 83 + ], + "type": "text", + "content": "directions:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 86, + 287, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 86, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 52, + 86, + 287, + 134 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {C} _ {\\mathrm {p b , d f a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {o}} (\\mathbf {r} (t), - \\mathbf {d}) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\frac {L _ {\\mathrm {d f a r}} (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s}) f (\\mathbf {r} (t) , \\boldsymbol {\\omega} _ {s} , - \\mathbf {d}) \\left(\\boldsymbol {\\omega} _ {s} \\cdot \\mathbf {n}\\right) ^ {+}}{p \\left(\\boldsymbol {\\omega} _ {s}\\right)}, \\tag {11} \\\\ \\end{array}", + "image_path": "b3178f9bfdff214e6e0db2de99f76b11114021929f631706773626489af371ea.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "content": " is the number of sampled directions, " + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\omega_{s}" + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "content": " a sampled incident direction, and " + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "inline_equation", + "content": "p(\\omega_{s})" + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "content": " the probability density function (PDF) for the sampled direction. We use multiple importance sampling [26] to combine evaluations from different strategies: BRDF importance sampling [13], SG lighting importance sampling [33], and cosine importance sampling. we follow differentiable rendering works [15] to use a small " + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "inline_equation", + "content": "S = 20" + }, + { + "bbox": [ + 46, + 137, + 287, + 258 + ], + "type": "text", + "content": ", as stochastic gradient descent handles noisy gradients. The incident radiance, now taking the visibility of light sources into account, is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 262, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 262, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 84, + 262, + 287, + 275 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d f a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}) = L _ {\\mathrm {f a r}} (\\boldsymbol {\\omega} _ {s}) V (\\mathbf {r} (t), \\boldsymbol {\\omega} _ {s}), \\tag {12}", + "image_path": "c3c543ca3d4f68a8bf42c36570f4124803f142c7da078c6bf4471d0d45ce388a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "V(\\mathbf{r}(t), \\boldsymbol{\\omega}_s) \\in [0,1]" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": " representing visibility, or the inverse of the accumulated opacity along a secondary ray " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{sec}}(t') = \\mathbf{r}(t) + t' \\boldsymbol{\\omega}_s" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": ", (" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "t' > 0" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": "), computed in a manner akin to Eq. (10). For direct near-field lighting " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dnear}}" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": ", the integral in Eq. (1) simplifies to a multiplication, as the light originates from a single direction:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 355, + 287, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 287, + 379 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d n e a r}} (\\mathbf {o}, \\mathbf {d}) = L _ {\\mathrm {d n e a r}} (\\mathbf {r} (t), \\boldsymbol {\\omega}) f (\\mathbf {r} (t), \\boldsymbol {\\omega}, - \\mathbf {d}) \\left(\\boldsymbol {\\omega} \\cdot \\mathbf {n}\\right) ^ {+}, \\tag {13}", + "image_path": "8ea59a95788b4bcb108b7f30a7af3993ad53ebfead2d1f85261ebfed51bec0b0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\omega = (\\mathbf{p} - \\mathbf{r}(t)) / \\| \\mathbf{p} - \\mathbf{r}(t)\\|_{2}" + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": " is the direction of incident light, and " + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dnear}}(\\mathbf{r}(t),\\omega)" + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": " is the potentially occluded incident radiance, computed similarly to Eq. (12)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": "Indirect illumination " + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": " relevant to both far-field and near-field lighting, considers radiance reflected from the scene itself, potentially from all directions, thus necessitating Monte Carlo integration. The neural radiance field " + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": " substitutes multi-bounce path tracing for indirect illumination, given its encapsulation of the scene radiance under infinite lighting bounces. The radiance " + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{pb,ind}\\{\\mathrm{far,near}\\}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": " from indirect illumination is evaluated similarly to Eq. (12), but with the " + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dfar}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": " replaced by the indirect lighting " + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}\\{\\mathrm{far,near}\\}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 537 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 540, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 540, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 540, + 287, + 554 + ], + "type": "interline_equation", + "content": "L _ {\\text {i n d} *} (\\mathbf {r} (t), \\omega_ {s}) = \\mathbf {c} _ {*} \\left(\\mathbf {x} ^ {\\prime}, - \\omega_ {s}\\right) \\left(1 - V (\\mathbf {r} (t), \\omega_ {s})\\right), \\tag {14}", + "image_path": "e9660dc9429505265e1ec588473931ddd2c27e9ab72176392648e86684732914.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "content": " is the intersection of the secondary ray " + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{sec}}(t')" + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "content": " with the scene geometry. The radiance from indirect illumination is also computed for each lighting condition separately, then combined with direct illumination to yield the complete PBR radiance for the camera ray under a specific lighting condition:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 634, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 634, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 52, + 634, + 287, + 648 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r , n e a r} \\}} = \\mathbf {C} _ {\\mathrm {p b}, \\mathrm {d} \\{\\text {f a r , n e a r} \\}} + \\mathbf {C} _ {\\mathrm {p b}, \\text {i n d} \\{\\text {f a r , n e a r} \\}}. \\tag {15}", + "image_path": "8fed0248104ab0e63ab3d40f61fee1f3dc93f3093923d6191baf9f91423322f4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 111, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 111, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 111, + 672 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Training Schemes. Our training process is divided into two sequential stages. In the first stage, the focus is on training the geometry and radiance field without employ" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "ing PBR. The objectives are: 1) to recover the scene's geometry, 2) to distinguish the appearance under each lighting condition, a necessity due to the potential presence of multiple light sources in one image, and 3) to utilize the reconstructed radiance field for indirect illumination. In the second stage, we train the material field and scene lighting using PBR, aiming to estimate material properties." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "text", + "content": "Loss Functions. We average the loss functions across batches of camera rays. During the first stage, our loss function compares the total re-rendered radiance under all lighting conditions present in an image – including a far-field light " + }, + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "text", + "content": " and any active near-field lights – with the ground-truth (GT) color " + }, + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 159, + 545, + 243 + ], + "type": "text", + "content": " from the input observation. This comparison trains the geometry and radiance field:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 247, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 247, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 320, + 247, + 545, + 273 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r f}} = \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\text {f a r} _ {i}} + \\sum_ {\\text {n e a r} i \\text {i s o n}} \\mathbf {C} _ {\\mathrm {r f}, \\text {n e a r} _ {i}} - \\mathbf {C} _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}. \\tag {16}", + "image_path": "8f6bc4e5a8cdad2cb1bd6996b7e2ba6c09243437d7469be316f6bc8cce47624f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 275, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 300 + ], + "type": "text", + "content": "The Eikonal loss [8] is applied to the gradients of the SDF values " + }, + { + "bbox": [ + 304, + 275, + 545, + 300 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 275, + 545, + 300 + ], + "type": "text", + "content": " for geometric regularization:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 304, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 340, + 304, + 545, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e i k}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\left\\| \\frac {\\partial s (\\mathbf {r} (t _ {j}))}{\\partial \\mathbf {r} (t _ {j})} \\right\\| _ {2} - 1\\right) ^ {2}, \\tag {17}", + "image_path": "0e2ae8f583e439ea8aca989b310709d3b7a2732806316b20ec6744294e36f5ee.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 341, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 365 + ], + "type": "text", + "content": "A silhouette loss using GT alpha " + }, + { + "bbox": [ + 304, + 341, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 341, + 545, + 365 + ], + "type": "text", + "content": " from the foreground mask and a normal smoothness loss [41] aid shape recovery:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 370, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 370, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 325, + 370, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s i l}} = \\left(\\left(1 - T _ {N + 1}\\right) - \\alpha_ {\\mathrm {g t}}\\right) ^ {2}. \\tag {18}", + "image_path": "6eb9fde22a1a686da8b8e1b9a14273843840128a1e5f76e6527addeb2ce400af.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 387, + 545, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 325, + 387, + 545, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n s}} = \\sum_ {j = 1} ^ {N} T _ {j} \\alpha_ {j} \\left(\\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right)\\right) - \\mathbf {n} \\left(\\mathbf {r} \\left(t _ {j}\\right) + \\boldsymbol {\\epsilon}\\right)\\right) ^ {2}, \\tag {19}", + "image_path": "74a4803bbfc35a8dbb4d286e85d6e0edc906d2b9b422dcefd38e9cfdd62efdf8.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " is a small random perturbation. The total loss for stage one is " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rf}} + \\lambda_{\\mathrm{eik}}\\mathcal{L}_{\\mathrm{eik}} + \\lambda_{\\mathrm{sil}}\\mathcal{L}_{\\mathrm{sil}} + \\lambda_{\\mathrm{ns}}\\mathcal{L}_{\\mathrm{ns}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\lambda_*" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " representing the loss weights. In stage two, PBR colors replace radiance field-rendered colors for comparison with " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": ", and a novel self-consistency loss is introduced between neural and PBR radiance under each lighting for additional supervision. The direct supervision loss " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pb}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " mirrors " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rf}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{rf},\\{\\mathrm{far,near}\\}_i}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " replaced by " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathrm{pb},\\{\\mathrm{far,near}\\}_i}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": ". The perlighting self-consistency loss " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 533, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 545, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n}} = \\sum_ {i = 1} ^ {N _ {\\{\\text {f a r}, \\text {n e a r} \\}}} \\left\\| \\mathbf {C} _ {\\mathrm {r f}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} - \\mathbf {C} _ {\\mathrm {p b}, \\{\\text {f a r}, \\text {n e a r} \\} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {20}", + "image_path": "772e00abdc88d15ea5a0ccc4845cabac9ba8b70a41e6745f4c7372e0661cebe4.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "text", + "content": "This loss is crucial for discerning contributions from each lighting in images with multiple active light sources, thereby reducing ambiguity and improving material estimation. A material smoothness loss " + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ms}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "text", + "content": " is used similar to " + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ns}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "text", + "content": ". The total loss for stage two is " + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pb}} + \\lambda_{\\mathrm{con}}\\mathcal{L}_{\\mathrm{con}} + \\lambda_{\\mathrm{ms}}\\mathcal{L}_{\\mathrm{ms}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Our model is trained on an NVIDIA RTX 3090 GPU for a total of 40,000 steps. The first 30,000 steps of stage one take about 20 minutes, while the subsequent 10,000 steps of stage two require about 40 minutes. Post-training, we follow a procedure similar to WildLight [6] to extract the fields into textured meshes, which can be easily integrated into industry-standard rendering software like Blender [1]," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11805" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 106, + 544, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Quantitative comparison results with state-of-the-art methods averaged on 6 synthetic scenes. We show results of surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, the specular reflection part of FV relit RGB, and training time on a single RTX 3090 GPU. We mark the best and the second best results in each column. " + }, + { + "bbox": [ + 47, + 72, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 47, + 72, + 547, + 105 + ], + "type": "text", + "content": " means bigger (smaller) is better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 106, + 544, + 244 + ], + "lines": [ + { + "bbox": [ + 49, + 106, + 544, + 244 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 544, + 244 + ], + "type": "table", + "html": "
MethodInput lighting conditionsNormal MAngE↓AlbedoView synthesisFV relitFV relit (spec)Time
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
(1) TensoIR [10]17.6626.480.92129.480.91228.180.90128.300.861300 mins
(2) NVDiffRecMC [9]Single far-field16.2426.520.91527.130.91326.610.90125.570.836150 mins
(3) Ours12.3924.500.88228.200.93427.460.92127.560.87145 mins
(4) WildLight [6]Single far-field + Flashlight11.4928.860.94029.870.92930.440.93027.710.8631440 mins
(5) Ours10.8931.620.95332.090.95332.000.95330.750.90660 mins
(6) TensoIR [10]Two far-field16.2427.180.92929.680.91228.660.90228.460.863300 mins
(7) Ours11.7026.070.90229.590.94229.270.93429.090.89045 mins
(8) OursTwo far + Single near10.7932.040.95732.100.95032.380.95431.400.91060 mins
", + "image_path": "ac4b766d47d1ca8f325230f955f1f0236da4f91fda954960aa459b98c77b5209.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 251, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 288, + 275 + ], + "type": "text", + "content": "enabling fast and high-quality rendering suitable for various applications. Check the supplement for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 128, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 128, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 128, + 298 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 304, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 287, + 365 + ], + "type": "text", + "content": "Our experiments involve a comprehensive comparison with state-of-the-art methods (Sec. 4.2), ablation studies (Sec. 4.3), and evaluations on both synthetic datasets (Sec. 4.1) and real-world images (Sec. 4.4). The supplement shows further data creation details and more experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 371, + 156, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 156, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 156, + 384 + ], + "type": "text", + "content": "4.1. Synthetic Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 605 + ], + "type": "text", + "content": "We collect 6 synthetic scenes comprising a variety of scene geometries and materials. For training, each scene is rendered under four distinct lighting setups: 1) single far-field light, 2) single far-field light with a camera-collocated flashlight, 3) two far-field lights, and 4) two far-field lights with a near-field light source. We render 100 training images per setting using random viewpoints and per-image lighting conditions. The far-field lighting from the first setting is used in subsequent settings, with the last two sharing their far-field lighting. For testing, 200 images under one common far-field lighting are rendered to assess novel view synthesis under seen lighting, and another 200 under unseen far-field lighting to evaluate free-viewpoint (FV) relighting, which largely depends on the geometry and material estimation quality. This methodology provides a relatively fair comparison across methods with different abilities regarding supported input lighting conditions, trained under similar conditions and tested all using the same sets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 270, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 270, + 624 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 715 + ], + "type": "text", + "content": "We compare our method with leading multi-view inverse rendering methods under each input lighting scenario. For single far-field lighting inputs (Tab. 1 row 1), we compare with NVDiffRecMC [9] and TensoIR [10] (single-lighting input). For multiple far-field lighting scenarios (Tab. 1 row 3), we compare with TensoIR [10]. For setups involving a far-field light and a flashlight (Tab. 1 row 4), we compare" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "type": "text", + "content": "with WildLight [6]. We use Blender [1] for high-quality relit results for our method, WildLight [6], and NVDiffRecMC [9], importing trained SDF and material fields as textured meshes. Since TensoIR [10] employs density-based geometry and cannot extract high-quality meshes, we utilize its differentiable renderer for relighting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 453 + ], + "type": "text", + "content": "Our quantitative comparison employs mean angular error (MAngE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) [30] metrics for surface normal, diffuse albedo, view synthesis RGB, free-viewpoint (FV) relit RGB, and specular reflection of FV relit RGB. We follow NeRFactor [41] in assuming albedo and lighting brightness as scale-invariant, necessitating metric comparisons using corresponding scales. For visualization and metrics, we scale each RGB channel of albedo and relit images from all methods by a global scale to minimize mean squared error against ground truth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 632 + ], + "type": "text", + "content": "The quantitative results are detailed in Tab. 2. We observe that when only input images under single far-field lighting are available (rows 1-3), our method gives comparable results as TensoIR [10] and NVDiffRecMC [9]. Adding a flashlight (rows 4, 5) significantly helps in geometry and material estimation (comparing rows 3, 5), where the greatest improvement is made regarding the diffuse albedo. Also, our method clearly surpasses WildLight [6] in both quality and speed in this setting. Leveraging another far-field lighting condition (rows 6, 7) brings similar effects as using a flashlight, but generally with a lower degree of improvement. Further adding a flashlight (row 8) gives the best results among all settings and methods, but the gain is marginal compared to the setting of single far-field lighting with a flashlight (comparing rows 5, 8)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 633, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 547, + 706 + ], + "type": "text", + "content": "Fig. 3 shows the qualitative comparison on two synthetic scenes: LEGO (left) with detailed geometry, and TROOPER (right) with highly-reflective SV materials. The results show that our method predicts more detailed and faithful shape, diffuse albedo, and specular parameters that could produce accurate and realistic relit results under unseen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11806" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 302, + 277 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 302, + 277 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 302, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 302, + 277 + ], + "type": "image", + "image_path": "b2fb9ddc485f21d9ae4368406011da2716409663ac6d6e3046cbad3fe42965c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "type": "text", + "content": "Figure 3. Comparison with state-of-the-art methods. We test NVDiffRecMC [9] (single far-field lighting in input), TensoIR [10] (two far-field lighting in input), and WildLight [6] and our method (single far-field lighting with a flashlight in input) on two synthetic scenes. The intensity of the specular reflection is multiplied by " + }, + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "type": "text", + "content": " for the visualization purpose. Please zoom in for details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 542, + 276 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 542, + 276 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 542, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 542, + 276 + ], + "type": "image", + "image_path": "ee09e28916cb666d5eed9b22d6bf9faf6bea8a608629c3c7f21f6da5d7825cd0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 341, + 284, + 419 + ], + "blocks": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "type": "text", + "content": "Table 3. Ablation study on our method using inputs under single far-field lighting with a flashlight." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 341, + 284, + 419 + ], + "lines": [ + { + "bbox": [ + 52, + 341, + 284, + 419 + ], + "spans": [ + { + "bbox": [ + 52, + 341, + 284, + 419 + ], + "type": "table", + "html": "
ModelNormalAlbedoFV RelitRelit (spec)
MAngE↓PSNR↑PSNR↑PSNR↑
(1) w/o Lcon8.45227.7029.8529.08
(2) Mod. cnear8.57227.7930.0328.87
(3) w/o encmat8.26327.7930.0728.92
(4) Full model8.29828.0530.1929.36
", + "image_path": "417858a2c896cfcfa89dcd3d9f8c12898f6afdae47db8df7a6c7db8a9ba88a1b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 470 + ], + "type": "text", + "content": "lighting. In contrast, WildLight [6] tends to overly smooth geometry and material. TensoIR [10] can not handle highly-reflective surfaces, and lighting is baked into its predicted diffuse albedo for TROOPER." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 147, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 147, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 147, + 487 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "content": "In Tab. 3, we quantitatively evaluate ablation models on a subset of our synthetic scenes: 1) the model excluding the self-consistency loss " + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "content": ", 2) the model where the output of the near-field radiance MLP is not explicitly multiplied by the cosine weakening factor and incident radiance (Eq. (8)), 3) the model with " + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\mathrm{mat}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "content": " replaced by " + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{enc}_{\\mathrm{geo}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "content": ". The results show that all the above techniques enhance the geometry and materials estimated by our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 160, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 160, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 160, + 607 + ], + "type": "text", + "content": "4.4. Real-world Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "type": "text", + "content": "Fig. 4 shows the results of our method on two real-world scenes captured under single far-field lighting and a flashlight. Our method faithfully recovers the geometry and material of the objects and produces realistic relit results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We introduce a versatile multi-view inverse rendering framework, distinguished by its ability to leverage input im" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 316, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 316, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 407 + ], + "type": "image", + "image_path": "cbf612ad1b7d69f94b9168c7654cd04c1bf2429980067f248be7b32e4ec3c024.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 408, + 545, + 500 + ], + "blocks": [ + { + "bbox": [ + 307, + 408, + 545, + 500 + ], + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 500 + ], + "type": "image", + "image_path": "2170a37d6d11c10633e96c8b249bafbbbe6b0a96d45ffda7305aa22d64cfe820.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 502, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 535 + ], + "type": "text", + "content": "Figure 4. The results of our method on two real-world scenes, using multi-view images under one ambient lighting and a flashlight. The specular reflection intensity is increased for visualization." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "content": "ages under varied far- and near-field light sources available in capture for better geometry and material estimation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "content": "Limitations. VMINer does not model the background, necessitating a foreground mask for each image. It also does not consider unknown tone-mapping curves applied during the image signal processing (ISP) stage, using a fixed curve with " + }, + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\gamma = 2.2" + }, + { + "bbox": [ + 304, + 566, + 545, + 662 + ], + "type": "text", + "content": " to the computed linear radiance. It depends on distinguishing the contribution from each lighting, which implies that, for optimal results, near-field light sources should be switched on/off during capture." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 674, + 403, + 688 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 403, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 403, + 688 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grand No. 62136001, 62088102)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11807" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Blender Foundation. The Blender project - free and open 3D creation software. https://www.blender.org. Accessed: 2023-11-07. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 179 + ], + "type": "text", + "content": "[2] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural reflectance decomposition from image collections. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "text", + "content": "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. NeuralPIL: Neural pre-integrated lighting for reflectance decomposition. In Proc. of Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "text", + "content": "[4] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P. A. Lensch, and Varun Jampani. SAMURAI: shape and material from unconstrained real-world arbitrary image collections. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 301, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 287, + 323 + ], + "type": "text", + "content": "[5] Brent Burley. Physically-based shading at disney. In SIGGRAPH 2012 Courses, 2012. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 369 + ], + "type": "text", + "content": "[6] Ziang Cheng, Junxuan Li, and Hongdong Li. WildLight: In-the-wild inverse rendering with a flashlight. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "text", + "content": "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proc. of International Conference on Machine Learning (ICML), 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "type": "text", + "content": "[9] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using Monte Carlo rendering and denoising. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[10] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[11] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[12] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG), 41(4):56:1-56:12, 2022. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "text", + "content": "[13] Eric P Lafortune and Yves D Willems. Using the modified phong reflectance model for physically based rendering. Katholieke Universiteit Leuven. Departement Computerwetenschappen, 1994. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates NeRFs. arXiv preprint arXiv:2305.04966, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[15] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable Monte Carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6): 222, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 216 + ], + "type": "text", + "content": "[16] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[17] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Milos Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In Proc. of European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 1, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "text", + "content": "[21] Jacob Munkberg, Wenzheng Chen, Jon Hasselgren, Alex Evans, Tianchang Shen, Thomas Müller, Jun Gao, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 536 + ], + "type": "text", + "content": "[22] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2001. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[23] Viktor Rudnev, Mohamed Elgharib, William A. P. Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 646 + ], + "type": "text", + "content": "[24] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "text", + "content": "[25] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. ReLight My NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11808" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[26] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[27] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. In Proc. of the Eurographics Symposium on Rendering Techniques, 2007. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "text", + "content": "[28] Jiaping Wang, Peiran Ren, Minmin Gong, John M. Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially-varying reflectance. ACM Transactions on Graphics (TOG), 28(5):133, 2009. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proc. of Neural Information Processing Systems (NeurIPS), 2021. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "text", + "content": "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[31] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "type": "text", + "content": "[32] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[33] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[34] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[36] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(8):10129-10142, 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "text", + "content": "[37] Chong Zeng, Guojun Chen, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Relighting neural radiance fields with shadow and highlight hints. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[38] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 632 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Inter-reflectable light fields for geometry and material estimation. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[40] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: inverse rendering by optimizing neural SDFs and materials from photometric images. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "type": "text", + "content": "[41] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "text", + "content": "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "text", + "content": "[43] Youjia Zhang, Teng Xu, Junqing Yu, Yuteng Ye, Yanqing Jing, Junle Wang, Jingyi Yu, and Wei Yang. NeMF: Inverse volume rendering with neural microflake field. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "text", + "content": "[44] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG), 41(6):235:1-235:17, 2022. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 441, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 518 + ], + "type": "text", + "content": "[45] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable Monte Carlo raytracing. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 519, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 586 + ], + "type": "text", + "content": "[46] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Di-anbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I" + }, + { + "bbox": [ + 307, + 519, + 545, + 586 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 307, + 519, + 545, + 586 + ], + "type": "text", + "content": "-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "text", + "content": "[47] Yongjie Zhu, Yinda Zhang, Si Li, and Boxin Shi. Spatially-varying outdoor lighting estimation from intrinsics. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11809" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_content_list.json b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02aa742455c25174310701897c85960256f5864c --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "VOODOO 3D: Volumetric Portrait Disentanglement for One-Shot 3D Head Reenactment", + "text_level": 1, + "bbox": [ + 202, + 130, + 769, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phong Tran1 Egor Zakharov2 Long-Nhat Ho1 Anh Tuan Tran3 Liwen Hu4 Hao Li1,4 \n1MBZUAI ETH Zurich 2VinAI Research 3Pinscreen \n{the.tran, long.ho}@mbzuai.ac.ae anhtt152@vinai.io ezakharov@ethz.ch liwen@pinsscreen.com hao@hao-li.com", + "bbox": [ + 107, + 202, + 861, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6c0e3214c7b78dceacfeb8013d7aef20ceb7d28032ac84ae7eb85de77de19ba.jpg", + "image_caption": [ + "Figure 1. We introduce VOODOO 3D: a high-fidelity 3D-aware one-shot head reenactment technique. Our method transfers the expression of a driver to a source and produces view consistent renderings for holographic displays." + ], + "image_footnote": [], + "bbox": [ + 117, + 309, + 849, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 590, + 313, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a 3D-aware one-shot head reenactment method based on a fully volumetric neural disentanglement framework for source appearance and driver expressions. Our method is real-time and produces high-fidelity and view-consistent output, suitable for 3D teleconferencing systems based on holographic displays. Existing cutting-edge 3D-aware reenactment methods often use neural radiance fields or 3D meshes to produce view-consistent appearance encoding, but, at the same time, they rely on linear face models, such as 3DMM, to achieve its disentanglement with facial expressions. As a result, their reenactment results often exhibit identity leakage from the driver or have unnatural expressions. To address these problems, we propose a neural self-supervised disentanglement approach that lifts both the source image and driver video frame into a shared 3D volumetric representation based on tri-planes. This representation can then be freely manipulated with expression tri-planes extracted from the driving images and rendered from an arbitrary view using neural radiance fields. We achieve this disentanglement via self", + "bbox": [ + 73, + 621, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervised learning on a large in-the-wild video dataset. We further introduce a highly effective fine-tuning approach to improve the generalizability of the 3D lifting using the same real-world data. We demonstrate state-of-the-art performance on a wide range of datasets, and also showcase high-quality 3D-aware head reenactment on highly challenging and diverse subjects, including non-frontal head poses and complex expressions for both source and driver.", + "bbox": [ + 496, + 592, + 893, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 737, + 632, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating 3D head avatars from a single photo is a core capability in making a wide range of consumer AR/VR and telepresence applications more accessible, and user experiences more engaging. Graphics engine-based 3D avatar digitization methods [9, 14, 33, 37, 46, 47, 49, 57] are suitable for today's video games and virtual worlds, and many commercial solutions exist (AvatarNeo [5], AvatarSDK [1], ReadyPlayerMe [6], in3D [2], etc.). However, the photorealism achieved by modern neural head reenactment tech-", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10336", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "niques is becoming increasingly appealing for advanced effects in video sharing apps and visual effects. For immersive telepresence systems that use AR/VR headsets, facial expression capture is typically achieved using tiny video cameras built into HMDs [29, 50, 55, 64, 69], while the identity of the source subject recorded using a separate process. However, the teleconferencing solutions based on holographic 3D displays (LookingGlass [4], LEIA [3], etc.) use regular webcams [81] or depth sensors [48]. As opposed to a video-based setting, head reenactment for immersive applications needs to be 3D-aware, meaning that in addition to generating the correct poses and expressions from a photo, multi-view consistency is critical.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While impressive facial reenactments results have been demonstrated using 2D approaches [26, 27, 83, 97, 102, 103], they typically struggle with preserving the likeness of the source and exhibit significant identity changes when varying the camera pose. More recently, 3D-aware one-shot head reenactment methods [36, 43, 53, 54, 59, 99] have used either 3D meshes or tri-plane neural radiance fields as a fast and memory efficient volumetric data representations for neural rendering. However, the expression and identity disentanglement in these methods is based on variants of linear face and expression models [15, 52] which lack expressiveness and high-frequency details. While these methods can achieve view consistency, facial expressions are often uncanny, and preserving the likeness of the input source portrait is challenging, especially for views different than the source image. Hence, input sources with extreme expressions and non-frontal poses are often avoided.", + "bbox": [ + 75, + 289, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce the first 3D aware one-shot head reenactment technique that disentangles source identities and the target expressions fully volumetrically, and without the use of explicit linear face models. Our method is real-time and designed with holographic displays in mind, where a large number of views (up to 45) can be rendered in parallel based on their viewing angle. We leverage the fact that real-time 3D lifting for human heads has recently been made possible [81] with the help of Vision Transformers (ViT) [25], which avoids the need for inefficient optimization-based GAN-inversion process [68]. In particular, 3D lifting allows us to map 2D face images into a canonical tri-plane representation for both source and target subjects and treat identity and expression disentanglement independently from the head pose.", + "bbox": [ + 75, + 550, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once the source image and driver frame are lifted into a pose-normalized tri-plane representation, we extract appearance features from the source subject and expressions from the driver. The pose of the driver is estimated separately using a 3D face tracker and used as input to a neural renderer. Tri-plane-based feature extraction ensures view-consistent rendering, while facial appearance and driver expression feature use frontalized views from the 3D lifting", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to enable robust and high-fidelity facial disentanglement. To handle highly diverse portraits (variations in facial appearance, hairstyle, head covering, eyewear, etc.), we propose a new method for fine-tuning Lp3D on real datasets by introducing a mixed loss function based on real and synthetic datasets. Our volumetric disentanglement and rendering framework is trained only using in-the-wild videos from the CelebV-HQ dataset [112] in a self-supervised fashion.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We not only demonstrate that our volumetric face disentanglement approach produces qualitative superior head reenactments than existing ones, but also show on a wide and diverse set of source images how non-frontal poses and extreme expressions can be handled. We have quantitatively assessed our method on multiple benchmarks and outperform existing 2D and 3D state-of-the-art techniques in terms of fidelity, expression, and likeness accuracy metrics. Our 3D aware head reenactment technique is therefore suitable for AR/VR-based immersive applications, and we also showcase a teleconferencing system using a holographic display from LookingGlass [4]. We summarize the main contributions as follows:", + "bbox": [ + 496, + 212, + 892, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First fully volumetric disentanglement approach for real-time 3D aware head reenactment from a single photo. This method combines 3D lifting into a canonical triplane representation and formalized facial appearance and expression feature extraction.", + "- A 3D lifting network that is fine-tuned on unconstrained real-world data instead of only generating synthetic ones.", + "- We demonstrate superior fidelity, identity preservation, and robustness w.r.t. current state-of-the-art methods for facial reenactment on a wide range of public datasets. We plan to release our code to the public." + ], + "bbox": [ + 500, + 407, + 890, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 587, + 640, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Neural Head Reenactment. The problem of generating animations of photorealistic human heads given images or video inputs has been thoroughly explored using various neural rendering techniques in the past few years, outperforming traditional 3DMM-based methods [8, 26, 31, 44, 63, 66, 78, 79, 96] which often appear uncanny due to their compressed linear space. These approaches can be categorized into one-shot and multi-shot ones. While multishot methods generally achieve high-fidelity results, they are not suitable for many consumer applications as they typically require an extensive amount of training data, such as a monocular video capture [10, 11, 18, 21, 30, 34, 93, 108-110, 113], and sometimes even a calibrated multi-view stereo setup [13, 29, 55, 58, 69]. More recently, few-shot techniques [104] have also been introduced.", + "bbox": [ + 496, + 613, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To maximize accessibility, a considerable number of methods [17, 26, 27, 32, 35, 39, 66, 72-74, 76, 77, 83, 86, 87, 89, 97, 101-103, 107] use a single portrait as input by leveraging advanced generative modeling techniques based", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10337", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on in-the-wild video training data. While most methods rely on linear face models to extract facial expressions, the head reenactment technique from Drobyshev et al. [27] directly extract expression features from cropped 2D face regions, allowing them to obtain better face disentanglements, which results in higher fidelity face synthesis. While similar to our proposed approach in avoiding the use of low dimensional linear face models, their method is purely 2D and strugglesly with ensuring identity and expression consistency when novel views are synthesized.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D-Aware One-Shot Head Reenactment. Due to potential inconsistencies when rendering from different views or poses, a number of 3D-aware single shot head reenactment techniques [7, 19, 20, 24, 62, 65, 70, 75, 88, 92, 94, 95] have been introduced. These methods generally use an efficient 3D representation, such as neural radiance fields or 3D mesh, to geometrically constraint the neural rendering and improve view consistency. ROME [43] for instance is a mesh-based method using FLAME blendshapes [51] and neural textures. While view-consistent results can be produced for both face and hair regions, the use of low resolution polygonal meshes hinders the neural renderer to generate high-fidelity geometric and appearance details.", + "bbox": [ + 75, + 261, + 472, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit representations such as HeadNeRF [36] and MofaNeRF [38] use a NeRF-based parametric model which supports direct control of the head pose of the generated images. While real-time rendering is possible, these methods require intensive test-time optimization and often fail to preserve the identity of the source due to the use of compact latent vectors. Most recent methods [53, 54, 99] adopt the highly efficient tri-plane-based neural fields representation [20] to encode the 3D structure and appearance of the avatars head. Compared to the previous works on view-consistent neural avatars [36, 43, 53, 54, 59, 99], we refrain from depending on parametric head models for motion synthesis and, instead, learn the volumetric motion model from the training data. This methodology enables us to narrow the identity gap between the source and generated images and yield a superior fidelity of the generated motion compared to competing approaches, and hence a higher quality disentanglement for reenactment.", + "bbox": [ + 75, + 458, + 472, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D GAN Inversion. When training a whole reconstruction and disentangled reenactment model end-to-end on facial performance videos, one can introduce substantial overfitting and reduce the quality of the results. To address this problems, we focus our training approach to an inversion of pre-trained 3D-aware generative models for human heads.", + "bbox": [ + 75, + 750, + 470, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use tri-plane-based generative network EG3D [20] as the foundational generator, due to its proficiency in producing high-fidelity and view-consistent synthesis of human heads. For a given image, an effective 3D GAN inver", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion method should leverage these properties for estimating latent representations, which can be decoded into outputs that maintain view consistency and faithfully replicate the contents of the input. One naive approach is to adapt GAN inversion methods that were initially designed for 2D GANs to the EG3D pre-trained network. These methods either do a time consuming but more precise optimization [42, 68] or train a fast but less accurate encoder network [67, 80] to obtain the corresponding latent vectors. They often produce incorrect depth prediction, leading to clear artifacts in novel view synthesis. Hence, some methods are specifically designed for inverting 3D GANs, which either do multi-view optimization [45, 91] or predict residual features/tri-plane maps for refining the initial inversion results [12, 81, 98, 100].", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we rely on the state-of-the-art EG3D inversion method Lp3D [81]. While achieving excellent novel-view synthesis results, it lacks disentanglement between the appearance and expression of the provided image and is unable to impose various driving expressions onto the input. To address this limitation, we propose a new method that introduces appearance-expression disentanglement in the latent space of tri-planes using our new self and cross-reenactment training pipeline while relying on a pre-trained but fine-tuned Lp3D network for regularization which enables highly consistent view synthesis.", + "bbox": [ + 496, + 316, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. 3D-Aware Head Reenactment", + "text_level": 1, + "bbox": [ + 498, + 494, + 774, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our head reenactment pipeline consists of three stages: 1) 3D Lifting, 2) Volumetric Disentanglement, and 3) Tri-plane Rendering. Given a pair of source and driver images, we first frontalize them using a pre-trained but fine-tuned tri-plane-based 3D lifting module [81]. This driver alignment step is crucial and allows our model to disentangle the expressions from the head pose, which prevents overfitting. Then, the frontalized faces are fed into two separate convolutional encoders to extract the face features $F_{s}$ and $F_{d}$ . These extracted features are concatenated with the ones extracted from the tri-planes of the source, and all are fed together into several transformer blocks [90] to produce the expression tri-plane residual, which is added to the tri-planes of the source image. The final target image can be rendered from the new tri-planes using a pre-trained tri-plane renderer using the driver's pose.", + "bbox": [ + 496, + 520, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Fine-Tuned 3D Lifting", + "text_level": 1, + "bbox": [ + 498, + 771, + 710, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt Lp3d [81] as a 3D face-lifting module, which predicts the radiance field of any given face image in real-time. Instead of using an implicit multi-layer perceptron [61] or sparse voxels [28, 71] for the radiance field, Lp3D [81] uses tri-planes [20], which can be computed using a single forward of a deep learning network. Specifically, for a given source image $x_{s}$ , we first extract the tri-planes $T$ using a", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10338", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5267cdbdd612a796023ff7ca70a6eeefd6865cc1442abf601c03b118b3eab723.jpg", + "image_caption": [ + "Figure 2. Given a pair of source and driver images, our method processes them in three steps: 3D Lifting into tri-plane representations, Volumetric Disentanglement, which consists of source and driver frontalization and tri-plane residual generation, and Tri-plane Rendering via volumetric ray marching with subsequent super-resolution." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 831, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transformer-based appearance encoder $\\mathbf{E}_{\\mathrm{app}}$ ..", + "bbox": [ + 76, + 347, + 372, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\text {a p p}} \\left(x _ {s}\\right) = T \\in \\mathbb {R} ^ {3 \\times H \\times W \\times C} = \\left\\{T _ {x y}, T _ {y z}, T _ {z x} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 369, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The color $c$ and density $\\sigma$ of each point $p = (x, y, z)$ in the radiance field can be obtained by projecting $p$ onto the three planes and by summing up the features at the projected positions:", + "bbox": [ + 75, + 396, + 470, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc, \\sigma = \\mathbf {D} \\left(F _ {x y} + F _ {y z} + F _ {z x}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 465, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{D}$ is a shallow MLP decoder for the tri-plane rendering, $F_{xy}, F_{yz}$ , and $F_{zx}$ are the feature vectors at the projected positions on $xy$ , $yz$ , and $zx$ planes, respectively, calculated using bilinear interpolation. The rendered $128 \\times 128$ image is then upsampled using a super-resolution module to produce a high-resolution output. To train the encoder $\\mathbf{E}_{\\mathrm{app}}$ , Lp3D [81] uses synthetic data generated from a 3D-aware face generative model [20]. While these synthetic data have ground truth camera poses, they are limited to the face distribution of the generative model. As a result, Lp3D can fail to generalize to in-the-wild images as shown in Fig. 5. To prevent this, we fine-tune the pre-trained Lp3D on a large-scale real-world dataset. We also replace the original super-resolution module in Lp3D with a pre-trained GF-PGAN [85], which is then fine-tuned together with Lp3D (see Sec. 3.4).", + "bbox": [ + 75, + 491, + 472, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Disentangling Appearance and Expression", + "text_level": 1, + "bbox": [ + 76, + 741, + 439, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Separating facial expression from the identity appearance in a 3D radiance field is very challenging especially when source and driver subjects have misaligned expressions. In order to simplify the problem, we use our 3D lifting approach to bring both source and driver heads into a pos-oriented space where faces are frontalized. Here, we denote frontalized source and driver images as $x_{s}^{f}$ and $x_{d}^{f}$ , respectively. These images are then fed into two separate convolutional source and driver encoders $\\mathbf{E}_s$ and $\\mathbf{E}_d$ to produce", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coarse feature maps $F_{s} = \\mathbf{E}_{s}(x_{s}^{f})$ and $F_{d} = \\mathbf{E}_{d}(x_{d}^{f})$ . Since we already have the source's tri-plane, which encodes the 3D shape of the source, we use another encoder to encode this tri-plane and concatenate it together with the coarse frontalized feature maps of the images to produce expression feature $F$ :", + "bbox": [ + 496, + 345, + 893, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t} = \\mathbf {E} _ {t} (T)\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 446, + 712, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = F _ {s} \\oplus F _ {d} \\oplus F _ {t}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 465, + 763, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even though face frontalization aligns the source and the driver, there is still some misalignment between the two faces, e.g., the positions of the eyes may be different, or one mouth is open while the other is closed. Therefore, we feed the concatenation of the feature maps into several transformer blocks to produce the final residual tri-plane $\\mathbf{E}_v(F)$ . This residual is then added back to the source's tri-planes to change the source's expression to the driver's expression $T' = T + \\mathbf{E}_v(F)$ . Unlike LPR [54], we do not use a 3D face model to compute the expression but instead use the RGB images of the source and the driver directly, allowing the model to learn high-fidelity and realistic expressions.", + "bbox": [ + 496, + 491, + 893, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tri-Plane Rendering", + "text_level": 1, + "bbox": [ + 500, + 681, + 694, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting tri-planes are then volumetrically rendered into one or multiple output images using pose parameters and viewing angles in the case of a holographic display. Following EG3D [20], we use a neural radiance fields (NeRFs)-based volumetric ray marching approach [60]. However, instead of encoding each point in space via positional encodings [60], the features of the points along rays are calculated using their projections onto tri-planes. Since tri-planes are aligned with the frontal face, we can compute these rays directly using camera extrinsics $\\mathbf{P}_{\\mathrm{driver}}$ predicted by an off-the-shelf 3D head pose estimator [23].", + "bbox": [ + 496, + 703, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the renderings are highly view-consistent, the large number of points evaluated for each ray still limits", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10339", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the output resolution for real-time performance. We therefore follow [54] and employ a 2D upsampling network [84] based on StyleGAN2 [41], which in our experiments produced higher quality results than the upsampling approach in EG3D [20]. Finally, for holographic displays, we generate a number of renderings based on their viewing angles and simply using the head pose parameter. Real-time performance is achieved using efficient inference libraries such as TensorRT, half-precision, and batched inference over multiple GPUs.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Strategy", + "text_level": 1, + "bbox": [ + 76, + 250, + 250, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-Tuning Lp3D. To make Lp3D work with in-the-wild images, we fine-tune it on a large-scale real-world video dataset [111]. Unlike the use of synthetic data, real-world data do not have ground-truth camera parameters and facial expressions in monocular videos are typically inconsistent over time. While the camera parameters can be estimated using standard 3D pose estimators, the expression differences are difficult to determine. However, we found that we can ignore this expression difference and fine-tune Lp3D using real data together with continuous training on synthetic data. In particular, our experiments indicate that the fine-tuned model can still faithfully reconstruct 3D faces from the input without changing expressions and still generalize successfully on in-the-wild images. Specifically, on real video data, we sample two frames $x_{s}^{r}$ and $x_{d}^{r}$ and estimate their camera parameters $P_{s}^{r}$ and $P_{d}^{r}$ . Similar to [20], we assume a fixed intrinsic for standard portraits for all images. Then we use $E_{app}$ from Lp3D to calculate the triplanes of $x_{s}^{r}$ , render it using the two poses, and calculate reconstruction losses on the two rendered images:", + "bbox": [ + 75, + 272, + 472, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e a l}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, P _ {d} ^ {r}\\right) - x _ {d} ^ {r} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, x _ {s} ^ {r}\\right) - x _ {s} ^ {r} \\right\\|,\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 584, + 459, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Lp3D}(x,P)$ is the face in $x$ re-rendered using camera pose $P$ and $\\mathcal{L}_{\\mathrm{real}}$ is the loss for real images. Simultaneously, we render two synthetic images employing an identical latent code but through varying camera views and calculate the synthetic loss $\\mathcal{L}_{\\mathrm{syn}}$ :", + "bbox": [ + 76, + 609, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s y n}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {f}, P _ {d} ^ {s}\\right) - x _ {d} ^ {s} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {s}, P _ {s} ^ {s}\\right) - x _ {s} ^ {s} \\right\\|\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 694, + 455, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t r i}} = \\left\\| E _ {\\mathrm {a p p}} \\left(x _ {s} ^ {f}\\right) - T \\right\\|,\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 715, + 254, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the ground-truth tri-planes returned by EG3D [20] and $\\mathcal{L}_{\\mathrm{tri}}$ is the tri-plane loss adopted directly from Lp3D. The final loss $\\mathcal{L}_{\\mathrm{app}}$ for fine-tuning Lp3D can be formulated as:", + "bbox": [ + 76, + 741, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a p p}} = \\mathcal {L} _ {\\text {r e a l}} + \\lambda_ {\\text {s y n}} \\mathcal {L} _ {\\text {s y n}} + \\lambda_ {\\text {t r i}} \\mathcal {L} _ {\\text {t r i}}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 811, + 383, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{syn}}$ and $\\lambda_{\\mathrm{tri}}$ are tunable hyperparameters.", + "bbox": [ + 76, + 837, + 397, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Disentangling Appearance and Expressions. In this stage, we also use real-world videos as training data. For a", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pair of source and driver images $x_{s}$ and $x_{d}$ sampled from the same video, we apply the reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ which is a combination of $L1$ , perceptual [105], and identity losses, between the reenacted image $x_{s \\to d}$ and the corresponding ground-truth $x_{d}$ :", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\left\\| x _ {s \\rightarrow d} - x _ {d} \\right\\| _ {1} + \\phi \\left(x _ {s \\rightarrow d}, x _ {d}\\right) \\\\ + \\mathcal {L} _ {\\cos} \\left(\\mathrm {I D} \\left(x _ {s \\rightarrow d}\\right), \\mathrm {I D} \\left(x _ {d}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 178, + 828, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi$ is the perceptual loss, $\\mathrm{ID}(\\cdot)$ is a pretrained face recognizer [82], and $\\mathcal{L}_{\\mathrm{cos}}$ is the cosine similarity loss:", + "bbox": [ + 498, + 224, + 890, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o s}} ((u, v) = \\frac {< u , v >}{m a x (\\| u \\| , \\epsilon) \\times m a x (\\| v \\| , \\epsilon)}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 265, + 841, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\epsilon$ is a small constant to void division by 0. Similar to other works that use RGB images directly to calculate expressions [27], our proposed encoder also suffers from an \"identity leaking\" issue. Since there is no cross-reenactment dataset, the expression module is trained with self-reenactment video data. Therefore, without proper augmentation and regularization, the expression module can leak identity information from the driver to the output, making the model fail to generalize to cross-reenactment tasks. Hence, we introduce a Cross Identity Regularization. Specifically, we further sample an additional driver frame $x_{d'}$ from another video. We incorporate a GAN loss where real samples are $\\mathrm{Lp3D}(x_s, P^d)$ and fake samples are $x_{s \\to d'}$ . This GAN loss is also conditioned on the identity vector of the source $\\mathrm{ID}(x_s)$ . Following [27], we also apply strong augmentation (random warping and color jittering) and additionally mask the border of the driver randomly to further reduce potential identity leaks. The loss for expression training can be summarized as:", + "bbox": [ + 496, + 305, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\exp} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {C I R}} \\mathcal {L} _ {\\text {C I R}},\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 604, + 784, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CIR}}$ and $\\lambda_{\\mathrm{CIR}}$ are cross identity regularization and its hyperparameter, respectively.", + "bbox": [ + 498, + 633, + 890, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global Fine-Tuning. After training both Lp3D and the expression module, we iteratively fine-tune the two modules using the same losses as the previous sections. Specifically, for every 10000 iterations, we freeze one module and fine-tune the other and vice versa. In addition, we add a GAN loss on the super-resolution output of the Lp3D module.", + "bbox": [ + 496, + 681, + 890, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 784, + 633, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We train our model on CelebV-HQ dataset [112] using 7 NVIDIA RTX A6000 ADA (50Gb memory each). We use AdamW [56] to optimize the parameters with a learning rate of $10^{-4}$ and batch size of 28. The Lp3D finetuning takes 5 days for 500K iterations to converge. Training the expression module takes 2 days, and the", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10340", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f513c2a949c0db7a267ecb66890172cd06d9528b0c93469c943e9d0e6036d37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 89, + 212, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e301983bfe1ea85c7eceaad90b91f3aaa3a783d15cb18adde325e44049b5b785.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 89, + 321, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfa8277f2d88edd09be2a72efe590400755ce2a6a1e75f843093c225297f1add.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 89, + 431, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e0d254cf5d9713a9da084dec1a7e43bcdee6b84a1467afcd5d39fc57fb5d5d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 89, + 542, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1b6fc341f84b10ed1d198f7b8b98111911296fc8472fbb905b04f3a75ade98d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 89, + 671, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2cc95820c5f0783680b1ad4d1f50219473b6bd20900ebc95901fb166dc8c04ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 89, + 748, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a617a6e81ba29d544c1cf20bc04bad8ab5ba62f282df4dfa3a2d97b8efbb4683.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 89, + 859, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82b1a70b4f7a25d4dcc48c777800da2b5dc58196b807796ebd80e58942c7962a.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 104, + 172, + 212, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2a4ae51a7174dc3c69582b1be432a1c223a6c39692faf3dc0afd033e49daabe.jpg", + "image_caption": [ + "Driver" + ], + "image_footnote": [], + "bbox": [ + 212, + 172, + 321, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c77cc9da8ba590134d7e9810596f7c7a1faf7a6bba3d162cd72d1ac8ac6e4a52.jpg", + "image_caption": [ + "Ref View" + ], + "image_footnote": [], + "bbox": [ + 323, + 172, + 431, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fbe6ec14fed6fbe1715765fdb3e921b4392f2b83ce9bf5bbe2fd8fc28495a233.jpg", + "image_caption": [ + "Novel View" + ], + "image_footnote": [], + "bbox": [ + 433, + 172, + 542, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/483934e4ed20d13d76b74fa22756740634daf3c1d931d198638a0be8e76550c4.jpg", + "image_caption": [ + "Wrinkles" + ], + "image_footnote": [], + "bbox": [ + 544, + 172, + 750, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2af9511cf063f08dc26d5f9d7907c065f5d0823dba0353f6931b4864f96bcd12.jpg", + "image_caption": [ + "Geometry" + ], + "image_footnote": [], + "bbox": [ + 751, + 172, + 859, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c084db6949b657f0c5fc4f6df5a478a3a06409c46b84be45156a7bec4905fe1a.jpg", + "image_caption": [ + "Figure 3. Expression dependent high-fidelity details, incl. eye and forehead wrinkles, as well as nasolabial folds (see zoom-ins)", + "Figure 4. A qualitative comparison with the baselines on in-the-wild photos. Notice that our method is capable of producing a variety of facial expressions, and handle highly diverse subjects, with and without accessories, as well as extreme head poses, such as rows 3 and 4." + ], + "image_footnote": [], + "bbox": [ + 104, + 305, + 859, + 828 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10341", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ee85584040d7cc47cb6e01a904490d287edb498fa4d821438268dece613a4722.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSelf-reenactmentCross-reenactment
PSNR ↑SSIM ↑LPIPS ↓NAKD ↓ECMD ↓FID ↓CSIM ↑ECMD ↓FID ↓
ROME [43]18.460.4880.3510.0300.5941380.5070.740172
StyleHeat [97]19.730.6890.2780.0350.74889.80.3980.74495.5
OTAvatar [59]19.280.7490.2890.0350.65167.00.4620.90172.4
MegaPortraits [27]21.100.7310.2910.0220.75552.00.7290.77161.7
Ours22.830.7680.1680.0120.42640.50.7540.75436.4
", + "bbox": [ + 76, + 92, + 624, + 164 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ef1c474ae904fd5cf789f1d885b1c978e253f17aba83fb411b43866856a86dd.jpg", + "table_caption": [ + "Table 1. Evaluation on HDTF [106] dataset. Our method outperforms the competitors across almost all of the metrics for both self- and cross-reenactment scenarios." + ], + "table_footnote": [], + "table_body": "
MethodCross-reenactment
CSIM ↑ECMD ↓FID ↓
ROME [43]0.5190.9152.6
HeadNeRF [36]0.3460.88113
StyleHeat [97]0.4670.8550.2
MegaPortraits [27]0.6470.7729.2
Ours0.6080.7923.6
", + "bbox": [ + 643, + 92, + 890, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9127b5160d117f322ec40bd912a6a4c7fee3ea2570ab6fc5683e6bd8a64807a6.jpg", + "image_caption": [ + "Figure 5. Our implementation of Lp3D [81] before and after CelebV-HQ [112] fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 78, + 217, + 264, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d22b1cc7dea3097f684e30b683b5abb295a406d8c5ed5ea94c7a1a83ebe3432e.jpg", + "table_caption": [ + "Table 2. Evaluation on CelebA-HQ [40] dataset." + ], + "table_footnote": [], + "table_body": "
CSIM ↑ECMD ↓
Lp3D0.5480.82
Lp3D-FT0.6700.76
w/o frontal0.6681.01
w/o CIR0.5700.97
Ours0.6080.79
", + "bbox": [ + 282, + 215, + 467, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies conducted on CelebA-HQ [40] dataset. FT is a fine-tuned version of Lp3D, and \"frontal\" denotes frontalization of the source and driver.", + "bbox": [ + 279, + 284, + 468, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45d7c05ab06300c43cd4eb787b4fe219213b251900e5e6adab50625e5c74508f.jpg", + "image_caption": [ + "Figure 6. Ablation study for source and driver frontalization and cross identity regularization (CIR)." + ], + "image_footnote": [], + "bbox": [ + 94, + 378, + 447, + 493 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27fd866d1831199fb47c81b48ca96773e7cfc37b6b76051393e9a9da4ed6f5cd.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with LPR [54] method on the samples from HDTF [106] dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 534, + 447, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iterative fine-tuning takes another 5 days. More training details, such as hyperparameter fine-tuning or architecture of the networks, can be found in the supplementary materials.", + "bbox": [ + 75, + 718, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike Lp3D, our method reenacts faces without relifting in 3D for every frame. The tri-plane encoding and frontalization of the source image are only performed once. For each driver, we perform a single frontalization for the driver image (0.0115 ms), one expression encoding (0.0034 s), one tri-plane rendering at $128 \\times 128$ resolution (0.0071 s), and one neural upsampling (0.0099 s). Each view runs at 31.9 fps on an Nvidia RTX 4090 GPU including I/O. More details are provided in the supplemental materials.", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with state-of-the-art 3D-based [36, 43, 59] and 2D-based [27, 97] models. For MegaPortraits [27], we use our own implementation that was trained on the CelebV-HQ dataset. Similar to previous works, we evaluate our method using public benchmarks, including CelebA-HQ [40] and HDTF [106]. For CelebAHQ, we split the data into two equal sets. Each set contains around $15\\mathrm{K}$ images. Then, we use one set as the source and the rest as driver images. For the HDTF dataset, we perform cross-reenactment by using the first frame of each video as source and 200 first frames of other videos as drivers, which is more than $60\\mathrm{K}$ data pairs. Similarly, to evaluate self-reenactment, we also use the first frames of each video as sources and the rest of the same video as the driver. Furthermore, we also collected 100 face images on the internet and around 100 high-quality videos for qualitative comparisons. Video results are provided in the supplementary materials.", + "bbox": [ + 496, + 215, + 893, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. Given a source image $x_{s}$ , a driver image $x_{d}$ and reenacted output $x_{s\\rightarrow d}$ we use EMOCAv2 [22] to extract the FLAME [52] expression coefficients of the prediction and the driver, as well as the shape coefficients of the source. We then compute 2 FLAME meshes using the predicted shape coefficients in world coordinates, one with the expression coefficients of the driver and one with the expression coefficients of the reenacted output. We measure the distance between the 2 meshes and denote this expression metric as ECMD. Moreover, we also use cosine similarity between the embeddings of a face recognition network (CSIM) [101], normalized average keypoint distance (NAKD) [16], perceptual image similarity (LPIPS) [105], peak signal-to-noise ratio (PSNR), and structure similarity index measure (SSIM).", + "bbox": [ + 496, + 492, + 892, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide quantitative comparisons on HDTF and CelebA-HQ datasets in Tab. 1 and Tab. 2, respectively, and show that our method outperforms existing methods on both datasets. We also note that our FID and CSIM scores are significantly more reliable than the others, while expression-based metrics such as NAKD and ECMD are either better or very close to the best baseline, w.r.t output quality, expression accuracy, and identity consistency.", + "bbox": [ + 496, + 719, + 893, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We showcase the qualitative results of cross-identity reenactment on in-the-wild images in Fig. 4 and Fig. 3. Compared to the baselines [27, 36, 43, 97], our reenactment faithfully reconstructs intricate and complex elements,", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10342", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ca4e4feee8c33ed28bd87b7574a7ac5634c1adec84b3c3c545194688e98973d.jpg", + "image_caption": [ + "Figure 8. Failure cases of our method include side views in the source, extreme expressions, modeling of cartoonish characters and paintings, as well as modeling the reflections and semitransparency of the eyewear." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 447, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "such as hairstyle, facial hair, glasses, and facial makeup. Furthermore, our method effectively generates realistic and fine-scale dynamic details that match the driver's expressions including substantial head pose rotations. We also conduct a comparative analysis of our results with the current state-of-the-art 3D-aware method LPR [54] in Fig. 7. Compared to LPR, our method achieves superior identity consistency. We further refer to the supplemental video for a live demonstration of our holographic telepresence system and animated head reenactment results and comparisons, with and without disentangled poses.", + "bbox": [ + 75, + 256, + 472, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We compare Lp3D with and without fine-tuning on the CelebA-HQ dataset in Tab. 3 and show several examples in Fig. 5. Without fine-tuning on real data, our implementation of Lp3D fails to preserve the identity of the input image, resulting in a considerably lower CSIM score. We also try without any facial frontalization in the expression module and instead use the source and driver images directly to calculate the expression tri-plane residual. We observe in Fig. 6 that without face frontalization, the model completely ignores the expression of the driver and keeps the expression of the input source instead. We show in Tab. 3, that facial frontalization leads to much better ECMD score. We then measure the effectiveness of the GAN-based cross-identity regularization on the CelebA-HQ dataset, $\\mathcal{L}_{\\mathrm{CIR}}$ . Without this loss, identity characteristics (hairstyle or color) can leak from the driver to the output. See column 4 in Fig. 6. Tab. 3 also shows that cross-identity regularization can reduce identity leaking and improve the CSIM score. Lastly, we have also attempted to train our model end-to-end using the same losses and optimization process instead of our proposed iterative fine-tuning. Even with a lower learning rate and the use of pre-trained Lp3D weights, we were unable to succeed.", + "bbox": [ + 75, + 435, + 472, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Limitations of our approach are illustrated in Figure 8. For source images that are extremely side ways (i.e., over $90^{\\circ}$ ), our method can produce a plausible frontal face, but the likeness cannot be guaranteed due to insufficient visibility. For very highly stylized portraits, such as cartoons, our framework often produces photorealistic facial elements such as teeth which can be inconsistent in", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "style. Due to the dependence on training data volume and diversity, accessories such as dental braces or glasses may disappear or look different during synthesis. We believe that providing more and better training data can further improve the performance of our algorithm.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 180, + 612, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have demonstrated that a fully volumetric disentanglement of facial appearance and expressions is possible through a shared canonical tri-plane representation. In particular, an improved disentanglement also leads to higher fidelity and more robust head reenactment, when compared to existing methods that use linear face models for expressions, especially for non-frontal poses. A critical insight of our approach is that head frontalization via 3D lifting is particularly effective for extracting features that can encode fine details and expressions such as wrinkles and folds. The resulting reenactment is also highly view-consistent for large angles, making our solution suitable for holographic displays. We have also shown that the 3D lifting model can still be successfully trained with real data despite the fact that different frames with the same subject have varying facial expressions. Without a fine-tuned 3D lifting model, our 3D-aware reenactment framework would struggle with preserving the identity of the source, especially for side views. Our experiments indicate that our results achieve better visual quality and are more robust to extreme poses, validated via an extensive evaluation on multiple datasets.", + "bbox": [ + 496, + 205, + 892, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Risks and Potential Misuse. The proposed method is intended to promote avatar-based 3D communication. Nevertheless, our AI-based reenactment solution produces synthetic but highly realistic face videos from only a single photo, which could be hard to distinguish from a real person. Like deepfakes and other facial manipulation methods, potential misuse is possible and hence, we refer to the supplemental material for more discussions.", + "bbox": [ + 496, + 532, + 890, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Work. We are also interested in expanding our work to upper and full body reenactment, where hand gestures can be used for more engaging communication. To this end, we plan to investigate the use of canonical representations for human bodies, such as T-poses. As our primary motivation, we have showcased a solution using holographic displays for immersive 3D teleconferencing. However, we believe that our approach can also be extended to AR/VR HMD-based settings where full $360^{\\circ}$ head views are possible. The recent work by An et al. [7] is a promising avenue for future exploration.", + "bbox": [ + 496, + 662, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the Meta-verse Center Grant from the MBZUAI Research Office and by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\".", + "bbox": [ + 496, + 847, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10343", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] ItSeez3D AvatarSDK, https://avatarsdk.com. 1", + "[2] in3D, https://in3d.io.1", + "[3] Leia, https://www.leiainc.com. 2", + "[4] Looking Glass Factory, https://lookingglassfactory.com. 2", + "[5] Pinscreen Avatar Neo, https://www.Avatarneo.com. 1", + "[6] ReadyPlayerMe, https://readyplayer.me. 1", + "[7] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 8", + "[8] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20364-20373, 2022. 2", + "[9] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 362–371, 2023. 1", + "[10] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhinitra Meka, Ruofei Du, Ming-song Dou, Sergio Orts-Escalano, Rohit Pandey, Ping Tan, Thabo Beeler, Sean Fanello, and Yinda Zhang. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[11] Shrisha Bharadwaj, Yufeng Zheng, Otmar Hilliges, Michael J. Black, and Victoria Fernandez Abrevaya. FLARE: Fast learning of animatable and relightable mesh avatars. ACM Transactions on Graphics, 42:15, 2023. 2", + "[12] Ananta R Bhattacharai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. arXiv preprint arXiv:2303.13497, 2023. 3", + "[13] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason M. Saragih. Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG), 40, 2021. 2", + "[14] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 1", + "[15] Volker Blanz and Thomas Vetter. A Morphable Model For The Synthesis Of 3D Faces. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 2", + "[16] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 7", + "[17] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2" + ], + "bbox": [ + 86, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41, 2022. 2", + "[19] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 3", + "[20] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3, 4, 5", + "[21] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 416-426, 2023. 2", + "[22] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 7", + "[23] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 4", + "[24] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022. 3", + "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2", + "[26] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[27] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2, 3, 5, 7", + "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3", + "[29] Yonggan Fu, Yuecheng Li, Chenghui Li, Jason Saragih, Peizhao Zhang, Xiaoliang Dai, and Yingyan (Celine) Lin. Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In Proceedings of the" + ], + "bbox": [ + 509, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10344", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21036-21045, 2023. 2", + "[30] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2", + "[31] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2021. 2", + "[32] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5609-5619, 2023. 2", + "[33] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos P Zafeiriou. Fast-ganfit: Generative adversarial network for high fidelity 3d face reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1", + "[34] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18653–18664, 2022. 2", + "[35] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2", + "[36] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2, 3, 7", + "[37] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 1", + "[38] Yiyu huang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 3", + "[39] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2", + "[40] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 7", + "[41] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[42] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[43] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference of Computer vision (ECCV), 2022. 2, 3, 7" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Nießner, P. Pérez, C. Richardt, M. Zollhöfer, and C. Theobalt. Deep video portraits. ACM Transactions on Graphics 2018 (TOG), 2018. 2", + "[45] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2967-2976, 2023. 3", + "[46] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Abhijeet Ghosh, and Stefanos P Zafeiriou. Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1", + "[47] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8629-8640, 2023. 1", + "[48] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 40(6), 2021. 2", + "[49] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 394-403, 2023. 1", + "[50] Hao Li, Laura Trutoiu, Kyle Olszewski, Lingyu Wei, Tristan Trutna, Pei-Lun Hsieh, Aaron Nicholls, and Chongyang Ma. Facial performance sensing head-mounted display. ACM Transactions on Graphics (Proceedings SIGGRAPH 2015), 34(4), 2015. 2", + "[51] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3", + "[52] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7", + "[53] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17969-17978, 2023. 2, 3", + "[54] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. arXiv preprint arXiv:2306.08768, 2023. 2, 3, 4, 5, 7, 8", + "[55] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4), 2018. 2" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10345", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5", + "[57] Huiwen Luo, Koki Nagano, Han-Wei Kung, Mclean Gold-white, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. CoRR, abs/2106.11423, 2021. 1", + "[58] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De la Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 2", + "[59] Zhiyuan Ma, Xiangyu Zhu, Guo-Jun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16910, 2023. 2, 3, 7", + "[60] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 4", + "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[62] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021. 3", + "[63] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2", + "[64] Kyle Olszewski, Joseph J. Lim, Shunsuke Saito, and Hao Li. High-fidelity facial and speech animation for vr hmds. ACM Transactions on Graphics (TOG), 35:1-14, 2016. 2", + "[65] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 3", + "[66] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13759-13768, 2021. 2", + "[67] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[68] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3", + "[69] Gabriel Schwartz, Shih-En Wei, Te-Li Wang, Stephen Lombardi, Tomas Simon, Jason Saragih, and Yaser Sheikh. The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Trans. Graph., 39(4), 2020. 2" + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020. 3", + "[71] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 3", + "[72] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2", + "[73] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Conference on Neural Information Processing Systems (NeurIPS), 2019.", + "[74] Aliaksandr Siarohin, Oliver Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In CVPR, 2021. 2", + "[75] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d GANs. In Advances in Neural Information Processing Systems, 2022. 3", + "[76] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Pareidolia face reenactment. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[77] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3637-3646, 2022. 2", + "[78] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2016. 2", + "[79] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Headon: Real-time reenactment of human portrait videos. ACM Transactions on Graphics 2018 (TOG), 2018. 2", + "[80] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40 (4):1-14, 2021. 3", + "[81] Alex Trevithick, Matthew Chan, Michael Stengel, Eric Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 2, 3, 4, 7", + "[82] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5", + "[83] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2", + "[84] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168-9178, 2021. 4", + "[86] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022. 2", + "[87] O. Wiles, A.S. Koepke, and A. Zisserman. X2face: A network for controlling face generation by using images, audio, and pose codes. In European Conference on Computer Vision, 2018. 2", + "[88] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2195-2205, 2023. 3", + "[89] Sitao Xiang, Yuming Gu, Pengda Xiang, Mingming He, Koki Nagano, Haiwei Chen, and Hao Li. One-shot identity-preserving portrait reenactment. arXiv preprint arXiv:2004.12452, 2020. 2", + "[90] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3", + "[91] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2023. 3", + "[92] Hongyi Xu, Guoxian Song, Zihang Jiang, Jianfeng Zhang, Yichun Shi, Jing Liu, Wanchun Ma, Jiashi Feng, and Linjie Luo. Omniavatar: Geometry-guided controllable 3d head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12814-12824, 2023. 3", + "[93] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings. Association for Computing Machinery, 2023. 2", + "[94] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3", + "[95] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18440-18449, 2022. 3", + "[96] Kewei Yang, Kang Chen, Daoliang Guo, Song-Hai Zhang, Yuan-Chen Guo, and Weidong Zhang. Face2face $\\rho$ : Real-time high-resolution one-shot face reenactment. 2022. 2", + "[97] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In ECCV, 2022. 2, 7", + "[98] Yu Yin, Kamran Ghasedi, HsiangTao Wu, Jiaolong Yang, Xin Tong, and Yun Fu. Nerfinvertor: High fidelity nerf-gan" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inversion for single-shot real image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8539-8548, 2023. 3", + "[99] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based oneshot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 3", + "[100] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. arXiv preprint arXiv:2303.12326, 2023. 3", + "[101] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2, 7", + "[102] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 2", + "[103] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaprotait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22096-22105, 2023. 2", + "[104] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022. 2", + "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7", + "[106] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7", + "[107] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, pages 3657–3666, 2022. 2", + "[108] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2", + "[109] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13545-13555, 2022.", + "[110] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[111] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In Eu" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10347", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ropean conference on computer vision, pages 650-667. Springer, 2022. 5", + "[112] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 5, 7", + "[113] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 210 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "10348", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_model.json b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..066c1c16a8a6daef2ab4d1fd702be86c355ee969 --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_model.json @@ -0,0 +1,2910 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.77, + 0.175 + ], + "angle": 0, + "content": "VOODOO 3D: Volumetric Portrait Disentanglement for One-Shot 3D Head Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.203, + 0.862, + 0.276 + ], + "angle": 0, + "content": "Phong Tran1 Egor Zakharov2 Long-Nhat Ho1 Anh Tuan Tran3 Liwen Hu4 Hao Li1,4 \n1MBZUAI ETH Zurich 2VinAI Research 3Pinscreen \n{the.tran, long.ho}@mbzuai.ac.ae anhtt152@vinai.io ezakharov@ethz.ch liwen@pinsscreen.com hao@hao-li.com" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.31, + 0.851, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.538, + 0.825, + 0.566 + ], + "angle": 0, + "content": "Figure 1. We introduce VOODOO 3D: a high-fidelity 3D-aware one-shot head reenactment technique. Our method transfers the expression of a driver to a source and produces view consistent renderings for holographic displays." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.591, + 0.314, + 0.607 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.622, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We present a 3D-aware one-shot head reenactment method based on a fully volumetric neural disentanglement framework for source appearance and driver expressions. Our method is real-time and produces high-fidelity and view-consistent output, suitable for 3D teleconferencing systems based on holographic displays. Existing cutting-edge 3D-aware reenactment methods often use neural radiance fields or 3D meshes to produce view-consistent appearance encoding, but, at the same time, they rely on linear face models, such as 3DMM, to achieve its disentanglement with facial expressions. As a result, their reenactment results often exhibit identity leakage from the driver or have unnatural expressions. To address these problems, we propose a neural self-supervised disentanglement approach that lifts both the source image and driver video frame into a shared 3D volumetric representation based on tri-planes. This representation can then be freely manipulated with expression tri-planes extracted from the driving images and rendered from an arbitrary view using neural radiance fields. We achieve this disentanglement via self" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.706 + ], + "angle": 0, + "content": "supervised learning on a large in-the-wild video dataset. We further introduce a highly effective fine-tuning approach to improve the generalizability of the 3D lifting using the same real-world data. We demonstrate state-of-the-art performance on a wide range of datasets, and also showcase high-quality 3D-aware head reenactment on highly challenging and diverse subjects, including non-frontal head poses and complex expressions for both source and driver." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.738, + 0.633, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Creating 3D head avatars from a single photo is a core capability in making a wide range of consumer AR/VR and telepresence applications more accessible, and user experiences more engaging. Graphics engine-based 3D avatar digitization methods [9, 14, 33, 37, 46, 47, 49, 57] are suitable for today's video games and virtual worlds, and many commercial solutions exist (AvatarNeo [5], AvatarSDK [1], ReadyPlayerMe [6], in3D [2], etc.). However, the photorealism achieved by modern neural head reenactment tech-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10336" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "niques is becoming increasingly appealing for advanced effects in video sharing apps and visual effects. For immersive telepresence systems that use AR/VR headsets, facial expression capture is typically achieved using tiny video cameras built into HMDs [29, 50, 55, 64, 69], while the identity of the source subject recorded using a separate process. However, the teleconferencing solutions based on holographic 3D displays (LookingGlass [4], LEIA [3], etc.) use regular webcams [81] or depth sensors [48]. As opposed to a video-based setting, head reenactment for immersive applications needs to be 3D-aware, meaning that in addition to generating the correct poses and expressions from a photo, multi-view consistency is critical." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.548 + ], + "angle": 0, + "content": "While impressive facial reenactments results have been demonstrated using 2D approaches [26, 27, 83, 97, 102, 103], they typically struggle with preserving the likeness of the source and exhibit significant identity changes when varying the camera pose. More recently, 3D-aware one-shot head reenactment methods [36, 43, 53, 54, 59, 99] have used either 3D meshes or tri-plane neural radiance fields as a fast and memory efficient volumetric data representations for neural rendering. However, the expression and identity disentanglement in these methods is based on variants of linear face and expression models [15, 52] which lack expressiveness and high-frequency details. While these methods can achieve view consistency, facial expressions are often uncanny, and preserving the likeness of the input source portrait is challenging, especially for views different than the source image. Hence, input sources with extreme expressions and non-frontal poses are often avoided." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.779 + ], + "angle": 0, + "content": "In this paper, we introduce the first 3D aware one-shot head reenactment technique that disentangles source identities and the target expressions fully volumetrically, and without the use of explicit linear face models. Our method is real-time and designed with holographic displays in mind, where a large number of views (up to 45) can be rendered in parallel based on their viewing angle. We leverage the fact that real-time 3D lifting for human heads has recently been made possible [81] with the help of Vision Transformers (ViT) [25], which avoids the need for inefficient optimization-based GAN-inversion process [68]. In particular, 3D lifting allows us to map 2D face images into a canonical tri-plane representation for both source and target subjects and treat identity and expression disentanglement independently from the head pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Once the source image and driver frame are lifted into a pose-normalized tri-plane representation, we extract appearance features from the source subject and expressions from the driver. The pose of the driver is estimated separately using a 3D face tracker and used as input to a neural renderer. Tri-plane-based feature extraction ensures view-consistent rendering, while facial appearance and driver expression feature use frontalized views from the 3D lifting" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "to enable robust and high-fidelity facial disentanglement. To handle highly diverse portraits (variations in facial appearance, hairstyle, head covering, eyewear, etc.), we propose a new method for fine-tuning Lp3D on real datasets by introducing a mixed loss function based on real and synthetic datasets. Our volumetric disentanglement and rendering framework is trained only using in-the-wild videos from the CelebV-HQ dataset [112] in a self-supervised fashion." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.213, + 0.893, + 0.408 + ], + "angle": 0, + "content": "We not only demonstrate that our volumetric face disentanglement approach produces qualitative superior head reenactments than existing ones, but also show on a wide and diverse set of source images how non-frontal poses and extreme expressions can be handled. We have quantitatively assessed our method on multiple benchmarks and outperform existing 2D and 3D state-of-the-art techniques in terms of fidelity, expression, and likeness accuracy metrics. Our 3D aware head reenactment technique is therefore suitable for AR/VR-based immersive applications, and we also showcase a teleconferencing system using a holographic display from LookingGlass [4]. We summarize the main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.484 + ], + "angle": 0, + "content": "- First fully volumetric disentanglement approach for real-time 3D aware head reenactment from a single photo. This method combines 3D lifting into a canonical triplane representation and formalized facial appearance and expression feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.484, + 0.892, + 0.514 + ], + "angle": 0, + "content": "- A 3D lifting network that is fine-tuned on unconstrained real-world data instead of only generating synthetic ones." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.515, + 0.892, + 0.575 + ], + "angle": 0, + "content": "- We demonstrate superior fidelity, identity preservation, and robustness w.r.t. current state-of-the-art methods for facial reenactment on a wide range of public datasets. We plan to release our code to the public." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.642, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.614, + 0.892, + 0.84 + ], + "angle": 0, + "content": "2D Neural Head Reenactment. The problem of generating animations of photorealistic human heads given images or video inputs has been thoroughly explored using various neural rendering techniques in the past few years, outperforming traditional 3DMM-based methods [8, 26, 31, 44, 63, 66, 78, 79, 96] which often appear uncanny due to their compressed linear space. These approaches can be categorized into one-shot and multi-shot ones. While multishot methods generally achieve high-fidelity results, they are not suitable for many consumer applications as they typically require an extensive amount of training data, such as a monocular video capture [10, 11, 18, 21, 30, 34, 93, 108-110, 113], and sometimes even a calibrated multi-view stereo setup [13, 29, 55, 58, 69]. More recently, few-shot techniques [104] have also been introduced." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To maximize accessibility, a considerable number of methods [17, 26, 27, 32, 35, 39, 66, 72-74, 76, 77, 83, 86, 87, 89, 97, 101-103, 107] use a single portrait as input by leveraging advanced generative modeling techniques based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "on in-the-wild video training data. While most methods rely on linear face models to extract facial expressions, the head reenactment technique from Drobyshev et al. [27] directly extract expression features from cropped 2D face regions, allowing them to obtain better face disentanglements, which results in higher fidelity face synthesis. While similar to our proposed approach in avoiding the use of low dimensional linear face models, their method is purely 2D and strugglesly with ensuring identity and expression consistency when novel views are synthesized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.459 + ], + "angle": 0, + "content": "3D-Aware One-Shot Head Reenactment. Due to potential inconsistencies when rendering from different views or poses, a number of 3D-aware single shot head reenactment techniques [7, 19, 20, 24, 62, 65, 70, 75, 88, 92, 94, 95] have been introduced. These methods generally use an efficient 3D representation, such as neural radiance fields or 3D mesh, to geometrically constraint the neural rendering and improve view consistency. ROME [43] for instance is a mesh-based method using FLAME blendshapes [51] and neural textures. While view-consistent results can be produced for both face and hair regions, the use of low resolution polygonal meshes hinders the neural renderer to generate high-fidelity geometric and appearance details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.473, + 0.732 + ], + "angle": 0, + "content": "Implicit representations such as HeadNeRF [36] and MofaNeRF [38] use a NeRF-based parametric model which supports direct control of the head pose of the generated images. While real-time rendering is possible, these methods require intensive test-time optimization and often fail to preserve the identity of the source due to the use of compact latent vectors. Most recent methods [53, 54, 99] adopt the highly efficient tri-plane-based neural fields representation [20] to encode the 3D structure and appearance of the avatars head. Compared to the previous works on view-consistent neural avatars [36, 43, 53, 54, 59, 99], we refrain from depending on parametric head models for motion synthesis and, instead, learn the volumetric motion model from the training data. This methodology enables us to narrow the identity gap between the source and generated images and yield a superior fidelity of the generated motion compared to competing approaches, and hence a higher quality disentanglement for reenactment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.472, + 0.84 + ], + "angle": 0, + "content": "3D GAN Inversion. When training a whole reconstruction and disentangled reenactment model end-to-end on facial performance videos, one can introduce substantial overfitting and reduce the quality of the results. To address this problems, we focus our training approach to an inversion of pre-trained 3D-aware generative models for human heads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We use tri-plane-based generative network EG3D [20] as the foundational generator, due to its proficiency in producing high-fidelity and view-consistent synthesis of human heads. For a given image, an effective 3D GAN inver" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "sion method should leverage these properties for estimating latent representations, which can be decoded into outputs that maintain view consistency and faithfully replicate the contents of the input. One naive approach is to adapt GAN inversion methods that were initially designed for 2D GANs to the EG3D pre-trained network. These methods either do a time consuming but more precise optimization [42, 68] or train a fast but less accurate encoder network [67, 80] to obtain the corresponding latent vectors. They often produce incorrect depth prediction, leading to clear artifacts in novel view synthesis. Hence, some methods are specifically designed for inverting 3D GANs, which either do multi-view optimization [45, 91] or predict residual features/tri-plane maps for refining the initial inversion results [12, 81, 98, 100]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.485 + ], + "angle": 0, + "content": "In this work, we rely on the state-of-the-art EG3D inversion method Lp3D [81]. While achieving excellent novel-view synthesis results, it lacks disentanglement between the appearance and expression of the provided image and is unable to impose various driving expressions onto the input. To address this limitation, we propose a new method that introduces appearance-expression disentanglement in the latent space of tri-planes using our new self and cross-reenactment training pipeline while relying on a pre-trained but fine-tuned Lp3D network for regularization which enables highly consistent view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.496, + 0.776, + 0.512 + ], + "angle": 0, + "content": "3. 3D-Aware Head Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.765 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our head reenactment pipeline consists of three stages: 1) 3D Lifting, 2) Volumetric Disentanglement, and 3) Tri-plane Rendering. Given a pair of source and driver images, we first frontalize them using a pre-trained but fine-tuned tri-plane-based 3D lifting module [81]. This driver alignment step is crucial and allows our model to disentangle the expressions from the head pose, which prevents overfitting. Then, the frontalized faces are fed into two separate convolutional encoders to extract the face features \\( F_{s} \\) and \\( F_{d} \\). These extracted features are concatenated with the ones extracted from the tri-planes of the source, and all are fed together into several transformer blocks [90] to produce the expression tri-plane residual, which is added to the tri-planes of the source image. The final target image can be rendered from the new tri-planes using a pre-trained tri-plane renderer using the driver's pose." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.712, + 0.789 + ], + "angle": 0, + "content": "3.1. Fine-Tuned 3D Lifting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We adopt Lp3d [81] as a 3D face-lifting module, which predicts the radiance field of any given face image in real-time. Instead of using an implicit multi-layer perceptron [61] or sparse voxels [28, 71] for the radiance field, Lp3D [81] uses tri-planes [20], which can be computed using a single forward of a deep learning network. Specifically, for a given source image \\( x_{s} \\), we first extract the tri-planes \\( T \\) using a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10338" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.089, + 0.833, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 2. Given a pair of source and driver images, our method processes them in three steps: 3D Lifting into tri-plane representations, Volumetric Disentanglement, which consists of source and driver frontalization and tri-plane residual generation, and Tri-plane Rendering via volumetric ray marching with subsequent super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.374, + 0.365 + ], + "angle": 0, + "content": "transformer-based appearance encoder \\(\\mathbf{E}_{\\mathrm{app}}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.37, + 0.47, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\text {a p p}} \\left(x _ {s}\\right) = T \\in \\mathbb {R} ^ {3 \\times H \\times W \\times C} = \\left\\{T _ {x y}, T _ {y z}, T _ {z x} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.457 + ], + "angle": 0, + "content": "The color \\( c \\) and density \\( \\sigma \\) of each point \\( p = (x, y, z) \\) in the radiance field can be obtained by projecting \\( p \\) onto the three planes and by summing up the features at the projected positions:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.467, + 0.47, + 0.485 + ], + "angle": 0, + "content": "\\[\nc, \\sigma = \\mathbf {D} \\left(F _ {x y} + F _ {y z} + F _ {z x}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.473, + 0.735 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}\\) is a shallow MLP decoder for the tri-plane rendering, \\(F_{xy}, F_{yz}\\), and \\(F_{zx}\\) are the feature vectors at the projected positions on \\(xy\\), \\(yz\\), and \\(zx\\) planes, respectively, calculated using bilinear interpolation. The rendered \\(128 \\times 128\\) image is then upsampled using a super-resolution module to produce a high-resolution output. To train the encoder \\(\\mathbf{E}_{\\mathrm{app}}\\), Lp3D [81] uses synthetic data generated from a 3D-aware face generative model [20]. While these synthetic data have ground truth camera poses, they are limited to the face distribution of the generative model. As a result, Lp3D can fail to generalize to in-the-wild images as shown in Fig. 5. To prevent this, we fine-tune the pre-trained Lp3D on a large-scale real-world dataset. We also replace the original super-resolution module in Lp3D with a pre-trained GF-PGAN [85], which is then fine-tuned together with Lp3D (see Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.441, + 0.759 + ], + "angle": 0, + "content": "3.2. Disentangling Appearance and Expression" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Separating facial expression from the identity appearance in a 3D radiance field is very challenging especially when source and driver subjects have misaligned expressions. In order to simplify the problem, we use our 3D lifting approach to bring both source and driver heads into a pos-oriented space where faces are frontalized. Here, we denote frontalized source and driver images as \\( x_{s}^{f} \\) and \\( x_{d}^{f} \\), respectively. These images are then fed into two separate convolutional source and driver encoders \\( \\mathbf{E}_s \\) and \\( \\mathbf{E}_d \\) to produce" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.895, + 0.437 + ], + "angle": 0, + "content": "coarse feature maps \\( F_{s} = \\mathbf{E}_{s}(x_{s}^{f}) \\) and \\( F_{d} = \\mathbf{E}_{d}(x_{d}^{f}) \\). Since we already have the source's tri-plane, which encodes the 3D shape of the source, we use another encoder to encode this tri-plane and concatenate it together with the coarse frontalized feature maps of the images to produce expression feature \\( F \\):" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.447, + 0.713, + 0.464 + ], + "angle": 0, + "content": "\\[\nF _ {t} = \\mathbf {E} _ {t} (T)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.466, + 0.764, + 0.482 + ], + "angle": 0, + "content": "\\[\nF = F _ {s} \\oplus F _ {d} \\oplus F _ {t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.675 + ], + "angle": 0, + "content": "Even though face frontalization aligns the source and the driver, there is still some misalignment between the two faces, e.g., the positions of the eyes may be different, or one mouth is open while the other is closed. Therefore, we feed the concatenation of the feature maps into several transformer blocks to produce the final residual tri-plane \\(\\mathbf{E}_v(F)\\). This residual is then added back to the source's tri-planes to change the source's expression to the driver's expression \\(T' = T + \\mathbf{E}_v(F)\\). Unlike LPR [54], we do not use a 3D face model to compute the expression but instead use the RGB images of the source and the driver directly, allowing the model to learn high-fidelity and realistic expressions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.696, + 0.699 + ], + "angle": 0, + "content": "3.3. Tri-Plane Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.87 + ], + "angle": 0, + "content": "The resulting tri-planes are then volumetrically rendered into one or multiple output images using pose parameters and viewing angles in the case of a holographic display. Following EG3D [20], we use a neural radiance fields (NeRFs)-based volumetric ray marching approach [60]. However, instead of encoding each point in space via positional encodings [60], the features of the points along rays are calculated using their projections onto tri-planes. Since tri-planes are aligned with the frontal face, we can compute these rays directly using camera extrinsics \\(\\mathbf{P}_{\\mathrm{driver}}\\) predicted by an off-the-shelf 3D head pose estimator [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "While the renderings are highly view-consistent, the large number of points evaluated for each ray still limits" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10339" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "the output resolution for real-time performance. We therefore follow [54] and employ a 2D upsampling network [84] based on StyleGAN2 [41], which in our experiments produced higher quality results than the upsampling approach in EG3D [20]. Finally, for holographic displays, we generate a number of renderings based on their viewing angles and simply using the head pose parameter. Real-time performance is achieved using efficient inference libraries such as TensorRT, half-precision, and batched inference over multiple GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.251, + 0.25, + 0.267 + ], + "angle": 0, + "content": "3.4. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.577 + ], + "angle": 0, + "content": "Fine-Tuning Lp3D. To make Lp3D work with in-the-wild images, we fine-tune it on a large-scale real-world video dataset [111]. Unlike the use of synthetic data, real-world data do not have ground-truth camera parameters and facial expressions in monocular videos are typically inconsistent over time. While the camera parameters can be estimated using standard 3D pose estimators, the expression differences are difficult to determine. However, we found that we can ignore this expression difference and fine-tune Lp3D using real data together with continuous training on synthetic data. In particular, our experiments indicate that the fine-tuned model can still faithfully reconstruct 3D faces from the input without changing expressions and still generalize successfully on in-the-wild images. Specifically, on real video data, we sample two frames \\( x_{s}^{r} \\) and \\( x_{d}^{r} \\) and estimate their camera parameters \\( P_{s}^{r} \\) and \\( P_{d}^{r} \\). Similar to [20], we assume a fixed intrinsic for standard portraits for all images. Then we use \\( E_{app} \\) from Lp3D to calculate the triplanes of \\( x_{s}^{r} \\), render it using the two poses, and calculate reconstruction losses on the two rendered images:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.585, + 0.46, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e a l}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, P _ {d} ^ {r}\\right) - x _ {d} ^ {r} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, x _ {s} ^ {r}\\right) - x _ {s} ^ {r} \\right\\|,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.47, + 0.687 + ], + "angle": 0, + "content": "where \\(\\mathrm{Lp3D}(x,P)\\) is the face in \\(x\\) re-rendered using camera pose \\(P\\) and \\(\\mathcal{L}_{\\mathrm{real}}\\) is the loss for real images. Simultaneously, we render two synthetic images employing an identical latent code but through varying camera views and calculate the synthetic loss \\(\\mathcal{L}_{\\mathrm{syn}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.695, + 0.457, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s y n}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {f}, P _ {d} ^ {s}\\right) - x _ {d} ^ {s} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {s}, P _ {s} ^ {s}\\right) - x _ {s} ^ {s} \\right\\|\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.717, + 0.255, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t r i}} = \\left\\| E _ {\\mathrm {a p p}} \\left(x _ {s} ^ {f}\\right) - T \\right\\|,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.47, + 0.803 + ], + "angle": 0, + "content": "where \\(T\\) is the ground-truth tri-planes returned by EG3D [20] and \\(\\mathcal{L}_{\\mathrm{tri}}\\) is the tri-plane loss adopted directly from Lp3D. The final loss \\(\\mathcal{L}_{\\mathrm{app}}\\) for fine-tuning Lp3D can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.813, + 0.384, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a p p}} = \\mathcal {L} _ {\\text {r e a l}} + \\lambda_ {\\text {s y n}} \\mathcal {L} _ {\\text {s y n}} + \\lambda_ {\\text {t r i}} \\mathcal {L} _ {\\text {t r i}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.398, + 0.854 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{syn}}\\) and \\(\\lambda_{\\mathrm{tri}}\\) are tunable hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Disentangling Appearance and Expressions. In this stage, we also use real-world videos as training data. For a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "pair of source and driver images \\( x_{s} \\) and \\( x_{d} \\) sampled from the same video, we apply the reconstruction loss \\( \\mathcal{L}_{\\mathrm{recon}} \\) which is a combination of \\( L1 \\), perceptual [105], and identity losses, between the reenacted image \\( x_{s \\to d} \\) and the corresponding ground-truth \\( x_{d} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.179, + 0.829, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\left\\| x _ {s \\rightarrow d} - x _ {d} \\right\\| _ {1} + \\phi \\left(x _ {s \\rightarrow d}, x _ {d}\\right) \\\\ + \\mathcal {L} _ {\\cos} \\left(\\mathrm {I D} \\left(x _ {s \\rightarrow d}\\right), \\mathrm {I D} \\left(x _ {d}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.257 + ], + "angle": 0, + "content": "where \\(\\phi\\) is the perceptual loss, \\(\\mathrm{ID}(\\cdot)\\) is a pretrained face recognizer [82], and \\(\\mathcal{L}_{\\mathrm{cos}}\\) is the cosine similarity loss:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.266, + 0.842, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o s}} ((u, v) = \\frac {< u , v >}{m a x (\\| u \\| , \\epsilon) \\times m a x (\\| v \\| , \\epsilon)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Here \\(\\epsilon\\) is a small constant to void division by 0. Similar to other works that use RGB images directly to calculate expressions [27], our proposed encoder also suffers from an \"identity leaking\" issue. Since there is no cross-reenactment dataset, the expression module is trained with self-reenactment video data. Therefore, without proper augmentation and regularization, the expression module can leak identity information from the driver to the output, making the model fail to generalize to cross-reenactment tasks. Hence, we introduce a Cross Identity Regularization. Specifically, we further sample an additional driver frame \\(x_{d'}\\) from another video. We incorporate a GAN loss where real samples are \\(\\mathrm{Lp3D}(x_s, P^d)\\) and fake samples are \\(x_{s \\to d'}\\). This GAN loss is also conditioned on the identity vector of the source \\(\\mathrm{ID}(x_s)\\). Following [27], we also apply strong augmentation (random warping and color jittering) and additionally mask the border of the driver randomly to further reduce potential identity leaks. The loss for expression training can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.606, + 0.785, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\exp} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {C I R}} \\mathcal {L} _ {\\text {C I R}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.665 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CIR}}\\) and \\(\\lambda_{\\mathrm{CIR}}\\) are cross identity regularization and its hyperparameter, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Global Fine-Tuning. After training both Lp3D and the expression module, we iteratively fine-tune the two modules using the same losses as the previous sections. Specifically, for every 10000 iterations, we freeze one module and fine-tune the other and vice versa. In addition, we add a GAN loss on the super-resolution output of the Lp3D module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.634, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We train our model on CelebV-HQ dataset [112] using 7 NVIDIA RTX A6000 ADA (50Gb memory each). We use AdamW [56] to optimize the parameters with a learning rate of \\(10^{-4}\\) and batch size of 28. The Lp3D finetuning takes 5 days for 500K iterations to converge. Training the expression module takes 2 days, and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.214, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.09, + 0.323, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.09, + 0.432, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.09, + 0.543, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.09, + 0.673, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.09, + 0.749, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.09, + 0.861, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.174, + 0.214, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.261, + 0.186, + 0.273 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.174, + 0.323, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.261, + 0.292, + 0.273 + ], + "angle": 0, + "content": "Driver" + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.174, + 0.432, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.261, + 0.416, + 0.273 + ], + "angle": 0, + "content": "Ref View" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.174, + 0.543, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.261, + 0.529, + 0.273 + ], + "angle": 0, + "content": "Novel View" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.174, + 0.75, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.261, + 0.681, + 0.273 + ], + "angle": 0, + "content": "Wrinkles" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.174, + 0.86, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.261, + 0.843, + 0.275 + ], + "angle": 0, + "content": "Geometry" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.28, + 0.86, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Expression dependent high-fidelity details, incl. eye and forehead wrinkles, as well as nasolabial folds (see zoom-ins)" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.306, + 0.861, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.835, + 0.893, + 0.863 + ], + "angle": 0, + "content": "Figure 4. A qualitative comparison with the baselines on in-the-wild photos. Notice that our method is capable of producing a variety of facial expressions, and handle highly diverse subjects, with and without accessories, as well as extreme head poses, such as rows 3 and 4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10341" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.093, + 0.625, + 0.165 + ], + "angle": 0, + "content": "
MethodSelf-reenactmentCross-reenactment
PSNR ↑SSIM ↑LPIPS ↓NAKD ↓ECMD ↓FID ↓CSIM ↑ECMD ↓FID ↓
ROME [43]18.460.4880.3510.0300.5941380.5070.740172
StyleHeat [97]19.730.6890.2780.0350.74889.80.3980.74495.5
OTAvatar [59]19.280.7490.2890.0350.65167.00.4620.90172.4
MegaPortraits [27]21.100.7310.2910.0220.75552.00.7290.77161.7
Ours22.830.7680.1680.0120.42640.50.7540.75436.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.173, + 0.627, + 0.199 + ], + "angle": 0, + "content": "Table 1. Evaluation on HDTF [106] dataset. Our method outperforms the competitors across almost all of the metrics for both self- and cross-reenactment scenarios." + }, + { + "type": "table", + "bbox": [ + 0.645, + 0.093, + 0.891, + 0.164 + ], + "angle": 0, + "content": "
MethodCross-reenactment
CSIM ↑ECMD ↓FID ↓
ROME [43]0.5190.9152.6
HeadNeRF [36]0.3460.88113
StyleHeat [97]0.4670.8550.2
MegaPortraits [27]0.6470.7729.2
Ours0.6080.7923.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.643, + 0.173, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Table 2. Evaluation on CelebA-HQ [40] dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.218, + 0.266, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.325, + 0.266, + 0.364 + ], + "angle": 0, + "content": "Figure 5. Our implementation of Lp3D [81] before and after CelebV-HQ [112] fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.217, + 0.468, + 0.282 + ], + "angle": 0, + "content": "
CSIM ↑ECMD ↓
Lp3D0.5480.82
Lp3D-FT0.6700.76
w/o frontal0.6681.01
w/o CIR0.5700.97
Ours0.6080.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.285, + 0.47, + 0.361 + ], + "angle": 0, + "content": "Table 3. Ablation studies conducted on CelebA-HQ [40] dataset. FT is a fine-tuned version of Lp3D, and \"frontal\" denotes frontalization of the source and driver." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.379, + 0.449, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.494, + 0.47, + 0.521 + ], + "angle": 0, + "content": "Figure 6. Ablation study for source and driver frontalization and cross identity regularization (CIR)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.535, + 0.449, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with LPR [54] method on the samples from HDTF [106] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.764 + ], + "angle": 0, + "content": "iterative fine-tuning takes another 5 days. More training details, such as hyperparameter fine-tuning or architecture of the networks, can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Unlike Lp3D, our method reenacts faces without relifting in 3D for every frame. The tri-plane encoding and frontalization of the source image are only performed once. For each driver, we perform a single frontalization for the driver image (0.0115 ms), one expression encoding (0.0034 s), one tri-plane rendering at \\(128 \\times 128\\) resolution (0.0071 s), and one neural upsampling (0.0099 s). Each view runs at 31.9 fps on an Nvidia RTX 4090 GPU including I/O. More details are provided in the supplemental materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.895, + 0.474 + ], + "angle": 0, + "content": "We compare our method with state-of-the-art 3D-based [36, 43, 59] and 2D-based [27, 97] models. For MegaPortraits [27], we use our own implementation that was trained on the CelebV-HQ dataset. Similar to previous works, we evaluate our method using public benchmarks, including CelebA-HQ [40] and HDTF [106]. For CelebAHQ, we split the data into two equal sets. Each set contains around \\(15\\mathrm{K}\\) images. Then, we use one set as the source and the rest as driver images. For the HDTF dataset, we perform cross-reenactment by using the first frame of each video as source and 200 first frames of other videos as drivers, which is more than \\(60\\mathrm{K}\\) data pairs. Similarly, to evaluate self-reenactment, we also use the first frames of each video as sources and the rest of the same video as the driver. Furthermore, we also collected 100 face images on the internet and around 100 high-quality videos for qualitative comparisons. Video results are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.719 + ], + "angle": 0, + "content": "Evaluation. Given a source image \\( x_{s} \\), a driver image \\( x_{d} \\) and reenacted output \\( x_{s\\rightarrow d} \\) we use EMOCAv2 [22] to extract the FLAME [52] expression coefficients of the prediction and the driver, as well as the shape coefficients of the source. We then compute 2 FLAME meshes using the predicted shape coefficients in world coordinates, one with the expression coefficients of the driver and one with the expression coefficients of the reenacted output. We measure the distance between the 2 meshes and denote this expression metric as ECMD. Moreover, we also use cosine similarity between the embeddings of a face recognition network (CSIM) [101], normalized average keypoint distance (NAKD) [16], perceptual image similarity (LPIPS) [105], peak signal-to-noise ratio (PSNR), and structure similarity index measure (SSIM)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.841 + ], + "angle": 0, + "content": "We provide quantitative comparisons on HDTF and CelebA-HQ datasets in Tab. 1 and Tab. 2, respectively, and show that our method outperforms existing methods on both datasets. We also note that our FID and CSIM scores are significantly more reliable than the others, while expression-based metrics such as NAKD and ECMD are either better or very close to the best baseline, w.r.t output quality, expression accuracy, and identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We showcase the qualitative results of cross-identity reenactment on in-the-wild images in Fig. 4 and Fig. 3. Compared to the baselines [27, 36, 43, 97], our reenactment faithfully reconstructs intricate and complex elements," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.449, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.187, + 0.47, + 0.24 + ], + "angle": 0, + "content": "Figure 8. Failure cases of our method include side views in the source, extreme expressions, modeling of cartoonish characters and paintings, as well as modeling the reflections and semitransparency of the eyewear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.424 + ], + "angle": 0, + "content": "such as hairstyle, facial hair, glasses, and facial makeup. Furthermore, our method effectively generates realistic and fine-scale dynamic details that match the driver's expressions including substantial head pose rotations. We also conduct a comparative analysis of our results with the current state-of-the-art 3D-aware method LPR [54] in Fig. 7. Compared to LPR, our method achieves superior identity consistency. We further refer to the supplemental video for a live demonstration of our holographic telepresence system and animated head reenactment results and comparisons, with and without disentangled poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.473, + 0.784 + ], + "angle": 0, + "content": "Ablation Study. We compare Lp3D with and without fine-tuning on the CelebA-HQ dataset in Tab. 3 and show several examples in Fig. 5. Without fine-tuning on real data, our implementation of Lp3D fails to preserve the identity of the input image, resulting in a considerably lower CSIM score. We also try without any facial frontalization in the expression module and instead use the source and driver images directly to calculate the expression tri-plane residual. We observe in Fig. 6 that without face frontalization, the model completely ignores the expression of the driver and keeps the expression of the input source instead. We show in Tab. 3, that facial frontalization leads to much better ECMD score. We then measure the effectiveness of the GAN-based cross-identity regularization on the CelebA-HQ dataset, \\(\\mathcal{L}_{\\mathrm{CIR}}\\). Without this loss, identity characteristics (hairstyle or color) can leak from the driver to the output. See column 4 in Fig. 6. Tab. 3 also shows that cross-identity regularization can reduce identity leaking and improve the CSIM score. Lastly, we have also attempted to train our model end-to-end using the same losses and optimization process instead of our proposed iterative fine-tuning. Even with a lower learning rate and the use of pre-trained Lp3D weights, we were unable to succeed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Limitations. Limitations of our approach are illustrated in Figure 8. For source images that are extremely side ways (i.e., over \\(90^{\\circ}\\)), our method can produce a plausible frontal face, but the likeness cannot be guaranteed due to insufficient visibility. For very highly stylized portraits, such as cartoons, our framework often produces photorealistic facial elements such as teeth which can be inconsistent in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "style. Due to the dependence on training data volume and diversity, accessories such as dental braces or glasses may disappear or look different during synthesis. We believe that providing more and better training data can further improve the performance of our algorithm." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.181, + 0.613, + 0.195 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.893, + 0.523 + ], + "angle": 0, + "content": "We have demonstrated that a fully volumetric disentanglement of facial appearance and expressions is possible through a shared canonical tri-plane representation. In particular, an improved disentanglement also leads to higher fidelity and more robust head reenactment, when compared to existing methods that use linear face models for expressions, especially for non-frontal poses. A critical insight of our approach is that head frontalization via 3D lifting is particularly effective for extracting features that can encode fine details and expressions such as wrinkles and folds. The resulting reenactment is also highly view-consistent for large angles, making our solution suitable for holographic displays. We have also shown that the 3D lifting model can still be successfully trained with real data despite the fact that different frames with the same subject have varying facial expressions. Without a fine-tuned 3D lifting model, our 3D-aware reenactment framework would struggle with preserving the identity of the source, especially for side views. Our experiments indicate that our results achieve better visual quality and are more robust to extreme poses, validated via an extensive evaluation on multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Risks and Potential Misuse. The proposed method is intended to promote avatar-based 3D communication. Nevertheless, our AI-based reenactment solution produces synthetic but highly realistic face videos from only a single photo, which could be hard to distinguish from a real person. Like deepfakes and other facial manipulation methods, potential misuse is possible and hence, we refer to the supplemental material for more discussions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Future Work. We are also interested in expanding our work to upper and full body reenactment, where hand gestures can be used for more engaging communication. To this end, we plan to investigate the use of canonical representations for human bodies, such as T-poses. As our primary motivation, we have showcased a solution using holographic displays for immersive 3D teleconferencing. However, we believe that our approach can also be extended to AR/VR HMD-based settings where full \\(360^{\\circ}\\) head views are possible. The recent work by An et al. [7] is a promising avenue for future exploration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.848, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the Meta-verse Center Grant from the MBZUAI Research Office and by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\"." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10343" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.115, + 0.396, + 0.127 + ], + "angle": 0, + "content": "[1] ItSeez3D AvatarSDK, https://avatarsdk.com. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.13, + 0.255, + 0.142 + ], + "angle": 0, + "content": "[2] in3D, https://in3d.io.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.145, + 0.315, + 0.157 + ], + "angle": 0, + "content": "[3] Leia, https://www.leiainc.com. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.16, + 0.463, + 0.172 + ], + "angle": 0, + "content": "[4] Looking Glass Factory, https://lookingglassfactory.com. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.175, + 0.432, + 0.187 + ], + "angle": 0, + "content": "[5] Pinscreen Avatar Neo, https://www.Avatarneo.com. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.19, + 0.37, + 0.202 + ], + "angle": 0, + "content": "[6] ReadyPlayerMe, https://readyplayer.me. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.204, + 0.469, + 0.267 + ], + "angle": 0, + "content": "[7] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.27, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[8] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20364-20373, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.336, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[9] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 362–371, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.402, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[10] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhinitra Meka, Ruofei Du, Ming-song Dou, Sergio Orts-Escalano, Rohit Pandey, Ping Tan, Thabo Beeler, Sean Fanello, and Yinda Zhang. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.506, + 0.468, + 0.556 + ], + "angle": 0, + "content": "[11] Shrisha Bharadwaj, Yufeng Zheng, Otmar Hilliges, Michael J. Black, and Victoria Fernandez Abrevaya. FLARE: Fast learning of animatable and relightable mesh avatars. ACM Transactions on Graphics, 42:15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.559, + 0.468, + 0.597 + ], + "angle": 0, + "content": "[12] Ananta R Bhattacharai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. arXiv preprint arXiv:2303.13497, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.599, + 0.468, + 0.662 + ], + "angle": 0, + "content": "[13] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason M. Saragih. Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG), 40, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.665, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[14] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.731, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[15] Volker Blanz and Thomas Vetter. A Morphable Model For The Synthesis Of 3D Faces. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.771, + 0.469, + 0.834 + ], + "angle": 0, + "content": "[16] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.837, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[17] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.115, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.094, + 0.891, + 0.157 + ], + "angle": 0, + "content": "[18] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.159, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.237, + 0.892, + 0.325 + ], + "angle": 0, + "content": "[20] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.328, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[21] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 416-426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.393, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[22] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.458, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[23] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.536, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[24] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.602, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[26] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.733, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[27] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.785, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.85, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Yonggan Fu, Yuecheng Li, Chenghui Li, Jason Saragih, Peizhao Zhang, Xiaoliang Dai, and Yingyan (Celine) Lin. Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10344" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21036-21045, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.12, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[30] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.184, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[31] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.235, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[32] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5609-5619, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.299, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[33] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos P Zafeiriou. Fast-ganfit: Generative adversarial network for high fidelity 3d face reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.362, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[34] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18653–18664, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.427, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[35] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.466, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[36] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.529, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[37] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.593, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[38] Yiyu huang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.62, + 0.468, + 0.682 + ], + "angle": 0, + "content": "[39] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.683, + 0.468, + 0.734 + ], + "angle": 0, + "content": "[40] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.735, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[41] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.786, + 0.47, + 0.849 + ], + "angle": 0, + "content": "[42] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.85, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[43] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference of Computer vision (ECCV), 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.145 + ], + "angle": 0, + "content": "[44] H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Nießner, P. Pérez, C. Richardt, M. Zollhöfer, and C. Theobalt. Deep video portraits. ACM Transactions on Graphics 2018 (TOG), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.146, + 0.892, + 0.209 + ], + "angle": 0, + "content": "[45] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2967-2976, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.211, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[46] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Abhijeet Ghosh, and Stefanos P Zafeiriou. Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[47] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8629-8640, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.367, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[48] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 40(6), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.471, + 0.892, + 0.548 + ], + "angle": 0, + "content": "[49] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 394-403, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.549, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[50] Hao Li, Laura Trutoiu, Kyle Olszewski, Lingyu Wei, Tristan Trutna, Pei-Lun Hsieh, Aaron Nicholls, and Chongyang Ma. Facial performance sensing head-mounted display. ACM Transactions on Graphics (Proceedings SIGGRAPH 2015), 34(4), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.613, + 0.892, + 0.665 + ], + "angle": 0, + "content": "[51] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.666, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[52] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.718, + 0.892, + 0.808 + ], + "angle": 0, + "content": "[53] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17969-17978, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.809, + 0.892, + 0.86 + ], + "angle": 0, + "content": "[54] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. arXiv preprint arXiv:2306.08768, 2023. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.862, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4), 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10345" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.131 + ], + "angle": 0, + "content": "[56] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.132, + 0.471, + 0.184 + ], + "angle": 0, + "content": "[57] Huiwen Luo, Koki Nagano, Han-Wei Kung, Mclean Gold-white, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. CoRR, abs/2106.11423, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[58] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De la Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.25, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[59] Zhiyuan Ma, Xiangyu Zhu, Guo-Jun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16910, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[60] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.367, + 0.471, + 0.43 + ], + "angle": 0, + "content": "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[62] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.471, + 0.548 + ], + "angle": 0, + "content": "[63] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.549, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[64] Kyle Olszewski, Joseph J. Lim, Shunsuke Saito, and Hao Li. High-fidelity facial and speech animation for vr hmds. ACM Transactions on Graphics (TOG), 35:1-14, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.666 + ], + "angle": 0, + "content": "[65] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.667, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[66] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13759-13768, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.471, + 0.796 + ], + "angle": 0, + "content": "[67] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.797, + 0.471, + 0.835 + ], + "angle": 0, + "content": "[68] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.836, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[69] Gabriel Schwartz, Shih-En Wei, Te-Li Wang, Stephen Lombardi, Tomas Simon, Jason Saragih, and Yaser Sheikh. The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Trans. Graph., 39(4), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.157 + ], + "angle": 0, + "content": "[70] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.158, + 0.892, + 0.209 + ], + "angle": 0, + "content": "[71] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.21, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[72] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.249, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[73] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Conference on Neural Information Processing Systems (NeurIPS), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.302, + 0.892, + 0.34 + ], + "angle": 0, + "content": "[74] Aliaksandr Siarohin, Oliver Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.342, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[75] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d GANs. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.393, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[76] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Pareidolia face reenactment. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.446, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[77] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3637-3646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.51, + 0.892, + 0.562 + ], + "angle": 0, + "content": "[78] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.563, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[79] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Headon: Real-time reenactment of human portrait videos. ACM Transactions on Graphics 2018 (TOG), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.614, + 0.892, + 0.666 + ], + "angle": 0, + "content": "[80] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40 (4):1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.667, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[81] Alex Trevithick, Matthew Chan, Michael Stengel, Eric Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.744, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[82] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.785, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[83] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.848, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[84] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10346" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[85] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168-9178, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.209 + ], + "angle": 0, + "content": "[86] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.209, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[87] O. Wiles, A.S. Koepke, and A. Zisserman. X2face: A network for controlling face generation by using images, audio, and pose codes. In European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[88] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2195-2205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.324, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[89] Sitao Xiang, Yuming Gu, Pengda Xiang, Mingming He, Koki Nagano, Haiwei Chen, and Hao Li. One-shot identity-preserving portrait reenactment. arXiv preprint arXiv:2004.12452, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[90] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[91] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[92] Hongyi Xu, Guoxian Song, Zihang Jiang, Jianfeng Zhang, Yichun Shi, Jing Liu, Wanchun Ma, Jiashi Feng, and Linjie Luo. Omniavatar: Geometry-guided controllable 3d head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12814-12824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.471, + 0.644 + ], + "angle": 0, + "content": "[93] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings. Association for Computing Machinery, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.471, + 0.707 + ], + "angle": 0, + "content": "[94] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.708, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[95] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18440-18449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.772, + 0.471, + 0.811 + ], + "angle": 0, + "content": "[96] Kewei Yang, Kang Chen, Daoliang Guo, Song-Hai Zhang, Yuan-Chen Guo, and Weidong Zhang. Face2face \\(\\rho\\): Real-time high-resolution one-shot face reenactment. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.811, + 0.471, + 0.874 + ], + "angle": 0, + "content": "[97] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In ECCV, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[98] Yu Yin, Kamran Ghasedi, HsiangTao Wu, Jiaolong Yang, Xin Tong, and Yun Fu. Nerfinvertor: High fidelity nerf-gan" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "inversion for single-shot real image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8539-8548, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.133, + 0.892, + 0.197 + ], + "angle": 0, + "content": "[99] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based oneshot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.198, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[100] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. arXiv preprint arXiv:2303.12326, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[101] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[102] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[103] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaprotait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22096-22105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.458, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[104] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.51, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[106] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.64, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[107] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, pages 3657–3666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.667, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[108] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.795 + ], + "angle": 0, + "content": "[109] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13545-13555, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.797, + 0.892, + 0.861 + ], + "angle": 0, + "content": "[110] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.862, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[111] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In Eu" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10347" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ropean conference on computer vision, pages 650-667. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.17 + ], + "angle": 0, + "content": "[112] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.172, + 0.469, + 0.211 + ], + "angle": 0, + "content": "[113] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10348" + } + ] +] \ No newline at end of file diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_origin.pdf b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..196f76f1c0483113cddfdfbc08d34d050c9d2ff4 --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/232fb53c-e6e5-4c0f-99d0-a2b74f05feaf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8672935f33c5fccf302f98384bf852f70daa9d7f3938164ff227a0b4bf5d6b1 +size 7271614 diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/full.md b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e744d9be03c229d2bf0734a24c9e9d67625e87e2 --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/full.md @@ -0,0 +1,374 @@ +# VOODOO 3D: Volumetric Portrait Disentanglement for One-Shot 3D Head Reenactment + +Phong Tran1 Egor Zakharov2 Long-Nhat Ho1 Anh Tuan Tran3 Liwen Hu4 Hao Li1,4 +1MBZUAI ETH Zurich 2VinAI Research 3Pinscreen +{the.tran, long.ho}@mbzuai.ac.ae anhtt152@vinai.io ezakharov@ethz.ch liwen@pinsscreen.com hao@hao-li.com + +![](images/a6c0e3214c7b78dceacfeb8013d7aef20ceb7d28032ac84ae7eb85de77de19ba.jpg) +Figure 1. We introduce VOODOO 3D: a high-fidelity 3D-aware one-shot head reenactment technique. Our method transfers the expression of a driver to a source and produces view consistent renderings for holographic displays. + +# Abstract + +We present a 3D-aware one-shot head reenactment method based on a fully volumetric neural disentanglement framework for source appearance and driver expressions. Our method is real-time and produces high-fidelity and view-consistent output, suitable for 3D teleconferencing systems based on holographic displays. Existing cutting-edge 3D-aware reenactment methods often use neural radiance fields or 3D meshes to produce view-consistent appearance encoding, but, at the same time, they rely on linear face models, such as 3DMM, to achieve its disentanglement with facial expressions. As a result, their reenactment results often exhibit identity leakage from the driver or have unnatural expressions. To address these problems, we propose a neural self-supervised disentanglement approach that lifts both the source image and driver video frame into a shared 3D volumetric representation based on tri-planes. This representation can then be freely manipulated with expression tri-planes extracted from the driving images and rendered from an arbitrary view using neural radiance fields. We achieve this disentanglement via self + +supervised learning on a large in-the-wild video dataset. We further introduce a highly effective fine-tuning approach to improve the generalizability of the 3D lifting using the same real-world data. We demonstrate state-of-the-art performance on a wide range of datasets, and also showcase high-quality 3D-aware head reenactment on highly challenging and diverse subjects, including non-frontal head poses and complex expressions for both source and driver. + +# 1. Introduction + +Creating 3D head avatars from a single photo is a core capability in making a wide range of consumer AR/VR and telepresence applications more accessible, and user experiences more engaging. Graphics engine-based 3D avatar digitization methods [9, 14, 33, 37, 46, 47, 49, 57] are suitable for today's video games and virtual worlds, and many commercial solutions exist (AvatarNeo [5], AvatarSDK [1], ReadyPlayerMe [6], in3D [2], etc.). However, the photorealism achieved by modern neural head reenactment tech- + +niques is becoming increasingly appealing for advanced effects in video sharing apps and visual effects. For immersive telepresence systems that use AR/VR headsets, facial expression capture is typically achieved using tiny video cameras built into HMDs [29, 50, 55, 64, 69], while the identity of the source subject recorded using a separate process. However, the teleconferencing solutions based on holographic 3D displays (LookingGlass [4], LEIA [3], etc.) use regular webcams [81] or depth sensors [48]. As opposed to a video-based setting, head reenactment for immersive applications needs to be 3D-aware, meaning that in addition to generating the correct poses and expressions from a photo, multi-view consistency is critical. + +While impressive facial reenactments results have been demonstrated using 2D approaches [26, 27, 83, 97, 102, 103], they typically struggle with preserving the likeness of the source and exhibit significant identity changes when varying the camera pose. More recently, 3D-aware one-shot head reenactment methods [36, 43, 53, 54, 59, 99] have used either 3D meshes or tri-plane neural radiance fields as a fast and memory efficient volumetric data representations for neural rendering. However, the expression and identity disentanglement in these methods is based on variants of linear face and expression models [15, 52] which lack expressiveness and high-frequency details. While these methods can achieve view consistency, facial expressions are often uncanny, and preserving the likeness of the input source portrait is challenging, especially for views different than the source image. Hence, input sources with extreme expressions and non-frontal poses are often avoided. + +In this paper, we introduce the first 3D aware one-shot head reenactment technique that disentangles source identities and the target expressions fully volumetrically, and without the use of explicit linear face models. Our method is real-time and designed with holographic displays in mind, where a large number of views (up to 45) can be rendered in parallel based on their viewing angle. We leverage the fact that real-time 3D lifting for human heads has recently been made possible [81] with the help of Vision Transformers (ViT) [25], which avoids the need for inefficient optimization-based GAN-inversion process [68]. In particular, 3D lifting allows us to map 2D face images into a canonical tri-plane representation for both source and target subjects and treat identity and expression disentanglement independently from the head pose. + +Once the source image and driver frame are lifted into a pose-normalized tri-plane representation, we extract appearance features from the source subject and expressions from the driver. The pose of the driver is estimated separately using a 3D face tracker and used as input to a neural renderer. Tri-plane-based feature extraction ensures view-consistent rendering, while facial appearance and driver expression feature use frontalized views from the 3D lifting + +to enable robust and high-fidelity facial disentanglement. To handle highly diverse portraits (variations in facial appearance, hairstyle, head covering, eyewear, etc.), we propose a new method for fine-tuning Lp3D on real datasets by introducing a mixed loss function based on real and synthetic datasets. Our volumetric disentanglement and rendering framework is trained only using in-the-wild videos from the CelebV-HQ dataset [112] in a self-supervised fashion. + +We not only demonstrate that our volumetric face disentanglement approach produces qualitative superior head reenactments than existing ones, but also show on a wide and diverse set of source images how non-frontal poses and extreme expressions can be handled. We have quantitatively assessed our method on multiple benchmarks and outperform existing 2D and 3D state-of-the-art techniques in terms of fidelity, expression, and likeness accuracy metrics. Our 3D aware head reenactment technique is therefore suitable for AR/VR-based immersive applications, and we also showcase a teleconferencing system using a holographic display from LookingGlass [4]. We summarize the main contributions as follows: + +- First fully volumetric disentanglement approach for real-time 3D aware head reenactment from a single photo. This method combines 3D lifting into a canonical triplane representation and formalized facial appearance and expression feature extraction. +- A 3D lifting network that is fine-tuned on unconstrained real-world data instead of only generating synthetic ones. +- We demonstrate superior fidelity, identity preservation, and robustness w.r.t. current state-of-the-art methods for facial reenactment on a wide range of public datasets. We plan to release our code to the public. + +# 2. Related Work + +2D Neural Head Reenactment. The problem of generating animations of photorealistic human heads given images or video inputs has been thoroughly explored using various neural rendering techniques in the past few years, outperforming traditional 3DMM-based methods [8, 26, 31, 44, 63, 66, 78, 79, 96] which often appear uncanny due to their compressed linear space. These approaches can be categorized into one-shot and multi-shot ones. While multishot methods generally achieve high-fidelity results, they are not suitable for many consumer applications as they typically require an extensive amount of training data, such as a monocular video capture [10, 11, 18, 21, 30, 34, 93, 108-110, 113], and sometimes even a calibrated multi-view stereo setup [13, 29, 55, 58, 69]. More recently, few-shot techniques [104] have also been introduced. + +To maximize accessibility, a considerable number of methods [17, 26, 27, 32, 35, 39, 66, 72-74, 76, 77, 83, 86, 87, 89, 97, 101-103, 107] use a single portrait as input by leveraging advanced generative modeling techniques based + +on in-the-wild video training data. While most methods rely on linear face models to extract facial expressions, the head reenactment technique from Drobyshev et al. [27] directly extract expression features from cropped 2D face regions, allowing them to obtain better face disentanglements, which results in higher fidelity face synthesis. While similar to our proposed approach in avoiding the use of low dimensional linear face models, their method is purely 2D and strugglesly with ensuring identity and expression consistency when novel views are synthesized. + +3D-Aware One-Shot Head Reenactment. Due to potential inconsistencies when rendering from different views or poses, a number of 3D-aware single shot head reenactment techniques [7, 19, 20, 24, 62, 65, 70, 75, 88, 92, 94, 95] have been introduced. These methods generally use an efficient 3D representation, such as neural radiance fields or 3D mesh, to geometrically constraint the neural rendering and improve view consistency. ROME [43] for instance is a mesh-based method using FLAME blendshapes [51] and neural textures. While view-consistent results can be produced for both face and hair regions, the use of low resolution polygonal meshes hinders the neural renderer to generate high-fidelity geometric and appearance details. + +Implicit representations such as HeadNeRF [36] and MofaNeRF [38] use a NeRF-based parametric model which supports direct control of the head pose of the generated images. While real-time rendering is possible, these methods require intensive test-time optimization and often fail to preserve the identity of the source due to the use of compact latent vectors. Most recent methods [53, 54, 99] adopt the highly efficient tri-plane-based neural fields representation [20] to encode the 3D structure and appearance of the avatars head. Compared to the previous works on view-consistent neural avatars [36, 43, 53, 54, 59, 99], we refrain from depending on parametric head models for motion synthesis and, instead, learn the volumetric motion model from the training data. This methodology enables us to narrow the identity gap between the source and generated images and yield a superior fidelity of the generated motion compared to competing approaches, and hence a higher quality disentanglement for reenactment. + +3D GAN Inversion. When training a whole reconstruction and disentangled reenactment model end-to-end on facial performance videos, one can introduce substantial overfitting and reduce the quality of the results. To address this problems, we focus our training approach to an inversion of pre-trained 3D-aware generative models for human heads. + +We use tri-plane-based generative network EG3D [20] as the foundational generator, due to its proficiency in producing high-fidelity and view-consistent synthesis of human heads. For a given image, an effective 3D GAN inver + +sion method should leverage these properties for estimating latent representations, which can be decoded into outputs that maintain view consistency and faithfully replicate the contents of the input. One naive approach is to adapt GAN inversion methods that were initially designed for 2D GANs to the EG3D pre-trained network. These methods either do a time consuming but more precise optimization [42, 68] or train a fast but less accurate encoder network [67, 80] to obtain the corresponding latent vectors. They often produce incorrect depth prediction, leading to clear artifacts in novel view synthesis. Hence, some methods are specifically designed for inverting 3D GANs, which either do multi-view optimization [45, 91] or predict residual features/tri-plane maps for refining the initial inversion results [12, 81, 98, 100]. + +In this work, we rely on the state-of-the-art EG3D inversion method Lp3D [81]. While achieving excellent novel-view synthesis results, it lacks disentanglement between the appearance and expression of the provided image and is unable to impose various driving expressions onto the input. To address this limitation, we propose a new method that introduces appearance-expression disentanglement in the latent space of tri-planes using our new self and cross-reenactment training pipeline while relying on a pre-trained but fine-tuned Lp3D network for regularization which enables highly consistent view synthesis. + +# 3. 3D-Aware Head Reenactment + +As illustrated in Fig. 2, our head reenactment pipeline consists of three stages: 1) 3D Lifting, 2) Volumetric Disentanglement, and 3) Tri-plane Rendering. Given a pair of source and driver images, we first frontalize them using a pre-trained but fine-tuned tri-plane-based 3D lifting module [81]. This driver alignment step is crucial and allows our model to disentangle the expressions from the head pose, which prevents overfitting. Then, the frontalized faces are fed into two separate convolutional encoders to extract the face features $F_{s}$ and $F_{d}$ . These extracted features are concatenated with the ones extracted from the tri-planes of the source, and all are fed together into several transformer blocks [90] to produce the expression tri-plane residual, which is added to the tri-planes of the source image. The final target image can be rendered from the new tri-planes using a pre-trained tri-plane renderer using the driver's pose. + +# 3.1. Fine-Tuned 3D Lifting + +We adopt Lp3d [81] as a 3D face-lifting module, which predicts the radiance field of any given face image in real-time. Instead of using an implicit multi-layer perceptron [61] or sparse voxels [28, 71] for the radiance field, Lp3D [81] uses tri-planes [20], which can be computed using a single forward of a deep learning network. Specifically, for a given source image $x_{s}$ , we first extract the tri-planes $T$ using a + +![](images/5267cdbdd612a796023ff7ca70a6eeefd6865cc1442abf601c03b118b3eab723.jpg) +Figure 2. Given a pair of source and driver images, our method processes them in three steps: 3D Lifting into tri-plane representations, Volumetric Disentanglement, which consists of source and driver frontalization and tri-plane residual generation, and Tri-plane Rendering via volumetric ray marching with subsequent super-resolution. + +transformer-based appearance encoder $\mathbf{E}_{\mathrm{app}}$ .. + +$$ +\mathbf {E} _ {\text {a p p}} \left(x _ {s}\right) = T \in \mathbb {R} ^ {3 \times H \times W \times C} = \left\{T _ {x y}, T _ {y z}, T _ {z x} \right\}. \tag {1} +$$ + +The color $c$ and density $\sigma$ of each point $p = (x, y, z)$ in the radiance field can be obtained by projecting $p$ onto the three planes and by summing up the features at the projected positions: + +$$ +c, \sigma = \mathbf {D} \left(F _ {x y} + F _ {y z} + F _ {z x}\right), \tag {2} +$$ + +where $\mathbf{D}$ is a shallow MLP decoder for the tri-plane rendering, $F_{xy}, F_{yz}$ , and $F_{zx}$ are the feature vectors at the projected positions on $xy$ , $yz$ , and $zx$ planes, respectively, calculated using bilinear interpolation. The rendered $128 \times 128$ image is then upsampled using a super-resolution module to produce a high-resolution output. To train the encoder $\mathbf{E}_{\mathrm{app}}$ , Lp3D [81] uses synthetic data generated from a 3D-aware face generative model [20]. While these synthetic data have ground truth camera poses, they are limited to the face distribution of the generative model. As a result, Lp3D can fail to generalize to in-the-wild images as shown in Fig. 5. To prevent this, we fine-tune the pre-trained Lp3D on a large-scale real-world dataset. We also replace the original super-resolution module in Lp3D with a pre-trained GF-PGAN [85], which is then fine-tuned together with Lp3D (see Sec. 3.4). + +# 3.2. Disentangling Appearance and Expression + +Separating facial expression from the identity appearance in a 3D radiance field is very challenging especially when source and driver subjects have misaligned expressions. In order to simplify the problem, we use our 3D lifting approach to bring both source and driver heads into a pos-oriented space where faces are frontalized. Here, we denote frontalized source and driver images as $x_{s}^{f}$ and $x_{d}^{f}$ , respectively. These images are then fed into two separate convolutional source and driver encoders $\mathbf{E}_s$ and $\mathbf{E}_d$ to produce + +coarse feature maps $F_{s} = \mathbf{E}_{s}(x_{s}^{f})$ and $F_{d} = \mathbf{E}_{d}(x_{d}^{f})$ . Since we already have the source's tri-plane, which encodes the 3D shape of the source, we use another encoder to encode this tri-plane and concatenate it together with the coarse frontalized feature maps of the images to produce expression feature $F$ : + +$$ +F _ {t} = \mathbf {E} _ {t} (T) +$$ + +$$ +F = F _ {s} \oplus F _ {d} \oplus F _ {t} +$$ + +Even though face frontalization aligns the source and the driver, there is still some misalignment between the two faces, e.g., the positions of the eyes may be different, or one mouth is open while the other is closed. Therefore, we feed the concatenation of the feature maps into several transformer blocks to produce the final residual tri-plane $\mathbf{E}_v(F)$ . This residual is then added back to the source's tri-planes to change the source's expression to the driver's expression $T' = T + \mathbf{E}_v(F)$ . Unlike LPR [54], we do not use a 3D face model to compute the expression but instead use the RGB images of the source and the driver directly, allowing the model to learn high-fidelity and realistic expressions. + +# 3.3. Tri-Plane Rendering + +The resulting tri-planes are then volumetrically rendered into one or multiple output images using pose parameters and viewing angles in the case of a holographic display. Following EG3D [20], we use a neural radiance fields (NeRFs)-based volumetric ray marching approach [60]. However, instead of encoding each point in space via positional encodings [60], the features of the points along rays are calculated using their projections onto tri-planes. Since tri-planes are aligned with the frontal face, we can compute these rays directly using camera extrinsics $\mathbf{P}_{\mathrm{driver}}$ predicted by an off-the-shelf 3D head pose estimator [23]. + +While the renderings are highly view-consistent, the large number of points evaluated for each ray still limits + +the output resolution for real-time performance. We therefore follow [54] and employ a 2D upsampling network [84] based on StyleGAN2 [41], which in our experiments produced higher quality results than the upsampling approach in EG3D [20]. Finally, for holographic displays, we generate a number of renderings based on their viewing angles and simply using the head pose parameter. Real-time performance is achieved using efficient inference libraries such as TensorRT, half-precision, and batched inference over multiple GPUs. + +# 3.4. Training Strategy + +Fine-Tuning Lp3D. To make Lp3D work with in-the-wild images, we fine-tune it on a large-scale real-world video dataset [111]. Unlike the use of synthetic data, real-world data do not have ground-truth camera parameters and facial expressions in monocular videos are typically inconsistent over time. While the camera parameters can be estimated using standard 3D pose estimators, the expression differences are difficult to determine. However, we found that we can ignore this expression difference and fine-tune Lp3D using real data together with continuous training on synthetic data. In particular, our experiments indicate that the fine-tuned model can still faithfully reconstruct 3D faces from the input without changing expressions and still generalize successfully on in-the-wild images. Specifically, on real video data, we sample two frames $x_{s}^{r}$ and $x_{d}^{r}$ and estimate their camera parameters $P_{s}^{r}$ and $P_{d}^{r}$ . Similar to [20], we assume a fixed intrinsic for standard portraits for all images. Then we use $E_{app}$ from Lp3D to calculate the triplanes of $x_{s}^{r}$ , render it using the two poses, and calculate reconstruction losses on the two rendered images: + +$$ +\mathcal {L} _ {\text {r e a l}} = \left\| \mathrm {L p 3 D} \left(x _ {s} ^ {r}, P _ {d} ^ {r}\right) - x _ {d} ^ {r} \right\| + \left\| \mathrm {L p 3 D} \left(x _ {s} ^ {r}, x _ {s} ^ {r}\right) - x _ {s} ^ {r} \right\|, +$$ + +where $\mathrm{Lp3D}(x,P)$ is the face in $x$ re-rendered using camera pose $P$ and $\mathcal{L}_{\mathrm{real}}$ is the loss for real images. Simultaneously, we render two synthetic images employing an identical latent code but through varying camera views and calculate the synthetic loss $\mathcal{L}_{\mathrm{syn}}$ : + +$$ +\mathcal {L} _ {\mathrm {s y n}} = \left\| \mathrm {L p 3 D} \left(x _ {s} ^ {f}, P _ {d} ^ {s}\right) - x _ {d} ^ {s} \right\| + \left\| \mathrm {L p 3 D} \left(x _ {s} ^ {s}, P _ {s} ^ {s}\right) - x _ {s} ^ {s} \right\| +$$ + +$$ +\mathcal {L} _ {\mathrm {t r i}} = \left\| E _ {\mathrm {a p p}} \left(x _ {s} ^ {f}\right) - T \right\|, +$$ + +where $T$ is the ground-truth tri-planes returned by EG3D [20] and $\mathcal{L}_{\mathrm{tri}}$ is the tri-plane loss adopted directly from Lp3D. The final loss $\mathcal{L}_{\mathrm{app}}$ for fine-tuning Lp3D can be formulated as: + +$$ +\mathcal {L} _ {\text {a p p}} = \mathcal {L} _ {\text {r e a l}} + \lambda_ {\text {s y n}} \mathcal {L} _ {\text {s y n}} + \lambda_ {\text {t r i}} \mathcal {L} _ {\text {t r i}} +$$ + +where $\lambda_{\mathrm{syn}}$ and $\lambda_{\mathrm{tri}}$ are tunable hyperparameters. + +Disentangling Appearance and Expressions. In this stage, we also use real-world videos as training data. For a + +pair of source and driver images $x_{s}$ and $x_{d}$ sampled from the same video, we apply the reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ which is a combination of $L1$ , perceptual [105], and identity losses, between the reenacted image $x_{s \to d}$ and the corresponding ground-truth $x_{d}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e c o n}} = \left\| x _ {s \rightarrow d} - x _ {d} \right\| _ {1} + \phi \left(x _ {s \rightarrow d}, x _ {d}\right) \\ + \mathcal {L} _ {\cos} \left(\mathrm {I D} \left(x _ {s \rightarrow d}\right), \mathrm {I D} \left(x _ {d}\right)\right), \\ \end{array} +$$ + +where $\phi$ is the perceptual loss, $\mathrm{ID}(\cdot)$ is a pretrained face recognizer [82], and $\mathcal{L}_{\mathrm{cos}}$ is the cosine similarity loss: + +$$ +\mathcal {L} _ {\mathrm {c o s}} ((u, v) = \frac {< u , v >}{m a x (\| u \| , \epsilon) \times m a x (\| v \| , \epsilon)} +$$ + +Here $\epsilon$ is a small constant to void division by 0. Similar to other works that use RGB images directly to calculate expressions [27], our proposed encoder also suffers from an "identity leaking" issue. Since there is no cross-reenactment dataset, the expression module is trained with self-reenactment video data. Therefore, without proper augmentation and regularization, the expression module can leak identity information from the driver to the output, making the model fail to generalize to cross-reenactment tasks. Hence, we introduce a Cross Identity Regularization. Specifically, we further sample an additional driver frame $x_{d'}$ from another video. We incorporate a GAN loss where real samples are $\mathrm{Lp3D}(x_s, P^d)$ and fake samples are $x_{s \to d'}$ . This GAN loss is also conditioned on the identity vector of the source $\mathrm{ID}(x_s)$ . Following [27], we also apply strong augmentation (random warping and color jittering) and additionally mask the border of the driver randomly to further reduce potential identity leaks. The loss for expression training can be summarized as: + +$$ +\mathcal {L} _ {\exp} = \mathcal {L} _ {\text {r e c o n}} + \lambda_ {\text {C I R}} \mathcal {L} _ {\text {C I R}}, +$$ + +where $\mathcal{L}_{\mathrm{CIR}}$ and $\lambda_{\mathrm{CIR}}$ are cross identity regularization and its hyperparameter, respectively. + +Global Fine-Tuning. After training both Lp3D and the expression module, we iteratively fine-tune the two modules using the same losses as the previous sections. Specifically, for every 10000 iterations, we freeze one module and fine-tune the other and vice versa. In addition, we add a GAN loss on the super-resolution output of the Lp3D module. + +# 4. Experiments + +Implementation Details. We train our model on CelebV-HQ dataset [112] using 7 NVIDIA RTX A6000 ADA (50Gb memory each). We use AdamW [56] to optimize the parameters with a learning rate of $10^{-4}$ and batch size of 28. The Lp3D finetuning takes 5 days for 500K iterations to converge. Training the expression module takes 2 days, and the + +![](images/3f513c2a949c0db7a267ecb66890172cd06d9528b0c93469c943e9d0e6036d37.jpg) + +![](images/e301983bfe1ea85c7eceaad90b91f3aaa3a783d15cb18adde325e44049b5b785.jpg) + +![](images/dfa8277f2d88edd09be2a72efe590400755ce2a6a1e75f843093c225297f1add.jpg) + +![](images/0e0d254cf5d9713a9da084dec1a7e43bcdee6b84a1467afcd5d39fc57fb5d5d6.jpg) + +![](images/c1b6fc341f84b10ed1d198f7b8b98111911296fc8472fbb905b04f3a75ade98d.jpg) + +![](images/2cc95820c5f0783680b1ad4d1f50219473b6bd20900ebc95901fb166dc8c04ec.jpg) + +![](images/a617a6e81ba29d544c1cf20bc04bad8ab5ba62f282df4dfa3a2d97b8efbb4683.jpg) + +![](images/82b1a70b4f7a25d4dcc48c777800da2b5dc58196b807796ebd80e58942c7962a.jpg) +Source + +![](images/c2a4ae51a7174dc3c69582b1be432a1c223a6c39692faf3dc0afd033e49daabe.jpg) +Driver + +![](images/c77cc9da8ba590134d7e9810596f7c7a1faf7a6bba3d162cd72d1ac8ac6e4a52.jpg) +Ref View + +![](images/fbe6ec14fed6fbe1715765fdb3e921b4392f2b83ce9bf5bbe2fd8fc28495a233.jpg) +Novel View + +![](images/483934e4ed20d13d76b74fa22756740634daf3c1d931d198638a0be8e76550c4.jpg) +Wrinkles + +![](images/2af9511cf063f08dc26d5f9d7907c065f5d0823dba0353f6931b4864f96bcd12.jpg) +Geometry + +![](images/c084db6949b657f0c5fc4f6df5a478a3a06409c46b84be45156a7bec4905fe1a.jpg) +Figure 3. Expression dependent high-fidelity details, incl. eye and forehead wrinkles, as well as nasolabial folds (see zoom-ins) +Figure 4. A qualitative comparison with the baselines on in-the-wild photos. Notice that our method is capable of producing a variety of facial expressions, and handle highly diverse subjects, with and without accessories, as well as extreme head poses, such as rows 3 and 4. + +
MethodSelf-reenactmentCross-reenactment
PSNR ↑SSIM ↑LPIPS ↓NAKD ↓ECMD ↓FID ↓CSIM ↑ECMD ↓FID ↓
ROME [43]18.460.4880.3510.0300.5941380.5070.740172
StyleHeat [97]19.730.6890.2780.0350.74889.80.3980.74495.5
OTAvatar [59]19.280.7490.2890.0350.65167.00.4620.90172.4
MegaPortraits [27]21.100.7310.2910.0220.75552.00.7290.77161.7
Ours22.830.7680.1680.0120.42640.50.7540.75436.4
+ +Table 1. Evaluation on HDTF [106] dataset. Our method outperforms the competitors across almost all of the metrics for both self- and cross-reenactment scenarios. + +
MethodCross-reenactment
CSIM ↑ECMD ↓FID ↓
ROME [43]0.5190.9152.6
HeadNeRF [36]0.3460.88113
StyleHeat [97]0.4670.8550.2
MegaPortraits [27]0.6470.7729.2
Ours0.6080.7923.6
+ +![](images/9127b5160d117f322ec40bd912a6a4c7fee3ea2570ab6fc5683e6bd8a64807a6.jpg) +Figure 5. Our implementation of Lp3D [81] before and after CelebV-HQ [112] fine-tuning. + +Table 2. Evaluation on CelebA-HQ [40] dataset. + +
CSIM ↑ECMD ↓
Lp3D0.5480.82
Lp3D-FT0.6700.76
w/o frontal0.6681.01
w/o CIR0.5700.97
Ours0.6080.79
+ +Table 3. Ablation studies conducted on CelebA-HQ [40] dataset. FT is a fine-tuned version of Lp3D, and "frontal" denotes frontalization of the source and driver. + +![](images/45d7c05ab06300c43cd4eb787b4fe219213b251900e5e6adab50625e5c74508f.jpg) +Figure 6. Ablation study for source and driver frontalization and cross identity regularization (CIR). + +![](images/27fd866d1831199fb47c81b48ca96773e7cfc37b6b76051393e9a9da4ed6f5cd.jpg) +Figure 7. Qualitative comparison with LPR [54] method on the samples from HDTF [106] dataset. + +iterative fine-tuning takes another 5 days. More training details, such as hyperparameter fine-tuning or architecture of the networks, can be found in the supplementary materials. + +Unlike Lp3D, our method reenacts faces without relifting in 3D for every frame. The tri-plane encoding and frontalization of the source image are only performed once. For each driver, we perform a single frontalization for the driver image (0.0115 ms), one expression encoding (0.0034 s), one tri-plane rendering at $128 \times 128$ resolution (0.0071 s), and one neural upsampling (0.0099 s). Each view runs at 31.9 fps on an Nvidia RTX 4090 GPU including I/O. More details are provided in the supplemental materials. + +We compare our method with state-of-the-art 3D-based [36, 43, 59] and 2D-based [27, 97] models. For MegaPortraits [27], we use our own implementation that was trained on the CelebV-HQ dataset. Similar to previous works, we evaluate our method using public benchmarks, including CelebA-HQ [40] and HDTF [106]. For CelebAHQ, we split the data into two equal sets. Each set contains around $15\mathrm{K}$ images. Then, we use one set as the source and the rest as driver images. For the HDTF dataset, we perform cross-reenactment by using the first frame of each video as source and 200 first frames of other videos as drivers, which is more than $60\mathrm{K}$ data pairs. Similarly, to evaluate self-reenactment, we also use the first frames of each video as sources and the rest of the same video as the driver. Furthermore, we also collected 100 face images on the internet and around 100 high-quality videos for qualitative comparisons. Video results are provided in the supplementary materials. + +Evaluation. Given a source image $x_{s}$ , a driver image $x_{d}$ and reenacted output $x_{s\rightarrow d}$ we use EMOCAv2 [22] to extract the FLAME [52] expression coefficients of the prediction and the driver, as well as the shape coefficients of the source. We then compute 2 FLAME meshes using the predicted shape coefficients in world coordinates, one with the expression coefficients of the driver and one with the expression coefficients of the reenacted output. We measure the distance between the 2 meshes and denote this expression metric as ECMD. Moreover, we also use cosine similarity between the embeddings of a face recognition network (CSIM) [101], normalized average keypoint distance (NAKD) [16], perceptual image similarity (LPIPS) [105], peak signal-to-noise ratio (PSNR), and structure similarity index measure (SSIM). + +We provide quantitative comparisons on HDTF and CelebA-HQ datasets in Tab. 1 and Tab. 2, respectively, and show that our method outperforms existing methods on both datasets. We also note that our FID and CSIM scores are significantly more reliable than the others, while expression-based metrics such as NAKD and ECMD are either better or very close to the best baseline, w.r.t output quality, expression accuracy, and identity consistency. + +We showcase the qualitative results of cross-identity reenactment on in-the-wild images in Fig. 4 and Fig. 3. Compared to the baselines [27, 36, 43, 97], our reenactment faithfully reconstructs intricate and complex elements, + +![](images/8ca4e4feee8c33ed28bd87b7574a7ac5634c1adec84b3c3c545194688e98973d.jpg) +Figure 8. Failure cases of our method include side views in the source, extreme expressions, modeling of cartoonish characters and paintings, as well as modeling the reflections and semitransparency of the eyewear. + +such as hairstyle, facial hair, glasses, and facial makeup. Furthermore, our method effectively generates realistic and fine-scale dynamic details that match the driver's expressions including substantial head pose rotations. We also conduct a comparative analysis of our results with the current state-of-the-art 3D-aware method LPR [54] in Fig. 7. Compared to LPR, our method achieves superior identity consistency. We further refer to the supplemental video for a live demonstration of our holographic telepresence system and animated head reenactment results and comparisons, with and without disentangled poses. + +Ablation Study. We compare Lp3D with and without fine-tuning on the CelebA-HQ dataset in Tab. 3 and show several examples in Fig. 5. Without fine-tuning on real data, our implementation of Lp3D fails to preserve the identity of the input image, resulting in a considerably lower CSIM score. We also try without any facial frontalization in the expression module and instead use the source and driver images directly to calculate the expression tri-plane residual. We observe in Fig. 6 that without face frontalization, the model completely ignores the expression of the driver and keeps the expression of the input source instead. We show in Tab. 3, that facial frontalization leads to much better ECMD score. We then measure the effectiveness of the GAN-based cross-identity regularization on the CelebA-HQ dataset, $\mathcal{L}_{\mathrm{CIR}}$ . Without this loss, identity characteristics (hairstyle or color) can leak from the driver to the output. See column 4 in Fig. 6. Tab. 3 also shows that cross-identity regularization can reduce identity leaking and improve the CSIM score. Lastly, we have also attempted to train our model end-to-end using the same losses and optimization process instead of our proposed iterative fine-tuning. Even with a lower learning rate and the use of pre-trained Lp3D weights, we were unable to succeed. + +Limitations. Limitations of our approach are illustrated in Figure 8. For source images that are extremely side ways (i.e., over $90^{\circ}$ ), our method can produce a plausible frontal face, but the likeness cannot be guaranteed due to insufficient visibility. For very highly stylized portraits, such as cartoons, our framework often produces photorealistic facial elements such as teeth which can be inconsistent in + +style. Due to the dependence on training data volume and diversity, accessories such as dental braces or glasses may disappear or look different during synthesis. We believe that providing more and better training data can further improve the performance of our algorithm. + +# 5. Discussion + +We have demonstrated that a fully volumetric disentanglement of facial appearance and expressions is possible through a shared canonical tri-plane representation. In particular, an improved disentanglement also leads to higher fidelity and more robust head reenactment, when compared to existing methods that use linear face models for expressions, especially for non-frontal poses. A critical insight of our approach is that head frontalization via 3D lifting is particularly effective for extracting features that can encode fine details and expressions such as wrinkles and folds. The resulting reenactment is also highly view-consistent for large angles, making our solution suitable for holographic displays. We have also shown that the 3D lifting model can still be successfully trained with real data despite the fact that different frames with the same subject have varying facial expressions. Without a fine-tuned 3D lifting model, our 3D-aware reenactment framework would struggle with preserving the identity of the source, especially for side views. Our experiments indicate that our results achieve better visual quality and are more robust to extreme poses, validated via an extensive evaluation on multiple datasets. + +Risks and Potential Misuse. The proposed method is intended to promote avatar-based 3D communication. Nevertheless, our AI-based reenactment solution produces synthetic but highly realistic face videos from only a single photo, which could be hard to distinguish from a real person. Like deepfakes and other facial manipulation methods, potential misuse is possible and hence, we refer to the supplemental material for more discussions. + +Future Work. We are also interested in expanding our work to upper and full body reenactment, where hand gestures can be used for more engaging communication. To this end, we plan to investigate the use of canonical representations for human bodies, such as T-poses. As our primary motivation, we have showcased a solution using holographic displays for immersive 3D teleconferencing. However, we believe that our approach can also be extended to AR/VR HMD-based settings where full $360^{\circ}$ head views are possible. The recent work by An et al. [7] is a promising avenue for future exploration. + +Acknowledgement. This work is supported by the Meta-verse Center Grant from the MBZUAI Research Office and by the Swiss SERI Consolidation Grant "AI-PERCEIVE". + +# References + +[1] ItSeez3D AvatarSDK, https://avatarsdk.com. 1 +[2] in3D, https://in3d.io.1 +[3] Leia, https://www.leiainc.com. 2 +[4] Looking Glass Factory, https://lookingglassfactory.com. 2 +[5] Pinscreen Avatar Neo, https://www.Avatarneo.com. 1 +[6] ReadyPlayerMe, https://readyplayer.me. 1 +[7] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 8 +[8] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20364-20373, 2022. 2 +[9] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 362–371, 2023. 1 +[10] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhinitra Meka, Ruofei Du, Ming-song Dou, Sergio Orts-Escalano, Rohit Pandey, Ping Tan, Thabo Beeler, Sean Fanello, and Yinda Zhang. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[11] Shrisha Bharadwaj, Yufeng Zheng, Otmar Hilliges, Michael J. Black, and Victoria Fernandez Abrevaya. FLARE: Fast learning of animatable and relightable mesh avatars. ACM Transactions on Graphics, 42:15, 2023. 2 +[12] Ananta R Bhattacharai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. arXiv preprint arXiv:2303.13497, 2023. 3 +[13] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason M. Saragih. Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG), 40, 2021. 2 +[14] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 1 +[15] Volker Blanz and Thomas Vetter. A Morphable Model For The Synthesis Of 3D Faces. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 2 +[16] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 7 +[17] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2 + +[18] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41, 2022. 2 +[19] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 3 +[20] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3, 4, 5 +[21] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 416-426, 2023. 2 +[22] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 7 +[23] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 4 +[24] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022. 3 +[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2 +[26] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[27] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2, 3, 5, 7 +[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3 +[29] Yonggan Fu, Yuecheng Li, Chenghui Li, Jason Saragih, Peizhao Zhang, Xiaoliang Dai, and Yingyan (Celine) Lin. Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21036-21045, 2023. 2 +[30] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2 +[31] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2021. 2 +[32] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5609-5619, 2023. 2 +[33] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos P Zafeiriou. Fast-ganfit: Generative adversarial network for high fidelity 3d face reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1 +[34] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18653–18664, 2022. 2 +[35] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2 +[36] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2, 3, 7 +[37] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 1 +[38] Yiyu huang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 3 +[39] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2 +[40] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 7 +[41] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[42] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[43] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference of Computer vision (ECCV), 2022. 2, 3, 7 + +[44] H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Nießner, P. Pérez, C. Richardt, M. Zollhöfer, and C. Theobalt. Deep video portraits. ACM Transactions on Graphics 2018 (TOG), 2018. 2 +[45] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2967-2976, 2023. 3 +[46] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Abhijeet Ghosh, and Stefanos P Zafeiriou. Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1 +[47] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8629-8640, 2023. 1 +[48] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 40(6), 2021. 2 +[49] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 394-403, 2023. 1 +[50] Hao Li, Laura Trutoiu, Kyle Olszewski, Lingyu Wei, Tristan Trutna, Pei-Lun Hsieh, Aaron Nicholls, and Chongyang Ma. Facial performance sensing head-mounted display. ACM Transactions on Graphics (Proceedings SIGGRAPH 2015), 34(4), 2015. 2 +[51] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3 +[52] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7 +[53] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17969-17978, 2023. 2, 3 +[54] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. arXiv preprint arXiv:2306.08768, 2023. 2, 3, 4, 5, 7, 8 +[55] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4), 2018. 2 + +[56] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5 +[57] Huiwen Luo, Koki Nagano, Han-Wei Kung, Mclean Gold-white, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. CoRR, abs/2106.11423, 2021. 1 +[58] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De la Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 2 +[59] Zhiyuan Ma, Xiangyu Zhu, Guo-Jun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16910, 2023. 2, 3, 7 +[60] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 4 +[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[62] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021. 3 +[63] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2 +[64] Kyle Olszewski, Joseph J. Lim, Shunsuke Saito, and Hao Li. High-fidelity facial and speech animation for vr hmds. ACM Transactions on Graphics (TOG), 35:1-14, 2016. 2 +[65] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 3 +[66] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13759-13768, 2021. 2 +[67] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[68] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3 +[69] Gabriel Schwartz, Shih-En Wei, Te-Li Wang, Stephen Lombardi, Tomas Simon, Jason Saragih, and Yaser Sheikh. The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Trans. Graph., 39(4), 2020. 2 + +[70] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020. 3 +[71] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 3 +[72] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2 +[73] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Conference on Neural Information Processing Systems (NeurIPS), 2019. +[74] Aliaksandr Siarohin, Oliver Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In CVPR, 2021. 2 +[75] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d GANs. In Advances in Neural Information Processing Systems, 2022. 3 +[76] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Pareidolia face reenactment. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[77] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3637-3646, 2022. 2 +[78] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2016. 2 +[79] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Headon: Real-time reenactment of human portrait videos. ACM Transactions on Graphics 2018 (TOG), 2018. 2 +[80] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40 (4):1-14, 2021. 3 +[81] Alex Trevithick, Matthew Chan, Michael Stengel, Eric Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 2, 3, 4, 7 +[82] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5 +[83] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2 +[84] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 + +[85] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168-9178, 2021. 4 +[86] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022. 2 +[87] O. Wiles, A.S. Koepke, and A. Zisserman. X2face: A network for controlling face generation by using images, audio, and pose codes. In European Conference on Computer Vision, 2018. 2 +[88] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2195-2205, 2023. 3 +[89] Sitao Xiang, Yuming Gu, Pengda Xiang, Mingming He, Koki Nagano, Haiwei Chen, and Hao Li. One-shot identity-preserving portrait reenactment. arXiv preprint arXiv:2004.12452, 2020. 2 +[90] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3 +[91] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2023. 3 +[92] Hongyi Xu, Guoxian Song, Zihang Jiang, Jianfeng Zhang, Yichun Shi, Jing Liu, Wanchun Ma, Jiashi Feng, and Linjie Luo. Omniavatar: Geometry-guided controllable 3d head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12814-12824, 2023. 3 +[93] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings. Association for Computing Machinery, 2023. 2 +[94] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3 +[95] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18440-18449, 2022. 3 +[96] Kewei Yang, Kang Chen, Daoliang Guo, Song-Hai Zhang, Yuan-Chen Guo, and Weidong Zhang. Face2face $\rho$ : Real-time high-resolution one-shot face reenactment. 2022. 2 +[97] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In ECCV, 2022. 2, 7 +[98] Yu Yin, Kamran Ghasedi, HsiangTao Wu, Jiaolong Yang, Xin Tong, and Yun Fu. Nerfinvertor: High fidelity nerf-gan + +inversion for single-shot real image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8539-8548, 2023. 3 +[99] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based oneshot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 3 +[100] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. arXiv preprint arXiv:2303.12326, 2023. 3 +[101] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2, 7 +[102] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 2 +[103] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaprotait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22096-22105, 2023. 2 +[104] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022. 2 +[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7 +[106] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7 +[107] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, pages 3657–3666, 2022. 2 +[108] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2 +[109] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13545-13555, 2022. +[110] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[111] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In Eu + +ropean conference on computer vision, pages 650-667. Springer, 2022. 5 +[112] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 5, 7 +[113] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 \ No newline at end of file diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/images.zip b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1936318d452b32dc8d5188d8f608ee2424e1dd57 --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8ea34b1a7366110c4606deda51c97edcf54ea81799973e0cc2b2ff64aa7ab17 +size 624634 diff --git a/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/layout.json b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..489d29216cba4aa7b171b25369aa8036f8eebac2 --- /dev/null +++ b/2024/VOODOO 3D_ Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment/layout.json @@ -0,0 +1,10295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 471, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 471, + 138 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 471, + 138 + ], + "type": "text", + "content": "VOODOO 3D: Volumetric Portrait Disentanglement for One-Shot 3D Head Reenactment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 160, + 527, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 160, + 527, + 218 + ], + "spans": [ + { + "bbox": [ + 66, + 160, + 527, + 218 + ], + "type": "text", + "content": "Phong Tran1 Egor Zakharov2 Long-Nhat Ho1 Anh Tuan Tran3 Liwen Hu4 Hao Li1,4 \n1MBZUAI ETH Zurich 2VinAI Research 3Pinscreen \n{the.tran, long.ho}@mbzuai.ac.ae anhtt152@vinai.io ezakharov@ethz.ch liwen@pinsscreen.com hao@hao-li.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 72, + 245, + 520, + 424 + ], + "blocks": [ + { + "bbox": [ + 72, + 245, + 520, + 424 + ], + "lines": [ + { + "bbox": [ + 72, + 245, + 520, + 424 + ], + "spans": [ + { + "bbox": [ + 72, + 245, + 520, + 424 + ], + "type": "image", + "image_path": "a6c0e3214c7b78dceacfeb8013d7aef20ceb7d28032ac84ae7eb85de77de19ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 426, + 504, + 448 + ], + "lines": [ + { + "bbox": [ + 87, + 426, + 504, + 448 + ], + "spans": [ + { + "bbox": [ + 87, + 426, + 504, + 448 + ], + "type": "text", + "content": "Figure 1. We introduce VOODOO 3D: a high-fidelity 3D-aware one-shot head reenactment technique. Our method transfers the expression of a driver to a source and produces view consistent renderings for holographic displays." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 468, + 192, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 468, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 143, + 468, + 192, + 480 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 492, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 492, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 492, + 290, + 715 + ], + "type": "text", + "content": "We present a 3D-aware one-shot head reenactment method based on a fully volumetric neural disentanglement framework for source appearance and driver expressions. Our method is real-time and produces high-fidelity and view-consistent output, suitable for 3D teleconferencing systems based on holographic displays. Existing cutting-edge 3D-aware reenactment methods often use neural radiance fields or 3D meshes to produce view-consistent appearance encoding, but, at the same time, they rely on linear face models, such as 3DMM, to achieve its disentanglement with facial expressions. As a result, their reenactment results often exhibit identity leakage from the driver or have unnatural expressions. To address these problems, we propose a neural self-supervised disentanglement approach that lifts both the source image and driver video frame into a shared 3D volumetric representation based on tri-planes. This representation can then be freely manipulated with expression tri-planes extracted from the driving images and rendered from an arbitrary view using neural radiance fields. We achieve this disentanglement via self" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 469, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 559 + ], + "type": "text", + "content": "supervised learning on a large in-the-wild video dataset. We further introduce a highly effective fine-tuning approach to improve the generalizability of the 3D lifting using the same real-world data. We demonstrate state-of-the-art performance on a wide range of datasets, and also showcase high-quality 3D-aware head reenactment on highly challenging and diverse subjects, including non-frontal head poses and complex expressions for both source and driver." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 584, + 387, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 387, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 387, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Creating 3D head avatars from a single photo is a core capability in making a wide range of consumer AR/VR and telepresence applications more accessible, and user experiences more engaging. Graphics engine-based 3D avatar digitization methods [9, 14, 33, 37, 46, 47, 49, 57] are suitable for today's video games and virtual worlds, and many commercial solutions exist (AvatarNeo [5], AvatarSDK [1], ReadyPlayerMe [6], in3D [2], etc.). However, the photorealism achieved by modern neural head reenactment tech-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10336" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "niques is becoming increasingly appealing for advanced effects in video sharing apps and visual effects. For immersive telepresence systems that use AR/VR headsets, facial expression capture is typically achieved using tiny video cameras built into HMDs [29, 50, 55, 64, 69], while the identity of the source subject recorded using a separate process. However, the teleconferencing solutions based on holographic 3D displays (LookingGlass [4], LEIA [3], etc.) use regular webcams [81] or depth sensors [48]. As opposed to a video-based setting, head reenactment for immersive applications needs to be 3D-aware, meaning that in addition to generating the correct poses and expressions from a photo, multi-view consistency is critical." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 434 + ], + "type": "text", + "content": "While impressive facial reenactments results have been demonstrated using 2D approaches [26, 27, 83, 97, 102, 103], they typically struggle with preserving the likeness of the source and exhibit significant identity changes when varying the camera pose. More recently, 3D-aware one-shot head reenactment methods [36, 43, 53, 54, 59, 99] have used either 3D meshes or tri-plane neural radiance fields as a fast and memory efficient volumetric data representations for neural rendering. However, the expression and identity disentanglement in these methods is based on variants of linear face and expression models [15, 52] which lack expressiveness and high-frequency details. While these methods can achieve view consistency, facial expressions are often uncanny, and preserving the likeness of the input source portrait is challenging, especially for views different than the source image. Hence, input sources with extreme expressions and non-frontal poses are often avoided." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 436, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 616 + ], + "type": "text", + "content": "In this paper, we introduce the first 3D aware one-shot head reenactment technique that disentangles source identities and the target expressions fully volumetrically, and without the use of explicit linear face models. Our method is real-time and designed with holographic displays in mind, where a large number of views (up to 45) can be rendered in parallel based on their viewing angle. We leverage the fact that real-time 3D lifting for human heads has recently been made possible [81] with the help of Vision Transformers (ViT) [25], which avoids the need for inefficient optimization-based GAN-inversion process [68]. In particular, 3D lifting allows us to map 2D face images into a canonical tri-plane representation for both source and target subjects and treat identity and expression disentanglement independently from the head pose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Once the source image and driver frame are lifted into a pose-normalized tri-plane representation, we extract appearance features from the source subject and expressions from the driver. The pose of the driver is estimated separately using a 3D face tracker and used as input to a neural renderer. Tri-plane-based feature extraction ensures view-consistent rendering, while facial appearance and driver expression feature use frontalized views from the 3D lifting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "to enable robust and high-fidelity facial disentanglement. To handle highly diverse portraits (variations in facial appearance, hairstyle, head covering, eyewear, etc.), we propose a new method for fine-tuning Lp3D on real datasets by introducing a mixed loss function based on real and synthetic datasets. Our volumetric disentanglement and rendering framework is trained only using in-the-wild videos from the CelebV-HQ dataset [112] in a self-supervised fashion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 168, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 323 + ], + "type": "text", + "content": "We not only demonstrate that our volumetric face disentanglement approach produces qualitative superior head reenactments than existing ones, but also show on a wide and diverse set of source images how non-frontal poses and extreme expressions can be handled. We have quantitatively assessed our method on multiple benchmarks and outperform existing 2D and 3D state-of-the-art techniques in terms of fidelity, expression, and likeness accuracy metrics. Our 3D aware head reenactment technique is therefore suitable for AR/VR-based immersive applications, and we also showcase a teleconferencing system using a holographic display from LookingGlass [4]. We summarize the main contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 323, + 545, + 455 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 383 + ], + "type": "text", + "content": "- First fully volumetric disentanglement approach for real-time 3D aware head reenactment from a single photo. This method combines 3D lifting into a canonical triplane representation and formalized facial appearance and expression feature extraction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 383, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 545, + 407 + ], + "type": "text", + "content": "- A 3D lifting network that is fine-tuned on unconstrained real-world data instead of only generating synthetic ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "type": "text", + "content": "- We demonstrate superior fidelity, identity preservation, and robustness w.r.t. current state-of-the-art methods for facial reenactment on a wide range of public datasets. We plan to release our code to the public." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "content": "2D Neural Head Reenactment. The problem of generating animations of photorealistic human heads given images or video inputs has been thoroughly explored using various neural rendering techniques in the past few years, outperforming traditional 3DMM-based methods [8, 26, 31, 44, 63, 66, 78, 79, 96] which often appear uncanny due to their compressed linear space. These approaches can be categorized into one-shot and multi-shot ones. While multishot methods generally achieve high-fidelity results, they are not suitable for many consumer applications as they typically require an extensive amount of training data, such as a monocular video capture [10, 11, 18, 21, 30, 34, 93, 108-110, 113], and sometimes even a calibrated multi-view stereo setup [13, 29, 55, 58, 69]. More recently, few-shot techniques [104] have also been introduced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To maximize accessibility, a considerable number of methods [17, 26, 27, 32, 35, 39, 66, 72-74, 76, 77, 83, 86, 87, 89, 97, 101-103, 107] use a single portrait as input by leveraging advanced generative modeling techniques based" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10337" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "on in-the-wild video training data. While most methods rely on linear face models to extract facial expressions, the head reenactment technique from Drobyshev et al. [27] directly extract expression features from cropped 2D face regions, allowing them to obtain better face disentanglements, which results in higher fidelity face synthesis. While similar to our proposed approach in avoiding the use of low dimensional linear face models, their method is purely 2D and strugglesly with ensuring identity and expression consistency when novel views are synthesized." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 363 + ], + "type": "text", + "content": "3D-Aware One-Shot Head Reenactment. Due to potential inconsistencies when rendering from different views or poses, a number of 3D-aware single shot head reenactment techniques [7, 19, 20, 24, 62, 65, 70, 75, 88, 92, 94, 95] have been introduced. These methods generally use an efficient 3D representation, such as neural radiance fields or 3D mesh, to geometrically constraint the neural rendering and improve view consistency. ROME [43] for instance is a mesh-based method using FLAME blendshapes [51] and neural textures. While view-consistent results can be produced for both face and hair regions, the use of low resolution polygonal meshes hinders the neural renderer to generate high-fidelity geometric and appearance details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 363, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 289, + 579 + ], + "type": "text", + "content": "Implicit representations such as HeadNeRF [36] and MofaNeRF [38] use a NeRF-based parametric model which supports direct control of the head pose of the generated images. While real-time rendering is possible, these methods require intensive test-time optimization and often fail to preserve the identity of the source due to the use of compact latent vectors. Most recent methods [53, 54, 99] adopt the highly efficient tri-plane-based neural fields representation [20] to encode the 3D structure and appearance of the avatars head. Compared to the previous works on view-consistent neural avatars [36, 43, 53, 54, 59, 99], we refrain from depending on parametric head models for motion synthesis and, instead, learn the volumetric motion model from the training data. This methodology enables us to narrow the identity gap between the source and generated images and yield a superior fidelity of the generated motion compared to competing approaches, and hence a higher quality disentanglement for reenactment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "content": "3D GAN Inversion. When training a whole reconstruction and disentangled reenactment model end-to-end on facial performance videos, one can introduce substantial overfitting and reduce the quality of the results. To address this problems, we focus our training approach to an inversion of pre-trained 3D-aware generative models for human heads." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "We use tri-plane-based generative network EG3D [20] as the foundational generator, due to its proficiency in producing high-fidelity and view-consistent synthesis of human heads. For a given image, an effective 3D GAN inver" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "sion method should leverage these properties for estimating latent representations, which can be decoded into outputs that maintain view consistency and faithfully replicate the contents of the input. One naive approach is to adapt GAN inversion methods that were initially designed for 2D GANs to the EG3D pre-trained network. These methods either do a time consuming but more precise optimization [42, 68] or train a fast but less accurate encoder network [67, 80] to obtain the corresponding latent vectors. They often produce incorrect depth prediction, leading to clear artifacts in novel view synthesis. Hence, some methods are specifically designed for inverting 3D GANs, which either do multi-view optimization [45, 91] or predict residual features/tri-plane maps for refining the initial inversion results [12, 81, 98, 100]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 384 + ], + "type": "text", + "content": "In this work, we rely on the state-of-the-art EG3D inversion method Lp3D [81]. While achieving excellent novel-view synthesis results, it lacks disentanglement between the appearance and expression of the provided image and is unable to impose various driving expressions onto the input. To address this limitation, we propose a new method that introduces appearance-expression disentanglement in the latent space of tri-planes using our new self and cross-reenactment training pipeline while relying on a pre-trained but fine-tuned Lp3D network for regularization which enables highly consistent view synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 392, + 474, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 474, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 474, + 405 + ], + "type": "text", + "content": "3. 3D-Aware Head Reenactment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our head reenactment pipeline consists of three stages: 1) 3D Lifting, 2) Volumetric Disentanglement, and 3) Tri-plane Rendering. Given a pair of source and driver images, we first frontalize them using a pre-trained but fine-tuned tri-plane-based 3D lifting module [81]. This driver alignment step is crucial and allows our model to disentangle the expressions from the head pose, which prevents overfitting. Then, the frontalized faces are fed into two separate convolutional encoders to extract the face features " + }, + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 304, + 412, + 545, + 605 + ], + "type": "text", + "content": ". These extracted features are concatenated with the ones extracted from the tri-planes of the source, and all are fed together into several transformer blocks [90] to produce the expression tri-plane residual, which is added to the tri-planes of the source image. The final target image can be rendered from the new tri-planes using a pre-trained tri-plane renderer using the driver's pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 611, + 435, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 435, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 435, + 624 + ], + "type": "text", + "content": "3.1. Fine-Tuned 3D Lifting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "We adopt Lp3d [81] as a 3D face-lifting module, which predicts the radiance field of any given face image in real-time. Instead of using an implicit multi-layer perceptron [61] or sparse voxels [28, 71] for the radiance field, Lp3D [81] uses tri-planes [20], which can be computed using a single forward of a deep learning network. Specifically, for a given source image " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": ", we first extract the tri-planes " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " using a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10338" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 509, + 214 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 214 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 214 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 214 + ], + "type": "image", + "image_path": "5267cdbdd612a796023ff7ca70a6eeefd6865cc1442abf601c03b118b3eab723.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. Given a pair of source and driver images, our method processes them in three steps: 3D Lifting into tri-plane representations, Volumetric Disentanglement, which consists of source and driver frontalization and tri-plane residual generation, and Tri-plane Rendering via volumetric ray marching with subsequent super-resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 275, + 228, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 228, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 228, + 289 + ], + "type": "text", + "content": "transformer-based appearance encoder " + }, + { + "bbox": [ + 47, + 275, + 228, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{app}}" + }, + { + "bbox": [ + 47, + 275, + 228, + 289 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 293, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 293, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 287, + 308 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\text {a p p}} \\left(x _ {s}\\right) = T \\in \\mathbb {R} ^ {3 \\times H \\times W \\times C} = \\left\\{T _ {x y}, T _ {y z}, T _ {z x} \\right\\}. \\tag {1}", + "image_path": "7eeaaf5674586e69ea1455ac2641d982d99d41548617809c7c7de01c1c4b35d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "content": "The color " + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "content": " of each point " + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "inline_equation", + "content": "p = (x, y, z)" + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "content": " in the radiance field can be obtained by projecting " + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 314, + 288, + 361 + ], + "type": "text", + "content": " onto the three planes and by summing up the features at the projected positions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 369, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 369, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 287, + 384 + ], + "type": "interline_equation", + "content": "c, \\sigma = \\mathbf {D} \\left(F _ {x y} + F _ {y z} + F _ {z x}\\right), \\tag {2}", + "image_path": "63110bcedfea2fb60798e681ce1da57ca2f6dbb15f73191eeee9b1cb4361b4e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": " is a shallow MLP decoder for the tri-plane rendering, " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "F_{xy}, F_{yz}" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "F_{zx}" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": " are the feature vectors at the projected positions on " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "yz" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "zx" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": " planes, respectively, calculated using bilinear interpolation. The rendered " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": " image is then upsampled using a super-resolution module to produce a high-resolution output. To train the encoder " + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{app}}" + }, + { + "bbox": [ + 46, + 389, + 289, + 582 + ], + "type": "text", + "content": ", Lp3D [81] uses synthetic data generated from a 3D-aware face generative model [20]. While these synthetic data have ground truth camera poses, they are limited to the face distribution of the generative model. As a result, Lp3D can fail to generalize to in-the-wild images as shown in Fig. 5. To prevent this, we fine-tune the pre-trained Lp3D on a large-scale real-world dataset. We also replace the original super-resolution module in Lp3D with a pre-trained GF-PGAN [85], which is then fine-tuned together with Lp3D (see Sec. 3.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 269, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 269, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 269, + 601 + ], + "type": "text", + "content": "3.2. Disentangling Appearance and Expression" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Separating facial expression from the identity appearance in a 3D radiance field is very challenging especially when source and driver subjects have misaligned expressions. In order to simplify the problem, we use our 3D lifting approach to bring both source and driver heads into a pos-oriented space where faces are frontalized. Here, we denote frontalized source and driver images as " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{s}^{f}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{d}^{f}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", respectively. These images are then fed into two separate convolutional source and driver encoders " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_s" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_d" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " to produce" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "coarse feature maps " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "F_{s} = \\mathbf{E}_{s}(x_{s}^{f})" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "F_{d} = \\mathbf{E}_{d}(x_{d}^{f})" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": ". Since we already have the source's tri-plane, which encodes the 3D shape of the source, we use another encoder to encode this tri-plane and concatenate it together with the coarse frontalized feature maps of the images to produce expression feature " + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 354, + 436, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 354, + 436, + 367 + ], + "spans": [ + { + "bbox": [ + 384, + 354, + 436, + 367 + ], + "type": "interline_equation", + "content": "F _ {t} = \\mathbf {E} _ {t} (T)", + "image_path": "f779cc8195b0cf8b4fa62b41f738d8d45d9738994f581517bbc602e11c5344aa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 369, + 467, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 369, + 467, + 381 + ], + "spans": [ + { + "bbox": [ + 388, + 369, + 467, + 381 + ], + "type": "interline_equation", + "content": "F = F _ {s} \\oplus F _ {d} \\oplus F _ {t}", + "image_path": "8da9e44b080f03e979cec3ca9829c07229e363641fb91c3dd3909cae2c8ebb50.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "text", + "content": "Even though face frontalization aligns the source and the driver, there is still some misalignment between the two faces, e.g., the positions of the eyes may be different, or one mouth is open while the other is closed. Therefore, we feed the concatenation of the feature maps into several transformer blocks to produce the final residual tri-plane " + }, + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v(F)" + }, + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "text", + "content": ". This residual is then added back to the source's tri-planes to change the source's expression to the driver's expression " + }, + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "inline_equation", + "content": "T' = T + \\mathbf{E}_v(F)" + }, + { + "bbox": [ + 304, + 389, + 547, + 534 + ], + "type": "text", + "content": ". Unlike LPR [54], we do not use a 3D face model to compute the expression but instead use the RGB images of the source and the driver directly, allowing the model to learn high-fidelity and realistic expressions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 540, + 425, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 425, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 425, + 553 + ], + "type": "text", + "content": "3.3. Tri-Plane Rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "content": "The resulting tri-planes are then volumetrically rendered into one or multiple output images using pose parameters and viewing angles in the case of a holographic display. Following EG3D [20], we use a neural radiance fields (NeRFs)-based volumetric ray marching approach [60]. However, instead of encoding each point in space via positional encodings [60], the features of the points along rays are calculated using their projections onto tri-planes. Since tri-planes are aligned with the frontal face, we can compute these rays directly using camera extrinsics " + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{driver}}" + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "content": " predicted by an off-the-shelf 3D head pose estimator [23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "While the renderings are highly view-consistent, the large number of points evaluated for each ray still limits" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10339" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "the output resolution for real-time performance. We therefore follow [54] and employ a 2D upsampling network [84] based on StyleGAN2 [41], which in our experiments produced higher quality results than the upsampling approach in EG3D [20]. Finally, for holographic displays, we generate a number of renderings based on their viewing angles and simply using the head pose parameter. Real-time performance is achieved using efficient inference libraries such as TensorRT, half-precision, and batched inference over multiple GPUs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 198, + 153, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 153, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 153, + 211 + ], + "type": "text", + "content": "3.4. Training Strategy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": "Fine-Tuning Lp3D. To make Lp3D work with in-the-wild images, we fine-tune it on a large-scale real-world video dataset [111]. Unlike the use of synthetic data, real-world data do not have ground-truth camera parameters and facial expressions in monocular videos are typically inconsistent over time. While the camera parameters can be estimated using standard 3D pose estimators, the expression differences are difficult to determine. However, we found that we can ignore this expression difference and fine-tune Lp3D using real data together with continuous training on synthetic data. In particular, our experiments indicate that the fine-tuned model can still faithfully reconstruct 3D faces from the input without changing expressions and still generalize successfully on in-the-wild images. Specifically, on real video data, we sample two frames " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "x_{s}^{r}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "x_{d}^{r}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": " and estimate their camera parameters " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "P_{s}^{r}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "P_{d}^{r}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": ". Similar to [20], we assume a fixed intrinsic for standard portraits for all images. Then we use " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "E_{app}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": " from Lp3D to calculate the triplanes of " + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "inline_equation", + "content": "x_{s}^{r}" + }, + { + "bbox": [ + 46, + 216, + 289, + 456 + ], + "type": "text", + "content": ", render it using the two poses, and calculate reconstruction losses on the two rendered images:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 463, + 281, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 463, + 281, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 463, + 281, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e a l}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, P _ {d} ^ {r}\\right) - x _ {d} ^ {r} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {r}, x _ {s} ^ {r}\\right) - x _ {s} ^ {r} \\right\\|,", + "image_path": "a162069e9c3172cd976aceec9067b2be768b30827f4b60ed9686610e15583b7f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{Lp3D}(x,P)" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": " is the face in " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": " re-rendered using camera pose " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{real}}" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": " is the loss for real images. Simultaneously, we render two synthetic images employing an identical latent code but through varying camera views and calculate the synthetic loss " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{syn}}" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 550, + 279, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 550, + 279, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 550, + 279, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s y n}} = \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {f}, P _ {d} ^ {s}\\right) - x _ {d} ^ {s} \\right\\| + \\left\\| \\mathrm {L p 3 D} \\left(x _ {s} ^ {s}, P _ {s} ^ {s}\\right) - x _ {s} ^ {s} \\right\\|", + "image_path": "877dfe75ab2cd514a8ce99e716a9708010e76a8f9cbb87ef85de339d55d656dc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 567, + 156, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 567, + 156, + 581 + ], + "spans": [ + { + "bbox": [ + 58, + 567, + 156, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t r i}} = \\left\\| E _ {\\mathrm {a p p}} \\left(x _ {s} ^ {f}\\right) - T \\right\\|,", + "image_path": "6aff8e4ece9c4344c54ebb57aed66837d5b1221e535dbc813d1f230af9c94cf9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": " is the ground-truth tri-planes returned by EG3D [20] and " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tri}}" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": " is the tri-plane loss adopted directly from Lp3D. The final loss " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{app}}" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": " for fine-tuning Lp3D can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 643, + 235, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 643, + 235, + 657 + ], + "spans": [ + { + "bbox": [ + 99, + 643, + 235, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a p p}} = \\mathcal {L} _ {\\text {r e a l}} + \\lambda_ {\\text {s y n}} \\mathcal {L} _ {\\text {s y n}} + \\lambda_ {\\text {t r i}} \\mathcal {L} _ {\\text {t r i}}", + "image_path": "5256cd6ea5e6d51f8a41af785acbad30716ca2108d7b70efc1e7ece87fabd5dd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{syn}}" + }, + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tri}}" + }, + { + "bbox": [ + 47, + 663, + 243, + 676 + ], + "type": "text", + "content": " are tunable hyperparameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Disentangling Appearance and Expressions. In this stage, we also use real-world videos as training data. For a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "pair of source and driver images " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " sampled from the same video, we apply the reconstruction loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " which is a combination of " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", perceptual [105], and identity losses, between the reenacted image " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{s \\to d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the corresponding ground-truth " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 141, + 507, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 141, + 507, + 171 + ], + "spans": [ + { + "bbox": [ + 344, + 141, + 507, + 171 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e c o n}} = \\left\\| x _ {s \\rightarrow d} - x _ {d} \\right\\| _ {1} + \\phi \\left(x _ {s \\rightarrow d}, x _ {d}\\right) \\\\ + \\mathcal {L} _ {\\cos} \\left(\\mathrm {I D} \\left(x _ {s \\rightarrow d}\\right), \\mathrm {I D} \\left(x _ {d}\\right)\\right), \\\\ \\end{array}", + "image_path": "692e790c4b887d5a5d70dabc1c9bd51a9eb9a0e88baecc02cf6613385dbecd22.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "text", + "content": " is the perceptual loss, " + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}(\\cdot)" + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "text", + "content": " is a pretrained face recognizer [82], and " + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 305, + 178, + 545, + 203 + ], + "type": "text", + "content": " is the cosine similarity loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 210, + 515, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 210, + 515, + 236 + ], + "spans": [ + { + "bbox": [ + 336, + 210, + 515, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o s}} ((u, v) = \\frac {< u , v >}{m a x (\\| u \\| , \\epsilon) \\times m a x (\\| v \\| , \\epsilon)}", + "image_path": "cad6deb4b89c94375229baf7359a4c9f5b65fe92bbb72f936e3d44be362adf40.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": " is a small constant to void division by 0. Similar to other works that use RGB images directly to calculate expressions [27], our proposed encoder also suffers from an \"identity leaking\" issue. Since there is no cross-reenactment dataset, the expression module is trained with self-reenactment video data. Therefore, without proper augmentation and regularization, the expression module can leak identity information from the driver to the output, making the model fail to generalize to cross-reenactment tasks. Hence, we introduce a Cross Identity Regularization. Specifically, we further sample an additional driver frame " + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "inline_equation", + "content": "x_{d'}" + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": " from another video. We incorporate a GAN loss where real samples are " + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{Lp3D}(x_s, P^d)" + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": " and fake samples are " + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "inline_equation", + "content": "x_{s \\to d'}" + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": ". This GAN loss is also conditioned on the identity vector of the source " + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}(x_s)" + }, + { + "bbox": [ + 304, + 242, + 545, + 469 + ], + "type": "text", + "content": ". Following [27], we also apply strong augmentation (random warping and color jittering) and additionally mask the border of the driver randomly to further reduce potential identity leaks. The loss for expression training can be summarized as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 479, + 480, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 479, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 371, + 479, + 480, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\exp} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {C I R}} \\mathcal {L} _ {\\text {C I R}},", + "image_path": "d97555c4064dfd9cf0bb8c32d6d0e30fdb3116a6be82cf5c15121d110a991ee0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CIR}}" + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CIR}}" + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": " are cross identity regularization and its hyperparameter, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": "Global Fine-Tuning. After training both Lp3D and the expression module, we iteratively fine-tune the two modules using the same losses as the previous sections. Specifically, for every 10000 iterations, we freeze one module and fine-tune the other and vice versa. In addition, we add a GAN loss on the super-resolution output of the Lp3D module." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We train our model on CelebV-HQ dataset [112] using 7 NVIDIA RTX A6000 ADA (50Gb memory each). We use AdamW [56] to optimize the parameters with a learning rate of " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and batch size of 28. The Lp3D finetuning takes 5 days for 500K iterations to converge. Training the expression module takes 2 days, and the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10340" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 130, + 137 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 130, + 137 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 130, + 137 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 130, + 137 + ], + "type": "image", + "image_path": "3f513c2a949c0db7a267ecb66890172cd06d9528b0c93469c943e9d0e6036d37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 71, + 197, + 137 + ], + "blocks": [ + { + "bbox": [ + 130, + 71, + 197, + 137 + ], + "lines": [ + { + "bbox": [ + 130, + 71, + 197, + 137 + ], + "spans": [ + { + "bbox": [ + 130, + 71, + 197, + 137 + ], + "type": "image", + "image_path": "e301983bfe1ea85c7eceaad90b91f3aaa3a783d15cb18adde325e44049b5b785.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 71, + 264, + 137 + ], + "blocks": [ + { + "bbox": [ + 197, + 71, + 264, + 137 + ], + "lines": [ + { + "bbox": [ + 197, + 71, + 264, + 137 + ], + "spans": [ + { + "bbox": [ + 197, + 71, + 264, + 137 + ], + "type": "image", + "image_path": "dfa8277f2d88edd09be2a72efe590400755ce2a6a1e75f843093c225297f1add.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 71, + 332, + 137 + ], + "blocks": [ + { + "bbox": [ + 265, + 71, + 332, + 137 + ], + "lines": [ + { + "bbox": [ + 265, + 71, + 332, + 137 + ], + "spans": [ + { + "bbox": [ + 265, + 71, + 332, + 137 + ], + "type": "image", + "image_path": "0e0d254cf5d9713a9da084dec1a7e43bcdee6b84a1467afcd5d39fc57fb5d5d6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 333, + 71, + 411, + 137 + ], + "blocks": [ + { + "bbox": [ + 333, + 71, + 411, + 137 + ], + "lines": [ + { + "bbox": [ + 333, + 71, + 411, + 137 + ], + "spans": [ + { + "bbox": [ + 333, + 71, + 411, + 137 + ], + "type": "image", + "image_path": "c1b6fc341f84b10ed1d198f7b8b98111911296fc8472fbb905b04f3a75ade98d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 411, + 71, + 458, + 137 + ], + "blocks": [ + { + "bbox": [ + 411, + 71, + 458, + 137 + ], + "lines": [ + { + "bbox": [ + 411, + 71, + 458, + 137 + ], + "spans": [ + { + "bbox": [ + 411, + 71, + 458, + 137 + ], + "type": "image", + "image_path": "2cc95820c5f0783680b1ad4d1f50219473b6bd20900ebc95901fb166dc8c04ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 460, + 71, + 526, + 137 + ], + "blocks": [ + { + "bbox": [ + 460, + 71, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 460, + 71, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 460, + 71, + 526, + 137 + ], + "type": "image", + "image_path": "a617a6e81ba29d544c1cf20bc04bad8ab5ba62f282df4dfa3a2d97b8efbb4683.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 64, + 137, + 130, + 204 + ], + "blocks": [ + { + "bbox": [ + 64, + 137, + 130, + 204 + ], + "lines": [ + { + "bbox": [ + 64, + 137, + 130, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 137, + 130, + 204 + ], + "type": "image", + "image_path": "82b1a70b4f7a25d4dcc48c777800da2b5dc58196b807796ebd80e58942c7962a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 206, + 113, + 216 + ], + "lines": [ + { + "bbox": [ + 82, + 206, + 113, + 216 + ], + "spans": [ + { + "bbox": [ + 82, + 206, + 113, + 216 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 130, + 137, + 197, + 204 + ], + "blocks": [ + { + "bbox": [ + 130, + 137, + 197, + 204 + ], + "lines": [ + { + "bbox": [ + 130, + 137, + 197, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 197, + 204 + ], + "type": "image", + "image_path": "c2a4ae51a7174dc3c69582b1be432a1c223a6c39692faf3dc0afd033e49daabe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 206, + 178, + 216 + ], + "lines": [ + { + "bbox": [ + 151, + 206, + 178, + 216 + ], + "spans": [ + { + "bbox": [ + 151, + 206, + 178, + 216 + ], + "type": "text", + "content": "Driver" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 198, + 137, + 264, + 204 + ], + "blocks": [ + { + "bbox": [ + 198, + 137, + 264, + 204 + ], + "lines": [ + { + "bbox": [ + 198, + 137, + 264, + 204 + ], + "spans": [ + { + "bbox": [ + 198, + 137, + 264, + 204 + ], + "type": "image", + "image_path": "c77cc9da8ba590134d7e9810596f7c7a1faf7a6bba3d162cd72d1ac8ac6e4a52.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 206, + 254, + 216 + ], + "lines": [ + { + "bbox": [ + 214, + 206, + 254, + 216 + ], + "spans": [ + { + "bbox": [ + 214, + 206, + 254, + 216 + ], + "type": "text", + "content": "Ref View" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 265, + 137, + 332, + 205 + ], + "blocks": [ + { + "bbox": [ + 265, + 137, + 332, + 205 + ], + "lines": [ + { + "bbox": [ + 265, + 137, + 332, + 205 + ], + "spans": [ + { + "bbox": [ + 265, + 137, + 332, + 205 + ], + "type": "image", + "image_path": "fbe6ec14fed6fbe1715765fdb3e921b4392f2b83ce9bf5bbe2fd8fc28495a233.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 206, + 323, + 216 + ], + "lines": [ + { + "bbox": [ + 272, + 206, + 323, + 216 + ], + "spans": [ + { + "bbox": [ + 272, + 206, + 323, + 216 + ], + "type": "text", + "content": "Novel View" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 333, + 137, + 459, + 205 + ], + "blocks": [ + { + "bbox": [ + 333, + 137, + 459, + 205 + ], + "lines": [ + { + "bbox": [ + 333, + 137, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 333, + 137, + 459, + 205 + ], + "type": "image", + "image_path": "483934e4ed20d13d76b74fa22756740634daf3c1d931d198638a0be8e76550c4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 206, + 416, + 216 + ], + "lines": [ + { + "bbox": [ + 378, + 206, + 416, + 216 + ], + "spans": [ + { + "bbox": [ + 378, + 206, + 416, + 216 + ], + "type": "text", + "content": "Wrinkles" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 460, + 137, + 526, + 205 + ], + "blocks": [ + { + "bbox": [ + 460, + 137, + 526, + 205 + ], + "lines": [ + { + "bbox": [ + 460, + 137, + 526, + 205 + ], + "spans": [ + { + "bbox": [ + 460, + 137, + 526, + 205 + ], + "type": "image", + "image_path": "2af9511cf063f08dc26d5f9d7907c065f5d0823dba0353f6931b4864f96bcd12.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 206, + 515, + 217 + ], + "lines": [ + { + "bbox": [ + 471, + 206, + 515, + 217 + ], + "spans": [ + { + "bbox": [ + 471, + 206, + 515, + 217 + ], + "type": "text", + "content": "Geometry" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 64, + 242, + 526, + 656 + ], + "blocks": [ + { + "bbox": [ + 66, + 221, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 66, + 221, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 66, + 221, + 526, + 232 + ], + "type": "text", + "content": "Figure 3. Expression dependent high-fidelity details, incl. eye and forehead wrinkles, as well as nasolabial folds (see zoom-ins)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 242, + 526, + 656 + ], + "lines": [ + { + "bbox": [ + 64, + 242, + 526, + 656 + ], + "spans": [ + { + "bbox": [ + 64, + 242, + 526, + 656 + ], + "type": "image", + "image_path": "c084db6949b657f0c5fc4f6df5a478a3a06409c46b84be45156a7bec4905fe1a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 661, + 546, + 683 + ], + "lines": [ + { + "bbox": [ + 46, + 661, + 546, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 661, + 546, + 683 + ], + "type": "text", + "content": "Figure 4. A qualitative comparison with the baselines on in-the-wild photos. Notice that our method is capable of producing a variety of facial expressions, and handle highly diverse subjects, with and without accessories, as well as extreme head poses, such as rows 3 and 4." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10341" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 73, + 382, + 130 + ], + "blocks": [ + { + "bbox": [ + 47, + 73, + 382, + 130 + ], + "lines": [ + { + "bbox": [ + 47, + 73, + 382, + 130 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 382, + 130 + ], + "type": "table", + "html": "
MethodSelf-reenactmentCross-reenactment
PSNR ↑SSIM ↑LPIPS ↓NAKD ↓ECMD ↓FID ↓CSIM ↑ECMD ↓FID ↓
ROME [43]18.460.4880.3510.0300.5941380.5070.740172
StyleHeat [97]19.730.6890.2780.0350.74889.80.3980.74495.5
OTAvatar [59]19.280.7490.2890.0350.65167.00.4620.90172.4
MegaPortraits [27]21.100.7310.2910.0220.75552.00.7290.77161.7
Ours22.830.7680.1680.0120.42640.50.7540.75436.4
", + "image_path": "ee85584040d7cc47cb6e01a904490d287edb498fa4d821438268dece613a4722.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 394, + 73, + 545, + 129 + ], + "blocks": [ + { + "bbox": [ + 46, + 137, + 383, + 157 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 383, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 383, + 157 + ], + "type": "text", + "content": "Table 1. Evaluation on HDTF [106] dataset. Our method outperforms the competitors across almost all of the metrics for both self- and cross-reenactment scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 394, + 73, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 394, + 73, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 394, + 73, + 545, + 129 + ], + "type": "table", + "html": "
MethodCross-reenactment
CSIM ↑ECMD ↓FID ↓
ROME [43]0.5190.9152.6
HeadNeRF [36]0.3460.88113
StyleHeat [97]0.4670.8550.2
MegaPortraits [27]0.6470.7729.2
Ours0.6080.7923.6
", + "image_path": "2ef1c474ae904fd5cf789f1d885b1c978e253f17aba83fb411b43866856a86dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 172, + 162, + 249 + ], + "blocks": [ + { + "bbox": [ + 48, + 172, + 162, + 249 + ], + "lines": [ + { + "bbox": [ + 48, + 172, + 162, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 162, + 249 + ], + "type": "image", + "image_path": "9127b5160d117f322ec40bd912a6a4c7fee3ea2570ab6fc5683e6bd8a64807a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 257, + 162, + 288 + ], + "lines": [ + { + "bbox": [ + 47, + 257, + 162, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 162, + 288 + ], + "type": "text", + "content": "Figure 5. Our implementation of Lp3D [81] before and after CelebV-HQ [112] fine-tuning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 173, + 171, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 393, + 137, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 393, + 137, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 393, + 137, + 545, + 156 + ], + "type": "text", + "content": "Table 2. Evaluation on CelebA-HQ [40] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 171, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 173, + 171, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 173, + 171, + 286, + 223 + ], + "type": "table", + "html": "
CSIM ↑ECMD ↓
Lp3D0.5480.82
Lp3D-FT0.6700.76
w/o frontal0.6681.01
w/o CIR0.5700.97
Ours0.6080.79
", + "image_path": "d22b1cc7dea3097f684e30b683b5abb295a406d8c5ed5ea94c7a1a83ebe3432e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 225, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 171, + 225, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 171, + 225, + 287, + 285 + ], + "type": "text", + "content": "Table 3. Ablation studies conducted on CelebA-HQ [40] dataset. FT is a fine-tuned version of Lp3D, and \"frontal\" denotes frontalization of the source and driver." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 300, + 274, + 391 + ], + "blocks": [ + { + "bbox": [ + 58, + 300, + 274, + 391 + ], + "lines": [ + { + "bbox": [ + 58, + 300, + 274, + 391 + ], + "spans": [ + { + "bbox": [ + 58, + 300, + 274, + 391 + ], + "type": "image", + "image_path": "45d7c05ab06300c43cd4eb787b4fe219213b251900e5e6adab50625e5c74508f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 391, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 47, + 391, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 287, + 412 + ], + "type": "text", + "content": "Figure 6. Ablation study for source and driver frontalization and cross identity regularization (CIR)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 423, + 274, + 540 + ], + "blocks": [ + { + "bbox": [ + 59, + 423, + 274, + 540 + ], + "lines": [ + { + "bbox": [ + 59, + 423, + 274, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 423, + 274, + 540 + ], + "type": "image", + "image_path": "27fd866d1831199fb47c81b48ca96773e7cfc37b6b76051393e9a9da4ed6f5cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 540, + 287, + 561 + ], + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 561 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with LPR [54] method on the samples from HDTF [106] dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 605 + ], + "type": "text", + "content": "iterative fine-tuning takes another 5 days. More training details, such as hyperparameter fine-tuning or architecture of the networks, can be found in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Unlike Lp3D, our method reenacts faces without relifting in 3D for every frame. The tri-plane encoding and frontalization of the source image are only performed once. For each driver, we perform a single frontalization for the driver image (0.0115 ms), one expression encoding (0.0034 s), one tri-plane rendering at " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " resolution (0.0071 s), and one neural upsampling (0.0099 s). Each view runs at 31.9 fps on an Nvidia RTX 4090 GPU including I/O. More details are provided in the supplemental materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "text", + "content": "We compare our method with state-of-the-art 3D-based [36, 43, 59] and 2D-based [27, 97] models. For MegaPortraits [27], we use our own implementation that was trained on the CelebV-HQ dataset. Similar to previous works, we evaluate our method using public benchmarks, including CelebA-HQ [40] and HDTF [106]. For CelebAHQ, we split the data into two equal sets. Each set contains around " + }, + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "inline_equation", + "content": "15\\mathrm{K}" + }, + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "text", + "content": " images. Then, we use one set as the source and the rest as driver images. For the HDTF dataset, we perform cross-reenactment by using the first frame of each video as source and 200 first frames of other videos as drivers, which is more than " + }, + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "inline_equation", + "content": "60\\mathrm{K}" + }, + { + "bbox": [ + 304, + 171, + 547, + 375 + ], + "type": "text", + "content": " data pairs. Similarly, to evaluate self-reenactment, we also use the first frames of each video as sources and the rest of the same video as the driver. Furthermore, we also collected 100 face images on the internet and around 100 high-quality videos for qualitative comparisons. Video results are provided in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "text", + "content": "Evaluation. Given a source image " + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "text", + "content": ", a driver image " + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "text", + "content": " and reenacted output " + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "inline_equation", + "content": "x_{s\\rightarrow d}" + }, + { + "bbox": [ + 304, + 390, + 546, + 569 + ], + "type": "text", + "content": " we use EMOCAv2 [22] to extract the FLAME [52] expression coefficients of the prediction and the driver, as well as the shape coefficients of the source. We then compute 2 FLAME meshes using the predicted shape coefficients in world coordinates, one with the expression coefficients of the driver and one with the expression coefficients of the reenacted output. We measure the distance between the 2 meshes and denote this expression metric as ECMD. Moreover, we also use cosine similarity between the embeddings of a face recognition network (CSIM) [101], normalized average keypoint distance (NAKD) [16], perceptual image similarity (LPIPS) [105], peak signal-to-noise ratio (PSNR), and structure similarity index measure (SSIM)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 666 + ], + "type": "text", + "content": "We provide quantitative comparisons on HDTF and CelebA-HQ datasets in Tab. 1 and Tab. 2, respectively, and show that our method outperforms existing methods on both datasets. We also note that our FID and CSIM scores are significantly more reliable than the others, while expression-based metrics such as NAKD and ECMD are either better or very close to the best baseline, w.r.t output quality, expression accuracy, and identity consistency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "We showcase the qualitative results of cross-identity reenactment on in-the-wild images in Fig. 4 and Fig. 3. Compared to the baselines [27, 36, 43, 97], our reenactment faithfully reconstructs intricate and complex elements," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10342" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 274, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 274, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 274, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 274, + 145 + ], + "type": "image", + "image_path": "8ca4e4feee8c33ed28bd87b7574a7ac5634c1adec84b3c3c545194688e98973d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 148, + 287, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 190 + ], + "type": "text", + "content": "Figure 8. Failure cases of our method include side views in the source, extreme expressions, modeling of cartoonish characters and paintings, as well as modeling the reflections and semitransparency of the eyewear." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 203, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 335 + ], + "type": "text", + "content": "such as hairstyle, facial hair, glasses, and facial makeup. Furthermore, our method effectively generates realistic and fine-scale dynamic details that match the driver's expressions including substantial head pose rotations. We also conduct a comparative analysis of our results with the current state-of-the-art 3D-aware method LPR [54] in Fig. 7. Compared to LPR, our method achieves superior identity consistency. We further refer to the supplemental video for a live demonstration of our holographic telepresence system and animated head reenactment results and comparisons, with and without disentangled poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 345, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 289, + 620 + ], + "type": "text", + "content": "Ablation Study. We compare Lp3D with and without fine-tuning on the CelebA-HQ dataset in Tab. 3 and show several examples in Fig. 5. Without fine-tuning on real data, our implementation of Lp3D fails to preserve the identity of the input image, resulting in a considerably lower CSIM score. We also try without any facial frontalization in the expression module and instead use the source and driver images directly to calculate the expression tri-plane residual. We observe in Fig. 6 that without face frontalization, the model completely ignores the expression of the driver and keeps the expression of the input source instead. We show in Tab. 3, that facial frontalization leads to much better ECMD score. We then measure the effectiveness of the GAN-based cross-identity regularization on the CelebA-HQ dataset, " + }, + { + "bbox": [ + 46, + 345, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CIR}}" + }, + { + "bbox": [ + 46, + 345, + 289, + 620 + ], + "type": "text", + "content": ". Without this loss, identity characteristics (hairstyle or color) can leak from the driver to the output. See column 4 in Fig. 6. Tab. 3 also shows that cross-identity regularization can reduce identity leaking and improve the CSIM score. Lastly, we have also attempted to train our model end-to-end using the same losses and optimization process instead of our proposed iterative fine-tuning. Even with a lower learning rate and the use of pre-trained Lp3D weights, we were unable to succeed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Limitations. Limitations of our approach are illustrated in Figure 8. For source images that are extremely side ways (i.e., over " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "), our method can produce a plausible frontal face, but the likeness cannot be guaranteed due to insufficient visibility. For very highly stylized portraits, such as cartoons, our framework often produces photorealistic facial elements such as teeth which can be inconsistent in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "style. Due to the dependence on training data volume and diversity, accessories such as dental braces or glasses may disappear or look different during synthesis. We believe that providing more and better training data can further improve the performance of our algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 143, + 375, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 375, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 375, + 154 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 163, + 546, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 414 + ], + "type": "text", + "content": "We have demonstrated that a fully volumetric disentanglement of facial appearance and expressions is possible through a shared canonical tri-plane representation. In particular, an improved disentanglement also leads to higher fidelity and more robust head reenactment, when compared to existing methods that use linear face models for expressions, especially for non-frontal poses. A critical insight of our approach is that head frontalization via 3D lifting is particularly effective for extracting features that can encode fine details and expressions such as wrinkles and folds. The resulting reenactment is also highly view-consistent for large angles, making our solution suitable for holographic displays. We have also shown that the 3D lifting model can still be successfully trained with real data despite the fact that different frames with the same subject have varying facial expressions. Without a fine-tuned 3D lifting model, our 3D-aware reenactment framework would struggle with preserving the identity of the source, especially for side views. Our experiments indicate that our results achieve better visual quality and are more robust to extreme poses, validated via an extensive evaluation on multiple datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "text", + "content": "Risks and Potential Misuse. The proposed method is intended to promote avatar-based 3D communication. Nevertheless, our AI-based reenactment solution produces synthetic but highly realistic face videos from only a single photo, which could be hard to distinguish from a real person. Like deepfakes and other facial manipulation methods, potential misuse is possible and hence, we refer to the supplemental material for more discussions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": "Future Work. We are also interested in expanding our work to upper and full body reenactment, where hand gestures can be used for more engaging communication. To this end, we plan to investigate the use of canonical representations for human bodies, such as T-poses. As our primary motivation, we have showcased a solution using holographic displays for immersive 3D teleconferencing. However, we believe that our approach can also be extended to AR/VR HMD-based settings where full " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " head views are possible. The recent work by An et al. [7] is a promising avenue for future exploration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 671, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 671, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 671, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the Meta-verse Center Grant from the MBZUAI Research Office and by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\"." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10343" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 58, + 91, + 242, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 242, + 100 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 242, + 100 + ], + "type": "text", + "content": "[1] ItSeez3D AvatarSDK, https://avatarsdk.com. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 102, + 156, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 156, + 112 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 156, + 112 + ], + "type": "text", + "content": "[2] in3D, https://in3d.io.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 114, + 192, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 114, + 192, + 124 + ], + "spans": [ + { + "bbox": [ + 58, + 114, + 192, + 124 + ], + "type": "text", + "content": "[3] Leia, https://www.leiainc.com. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 126, + 283, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 126, + 283, + 136 + ], + "spans": [ + { + "bbox": [ + 58, + 126, + 283, + 136 + ], + "type": "text", + "content": "[4] Looking Glass Factory, https://lookingglassfactory.com. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 138, + 264, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 138, + 264, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 138, + 264, + 148 + ], + "type": "text", + "content": "[5] Pinscreen Avatar Neo, https://www.Avatarneo.com. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 150, + 226, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 150, + 226, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 226, + 159 + ], + "type": "text", + "content": "[6] ReadyPlayerMe, https://readyplayer.me. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 161, + 287, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 161, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 58, + 161, + 287, + 211 + ], + "type": "text", + "content": "[7] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 213, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 213, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 287, + 263 + ], + "type": "text", + "content": "[8] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20364-20373, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 266, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 266, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 266, + 287, + 316 + ], + "type": "text", + "content": "[9] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 362–371, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 318, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 398 + ], + "type": "text", + "content": "[10] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhinitra Meka, Ruofei Du, Ming-song Dou, Sergio Orts-Escalano, Rohit Pandey, Ping Tan, Thabo Beeler, Sean Fanello, and Yinda Zhang. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 400, + 286, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 400, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 400, + 286, + 440 + ], + "type": "text", + "content": "[11] Shrisha Bharadwaj, Yufeng Zheng, Otmar Hilliges, Michael J. Black, and Victoria Fernandez Abrevaya. FLARE: Fast learning of animatable and relightable mesh avatars. ACM Transactions on Graphics, 42:15, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 442, + 286, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 286, + 472 + ], + "type": "text", + "content": "[12] Ananta R Bhattacharai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. arXiv preprint arXiv:2303.13497, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 474, + 286, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 474, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 474, + 286, + 524 + ], + "type": "text", + "content": "[13] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason M. Saragih. Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG), 40, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 526, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 577 + ], + "type": "text", + "content": "[14] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 578, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 608 + ], + "type": "text", + "content": "[15] Volker Blanz and Thomas Vetter. A Morphable Model For The Synthesis Of 3D Faces. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 610, + 287, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 610, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 53, + 610, + 287, + 660 + ], + "type": "text", + "content": "[16] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE international conference on computer vision, pages 1021-1030, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 662, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 287, + 712 + ], + "type": "text", + "content": "[17] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 312, + 74, + 545, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 74, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 312, + 74, + 545, + 124 + ], + "type": "text", + "content": "[18] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 125, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 125, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 545, + 184 + ], + "type": "text", + "content": "[19] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 187, + 545, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 187, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 312, + 187, + 545, + 257 + ], + "type": "text", + "content": "[20] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 259, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 259, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 259, + 545, + 308 + ], + "type": "text", + "content": "[21] Chuhan Chen, Matthew O'Toole, Gaurav Bharaj, and Pablo Garrido. Implicit neural head synthesis via controllable local deformation fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 416-426, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 311, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 311, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 311, + 545, + 360 + ], + "type": "text", + "content": "[22] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 362, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 362, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 312, + 362, + 545, + 422 + ], + "type": "text", + "content": "[23] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 424, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 424, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 545, + 474 + ], + "type": "text", + "content": "[24] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 476, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 476, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 476, + 545, + 536 + ], + "type": "text", + "content": "[25] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 537, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 578 + ], + "type": "text", + "content": "[26] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 580, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 545, + 619 + ], + "type": "text", + "content": "[27] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "type": "text", + "content": "[28] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 673, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 673, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 673, + 545, + 712 + ], + "type": "text", + "content": "[29] Yonggan Fu, Yuecheng Li, Chenghui Li, Jason Saragih, Peizhao Zhang, Xiaoliang Dai, and Yingyan (Celine) Lin. Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In Proceedings of the" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10344" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21036-21045, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 144 + ], + "type": "text", + "content": "[30] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 145, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 145, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 54, + 145, + 288, + 185 + ], + "type": "text", + "content": "[31] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 186, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 186, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 288, + 236 + ], + "type": "text", + "content": "[32] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5609-5619, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 236, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 288, + 285 + ], + "type": "text", + "content": "[33] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos P Zafeiriou. Fast-ganfit: Generative adversarial network for high fidelity 3d face reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 286, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 288, + 337 + ], + "type": "text", + "content": "[34] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18653–18664, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 338, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 288, + 368 + ], + "type": "text", + "content": "[35] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 369, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 288, + 418 + ], + "type": "text", + "content": "[36] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 418, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 288, + 468 + ], + "type": "text", + "content": "[37] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Trans. Graph., 36(6), 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 469, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 469, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 469, + 286, + 490 + ], + "type": "text", + "content": "[38] Yiyu huang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 491, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 286, + 540 + ], + "type": "text", + "content": "[39] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 540, + 286, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 286, + 581 + ], + "type": "text", + "content": "[40] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 582, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 286, + 621 + ], + "type": "text", + "content": "[41] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 622, + 287, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 622, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 54, + 622, + 287, + 672 + ], + "type": "text", + "content": "[42] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 673, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 673, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 673, + 286, + 713 + ], + "type": "text", + "content": "[43] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference of Computer vision (ECCV), 2022. 2, 3, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 114 + ], + "type": "text", + "content": "[44] H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Nießner, P. Pérez, C. Richardt, M. Zollhöfer, and C. Theobalt. Deep video portraits. ACM Transactions on Graphics 2018 (TOG), 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 115, + 545, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 115, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 312, + 115, + 545, + 165 + ], + "type": "text", + "content": "[45] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2967-2976, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 167, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 167, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 167, + 545, + 227 + ], + "type": "text", + "content": "[46] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Abhijeet Ghosh, and Stefanos P Zafeiriou. Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 228, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 289 + ], + "type": "text", + "content": "[47] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8629-8640, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 290, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 290, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 290, + 545, + 371 + ], + "type": "text", + "content": "[48] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 40(6), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 373, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 545, + 434 + ], + "type": "text", + "content": "[49] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 394-403, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 434, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 545, + 484 + ], + "type": "text", + "content": "[50] Hao Li, Laura Trutoiu, Kyle Olszewski, Lingyu Wei, Tristan Trutna, Pei-Lun Hsieh, Aaron Nicholls, and Chongyang Ma. Facial performance sensing head-mounted display. ACM Transactions on Graphics (Proceedings SIGGRAPH 2015), 34(4), 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 485, + 545, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 545, + 526 + ], + "type": "text", + "content": "[51] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 527, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 527, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 527, + 545, + 567 + ], + "type": "text", + "content": "[52] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 568, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 545, + 639 + ], + "type": "text", + "content": "[53] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17969-17978, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 640, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 640, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 312, + 640, + 545, + 681 + ], + "type": "text", + "content": "[54] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. arXiv preprint arXiv:2306.08768, 2023. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 682, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 682, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 682, + 545, + 713 + ], + "type": "text", + "content": "[55] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep appearance models for face rendering. ACM Trans. Graph., 37(4), 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10345" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 103 + ], + "type": "text", + "content": "[56] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 104, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 104, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 104, + 288, + 145 + ], + "type": "text", + "content": "[57] Huiwen Luo, Koki Nagano, Han-Wei Kung, Mclean Gold-white, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. CoRR, abs/2106.11423, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 196 + ], + "type": "text", + "content": "[58] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De la Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 64-73, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 248 + ], + "type": "text", + "content": "[59] Zhiyuan Ma, Xiangyu Zhu, Guo-Jun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16910, 2023. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 289 + ], + "type": "text", + "content": "[60] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 290, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 290, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 288, + 340 + ], + "type": "text", + "content": "[61] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 392 + ], + "type": "text", + "content": "[62] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 393, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 434 + ], + "type": "text", + "content": "[63] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 465 + ], + "type": "text", + "content": "[64] Kyle Olszewski, Joseph J. Lim, Shunsuke Saito, and Hao Li. High-fidelity facial and speech animation for vr hmds. ACM Transactions on Graphics (TOG), 35:1-14, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 527 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 527 + ], + "type": "text", + "content": "[65] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 528, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 528, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 528, + 288, + 578 + ], + "type": "text", + "content": "[66] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13759-13768, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 579, + 288, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 288, + 630 + ], + "type": "text", + "content": "[67] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 631, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 631, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 631, + 288, + 661 + ], + "type": "text", + "content": "[68] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 662, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 662, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 662, + 288, + 712 + ], + "type": "text", + "content": "[69] Gabriel Schwartz, Shih-En Wei, Te-Li Wang, Stephen Lombardi, Tomas Simon, Jason Saragih, and Yaser Sheikh. The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Trans. Graph., 39(4), 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 124 + ], + "type": "text", + "content": "[70] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 125, + 545, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 125, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 545, + 165 + ], + "type": "text", + "content": "[71] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 166, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 166, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 166, + 545, + 196 + ], + "type": "text", + "content": "[72] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In CVPR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 197, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 197, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 197, + 545, + 238 + ], + "type": "text", + "content": "[73] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Conference on Neural Information Processing Systems (NeurIPS), 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 545, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 269 + ], + "type": "text", + "content": "[74] Aliaksandr Siarohin, Oliver Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 270, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 270, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 312, + 270, + 545, + 310 + ], + "type": "text", + "content": "[75] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d GANs. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 311, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 311, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 311, + 545, + 352 + ], + "type": "text", + "content": "[76] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Pareidolia face reenactment. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 353, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 353, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 353, + 545, + 403 + ], + "type": "text", + "content": "[77] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3637-3646, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 403, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 403, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 545, + 445 + ], + "type": "text", + "content": "[78] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 445, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 485 + ], + "type": "text", + "content": "[79] J. Thies, M. Zollhöfer, M. Stamminger, C. Theobalt, and M. Nießner. Headon: Real-time reenactment of human portrait videos. ACM Transactions on Graphics 2018 (TOG), 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 486, + 545, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 545, + 527 + ], + "type": "text", + "content": "[80] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40 (4):1-14, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 528, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 528, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 528, + 545, + 588 + ], + "type": "text", + "content": "[81] Alex Trevithick, Matthew Chan, Michael Stengel, Eric Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 589, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 589, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 312, + 589, + 545, + 620 + ], + "type": "text", + "content": "[82] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 312, + 621, + 545, + 670 + ], + "type": "text", + "content": "[83] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 671, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 671, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 671, + 545, + 712 + ], + "type": "text", + "content": "[84] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10346" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 123 + ], + "type": "text", + "content": "[85] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168-9178, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 124, + 288, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 165 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 165 + ], + "type": "text", + "content": "[86] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 165, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 165, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 165, + 288, + 205 + ], + "type": "text", + "content": "[87] O. Wiles, A.S. Koepke, and A. Zisserman. X2face: A network for controlling face generation by using images, audio, and pose codes. In European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 255 + ], + "type": "text", + "content": "[88] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2195-2205, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 256, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 288, + 297 + ], + "type": "text", + "content": "[89] Sitao Xiang, Yuming Gu, Pengda Xiang, Mingming He, Koki Nagano, Haiwei Chen, and Hao Li. One-shot identity-preserving portrait reenactment. arXiv preprint arXiv:2004.12452, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 346 + ], + "type": "text", + "content": "[90] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 398 + ], + "type": "text", + "content": "[91] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 458 + ], + "type": "text", + "content": "[92] Hongyi Xu, Guoxian Song, Zihang Jiang, Jianfeng Zhang, Yichun Shi, Jing Liu, Wanchun Ma, Jiashi Feng, and Linjie Luo. Omniavatar: Geometry-guided controllable 3d head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12814-12824, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 510 + ], + "type": "text", + "content": "[93] Yuelang Xu, Hongwen Zhang, Lizhen Wang, Xiaochen Zhao, Han Huang, Guojun Qi, and Yebin Liu. Latentavatar: Learning latent expression code for expressive neural head avatar. In ACM SIGGRAPH 2023 Conference Proceedings. Association for Computing Machinery, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 559 + ], + "type": "text", + "content": "[94] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 560, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 560, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 560, + 288, + 609 + ], + "type": "text", + "content": "[95] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18440-18449, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 611, + 288, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 611, + 288, + 642 + ], + "spans": [ + { + "bbox": [ + 53, + 611, + 288, + 642 + ], + "type": "text", + "content": "[96] Kewei Yang, Kang Chen, Daoliang Guo, Song-Hai Zhang, Yuan-Chen Guo, and Weidong Zhang. Face2face " + }, + { + "bbox": [ + 53, + 611, + 288, + 642 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 53, + 611, + 288, + 642 + ], + "type": "text", + "content": ": Real-time high-resolution one-shot face reenactment. 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 642, + 288, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 642, + 288, + 692 + ], + "spans": [ + { + "bbox": [ + 53, + 642, + 288, + 692 + ], + "type": "text", + "content": "[97] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In ECCV, 2022. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "text", + "content": "[98] Yu Yin, Kamran Ghasedi, HsiangTao Wu, Jiaolong Yang, Xin Tong, and Yun Fu. Nerfinvertor: High fidelity nerf-gan" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "inversion for single-shot real image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8539-8548, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 105, + 545, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 105, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 312, + 105, + 545, + 156 + ], + "type": "text", + "content": "[99] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based oneshot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 156, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 156, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 308, + 156, + 545, + 197 + ], + "type": "text", + "content": "[100] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3d gan inversion through geometry and occlusion-aware encoding. arXiv preprint arXiv:2303.12326, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 248 + ], + "type": "text", + "content": "[101] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 249, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 299 + ], + "type": "text", + "content": "[102] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 300, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 361 + ], + "type": "text", + "content": "[103] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaprotait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22096-22105, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 362, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 362, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 362, + 545, + 403 + ], + "type": "text", + "content": "[104] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 403, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 403, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 403, + 545, + 454 + ], + "type": "text", + "content": "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 455, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 506 + ], + "type": "text", + "content": "[106] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 506, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 506, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 506, + 545, + 528 + ], + "type": "text", + "content": "[107] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, pages 3657–3666, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 528, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 528, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 528, + 545, + 568 + ], + "type": "text", + "content": "[108] Xiaochen Zhao, Lizhen Wang, Jingxiang Sun, Hongwen Zhang, Jinli Suo, and Yebin Liu. Havatar: High-fidelity head avatar via facial model conditioned neural radiance field. ACM Trans. Graph., 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 569, + 545, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 629 + ], + "type": "text", + "content": "[109] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13545-13555, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 631, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 631, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 308, + 631, + 545, + 681 + ], + "type": "text", + "content": "[110] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J. Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 682, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 682, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 682, + 545, + 713 + ], + "type": "text", + "content": "[111] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In Eu" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10347" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 167 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "ropean conference on computer vision, pages 650-667. Springer, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 134 + ], + "type": "text", + "content": "[112] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 136, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 136, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 136, + 287, + 167 + ], + "type": "text", + "content": "[113] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10348" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_content_list.json b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2aa5a6182c9e0416d6fdfcf3406090fad1e9baf --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_content_list.json @@ -0,0 +1,1296 @@ +[ + { + "type": "text", + "text": "VP3D: Unleashing 2D Visual Prompt for Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Chen†, Yingwei Pan†, Haibo Yang§, Ting Yao†, and Tao Mei† \n†HiDream.ai Inc.", + "bbox": [ + 218, + 179, + 746, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{8}$ School of Computer Science, Fudan University, China", + "bbox": [ + 300, + 215, + 668, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{clenyang, pandy}@hidream.ai, yanghaibo.fdu@gmail.com, {tiyao, tmei}@hidream.ai", + "bbox": [ + 133, + 236, + 828, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent innovations on text-to-3D generation have featured Score Distillation Sampling (SDS), which enables the zero-shot learning of implicit 3D models (NeRF) by directly distilling prior knowledge from 2D diffusion models. However, current SDS-based models still struggle with intricate text prompts and commonly result in distorted 3D models with unrealistic textures or cross-view inconsistency issues. In this work, we introduce a novel Visual Prompt-guided text-to-3D diffusion model (VP3D) that explicitly unleashes the visual appearance knowledge in 2D visual prompt to boost text-to-3D generation. Instead of solely supervising SDS with text prompt, VP3D first capitalizes on 2D diffusion model to generate a high-quality image from input text, which subsequently acts as visual prompt to strengthen SDS optimization with explicit visual appearance. Meanwhile, we couple the SDS optimization with additional differentiable reward function that encourages rendering images of 3D models to better visually align with 2D visual prompt and semantically match with text prompt. Through extensive experiments, we show that the 2D Visual Prompt in our VP3D significantly eases the learning of visual appearance of 3D models and thus leads to higher visual fidelity with more detailed textures. It is also appealing in view that when replacing the self-generating visual prompt with a given reference image, VP3D is able to trigger a new task of stylized text-to-3D generation. Our project page is available at https://vp3d-cvpr24.github.io.", + "bbox": [ + 75, + 318, + 473, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Artificial Intelligence (especially for vision content generation) has aroused great attention in computer vision field [5, 6, 20, 26], leading to impressive advancements in text-to-image [30-32] and text-to-video generation [10, 14, 34]. These accomplishments can be attributed to the availability of large-scale image-text and video-text pair data [1, 33] and the emergence of robust diffusion-based generative models [12, 13, 25, 35]. Recently, researchers", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cdaf6826e003cb2f350a3b5c5f2ab8488a218d4a642f82154b530ff531c1345.jpg", + "image_caption": [ + "Text prompt: \"A florist is making a bouquet with fresh flowers\"" + ], + "image_footnote": [], + "bbox": [ + 504, + 294, + 694, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d77737463c8c506849b33453d7653609f379d03241a87bce959eb967f923a379.jpg", + "image_caption": [ + "(b)ProlificDreamer" + ], + "image_footnote": [], + "bbox": [ + 694, + 294, + 885, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09c4e66c1563e8cd39c1379ae4c26bcdee93215b40340be42e9dcbd23e726bc4.jpg", + "image_caption": [ + "(a) Magic3D", + "Visual prompt" + ], + "image_footnote": [], + "bbox": [ + 537, + 412, + 656, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ce9398ccec78e2746453dd7b53f0b4229098a936a39a7c64bfd9633c3e40178.jpg", + "image_caption": [ + "(c) VP3D (Ours)", + "Figure 1. Existing text-to-3D generation techniques (e.g., Magic3D [17] and ProlificDreamer [39]) often suffer from degenerated results (e.g., over-saturated appearances or inaccurate geometries). Our VP3D novelly integrates a visual prompt to strength score distillation sampling, leading to better 3D results." + ], + "image_footnote": [], + "bbox": [ + 666, + 412, + 795, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b622e2443acf43447a926aef5fb1eb2ebeeba709d72ac81186488496c75c6672.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 412, + 862, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have gone beyond text-driven image/video generation, and begun exploring diffusion models for text-driven content creation of 3D assets (e.g., text-to-3D generation). This direction paves a new way for practical 3D content creation and has a great potential impact for numerous applications like virtual reality, gaming and Metaverse. Compared to image generation, text-to-3D generation, however, is more challenging, due to the complexities associated with intricate 3D geometry and appearance (i.e., object shapes and textures). Moreover, the collection and annotation of 3D data are somewhat resourcefully expensive and thus cannot be easily scaled up to billion level as image data.", + "bbox": [ + 496, + 626, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this issue, a pioneering text-to-3D work (DreamFusion [27]) presents the first attempt of exploiting an off-the-shelf text-to-image diffusion model to generate promising 3D assets in a zero-shot fashion. The key design behind such success is Score Distillation Sampling (SDS), which directly optimizes the implicit 3D model of", + "bbox": [ + 496, + 810, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4896", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) with prior knowledge distilled from 2D diffusion model. Nevertheless, such distilled prior knowledge is merely driven by the input text prompt, and it is non-trivial to learn high-quality NeRF with distilled SDS supervision. Although several subsequent works [4, 17, 22, 38, 39] further upgrade SDS, this kind of SDS-based solution still results in degenerated 3D models with unrealistic/less-detailed textures, especially when feeding intricate text prompts (as seen in Figure 1 (a-b)).", + "bbox": [ + 75, + 90, + 468, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to mitigate this limitation through a unique design of visual prompt-guided text-to-3D diffusion model, namely VP3D. Intuitively, \"a picture is worth a thousand words.\" That is, a single image can convey human intentions of visual content creation (e.g., the visual appearance or semantic structure) more effectively than textual sentences. This motivates us to introduce additional guidance of visual prompt, and thus decompose the typical single-shot text-to-3D process into two cascaded processes: first text-to-image generation, and then (text plus image)-to-3D generation. In particular, VP3D first leverages off-the-shelf text-to-image diffusion models to produce a high-fidelity image that reflects extremely realistic appearance with rich details. In the latter process, this synthetic image further acts as 2D visual prompt to supervise SDS optimization of NeRF, coupled with the input text prompt. At the same time, a differentiable reward function is additionally utilized to encourage the rendering images of NeRF to be better aligned with 2D visual prompt (visual appearance consistency) and text prompt (semantic consistency). As illustrated in Figure 1 (c), we show that the novel visual prompt-guided diffusion process in VP3D significantly enhances the visual fidelity of 3D assets with realistic and rich texture details. Meanwhile, when easing the learning of visual appearance of 3D assets via visual prompt guidance, the optimization of NeRF will focus more on the modeling of geometry, leading to better 3D sharps with cross-view consistency. We believe that the ability of unleashing high-quality visual knowledge in 2D visual prompt is potentially a new paradigm of text-to-3D generation.", + "bbox": [ + 75, + 226, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a by-product, we also demonstrate that our VP3D can be readily adapted for a new task of stylized text-to-3D generation. Intriguingly, we simply replace the self-generating image in VP3D with a user-given reference image, and treat it as a new visual prompt to trigger (text plus image)-to-3D generation. In this way, our VP3D is able to produce a stylized 3D asset, which not only semantically aligns with text prompt but also shares similar geometric & visual style as the reference image.", + "bbox": [ + 75, + 680, + 468, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 830, + 218, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D generation. Significant advancements have been witnessed in text-to-image generation with 2D diffusion models in recent years [3, 12, 13, 25, 30-32, 35].", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, extending these capabilities to 3D content generation poses a substantial challenge, primarily due to the absence of large-scale paired text-3D datasets. To mitigate the reliance on extensive training data, recent works try to accomplish zero-shot text-to-3D generation [4, 7, 8, 17, 22, 27, 38, 39, 42]. Specifically, the pioneering work DreamFusion [27] showcased remarkable achievements in text-to-3D generation through pre-trained text-to-image diffusion models. SJC [38] concurrently addressed the out-of-distribution problem in lifting 2D diffusion models to perform text-to-3D generation. Following these, several subsequent works have strived to enhance text-to-3D generation further. For instance, Latent-NeRF [22] proposed to incorporate a sketch shape to guide the 3D generation directly in the latent space of a latent diffusion model. Magic3D [17] presented a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors to learn the underlying 3D representation. Control3D [8] proposed to enhance user controllability in text-to-3D generation by incorporating additional hand-drawn sketch conditions. ProlificDreamer [39] presented a principled particle-based variational framework to improve the generation quality.", + "bbox": [ + 496, + 90, + 890, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike previous works, we formulate the text-to-3D generation process from a new perspective. We first leverage the off-the-shelf text-to-image diffusion models to generate a high-quality image that faithfully matches the input text prompt. This synthetic reference image then serves as a complementary input alongside the text, synergistically guiding the 3D learning process. Moreover, we showcase the remarkable versatility of this novel architecture by effortlessly extending its capabilities to the realm of stylized text-to-3D generation. The resulting 3D asset not only exhibits semantic alignment with the provided text prompt but also masterfully captures the visual style of the reference image. This capability marks another pivotal distinction between our VP3D and previous text-to-3D approaches.", + "bbox": [ + 496, + 428, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-3D generation. Recently, prior works RealFusion [21], NeuralLift-360 [40] and NeRDi [9] leverage 2D diffusion models to achieve image-to-3D generation. The following work Make-IT-3D [37] proposed a two-stage optimization framework to improve the generation quality further. Zero-1-to-3 [19] finetuned the Stable Diffusion model to enable generating novel views of the input image. It can then be used as a 3D prior model to achieve high-quality image-to-3D generation. Inspired by this, Magic123 [28] proposed to use 2D and 3D priors simultaneously to generate faithful 3D content from the given image. One-2-3-45 [18] integrated Zero-1-to-3 and a multi-view reconstruction model to accelerate the 3D generation process. It is worth noting that our work is not targeting image-to-3D generation. We utilize a reference image to guide the text-to-3D learning process, instead of directly turning the reference image into 3D content.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. VP3D", + "text_level": 1, + "bbox": [ + 76, + 89, + 151, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate the architecture of our VP3D, which introduces a novel visual prompt-guided text-to-3D diffusion model. An overview of our VP3D architecture is depicted in Figure 2.", + "bbox": [ + 75, + 114, + 468, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 76, + 184, + 207, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Models. Diffusion models are a family of generative models that are trained to gradually transform Gaussian noise into samples from a target distribution [13]. Given a target data distribution $q(\\mathbf{x})$ , a forward diffusion process is defined to progressively add a small amount of Gaussian noise to the data $\\mathbf{x}_0$ sampled from $q(\\mathbf{x})$ . This process follows a Markov chain $q(\\mathbf{x}_{1:T}) = \\prod_{t=1}^{T} q(\\mathbf{x}_t | \\mathbf{x}_{t-1})$ and produces a sequence of latent variables $\\mathbf{x}_1, \\ldots, \\mathbf{x}_T$ after $T$ time steps. The marginal distribution of latent variables at time step $t$ is given by $q(\\mathbf{x}_t | \\mathbf{x}) = \\mathcal{N}(\\mathbf{x}_t; \\alpha_t \\mathbf{x}, \\sigma_t^2 \\mathbf{I})$ . Thus the noisy sample $\\mathbf{x}_t$ can be directly generated through the equation $\\mathbf{x}_t = \\alpha_t \\mathbf{x} + \\sigma_t^2 \\epsilon$ , where $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ , $\\alpha_t$ and $\\sigma_t$ are chosen parameters such that $\\alpha_t^2 + \\sigma_t^2 = 1$ . After $T$ noise adding steps, $\\mathbf{x}_T$ is equivalent to an isotropic Gaussian distribution. Then, a reverse generative process is defined to gradually \"denoise\" $X_T$ to reconstruct the original sample. This can be described by a Markov process $p_\\phi(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t)$ , with the conditional probability $p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t-1}; \\boldsymbol{\\mu}_\\phi(\\mathbf{x}_t, t), \\boldsymbol{\\Sigma}_\\phi(\\mathbf{x}_t, t))$ . Commonly, a UNet neural network $\\epsilon_\\phi(\\mathbf{x}_t; t)$ with parameters $\\phi$ is used to predict the noise that was used to produce $\\mathbf{x}_t$ at time step $t$ . Text-to-image diffusion models build upon the above theory to condition the diffusion process with a given text prompt $y$ using classifier-free guidance (CFG) [12]. The corresponding noise predictor is remodeled by:", + "bbox": [ + 73, + 207, + 472, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; \\mathbf {z} _ {y}, t) = \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset) + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 628, + 472, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s$ is a scale that denotes the classifier-free guidance weight, $\\mathbf{z}_y$ is the corresponding text embedding of the text prompt $y$ and $\\varnothing$ indicates the noise prediction without conditioning. The diffusion model $\\epsilon_{\\phi}$ is typically optimized by a simplified variant of the variational lower bound of the log data likelihood, which is a Mean Squared Error criterion:", + "bbox": [ + 75, + 659, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} (\\phi) = \\mathbb {E} _ {\\mathbf {x}, t, \\epsilon} \\left[ w (t) \\| \\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 758, + 468, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function that depends on the timestep $t \\sim \\mathcal{U}(0,1)$ and $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ .", + "bbox": [ + 75, + 794, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling. A recent pioneering work called DreamFusion [27] introduces Score Distillation Sampling (SDS) that enables leveraging the priors of pre-trained text-to-image diffusion models to facilitate text-to-3D generation. Specifically, let $\\theta$ be the learnable parameters of a", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D model (e.g., NeRF) and $g$ be a differentiable rendering function that can render an image $\\mathbf{x} = g(\\theta; \\mathbf{c})$ from the 3D model $\\theta$ at a camera viewpoint $\\mathbf{c}$ . SDS introduces a loss function $\\mathcal{L}_{SDS}$ to optimize the parameters $\\theta$ . Its gradient is defined as follows:", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 170, + 890, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t$ is obtained by perturbing the rendered image $\\mathbf{x}$ with a Gaussian noise $\\epsilon$ corresponding to the $t$ -th timestep of the forward diffusion process, $\\mathbf{z}_y$ is the conditioned text embedding of given text prompt $y$ . Intuitively, the SDS loss estimates an update direction in which the noised version of rendered image $\\mathbf{x}$ should be moved towards a denser region in the distribution of real images (aligned with the conditional text prompt $y$ ). By randomly sampling views and backpropagating the gradient in Eq. 3 to the parameters $\\theta$ through the differential parametric function $g$ , this approach eventually results in a 3D model that resembles the text.", + "bbox": [ + 496, + 205, + 893, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual-prompted Score Distillation Sampling", + "text_level": 1, + "bbox": [ + 498, + 378, + 879, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompt Generation. As aforementioned, score distillation sampling plays a key role in text-to-3D generation. Nevertheless, empirical observations [11, 27, 39] reveal that SDS still results in degenerated 3D models especially when feeding intricate text prompts. First, SDS-generated results often suffer from over-saturation issues. These issues are, in part, attributed to the necessity of employing a large CFG value (i.e., 100) within the SDS framework [27, 39]. A Large CFG value narrows down the score distillation space to more text-relevant areas. This can mitigate the divergence of diffusion priors in the optimization process, thereby fostering enhanced stability in 3D representation learning. However, this is at the cost of less realistic and diversity generation results, as large CFG values are known to yield over-saturated results [39]. Second, results generated by SDS still face the risk of text-3D misalignment, such as missing key elements in the scene, especially when text prompts contain multiple objects with specific attributes.", + "bbox": [ + 496, + 402, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A fundamental reason behind the aforementioned issues may lie in the substantial distribution gap between text and 3D modalities. Thus it is non-trivial to directly learn a meaningful 3D scene solely based on a single text prompt. This insight motivates us to introduce an additional visual prompt as a bridge to explicitly establish a connection between the text input and the desired 3D output. Particularly, we leverage off-the-shelf text-to-image diffusion models (e.g., Stable Diffusion) to produce a high-fidelity image that faithfully matches the input text prompt and has an extremely realistic appearance. This image is then used as a visual prompt in conjunction with the input text prompt to jointly supervise the 3D generation process.", + "bbox": [ + 496, + 674, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling with Visual Prompt. We now present visual-prompted score distillation sampling that dis", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4898", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e0a0fc06ceed63935551bc5d46de1b59510a36ec250a8b1719bf89054701b48.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed VP3D framework for visual prompted text-to-3D generation." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 848, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tillation knowledge from a pre-trained diffusion model to optimize a 3D model by considering inputs not only from a text prompt $y$ but also from a visual prompt $v$ . To be clear, we restructure the standard SDS-based text-to-3D pipeline by utilizing an image-conditioned diffusion model [43] to trigger visual prompt-guided text-to-3D generation. Technically, the visual prompt is first converted to a global image embedding $\\mathbf{z}_v$ by the CLIP image encoder [29] and a following projection network. This image embedding represents the rich content and style of the visual prompt and has the same dimension as the text embedding $\\mathbf{z}_y$ used in the pre-trained text-to-image diffusion model (Stable Diffusion). Following SDS, we first add noise $\\epsilon$ to the rendered image of the underlying 3D model according to the random sampled time step $t$ to get a noised image $\\mathbf{x}_t$ . Then $\\mathbf{x}_t$ is input to the diffusion model along with the conditional visual prompt embedding $\\mathbf{z}_v$ and text prompt embedding $\\mathbf{z}_y$ to estimate the added noise as follows:", + "bbox": [ + 75, + 349, + 473, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {e}} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}\\right) = \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right) \\\\ + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\lambda * \\mathbf {z} _ {v}\\right)\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right)), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 627, + 467, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s$ is the classifier-free guidance weight, $\\lambda \\in [0,1]$ is the visual prompt condition weight, $\\phi$ is the parameter of the pre-trained noise predictor $\\epsilon_{\\phi}$ and $\\epsilon_{\\phi}(\\mathbf{x}_t; t, \\emptyset, \\emptyset)$ denotes the noise prediction without conditioning. In this way, our proposed method explicitly incorporates the visual prompt and text prompt in a unified fashion for text-to-3D generation. Consequently, the final gradient of our introduced visual-prompted score distillation sampling (VP-SDS) loss $\\theta$ is expressed as:", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {V P - S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\tilde {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 816, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(t)$ is a scheduling coefficient.", + "bbox": [ + 76, + 854, + 334, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with SDS. Comparing the update gradient of SDS (Eq. 3) and VP-SDS (Eq. 5), SDS is a special case", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of our VP-SDS by setting $\\lambda = 0$ where the visual prompt condition is neglected. In accordance with the theoretical analysis presented in [27, 39], the mode-seeking nature of SDS necessitates a large CFG to ensure that the pre-trained diffusion model $\\epsilon_{\\phi}$ delivers a \"sharp\" updating direction for the underlying 3D model. Nevertheless, a large CFG, in turn, results in poor-quality samples and thus a \"degraded\" update direction. In contrast, VP-SDS leverages the additional visual prompt to narrow down the distillation space of $\\epsilon_{\\phi}$ into a more compact region that aligns tightly with the visual prompt. Meanwhile, the distillation space is also refined by the visual prompt as it reflects realistic appearances with rich details. Therefore, the updating direction derived from our VP-SDS is not only \"sharp\" but also \"fine\", which can obtain much better 3D generation results than SDS.", + "bbox": [ + 496, + 349, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, a recent work ProlificDreamer [39] presents variational score distillation (VSD) to address the aforementioned issues in SDS. However, VSD needs to train an additional diffusion model using LoRA [15] during the optimization process, which incurs a considerable computational overhead compared to SDS. Instead, the additional computational cost of our VP-SDS is nearly negligible, making it computationally more efficient than VSD.", + "bbox": [ + 496, + 579, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "View-dependent Visual Prompting. Apart from the oversaturation problem discussed above, existing text-to-3D methods are known to also suffer from the multi-view inconsistency problem (e.g., the multi-face Janus problem). This arises from the fact that the underlying prior diffusion model is exclusively trained on individual 2D images and therefore lacks 3D awareness. To alleviate this issue, existing text-to-3D methods [17, 27, 38, 39] always employ diffusion loss with view-dependent text conditioning, which is to append \"front view\", \"side view\", or \"back view\" to the input text based on the location of the randomly sampled camera. Inspired by this, we devise a view-dependent visual prompting strategy to further mitigate the view incon", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4899", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sistency problem in collaboration with our introduced VP-SDS. Technically, given the input visual prompt (assuming it is shot from the front view), we use a view-conditioned 2D diffusion model, Zero-1-to-3 [19], to transform it into left-side, right-side and backward views. Then we fed different visual prompts into VP-SDS (Eq. 5) depending on the corresponding sampled camera viewpoint. For instance, when the azimuth angle $\\gamma_{cam} \\in [0^{\\circ}, 360^{\\circ}]$ of the camera position falls in the range near $180^{\\circ}$ ( $0^{\\circ}$ denotes the front view), we feed the generated back view counterpart of the input visual prompt into Eq 5. In this way, the inherent 3D geometry information contained in the multi-view visual prompts is encoded into the 3D representation learning through view-dependent VP-SDS, leading to better view consistency in the 3D generation.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning with Reward Feedback", + "text_level": 1, + "bbox": [ + 76, + 325, + 366, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further encourage rendered images of the underlying 3D model that are high fidelity and well aligned with the input visual prompt and text prompt, we devise two types of differentiable reward functions that complement the aforementioned VP-SDS objective.", + "bbox": [ + 75, + 349, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Feedback Reward. Recent practice has shown the capability of improving text-to-image models with human feedback [41]. Particularly, it first trains a reward model on a large dataset comprised of human assessments of text-image pairs. Such a reward model thus has the ability to measure the quality of the generated samples in terms of both image fidelity and image-text alignment. Consequently, it can then be used to fine-tune diffusion models to maximize the predicted scores of the reward model through differentiable reward functions, leading to better generation results. Motivated by this, we go one step further to utilize the open-sourced reward model $\\mathbf{r}$ in ImageReward [41] for text-to-3D generation. Specifically, we introduce a human feedback reward loss as follows:", + "bbox": [ + 75, + 425, + 472, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h f - r e w a r d} = \\mathbb {E} _ {\\mathbf {c}} [ \\psi (\\mathbf {r} (\\mathbf {x}, y)) ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 647, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x} = g(\\theta; \\mathbf{c})$ is a rendered image by the underlying 3D model $\\theta$ from an arbitrary viewpoint $\\mathbf{c}$ , $y$ is the conditional text prompt and $\\psi$ is a differentiable reward-to-loss map function as in [41]. Intuitively, minimizing the loss in Eq. 6 encourages the rendered image $\\mathbf{x}$ to get a higher reward score from the reward model $\\mathbf{r}$ , which means the underlying 3D model should update towards the refined direction where the renderings have high appearance fidelity and faithfully match the input text prompt.", + "bbox": [ + 75, + 672, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Consistency Reward. Given that the above human feedback reward only takes into account the input text prompt, we further devised a visual consistency reward to fully leverage the visual prompt as well, since text prompts cannot capture all appearance details. Technically, we adopt a pre-trained self-supervised vision transformer DINO-ViT", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[2] to extract the visual features $F_{dino}(v)$ and $F_{dino}(\\mathbf{x})$ of the input visual prompt $v$ and rendered image $\\mathbf{x}$ , respectively. Then we penalize the feature-wise difference between them at the visual prompt viewpoint:", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v c - r e w a r d} = \\left\\| F _ {d i n o} (\\mathbf {x}) - F _ {d i n o} (\\mathbf {v}) \\right\\| ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 162, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By imposing such visual consistency loss, we encourage the underlying 3D model to adhere to the plausible shape and appearance properties conveyed by the visual prompt.", + "bbox": [ + 496, + 191, + 892, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. 3D Representation and Training", + "text_level": 1, + "bbox": [ + 498, + 247, + 784, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [17], we adopt a two-stage coarse-to-fine framework for text-to-3D generation with two different 3D scene representations. At the coarse stage, we leverage InstantNGP [24] as 3D representation, which is much faster to optimize compared to the vanilla NeRF [23] and can recover complex geometry. In the fine stage, we leverage DMTet as the 3D representation to further optimize a high-fidelity mesh and texture. Specifically, the 3D shape and texture represented in DMTet are first initialized from the density field and color field of the coarse stage, respectively [17].", + "bbox": [ + 496, + 270, + 890, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the optimization process in each stage, we first render images from the underlying 3D model through differentiable rasterizers at arbitrary camera poses and optimize the 3D model with a combination of losses:", + "bbox": [ + 496, + 422, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n e}} = \\mathcal {L} _ {V P - S D S} + \\lambda_ {1} \\mathcal {L} _ {v c - r e w a r d} + \\lambda_ {2} \\mathcal {L} _ {h f - r e w a r d}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 494, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are the trade-off parameters.", + "bbox": [ + 500, + 526, + 777, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 554, + 633, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of our VP3D for text-to-3D generation via extensive empirical evaluations. We first show both quantitative and qualitative results of VP3D in comparison to existing techniques on the newly released text-to-3D benchmark (T $^3$ Bench [11]). Next, we conduct ablation studies to validate each design in VP3D. Finally, we demonstrate the extended capability of VP3D for stylized text-to-3D generation.", + "bbox": [ + 496, + 580, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 710, + 705, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In the coarse and fine stage, the underlying 3D models are both optimized for 5000 iterations using Adam optimizer with 0.001 learning rate. The rendering resolutions are set to $128 \\times 128$ and $512 \\times 512$ for coarse and fine stage, respectively. We implement the underlying Instant-NGP and DMTet 3D representation mainly based on the Stable-DreamFusion codebase [36]. $\\lambda_{1}$ is set to 0.1 in the coarse stage and 0.01 in the fine stage. $\\lambda_{2}$ is linearly increased from 0.001 to 0.01 during the optimization process. The visual prompt condition weight is set to 0.5 in all experiments.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4900", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/52cf09fe66da209a293abadce9c41a7cea00c27f0264fd0dac6c217f0aa131cd.jpg", + "table_caption": [ + "Table 1. The quantitative results of our method and baselines on ${\\mathrm{T}}^{3}$ Bench [11]." + ], + "table_footnote": [], + "table_body": "
MethodSingle ObjectSingle Object with SurroundingsMultiple Objects
QualityAlignmentAverageQualityAlignmentAverageQualityAlignmentAverage
DreamFusion [27]24.924.024.419.329.824.617.314.816.1
SJC [38]26.323.024.717.322.319.817.75.811.7
LatentNeRF [22]34.232.033.123.737.530.621.719.520.6
Fantasia3D [4]29.223.526.421.932.027.022.714.318.5
ProlificDreamer [39]51.147.849.442.547.044.845.725.835.8
Magic3D [17]38.735.337.029.841.035.426.624.825.7
VP3D (Ours)54.852.253.545.450.848.149.131.540.3
", + "bbox": [ + 81, + 106, + 893, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. Existing text-to-3D generation works commonly examine their methods over the CLIP R-Precision score [16], which is an automated metric for the consistency of rendered images with respect to the input text. However, this text-image alignment-based metric cannot faithfully represent the overall 3D quality. For example, CLIP-based text-to-3D methods can also achieve high CLIP R-Precision scores even if the resulting 3D scenes are unrealistic and severely distorted [27]. Taking this into account, we instead conduct experiments on a newly open-sourced benchmark: $\\mathrm{T}^3$ Bench [11], which is the first comprehensive text-to-3D benchmark containing 300 diverse text prompts of three categories (single object, single object with surroundings, and multiple objects).", + "bbox": [ + 75, + 275, + 472, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathrm{T}^{3}$ Bench provides two automatic metrics (quality and alignment) based on the rendered multi-view images to assess the subjective quality and text alignment. The quality metric utilizes a combination of multi-view text-image scores and regional convolution to effectively identify quality and view inconsistency. The alignment metric employs a 3D captioning model and a Large Language Model (i.e., GPT-4) to access text-3D consistency. Following this, we also leverage the quality and alignment metric to quantitatively compare our VP3D against baseline methods.", + "bbox": [ + 75, + 489, + 468, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To evaluate our method, we compare our VP3D with six state-of-the-art text-to-3D generation methods: DreamFusion [27], SJC [38], LatentNeRF [22], Fantasy3D [4], Magic3D [17] and ProlificDreamer [39]. Specifically, DreamFusion [27] firstly introduces score distillation sampling (SDS) that enables leveraging 2D diffusion model (Imagen [14]) to optimize a NeRF [23]. SJC [38] concurrently addresses the out-of-distribution problem in SDS and utilizes an open-sourced diffusion model (Stable Diffusion) to optimize a voxel NeRF. Latent-NeRF [22] first brings NeRF to the latent space to harmonize with latent diffusion models, then refines it in pixel space. Magic3D [17] extends DreamFusion with a coarse-to-fine framework that first optimizes a low-resolution NeRF model and then a high-resolution DMTet model via SDS. Fantasy3D [4] disentangles the SDS-based 3D learning into geometry and appearance learning. ProlificDreamer [39] upgrades Dream", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fusion by a variational score distillation (VSD) loss that treats the underlying 3D scene as a random variable instead of a single point as in SDS.", + "bbox": [ + 496, + 275, + 893, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 332, + 694, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative performance comparisons of different methods for text-to-3D generation are summarized in Table 1. Overall, our VP3D consistently achieves better performances against existing techniques across all evaluation metrics and prompt categories. Remarkably, VP3D achieves an absolute quality-alignment average score improvement of $4.1\\%$ , $3.3\\%$ , and $4.5\\%$ against the best competitor ProlificDreamer across the three text prompt categories, respectively, which validates the effectiveness of our overall proposals. More importantly, while VP3D employs the same NeRF & DMTet 3D representation and coarse-t fine training scheme as the baseline method Magic3D, it significantly outperforms Magic3D by achieving $53.5\\%$ , $48.1\\%$ , and $40.3\\%$ average scores, representing a substantial improvement over Magic3D's average scores of $37.0\\%$ , $35.4\\%$ , and $25.7\\%$ . The results generally highlight the key advantage of introducing visual prompts in lifting 2D diffusion models to perform text-to-3D generation.", + "bbox": [ + 496, + 356, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, DreamFusion and SJC enable the zero-shot learning of implicit 3D models by distilling prior knowledge from 2D diffusion models. However, the generated 3D scenes have relatively low quality and alignment scores, especially in complex scenarios where the text prompt contains multiple objects or surroundings. Latent-NeRF employs score distillation in the latent space and then back to pixel space to further refine the 3D model, leading to better results. The aforementioned three methods only utilize implicit 3D representations (NeRFs). In contrast, Magic3D adopts textured mesh DMTet as 3D representation for enabling high-resolution optimization and exhibits better performances across all three prompt categories. Fantasy3D also capitalizes on DMTet for geometry learning and then leverages BRDF for appearance learning in a disentangled manner. While Fantasy3D achieves better average scores than DreamFusion and SJC, it fails to create high-fidelity results in complex scenes (e.g., \"multiple objects\"). Pro", + "bbox": [ + 496, + 628, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4901", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f305db8666f607d222b0a1e12203c3543cd08209a0a63d6d3d70456663f5233.jpg", + "image_caption": [ + "(a)", + "Figure 3. Comparisons on qualitative results of our VP3D with other text-to-3D techniques on $\\mathrm{T}^3$ Bench [11]. The prompts are (a) \"A fuzzy pink flamingo lawn ornament\", (b) \"A blooming potted orchid with purple flowers\", (c) \"A blue butterfly on a pink flower\", (d) \"A lighthouse on a rocky shore\", (e) \"Hot popcorn jump out from the red striped popcorn maker\", (f) \"A chef is making pizza dough in the kitchen\". (a-b), (c-d), (e-f) belongs to the Single Object, Single Object with Surr and Multi Objects category in $\\mathrm{T}^3$ Bench, respectively." + ], + "image_footnote": [], + "bbox": [ + 148, + 90, + 848, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lificDreamer further boosts the performance by training an additional diffusion model during the optimization process to realize a principled particle-based variational score distillation loss. However, our VP3D still outperforms ProlificDreamer across all evaluation metrics and prompt sets, which confirms the effectiveness of our VP3D.", + "bbox": [ + 75, + 672, + 470, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 771, + 261, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative comparisons for text-to-3D generation are presented in Figure 3. As can be seen, our VP3D generally produces superior 3D scenes with plausible geometry and realistic textures when compared with baseline methods. Specifically, DreamFusion suffers from a severely over-saturated problem and has difficulty generating complex geometry. Magic3D and Latent-NeRF slightly allevi", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ate these issues through their higher-resolution DMTet and pixel space refinement, respectively. While Fantasia3D and SJC can generate richer textures than DreamFusion, the geometric quality of the generated 3D scenes falls short of expectations. Notably, ProlificDreamer trains an additional diffusion model during the optimization process to perform variational score distillation (VSD) instead of SDS, achieving satisfactory single-object objects. However, the use of VSD at times introduces excessive irrelevant information or geometry noise in more complex scenarios. In contrast, we can clearly observe that the generated 3D scenes by VP3D faithfully match the input text prompt with plausible geometry and realistic appearance, which demonstrates the superiority of VP3D over state-of-the-art methods and its ability to generate high-quality 3D content.", + "bbox": [ + 496, + 672, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4902", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d9ff906e36fe6ccc62a74c87c4a97996b88a72876a2ccd17f595d58d23847ae.jpg", + "image_caption": [ + "Figure 4. Stylized text-to-3D generation results of our VP3D." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 468, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 452, + 230, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we investigate how each design in our VP3D influences the overall generation performance. We depict the qualitative results of each ablated run in Figure 5. $\\mathcal{L}_{SDS}$ is our baseline model that employs vanilla score distillation sampling loss. As can be seen, the generated 3D scene is over-saturated and geometry unreasonable. Instead, when $\\mathcal{L}_{VP - SDS}$ is employed, the generation quality is clearly enhanced in terms of both geometry and appearance. This highlights the critical effectiveness of our proposed visual-prompted score distillation sampling. Nevertheless, the resulting 3D scenes by $\\mathcal{L}_{VP - SDS}$ are still not satisfying enough. By utilizing additional visual consistency and human feedback reward functions $\\mathcal{L}_{vc - reward}$ (Eq. 7) and $\\mathcal{L}_{hf - reward}$ (Eq. 6), the generation quality is gradually improved. The results basically validate the effectiveness of these two complementary factors.", + "bbox": [ + 75, + 474, + 472, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Extension to Stylized Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 76, + 726, + 452, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we demonstrate that another advantage of our VP3D is its remarkable versatility in 3D generation as it can be readily adapted for a new task of stylized text-to-3D generation. The main difference is that the visual prompt is no longer generated from the text prompt but from a user-specified reference image. We also empirically discard the loss in Eq. 6 to eliminate the strictly text-image alignment constraint. In this way, our VP3D can integrate the visual cues contained in the reference image into text-to-3D generation and produce a stylized 3D asset. This asset not only", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7574029076a2fe215d2e2a48c5741efcedb9e842e04e073f781d767c496f231a.jpg", + "image_caption": [ + "Figure 5. Comparisons on qualitative results by using different ablated runs of our VP3D. The text prompts are (a) \"A broken tablespoon lies next to an empty sugar bowl\" and (b) \"A chameleon perched on a tree branch\"." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 883, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "semantically aligns with the text prompt but also reflects visual and geometry properties in the reference image.", + "bbox": [ + 498, + 303, + 890, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows our stylized text-to-3D generation results. Our VP3D can generate diverse and stylized 3D assets by giving different visual prompts to the same text prompt. As shown in Figure 4 (a-b), the generated result semantically is a rabbit that adheres to the text prompt but also inherits some visual cues of the visual prompt. To be clear, the generated 3D rabbits have somewhat consistent geometry pose and appearance texture with the object in the visual prompt. For example, in Figure 4 (b), the generated rabbit mirrors the \"hugging pose\" of the reference image and also has the same style of \"crescent-shaped eyebrows\" and \"yellow plaid jacket\" as in the reference image. In Figure 4 (c-d), we showcase the versatility of our VP3D by seamlessly blending styles from different visual prompts. Take Figure 4 (d) as an instance, we use the leopard image as a visual prompt in the coarse stage and then replace it with an oil painting image in the fine stage. Our VP3D finally resulted in a 3D rabbit that not only has a consistent pose with the leopard but also a colorful oil painting style texture. The stylized 3D generation ability distinct our VP3D from previous text-to-3D approaches and can lead to more creative and diverse 3D content creation.", + "bbox": [ + 496, + 335, + 892, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 691, + 619, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose VP3D, a new paradigm for text-to-3D generation by leveraging 2D visual prompts. We first capitalize on 2D diffusion models to generate a high-quality image from input text. This image then acts as a visual prompt to strengthen the 3D model learning with our devised visual-prompted score distillation sampling. Meanwhile, we introduce additional human feedback and visual consistency reward functions to encourage the semantic and appearance consistency between the 3D model and input visual & text prompt. Both qualitative and quantitative comparisons on the $\\mathrm{T}^3$ Bench benchmark demonstrate the superiority of our VP3D over existing SOTA techniques.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4903", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5", + "[3] Jingwen Chen, Yingwei Pan, Ting Yao, and Tao Mei. Controlstyle: Text-driven stylized image generation using diffusion priors. In ACM Multimedia, 2023. 2", + "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 2, 6", + "[5] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Animating your life: Real-time video-to-animation translation. In ACM MM, 2019. 1", + "[6] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM MM, 2019. 1", + "[7] Yang Chen, Jingwen Chen, Yingwei Pan, Xinmei Tian, and Tao Mei. 3d creation at your fingertips: From text or image to 3d assets. In ACM MM, 2023. 2", + "[8] Yang Chen, Yingwei Pan, Yehao Li, Ting Yao, and Tao Mei. Control3d: Towards controllable text-to-3d generation. In ACM MM, 2023. 2", + "[9] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2", + "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1", + "[11] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 3, 5, 6, 7", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2022. 1, 2, 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3", + "[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 6", + "[15] EJ Hu, Y Shen, P Wallis, Z Allen-Zhu, Y Li, S Wang, L Wang, and W Chen. Low-rank adaptation of large language models, arxiv, 2021. arXiv preprint arXiv:2106.09685, 10.4", + "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 6", + "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2, 4, 5, 6" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. In NeurIPS, 2024. 2", + "[19] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 5", + "[20] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-conditional diffusion networks for image captioning. In CVPR, 2023. 1", + "[21] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 2", + "[22] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5, 6", + "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 5", + "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICLR, 2021. 1, 2", + "[26] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 1", + "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2, 3, 4, 6", + "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022.", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2", + "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 1", + "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1", + "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2", + "[36] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.5", + "[37] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 2", + "[38] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 2, 4, 6", + "[39] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 1, 2, 3, 4, 6", + "[40] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 2", + "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.5", + "[42] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In ACM MM, 2023. 2", + "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + ], + "bbox": [ + 78, + 90, + 468, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4905", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_model.json b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb5ed249e4069c574399fbc7b5f2d11f9f03f27c --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_model.json @@ -0,0 +1,1815 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.807, + 0.152 + ], + "angle": 0, + "content": "VP3D: Unleashing 2D Visual Prompt for Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.18, + 0.748, + 0.216 + ], + "angle": 0, + "content": "Yang Chen†, Yingwei Pan†, Haibo Yang§, Ting Yao†, and Tao Mei† \n†HiDream.ai Inc." + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.217, + 0.669, + 0.235 + ], + "angle": 0, + "content": "\\(^{8}\\)School of Computer Science, Fudan University, China" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.237, + 0.829, + 0.253 + ], + "angle": 0, + "content": "{clenyang, pandy}@hidream.ai, yanghaibo.fdu@gmail.com, {tiyao, tmei}@hidream.ai" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.728 + ], + "angle": 0, + "content": "Recent innovations on text-to-3D generation have featured Score Distillation Sampling (SDS), which enables the zero-shot learning of implicit 3D models (NeRF) by directly distilling prior knowledge from 2D diffusion models. However, current SDS-based models still struggle with intricate text prompts and commonly result in distorted 3D models with unrealistic textures or cross-view inconsistency issues. In this work, we introduce a novel Visual Prompt-guided text-to-3D diffusion model (VP3D) that explicitly unleashes the visual appearance knowledge in 2D visual prompt to boost text-to-3D generation. Instead of solely supervising SDS with text prompt, VP3D first capitalizes on 2D diffusion model to generate a high-quality image from input text, which subsequently acts as visual prompt to strengthen SDS optimization with explicit visual appearance. Meanwhile, we couple the SDS optimization with additional differentiable reward function that encourages rendering images of 3D models to better visually align with 2D visual prompt and semantically match with text prompt. Through extensive experiments, we show that the 2D Visual Prompt in our VP3D significantly eases the learning of visual appearance of 3D models and thus leads to higher visual fidelity with more detailed textures. It is also appealing in view that when replacing the self-generating visual prompt with a given reference image, VP3D is able to trigger a new task of stylized text-to-3D generation. Our project page is available at https://vp3d-cvpr24.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Generative Artificial Intelligence (especially for vision content generation) has aroused great attention in computer vision field [5, 6, 20, 26], leading to impressive advancements in text-to-image [30-32] and text-to-video generation [10, 14, 34]. These accomplishments can be attributed to the availability of large-scale image-text and video-text pair data [1, 33] and the emergence of robust diffusion-based generative models [12, 13, 25, 35]. Recently, researchers" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.282, + 0.857, + 0.294 + ], + "angle": 0, + "content": "Text prompt: \"A florist is making a bouquet with fresh flowers\"" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.295, + 0.695, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.398, + 0.631, + 0.41 + ], + "angle": 0, + "content": "(a) Magic3D" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.295, + 0.887, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.399, + 0.84, + 0.411 + ], + "angle": 0, + "content": "(b)ProlificDreamer" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.414, + 0.658, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.51, + 0.635, + 0.521 + ], + "angle": 0, + "content": "Visual prompt" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.414, + 0.796, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.414, + 0.864, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.522, + 0.756, + 0.534 + ], + "angle": 0, + "content": "(c) VP3D (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.544, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 1. Existing text-to-3D generation techniques (e.g., Magic3D [17] and ProlificDreamer [39]) often suffer from degenerated results (e.g., over-saturated appearances or inaccurate geometries). Our VP3D novelly integrates a visual prompt to strength score distillation sampling, leading to better 3D results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.809 + ], + "angle": 0, + "content": "have gone beyond text-driven image/video generation, and begun exploring diffusion models for text-driven content creation of 3D assets (e.g., text-to-3D generation). This direction paves a new way for practical 3D content creation and has a great potential impact for numerous applications like virtual reality, gaming and Metaverse. Compared to image generation, text-to-3D generation, however, is more challenging, due to the complexities associated with intricate 3D geometry and appearance (i.e., object shapes and textures). Moreover, the collection and annotation of 3D data are somewhat resourcefully expensive and thus cannot be easily scaled up to billion level as image data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.896, + 0.903 + ], + "angle": 0, + "content": "To tackle this issue, a pioneering text-to-3D work (DreamFusion [27]) presents the first attempt of exploiting an off-the-shelf text-to-image diffusion model to generate promising 3D assets in a zero-shot fashion. The key design behind such success is Score Distillation Sampling (SDS), which directly optimizes the implicit 3D model of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.226 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) with prior knowledge distilled from 2D diffusion model. Nevertheless, such distilled prior knowledge is merely driven by the input text prompt, and it is non-trivial to learn high-quality NeRF with distilled SDS supervision. Although several subsequent works [4, 17, 22, 38, 39] further upgrade SDS, this kind of SDS-based solution still results in degenerated 3D models with unrealistic/less-detailed textures, especially when feeding intricate text prompts (as seen in Figure 1 (a-b))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.681 + ], + "angle": 0, + "content": "In this work, we propose to mitigate this limitation through a unique design of visual prompt-guided text-to-3D diffusion model, namely VP3D. Intuitively, \"a picture is worth a thousand words.\" That is, a single image can convey human intentions of visual content creation (e.g., the visual appearance or semantic structure) more effectively than textual sentences. This motivates us to introduce additional guidance of visual prompt, and thus decompose the typical single-shot text-to-3D process into two cascaded processes: first text-to-image generation, and then (text plus image)-to-3D generation. In particular, VP3D first leverages off-the-shelf text-to-image diffusion models to produce a high-fidelity image that reflects extremely realistic appearance with rich details. In the latter process, this synthetic image further acts as 2D visual prompt to supervise SDS optimization of NeRF, coupled with the input text prompt. At the same time, a differentiable reward function is additionally utilized to encourage the rendering images of NeRF to be better aligned with 2D visual prompt (visual appearance consistency) and text prompt (semantic consistency). As illustrated in Figure 1 (c), we show that the novel visual prompt-guided diffusion process in VP3D significantly enhances the visual fidelity of 3D assets with realistic and rich texture details. Meanwhile, when easing the learning of visual appearance of 3D assets via visual prompt guidance, the optimization of NeRF will focus more on the modeling of geometry, leading to better 3D sharps with cross-view consistency. We believe that the ability of unleashing high-quality visual knowledge in 2D visual prompt is potentially a new paradigm of text-to-3D generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.817 + ], + "angle": 0, + "content": "As a by-product, we also demonstrate that our VP3D can be readily adapted for a new task of stylized text-to-3D generation. Intriguingly, we simply replace the self-generating image in VP3D with a user-given reference image, and treat it as a new visual prompt to trigger (text plus image)-to-3D generation. In this way, our VP3D is able to produce a stylized 3D asset, which not only semantically aligns with text prompt but also shares similar geometric & visual style as the reference image." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.22, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Text-to-3D generation. Significant advancements have been witnessed in text-to-image generation with 2D diffusion models in recent years [3, 12, 13, 25, 30-32, 35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.426 + ], + "angle": 0, + "content": "However, extending these capabilities to 3D content generation poses a substantial challenge, primarily due to the absence of large-scale paired text-3D datasets. To mitigate the reliance on extensive training data, recent works try to accomplish zero-shot text-to-3D generation [4, 7, 8, 17, 22, 27, 38, 39, 42]. Specifically, the pioneering work DreamFusion [27] showcased remarkable achievements in text-to-3D generation through pre-trained text-to-image diffusion models. SJC [38] concurrently addressed the out-of-distribution problem in lifting 2D diffusion models to perform text-to-3D generation. Following these, several subsequent works have strived to enhance text-to-3D generation further. For instance, Latent-NeRF [22] proposed to incorporate a sketch shape to guide the 3D generation directly in the latent space of a latent diffusion model. Magic3D [17] presented a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors to learn the underlying 3D representation. Control3D [8] proposed to enhance user controllability in text-to-3D generation by incorporating additional hand-drawn sketch conditions. ProlificDreamer [39] presented a principled particle-based variational framework to improve the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Unlike previous works, we formulate the text-to-3D generation process from a new perspective. We first leverage the off-the-shelf text-to-image diffusion models to generate a high-quality image that faithfully matches the input text prompt. This synthetic reference image then serves as a complementary input alongside the text, synergistically guiding the 3D learning process. Moreover, we showcase the remarkable versatility of this novel architecture by effortlessly extending its capabilities to the realm of stylized text-to-3D generation. The resulting 3D asset not only exhibits semantic alignment with the provided text prompt but also masterfully captures the visual style of the reference image. This capability marks another pivotal distinction between our VP3D and previous text-to-3D approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image-to-3D generation. Recently, prior works RealFusion [21], NeuralLift-360 [40] and NeRDi [9] leverage 2D diffusion models to achieve image-to-3D generation. The following work Make-IT-3D [37] proposed a two-stage optimization framework to improve the generation quality further. Zero-1-to-3 [19] finetuned the Stable Diffusion model to enable generating novel views of the input image. It can then be used as a 3D prior model to achieve high-quality image-to-3D generation. Inspired by this, Magic123 [28] proposed to use 2D and 3D priors simultaneously to generate faithful 3D content from the given image. One-2-3-45 [18] integrated Zero-1-to-3 and a multi-view reconstruction model to accelerate the 3D generation process. It is worth noting that our work is not targeting image-to-3D generation. We utilize a reference image to guide the text-to-3D learning process, instead of directly turning the reference image into 3D content." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4897" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.152, + 0.105 + ], + "angle": 0, + "content": "3. VP3D" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.177 + ], + "angle": 0, + "content": "In this section, we elaborate the architecture of our VP3D, which introduces a novel visual prompt-guided text-to-3D diffusion model. An overview of our VP3D architecture is depicted in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.185, + 0.208, + 0.2 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.208, + 0.473, + 0.604 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Models. Diffusion models are a family of generative models that are trained to gradually transform Gaussian noise into samples from a target distribution [13]. Given a target data distribution \\( q(\\mathbf{x}) \\), a forward diffusion process is defined to progressively add a small amount of Gaussian noise to the data \\( \\mathbf{x}_0 \\) sampled from \\( q(\\mathbf{x}) \\). This process follows a Markov chain \\( q(\\mathbf{x}_{1:T}) = \\prod_{t=1}^{T} q(\\mathbf{x}_t | \\mathbf{x}_{t-1}) \\) and produces a sequence of latent variables \\( \\mathbf{x}_1, \\ldots, \\mathbf{x}_T \\) after \\( T \\) time steps. The marginal distribution of latent variables at time step \\( t \\) is given by \\( q(\\mathbf{x}_t | \\mathbf{x}) = \\mathcal{N}(\\mathbf{x}_t; \\alpha_t \\mathbf{x}, \\sigma_t^2 \\mathbf{I}) \\). Thus the noisy sample \\( \\mathbf{x}_t \\) can be directly generated through the equation \\( \\mathbf{x}_t = \\alpha_t \\mathbf{x} + \\sigma_t^2 \\epsilon \\), where \\( \\epsilon \\sim \\mathcal{N}(0, \\mathbf{I}) \\), \\( \\alpha_t \\) and \\( \\sigma_t \\) are chosen parameters such that \\( \\alpha_t^2 + \\sigma_t^2 = 1 \\). After \\( T \\) noise adding steps, \\( \\mathbf{x}_T \\) is equivalent to an isotropic Gaussian distribution. Then, a reverse generative process is defined to gradually \"denoise\" \\( X_T \\) to reconstruct the original sample. This can be described by a Markov process \\( p_\\phi(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t) \\), with the conditional probability \\( p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t-1}; \\boldsymbol{\\mu}_\\phi(\\mathbf{x}_t, t), \\boldsymbol{\\Sigma}_\\phi(\\mathbf{x}_t, t)) \\). Commonly, a UNet neural network \\( \\epsilon_\\phi(\\mathbf{x}_t; t) \\) with parameters \\( \\phi \\) is used to predict the noise that was used to produce \\( \\mathbf{x}_t \\) at time step \\( t \\). Text-to-image diffusion models build upon the above theory to condition the diffusion process with a given text prompt \\( y \\) using classifier-free guidance (CFG) [12]. The corresponding noise predictor is remodeled by:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.629, + 0.473, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; \\mathbf {z} _ {y}, t) = \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset) + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.751 + ], + "angle": 0, + "content": "where \\(s\\) is a scale that denotes the classifier-free guidance weight, \\(\\mathbf{z}_y\\) is the corresponding text embedding of the text prompt \\(y\\) and \\(\\varnothing\\) indicates the noise prediction without conditioning. The diffusion model \\(\\epsilon_{\\phi}\\) is typically optimized by a simplified variant of the variational lower bound of the log data likelihood, which is a Mean Squared Error criterion:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.76, + 0.47, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} (\\phi) = \\mathbb {E} _ {\\mathbf {x}, t, \\epsilon} \\left[ w (t) \\| \\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function that depends on the timestep \\( t \\sim \\mathcal{U}(0,1) \\) and \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Score Distillation Sampling. A recent pioneering work called DreamFusion [27] introduces Score Distillation Sampling (SDS) that enables leveraging the priors of pre-trained text-to-image diffusion models to facilitate text-to-3D generation. Specifically, let \\(\\theta\\) be the learnable parameters of a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "3D model (e.g., NeRF) and \\( g \\) be a differentiable rendering function that can render an image \\( \\mathbf{x} = g(\\theta; \\mathbf{c}) \\) from the 3D model \\( \\theta \\) at a camera viewpoint \\( \\mathbf{c} \\). SDS introduces a loss function \\( \\mathcal{L}_{SDS} \\) to optimize the parameters \\( \\theta \\). Its gradient is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.171, + 0.892, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.895, + 0.372 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t\\) is obtained by perturbing the rendered image \\(\\mathbf{x}\\) with a Gaussian noise \\(\\epsilon\\) corresponding to the \\(t\\)-th timestep of the forward diffusion process, \\(\\mathbf{z}_y\\) is the conditioned text embedding of given text prompt \\(y\\). Intuitively, the SDS loss estimates an update direction in which the noised version of rendered image \\(\\mathbf{x}\\) should be moved towards a denser region in the distribution of real images (aligned with the conditional text prompt \\(y\\)). By randomly sampling views and backpropagating the gradient in Eq. 3 to the parameters \\(\\theta\\) through the differential parametric function \\(g\\), this approach eventually results in a 3D model that resembles the text." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.38, + 0.88, + 0.396 + ], + "angle": 0, + "content": "3.2. Visual-prompted Score Distillation Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Visual Prompt Generation. As aforementioned, score distillation sampling plays a key role in text-to-3D generation. Nevertheless, empirical observations [11, 27, 39] reveal that SDS still results in degenerated 3D models especially when feeding intricate text prompts. First, SDS-generated results often suffer from over-saturation issues. These issues are, in part, attributed to the necessity of employing a large CFG value (i.e., 100) within the SDS framework [27, 39]. A Large CFG value narrows down the score distillation space to more text-relevant areas. This can mitigate the divergence of diffusion priors in the optimization process, thereby fostering enhanced stability in 3D representation learning. However, this is at the cost of less realistic and diversity generation results, as large CFG values are known to yield over-saturated results [39]. Second, results generated by SDS still face the risk of text-3D misalignment, such as missing key elements in the scene, especially when text prompts contain multiple objects with specific attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.87 + ], + "angle": 0, + "content": "A fundamental reason behind the aforementioned issues may lie in the substantial distribution gap between text and 3D modalities. Thus it is non-trivial to directly learn a meaningful 3D scene solely based on a single text prompt. This insight motivates us to introduce an additional visual prompt as a bridge to explicitly establish a connection between the text input and the desired 3D output. Particularly, we leverage off-the-shelf text-to-image diffusion models (e.g., Stable Diffusion) to produce a high-fidelity image that faithfully matches the input text prompt and has an extremely realistic appearance. This image is then used as a visual prompt in conjunction with the input text prompt to jointly supervise the 3D generation process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Score Distillation Sampling with Visual Prompt. We now present visual-prompted score distillation sampling that dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.849, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.328, + 0.782, + 0.342 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed VP3D framework for visual prompted text-to-3D generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.474, + 0.622 + ], + "angle": 0, + "content": "tillation knowledge from a pre-trained diffusion model to optimize a 3D model by considering inputs not only from a text prompt \\(y\\) but also from a visual prompt \\(v\\). To be clear, we restructure the standard SDS-based text-to-3D pipeline by utilizing an image-conditioned diffusion model [43] to trigger visual prompt-guided text-to-3D generation. Technically, the visual prompt is first converted to a global image embedding \\(\\mathbf{z}_v\\) by the CLIP image encoder [29] and a following projection network. This image embedding represents the rich content and style of the visual prompt and has the same dimension as the text embedding \\(\\mathbf{z}_y\\) used in the pre-trained text-to-image diffusion model (Stable Diffusion). Following SDS, we first add noise \\(\\epsilon\\) to the rendered image of the underlying 3D model according to the random sampled time step \\(t\\) to get a noised image \\(\\mathbf{x}_t\\). Then \\(\\mathbf{x}_t\\) is input to the diffusion model along with the conditional visual prompt embedding \\(\\mathbf{z}_v\\) and text prompt embedding \\(\\mathbf{z}_y\\) to estimate the added noise as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.628, + 0.468, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {e}} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}\\right) = \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right) \\\\ + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\lambda * \\mathbf {z} _ {v}\\right)\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right)), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.809 + ], + "angle": 0, + "content": "where \\(s\\) is the classifier-free guidance weight, \\(\\lambda \\in [0,1]\\) is the visual prompt condition weight, \\(\\phi\\) is the parameter of the pre-trained noise predictor \\(\\epsilon_{\\phi}\\) and \\(\\epsilon_{\\phi}(\\mathbf{x}_t; t, \\emptyset, \\emptyset)\\) denotes the noise prediction without conditioning. In this way, our proposed method explicitly incorporates the visual prompt and text prompt in a unified fashion for text-to-3D generation. Consequently, the final gradient of our introduced visual-prompted score distillation sampling (VP-SDS) loss \\(\\theta\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.818, + 0.47, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {V P - S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\tilde {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.336, + 0.871 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a scheduling coefficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Comparison with SDS. Comparing the update gradient of SDS (Eq. 3) and VP-SDS (Eq. 5), SDS is a special case" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.577 + ], + "angle": 0, + "content": "of our VP-SDS by setting \\(\\lambda = 0\\) where the visual prompt condition is neglected. In accordance with the theoretical analysis presented in [27, 39], the mode-seeking nature of SDS necessitates a large CFG to ensure that the pre-trained diffusion model \\(\\epsilon_{\\phi}\\) delivers a \"sharp\" updating direction for the underlying 3D model. Nevertheless, a large CFG, in turn, results in poor-quality samples and thus a \"degraded\" update direction. In contrast, VP-SDS leverages the additional visual prompt to narrow down the distillation space of \\(\\epsilon_{\\phi}\\) into a more compact region that aligns tightly with the visual prompt. Meanwhile, the distillation space is also refined by the visual prompt as it reflects realistic appearances with rich details. Therefore, the updating direction derived from our VP-SDS is not only \"sharp\" but also \"fine\", which can obtain much better 3D generation results than SDS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Notably, a recent work ProlificDreamer [39] presents variational score distillation (VSD) to address the aforementioned issues in SDS. However, VSD needs to train an additional diffusion model using LoRA [15] during the optimization process, which incurs a considerable computational overhead compared to SDS. Instead, the additional computational cost of our VP-SDS is nearly negligible, making it computationally more efficient than VSD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "View-dependent Visual Prompting. Apart from the oversaturation problem discussed above, existing text-to-3D methods are known to also suffer from the multi-view inconsistency problem (e.g., the multi-face Janus problem). This arises from the fact that the underlying prior diffusion model is exclusively trained on individual 2D images and therefore lacks 3D awareness. To alleviate this issue, existing text-to-3D methods [17, 27, 38, 39] always employ diffusion loss with view-dependent text conditioning, which is to append \"front view\", \"side view\", or \"back view\" to the input text based on the location of the randomly sampled camera. Inspired by this, we devise a view-dependent visual prompting strategy to further mitigate the view incon" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "sistency problem in collaboration with our introduced VP-SDS. Technically, given the input visual prompt (assuming it is shot from the front view), we use a view-conditioned 2D diffusion model, Zero-1-to-3 [19], to transform it into left-side, right-side and backward views. Then we fed different visual prompts into VP-SDS (Eq. 5) depending on the corresponding sampled camera viewpoint. For instance, when the azimuth angle \\(\\gamma_{cam} \\in [0^{\\circ}, 360^{\\circ}]\\) of the camera position falls in the range near \\(180^{\\circ}\\) (\\(0^{\\circ}\\) denotes the front view), we feed the generated back view counterpart of the input visual prompt into Eq 5. In this way, the inherent 3D geometry information contained in the multi-view visual prompts is encoded into the 3D representation learning through view-dependent VP-SDS, leading to better view consistency in the 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.327, + 0.367, + 0.343 + ], + "angle": 0, + "content": "3.3. Learning with Reward Feedback" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.426 + ], + "angle": 0, + "content": "To further encourage rendered images of the underlying 3D model that are high fidelity and well aligned with the input visual prompt and text prompt, we devise two types of differentiable reward functions that complement the aforementioned VP-SDS objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Human Feedback Reward. Recent practice has shown the capability of improving text-to-image models with human feedback [41]. Particularly, it first trains a reward model on a large dataset comprised of human assessments of text-image pairs. Such a reward model thus has the ability to measure the quality of the generated samples in terms of both image fidelity and image-text alignment. Consequently, it can then be used to fine-tune diffusion models to maximize the predicted scores of the reward model through differentiable reward functions, leading to better generation results. Motivated by this, we go one step further to utilize the open-sourced reward model \\(\\mathbf{r}\\) in ImageReward [41] for text-to-3D generation. Specifically, we introduce a human feedback reward loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.648, + 0.47, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h f - r e w a r d} = \\mathbb {E} _ {\\mathbf {c}} [ \\psi (\\mathbf {r} (\\mathbf {x}, y)) ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} = g(\\theta; \\mathbf{c})\\) is a rendered image by the underlying 3D model \\(\\theta\\) from an arbitrary viewpoint \\(\\mathbf{c}\\), \\(y\\) is the conditional text prompt and \\(\\psi\\) is a differentiable reward-to-loss map function as in [41]. Intuitively, minimizing the loss in Eq. 6 encourages the rendered image \\(\\mathbf{x}\\) to get a higher reward score from the reward model \\(\\mathbf{r}\\), which means the underlying 3D model should update towards the refined direction where the renderings have high appearance fidelity and faithfully match the input text prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Visual Consistency Reward. Given that the above human feedback reward only takes into account the input text prompt, we further devised a visual consistency reward to fully leverage the visual prompt as well, since text prompts cannot capture all appearance details. Technically, we adopt a pre-trained self-supervised vision transformer DINO-ViT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.153 + ], + "angle": 0, + "content": "[2] to extract the visual features \\( F_{dino}(v) \\) and \\( F_{dino}(\\mathbf{x}) \\) of the input visual prompt \\( v \\) and rendered image \\( \\mathbf{x} \\), respectively. Then we penalize the feature-wise difference between them at the visual prompt viewpoint:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.163, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v c - r e w a r d} = \\left\\| F _ {d i n o} (\\mathbf {x}) - F _ {d i n o} (\\mathbf {v}) \\right\\| ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.893, + 0.239 + ], + "angle": 0, + "content": "By imposing such visual consistency loss, we encourage the underlying 3D model to adhere to the plausible shape and appearance properties conveyed by the visual prompt." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.248, + 0.785, + 0.264 + ], + "angle": 0, + "content": "3.4. 3D Representation and Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Inspired by [17], we adopt a two-stage coarse-to-fine framework for text-to-3D generation with two different 3D scene representations. At the coarse stage, we leverage InstantNGP [24] as 3D representation, which is much faster to optimize compared to the vanilla NeRF [23] and can recover complex geometry. In the fine stage, we leverage DMTet as the 3D representation to further optimize a high-fidelity mesh and texture. Specifically, the 3D shape and texture represented in DMTet are first initialized from the density field and color field of the coarse stage, respectively [17]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.482 + ], + "angle": 0, + "content": "During the optimization process in each stage, we first render images from the underlying 3D model through differentiable rasterizers at arbitrary camera poses and optimize the 3D model with a combination of losses:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.496, + 0.891, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n e}} = \\mathcal {L} _ {V P - S D S} + \\lambda_ {1} \\mathcal {L} _ {v c - r e w a r d} + \\lambda_ {2} \\mathcal {L} _ {h f - r e w a r d}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.527, + 0.778, + 0.542 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are the trade-off parameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.555, + 0.634, + 0.573 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.703 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of our VP3D for text-to-3D generation via extensive empirical evaluations. We first show both quantitative and qualitative results of VP3D in comparison to existing techniques on the newly released text-to-3D benchmark (T\\(^3\\)Bench [11]). Next, we conduct ablation studies to validate each design in VP3D. Finally, we demonstrate the extended capability of VP3D for stylized text-to-3D generation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.712, + 0.707, + 0.728 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In the coarse and fine stage, the underlying 3D models are both optimized for 5000 iterations using Adam optimizer with 0.001 learning rate. The rendering resolutions are set to \\(128 \\times 128\\) and \\(512 \\times 512\\) for coarse and fine stage, respectively. We implement the underlying Instant-NGP and DMTet 3D representation mainly based on the Stable-DreamFusion codebase [36]. \\(\\lambda_{1}\\) is set to 0.1 in the coarse stage and 0.01 in the fine stage. \\(\\lambda_{2}\\) is linearly increased from 0.001 to 0.01 during the optimization process. The visual prompt condition weight is set to 0.5 in all experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4900" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.089, + 0.72, + 0.104 + ], + "angle": 0, + "content": "Table 1. The quantitative results of our method and baselines on \\( {\\mathrm{T}}^{3} \\) Bench [11]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.107, + 0.895, + 0.261 + ], + "angle": 0, + "content": "
MethodSingle ObjectSingle Object with SurroundingsMultiple Objects
QualityAlignmentAverageQualityAlignmentAverageQualityAlignmentAverage
DreamFusion [27]24.924.024.419.329.824.617.314.816.1
SJC [38]26.323.024.717.322.319.817.75.811.7
LatentNeRF [22]34.232.033.123.737.530.621.719.520.6
Fantasia3D [4]29.223.526.421.932.027.022.714.318.5
ProlificDreamer [39]51.147.849.442.547.044.845.725.835.8
Magic3D [17]38.735.337.029.841.035.426.624.825.7
VP3D (Ours)54.852.253.545.450.848.149.131.540.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Evaluation Protocol. Existing text-to-3D generation works commonly examine their methods over the CLIP R-Precision score [16], which is an automated metric for the consistency of rendered images with respect to the input text. However, this text-image alignment-based metric cannot faithfully represent the overall 3D quality. For example, CLIP-based text-to-3D methods can also achieve high CLIP R-Precision scores even if the resulting 3D scenes are unrealistic and severely distorted [27]. Taking this into account, we instead conduct experiments on a newly open-sourced benchmark: \\(\\mathrm{T}^3\\) Bench [11], which is the first comprehensive text-to-3D benchmark containing 300 diverse text prompts of three categories (single object, single object with surroundings, and multiple objects)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.642 + ], + "angle": 0, + "content": "\\(\\mathrm{T}^{3}\\) Bench provides two automatic metrics (quality and alignment) based on the rendered multi-view images to assess the subjective quality and text alignment. The quality metric utilizes a combination of multi-view text-image scores and regional convolution to effectively identify quality and view inconsistency. The alignment metric employs a 3D captioning model and a Large Language Model (i.e., GPT-4) to access text-3D consistency. Following this, we also leverage the quality and alignment metric to quantitatively compare our VP3D against baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baselines. To evaluate our method, we compare our VP3D with six state-of-the-art text-to-3D generation methods: DreamFusion [27], SJC [38], LatentNeRF [22], Fantasy3D [4], Magic3D [17] and ProlificDreamer [39]. Specifically, DreamFusion [27] firstly introduces score distillation sampling (SDS) that enables leveraging 2D diffusion model (Imagen [14]) to optimize a NeRF [23]. SJC [38] concurrently addresses the out-of-distribution problem in SDS and utilizes an open-sourced diffusion model (Stable Diffusion) to optimize a voxel NeRF. Latent-NeRF [22] first brings NeRF to the latent space to harmonize with latent diffusion models, then refines it in pixel space. Magic3D [17] extends DreamFusion with a coarse-to-fine framework that first optimizes a low-resolution NeRF model and then a high-resolution DMTet model via SDS. Fantasy3D [4] disentangles the SDS-based 3D learning into geometry and appearance learning. ProlificDreamer [39] upgrades Dream" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.895, + 0.323 + ], + "angle": 0, + "content": "Fusion by a variational score distillation (VSD) loss that treats the underlying 3D scene as a random variable instead of a single point as in SDS." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.333, + 0.695, + 0.349 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.629 + ], + "angle": 0, + "content": "The quantitative performance comparisons of different methods for text-to-3D generation are summarized in Table 1. Overall, our VP3D consistently achieves better performances against existing techniques across all evaluation metrics and prompt categories. Remarkably, VP3D achieves an absolute quality-alignment average score improvement of \\(4.1\\%\\), \\(3.3\\%\\), and \\(4.5\\%\\) against the best competitor ProlificDreamer across the three text prompt categories, respectively, which validates the effectiveness of our overall proposals. More importantly, while VP3D employs the same NeRF & DMTet 3D representation and coarse-t fine training scheme as the baseline method Magic3D, it significantly outperforms Magic3D by achieving \\(53.5\\%\\), \\(48.1\\%\\), and \\(40.3\\%\\) average scores, representing a substantial improvement over Magic3D's average scores of \\(37.0\\%\\), \\(35.4\\%\\), and \\(25.7\\%\\). The results generally highlight the key advantage of introducing visual prompts in lifting 2D diffusion models to perform text-to-3D generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Specifically, DreamFusion and SJC enable the zero-shot learning of implicit 3D models by distilling prior knowledge from 2D diffusion models. However, the generated 3D scenes have relatively low quality and alignment scores, especially in complex scenarios where the text prompt contains multiple objects or surroundings. Latent-NeRF employs score distillation in the latent space and then back to pixel space to further refine the 3D model, leading to better results. The aforementioned three methods only utilize implicit 3D representations (NeRFs). In contrast, Magic3D adopts textured mesh DMTet as 3D representation for enabling high-resolution optimization and exhibits better performances across all three prompt categories. Fantasy3D also capitalizes on DMTet for geometry learning and then leverages BRDF for appearance learning in a disentangled manner. While Fantasy3D achieves better average scores than DreamFusion and SJC, it fails to create high-fidelity results in complex scenes (e.g., \"multiple objects\"). Pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "4901" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.123, + 0.147, + 0.138 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.091, + 0.849, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.6, + 0.893, + 0.658 + ], + "angle": 0, + "content": "Figure 3. Comparisons on qualitative results of our VP3D with other text-to-3D techniques on \\(\\mathrm{T}^3\\) Bench [11]. The prompts are (a) \"A fuzzy pink flamingo lawn ornament\", (b) \"A blooming potted orchid with purple flowers\", (c) \"A blue butterfly on a pink flower\", (d) \"A lighthouse on a rocky shore\", (e) \"Hot popcorn jump out from the red striped popcorn maker\", (f) \"A chef is making pizza dough in the kitchen\". (a-b), (c-d), (e-f) belongs to the Single Object, Single Object with Surr and Multi Objects category in \\(\\mathrm{T}^3\\) Bench, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.765 + ], + "angle": 0, + "content": "lificDreamer further boosts the performance by training an additional diffusion model during the optimization process to realize a principled particle-based variational score distillation loss. However, our VP3D still outperforms ProlificDreamer across all evaluation metrics and prompt sets, which confirms the effectiveness of our VP3D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.262, + 0.789 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The qualitative comparisons for text-to-3D generation are presented in Figure 3. As can be seen, our VP3D generally produces superior 3D scenes with plausible geometry and realistic textures when compared with baseline methods. Specifically, DreamFusion suffers from a severely over-saturated problem and has difficulty generating complex geometry. Magic3D and Latent-NeRF slightly allevi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.901 + ], + "angle": 0, + "content": "ate these issues through their higher-resolution DMTet and pixel space refinement, respectively. While Fantasia3D and SJC can generate richer textures than DreamFusion, the geometric quality of the generated 3D scenes falls short of expectations. Notably, ProlificDreamer trains an additional diffusion model during the optimization process to perform variational score distillation (VSD) instead of SDS, achieving satisfactory single-object objects. However, the use of VSD at times introduces excessive irrelevant information or geometry noise in more complex scenarios. In contrast, we can clearly observe that the generated 3D scenes by VP3D faithfully match the input text prompt with plausible geometry and realistic appearance, which demonstrates the superiority of VP3D over state-of-the-art methods and its ability to generate high-quality 3D content." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "4902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.47, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.456, + 0.436 + ], + "angle": 0, + "content": "Figure 4. Stylized text-to-3D generation results of our VP3D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.232, + 0.469 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.718 + ], + "angle": 0, + "content": "Here we investigate how each design in our VP3D influences the overall generation performance. We depict the qualitative results of each ablated run in Figure 5. \\(\\mathcal{L}_{SDS}\\) is our baseline model that employs vanilla score distillation sampling loss. As can be seen, the generated 3D scene is over-saturated and geometry unreasonable. Instead, when \\(\\mathcal{L}_{VP - SDS}\\) is employed, the generation quality is clearly enhanced in terms of both geometry and appearance. This highlights the critical effectiveness of our proposed visual-prompted score distillation sampling. Nevertheless, the resulting 3D scenes by \\(\\mathcal{L}_{VP - SDS}\\) are still not satisfying enough. By utilizing additional visual consistency and human feedback reward functions \\(\\mathcal{L}_{vc - reward}\\) (Eq. 7) and \\(\\mathcal{L}_{hf - reward}\\) (Eq. 6), the generation quality is gradually improved. The results basically validate the effectiveness of these two complementary factors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.453, + 0.743 + ], + "angle": 0, + "content": "4.5. Extension to Stylized Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this section, we demonstrate that another advantage of our VP3D is its remarkable versatility in 3D generation as it can be readily adapted for a new task of stylized text-to-3D generation. The main difference is that the visual prompt is no longer generated from the text prompt but from a user-specified reference image. We also empirically discard the loss in Eq. 6 to eliminate the strictly text-image alignment constraint. In this way, our VP3D can integrate the visual cues contained in the reference image into text-to-3D generation and produce a stylized 3D asset. This asset not only" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.885, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Figure 5. Comparisons on qualitative results by using different ablated runs of our VP3D. The text prompts are (a) \"A broken tablespoon lies next to an empty sugar bowl\" and (b) \"A chameleon perched on a tree branch\"." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.334 + ], + "angle": 0, + "content": "semantically aligns with the text prompt but also reflects visual and geometry properties in the reference image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Figure 4 shows our stylized text-to-3D generation results. Our VP3D can generate diverse and stylized 3D assets by giving different visual prompts to the same text prompt. As shown in Figure 4 (a-b), the generated result semantically is a rabbit that adheres to the text prompt but also inherits some visual cues of the visual prompt. To be clear, the generated 3D rabbits have somewhat consistent geometry pose and appearance texture with the object in the visual prompt. For example, in Figure 4 (b), the generated rabbit mirrors the \"hugging pose\" of the reference image and also has the same style of \"crescent-shaped eyebrows\" and \"yellow plaid jacket\" as in the reference image. In Figure 4 (c-d), we showcase the versatility of our VP3D by seamlessly blending styles from different visual prompts. Take Figure 4 (d) as an instance, we use the leopard image as a visual prompt in the coarse stage and then replace it with an oil painting image in the fine stage. Our VP3D finally resulted in a 3D rabbit that not only has a consistent pose with the leopard but also a colorful oil painting style texture. The stylized 3D generation ability distinct our VP3D from previous text-to-3D approaches and can lead to more creative and diverse 3D content creation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.62, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose VP3D, a new paradigm for text-to-3D generation by leveraging 2D visual prompts. We first capitalize on 2D diffusion models to generate a high-quality image from input text. This image then acts as a visual prompt to strengthen the 3D model learning with our devised visual-prompted score distillation sampling. Meanwhile, we introduce additional human feedback and visual consistency reward functions to encourage the semantic and appearance consistency between the 3D model and input visual & text prompt. Both qualitative and quantitative comparisons on the \\(\\mathrm{T}^3\\) Bench benchmark demonstrate the superiority of our VP3D over existing SOTA techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4903" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Jingwen Chen, Yingwei Pan, Ting Yao, and Tao Mei. Controlstyle: Text-driven stylized image generation using diffusion priors. In ACM Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.298 + ], + "angle": 0, + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[5] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Animating your life: Real-time video-to-animation translation. In ACM MM, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[6] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM MM, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.423 + ], + "angle": 0, + "content": "[7] Yang Chen, Jingwen Chen, Yingwei Pan, Xinmei Tian, and Tao Mei. 3d creation at your fingertips: From text or image to 3d assets. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[8] Yang Chen, Yingwei Pan, Yehao Li, Ting Yao, and Tao Mei. Control3d: Towards controllable text-to-3d generation. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[9] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[11] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[15] EJ Hu, Y Shen, P Wallis, Z Allen-Zhu, Y Li, S Wang, L Wang, and W Chen. Low-rank adaptation of large language models, arxiv, 2021. arXiv preprint arXiv:2106.09685, 10.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[20] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-conditional diffusion networks for image captioning. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[21] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[22] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[26] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4904" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[36] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[37] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[38] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[39] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[40] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[42] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4905" + } + ] +] \ No newline at end of file diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_origin.pdf b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aa024903df3cc0434cd202f4024149a01ce72fb3 --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/d6596473-eeb8-454d-940e-3cbe50e191d2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80da17778115ef0a7a7a845d8ce407730fed17db84ae6a13fe29a1d06d0bfc6a +size 3325586 diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/full.md b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7442909934f2f71b8bb527b01694b134e0e00857 --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/full.md @@ -0,0 +1,260 @@ +# VP3D: Unleashing 2D Visual Prompt for Text-to-3D Generation + +Yang Chen†, Yingwei Pan†, Haibo Yang§, Ting Yao†, and Tao Mei† +†HiDream.ai Inc. + +$^{8}$ School of Computer Science, Fudan University, China + +{clenyang, pandy}@hidream.ai, yanghaibo.fdu@gmail.com, {tiyao, tmei}@hidream.ai + +# Abstract + +Recent innovations on text-to-3D generation have featured Score Distillation Sampling (SDS), which enables the zero-shot learning of implicit 3D models (NeRF) by directly distilling prior knowledge from 2D diffusion models. However, current SDS-based models still struggle with intricate text prompts and commonly result in distorted 3D models with unrealistic textures or cross-view inconsistency issues. In this work, we introduce a novel Visual Prompt-guided text-to-3D diffusion model (VP3D) that explicitly unleashes the visual appearance knowledge in 2D visual prompt to boost text-to-3D generation. Instead of solely supervising SDS with text prompt, VP3D first capitalizes on 2D diffusion model to generate a high-quality image from input text, which subsequently acts as visual prompt to strengthen SDS optimization with explicit visual appearance. Meanwhile, we couple the SDS optimization with additional differentiable reward function that encourages rendering images of 3D models to better visually align with 2D visual prompt and semantically match with text prompt. Through extensive experiments, we show that the 2D Visual Prompt in our VP3D significantly eases the learning of visual appearance of 3D models and thus leads to higher visual fidelity with more detailed textures. It is also appealing in view that when replacing the self-generating visual prompt with a given reference image, VP3D is able to trigger a new task of stylized text-to-3D generation. Our project page is available at https://vp3d-cvpr24.github.io. + +# 1. Introduction + +Generative Artificial Intelligence (especially for vision content generation) has aroused great attention in computer vision field [5, 6, 20, 26], leading to impressive advancements in text-to-image [30-32] and text-to-video generation [10, 14, 34]. These accomplishments can be attributed to the availability of large-scale image-text and video-text pair data [1, 33] and the emergence of robust diffusion-based generative models [12, 13, 25, 35]. Recently, researchers + +![](images/7cdaf6826e003cb2f350a3b5c5f2ab8488a218d4a642f82154b530ff531c1345.jpg) +Text prompt: "A florist is making a bouquet with fresh flowers" + +![](images/d77737463c8c506849b33453d7653609f379d03241a87bce959eb967f923a379.jpg) +(b)ProlificDreamer + +![](images/09c4e66c1563e8cd39c1379ae4c26bcdee93215b40340be42e9dcbd23e726bc4.jpg) +(a) Magic3D +Visual prompt + +![](images/5ce9398ccec78e2746453dd7b53f0b4229098a936a39a7c64bfd9633c3e40178.jpg) +(c) VP3D (Ours) +Figure 1. Existing text-to-3D generation techniques (e.g., Magic3D [17] and ProlificDreamer [39]) often suffer from degenerated results (e.g., over-saturated appearances or inaccurate geometries). Our VP3D novelly integrates a visual prompt to strength score distillation sampling, leading to better 3D results. + +![](images/b622e2443acf43447a926aef5fb1eb2ebeeba709d72ac81186488496c75c6672.jpg) + +have gone beyond text-driven image/video generation, and begun exploring diffusion models for text-driven content creation of 3D assets (e.g., text-to-3D generation). This direction paves a new way for practical 3D content creation and has a great potential impact for numerous applications like virtual reality, gaming and Metaverse. Compared to image generation, text-to-3D generation, however, is more challenging, due to the complexities associated with intricate 3D geometry and appearance (i.e., object shapes and textures). Moreover, the collection and annotation of 3D data are somewhat resourcefully expensive and thus cannot be easily scaled up to billion level as image data. + +To tackle this issue, a pioneering text-to-3D work (DreamFusion [27]) presents the first attempt of exploiting an off-the-shelf text-to-image diffusion model to generate promising 3D assets in a zero-shot fashion. The key design behind such success is Score Distillation Sampling (SDS), which directly optimizes the implicit 3D model of + +Neural Radiance Field (NeRF) with prior knowledge distilled from 2D diffusion model. Nevertheless, such distilled prior knowledge is merely driven by the input text prompt, and it is non-trivial to learn high-quality NeRF with distilled SDS supervision. Although several subsequent works [4, 17, 22, 38, 39] further upgrade SDS, this kind of SDS-based solution still results in degenerated 3D models with unrealistic/less-detailed textures, especially when feeding intricate text prompts (as seen in Figure 1 (a-b)). + +In this work, we propose to mitigate this limitation through a unique design of visual prompt-guided text-to-3D diffusion model, namely VP3D. Intuitively, "a picture is worth a thousand words." That is, a single image can convey human intentions of visual content creation (e.g., the visual appearance or semantic structure) more effectively than textual sentences. This motivates us to introduce additional guidance of visual prompt, and thus decompose the typical single-shot text-to-3D process into two cascaded processes: first text-to-image generation, and then (text plus image)-to-3D generation. In particular, VP3D first leverages off-the-shelf text-to-image diffusion models to produce a high-fidelity image that reflects extremely realistic appearance with rich details. In the latter process, this synthetic image further acts as 2D visual prompt to supervise SDS optimization of NeRF, coupled with the input text prompt. At the same time, a differentiable reward function is additionally utilized to encourage the rendering images of NeRF to be better aligned with 2D visual prompt (visual appearance consistency) and text prompt (semantic consistency). As illustrated in Figure 1 (c), we show that the novel visual prompt-guided diffusion process in VP3D significantly enhances the visual fidelity of 3D assets with realistic and rich texture details. Meanwhile, when easing the learning of visual appearance of 3D assets via visual prompt guidance, the optimization of NeRF will focus more on the modeling of geometry, leading to better 3D sharps with cross-view consistency. We believe that the ability of unleashing high-quality visual knowledge in 2D visual prompt is potentially a new paradigm of text-to-3D generation. + +As a by-product, we also demonstrate that our VP3D can be readily adapted for a new task of stylized text-to-3D generation. Intriguingly, we simply replace the self-generating image in VP3D with a user-given reference image, and treat it as a new visual prompt to trigger (text plus image)-to-3D generation. In this way, our VP3D is able to produce a stylized 3D asset, which not only semantically aligns with text prompt but also shares similar geometric & visual style as the reference image. + +# 2. Related Work + +Text-to-3D generation. Significant advancements have been witnessed in text-to-image generation with 2D diffusion models in recent years [3, 12, 13, 25, 30-32, 35]. + +However, extending these capabilities to 3D content generation poses a substantial challenge, primarily due to the absence of large-scale paired text-3D datasets. To mitigate the reliance on extensive training data, recent works try to accomplish zero-shot text-to-3D generation [4, 7, 8, 17, 22, 27, 38, 39, 42]. Specifically, the pioneering work DreamFusion [27] showcased remarkable achievements in text-to-3D generation through pre-trained text-to-image diffusion models. SJC [38] concurrently addressed the out-of-distribution problem in lifting 2D diffusion models to perform text-to-3D generation. Following these, several subsequent works have strived to enhance text-to-3D generation further. For instance, Latent-NeRF [22] proposed to incorporate a sketch shape to guide the 3D generation directly in the latent space of a latent diffusion model. Magic3D [17] presented a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors to learn the underlying 3D representation. Control3D [8] proposed to enhance user controllability in text-to-3D generation by incorporating additional hand-drawn sketch conditions. ProlificDreamer [39] presented a principled particle-based variational framework to improve the generation quality. + +Unlike previous works, we formulate the text-to-3D generation process from a new perspective. We first leverage the off-the-shelf text-to-image diffusion models to generate a high-quality image that faithfully matches the input text prompt. This synthetic reference image then serves as a complementary input alongside the text, synergistically guiding the 3D learning process. Moreover, we showcase the remarkable versatility of this novel architecture by effortlessly extending its capabilities to the realm of stylized text-to-3D generation. The resulting 3D asset not only exhibits semantic alignment with the provided text prompt but also masterfully captures the visual style of the reference image. This capability marks another pivotal distinction between our VP3D and previous text-to-3D approaches. + +Image-to-3D generation. Recently, prior works RealFusion [21], NeuralLift-360 [40] and NeRDi [9] leverage 2D diffusion models to achieve image-to-3D generation. The following work Make-IT-3D [37] proposed a two-stage optimization framework to improve the generation quality further. Zero-1-to-3 [19] finetuned the Stable Diffusion model to enable generating novel views of the input image. It can then be used as a 3D prior model to achieve high-quality image-to-3D generation. Inspired by this, Magic123 [28] proposed to use 2D and 3D priors simultaneously to generate faithful 3D content from the given image. One-2-3-45 [18] integrated Zero-1-to-3 and a multi-view reconstruction model to accelerate the 3D generation process. It is worth noting that our work is not targeting image-to-3D generation. We utilize a reference image to guide the text-to-3D learning process, instead of directly turning the reference image into 3D content. + +# 3. VP3D + +In this section, we elaborate the architecture of our VP3D, which introduces a novel visual prompt-guided text-to-3D diffusion model. An overview of our VP3D architecture is depicted in Figure 2. + +# 3.1. Background + +Text-to-Image Diffusion Models. Diffusion models are a family of generative models that are trained to gradually transform Gaussian noise into samples from a target distribution [13]. Given a target data distribution $q(\mathbf{x})$ , a forward diffusion process is defined to progressively add a small amount of Gaussian noise to the data $\mathbf{x}_0$ sampled from $q(\mathbf{x})$ . This process follows a Markov chain $q(\mathbf{x}_{1:T}) = \prod_{t=1}^{T} q(\mathbf{x}_t | \mathbf{x}_{t-1})$ and produces a sequence of latent variables $\mathbf{x}_1, \ldots, \mathbf{x}_T$ after $T$ time steps. The marginal distribution of latent variables at time step $t$ is given by $q(\mathbf{x}_t | \mathbf{x}) = \mathcal{N}(\mathbf{x}_t; \alpha_t \mathbf{x}, \sigma_t^2 \mathbf{I})$ . Thus the noisy sample $\mathbf{x}_t$ can be directly generated through the equation $\mathbf{x}_t = \alpha_t \mathbf{x} + \sigma_t^2 \epsilon$ , where $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ , $\alpha_t$ and $\sigma_t$ are chosen parameters such that $\alpha_t^2 + \sigma_t^2 = 1$ . After $T$ noise adding steps, $\mathbf{x}_T$ is equivalent to an isotropic Gaussian distribution. Then, a reverse generative process is defined to gradually "denoise" $X_T$ to reconstruct the original sample. This can be described by a Markov process $p_\phi(\mathbf{x}_{0:T}) = p(\mathbf{x}_T) \prod_{t=1}^{T} p_\phi(\mathbf{x}_{t-1} | \mathbf{x}_t)$ , with the conditional probability $p_\phi(\mathbf{x}_{t-1} | \mathbf{x}_t) = \mathcal{N}(\mathbf{x}_{t-1}; \boldsymbol{\mu}_\phi(\mathbf{x}_t, t), \boldsymbol{\Sigma}_\phi(\mathbf{x}_t, t))$ . Commonly, a UNet neural network $\epsilon_\phi(\mathbf{x}_t; t)$ with parameters $\phi$ is used to predict the noise that was used to produce $\mathbf{x}_t$ at time step $t$ . Text-to-image diffusion models build upon the above theory to condition the diffusion process with a given text prompt $y$ using classifier-free guidance (CFG) [12]. The corresponding noise predictor is remodeled by: + +$$ +\hat {\boldsymbol {\epsilon}} _ {\phi} (\mathbf {x} _ {t}; \mathbf {z} _ {y}, t) = \boldsymbol {\epsilon} _ {\phi} (\mathbf {x} _ {t}; t, \emptyset) + s * \left(\boldsymbol {\epsilon} _ {\phi} (\mathbf {x} _ {t}; t, \mathbf {z} _ {y}) - \boldsymbol {\epsilon} _ {\phi} (\mathbf {x} _ {t}; t, \emptyset)\right), \tag {1} +$$ + +where $s$ is a scale that denotes the classifier-free guidance weight, $\mathbf{z}_y$ is the corresponding text embedding of the text prompt $y$ and $\varnothing$ indicates the noise prediction without conditioning. The diffusion model $\epsilon_{\phi}$ is typically optimized by a simplified variant of the variational lower bound of the log data likelihood, which is a Mean Squared Error criterion: + +$$ +\mathcal {L} _ {\text {d i f f}} (\phi) = \mathbb {E} _ {\mathbf {x}, t, \epsilon} \left[ w (t) \| \hat {\boldsymbol {\epsilon}} _ {\phi} (\mathbf {x} _ {t}; y, t) - \epsilon \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $w(t)$ is a weighting function that depends on the timestep $t \sim \mathcal{U}(0,1)$ and $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . + +Score Distillation Sampling. A recent pioneering work called DreamFusion [27] introduces Score Distillation Sampling (SDS) that enables leveraging the priors of pre-trained text-to-image diffusion models to facilitate text-to-3D generation. Specifically, let $\theta$ be the learnable parameters of a + +3D model (e.g., NeRF) and $g$ be a differentiable rendering function that can render an image $\mathbf{x} = g(\theta; \mathbf{c})$ from the 3D model $\theta$ at a camera viewpoint $\mathbf{c}$ . SDS introduces a loss function $\mathcal{L}_{SDS}$ to optimize the parameters $\theta$ . Its gradient is defined as follows: + +$$ +\nabla_ {\theta} \mathcal {L} _ {S D S} = \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} (\mathbf {x} _ {t}; t, \mathbf {z} _ {y}) - \epsilon) \frac {\partial \mathbf {x}}{\partial \theta} ], \tag {3} +$$ + +where $\mathbf{x}_t$ is obtained by perturbing the rendered image $\mathbf{x}$ with a Gaussian noise $\epsilon$ corresponding to the $t$ -th timestep of the forward diffusion process, $\mathbf{z}_y$ is the conditioned text embedding of given text prompt $y$ . Intuitively, the SDS loss estimates an update direction in which the noised version of rendered image $\mathbf{x}$ should be moved towards a denser region in the distribution of real images (aligned with the conditional text prompt $y$ ). By randomly sampling views and backpropagating the gradient in Eq. 3 to the parameters $\theta$ through the differential parametric function $g$ , this approach eventually results in a 3D model that resembles the text. + +# 3.2. Visual-prompted Score Distillation Sampling + +Visual Prompt Generation. As aforementioned, score distillation sampling plays a key role in text-to-3D generation. Nevertheless, empirical observations [11, 27, 39] reveal that SDS still results in degenerated 3D models especially when feeding intricate text prompts. First, SDS-generated results often suffer from over-saturation issues. These issues are, in part, attributed to the necessity of employing a large CFG value (i.e., 100) within the SDS framework [27, 39]. A Large CFG value narrows down the score distillation space to more text-relevant areas. This can mitigate the divergence of diffusion priors in the optimization process, thereby fostering enhanced stability in 3D representation learning. However, this is at the cost of less realistic and diversity generation results, as large CFG values are known to yield over-saturated results [39]. Second, results generated by SDS still face the risk of text-3D misalignment, such as missing key elements in the scene, especially when text prompts contain multiple objects with specific attributes. + +A fundamental reason behind the aforementioned issues may lie in the substantial distribution gap between text and 3D modalities. Thus it is non-trivial to directly learn a meaningful 3D scene solely based on a single text prompt. This insight motivates us to introduce an additional visual prompt as a bridge to explicitly establish a connection between the text input and the desired 3D output. Particularly, we leverage off-the-shelf text-to-image diffusion models (e.g., Stable Diffusion) to produce a high-fidelity image that faithfully matches the input text prompt and has an extremely realistic appearance. This image is then used as a visual prompt in conjunction with the input text prompt to jointly supervise the 3D generation process. + +Score Distillation Sampling with Visual Prompt. We now present visual-prompted score distillation sampling that dis + +![](images/8e0a0fc06ceed63935551bc5d46de1b59510a36ec250a8b1719bf89054701b48.jpg) +Figure 2. An overview of the proposed VP3D framework for visual prompted text-to-3D generation. + +tillation knowledge from a pre-trained diffusion model to optimize a 3D model by considering inputs not only from a text prompt $y$ but also from a visual prompt $v$ . To be clear, we restructure the standard SDS-based text-to-3D pipeline by utilizing an image-conditioned diffusion model [43] to trigger visual prompt-guided text-to-3D generation. Technically, the visual prompt is first converted to a global image embedding $\mathbf{z}_v$ by the CLIP image encoder [29] and a following projection network. This image embedding represents the rich content and style of the visual prompt and has the same dimension as the text embedding $\mathbf{z}_y$ used in the pre-trained text-to-image diffusion model (Stable Diffusion). Following SDS, we first add noise $\epsilon$ to the rendered image of the underlying 3D model according to the random sampled time step $t$ to get a noised image $\mathbf{x}_t$ . Then $\mathbf{x}_t$ is input to the diffusion model along with the conditional visual prompt embedding $\mathbf{z}_v$ and text prompt embedding $\mathbf{z}_y$ to estimate the added noise as follows: + +$$ +\begin{array}{l} \tilde {\mathbf {e}} _ {\phi} \left(\mathbf {x} _ {t}; t, \mathbf {z} _ {y}, \mathbf {z} _ {v}\right) = \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {x} _ {t}; t, \emptyset , \emptyset\right) \\ + s * \left(\boldsymbol {\epsilon} _ {\phi} \left(\mathbf {x} _ {t}; t, \mathbf {z} _ {y}, \lambda * \mathbf {z} _ {v}\right)\right) - \boldsymbol {\epsilon} _ {\phi} \left(\mathbf {x} _ {t}; t, \emptyset , \emptyset\right)), \tag {4} \\ \end{array} +$$ + +where $s$ is the classifier-free guidance weight, $\lambda \in [0,1]$ is the visual prompt condition weight, $\phi$ is the parameter of the pre-trained noise predictor $\epsilon_{\phi}$ and $\epsilon_{\phi}(\mathbf{x}_t; t, \emptyset, \emptyset)$ denotes the noise prediction without conditioning. In this way, our proposed method explicitly incorporates the visual prompt and text prompt in a unified fashion for text-to-3D generation. Consequently, the final gradient of our introduced visual-prompted score distillation sampling (VP-SDS) loss $\theta$ is expressed as: + +$$ +\nabla_ {\theta} \mathcal {L} _ {V P - S D S} = \mathbb {E} _ {t, \epsilon} [ w (t) (\tilde {\epsilon} _ {\phi} (\mathbf {x} _ {t}; t, \mathbf {z} _ {y}, \mathbf {z} _ {v}) - \epsilon) \frac {\partial \mathbf {x}}{\partial \theta} ], (5) +$$ + +where $w(t)$ is a scheduling coefficient. + +Comparison with SDS. Comparing the update gradient of SDS (Eq. 3) and VP-SDS (Eq. 5), SDS is a special case + +of our VP-SDS by setting $\lambda = 0$ where the visual prompt condition is neglected. In accordance with the theoretical analysis presented in [27, 39], the mode-seeking nature of SDS necessitates a large CFG to ensure that the pre-trained diffusion model $\epsilon_{\phi}$ delivers a "sharp" updating direction for the underlying 3D model. Nevertheless, a large CFG, in turn, results in poor-quality samples and thus a "degraded" update direction. In contrast, VP-SDS leverages the additional visual prompt to narrow down the distillation space of $\epsilon_{\phi}$ into a more compact region that aligns tightly with the visual prompt. Meanwhile, the distillation space is also refined by the visual prompt as it reflects realistic appearances with rich details. Therefore, the updating direction derived from our VP-SDS is not only "sharp" but also "fine", which can obtain much better 3D generation results than SDS. + +Notably, a recent work ProlificDreamer [39] presents variational score distillation (VSD) to address the aforementioned issues in SDS. However, VSD needs to train an additional diffusion model using LoRA [15] during the optimization process, which incurs a considerable computational overhead compared to SDS. Instead, the additional computational cost of our VP-SDS is nearly negligible, making it computationally more efficient than VSD. + +View-dependent Visual Prompting. Apart from the oversaturation problem discussed above, existing text-to-3D methods are known to also suffer from the multi-view inconsistency problem (e.g., the multi-face Janus problem). This arises from the fact that the underlying prior diffusion model is exclusively trained on individual 2D images and therefore lacks 3D awareness. To alleviate this issue, existing text-to-3D methods [17, 27, 38, 39] always employ diffusion loss with view-dependent text conditioning, which is to append "front view", "side view", or "back view" to the input text based on the location of the randomly sampled camera. Inspired by this, we devise a view-dependent visual prompting strategy to further mitigate the view incon + +sistency problem in collaboration with our introduced VP-SDS. Technically, given the input visual prompt (assuming it is shot from the front view), we use a view-conditioned 2D diffusion model, Zero-1-to-3 [19], to transform it into left-side, right-side and backward views. Then we fed different visual prompts into VP-SDS (Eq. 5) depending on the corresponding sampled camera viewpoint. For instance, when the azimuth angle $\gamma_{cam} \in [0^{\circ}, 360^{\circ}]$ of the camera position falls in the range near $180^{\circ}$ ( $0^{\circ}$ denotes the front view), we feed the generated back view counterpart of the input visual prompt into Eq 5. In this way, the inherent 3D geometry information contained in the multi-view visual prompts is encoded into the 3D representation learning through view-dependent VP-SDS, leading to better view consistency in the 3D generation. + +# 3.3. Learning with Reward Feedback + +To further encourage rendered images of the underlying 3D model that are high fidelity and well aligned with the input visual prompt and text prompt, we devise two types of differentiable reward functions that complement the aforementioned VP-SDS objective. + +Human Feedback Reward. Recent practice has shown the capability of improving text-to-image models with human feedback [41]. Particularly, it first trains a reward model on a large dataset comprised of human assessments of text-image pairs. Such a reward model thus has the ability to measure the quality of the generated samples in terms of both image fidelity and image-text alignment. Consequently, it can then be used to fine-tune diffusion models to maximize the predicted scores of the reward model through differentiable reward functions, leading to better generation results. Motivated by this, we go one step further to utilize the open-sourced reward model $\mathbf{r}$ in ImageReward [41] for text-to-3D generation. Specifically, we introduce a human feedback reward loss as follows: + +$$ +\mathcal {L} _ {h f - r e w a r d} = \mathbb {E} _ {\mathbf {c}} [ \psi (\mathbf {r} (\mathbf {x}, y)) ], \tag {6} +$$ + +where $\mathbf{x} = g(\theta; \mathbf{c})$ is a rendered image by the underlying 3D model $\theta$ from an arbitrary viewpoint $\mathbf{c}$ , $y$ is the conditional text prompt and $\psi$ is a differentiable reward-to-loss map function as in [41]. Intuitively, minimizing the loss in Eq. 6 encourages the rendered image $\mathbf{x}$ to get a higher reward score from the reward model $\mathbf{r}$ , which means the underlying 3D model should update towards the refined direction where the renderings have high appearance fidelity and faithfully match the input text prompt. + +Visual Consistency Reward. Given that the above human feedback reward only takes into account the input text prompt, we further devised a visual consistency reward to fully leverage the visual prompt as well, since text prompts cannot capture all appearance details. Technically, we adopt a pre-trained self-supervised vision transformer DINO-ViT + +[2] to extract the visual features $F_{dino}(v)$ and $F_{dino}(\mathbf{x})$ of the input visual prompt $v$ and rendered image $\mathbf{x}$ , respectively. Then we penalize the feature-wise difference between them at the visual prompt viewpoint: + +$$ +\mathcal {L} _ {v c - r e w a r d} = \left\| F _ {d i n o} (\mathbf {x}) - F _ {d i n o} (\mathbf {v}) \right\| ^ {2}. \tag {7} +$$ + +By imposing such visual consistency loss, we encourage the underlying 3D model to adhere to the plausible shape and appearance properties conveyed by the visual prompt. + +# 3.4. 3D Representation and Training + +Inspired by [17], we adopt a two-stage coarse-to-fine framework for text-to-3D generation with two different 3D scene representations. At the coarse stage, we leverage InstantNGP [24] as 3D representation, which is much faster to optimize compared to the vanilla NeRF [23] and can recover complex geometry. In the fine stage, we leverage DMTet as the 3D representation to further optimize a high-fidelity mesh and texture. Specifically, the 3D shape and texture represented in DMTet are first initialized from the density field and color field of the coarse stage, respectively [17]. + +During the optimization process in each stage, we first render images from the underlying 3D model through differentiable rasterizers at arbitrary camera poses and optimize the 3D model with a combination of losses: + +$$ +\mathcal {L} _ {\text {f i n e}} = \mathcal {L} _ {V P - S D S} + \lambda_ {1} \mathcal {L} _ {v c - r e w a r d} + \lambda_ {2} \mathcal {L} _ {h f - r e w a r d}, \tag {8} +$$ + +where $\lambda_1, \lambda_2$ are the trade-off parameters. + +# 4. Experiments + +In this section, we evaluate the effectiveness of our VP3D for text-to-3D generation via extensive empirical evaluations. We first show both quantitative and qualitative results of VP3D in comparison to existing techniques on the newly released text-to-3D benchmark (T $^3$ Bench [11]). Next, we conduct ablation studies to validate each design in VP3D. Finally, we demonstrate the extended capability of VP3D for stylized text-to-3D generation. + +# 4.1. Experimental Settings + +Implementation Details. In the coarse and fine stage, the underlying 3D models are both optimized for 5000 iterations using Adam optimizer with 0.001 learning rate. The rendering resolutions are set to $128 \times 128$ and $512 \times 512$ for coarse and fine stage, respectively. We implement the underlying Instant-NGP and DMTet 3D representation mainly based on the Stable-DreamFusion codebase [36]. $\lambda_{1}$ is set to 0.1 in the coarse stage and 0.01 in the fine stage. $\lambda_{2}$ is linearly increased from 0.001 to 0.01 during the optimization process. The visual prompt condition weight is set to 0.5 in all experiments. + +Table 1. The quantitative results of our method and baselines on ${\mathrm{T}}^{3}$ Bench [11]. + +
MethodSingle ObjectSingle Object with SurroundingsMultiple Objects
QualityAlignmentAverageQualityAlignmentAverageQualityAlignmentAverage
DreamFusion [27]24.924.024.419.329.824.617.314.816.1
SJC [38]26.323.024.717.322.319.817.75.811.7
LatentNeRF [22]34.232.033.123.737.530.621.719.520.6
Fantasia3D [4]29.223.526.421.932.027.022.714.318.5
ProlificDreamer [39]51.147.849.442.547.044.845.725.835.8
Magic3D [17]38.735.337.029.841.035.426.624.825.7
VP3D (Ours)54.852.253.545.450.848.149.131.540.3
+ +Evaluation Protocol. Existing text-to-3D generation works commonly examine their methods over the CLIP R-Precision score [16], which is an automated metric for the consistency of rendered images with respect to the input text. However, this text-image alignment-based metric cannot faithfully represent the overall 3D quality. For example, CLIP-based text-to-3D methods can also achieve high CLIP R-Precision scores even if the resulting 3D scenes are unrealistic and severely distorted [27]. Taking this into account, we instead conduct experiments on a newly open-sourced benchmark: $\mathrm{T}^3$ Bench [11], which is the first comprehensive text-to-3D benchmark containing 300 diverse text prompts of three categories (single object, single object with surroundings, and multiple objects). + +$\mathrm{T}^{3}$ Bench provides two automatic metrics (quality and alignment) based on the rendered multi-view images to assess the subjective quality and text alignment. The quality metric utilizes a combination of multi-view text-image scores and regional convolution to effectively identify quality and view inconsistency. The alignment metric employs a 3D captioning model and a Large Language Model (i.e., GPT-4) to access text-3D consistency. Following this, we also leverage the quality and alignment metric to quantitatively compare our VP3D against baseline methods. + +Baselines. To evaluate our method, we compare our VP3D with six state-of-the-art text-to-3D generation methods: DreamFusion [27], SJC [38], LatentNeRF [22], Fantasy3D [4], Magic3D [17] and ProlificDreamer [39]. Specifically, DreamFusion [27] firstly introduces score distillation sampling (SDS) that enables leveraging 2D diffusion model (Imagen [14]) to optimize a NeRF [23]. SJC [38] concurrently addresses the out-of-distribution problem in SDS and utilizes an open-sourced diffusion model (Stable Diffusion) to optimize a voxel NeRF. Latent-NeRF [22] first brings NeRF to the latent space to harmonize with latent diffusion models, then refines it in pixel space. Magic3D [17] extends DreamFusion with a coarse-to-fine framework that first optimizes a low-resolution NeRF model and then a high-resolution DMTet model via SDS. Fantasy3D [4] disentangles the SDS-based 3D learning into geometry and appearance learning. ProlificDreamer [39] upgrades Dream + +Fusion by a variational score distillation (VSD) loss that treats the underlying 3D scene as a random variable instead of a single point as in SDS. + +# 4.2. Quantitative Results + +The quantitative performance comparisons of different methods for text-to-3D generation are summarized in Table 1. Overall, our VP3D consistently achieves better performances against existing techniques across all evaluation metrics and prompt categories. Remarkably, VP3D achieves an absolute quality-alignment average score improvement of $4.1\%$ , $3.3\%$ , and $4.5\%$ against the best competitor ProlificDreamer across the three text prompt categories, respectively, which validates the effectiveness of our overall proposals. More importantly, while VP3D employs the same NeRF & DMTet 3D representation and coarse-t fine training scheme as the baseline method Magic3D, it significantly outperforms Magic3D by achieving $53.5\%$ , $48.1\%$ , and $40.3\%$ average scores, representing a substantial improvement over Magic3D's average scores of $37.0\%$ , $35.4\%$ , and $25.7\%$ . The results generally highlight the key advantage of introducing visual prompts in lifting 2D diffusion models to perform text-to-3D generation. + +Specifically, DreamFusion and SJC enable the zero-shot learning of implicit 3D models by distilling prior knowledge from 2D diffusion models. However, the generated 3D scenes have relatively low quality and alignment scores, especially in complex scenarios where the text prompt contains multiple objects or surroundings. Latent-NeRF employs score distillation in the latent space and then back to pixel space to further refine the 3D model, leading to better results. The aforementioned three methods only utilize implicit 3D representations (NeRFs). In contrast, Magic3D adopts textured mesh DMTet as 3D representation for enabling high-resolution optimization and exhibits better performances across all three prompt categories. Fantasy3D also capitalizes on DMTet for geometry learning and then leverages BRDF for appearance learning in a disentangled manner. While Fantasy3D achieves better average scores than DreamFusion and SJC, it fails to create high-fidelity results in complex scenes (e.g., "multiple objects"). Pro + +![](images/4f305db8666f607d222b0a1e12203c3543cd08209a0a63d6d3d70456663f5233.jpg) +(a) +Figure 3. Comparisons on qualitative results of our VP3D with other text-to-3D techniques on $\mathrm{T}^3$ Bench [11]. The prompts are (a) "A fuzzy pink flamingo lawn ornament", (b) "A blooming potted orchid with purple flowers", (c) "A blue butterfly on a pink flower", (d) "A lighthouse on a rocky shore", (e) "Hot popcorn jump out from the red striped popcorn maker", (f) "A chef is making pizza dough in the kitchen". (a-b), (c-d), (e-f) belongs to the Single Object, Single Object with Surr and Multi Objects category in $\mathrm{T}^3$ Bench, respectively. + +lificDreamer further boosts the performance by training an additional diffusion model during the optimization process to realize a principled particle-based variational score distillation loss. However, our VP3D still outperforms ProlificDreamer across all evaluation metrics and prompt sets, which confirms the effectiveness of our VP3D. + +# 4.3. Qualitative Results + +The qualitative comparisons for text-to-3D generation are presented in Figure 3. As can be seen, our VP3D generally produces superior 3D scenes with plausible geometry and realistic textures when compared with baseline methods. Specifically, DreamFusion suffers from a severely over-saturated problem and has difficulty generating complex geometry. Magic3D and Latent-NeRF slightly allevi + +ate these issues through their higher-resolution DMTet and pixel space refinement, respectively. While Fantasia3D and SJC can generate richer textures than DreamFusion, the geometric quality of the generated 3D scenes falls short of expectations. Notably, ProlificDreamer trains an additional diffusion model during the optimization process to perform variational score distillation (VSD) instead of SDS, achieving satisfactory single-object objects. However, the use of VSD at times introduces excessive irrelevant information or geometry noise in more complex scenarios. In contrast, we can clearly observe that the generated 3D scenes by VP3D faithfully match the input text prompt with plausible geometry and realistic appearance, which demonstrates the superiority of VP3D over state-of-the-art methods and its ability to generate high-quality 3D content. + +![](images/9d9ff906e36fe6ccc62a74c87c4a97996b88a72876a2ccd17f595d58d23847ae.jpg) +Figure 4. Stylized text-to-3D generation results of our VP3D. + +# 4.4. Ablation Study + +Here we investigate how each design in our VP3D influences the overall generation performance. We depict the qualitative results of each ablated run in Figure 5. $\mathcal{L}_{SDS}$ is our baseline model that employs vanilla score distillation sampling loss. As can be seen, the generated 3D scene is over-saturated and geometry unreasonable. Instead, when $\mathcal{L}_{VP - SDS}$ is employed, the generation quality is clearly enhanced in terms of both geometry and appearance. This highlights the critical effectiveness of our proposed visual-prompted score distillation sampling. Nevertheless, the resulting 3D scenes by $\mathcal{L}_{VP - SDS}$ are still not satisfying enough. By utilizing additional visual consistency and human feedback reward functions $\mathcal{L}_{vc - reward}$ (Eq. 7) and $\mathcal{L}_{hf - reward}$ (Eq. 6), the generation quality is gradually improved. The results basically validate the effectiveness of these two complementary factors. + +# 4.5. Extension to Stylized Text-to-3D Generation + +In this section, we demonstrate that another advantage of our VP3D is its remarkable versatility in 3D generation as it can be readily adapted for a new task of stylized text-to-3D generation. The main difference is that the visual prompt is no longer generated from the text prompt but from a user-specified reference image. We also empirically discard the loss in Eq. 6 to eliminate the strictly text-image alignment constraint. In this way, our VP3D can integrate the visual cues contained in the reference image into text-to-3D generation and produce a stylized 3D asset. This asset not only + +![](images/7574029076a2fe215d2e2a48c5741efcedb9e842e04e073f781d767c496f231a.jpg) +Figure 5. Comparisons on qualitative results by using different ablated runs of our VP3D. The text prompts are (a) "A broken tablespoon lies next to an empty sugar bowl" and (b) "A chameleon perched on a tree branch". + +semantically aligns with the text prompt but also reflects visual and geometry properties in the reference image. + +Figure 4 shows our stylized text-to-3D generation results. Our VP3D can generate diverse and stylized 3D assets by giving different visual prompts to the same text prompt. As shown in Figure 4 (a-b), the generated result semantically is a rabbit that adheres to the text prompt but also inherits some visual cues of the visual prompt. To be clear, the generated 3D rabbits have somewhat consistent geometry pose and appearance texture with the object in the visual prompt. For example, in Figure 4 (b), the generated rabbit mirrors the "hugging pose" of the reference image and also has the same style of "crescent-shaped eyebrows" and "yellow plaid jacket" as in the reference image. In Figure 4 (c-d), we showcase the versatility of our VP3D by seamlessly blending styles from different visual prompts. Take Figure 4 (d) as an instance, we use the leopard image as a visual prompt in the coarse stage and then replace it with an oil painting image in the fine stage. Our VP3D finally resulted in a 3D rabbit that not only has a consistent pose with the leopard but also a colorful oil painting style texture. The stylized 3D generation ability distinct our VP3D from previous text-to-3D approaches and can lead to more creative and diverse 3D content creation. + +# 5. Conclusion + +In this work, we propose VP3D, a new paradigm for text-to-3D generation by leveraging 2D visual prompts. We first capitalize on 2D diffusion models to generate a high-quality image from input text. This image then acts as a visual prompt to strengthen the 3D model learning with our devised visual-prompted score distillation sampling. Meanwhile, we introduce additional human feedback and visual consistency reward functions to encourage the semantic and appearance consistency between the 3D model and input visual & text prompt. Both qualitative and quantitative comparisons on the $\mathrm{T}^3$ Bench benchmark demonstrate the superiority of our VP3D over existing SOTA techniques. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5 +[3] Jingwen Chen, Yingwei Pan, Ting Yao, and Tao Mei. Controlstyle: Text-driven stylized image generation using diffusion priors. In ACM Multimedia, 2023. 2 +[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 2, 6 +[5] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Animating your life: Real-time video-to-animation translation. In ACM MM, 2019. 1 +[6] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM MM, 2019. 1 +[7] Yang Chen, Jingwen Chen, Yingwei Pan, Xinmei Tian, and Tao Mei. 3d creation at your fingertips: From text or image to 3d assets. In ACM MM, 2023. 2 +[8] Yang Chen, Yingwei Pan, Yehao Li, Ting Yao, and Tao Mei. Control3d: Towards controllable text-to-3d generation. In ACM MM, 2023. 2 +[9] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2 +[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1 +[11] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 3, 5, 6, 7 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2022. 1, 2, 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3 +[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 6 +[15] EJ Hu, Y Shen, P Wallis, Z Allen-Zhu, Y Li, S Wang, L Wang, and W Chen. Low-rank adaptation of large language models, arxiv, 2021. arXiv preprint arXiv:2106.09685, 10.4 +[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 6 +[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2, 4, 5, 6 + +[18] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. In NeurIPS, 2024. 2 +[19] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 5 +[20] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-conditional diffusion networks for image captioning. In CVPR, 2023. 1 +[21] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 2 +[22] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5, 6 +[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 5 +[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICLR, 2021. 1, 2 +[26] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 1 +[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2, 3, 4, 6 +[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2 +[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 1 +[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1 +[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2 +[36] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.5 +[37] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 2 +[38] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 2, 4, 6 +[39] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 1, 2, 3, 4, 6 +[40] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 2 +[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.5 +[42] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In ACM MM, 2023. 2 +[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4 \ No newline at end of file diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/images.zip b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..daa964ea7de3bbf7debeeae7ebb8942b6c1721eb --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f288858a9d620696c59993955b0821937e670bba6bf58c8e9032f0057b63f60e +size 545872 diff --git a/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/layout.json b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5616e0263e367463d8da940962cd403e97faa56 --- /dev/null +++ b/2024/VP3D_ Unleashing 2D Visual Prompt for Text-to-3D Generation/layout.json @@ -0,0 +1,7767 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 120 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 120 + ], + "type": "text", + "content": "VP3D: Unleashing 2D Visual Prompt for Text-to-3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 457, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 142, + 457, + 171 + ], + "spans": [ + { + "bbox": [ + 134, + 142, + 457, + 171 + ], + "type": "text", + "content": "Yang Chen†, Yingwei Pan†, Haibo Yang§, Ting Yao†, and Tao Mei† \n†HiDream.ai Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 171, + 409, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 171, + 409, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 171, + 409, + 186 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 184, + 171, + 409, + 186 + ], + "type": "text", + "content": "School of Computer Science, Fudan University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 187, + 507, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 187, + 507, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 187, + 507, + 200 + ], + "type": "text", + "content": "{clenyang, pandy}@hidream.ai, yanghaibo.fdu@gmail.com, {tiyao, tmei}@hidream.ai" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 576 + ], + "type": "text", + "content": "Recent innovations on text-to-3D generation have featured Score Distillation Sampling (SDS), which enables the zero-shot learning of implicit 3D models (NeRF) by directly distilling prior knowledge from 2D diffusion models. However, current SDS-based models still struggle with intricate text prompts and commonly result in distorted 3D models with unrealistic textures or cross-view inconsistency issues. In this work, we introduce a novel Visual Prompt-guided text-to-3D diffusion model (VP3D) that explicitly unleashes the visual appearance knowledge in 2D visual prompt to boost text-to-3D generation. Instead of solely supervising SDS with text prompt, VP3D first capitalizes on 2D diffusion model to generate a high-quality image from input text, which subsequently acts as visual prompt to strengthen SDS optimization with explicit visual appearance. Meanwhile, we couple the SDS optimization with additional differentiable reward function that encourages rendering images of 3D models to better visually align with 2D visual prompt and semantically match with text prompt. Through extensive experiments, we show that the 2D Visual Prompt in our VP3D significantly eases the learning of visual appearance of 3D models and thus leads to higher visual fidelity with more detailed textures. It is also appealing in view that when replacing the self-generating visual prompt with a given reference image, VP3D is able to trigger a new task of stylized text-to-3D generation. Our project page is available at https://vp3d-cvpr24.github.io." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Generative Artificial Intelligence (especially for vision content generation) has aroused great attention in computer vision field [5, 6, 20, 26], leading to impressive advancements in text-to-image [30-32] and text-to-video generation [10, 14, 34]. These accomplishments can be attributed to the availability of large-scale image-text and video-text pair data [1, 33] and the emergence of robust diffusion-based generative models [12, 13, 25, 35]. Recently, researchers" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 233, + 425, + 312 + ], + "blocks": [ + { + "bbox": [ + 326, + 223, + 524, + 232 + ], + "lines": [ + { + "bbox": [ + 326, + 223, + 524, + 232 + ], + "spans": [ + { + "bbox": [ + 326, + 223, + 524, + 232 + ], + "type": "text", + "content": "Text prompt: \"A florist is making a bouquet with fresh flowers\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 233, + 425, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 233, + 425, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 233, + 425, + 312 + ], + "type": "image", + "image_path": "7cdaf6826e003cb2f350a3b5c5f2ab8488a218d4a642f82154b530ff531c1345.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 233, + 542, + 312 + ], + "blocks": [ + { + "bbox": [ + 425, + 233, + 542, + 312 + ], + "lines": [ + { + "bbox": [ + 425, + 233, + 542, + 312 + ], + "spans": [ + { + "bbox": [ + 425, + 233, + 542, + 312 + ], + "type": "image", + "image_path": "d77737463c8c506849b33453d7653609f379d03241a87bce959eb967f923a379.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 316, + 514, + 325 + ], + "lines": [ + { + "bbox": [ + 453, + 316, + 514, + 325 + ], + "spans": [ + { + "bbox": [ + 453, + 316, + 514, + 325 + ], + "type": "text", + "content": "(b)ProlificDreamer" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 327, + 402, + 400 + ], + "blocks": [ + { + "bbox": [ + 347, + 315, + 386, + 324 + ], + "lines": [ + { + "bbox": [ + 347, + 315, + 386, + 324 + ], + "spans": [ + { + "bbox": [ + 347, + 315, + 386, + 324 + ], + "type": "text", + "content": "(a) Magic3D" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 327, + 402, + 400 + ], + "lines": [ + { + "bbox": [ + 329, + 327, + 402, + 400 + ], + "spans": [ + { + "bbox": [ + 329, + 327, + 402, + 400 + ], + "type": "image", + "image_path": "09c4e66c1563e8cd39c1379ae4c26bcdee93215b40340be42e9dcbd23e726bc4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 403, + 388, + 412 + ], + "lines": [ + { + "bbox": [ + 343, + 403, + 388, + 412 + ], + "spans": [ + { + "bbox": [ + 343, + 403, + 388, + 412 + ], + "type": "text", + "content": "Visual prompt" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 408, + 327, + 487, + 408 + ], + "blocks": [ + { + "bbox": [ + 408, + 327, + 487, + 408 + ], + "lines": [ + { + "bbox": [ + 408, + 327, + 487, + 408 + ], + "spans": [ + { + "bbox": [ + 408, + 327, + 487, + 408 + ], + "type": "image", + "image_path": "5ce9398ccec78e2746453dd7b53f0b4229098a936a39a7c64bfd9633c3e40178.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 413, + 462, + 422 + ], + "lines": [ + { + "bbox": [ + 413, + 413, + 462, + 422 + ], + "spans": [ + { + "bbox": [ + 413, + 413, + 462, + 422 + ], + "type": "text", + "content": "(c) VP3D (Ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 430, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 486 + ], + "type": "text", + "content": "Figure 1. Existing text-to-3D generation techniques (e.g., Magic3D [17] and ProlificDreamer [39]) often suffer from degenerated results (e.g., over-saturated appearances or inaccurate geometries). Our VP3D novelly integrates a visual prompt to strength score distillation sampling, leading to better 3D results." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 488, + 327, + 528, + 408 + ], + "blocks": [ + { + "bbox": [ + 488, + 327, + 528, + 408 + ], + "lines": [ + { + "bbox": [ + 488, + 327, + 528, + 408 + ], + "spans": [ + { + "bbox": [ + 488, + 327, + 528, + 408 + ], + "type": "image", + "image_path": "b622e2443acf43447a926aef5fb1eb2ebeeba709d72ac81186488496c75c6672.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": "have gone beyond text-driven image/video generation, and begun exploring diffusion models for text-driven content creation of 3D assets (e.g., text-to-3D generation). This direction paves a new way for practical 3D content creation and has a great potential impact for numerous applications like virtual reality, gaming and Metaverse. Compared to image generation, text-to-3D generation, however, is more challenging, due to the complexities associated with intricate 3D geometry and appearance (i.e., object shapes and textures). Moreover, the collection and annotation of 3D data are somewhat resourcefully expensive and thus cannot be easily scaled up to billion level as image data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 642, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 715 + ], + "type": "text", + "content": "To tackle this issue, a pioneering text-to-3D work (DreamFusion [27]) presents the first attempt of exploiting an off-the-shelf text-to-image diffusion model to generate promising 3D assets in a zero-shot fashion. The key design behind such success is Score Distillation Sampling (SDS), which directly optimizes the implicit 3D model of" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4896" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) with prior knowledge distilled from 2D diffusion model. Nevertheless, such distilled prior knowledge is merely driven by the input text prompt, and it is non-trivial to learn high-quality NeRF with distilled SDS supervision. Although several subsequent works [4, 17, 22, 38, 39] further upgrade SDS, this kind of SDS-based solution still results in degenerated 3D models with unrealistic/less-detailed textures, especially when feeding intricate text prompts (as seen in Figure 1 (a-b))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 539 + ], + "type": "text", + "content": "In this work, we propose to mitigate this limitation through a unique design of visual prompt-guided text-to-3D diffusion model, namely VP3D. Intuitively, \"a picture is worth a thousand words.\" That is, a single image can convey human intentions of visual content creation (e.g., the visual appearance or semantic structure) more effectively than textual sentences. This motivates us to introduce additional guidance of visual prompt, and thus decompose the typical single-shot text-to-3D process into two cascaded processes: first text-to-image generation, and then (text plus image)-to-3D generation. In particular, VP3D first leverages off-the-shelf text-to-image diffusion models to produce a high-fidelity image that reflects extremely realistic appearance with rich details. In the latter process, this synthetic image further acts as 2D visual prompt to supervise SDS optimization of NeRF, coupled with the input text prompt. At the same time, a differentiable reward function is additionally utilized to encourage the rendering images of NeRF to be better aligned with 2D visual prompt (visual appearance consistency) and text prompt (semantic consistency). As illustrated in Figure 1 (c), we show that the novel visual prompt-guided diffusion process in VP3D significantly enhances the visual fidelity of 3D assets with realistic and rich texture details. Meanwhile, when easing the learning of visual appearance of 3D assets via visual prompt guidance, the optimization of NeRF will focus more on the modeling of geometry, leading to better 3D sharps with cross-view consistency. We believe that the ability of unleashing high-quality visual knowledge in 2D visual prompt is potentially a new paradigm of text-to-3D generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 539, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 647 + ], + "type": "text", + "content": "As a by-product, we also demonstrate that our VP3D can be readily adapted for a new task of stylized text-to-3D generation. Intriguingly, we simply replace the self-generating image in VP3D with a user-given reference image, and treat it as a new visual prompt to trigger (text plus image)-to-3D generation. In this way, our VP3D is able to produce a stylized 3D asset, which not only semantically aligns with text prompt but also shares similar geometric & visual style as the reference image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 134, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 134, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 134, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Text-to-3D generation. Significant advancements have been witnessed in text-to-image generation with 2D diffusion models in recent years [3, 12, 13, 25, 30-32, 35]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 337 + ], + "type": "text", + "content": "However, extending these capabilities to 3D content generation poses a substantial challenge, primarily due to the absence of large-scale paired text-3D datasets. To mitigate the reliance on extensive training data, recent works try to accomplish zero-shot text-to-3D generation [4, 7, 8, 17, 22, 27, 38, 39, 42]. Specifically, the pioneering work DreamFusion [27] showcased remarkable achievements in text-to-3D generation through pre-trained text-to-image diffusion models. SJC [38] concurrently addressed the out-of-distribution problem in lifting 2D diffusion models to perform text-to-3D generation. Following these, several subsequent works have strived to enhance text-to-3D generation further. For instance, Latent-NeRF [22] proposed to incorporate a sketch shape to guide the 3D generation directly in the latent space of a latent diffusion model. Magic3D [17] presented a coarse-to-fine strategy that leverages both low- and high-resolution diffusion priors to learn the underlying 3D representation. Control3D [8] proposed to enhance user controllability in text-to-3D generation by incorporating additional hand-drawn sketch conditions. ProlificDreamer [39] presented a principled particle-based variational framework to improve the generation quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "content": "Unlike previous works, we formulate the text-to-3D generation process from a new perspective. We first leverage the off-the-shelf text-to-image diffusion models to generate a high-quality image that faithfully matches the input text prompt. This synthetic reference image then serves as a complementary input alongside the text, synergistically guiding the 3D learning process. Moreover, we showcase the remarkable versatility of this novel architecture by effortlessly extending its capabilities to the realm of stylized text-to-3D generation. The resulting 3D asset not only exhibits semantic alignment with the provided text prompt but also masterfully captures the visual style of the reference image. This capability marks another pivotal distinction between our VP3D and previous text-to-3D approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Image-to-3D generation. Recently, prior works RealFusion [21], NeuralLift-360 [40] and NeRDi [9] leverage 2D diffusion models to achieve image-to-3D generation. The following work Make-IT-3D [37] proposed a two-stage optimization framework to improve the generation quality further. Zero-1-to-3 [19] finetuned the Stable Diffusion model to enable generating novel views of the input image. It can then be used as a 3D prior model to achieve high-quality image-to-3D generation. Inspired by this, Magic123 [28] proposed to use 2D and 3D priors simultaneously to generate faithful 3D content from the given image. One-2-3-45 [18] integrated Zero-1-to-3 and a multi-view reconstruction model to accelerate the 3D generation process. It is worth noting that our work is not targeting image-to-3D generation. We utilize a reference image to guide the text-to-3D learning process, instead of directly turning the reference image into 3D content." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4897" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 93, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 93, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 93, + 83 + ], + "type": "text", + "content": "3. VP3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "content": "In this section, we elaborate the architecture of our VP3D, which introduces a novel visual prompt-guided text-to-3D diffusion model. An overview of our VP3D architecture is depicted in Figure 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 146, + 127, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 127, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 127, + 158 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": "Text-to-Image Diffusion Models. Diffusion models are a family of generative models that are trained to gradually transform Gaussian noise into samples from a target distribution [13]. Given a target data distribution " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x})" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ", a forward diffusion process is defined to progressively add a small amount of Gaussian noise to the data " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x})" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ". This process follows a Markov chain " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T}) = \\prod_{t=1}^{T} q(\\mathbf{x}_t | \\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " and produces a sequence of latent variables " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\ldots, \\mathbf{x}_T" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " time steps. The marginal distribution of latent variables at time step " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t | \\mathbf{x}) = \\mathcal{N}(\\mathbf{x}_t; \\alpha_t \\mathbf{x}, \\sigma_t^2 \\mathbf{I})" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ". Thus the noisy sample " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " can be directly generated through the equation " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\alpha_t \\mathbf{x} + \\sigma_t^2 \\epsilon" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " are chosen parameters such that " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\alpha_t^2 + \\sigma_t^2 = 1" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " noise adding steps, " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " is equivalent to an isotropic Gaussian distribution. Then, a reverse generative process is defined to gradually \"denoise\" " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "X_T" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " to reconstruct the original sample. This can be described by a Markov process " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "p_\\phi(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t)" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ", with the conditional probability " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "p_\\phi(\\mathbf{x}_{t-1} | \\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t-1}; \\boldsymbol{\\mu}_\\phi(\\mathbf{x}_t, t), \\boldsymbol{\\Sigma}_\\phi(\\mathbf{x}_t, t))" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ". Commonly, a UNet neural network " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon_\\phi(\\mathbf{x}_t; t)" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " is used to predict the noise that was used to produce " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": ". Text-to-image diffusion models build upon the above theory to condition the diffusion process with a given text prompt " + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 45, + 164, + 289, + 478 + ], + "type": "text", + "content": " using classifier-free guidance (CFG) [12]. The corresponding noise predictor is remodeled by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 289, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 289, + 521 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; \\mathbf {z} _ {y}, t) = \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset) + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\emptyset)\\right), \\tag {1}", + "image_path": "d89e050a8dccfa517c57757e13e2df9dab074f59c1874541bb5847161cbe5b79.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " is a scale that denotes the classifier-free guidance weight, " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_y" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " is the corresponding text embedding of the text prompt " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " indicates the noise prediction without conditioning. The diffusion model " + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": " is typically optimized by a simplified variant of the variational lower bound of the log data likelihood, which is a Mean Squared Error criterion:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 601, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 601, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 601, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} (\\phi) = \\mathbb {E} _ {\\mathbf {x}, t, \\epsilon} \\left[ w (t) \\| \\hat {\\boldsymbol {\\epsilon}} _ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "b03930100ba1f2f1d0784dcbc81664c57eb0f19661cba4285e47c3b64755de08.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "text", + "content": " is a weighting function that depends on the timestep " + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 629, + 287, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Score Distillation Sampling. A recent pioneering work called DreamFusion [27] introduces Score Distillation Sampling (SDS) that enables leveraging the priors of pre-trained text-to-image diffusion models to facilitate text-to-3D generation. Specifically, let " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " be the learnable parameters of a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "3D model (e.g., NeRF) and " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " be a differentiable rendering function that can render an image " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta; \\mathbf{c})" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " from the 3D model " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " at a camera viewpoint " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": ". SDS introduces a loss function " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " to optimize the parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": ". Its gradient is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 332, + 135, + 545, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 135, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 332, + 135, + 545, + 159 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {3}", + "image_path": "6991c7a17e99d0754c8e3f23e81390421cb789206373ea3cb6a3151293ee359e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " is obtained by perturbing the rendered image " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " with a Gaussian noise " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": "-th timestep of the forward diffusion process, " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_y" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " is the conditioned text embedding of given text prompt " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": ". Intuitively, the SDS loss estimates an update direction in which the noised version of rendered image " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " should be moved towards a denser region in the distribution of real images (aligned with the conditional text prompt " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": "). By randomly sampling views and backpropagating the gradient in Eq. 3 to the parameters " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": " through the differential parametric function " + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 163, + 547, + 294 + ], + "type": "text", + "content": ", this approach eventually results in a 3D model that resembles the text." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 300, + 538, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 538, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 538, + 313 + ], + "type": "text", + "content": "3.2. Visual-prompted Score Distillation Sampling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 534 + ], + "type": "text", + "content": "Visual Prompt Generation. As aforementioned, score distillation sampling plays a key role in text-to-3D generation. Nevertheless, empirical observations [11, 27, 39] reveal that SDS still results in degenerated 3D models especially when feeding intricate text prompts. First, SDS-generated results often suffer from over-saturation issues. These issues are, in part, attributed to the necessity of employing a large CFG value (i.e., 100) within the SDS framework [27, 39]. A Large CFG value narrows down the score distillation space to more text-relevant areas. This can mitigate the divergence of diffusion priors in the optimization process, thereby fostering enhanced stability in 3D representation learning. However, this is at the cost of less realistic and diversity generation results, as large CFG values are known to yield over-saturated results [39]. Second, results generated by SDS still face the risk of text-3D misalignment, such as missing key elements in the scene, especially when text prompts contain multiple objects with specific attributes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": "A fundamental reason behind the aforementioned issues may lie in the substantial distribution gap between text and 3D modalities. Thus it is non-trivial to directly learn a meaningful 3D scene solely based on a single text prompt. This insight motivates us to introduce an additional visual prompt as a bridge to explicitly establish a connection between the text input and the desired 3D output. Particularly, we leverage off-the-shelf text-to-image diffusion models (e.g., Stable Diffusion) to produce a high-fidelity image that faithfully matches the input text prompt and has an extremely realistic appearance. This image is then used as a visual prompt in conjunction with the input text prompt to jointly supervise the 3D generation process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Score Distillation Sampling with Visual Prompt. We now present visual-prompted score distillation sampling that dis" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4898" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 519, + 257 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 519, + 257 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 519, + 257 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 519, + 257 + ], + "type": "image", + "image_path": "8e0a0fc06ceed63935551bc5d46de1b59510a36ec250a8b1719bf89054701b48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 259, + 478, + 270 + ], + "lines": [ + { + "bbox": [ + 113, + 259, + 478, + 270 + ], + "spans": [ + { + "bbox": [ + 113, + 259, + 478, + 270 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed VP3D framework for visual prompted text-to-3D generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": "tillation knowledge from a pre-trained diffusion model to optimize a 3D model by considering inputs not only from a text prompt " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " but also from a visual prompt " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": ". To be clear, we restructure the standard SDS-based text-to-3D pipeline by utilizing an image-conditioned diffusion model [43] to trigger visual prompt-guided text-to-3D generation. Technically, the visual prompt is first converted to a global image embedding " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_v" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " by the CLIP image encoder [29] and a following projection network. This image embedding represents the rich content and style of the visual prompt and has the same dimension as the text embedding " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_y" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " used in the pre-trained text-to-image diffusion model (Stable Diffusion). Following SDS, we first add noise " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " to the rendered image of the underlying 3D model according to the random sampled time step " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " to get a noised image " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " is input to the diffusion model along with the conditional visual prompt embedding " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_v" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " and text prompt embedding " + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_y" + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": " to estimate the added noise as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 497, + 286, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 286, + 531 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {e}} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}\\right) = \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right) \\\\ + s * \\left(\\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\lambda * \\mathbf {z} _ {v}\\right)\\right) - \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\mathbf {x} _ {t}; t, \\emptyset , \\emptyset\\right)), \\tag {4} \\\\ \\end{array}", + "image_path": "a79bb64fc04360ea5f2f7e8d4dc0fc30b22e47b8fb76135d2666515f5a254ff8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " is the classifier-free guidance weight, " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " is the visual prompt condition weight, " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " is the parameter of the pre-trained noise predictor " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\mathbf{x}_t; t, \\emptyset, \\emptyset)" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " denotes the noise prediction without conditioning. In this way, our proposed method explicitly incorporates the visual prompt and text prompt in a unified fashion for text-to-3D generation. Consequently, the final gradient of our introduced visual-prompted score distillation sampling (VP-SDS) loss " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 647, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 647, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 52, + 647, + 287, + 671 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {V P - S D S} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\tilde {\\epsilon} _ {\\phi} (\\mathbf {x} _ {t}; t, \\mathbf {z} _ {y}, \\mathbf {z} _ {v}) - \\epsilon) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], (5)", + "image_path": "4a26cfae2a2c07ea034bf49a363fdc525619ee651e8ae2a207219caaaecc041a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 205, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 205, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 205, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 205, + 689 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 47, + 677, + 205, + 689 + ], + "type": "text", + "content": " is a scheduling coefficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Comparison with SDS. Comparing the update gradient of SDS (Eq. 3) and VP-SDS (Eq. 5), SDS is a special case" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "content": "of our VP-SDS by setting " + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "content": " where the visual prompt condition is neglected. In accordance with the theoretical analysis presented in [27, 39], the mode-seeking nature of SDS necessitates a large CFG to ensure that the pre-trained diffusion model " + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "content": " delivers a \"sharp\" updating direction for the underlying 3D model. Nevertheless, a large CFG, in turn, results in poor-quality samples and thus a \"degraded\" update direction. In contrast, VP-SDS leverages the additional visual prompt to narrow down the distillation space of " + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 277, + 545, + 456 + ], + "type": "text", + "content": " into a more compact region that aligns tightly with the visual prompt. Meanwhile, the distillation space is also refined by the visual prompt as it reflects realistic appearances with rich details. Therefore, the updating direction derived from our VP-SDS is not only \"sharp\" but also \"fine\", which can obtain much better 3D generation results than SDS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "content": "Notably, a recent work ProlificDreamer [39] presents variational score distillation (VSD) to address the aforementioned issues in SDS. However, VSD needs to train an additional diffusion model using LoRA [15] during the optimization process, which incurs a considerable computational overhead compared to SDS. Instead, the additional computational cost of our VP-SDS is nearly negligible, making it computationally more efficient than VSD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "View-dependent Visual Prompting. Apart from the oversaturation problem discussed above, existing text-to-3D methods are known to also suffer from the multi-view inconsistency problem (e.g., the multi-face Janus problem). This arises from the fact that the underlying prior diffusion model is exclusively trained on individual 2D images and therefore lacks 3D awareness. To alleviate this issue, existing text-to-3D methods [17, 27, 38, 39] always employ diffusion loss with view-dependent text conditioning, which is to append \"front view\", \"side view\", or \"back view\" to the input text based on the location of the randomly sampled camera. Inspired by this, we devise a view-dependent visual prompting strategy to further mitigate the view incon" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4899" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "sistency problem in collaboration with our introduced VP-SDS. Technically, given the input visual prompt (assuming it is shot from the front view), we use a view-conditioned 2D diffusion model, Zero-1-to-3 [19], to transform it into left-side, right-side and backward views. Then we fed different visual prompts into VP-SDS (Eq. 5) depending on the corresponding sampled camera viewpoint. For instance, when the azimuth angle " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\gamma_{cam} \\in [0^{\\circ}, 360^{\\circ}]" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " of the camera position falls in the range near " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " denotes the front view), we feed the generated back view counterpart of the input visual prompt into Eq 5. In this way, the inherent 3D geometry information contained in the multi-view visual prompts is encoded into the 3D representation learning through view-dependent VP-SDS, leading to better view consistency in the 3D generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 258, + 224, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 224, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 224, + 271 + ], + "type": "text", + "content": "3.3. Learning with Reward Feedback" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": "To further encourage rendered images of the underlying 3D model that are high fidelity and well aligned with the input visual prompt and text prompt, we devise two types of differentiable reward functions that complement the aforementioned VP-SDS objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 504 + ], + "type": "text", + "content": "Human Feedback Reward. Recent practice has shown the capability of improving text-to-image models with human feedback [41]. Particularly, it first trains a reward model on a large dataset comprised of human assessments of text-image pairs. Such a reward model thus has the ability to measure the quality of the generated samples in terms of both image fidelity and image-text alignment. Consequently, it can then be used to fine-tune diffusion models to maximize the predicted scores of the reward model through differentiable reward functions, leading to better generation results. Motivated by this, we go one step further to utilize the open-sourced reward model " + }, + { + "bbox": [ + 46, + 337, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 337, + 289, + 504 + ], + "type": "text", + "content": " in ImageReward [41] for text-to-3D generation. Specifically, we introduce a human feedback reward loss as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 513, + 287, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 513, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 102, + 513, + 287, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h f - r e w a r d} = \\mathbb {E} _ {\\mathbf {c}} [ \\psi (\\mathbf {r} (\\mathbf {x}, y)) ], \\tag {6}", + "image_path": "708b8b44ab71f0e26eebef8786ccb10a4141105c2d30c749acb0c2bbe9f15af6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta; \\mathbf{c})" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " is a rendered image by the underlying 3D model " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " from an arbitrary viewpoint " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " is the conditional text prompt and " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " is a differentiable reward-to-loss map function as in [41]. Intuitively, minimizing the loss in Eq. 6 encourages the rendered image " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " to get a higher reward score from the reward model " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": ", which means the underlying 3D model should update towards the refined direction where the renderings have high appearance fidelity and faithfully match the input text prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Visual Consistency Reward. Given that the above human feedback reward only takes into account the input text prompt, we further devised a visual consistency reward to fully leverage the visual prompt as well, since text prompts cannot capture all appearance details. Technically, we adopt a pre-trained self-supervised vision transformer DINO-ViT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "[2] to extract the visual features " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "F_{dino}(v)" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "F_{dino}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " of the input visual prompt " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " and rendered image " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": ", respectively. Then we penalize the feature-wise difference between them at the visual prompt viewpoint:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 129, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 129, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 341, + 129, + 545, + 143 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v c - r e w a r d} = \\left\\| F _ {d i n o} (\\mathbf {x}) - F _ {d i n o} (\\mathbf {v}) \\right\\| ^ {2}. \\tag {7}", + "image_path": "8f14a3adafaaaacd406b15c4f678739b2f7e5814e8b64611a23a954ead71e599.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 152, + 546, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 546, + 189 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 546, + 189 + ], + "type": "text", + "content": "By imposing such visual consistency loss, we encourage the underlying 3D model to adhere to the plausible shape and appearance properties conveyed by the visual prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 196, + 480, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 480, + 209 + ], + "type": "text", + "content": "3.4. 3D Representation and Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 214, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 334 + ], + "type": "text", + "content": "Inspired by [17], we adopt a two-stage coarse-to-fine framework for text-to-3D generation with two different 3D scene representations. At the coarse stage, we leverage InstantNGP [24] as 3D representation, which is much faster to optimize compared to the vanilla NeRF [23] and can recover complex geometry. In the fine stage, we leverage DMTet as the 3D representation to further optimize a high-fidelity mesh and texture. Specifically, the 3D shape and texture represented in DMTet are first initialized from the density field and color field of the coarse stage, respectively [17]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 335, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 381 + ], + "type": "text", + "content": "During the optimization process in each stage, we first render images from the underlying 3D model through differentiable rasterizers at arbitrary camera poses and optimize the 3D model with a combination of losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 392, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 392, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 311, + 392, + 545, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n e}} = \\mathcal {L} _ {V P - S D S} + \\lambda_ {1} \\mathcal {L} _ {v c - r e w a r d} + \\lambda_ {2} \\mathcal {L} _ {h f - r e w a r d}, \\tag {8}", + "image_path": "f30976cf04f6d0ec2d05fecdbd485f4c1730556937daacabc415ce91ab00010a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 417, + 476, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 476, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 476, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 417, + 476, + 429 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 306, + 417, + 476, + 429 + ], + "type": "text", + "content": " are the trade-off parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 439, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 388, + 453 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of our VP3D for text-to-3D generation via extensive empirical evaluations. We first show both quantitative and qualitative results of VP3D in comparison to existing techniques on the newly released text-to-3D benchmark (T" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": "Bench [11]). Next, we conduct ablation studies to validate each design in VP3D. Finally, we demonstrate the extended capability of VP3D for stylized text-to-3D generation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 563, + 432, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 432, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 432, + 576 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. In the coarse and fine stage, the underlying 3D models are both optimized for 5000 iterations using Adam optimizer with 0.001 learning rate. The rendering resolutions are set to " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " for coarse and fine stage, respectively. We implement the underlying Instant-NGP and DMTet 3D representation mainly based on the Stable-DreamFusion codebase [36]. " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is set to 0.1 in the coarse stage and 0.01 in the fine stage. " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is linearly increased from 0.001 to 0.01 during the optimization process. The visual prompt condition weight is set to 0.5 in all experiments." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4900" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 84, + 547, + 206 + ], + "blocks": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "type": "text", + "content": "Table 1. The quantitative results of our method and baselines on " + }, + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "type": "inline_equation", + "content": "{\\mathrm{T}}^{3}" + }, + { + "bbox": [ + 153, + 70, + 440, + 82 + ], + "type": "text", + "content": " Bench [11]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 84, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 547, + 206 + ], + "type": "table", + "html": "
MethodSingle ObjectSingle Object with SurroundingsMultiple Objects
QualityAlignmentAverageQualityAlignmentAverageQualityAlignmentAverage
DreamFusion [27]24.924.024.419.329.824.617.314.816.1
SJC [38]26.323.024.717.322.319.817.75.811.7
LatentNeRF [22]34.232.033.123.737.530.621.719.520.6
Fantasia3D [4]29.223.526.421.932.027.022.714.318.5
ProlificDreamer [39]51.147.849.442.547.044.845.725.835.8
Magic3D [17]38.735.337.029.841.035.426.624.825.7
VP3D (Ours)54.852.253.545.450.848.149.131.540.3
", + "image_path": "52cf09fe66da209a293abadce9c41a7cea00c27f0264fd0dac6c217f0aa131cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "text", + "content": "Evaluation Protocol. Existing text-to-3D generation works commonly examine their methods over the CLIP R-Precision score [16], which is an automated metric for the consistency of rendered images with respect to the input text. However, this text-image alignment-based metric cannot faithfully represent the overall 3D quality. For example, CLIP-based text-to-3D methods can also achieve high CLIP R-Precision scores even if the resulting 3D scenes are unrealistic and severely distorted [27]. Taking this into account, we instead conduct experiments on a newly open-sourced benchmark: " + }, + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "text", + "content": " Bench [11], which is the first comprehensive text-to-3D benchmark containing 300 diverse text prompts of three categories (single object, single object with surroundings, and multiple objects)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^{3}" + }, + { + "bbox": [ + 46, + 388, + 287, + 508 + ], + "type": "text", + "content": " Bench provides two automatic metrics (quality and alignment) based on the rendered multi-view images to assess the subjective quality and text alignment. The quality metric utilizes a combination of multi-view text-image scores and regional convolution to effectively identify quality and view inconsistency. The alignment metric employs a 3D captioning model and a Large Language Model (i.e., GPT-4) to access text-3D consistency. Following this, we also leverage the quality and alignment metric to quantitatively compare our VP3D against baseline methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Baselines. To evaluate our method, we compare our VP3D with six state-of-the-art text-to-3D generation methods: DreamFusion [27], SJC [38], LatentNeRF [22], Fantasy3D [4], Magic3D [17] and ProlificDreamer [39]. Specifically, DreamFusion [27] firstly introduces score distillation sampling (SDS) that enables leveraging 2D diffusion model (Imagen [14]) to optimize a NeRF [23]. SJC [38] concurrently addresses the out-of-distribution problem in SDS and utilizes an open-sourced diffusion model (Stable Diffusion) to optimize a voxel NeRF. Latent-NeRF [22] first brings NeRF to the latent space to harmonize with latent diffusion models, then refines it in pixel space. Magic3D [17] extends DreamFusion with a coarse-to-fine framework that first optimizes a low-resolution NeRF model and then a high-resolution DMTet model via SDS. Fantasy3D [4] disentangles the SDS-based 3D learning into geometry and appearance learning. ProlificDreamer [39] upgrades Dream" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 218, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 255 + ], + "type": "text", + "content": "Fusion by a variational score distillation (VSD) loss that treats the underlying 3D scene as a random variable instead of a single point as in SDS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 263, + 425, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 425, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 425, + 276 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": "The quantitative performance comparisons of different methods for text-to-3D generation are summarized in Table 1. Overall, our VP3D consistently achieves better performances against existing techniques across all evaluation metrics and prompt categories. Remarkably, VP3D achieves an absolute quality-alignment average score improvement of " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": " against the best competitor ProlificDreamer across the three text prompt categories, respectively, which validates the effectiveness of our overall proposals. More importantly, while VP3D employs the same NeRF & DMTet 3D representation and coarse-t fine training scheme as the baseline method Magic3D, it significantly outperforms Magic3D by achieving " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "53.5\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "48.1\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "40.3\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": " average scores, representing a substantial improvement over Magic3D's average scores of " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "37.0\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "35.4\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "inline_equation", + "content": "25.7\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 498 + ], + "type": "text", + "content": ". The results generally highlight the key advantage of introducing visual prompts in lifting 2D diffusion models to perform text-to-3D generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "Specifically, DreamFusion and SJC enable the zero-shot learning of implicit 3D models by distilling prior knowledge from 2D diffusion models. However, the generated 3D scenes have relatively low quality and alignment scores, especially in complex scenarios where the text prompt contains multiple objects or surroundings. Latent-NeRF employs score distillation in the latent space and then back to pixel space to further refine the 3D model, leading to better results. The aforementioned three methods only utilize implicit 3D representations (NeRFs). In contrast, Magic3D adopts textured mesh DMTet as 3D representation for enabling high-resolution optimization and exhibits better performances across all three prompt categories. Fantasy3D also capitalizes on DMTet for geometry learning and then leverages BRDF for appearance learning in a disentangled manner. While Fantasy3D achieves better average scores than DreamFusion and SJC, it fails to create high-fidelity results in complex scenes (e.g., \"multiple objects\"). Pro" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4901" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 519, + 468 + ], + "blocks": [ + { + "bbox": [ + 76, + 97, + 89, + 109 + ], + "lines": [ + { + "bbox": [ + 76, + 97, + 89, + 109 + ], + "spans": [ + { + "bbox": [ + 76, + 97, + 89, + 109 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 72, + 519, + 468 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 519, + 468 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 519, + 468 + ], + "type": "image", + "image_path": "4f305db8666f607d222b0a1e12203c3543cd08209a0a63d6d3d70456663f5233.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "type": "text", + "content": "Figure 3. Comparisons on qualitative results of our VP3D with other text-to-3D techniques on " + }, + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "type": "text", + "content": " Bench [11]. The prompts are (a) \"A fuzzy pink flamingo lawn ornament\", (b) \"A blooming potted orchid with purple flowers\", (c) \"A blue butterfly on a pink flower\", (d) \"A lighthouse on a rocky shore\", (e) \"Hot popcorn jump out from the red striped popcorn maker\", (f) \"A chef is making pizza dough in the kitchen\". (a-b), (c-d), (e-f) belongs to the Single Object, Single Object with Surr and Multi Objects category in " + }, + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 46, + 475, + 546, + 521 + ], + "type": "text", + "content": " Bench, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "content": "lificDreamer further boosts the performance by training an additional diffusion model during the optimization process to realize a principled particle-based variational score distillation loss. However, our VP3D still outperforms ProlificDreamer across all evaluation metrics and prompt sets, which confirms the effectiveness of our VP3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 160, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 160, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 160, + 624 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "The qualitative comparisons for text-to-3D generation are presented in Figure 3. As can be seen, our VP3D generally produces superior 3D scenes with plausible geometry and realistic textures when compared with baseline methods. Specifically, DreamFusion suffers from a severely over-saturated problem and has difficulty generating complex geometry. Magic3D and Latent-NeRF slightly allevi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "content": "ate these issues through their higher-resolution DMTet and pixel space refinement, respectively. While Fantasia3D and SJC can generate richer textures than DreamFusion, the geometric quality of the generated 3D scenes falls short of expectations. Notably, ProlificDreamer trains an additional diffusion model during the optimization process to perform variational score distillation (VSD) instead of SDS, achieving satisfactory single-object objects. However, the use of VSD at times introduces excessive irrelevant information or geometry noise in more complex scenarios. In contrast, we can clearly observe that the generated 3D scenes by VP3D faithfully match the input text prompt with plausible geometry and realistic appearance, which demonstrates the superiority of VP3D over state-of-the-art methods and its ability to generate high-quality 3D content." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4902" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 287, + 324 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 287, + 324 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 287, + 324 + ], + "type": "image", + "image_path": "9d9ff906e36fe6ccc62a74c87c4a97996b88a72876a2ccd17f595d58d23847ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 279, + 345 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 279, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 279, + 345 + ], + "type": "text", + "content": "Figure 4. Stylized text-to-3D generation results of our VP3D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 358, + 141, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 141, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 141, + 371 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": "Here we investigate how each design in our VP3D influences the overall generation performance. We depict the qualitative results of each ablated run in Figure 5. " + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": " is our baseline model that employs vanilla score distillation sampling loss. As can be seen, the generated 3D scene is over-saturated and geometry unreasonable. Instead, when " + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VP - SDS}" + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": " is employed, the generation quality is clearly enhanced in terms of both geometry and appearance. This highlights the critical effectiveness of our proposed visual-prompted score distillation sampling. Nevertheless, the resulting 3D scenes by " + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VP - SDS}" + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": " are still not satisfying enough. By utilizing additional visual consistency and human feedback reward functions " + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vc - reward}" + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": " (Eq. 7) and " + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hf - reward}" + }, + { + "bbox": [ + 46, + 376, + 289, + 568 + ], + "type": "text", + "content": " (Eq. 6), the generation quality is gradually improved. The results basically validate the effectiveness of these two complementary factors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 575, + 277, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 277, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 277, + 588 + ], + "type": "text", + "content": "4.5. Extension to Stylized Text-to-3D Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In this section, we demonstrate that another advantage of our VP3D is its remarkable versatility in 3D generation as it can be readily adapted for a new task of stylized text-to-3D generation. The main difference is that the visual prompt is no longer generated from the text prompt but from a user-specified reference image. We also empirically discard the loss in Eq. 6 to eliminate the strictly text-image alignment constraint. In this way, our VP3D can integrate the visual cues contained in the reference image into text-to-3D generation and produce a stylized 3D asset. This asset not only" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 541, + 175 + ], + "type": "image", + "image_path": "7574029076a2fe215d2e2a48c5741efcedb9e842e04e073f781d767c496f231a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 185, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 229 + ], + "type": "text", + "content": "Figure 5. Comparisons on qualitative results by using different ablated runs of our VP3D. The text prompts are (a) \"A broken tablespoon lies next to an empty sugar bowl\" and (b) \"A chameleon perched on a tree branch\"." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "type": "text", + "content": "semantically aligns with the text prompt but also reflects visual and geometry properties in the reference image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 266, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 546, + 529 + ], + "type": "text", + "content": "Figure 4 shows our stylized text-to-3D generation results. Our VP3D can generate diverse and stylized 3D assets by giving different visual prompts to the same text prompt. As shown in Figure 4 (a-b), the generated result semantically is a rabbit that adheres to the text prompt but also inherits some visual cues of the visual prompt. To be clear, the generated 3D rabbits have somewhat consistent geometry pose and appearance texture with the object in the visual prompt. For example, in Figure 4 (b), the generated rabbit mirrors the \"hugging pose\" of the reference image and also has the same style of \"crescent-shaped eyebrows\" and \"yellow plaid jacket\" as in the reference image. In Figure 4 (c-d), we showcase the versatility of our VP3D by seamlessly blending styles from different visual prompts. Take Figure 4 (d) as an instance, we use the leopard image as a visual prompt in the coarse stage and then replace it with an oil painting image in the fine stage. Our VP3D finally resulted in a 3D rabbit that not only has a consistent pose with the leopard but also a colorful oil painting style texture. The stylized 3D generation ability distinct our VP3D from previous text-to-3D approaches and can lead to more creative and diverse 3D content creation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 548, + 379, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 379, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 379, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose VP3D, a new paradigm for text-to-3D generation by leveraging 2D visual prompts. We first capitalize on 2D diffusion models to generate a high-quality image from input text. This image then acts as a visual prompt to strengthen the 3D model learning with our devised visual-prompted score distillation sampling. Meanwhile, we introduce additional human feedback and visual consistency reward functions to encourage the semantic and appearance consistency between the 3D model and input visual & text prompt. Both qualitative and quantitative comparisons on the " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " Bench benchmark demonstrate the superiority of our VP3D over existing SOTA techniques." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4903" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Jingwen Chen, Yingwei Pan, Ting Yao, and Tao Mei. Controlstyle: Text-driven stylized image generation using diffusion priors. In ACM Multimedia, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 236 + ], + "type": "text", + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "text", + "content": "[5] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Animating your life: Real-time video-to-animation translation. In ACM MM, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "type": "text", + "content": "[6] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM MM, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[7] Yang Chen, Jingwen Chen, Yingwei Pan, Xinmei Tian, and Tao Mei. 3d creation at your fingertips: From text or image to 3d assets. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "text", + "content": "[8] Yang Chen, Yingwei Pan, Yehao Li, Ting Yao, and Tao Mei. Control3d: Towards controllable text-to-3d generation. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[9] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "type": "text", + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 502 + ], + "type": "text", + "content": "[11] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 525 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2022. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[14] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[15] EJ Hu, Y Shen, P Wallis, Z Allen-Zhu, Y Li, S Wang, L Wang, and W Chen. Low-rank adaptation of large language models, arxiv, 2021. arXiv preprint arXiv:2106.09685, 10.4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[19] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[20] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Jianlin Feng, Hongyang Chao, and Tao Mei. Semantic-conditional diffusion networks for image captioning. In CVPR, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[21] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[22] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "text", + "content": "[25] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[26] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 444, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 499 + ], + "type": "text", + "content": "[28] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4904" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 542 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "text", + "content": "[36] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[37] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "type": "text", + "content": "[38] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[39] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 396 + ], + "type": "text", + "content": "[40] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 497 + ], + "type": "text", + "content": "[42] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "text", + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4905" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_content_list.json b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c78009aa5ae3bfa2ca48fbbcd4087ffb327f98d --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_content_list.json @@ -0,0 +1,1918 @@ +[ + { + "type": "text", + "text": "VRP-SAM: SAM with Visual Reference Prompt", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanpeng Sun $^{1,2*}$ , Jiahui Chen $^{2,3*}$ , Shan Zhang $^{4}$ , Xinyu Zhang $^{2}$ , Qiang Chen $^{2}$ , Gang Zhang $^{2}$ , Errui Ding $^{2}$ , Jingdong Wang $^{2}$ , Zechao Li $^{1\\dagger}$", + "bbox": [ + 114, + 178, + 854, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University of Science and Technology, $^{2}$ Baidu VIS, $^{3}$ Beihang University, $^{4}$ Australian National University {yanpeng.sun, zechao.li}@njust.edu.cn", + "bbox": [ + 187, + 224, + 779, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 314, + 313, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Visual Reference Prompt (VRP) encoder that empowers the Segment Anything Model (SAM) to utilize annotated reference images as prompts for segmentation, creating the VRP-SAM model. In essence, VRP-SAM can utilize annotated reference images to comprehend specific objects and perform segmentation of specific objects in target image. It is note that the VRP encoder can support a variety of annotation formats for reference images, including point, box, scribble, and mask. VRP-SAM achieves a breakthrough within the SAM framework by extending its versatility and applicability while preserving SAM's inherent strengths, thus enhancing user-friendliness. To enhance the generalization ability of VRP-SAM, the VRP encoder adopts a meta-learning strategy. To validate the effectiveness of VRP-SAM, we conducted extensive empirical studies on the Pascal and COCO datasets. Remarkably, VRP-SAM achieved state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Furthermore, VRP-SAM demonstrates strong generalization capabilities, allowing it to perform segmentation of unseen objects and enabling cross-domain segmentation. The source code and models will be available at https://github.com/syp2ysy/VRPSAM", + "bbox": [ + 75, + 345, + 473, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 736, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent, the Segment Anything Model (SAM) [13] has emerged as a foundational visual model for image segmentation. Trained on an extensive datasets comprising billions of labels, SAM has demonstrated remarkable versatility in the realm of universal segmentation. What sets SAM apart is its human-interactive design, allowing segmentation based on user-provided prompts, be they in the form of", + "bbox": [ + 75, + 760, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/279cd4ad8789425d21e414ca504cc98936cbc938093da7b046c4f32cfcb431d9.jpg", + "image_caption": [ + "Figure 1. Comparison of SAM's built-in Point and Box prompt modes with Visual Reference Prompt when handling numerous images. The prompts are all provided by users." + ], + "image_footnote": [], + "bbox": [ + 503, + 314, + 890, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "points, bounding boxes, or coarse masks. This distinctive feature positions SAM as a robust tool that can be adaptable to various tasks and requirements [12, 48].", + "bbox": [ + 498, + 532, + 890, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the existing prompt formats of SAM present significant challenges in practical applications, especially when dealing with complex scenes and numerous images. As shown in Figure 2(a), SAM relies on user-provided prompts (points, boxes, coarse mask) to segment objects in the target image, demanding users to possess a comprehensive understanding of the target objects. In real-world applications, especially in complex scenarios, the level of user familiarity with the target objects can significantly impact the effectiveness of providing specific prompts. Furthermore, variations in the position, size, and quantity of target objects across different images require custom prompts for each image. As illustrated in Figure 1 with the aim of segmenting 'bicycles', users need to customize different prompts for each image, significantly impacting efficiency of SAM. Therefore, we propose integrating visual reference prompts to overcome these limitations and enhance adaptability of SAM.", + "bbox": [ + 496, + 580, + 892, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual reference prompts is annotated reference images that delineate the objects users expect to segment. As shown in Figure 1, by simply providing a visual reference prompt", + "bbox": [ + 498, + 854, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution", + "bbox": [ + 94, + 875, + 223, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23565", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/692023180000c933416f24d6830d1c26a4a0fdd99fb2f92e00cf025279a266c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 89, + 462, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef0ce84e843f6b8e08e4a28818863278170bc594756302651c4ae41d7f47f21a.jpg", + "image_caption": [ + "Figure 2. A Comparison between SAM and VRP-SAM. VRP-SAM introduces a visual reference prompt encoder, accepting annotated reference images with point, scribble, box, and mask formats, offering a distinct enhancement over SAM." + ], + "image_footnote": [], + "bbox": [ + 84, + 183, + 462, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with bicycle, we can segment bicycle in different images without requiring users to provide specific prompts for each image. It significantly enhances the efficiency of SAM while reducing reliance on user familiarity with objects. To achieve this goal, some methods [19, 46] incorporate semantic correlation models [9, 25] to establish reference-target correlation1, obtaining pseudo-masks for the target objects. Following this, a sampling strategy is devised to extract a set of points and bounding boxes from the pseudomasks, serving as prompts for SAM segment the target image. These methods overlook false positives within the pseudo-mask and exhibit high sensitivity to hyperparameters. Consequently, it heavily relies on the quality of the pseudo-mask and has poor generalization.", + "bbox": [ + 75, + 402, + 468, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Toward this end, we propose a straightforward and effective Visual Reference Prompt (VRP) encoder using meta-learning technique, integrated with SAM to create VRPSAM. It leverages annotated reference images as prompts to segment similar semantic objects in the target image. As illustrated in Figure 2(b), the VRP encoder accepts inputs in various annotation formats, including points, scribbles, boxes, and masks. Specifically, the VRP encoder introduces a semantic-related model to encode reference and target images into the same space. Following meta-learning methods, we first extract prototypes of target objects from annotated information in reference images, enhancing the representation of the target objects in both images. Following meta-learning methods, the prototypes of target objects (users' markers) are first generated from annotated reference images, which aim to highlight such target instances", + "bbox": [ + 75, + 616, + 468, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in both reference and target images. Then, we introduce a set of learnable queries to extract the semantic cues of the target objects from attentive enhanced reference features. Then, these queries interact with target images, generating prompt embeddings usable by mask decoder to segment semantically specific objects in the target image. Building upon the foundation of SAM's inherent capabilities, VRPSAM enhances the model's visual reference segmentation prowess. The introduction of Visual Reference Prompts (VRP) not only diversifies the prompts, enabling the model to swiftly segment objects with identical semantics, but also incorporates a meta-learning mechanism that significantly boosts the generalization of model, particularly in dealing with novel objects and cross-domain scenarios.", + "bbox": [ + 500, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To quantitatively assess the generalization capability of VRP-SAM, we follow the dataset configurations commonly used in few-shot segmentation and evaluate our VRP-SAM on Pascal- $5^{i}$ and COCO-20i datasets. Extensive experiments show that VRP-SAM overcomes the limitations of SAM in prompt format, enabling efficient visual reference segmentation. It is significantly better than sampling-based methods, achieving state-of-the-art results on Pascal- $5^{i}$ and COCO-20i datasets. Moreover, we present solid evidence of VRP-SAM's superior performance in handling unknown objects and cross-domain scenarios. Our experiments highlight that VRP-SAM achieves the rapid segmentation of a large number of images based on semantics. Furthermore, the incorporation of meta-learning principles significantly enhances its generalization, making it applicable across various scenarios.", + "bbox": [ + 496, + 303, + 892, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 561, + 640, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Application of SAM", + "text_level": 1, + "bbox": [ + 500, + 587, + 689, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segmentation Anything model (SAM) [13] is a recently introduced category-agnostic interactive segmentation model by Meta. It leverages user-guided instructions for segmentation, such as points, bounding boxes, and coarse masks. Currently, SAM has two primary application approaches. One involves using SAM's segmentation results as prior information to assist downstream tasks. For instance, Inpaint Anything (IA) [43] and Edit Everything [39] leverage SAM's segmentation results for image editing and restoration [42, 47] within the masked regions. Additionally, SEPL [5] employs SAM's segmentation results to enhance pseudo-labels in weakly supervised segmentation tasks [27, 30]. Furthermore, SAM segmentation results serve as prior information in tasks such as crack and volcano crater detection [1, 8].", + "bbox": [ + 496, + 612, + 890, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second involves guiding SAM's segmentation through various prompt combinations. For example, in the context of visual reference segmentation [31], Matcher [19] samples points and boxes from pseudo-masks, utilizing", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1SAM, being a category-agnostic model, encounters challenges in adequately capturing reference-target correlations.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23566", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SAM to refine these pseudo-masks. TAM [41] is applied in object tracking tasks, where it uses point prompts to initialize object masks and subsequently uses SAM to refine low-quality masks. Additionally, SAMAug [6] introduces a visual point enhancement method tailored for SAM, facilitating automatic image annotation. These examples illustrate SAM's effectiveness across different tasks. However, existing methods are constrained by SAM's existing prompt modalities, and they may struggle when confronted with complex objects and unfamiliar scenes. To break these limitations, we have designed a visual prompt encoder for SAM, expanding its applicability to a wider range of scenarios.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual reference segmentation", + "text_level": 1, + "bbox": [ + 76, + 295, + 344, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual reference segmentation [31, 33, 48] aims to guide the segmentation of a target image using a reference image. The goal of this task is to utilize a semantically annotated reference image to instruct the segmentation of objects or regions in the target image that share the same semantics as those in the reference image. In current research, methods can be broadly categorized into two groups: prototype-based and feature-matching-based. Prototype-based methods, such as PFENet [33], PANet [36], and CWT [21], usually focus on distinguishing prototypes with different class-specific features. ASGNet [15], on the other hand, improves the segmentation performance by increasing the number of prototypes. Another feature-matching approach [22, 35] leverages the pixel-level correlations between reference and target images to significantly enhance segmentation performance, as demonstrated by methods like CyCTR [44] and HDMNet [26]. Moreover, modern large-scale vision models [2, 37, 48] have recognized visual reference segmentation as a primary task, given its indispensable role in handling complex objects and unknown scenes. However, it's important to note that SAM [13] does not possess the capability to perform this task, underscoring the necessity of introducing VPR-SAM.", + "bbox": [ + 75, + 318, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 678, + 204, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first review the architecture of Segment Anything Model. Then we formulate the problem setting in this paper.", + "bbox": [ + 75, + 703, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SAM Architecture", + "text_level": 1, + "bbox": [ + 76, + 756, + 256, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM is an interactive segmentation model composed of an image encoder, a prompt encoder, and a mask decoder. Given a target image $I_{t}$ and some geometric prompts, SAM first employs a Vision Transformer (ViT) [34] as image encoder to extract image features. Subsequently, the prompt encoder is utilized to generate the prompt embeddings derived from the user-provided points, boxes, or masks. Finally, the mask decoder integrates the image features and", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompt embeddings, and generates a high-quality mask in a class-agnostic manner.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, current research based on SAM still relies on geometric prompts for segmentation. Thus far, there have been no efforts to introduce new forms of prompts to enhance SAM. We design a novel visual reference prompt to extend the visual reference segmentation capabilities of SAM.", + "bbox": [ + 496, + 122, + 893, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 224, + 702, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual reference segmentation aims to segment objects in target image that same semantic as the annotated object in reference image. Specifically, let $I_{t}$ represent the target image and $I_{r}$ represent the reference image. Given a reference image $I_{r}$ and its annotation $M_{r}^{i}$ , where $i$ denotes the category of the annotated object, we leverage this information to segment all regions in the target image belonging to category $i$ . Depending on the annotation granularity, the reference image have four annotation formats: point, scribble, box, and mask. Point annotation involves providing specific points on the target, scribble requires outlining the target region with arbitrary curves, boxes correspond to bounding boxes around the target, and masks provide pixel-level annotations for the entire target.", + "bbox": [ + 496, + 248, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. VRP-SAM", + "text_level": 1, + "bbox": [ + 500, + 474, + 616, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VRP-SAM extends SAM to perform visual reference segmentation without compromising its original functionality. We propose a training-efficient visual reference prompt encoder that, firstly, accommodates various granularities of visual references and, secondly, directly encodes these visual references into prompt embeddings rather than geometric prompts. These prompt embeddings are then fed directly into the mask decoder of SAM, resulting in the generation of the target mask.", + "bbox": [ + 496, + 501, + 890, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Figure 3, the VRP Encoder consists of feature augmenter and prompt generator. Next, we will delve into the details of VRP Encoder and the loss function.", + "bbox": [ + 496, + 638, + 893, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Feature Augmenter", + "text_level": 1, + "bbox": [ + 500, + 695, + 687, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by meta-learning, the Feature Augmenter separately encodes reference annotations $M_r^i$ into features of both the reference and target images. This process is designed to distinguish between foreground and background representations. To capture semantic correlations between reference and target images, we introduce a semantic-aware image encoder within the VRP encoder, encoding them into the same latent space. To prevent overfitting of the VRP encoder, we freeze the image encoder during the training phase. This ensures a balance between capturing semantic relevance and preventing excessive specialization of the VRP encoder.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23567", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1200482160451724d569485715c0f72ffa0d9b6becd63a62d4df9c425cdcf670.jpg", + "image_caption": [ + "Figure 3. Proposed VRP-SAM framework. Our approach enables SAM to perform visual reference segmentation by extends a VRP encoder. It takes various granularities of visual references as inputs and encodes these visual references into prompt embeddings. Our VRP encoder consists of a feature augmenter and a prompt generator." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Feature Augmenter is illustrated in Figure 3. Initially, leveraging a semantic-aware image encoder (e.g. ResNet-50), we encode $I_{r}$ and $I_{t}$ separately, resulting in the feature map $F_{r} \\in \\mathbf{R}^{C \\times H \\times W}$ and $F_{t} \\in \\mathbf{R}^{C \\times H \\times W}$ . Subsequently, we extract the prototype feature $P_{i}$ corresponding to class $i$ from $F_{r}$ using the mask $M_{r}^{i}$ . This process can be summarized as follows:", + "bbox": [ + 75, + 433, + 470, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\operatorname {M a s k A v g P o o l} \\left(F _ {r}, M _ {r} ^ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 549, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_r^i$ represents one of following annotation formats: point, scribble, box, or mask. To enhance the context information about class $i$ , we concatenate the prototype features and mask with $F_r$ and $F_t$ . $F_r$ is concatenated with mask $m_i$ , and $F_t$ is concatenated with pseudo-mask $m_i^{\\text{pseudo}}$ . $m_i^{\\text{pseudo}}$ is obtained using a common training-free approach, and detailed descriptions are provided in the Appendix. Subsequently, we employ a shared $1 \\times 1$ convolution layer to reduce dimensionality of enhanced features. This process can be summarized as follows:", + "bbox": [ + 75, + 578, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {r} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {r}, P _ {i}, m _ {i}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 739, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {t}, P _ {i}, m _ {i} ^ {\\text {p s e u d o}}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 762, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ultimately, we obtain enhanced image features $F_r' \\in \\mathbf{R}^{C \\times H \\times W}$ and $F_t' \\in \\mathbf{R}^{C \\times H \\times W}$ , which have enhanced context information for the category $i$ . Thus, the enhanced features comprise foreground representations for class $i$ and background representations for the other classes. Subsequently, we feed the enhanced features into the Prompt Generator to obtain a set of visual reference prompts.", + "bbox": [ + 75, + 792, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Prompt Generator", + "text_level": 1, + "bbox": [ + 498, + 433, + 679, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The purpose of the Prompt Generator is to obtain a set of visual reference prompts embedding for the SAM mask decoder. As depicted in Figure 3, the process commences with the introduction of a set of learnable queries denoted as $Q \\in \\mathbf{R}^{N \\times C}$ , where $N$ indicates the number of visual reference prompts. These queries first engage with the reference features $F_r'$ to obtain the category-specific information through cross-attention and self-attention layer:", + "bbox": [ + 496, + 455, + 892, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {r} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {1} \\left(\\operatorname {C r o s s A t t n} _ {1} \\left(Q, F _ {r} ^ {\\prime}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 584, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we obtain a set of queries, $Q_r'$ , possessing knowledge about the object to be segmented. Subsequently, we employ cross-attention to interact these queries with target image feature to obtain the foreground information in target image. Following this, a self-attention layer is used to update the queries, generating a set of prompts $Q_t'$ that align with the representation of SAM:", + "bbox": [ + 496, + 609, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {2} \\left(\\operatorname {C r o s s A t t n} _ {2} \\left(Q _ {r} ^ {\\prime}, F _ {t} ^ {\\prime}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 722, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final $Q_{t}^{\\prime}$ serve as the visual reference prompt embeddings for SAM, equipped with the capability to guide the segmentation of the foreground in the target image. By inputting this set of visual reference prompt embeddings into the mask decoder, the mask $M_{t}^{i}\\in \\mathbf{R}^{1\\times H\\times W}$ for category $i$ in the target image can be obtained.", + "bbox": [ + 496, + 747, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Loss Function", + "text_level": 1, + "bbox": [ + 500, + 847, + 643, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ Binary Cross-Entropy (BCE) loss and Dice loss to supervise the learning of the Visual Reference Prompt", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23568", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/64e041d8ad3002fc5827f74ce1df680c355e73c01bc4cf17cc6ee01e2879acab.jpg", + "table_caption": [ + "Table 1. Compare with other foundation models. Results of one-shot semantic segmentation on COCO- $20^{i}$ . Gray indicates the model is trained by in-domain datasets. $\\dagger$ indicates the method using SAM." + ], + "table_footnote": [], + "table_body": "
MethodsLabel typeF-0F-1F-2F-3Means
Painter [37]mask.31.235.333.532.433.1
SegGPT [38]56.357.458.951.756.1
PerSAM† [46]23.123.622.023.423.0
PerSAM-F† [46]22.324.023.424.123.5
Masker† [19]52.753.552.652.152.7
VRP-SAM†point.30.139.243.040.438.2
scribble.40.252.052.444.447.2
box.44.549.355.749.149.7
mask.48.155.860.051.653.9
", + "bbox": [ + 81, + 148, + 468, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoder. The BCE loss ensures pixel-wise accuracy, while the Dice loss provides additional context for pixel-level segmentation. Therefore, the total loss of VRP-SAM is:", + "bbox": [ + 76, + 299, + 468, + 344 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {t o t a l}} = \\underbrace {- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log (p _ {i}) + (1 - y _ {i}) \\log (1 - p _ {i}) \\right]} _ {\\mathrm {B C E L o s s}} \\\\ + \\underbrace {1 - \\frac {2 \\sum_ {i = 1} ^ {N} \\left(p _ {i} \\cdot y _ {i}\\right)}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} y _ {i} ^ {2}}} _ {\\text {D i c e L o s s}} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 345, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the total number of pixels, $y_{i}$ is the ground truth label for pixel $i$ , and $p_i$ is the predicted probability of pixel $i$ belonging to the object. By combining these two losses, we comprehensively consider both accuracy and contextual effects, thus guiding the Visual Reference Prompt Encoder more effectively in generating precise segmentation results.", + "bbox": [ + 75, + 469, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setting", + "text_level": 1, + "bbox": [ + 76, + 604, + 169, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: To validate segmentation performance and generalization capability of VRP-SAM, we conducted extensive experiments following the few-shot setting [14, 31, 44] on COCO-20 $^i$ [24] and PASCAL-5 $^i$ [29] datasets. Specifically, we organized all classes from both datasets into 4 folds. For each fold, PASCAL-5 $^i$ [29] comprises 15 base classes for training and 5 novel classes for testing, while COCO-20 $^i$ [24] includes 60 training base classes and 20 testing novel classes. To assess performance of model, we randomly sampled 1000 reference-target pairs in each fold. In each fold, As the mentioned datasets lack labels for point, scribble, and box annotations, we followed the SEEM [48] to generate these annotation labels by simulating user inputs randomly based on the reference ground truth masks.", + "bbox": [ + 75, + 628, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details: In visual reference prompt encoder, we use VGG-16 [3] and ResNet-50 [9] as the im", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "age encoder and initialize it with ImageNet [28] pre-trained weights. We employed the AdamW optimizer [20] along with a cosine learning rate decay strategy for training VRPSAM. Specifically, on the COCO- $20^{i}$ dataset, we conducted 50 epochs of training with an initial learning rate of 1e-4 and a batch size of 8. For the PASCAL- $5^{i}$ dataset, the model was trained for 100 epochs with an initial learning rate of 2e-4 and a batch size of 8. In VRP, the number of queries is set to 50 by default, and the input image size of all experiments needs to be adjusted to $512 \\times 512$ . Following SEEM [48], during training, we obtain annotations for points, scribbles, and boxes based on mask annotations. We provide detailed descriptions of this process in the Appendix. To ensure a fair comparison, VRP-SAM is exclusively compared to previous works [16, 26, 37] using visual reference prompts based on the mean intersection over union (mIoU).", + "bbox": [ + 496, + 90, + 893, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparison with the State-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 359, + 821, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with other foundation models. Leveraging the foundation model enables few-shot segmentation. We compared various approaches utilizing Painter [37], SegGPT [38], and SAM as foundation models for few-shot segmentation, providing evaluation metrics on the COCO-20 $^i$ dataset. As shown in Table 1, VRP-SAM achieves $53.9\\%$ mean mIoU without training on novel classes, achieving comparable with SegGPT. Note that the training data of SegGPT include all classes in COCO. Furthermore, our VRP-SAM outperforms other SAM-base methods including Matcher [19], which employs a DINOv2 [25] pretrained ViT-L [34] as image encoder.", + "bbox": [ + 496, + 383, + 890, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with few-shot methods. To validate the effectiveness of VRP-SAM, we compared it with state-of-the-art few-shot segmentors on the novel set of COCO- $20^{i}$ and PASCAL- $5^{i}$ datasets. To ensure a fair comparison, VRP-SAM is trained and tested separately in each fold, ensuring no overlap in classes between the training and test sets. The results in Table 2 demonstrate that VRP-SAM achieves state-of-the-art results on COCO $20^{i}$ and PASCAL- $5^{i}$ , with mIoU scores of 53.9 and 71.9, respectively. Notably, when using VGG-16 as the image encoder for VRP, VRP-SAM achieves mIoU scores of 48.0 and 68.7 on COCO- $20^{i}$ and PASCAL- $5^{i}$ datasets. It indicates that VRP-SAM achieves optimal performance on the novel set with only 1.6M learnable parameters, highlighting its powerful generalization capability.", + "bbox": [ + 496, + 566, + 892, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Comparison with Geometric Prompts", + "text_level": 1, + "bbox": [ + 498, + 800, + 823, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we design Visual Reference Prompts to represent target information. Different from the Geometric Prompts (GP) provided by SAM, our VRP is more flexible and robust. We conducted experiments comparing GP and VRP to validate the superiority of our method (see Table 3).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23569", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ba7aa129b550ea44ffc57401ad97af1bb341afbfb39f814d41c98f1c61e4395.jpg", + "table_caption": [ + "Table 2. Performance of one-shot semantic segmentation on COCO- ${20}^{i}$ and PASCAL- ${5}^{i}$ . The red and blue colors respectively represent the optimal and suboptimal results." + ], + "table_footnote": [], + "table_body": "
MethodImage encoderLearnable paramsCOCO-20iPASCAL-5i
F-0F-1F-2F-3MeanF-0F-1F-2F-3Mean
PFENet [33]VGG-1610.4M35.438.136.834.736.356.968.254.552.458.0
BAM [14]4.9M36.447.143.341.742.163.270.866.157.564.4
HDMNet [26]4.2M40.750.648.244.045.964.871.467.756.465.1
VRP-SAM1.6M43.651.750.046.548.070.074.768.361.968.7
PFENet [33]ResNet-5010.4M36.538.634.533.835.861.769.555.456.360.8
HSNet [22]2.6M36.343.138.738.739.264.370.760.360.564.0
CyCTR [44]15.4M38.943.039.639.840.365.771.059.559.764.0
SSP [7]8.7M35.539.637.936.737.460.567.866.451.061.4
NTRENet [18]19.9M36.842.639.937.939.365.472.359.459.864.2
DPCN [17]-42.047.043.339.743.065.771.669.160.666.7
VAT [10]3.2M39.043.842.639.741.367.672.062.360.165.5
BAM [14]4.9M39.449.946.245.245.269.073.667.661.167.8
HDMNet [26]4.2M43.855.351.649.450.071.075.468.962.169.4
VRP-SAM1.6M48.155.860.051.653.973.978.370.665.071.9
DCAMASwin-B47.7M49.552.752.848.750.972.273.864.367.169.3
", + "bbox": [ + 80, + 123, + 893, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/57951279882aec9b3f90783f37cc2322b289f674b7ccf96bca57cfff9f77fa11.jpg", + "table_caption": [ + "Table 3. Comparison with geometric prompts. Geometric prompts are randomly sampled from the pseudo-mask. $\\dagger$ indicates the carefully designed sampling strategy proposed in [19]." + ], + "table_footnote": [], + "table_body": "
MethodImage EncoderPromptsMean IoU
GP-SAMResNet-50box.19.7
point.21.7
box. + point.23.2
DINOv2box.31.3
point.36.6
box. + point.37.8
box. + point. † [19]52.7
VRP-SAMResNet-50point.38.4
scribble.47.3
box.49.7
mask.53.9
DINOv2-B/14mask.60.4
", + "bbox": [ + 80, + 429, + 465, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the GP-SAM experiments, we initially used an Image Encoder to generate a pseudo-mask for the target image following [33], and subsequently obtained points or bounding boxes from the pseudo-mask as geometric prompts. Experimental results consistently demonstrate the ongoing superiority of our VRP over the GP approach. We visualize the segmentation results of GP-SAM and our VRP-SAM in Figure 4. Visualization results indicate that the GP approach is prone to generating false-positive prompts, significantly impacting segmentation performance. In contrast, our VRP effectively avoids such issues.", + "bbox": [ + 75, + 635, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Generalization Evaluation", + "text_level": 1, + "bbox": [ + 76, + 808, + 316, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Domain shift: Next, we assess the effectiveness of VRPSAM in a domain shift scenario, which necessitates signif", + "bbox": [ + 76, + 832, + 468, + 862 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bcd4cb3092239f913c25dc0beb955a87c28b8944125d4e25fefa2acc636beef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 375, + 594, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a0d83cd3fd737170c45a218170bf81f94e28c0a14dd2af163c23b80cad60d12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 436, + 594, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49c984490b6632e211a9d199f10b1b7321ac142d499abe111d699a47740d64d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 506, + 594, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/246eaed965df90310858d4cf1967a3a6e7f460df87435f1af853c772be1834d3.jpg", + "image_caption": [ + "Visual Reference" + ], + "image_footnote": [], + "bbox": [ + 508, + 575, + 594, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2288c72273e4b217ab76c97dd49b7e950f62ef71d613960ee04e39fc1b411d77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 375, + 687, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd1334a4f875164eeacc12e1e6f4ffb643bcb14ad493cd1d25a678f8c896023f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 436, + 687, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f837fc9d7de39914eae01792ad01bd68fad350c47af3191eab0a7c134ea314ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 506, + 687, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14d799081ecd8daa077f5395ffff76aa2f3495feba55114608424e0d453ea8d1.jpg", + "image_caption": [ + "GP-SAM-RN50" + ], + "image_footnote": [], + "bbox": [ + 602, + 575, + 687, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d87dd1cd65efd2c715a7b6a14a93e7e2bea2730b863dff0b68c6f2a647481f31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 375, + 777, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ab4534d999a15ae392d8fd7e3d56e417612e99c6db0af8b116a71a95160155a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 436, + 777, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f51aeb8de1305a24306bbb087480d778617e39b1b2ec5e4c675483ff9c75c7a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 506, + 777, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9863e7291521b2a40cd51d5f602289c1c5e7583c0556573922b37afb26225bcc.jpg", + "image_caption": [ + "GP-SAM-DINOv2" + ], + "image_footnote": [], + "bbox": [ + 692, + 575, + 777, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/222840ba44ce1080c1f40bdd2a15f0d8b071d015da2202b495587cc685e10132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 375, + 867, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ab87febacbd1c910b62f682511d6376d21560beb44ce5fbc1e0623ae533e729.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 436, + 867, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8af1b9e16ae86437828b11c40198bbf1f5e11892a1189fa480e30992f8244a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 506, + 867, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f29b352fed25b2d46515260f7b3151c14a8c046bf2f1bc0fae1089afbf9a0a40.jpg", + "image_caption": [ + "VRP-SAM-RN50", + "Figure 4. The visualization results of VRP-SAM and GP-SAM." + ], + "image_footnote": [], + "bbox": [ + 782, + 575, + 867, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "icant domain differences between training and testing sets. Following prior work [22, 32, 33], we trained on COCO- $20^{i}$ and tested on PASCAL- $5^{i}$ , where the classes in training set do not overlap with those in test set. The results in Table 4, representing the average across four folds, clearly demonstrate the superior performance of VRP-SAM in domain shift scenario. Notably, VRP-SAM with scribble annotations outperforms FP-Trans [45] by $2.8\\%$ , while VRP-SAM with mask annotations surpasses DGPNet [11] by $5.8\\%$ . This affirms the robust generalization capability of VRPSAM and its effectiveness in domain transfer scenarios.", + "bbox": [ + 496, + 688, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization: To assess the generalization capability of VRP-SAM across diverse image styles, we curated a collection of target images from web, spanning various gen", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2For point prompts, we randomly sample 5 points from the pseudomask. For box prompt, we adopt the bounding box of the pseudo-mask.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23570", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80652b85f5f7f83b00f5695e5d600e4dec217ca35099b5b37bd926294b0b86a6.jpg", + "image_caption": [ + "JIOd" + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 230, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b81ba2695a92d5437bc87437341b00615ac7e52e28bc315e638a7d8ac62d7ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 95, + 362, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dd71cd829e2a5a4bfa435ff0fe8c09bff84b7a2a7d6fdf159835edac1018225.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 95, + 493, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a897af2669cd10a89d0452c5ffdce5118bbaf97be580f7df978177f3a12b1b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 95, + 622, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f3c329154d485d4af9bb1f4e12566bc038ce20c73df509cdcbeda1c66db2948.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 95, + 751, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27639975f24ff2b816b35af3ca2e62ff8266a3f3e6975aa46c16991db4e20b42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 95, + 882, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fd91bc4d07435e0cfd07d65b7a5a09c16a055212e6eda90ccbb5eae5132576d.jpg", + "image_caption": [ + "S" + ], + "image_footnote": [], + "bbox": [ + 81, + 196, + 230, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b61d50a7148940ddf68cedf883129879d89f3974b3be04cad5320869a33f6a9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 196, + 364, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a99809c44c45c18a4b57e4d836772b166f1059003f811ad36f86675f18fa32d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 196, + 493, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42af8c625ce8c59d0360b6b381a3fd89198a1415bffb6467ff94e1294a7d192f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 196, + 622, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c3859036847fc6cc75094782a41f91a92fc5095e47049da6ba3be65dec2876b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 196, + 751, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e464d08743027204ca2b7f7ab3ae33c6f4c8ce1580f3edb577f5a21c1a4fa0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 196, + 880, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8b7fc657345272017af5182b6eafe5846e20ab226ba50c0bf3d0be193a252b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 296, + 230, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a442cb22db880e4eee1a97386ed0c728eb9c3facdc41c546fd62755176eb68e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 296, + 364, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8dbffd86720765f7e4682f4b5b7790a3d8be1406a1da46d2a0b7452beba61ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 296, + 493, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4532c83c7bf26e7647ec2cafe03f9a3d13f7bce3e09af07116064478af3872a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 296, + 622, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f3dc69ecfeb8643dd5b93149f7815d24a81db00884253607dd439a7760c9921.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 296, + 751, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06b324a7d31b1b11a80b04ca0fb60368552c3075388330ba3766a8dcbfe8d161.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 296, + 880, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e041444f742e32b85f737a51fedf0574c170b5a7e296c2dffdc4a1ed3ce7ff0.jpg", + "image_caption": [ + "VRP", + "Figure 5. Qualitative results of VRP-SAM across diverse image styles is presented. The target images were sourced from the internet." + ], + "image_footnote": [], + "bbox": [ + 81, + 396, + 230, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b08169c45522b6f22eff1e42f3dd26e7fff2bb541e0d38ce5c48e1791e9bb41.jpg", + "image_caption": [ + "Image-1" + ], + "image_footnote": [], + "bbox": [ + 235, + 396, + 364, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38c2788b42e327737dec89980469bfb142bdbfe9772264927362bec3ea1cbd69.jpg", + "image_caption": [ + "Image-2" + ], + "image_footnote": [], + "bbox": [ + 367, + 396, + 493, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a4017350ddf98c60aa16176e6a44a95491249c073531cf151254f5dc156839d.jpg", + "image_caption": [ + "Image-3" + ], + "image_footnote": [], + "bbox": [ + 495, + 396, + 622, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8aa470003623f34b70d56547105e222cf8591414cca5656ac2af956f5207d295.jpg", + "image_caption": [ + "Image-4" + ], + "image_footnote": [], + "bbox": [ + 624, + 396, + 751, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7f5f00a4c2bd6bd2ff7157cfcb467fe8b2547563e9d55a1f23fc01c2f305e5c.jpg", + "image_caption": [ + "Image-5" + ], + "image_footnote": [], + "bbox": [ + 754, + 396, + 880, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "res such as natural landscapes, artworks, and complex environments. Visual Reference Prompts (VRPs) were selected from the COCO dataset to guide the segmentation of these target images. The segmentation results on different image styles are presented in Figure 5. It demonstrate that VRP-SAM adeptly adapts to varied image styles, accurately delineating target objects. Notably, VRP-SAM excels in handling ink-style images, showcasing refined segmentation performance. This compellingly underscores the robustness and effectiveness of VRP-SAM when confronted with images of unknown styles.", + "bbox": [ + 75, + 541, + 468, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 710, + 228, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of VRP-SAM, we conducted extensive ablation studies on PASCAL- $5^{i}$ . ResNet-50 was chosen as the image encoder for VRP to ensure experimental consistency. These experiments aimed to thoroughly investigate the performance of VRP-SAM under various conditions.", + "bbox": [ + 75, + 734, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss: To assess the impact of Binary Cross-Entropy (BCE) and Dice losses on VRP-SAM, experiments were conducted on the PASCAL- $5^{i}$ dataset. Results in Table 5 indicate comparable performance when using BCE or Dice loss individually. Notably, the optimal performance for", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c2c6c28c8e3d3435185dcd0dfe040d79011d904a9ea84b3e66d736cadc2f9d4.jpg", + "table_caption": [ + "Table 4. Evaluation (Mean IoU $(\\%)$ ) under the domain shift from COCO-20 to PASCAL-5." + ], + "table_footnote": [], + "table_body": "
MethodImage encoderLabel typeMean
RPMM [40]ResNet-50mask.49.6
PFENet [33]61.1
RePRI [4]63.2
VAT-HM [23]65.1
HSNet [22]ResNet-101mask.64.1
DGPNet [11]70.1
FP-Trans [45]DeiT-B/16mask.69.7
VRP-SAMResNet-50point.63.5
scribble.72.5
box.72.3
mask.75.9
", + "bbox": [ + 504, + 573, + 890, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VRP-SAM is achieved when both losses are combined. It emphasizes the importance of BCE and Dice losses in guiding VRP-SAM to produce more robust and accurate masks. BCE loss ensures precise pixel-level classification, while Dice loss facilitates accurate spatial localization of objects. Simultaneously employing BCE and Dice losses in VRP-SAM maximizes their complementary advantages.", + "bbox": [ + 496, + 752, + 890, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of query: In Figure 6, we conducted a comprehensive analysis of the impact of varying query quantities on VRP-SAM's performance. We observed", + "bbox": [ + 496, + 859, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23571", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/227fac2ac9f4b2027bfc56f7b2f1400f29a66f484303c2cf95d5e4b722b68530.jpg", + "table_caption": [ + "Table 5. Ablation study on different loss function in VRP-SAM." + ], + "table_footnote": [], + "table_body": "
MethodLabel typeBce lossDice lossMeans
VRP-SAMmask.✓✗70.1
✗✓70.3
✓✓71.9
", + "bbox": [ + 81, + 107, + 468, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1d89d230e8993a88026585303bcab175c30f91996180c870bd4a2ff8d009941.jpg", + "image_caption": [ + "Figure 6. Ablation study on VRP-SAM with different query numbers. The x-axis shows the number of queries, and the y-axis represents model performance." + ], + "image_footnote": [], + "bbox": [ + 81, + 172, + 460, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a positive correlation between increased query numbers and improved segmentation quality. However, once the query count surpassed 50, the performance gain started to diminish. This phenomenon suggests that 50 queries provide ample effective guidance for the model, and further increases in query count do not yield significant improvements, leading to performance saturation. To maintain high performance while minimizing model learnable parameters, we set the query quantity in VRP-SAM to 50. This decision optimally balances guidance information and model efficiency.", + "bbox": [ + 75, + 405, + 468, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization of query: In Table 6, we compared various query initialization strategies, such as random initialization, foreground prototype (FP), background prototype (BP), and a hybrid prototype with half foreground and half background (half-FP & half-BP). Surprisingly, random initialization outperformed all other strategies, showcasing superior performance. This unexpected outcome can be attributed to the intricate segmentation task and the dataset's object diversity. Random initialization enables the model to explore a broader range of states, avoiding potential local minima. In contrast, prototype-based initializations, whether foreground, background, or a combination, might introduce biases, limiting adaptability to diverse object characteristics.", + "bbox": [ + 75, + 571, + 468, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of VRPs: In our investigation of the influence of number of visual reference prompts on segmentation results, we explored utilization of few visual reference prompts. This approach involves sending several visual reference prompts to the VRP encoder, generating multiple prompt embeddings. These embeddings are concatenated and forwarded to the mask decoder, resulting in the final mask. The result in Table 7, we observed a substantial en", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7bc7f61c8f44e17fdd64c2c715fa8200f882cc917db77e91644d5370e4afc465.jpg", + "table_caption": [ + "Table 6. Ablation study of different query initialization methods on VRP-SAM." + ], + "table_footnote": [], + "table_body": "
MethodLabel typeImage encoderMean IoU
VRP-SAMmask.FP68.2
BP62.6
half-FP & half-BP67.4
random71.9
", + "bbox": [ + 517, + 119, + 870, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb0c7f11ea7557bf172608879da9254c25215c73fdaff7defa8f62ee134f227a.jpg", + "table_caption": [ + "Table 7. Ablation Study on Few Visual Reference prompts for VRP-SAM." + ], + "table_footnote": [], + "table_body": "
MethodLabel type1-VRP5-VRP
VRP-SAMpoint.62.964.1(+1.2)
scribble.66.868.4(+1.6)
box.69.470.5(+1.1)
mask.71.972.9(+1.0)
", + "bbox": [ + 504, + 234, + 890, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hancement in segmentation performance with an increase in the number of visual reference prompts. The results emphasize positive influence of incorporating multiple visual reference prompts, enhancing the model's ability to accurately capture intricate details and nuances associated with target object. The diverse prompts contribute enriched contextual information, facilitating a more comprehensive understanding of the object's characteristics and leading to generation of precise and nuanced segmentation masks.", + "bbox": [ + 496, + 325, + 890, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 479, + 617, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present VRP-SAM, an innovative extension of the SAM framework achieved through integration of a Visual Reference Prompt (VRP) encoder. This addition empowers SAM to leverage visual reference prompts for guided segmentation. The core methodology involves encoding annotated reference images through the VRP encoder, which then interacts with the target image to generate meaningful prompts for segmentation within the SAM framework. The seamless fusion of VRP encoder with SAM, resulting in VRP-SAM, enhances the model's generality and adaptability. VRP-SAM overcomes limitations posed by SAM's existing prompt formats, especially in complex scenarios and large datasets. The introduced visual reference prompts, including point, box, scribble, and mask annotations, offer a flexible solution, expanding the model's applicability. Extensive empirical studies conducted showcase VRP-SAM's state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Notably, VRP-SAM demonstrates robust generalization capabilities, excelling in segmentation tasks for novel objects and cross-domain scenarios.", + "bbox": [ + 496, + 507, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work was partially supported by the National Key Research and Development Program of China under Grant 2022ZD0118802 and the National Natural Science Foundation of China (Grant No. U20B2064 and U21B2043).", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23572", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohsen Ahmadi, Ahmad Gholizadeh Lonbar, Abbas Sharifi, Ali Tarlani Beris, Mohammadsadegh Nouri, and Amir Sharifzadeh Javidi. Application of segment anything model for civil infrastructure defect assessment. arXiv preprint arXiv:2304.12600, 2023.", + "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In Advances in Neural Information Processing Systems, pages 25005-25017, 2022.", + "[3] Yoshua Bengio and Yann LeCun. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015.", + "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021.", + "[5] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023.", + "[6] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, Wei Liu, et al. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023.", + "[7] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022.", + "[8] Iraklis Giannakis, Anshuman Bhardwaj, Lydia Sam, and Georgios Leontidis. Deep learning universal crater detection using segment anything model (sam). arXiv preprint arXiv:2304.07764, 2023.", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.", + "[10] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126, 2022.", + "[11] Joakim Johnander, Johan Edstedt, Michael Felsberg, Fahad Shahbaz Khan, and Martin Danelljan. Dense gaussian processes for few-shot segmentation. In European Conference on Computer Vision, pages 217-234, 2022.", + "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. arXiv preprint arXiv:2306.01567, 2023.", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.", + "[14] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE conference on computer vision and pattern recognition, pages 8057-8067, 2022." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021.", + "[16] Zechao Li, Yanpeng Sun, Liyan Zhang, and Jinhui Tang. Cnt-net: Context-based tandem network for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9904-9917, 2021.", + "[17] Jie Liu, Yanqi Bao, Guo-Sen Xie, Huan Xiong, Jan-Jakob Sonke, and Efstratios Gavves. Dynamic prototype convolution network for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2022.", + "[18] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022.", + "[19] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023.", + "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018.", + "[21] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In IEEE International Conference on Computer Vision, pages 8741-8750, 2021.", + "[22] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In IEEE international conference on computer vision, pages 6941-6952, 2021.", + "[23] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Hm: Hybrid masking for few-shot segmentation. In European Conference on Computer Vision, pages 506-523, 2022.", + "[24] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In IEEE International Conference on Computer Vision, pages 622-631, 2019.", + "[25] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", + "[26] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023.", + "[27] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023.", + "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23573", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015.", + "[29] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017.", + "[30] Yanpeng Sun and Zechao Li. Ssa: Semantic structure aware inference for weakly pixel-wise dense predictions without cost. arXiv preprint arXiv:2111.03392, 2021.", + "[31] Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, and Jingdong Wang. Singular value fine-tuning: Few-shot segmentation requires few-parameters fine-tuning. In Advances in Neural Information Processing Systems, pages 37484–37496, 2022.", + "[32] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023.", + "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020.", + "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017.", + "[35] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, 2016.", + "[36] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In IEEE international conference on computer vision, pages 9197-9206, 2019.", + "[37] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023.", + "[38] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In IEEE International Conference on Computer Vision, pages 1130-1140, 2023.", + "[39] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023.", + "[40] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 763-778, 2020.", + "[41] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023.", + "[42] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with seg-" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ment anything models. arXiv preprint arXiv:2306.04121, 2023.", + "[43] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023.", + "[44] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. In Advances in Neural Information Processing Systems, pages 21984-21996, 2021.", + "[45] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. In Advances in Neural Information Processing Systems, pages 6575–6588, 2022.", + "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023.", + "[47] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023.", + "[48] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023." + ], + "bbox": [ + 501, + 92, + 890, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23574", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_model.json b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dffe2389ad0a1eb4ad3678fc4d116115f738e873 --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_model.json @@ -0,0 +1,2530 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.73, + 0.154 + ], + "angle": 0, + "content": "VRP-SAM: SAM with Visual Reference Prompt" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.179, + 0.856, + 0.22 + ], + "angle": 0, + "content": "Yanpeng Sun\\(^{1,2*}\\), Jiahui Chen\\(^{2,3*}\\), Shan Zhang\\(^{4}\\), Xinyu Zhang\\(^{2}\\), Qiang Chen\\(^{2}\\), Gang Zhang\\(^{2}\\), Errui Ding\\(^{2}\\), Jingdong Wang\\(^{2}\\), Zechao Li\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.226, + 0.781, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanjing University of Science and Technology, \\(^{2}\\)Baidu VIS, \\(^{3}\\)Beihang University, \\(^{4}\\)Australian National University {yanpeng.sun, zechao.li}@njust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.315, + 0.314, + 0.331 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.474, + 0.709 + ], + "angle": 0, + "content": "In this paper, we propose a novel Visual Reference Prompt (VRP) encoder that empowers the Segment Anything Model (SAM) to utilize annotated reference images as prompts for segmentation, creating the VRP-SAM model. In essence, VRP-SAM can utilize annotated reference images to comprehend specific objects and perform segmentation of specific objects in target image. It is note that the VRP encoder can support a variety of annotation formats for reference images, including point, box, scribble, and mask. VRP-SAM achieves a breakthrough within the SAM framework by extending its versatility and applicability while preserving SAM's inherent strengths, thus enhancing user-friendliness. To enhance the generalization ability of VRP-SAM, the VRP encoder adopts a meta-learning strategy. To validate the effectiveness of VRP-SAM, we conducted extensive empirical studies on the Pascal and COCO datasets. Remarkably, VRP-SAM achieved state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Furthermore, VRP-SAM demonstrates strong generalization capabilities, allowing it to perform segmentation of unseen objects and enabling cross-domain segmentation. The source code and models will be available at https://github.com/syp2ysy/VRPSAM" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.737, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.868 + ], + "angle": 0, + "content": "In recent, the Segment Anything Model (SAM) [13] has emerged as a foundational visual model for image segmentation. Trained on an extensive datasets comprising billions of labels, SAM has demonstrated remarkable versatility in the realm of universal segmentation. What sets SAM apart is its human-interactive design, allowing segmentation based on user-provided prompts, be they in the form of" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.315, + 0.891, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.478, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Comparison of SAM's built-in Point and Box prompt modes with Visual Reference Prompt when handling numerous images. The prompts are all provided by users." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.579 + ], + "angle": 0, + "content": "points, bounding boxes, or coarse masks. This distinctive feature positions SAM as a robust tool that can be adaptable to various tasks and requirements [12, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.852 + ], + "angle": 0, + "content": "However, the existing prompt formats of SAM present significant challenges in practical applications, especially when dealing with complex scenes and numerous images. As shown in Figure 2(a), SAM relies on user-provided prompts (points, boxes, coarse mask) to segment objects in the target image, demanding users to possess a comprehensive understanding of the target objects. In real-world applications, especially in complex scenarios, the level of user familiarity with the target objects can significantly impact the effectiveness of providing specific prompts. Furthermore, variations in the position, size, and quantity of target objects across different images require custom prompts for each image. As illustrated in Figure 1 with the aim of segmenting 'bicycles', users need to customize different prompts for each image, significantly impacting efficiency of SAM. Therefore, we propose integrating visual reference prompts to overcome these limitations and enhance adaptability of SAM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Visual reference prompts is annotated reference images that delineate the objects users expect to segment. As shown in Figure 1, by simply providing a visual reference prompt" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.224, + 0.887 + ], + "angle": 0, + "content": "*equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.224, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.464, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.184, + 0.464, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Figure 2. A Comparison between SAM and VRP-SAM. VRP-SAM introduces a visual reference prompt encoder, accepting annotated reference images with point, scribble, box, and mask formats, offering a distinct enhancement over SAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.469, + 0.617 + ], + "angle": 0, + "content": "with bicycle, we can segment bicycle in different images without requiring users to provide specific prompts for each image. It significantly enhances the efficiency of SAM while reducing reliance on user familiarity with objects. To achieve this goal, some methods [19, 46] incorporate semantic correlation models [9, 25] to establish reference-target correlation1, obtaining pseudo-masks for the target objects. Following this, a sampling strategy is devised to extract a set of points and bounding boxes from the pseudomasks, serving as prompts for SAM segment the target image. These methods overlook false positives within the pseudo-mask and exhibit high sensitivity to hyperparameters. Consequently, it heavily relies on the quality of the pseudo-mask and has poor generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Toward this end, we propose a straightforward and effective Visual Reference Prompt (VRP) encoder using meta-learning technique, integrated with SAM to create VRPSAM. It leverages annotated reference images as prompts to segment similar semantic objects in the target image. As illustrated in Figure 2(b), the VRP encoder accepts inputs in various annotation formats, including points, scribbles, boxes, and masks. Specifically, the VRP encoder introduces a semantic-related model to encode reference and target images into the same space. Following meta-learning methods, we first extract prototypes of target objects from annotated information in reference images, enhancing the representation of the target objects in both images. Following meta-learning methods, the prototypes of target objects (users' markers) are first generated from annotated reference images, which aim to highlight such target instances" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "in both reference and target images. Then, we introduce a set of learnable queries to extract the semantic cues of the target objects from attentive enhanced reference features. Then, these queries interact with target images, generating prompt embeddings usable by mask decoder to segment semantically specific objects in the target image. Building upon the foundation of SAM's inherent capabilities, VRPSAM enhances the model's visual reference segmentation prowess. The introduction of Visual Reference Prompts (VRP) not only diversifies the prompts, enabling the model to swiftly segment objects with identical semantics, but also incorporates a meta-learning mechanism that significantly boosts the generalization of model, particularly in dealing with novel objects and cross-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.893, + 0.545 + ], + "angle": 0, + "content": "To quantitatively assess the generalization capability of VRP-SAM, we follow the dataset configurations commonly used in few-shot segmentation and evaluate our VRP-SAM on Pascal- \\(5^{i}\\) and COCO-20i datasets. Extensive experiments show that VRP-SAM overcomes the limitations of SAM in prompt format, enabling efficient visual reference segmentation. It is significantly better than sampling-based methods, achieving state-of-the-art results on Pascal- \\(5^{i}\\) and COCO-20i datasets. Moreover, we present solid evidence of VRP-SAM's superior performance in handling unknown objects and cross-domain scenarios. Our experiments highlight that VRP-SAM achieves the rapid segmentation of a large number of images based on semantics. Furthermore, the incorporation of meta-learning principles significantly enhances its generalization, making it applicable across various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.563, + 0.642, + 0.578 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.691, + 0.604 + ], + "angle": 0, + "content": "2.1. Application of SAM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Segmentation Anything model (SAM) [13] is a recently introduced category-agnostic interactive segmentation model by Meta. It leverages user-guided instructions for segmentation, such as points, bounding boxes, and coarse masks. Currently, SAM has two primary application approaches. One involves using SAM's segmentation results as prior information to assist downstream tasks. For instance, Inpaint Anything (IA) [43] and Edit Everything [39] leverage SAM's segmentation results for image editing and restoration [42, 47] within the masked regions. Additionally, SEPL [5] employs SAM's segmentation results to enhance pseudo-labels in weakly supervised segmentation tasks [27, 30]. Furthermore, SAM segmentation results serve as prior information in tasks such as crack and volcano crater detection [1, 8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The second involves guiding SAM's segmentation through various prompt combinations. For example, in the context of visual reference segmentation [31], Matcher [19] samples points and boxes from pseudo-masks, utilizing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "1SAM, being a category-agnostic model, encounters challenges in adequately capturing reference-target correlations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "SAM to refine these pseudo-masks. TAM [41] is applied in object tracking tasks, where it uses point prompts to initialize object masks and subsequently uses SAM to refine low-quality masks. Additionally, SAMAug [6] introduces a visual point enhancement method tailored for SAM, facilitating automatic image annotation. These examples illustrate SAM's effectiveness across different tasks. However, existing methods are constrained by SAM's existing prompt modalities, and they may struggle when confronted with complex objects and unfamiliar scenes. To break these limitations, we have designed a visual prompt encoder for SAM, expanding its applicability to a wider range of scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.296, + 0.345, + 0.312 + ], + "angle": 0, + "content": "2.2. Visual reference segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Visual reference segmentation [31, 33, 48] aims to guide the segmentation of a target image using a reference image. The goal of this task is to utilize a semantically annotated reference image to instruct the segmentation of objects or regions in the target image that share the same semantics as those in the reference image. In current research, methods can be broadly categorized into two groups: prototype-based and feature-matching-based. Prototype-based methods, such as PFENet [33], PANet [36], and CWT [21], usually focus on distinguishing prototypes with different class-specific features. ASGNet [15], on the other hand, improves the segmentation performance by increasing the number of prototypes. Another feature-matching approach [22, 35] leverages the pixel-level correlations between reference and target images to significantly enhance segmentation performance, as demonstrated by methods like CyCTR [44] and HDMNet [26]. Moreover, modern large-scale vision models [2, 37, 48] have recognized visual reference segmentation as a primary task, given its indispensable role in handling complex objects and unknown scenes. However, it's important to note that SAM [13] does not possess the capability to perform this task, underscoring the necessity of introducing VPR-SAM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.205, + 0.696 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.75 + ], + "angle": 0, + "content": "In this section, we first review the architecture of Segment Anything Model. Then we formulate the problem setting in this paper." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.258, + 0.772 + ], + "angle": 0, + "content": "3.1. SAM Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "SAM is an interactive segmentation model composed of an image encoder, a prompt encoder, and a mask decoder. Given a target image \\( I_{t} \\) and some geometric prompts, SAM first employs a Vision Transformer (ViT) [34] as image encoder to extract image features. Subsequently, the prompt encoder is utilized to generate the prompt embeddings derived from the user-provided points, boxes, or masks. Finally, the mask decoder integrates the image features and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "prompt embeddings, and generates a high-quality mask in a class-agnostic manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.894, + 0.214 + ], + "angle": 0, + "content": "To the best of our knowledge, current research based on SAM still relies on geometric prompts for segmentation. Thus far, there have been no efforts to introduce new forms of prompts to enhance SAM. We design a novel visual reference prompt to extend the visual reference segmentation capabilities of SAM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.704, + 0.24 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Visual reference segmentation aims to segment objects in target image that same semantic as the annotated object in reference image. Specifically, let \\( I_{t} \\) represent the target image and \\( I_{r} \\) represent the reference image. Given a reference image \\( I_{r} \\) and its annotation \\( M_{r}^{i} \\), where \\( i \\) denotes the category of the annotated object, we leverage this information to segment all regions in the target image belonging to category \\( i \\). Depending on the annotation granularity, the reference image have four annotation formats: point, scribble, box, and mask. Point annotation involves providing specific points on the target, scribble requires outlining the target region with arbitrary curves, boxes correspond to bounding boxes around the target, and masks provide pixel-level annotations for the entire target." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.617, + 0.492 + ], + "angle": 0, + "content": "4. VRP-SAM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.638 + ], + "angle": 0, + "content": "VRP-SAM extends SAM to perform visual reference segmentation without compromising its original functionality. We propose a training-efficient visual reference prompt encoder that, firstly, accommodates various granularities of visual references and, secondly, directly encodes these visual references into prompt embeddings rather than geometric prompts. These prompt embeddings are then fed directly into the mask decoder of SAM, resulting in the generation of the target mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.684 + ], + "angle": 0, + "content": "As depicted in Figure 3, the VRP Encoder consists of feature augmenter and prompt generator. Next, we will delve into the details of VRP Encoder and the loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.688, + 0.712 + ], + "angle": 0, + "content": "4.1. Feature Augmenter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Inspired by meta-learning, the Feature Augmenter separately encodes reference annotations \\( M_r^i \\) into features of both the reference and target images. This process is designed to distinguish between foreground and background representations. To capture semantic correlations between reference and target images, we introduce a semantic-aware image encoder within the VRP encoder, encoding them into the same latent space. To prevent overfitting of the VRP encoder, we freeze the image encoder during the training phase. This ensures a balance between capturing semantic relevance and preventing excessive specialization of the VRP encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.891, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.893, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Proposed VRP-SAM framework. Our approach enables SAM to perform visual reference segmentation by extends a VRP encoder. It takes various granularities of visual references as inputs and encodes these visual references into prompt embeddings. Our VRP encoder consists of a feature augmenter and a prompt generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.471, + 0.54 + ], + "angle": 0, + "content": "The Feature Augmenter is illustrated in Figure 3. Initially, leveraging a semantic-aware image encoder (e.g. ResNet-50), we encode \\( I_{r} \\) and \\( I_{t} \\) separately, resulting in the feature map \\( F_{r} \\in \\mathbf{R}^{C \\times H \\times W} \\) and \\( F_{t} \\in \\mathbf{R}^{C \\times H \\times W} \\). Subsequently, we extract the prototype feature \\( P_{i} \\) corresponding to class \\( i \\) from \\( F_{r} \\) using the mask \\( M_{r}^{i} \\). This process can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.55, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\operatorname {M a s k A v g P o o l} \\left(F _ {r}, M _ {r} ^ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.732 + ], + "angle": 0, + "content": "where \\( M_r^i \\) represents one of following annotation formats: point, scribble, box, or mask. To enhance the context information about class \\( i \\), we concatenate the prototype features and mask with \\( F_r \\) and \\( F_t \\). \\( F_r \\) is concatenated with mask \\( m_i \\), and \\( F_t \\) is concatenated with pseudo-mask \\( m_i^{\\text{pseudo}} \\). \\( m_i^{\\text{pseudo}} \\) is obtained using a common training-free approach, and detailed descriptions are provided in the Appendix. Subsequently, we employ a shared \\( 1 \\times 1 \\) convolution layer to reduce dimensionality of enhanced features. This process can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.741, + 0.469, + 0.76 + ], + "angle": 0, + "content": "\\[\nF _ {r} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {r}, P _ {i}, m _ {i}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.763, + 0.469, + 0.782 + ], + "angle": 0, + "content": "\\[\nF _ {t} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {t}, P _ {i}, m _ {i} ^ {\\text {p s e u d o}}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Ultimately, we obtain enhanced image features \\( F_r' \\in \\mathbf{R}^{C \\times H \\times W} \\) and \\( F_t' \\in \\mathbf{R}^{C \\times H \\times W} \\), which have enhanced context information for the category \\( i \\). Thus, the enhanced features comprise foreground representations for class \\( i \\) and background representations for the other classes. Subsequently, we feed the enhanced features into the Prompt Generator to obtain a set of visual reference prompts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.434, + 0.68, + 0.451 + ], + "angle": 0, + "content": "4.2. Prompt Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.578 + ], + "angle": 0, + "content": "The purpose of the Prompt Generator is to obtain a set of visual reference prompts embedding for the SAM mask decoder. As depicted in Figure 3, the process commences with the introduction of a set of learnable queries denoted as \\( Q \\in \\mathbf{R}^{N \\times C} \\), where \\( N \\) indicates the number of visual reference prompts. These queries first engage with the reference features \\( F_r' \\) to obtain the category-specific information through cross-attention and self-attention layer:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.585, + 0.891, + 0.604 + ], + "angle": 0, + "content": "\\[\nQ _ {r} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {1} \\left(\\operatorname {C r o s s A t t n} _ {1} \\left(Q, F _ {r} ^ {\\prime}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Here, we obtain a set of queries, \\( Q_r' \\), possessing knowledge about the object to be segmented. Subsequently, we employ cross-attention to interact these queries with target image feature to obtain the foreground information in target image. Following this, a self-attention layer is used to update the queries, generating a set of prompts \\( Q_t' \\) that align with the representation of SAM:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.723, + 0.891, + 0.743 + ], + "angle": 0, + "content": "\\[\nQ _ {t} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {2} \\left(\\operatorname {C r o s s A t t n} _ {2} \\left(Q _ {r} ^ {\\prime}, F _ {t} ^ {\\prime}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The final \\( Q_{t}^{\\prime} \\) serve as the visual reference prompt embeddings for SAM, equipped with the capability to guide the segmentation of the foreground in the target image. By inputting this set of visual reference prompt embeddings into the mask decoder, the mask \\( M_{t}^{i}\\in \\mathbf{R}^{1\\times H\\times W} \\) for category \\( i \\) in the target image can be obtained." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.645, + 0.862 + ], + "angle": 0, + "content": "4.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We employ Binary Cross-Entropy (BCE) loss and Dice loss to supervise the learning of the Visual Reference Prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23568" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 1. Compare with other foundation models. Results of one-shot semantic segmentation on COCO- \\(20^{i}\\). Gray indicates the model is trained by in-domain datasets. \\(\\dagger\\) indicates the method using SAM." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.149, + 0.47, + 0.296 + ], + "angle": 0, + "content": "
MethodsLabel typeF-0F-1F-2F-3Means
Painter [37]mask.31.235.333.532.433.1
SegGPT [38]56.357.458.951.756.1
PerSAM† [46]23.123.622.023.423.0
PerSAM-F† [46]22.324.023.424.123.5
Masker† [19]52.753.552.652.152.7
VRP-SAM†point.30.139.243.040.438.2
scribble.40.252.052.444.447.2
box.44.549.355.749.149.7
mask.48.155.860.051.653.9
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.3, + 0.47, + 0.345 + ], + "angle": 0, + "content": "Encoder. The BCE loss ensures pixel-wise accuracy, while the Dice loss provides additional context for pixel-level segmentation. Therefore, the total loss of VRP-SAM is:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.346, + 0.47, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {t o t a l}} = \\underbrace {- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log (p _ {i}) + (1 - y _ {i}) \\log (1 - p _ {i}) \\right]} _ {\\mathrm {B C E L o s s}} \\\\ + \\underbrace {1 - \\frac {2 \\sum_ {i = 1} ^ {N} \\left(p _ {i} \\cdot y _ {i}\\right)}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} y _ {i} ^ {2}}} _ {\\text {D i c e L o s s}} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where \\(N\\) represents the total number of pixels, \\(y_{i}\\) is the ground truth label for pixel \\(i\\), and \\(p_i\\) is the predicted probability of pixel \\(i\\) belonging to the object. By combining these two losses, we comprehensively consider both accuracy and contextual effects, thus guiding the Visual Reference Prompt Encoder more effectively in generating precise segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.21, + 0.599 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.17, + 0.623 + ], + "angle": 0, + "content": "5.1. Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Datasets: To validate segmentation performance and generalization capability of VRP-SAM, we conducted extensive experiments following the few-shot setting [14, 31, 44] on COCO-20\\(^i\\) [24] and PASCAL-5\\(^i\\) [29] datasets. Specifically, we organized all classes from both datasets into 4 folds. For each fold, PASCAL-5\\(^i\\) [29] comprises 15 base classes for training and 5 novel classes for testing, while COCO-20\\(^i\\) [24] includes 60 training base classes and 20 testing novel classes. To assess performance of model, we randomly sampled 1000 reference-target pairs in each fold. In each fold, As the mentioned datasets lack labels for point, scribble, and box annotations, we followed the SEEM [48] to generate these annotation labels by simulating user inputs randomly based on the reference ground truth masks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Implementation details: In visual reference prompt encoder, we use VGG-16 [3] and ResNet-50 [9] as the im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.349 + ], + "angle": 0, + "content": "age encoder and initialize it with ImageNet [28] pre-trained weights. We employed the AdamW optimizer [20] along with a cosine learning rate decay strategy for training VRPSAM. Specifically, on the COCO- \\(20^{i}\\) dataset, we conducted 50 epochs of training with an initial learning rate of 1e-4 and a batch size of 8. For the PASCAL- \\(5^{i}\\) dataset, the model was trained for 100 epochs with an initial learning rate of 2e-4 and a batch size of 8. In VRP, the number of queries is set to 50 by default, and the input image size of all experiments needs to be adjusted to \\(512 \\times 512\\). Following SEEM [48], during training, we obtain annotations for points, scribbles, and boxes based on mask annotations. We provide detailed descriptions of this process in the Appendix. To ensure a fair comparison, VRP-SAM is exclusively compared to previous works [16, 26, 37] using visual reference prompts based on the mean intersection over union (mIoU)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.36, + 0.822, + 0.377 + ], + "angle": 0, + "content": "5.2. Comparison with the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Comparison with other foundation models. Leveraging the foundation model enables few-shot segmentation. We compared various approaches utilizing Painter [37], SegGPT [38], and SAM as foundation models for few-shot segmentation, providing evaluation metrics on the COCO-20\\(^i\\) dataset. As shown in Table 1, VRP-SAM achieves \\(53.9\\%\\) mean mIoU without training on novel classes, achieving comparable with SegGPT. Note that the training data of SegGPT include all classes in COCO. Furthermore, our VRP-SAM outperforms other SAM-base methods including Matcher [19], which employs a DINOv2 [25] pretrained ViT-L [34] as image encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Comparison with few-shot methods. To validate the effectiveness of VRP-SAM, we compared it with state-of-the-art few-shot segmentors on the novel set of COCO- \\(20^{i}\\) and PASCAL- \\(5^{i}\\) datasets. To ensure a fair comparison, VRP-SAM is trained and tested separately in each fold, ensuring no overlap in classes between the training and test sets. The results in Table 2 demonstrate that VRP-SAM achieves state-of-the-art results on COCO\\(20^{i}\\) and PASCAL- \\(5^{i}\\), with mIoU scores of 53.9 and 71.9, respectively. Notably, when using VGG-16 as the image encoder for VRP, VRP-SAM achieves mIoU scores of 48.0 and 68.7 on COCO- \\(20^{i}\\) and PASCAL- \\(5^{i}\\) datasets. It indicates that VRP-SAM achieves optimal performance on the novel set with only 1.6M learnable parameters, highlighting its powerful generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.801, + 0.825, + 0.818 + ], + "angle": 0, + "content": "5.3. Comparison with Geometric Prompts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we design Visual Reference Prompts to represent target information. Different from the Geometric Prompts (GP) provided by SAM, our VRP is more flexible and robust. We conducted experiments comparing GP and VRP to validate the superiority of our method (see Table 3)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23569" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 2. Performance of one-shot semantic segmentation on COCO- \\( {20}^{i} \\) and PASCAL- \\( {5}^{i} \\) . The red and blue colors respectively represent the optimal and suboptimal results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.124, + 0.895, + 0.364 + ], + "angle": 0, + "content": "
MethodImage encoderLearnable paramsCOCO-20iPASCAL-5i
F-0F-1F-2F-3MeanF-0F-1F-2F-3Mean
PFENet [33]VGG-1610.4M35.438.136.834.736.356.968.254.552.458.0
BAM [14]4.9M36.447.143.341.742.163.270.866.157.564.4
HDMNet [26]4.2M40.750.648.244.045.964.871.467.756.465.1
VRP-SAM1.6M43.651.750.046.548.070.074.768.361.968.7
PFENet [33]ResNet-5010.4M36.538.634.533.835.861.769.555.456.360.8
HSNet [22]2.6M36.343.138.738.739.264.370.760.360.564.0
CyCTR [44]15.4M38.943.039.639.840.365.771.059.559.764.0
SSP [7]8.7M35.539.637.936.737.460.567.866.451.061.4
NTRENet [18]19.9M36.842.639.937.939.365.472.359.459.864.2
DPCN [17]-42.047.043.339.743.065.771.669.160.666.7
VAT [10]3.2M39.043.842.639.741.367.672.062.360.165.5
BAM [14]4.9M39.449.946.245.245.269.073.667.661.167.8
HDMNet [26]4.2M43.855.351.649.450.071.075.468.962.169.4
VRP-SAM1.6M48.155.860.051.653.973.978.370.665.071.9
DCAMASwin-B47.7M49.552.752.848.750.972.273.864.367.169.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.376, + 0.471, + 0.419 + ], + "angle": 0, + "content": "Table 3. Comparison with geometric prompts. Geometric prompts are randomly sampled from the pseudo-mask. \\(\\dagger\\) indicates the carefully designed sampling strategy proposed in [19]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.43, + 0.466, + 0.624 + ], + "angle": 0, + "content": "
MethodImage EncoderPromptsMean IoU
GP-SAMResNet-50box.19.7
point.21.7
box. + point.23.2
DINOv2box.31.3
point.36.6
box. + point.37.8
box. + point. † [19]52.7
VRP-SAMResNet-50point.38.4
scribble.47.3
box.49.7
mask.53.9
DINOv2-B/14mask.60.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.803 + ], + "angle": 0, + "content": "In the GP-SAM experiments, we initially used an Image Encoder to generate a pseudo-mask for the target image following [33], and subsequently obtained points or bounding boxes from the pseudo-mask as geometric prompts. Experimental results consistently demonstrate the ongoing superiority of our VRP over the GP approach. We visualize the segmentation results of GP-SAM and our VRP-SAM in Figure 4. Visualization results indicate that the GP approach is prone to generating false-positive prompts, significantly impacting segmentation performance. In contrast, our VRP effectively avoids such issues." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.809, + 0.317, + 0.823 + ], + "angle": 0, + "content": "5.4. Generalization Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.833, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Domain shift: Next, we assess the effectiveness of VRPSAM in a domain shift scenario, which necessitates signif" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.376, + 0.596, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.438, + 0.596, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.507, + 0.596, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.576, + 0.596, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.644, + 0.59, + 0.653 + ], + "angle": 0, + "content": "Visual Reference" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.376, + 0.688, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.438, + 0.688, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.507, + 0.688, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.576, + 0.688, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.644, + 0.684, + 0.653 + ], + "angle": 0, + "content": "GP-SAM-RN50" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.376, + 0.778, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.438, + 0.778, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.507, + 0.778, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.576, + 0.778, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.644, + 0.778, + 0.653 + ], + "angle": 0, + "content": "GP-SAM-DINOv2" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.376, + 0.868, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.438, + 0.868, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.507, + 0.868, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.576, + 0.868, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.644, + 0.865, + 0.653 + ], + "angle": 0, + "content": "VRP-SAM-RN50" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.663, + 0.885, + 0.677 + ], + "angle": 0, + "content": "Figure 4. The visualization results of VRP-SAM and GP-SAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.854 + ], + "angle": 0, + "content": "icant domain differences between training and testing sets. Following prior work [22, 32, 33], we trained on COCO- \\(20^{i}\\) and tested on PASCAL- \\(5^{i}\\), where the classes in training set do not overlap with those in test set. The results in Table 4, representing the average across four folds, clearly demonstrate the superior performance of VRP-SAM in domain shift scenario. Notably, VRP-SAM with scribble annotations outperforms FP-Trans [45] by \\(2.8\\%\\), while VRP-SAM with mask annotations surpasses DGPNet [11] by \\(5.8\\%\\). This affirms the robust generalization capability of VRPSAM and its effectiveness in domain transfer scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Visualization: To assess the generalization capability of VRP-SAM across diverse image styles, we curated a collection of target images from web, spanning various gen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2For point prompts, we randomly sample 5 points from the pseudomask. For box prompt, we adopt the bounding box of the pseudo-mask." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23570" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.131, + 0.1, + 0.161 + ], + "angle": 270, + "content": "JIOd" + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.228, + 0.101, + 0.275 + ], + "angle": 270, + "content": "S" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.097, + 0.232, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.096, + 0.364, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.096, + 0.494, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.096, + 0.623, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.096, + 0.752, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.096, + 0.883, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.197, + 0.232, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.197, + 0.365, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.197, + 0.495, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.197, + 0.623, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.197, + 0.752, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.197, + 0.882, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.297, + 0.232, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.297, + 0.365, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.297, + 0.495, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.297, + 0.624, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.297, + 0.752, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.297, + 0.882, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.397, + 0.232, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.497, + 0.184, + 0.508 + ], + "angle": 0, + "content": "VRP" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.397, + 0.365, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.497, + 0.327, + 0.511 + ], + "angle": 0, + "content": "Image-1" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.397, + 0.495, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.497, + 0.458, + 0.509 + ], + "angle": 0, + "content": "Image-2" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.397, + 0.624, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.497, + 0.588, + 0.51 + ], + "angle": 0, + "content": "Image-3" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.397, + 0.752, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.497, + 0.717, + 0.51 + ], + "angle": 0, + "content": "Image-4" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.397, + 0.882, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.497, + 0.853, + 0.509 + ], + "angle": 0, + "content": "Image-5" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.52, + 0.88, + 0.534 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of VRP-SAM across diverse image styles is presented. The target images were sourced from the internet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.708 + ], + "angle": 0, + "content": "res such as natural landscapes, artworks, and complex environments. Visual Reference Prompts (VRPs) were selected from the COCO dataset to guide the segmentation of these target images. The segmentation results on different image styles are presented in Figure 5. It demonstrate that VRP-SAM adeptly adapts to varied image styles, accurately delineating target objects. Notably, VRP-SAM excels in handling ink-style images, showcasing refined segmentation performance. This compellingly underscores the robustness and effectiveness of VRP-SAM when confronted with images of unknown styles." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.23, + 0.728 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.824 + ], + "angle": 0, + "content": "To validate the effectiveness of VRP-SAM, we conducted extensive ablation studies on PASCAL- \\(5^{i}\\). ResNet-50 was chosen as the image encoder for VRP to ensure experimental consistency. These experiments aimed to thoroughly investigate the performance of VRP-SAM under various conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Loss: To assess the impact of Binary Cross-Entropy (BCE) and Dice losses on VRP-SAM, experiments were conducted on the PASCAL- \\(5^{i}\\) dataset. Results in Table 5 indicate comparable performance when using BCE or Dice loss individually. Notably, the optimal performance for" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.54, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Table 4. Evaluation (Mean IoU \\((\\%)\\)) under the domain shift from COCO-20 to PASCAL-5." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.574, + 0.892, + 0.744 + ], + "angle": 0, + "content": "
MethodImage encoderLabel typeMean
RPMM [40]ResNet-50mask.49.6
PFENet [33]61.1
RePRI [4]63.2
VAT-HM [23]65.1
HSNet [22]ResNet-101mask.64.1
DGPNet [11]70.1
FP-Trans [45]DeiT-B/16mask.69.7
VRP-SAMResNet-50point.63.5
scribble.72.5
box.72.3
mask.75.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.859 + ], + "angle": 0, + "content": "VRP-SAM is achieved when both losses are combined. It emphasizes the importance of BCE and Dice losses in guiding VRP-SAM to produce more robust and accurate masks. BCE loss ensures precise pixel-level classification, while Dice loss facilitates accurate spatial localization of objects. Simultaneously employing BCE and Dice losses in VRP-SAM maximizes their complementary advantages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The number of query: In Figure 6, we conducted a comprehensive analysis of the impact of varying query quantities on VRP-SAM's performance. We observed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23571" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.089, + 0.465, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation study on different loss function in VRP-SAM." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.108, + 0.47, + 0.173 + ], + "angle": 0, + "content": "
MethodLabel typeBce lossDice lossMeans
VRP-SAMmask.✓✗70.1
✗✓70.3
✓✓71.9
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.174, + 0.462, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Figure 6. Ablation study on VRP-SAM with different query numbers. The x-axis shows the number of queries, and the y-axis represents model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.572 + ], + "angle": 0, + "content": "a positive correlation between increased query numbers and improved segmentation quality. However, once the query count surpassed 50, the performance gain started to diminish. This phenomenon suggests that 50 queries provide ample effective guidance for the model, and further increases in query count do not yield significant improvements, leading to performance saturation. To maintain high performance while minimizing model learnable parameters, we set the query quantity in VRP-SAM to 50. This decision optimally balances guidance information and model efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.781 + ], + "angle": 0, + "content": "Initialization of query: In Table 6, we compared various query initialization strategies, such as random initialization, foreground prototype (FP), background prototype (BP), and a hybrid prototype with half foreground and half background (half-FP & half-BP). Surprisingly, random initialization outperformed all other strategies, showcasing superior performance. This unexpected outcome can be attributed to the intricate segmentation task and the dataset's object diversity. Random initialization enables the model to explore a broader range of states, avoiding potential local minima. In contrast, prototype-based initializations, whether foreground, background, or a combination, might introduce biases, limiting adaptability to diverse object characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Number of VRPs: In our investigation of the influence of number of visual reference prompts on segmentation results, we explored utilization of few visual reference prompts. This approach involves sending several visual reference prompts to the VRP encoder, generating multiple prompt embeddings. These embeddings are concatenated and forwarded to the mask decoder, resulting in the final mask. The result in Table 7, we observed a substantial en" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 6. Ablation study of different query initialization methods on VRP-SAM." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.121, + 0.872, + 0.198 + ], + "angle": 0, + "content": "
MethodLabel typeImage encoderMean IoU
VRP-SAMmask.FP68.2
BP62.6
half-FP & half-BP67.4
random71.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.203, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Table 7. Ablation Study on Few Visual Reference prompts for VRP-SAM." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.235, + 0.892, + 0.32 + ], + "angle": 0, + "content": "
MethodLabel type1-VRP5-VRP
VRP-SAMpoint.62.964.1(+1.2)
scribble.66.868.4(+1.6)
box.69.470.5(+1.1)
mask.71.972.9(+1.0)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.463 + ], + "angle": 0, + "content": "hancement in segmentation performance with an increase in the number of visual reference prompts. The results emphasize positive influence of incorporating multiple visual reference prompts, enhancing the model's ability to accurately capture intricate details and nuances associated with target object. The diverse prompts contribute enriched contextual information, facilitating a more comprehensive understanding of the object's characteristics and leading to generation of precise and nuanced segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.619, + 0.497 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.824 + ], + "angle": 0, + "content": "In this paper, we present VRP-SAM, an innovative extension of the SAM framework achieved through integration of a Visual Reference Prompt (VRP) encoder. This addition empowers SAM to leverage visual reference prompts for guided segmentation. The core methodology involves encoding annotated reference images through the VRP encoder, which then interacts with the target image to generate meaningful prompts for segmentation within the SAM framework. The seamless fusion of VRP encoder with SAM, resulting in VRP-SAM, enhances the model's generality and adaptability. VRP-SAM overcomes limitations posed by SAM's existing prompt formats, especially in complex scenarios and large datasets. The introduced visual reference prompts, including point, box, scribble, and mask annotations, offer a flexible solution, expanding the model's applicability. Extensive empirical studies conducted showcase VRP-SAM's state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Notably, VRP-SAM demonstrates robust generalization capabilities, excelling in segmentation tasks for novel objects and cross-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgment This work was partially supported by the National Key Research and Development Program of China under Grant 2022ZD0118802 and the National Natural Science Foundation of China (Grant No. U20B2064 and U21B2043)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23572" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mohsen Ahmadi, Ahmad Gholizadeh Lonbar, Abbas Sharifi, Ali Tarlani Beris, Mohammadsadegh Nouri, and Amir Sharifzadeh Javidi. Application of segment anything model for civil infrastructure defect assessment. arXiv preprint arXiv:2304.12600, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In Advances in Neural Information Processing Systems, pages 25005-25017, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Yoshua Bengio and Yann LeCun. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, Wei Liu, et al. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[8] Iraklis Giannakis, Anshuman Bhardwaj, Lydia Sam, and Georgios Leontidis. Deep learning universal crater detection using segment anything model (sam). arXiv preprint arXiv:2304.07764, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[11] Joakim Johnander, Johan Edstedt, Michael Felsberg, Fahad Shahbaz Khan, and Martin Danelljan. Dense gaussian processes for few-shot segmentation. In European Conference on Computer Vision, pages 217-234, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. arXiv preprint arXiv:2306.01567, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE conference on computer vision and pattern recognition, pages 8057-8067, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[15] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[16] Zechao Li, Yanpeng Sun, Liyan Zhang, and Jinhui Tang. Cnt-net: Context-based tandem network for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9904-9917, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.287 + ], + "angle": 0, + "content": "[17] Jie Liu, Yanqi Bao, Guo-Sen Xie, Huan Xiong, Jan-Jakob Sonke, and Efstratios Gavves. Dynamic prototype convolution network for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.355 + ], + "angle": 0, + "content": "[18] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.893, + 0.412 + ], + "angle": 0, + "content": "[19] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[21] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In IEEE International Conference on Computer Vision, pages 8741-8750, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.565 + ], + "angle": 0, + "content": "[22] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In IEEE international conference on computer vision, pages 6941-6952, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.635 + ], + "angle": 0, + "content": "[23] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Hm: Hybrid masking for few-shot segmentation. In European Conference on Computer Vision, pages 506-523, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.893, + 0.678 + ], + "angle": 0, + "content": "[24] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In IEEE International Conference on Computer Vision, pages 622-631, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[25] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.893, + 0.816 + ], + "angle": 0, + "content": "[26] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.873 + ], + "angle": 0, + "content": "[27] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23573" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[29] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[30] Yanpeng Sun and Zechao Li. Ssa: Semantic structure aware inference for weakly pixel-wise dense predictions without cost. arXiv preprint arXiv:2111.03392, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[31] Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, and Jingdong Wang. Singular value fine-tuning: Few-shot segmentation requires few-parameters fine-tuning. In Advances in Neural Information Processing Systems, pages 37484–37496, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[32] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[35] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[36] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In IEEE international conference on computer vision, pages 9197-9206, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[37] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[38] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In IEEE International Conference on Computer Vision, pages 1130-1140, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[39] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[40] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 763-778, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[41] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[42] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with seg-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ment anything models. arXiv preprint arXiv:2306.04121, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[43] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[44] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. In Advances in Neural Information Processing Systems, pages 21984-21996, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[45] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. In Advances in Neural Information Processing Systems, pages 6575–6588, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[47] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[48] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23574" + } + ] +] \ No newline at end of file diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_origin.pdf b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26236bfb5a372dbe33bbb93e7f46ad20b733fa6e --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/7f3ecbdb-6106-4e8e-86ff-9c2733f56191_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef3e52174475d181eb41ecabeed26249ad6a39255d273fb2d9ae64cfb052646a +size 4311283 diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/full.md b/2024/VRP-SAM_ SAM with Visual Reference Prompt/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a305a6e10109cb6a5dae91cf20a09948321afce --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/full.md @@ -0,0 +1,361 @@ +# VRP-SAM: SAM with Visual Reference Prompt + +Yanpeng Sun $^{1,2*}$ , Jiahui Chen $^{2,3*}$ , Shan Zhang $^{4}$ , Xinyu Zhang $^{2}$ , Qiang Chen $^{2}$ , Gang Zhang $^{2}$ , Errui Ding $^{2}$ , Jingdong Wang $^{2}$ , Zechao Li $^{1\dagger}$ + +$^{1}$ Nanjing University of Science and Technology, $^{2}$ Baidu VIS, $^{3}$ Beihang University, $^{4}$ Australian National University {yanpeng.sun, zechao.li}@njust.edu.cn + +# Abstract + +In this paper, we propose a novel Visual Reference Prompt (VRP) encoder that empowers the Segment Anything Model (SAM) to utilize annotated reference images as prompts for segmentation, creating the VRP-SAM model. In essence, VRP-SAM can utilize annotated reference images to comprehend specific objects and perform segmentation of specific objects in target image. It is note that the VRP encoder can support a variety of annotation formats for reference images, including point, box, scribble, and mask. VRP-SAM achieves a breakthrough within the SAM framework by extending its versatility and applicability while preserving SAM's inherent strengths, thus enhancing user-friendliness. To enhance the generalization ability of VRP-SAM, the VRP encoder adopts a meta-learning strategy. To validate the effectiveness of VRP-SAM, we conducted extensive empirical studies on the Pascal and COCO datasets. Remarkably, VRP-SAM achieved state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Furthermore, VRP-SAM demonstrates strong generalization capabilities, allowing it to perform segmentation of unseen objects and enabling cross-domain segmentation. The source code and models will be available at https://github.com/syp2ysy/VRPSAM + +# 1. Introduction + +In recent, the Segment Anything Model (SAM) [13] has emerged as a foundational visual model for image segmentation. Trained on an extensive datasets comprising billions of labels, SAM has demonstrated remarkable versatility in the realm of universal segmentation. What sets SAM apart is its human-interactive design, allowing segmentation based on user-provided prompts, be they in the form of + +![](images/279cd4ad8789425d21e414ca504cc98936cbc938093da7b046c4f32cfcb431d9.jpg) +Figure 1. Comparison of SAM's built-in Point and Box prompt modes with Visual Reference Prompt when handling numerous images. The prompts are all provided by users. + +points, bounding boxes, or coarse masks. This distinctive feature positions SAM as a robust tool that can be adaptable to various tasks and requirements [12, 48]. + +However, the existing prompt formats of SAM present significant challenges in practical applications, especially when dealing with complex scenes and numerous images. As shown in Figure 2(a), SAM relies on user-provided prompts (points, boxes, coarse mask) to segment objects in the target image, demanding users to possess a comprehensive understanding of the target objects. In real-world applications, especially in complex scenarios, the level of user familiarity with the target objects can significantly impact the effectiveness of providing specific prompts. Furthermore, variations in the position, size, and quantity of target objects across different images require custom prompts for each image. As illustrated in Figure 1 with the aim of segmenting 'bicycles', users need to customize different prompts for each image, significantly impacting efficiency of SAM. Therefore, we propose integrating visual reference prompts to overcome these limitations and enhance adaptability of SAM. + +Visual reference prompts is annotated reference images that delineate the objects users expect to segment. As shown in Figure 1, by simply providing a visual reference prompt + +![](images/692023180000c933416f24d6830d1c26a4a0fdd99fb2f92e00cf025279a266c2.jpg) + +![](images/ef0ce84e843f6b8e08e4a28818863278170bc594756302651c4ae41d7f47f21a.jpg) +Figure 2. A Comparison between SAM and VRP-SAM. VRP-SAM introduces a visual reference prompt encoder, accepting annotated reference images with point, scribble, box, and mask formats, offering a distinct enhancement over SAM. + +with bicycle, we can segment bicycle in different images without requiring users to provide specific prompts for each image. It significantly enhances the efficiency of SAM while reducing reliance on user familiarity with objects. To achieve this goal, some methods [19, 46] incorporate semantic correlation models [9, 25] to establish reference-target correlation1, obtaining pseudo-masks for the target objects. Following this, a sampling strategy is devised to extract a set of points and bounding boxes from the pseudomasks, serving as prompts for SAM segment the target image. These methods overlook false positives within the pseudo-mask and exhibit high sensitivity to hyperparameters. Consequently, it heavily relies on the quality of the pseudo-mask and has poor generalization. + +Toward this end, we propose a straightforward and effective Visual Reference Prompt (VRP) encoder using meta-learning technique, integrated with SAM to create VRPSAM. It leverages annotated reference images as prompts to segment similar semantic objects in the target image. As illustrated in Figure 2(b), the VRP encoder accepts inputs in various annotation formats, including points, scribbles, boxes, and masks. Specifically, the VRP encoder introduces a semantic-related model to encode reference and target images into the same space. Following meta-learning methods, we first extract prototypes of target objects from annotated information in reference images, enhancing the representation of the target objects in both images. Following meta-learning methods, the prototypes of target objects (users' markers) are first generated from annotated reference images, which aim to highlight such target instances + +in both reference and target images. Then, we introduce a set of learnable queries to extract the semantic cues of the target objects from attentive enhanced reference features. Then, these queries interact with target images, generating prompt embeddings usable by mask decoder to segment semantically specific objects in the target image. Building upon the foundation of SAM's inherent capabilities, VRPSAM enhances the model's visual reference segmentation prowess. The introduction of Visual Reference Prompts (VRP) not only diversifies the prompts, enabling the model to swiftly segment objects with identical semantics, but also incorporates a meta-learning mechanism that significantly boosts the generalization of model, particularly in dealing with novel objects and cross-domain scenarios. + +To quantitatively assess the generalization capability of VRP-SAM, we follow the dataset configurations commonly used in few-shot segmentation and evaluate our VRP-SAM on Pascal- $5^{i}$ and COCO-20i datasets. Extensive experiments show that VRP-SAM overcomes the limitations of SAM in prompt format, enabling efficient visual reference segmentation. It is significantly better than sampling-based methods, achieving state-of-the-art results on Pascal- $5^{i}$ and COCO-20i datasets. Moreover, we present solid evidence of VRP-SAM's superior performance in handling unknown objects and cross-domain scenarios. Our experiments highlight that VRP-SAM achieves the rapid segmentation of a large number of images based on semantics. Furthermore, the incorporation of meta-learning principles significantly enhances its generalization, making it applicable across various scenarios. + +# 2. Related Work + +# 2.1. Application of SAM + +Segmentation Anything model (SAM) [13] is a recently introduced category-agnostic interactive segmentation model by Meta. It leverages user-guided instructions for segmentation, such as points, bounding boxes, and coarse masks. Currently, SAM has two primary application approaches. One involves using SAM's segmentation results as prior information to assist downstream tasks. For instance, Inpaint Anything (IA) [43] and Edit Everything [39] leverage SAM's segmentation results for image editing and restoration [42, 47] within the masked regions. Additionally, SEPL [5] employs SAM's segmentation results to enhance pseudo-labels in weakly supervised segmentation tasks [27, 30]. Furthermore, SAM segmentation results serve as prior information in tasks such as crack and volcano crater detection [1, 8]. + +The second involves guiding SAM's segmentation through various prompt combinations. For example, in the context of visual reference segmentation [31], Matcher [19] samples points and boxes from pseudo-masks, utilizing + +SAM to refine these pseudo-masks. TAM [41] is applied in object tracking tasks, where it uses point prompts to initialize object masks and subsequently uses SAM to refine low-quality masks. Additionally, SAMAug [6] introduces a visual point enhancement method tailored for SAM, facilitating automatic image annotation. These examples illustrate SAM's effectiveness across different tasks. However, existing methods are constrained by SAM's existing prompt modalities, and they may struggle when confronted with complex objects and unfamiliar scenes. To break these limitations, we have designed a visual prompt encoder for SAM, expanding its applicability to a wider range of scenarios. + +# 2.2. Visual reference segmentation + +Visual reference segmentation [31, 33, 48] aims to guide the segmentation of a target image using a reference image. The goal of this task is to utilize a semantically annotated reference image to instruct the segmentation of objects or regions in the target image that share the same semantics as those in the reference image. In current research, methods can be broadly categorized into two groups: prototype-based and feature-matching-based. Prototype-based methods, such as PFENet [33], PANet [36], and CWT [21], usually focus on distinguishing prototypes with different class-specific features. ASGNet [15], on the other hand, improves the segmentation performance by increasing the number of prototypes. Another feature-matching approach [22, 35] leverages the pixel-level correlations between reference and target images to significantly enhance segmentation performance, as demonstrated by methods like CyCTR [44] and HDMNet [26]. Moreover, modern large-scale vision models [2, 37, 48] have recognized visual reference segmentation as a primary task, given its indispensable role in handling complex objects and unknown scenes. However, it's important to note that SAM [13] does not possess the capability to perform this task, underscoring the necessity of introducing VPR-SAM. + +# 3. Preliminary + +In this section, we first review the architecture of Segment Anything Model. Then we formulate the problem setting in this paper. + +# 3.1. SAM Architecture + +SAM is an interactive segmentation model composed of an image encoder, a prompt encoder, and a mask decoder. Given a target image $I_{t}$ and some geometric prompts, SAM first employs a Vision Transformer (ViT) [34] as image encoder to extract image features. Subsequently, the prompt encoder is utilized to generate the prompt embeddings derived from the user-provided points, boxes, or masks. Finally, the mask decoder integrates the image features and + +prompt embeddings, and generates a high-quality mask in a class-agnostic manner. + +To the best of our knowledge, current research based on SAM still relies on geometric prompts for segmentation. Thus far, there have been no efforts to introduce new forms of prompts to enhance SAM. We design a novel visual reference prompt to extend the visual reference segmentation capabilities of SAM. + +# 3.2. Problem Formulation + +Visual reference segmentation aims to segment objects in target image that same semantic as the annotated object in reference image. Specifically, let $I_{t}$ represent the target image and $I_{r}$ represent the reference image. Given a reference image $I_{r}$ and its annotation $M_{r}^{i}$ , where $i$ denotes the category of the annotated object, we leverage this information to segment all regions in the target image belonging to category $i$ . Depending on the annotation granularity, the reference image have four annotation formats: point, scribble, box, and mask. Point annotation involves providing specific points on the target, scribble requires outlining the target region with arbitrary curves, boxes correspond to bounding boxes around the target, and masks provide pixel-level annotations for the entire target. + +# 4. VRP-SAM + +VRP-SAM extends SAM to perform visual reference segmentation without compromising its original functionality. We propose a training-efficient visual reference prompt encoder that, firstly, accommodates various granularities of visual references and, secondly, directly encodes these visual references into prompt embeddings rather than geometric prompts. These prompt embeddings are then fed directly into the mask decoder of SAM, resulting in the generation of the target mask. + +As depicted in Figure 3, the VRP Encoder consists of feature augmenter and prompt generator. Next, we will delve into the details of VRP Encoder and the loss function. + +# 4.1. Feature Augmenter + +Inspired by meta-learning, the Feature Augmenter separately encodes reference annotations $M_r^i$ into features of both the reference and target images. This process is designed to distinguish between foreground and background representations. To capture semantic correlations between reference and target images, we introduce a semantic-aware image encoder within the VRP encoder, encoding them into the same latent space. To prevent overfitting of the VRP encoder, we freeze the image encoder during the training phase. This ensures a balance between capturing semantic relevance and preventing excessive specialization of the VRP encoder. + +![](images/1200482160451724d569485715c0f72ffa0d9b6becd63a62d4df9c425cdcf670.jpg) +Figure 3. Proposed VRP-SAM framework. Our approach enables SAM to perform visual reference segmentation by extends a VRP encoder. It takes various granularities of visual references as inputs and encodes these visual references into prompt embeddings. Our VRP encoder consists of a feature augmenter and a prompt generator. + +The Feature Augmenter is illustrated in Figure 3. Initially, leveraging a semantic-aware image encoder (e.g. ResNet-50), we encode $I_{r}$ and $I_{t}$ separately, resulting in the feature map $F_{r} \in \mathbf{R}^{C \times H \times W}$ and $F_{t} \in \mathbf{R}^{C \times H \times W}$ . Subsequently, we extract the prototype feature $P_{i}$ corresponding to class $i$ from $F_{r}$ using the mask $M_{r}^{i}$ . This process can be summarized as follows: + +$$ +P _ {i} = \operatorname {M a s k A v g P o o l} \left(F _ {r}, M _ {r} ^ {i}\right) \tag {1} +$$ + +where $M_r^i$ represents one of following annotation formats: point, scribble, box, or mask. To enhance the context information about class $i$ , we concatenate the prototype features and mask with $F_r$ and $F_t$ . $F_r$ is concatenated with mask $m_i$ , and $F_t$ is concatenated with pseudo-mask $m_i^{\text{pseudo}}$ . $m_i^{\text{pseudo}}$ is obtained using a common training-free approach, and detailed descriptions are provided in the Appendix. Subsequently, we employ a shared $1 \times 1$ convolution layer to reduce dimensionality of enhanced features. This process can be summarized as follows: + +$$ +F _ {r} ^ {\prime} = \operatorname {C o n v} \left(\operatorname {c o n c a t} \left(F _ {r}, P _ {i}, m _ {i}\right)\right) \tag {2} +$$ + +$$ +F _ {t} ^ {\prime} = \operatorname {C o n v} \left(\operatorname {c o n c a t} \left(F _ {t}, P _ {i}, m _ {i} ^ {\text {p s e u d o}}\right)\right) \tag {3} +$$ + +Ultimately, we obtain enhanced image features $F_r' \in \mathbf{R}^{C \times H \times W}$ and $F_t' \in \mathbf{R}^{C \times H \times W}$ , which have enhanced context information for the category $i$ . Thus, the enhanced features comprise foreground representations for class $i$ and background representations for the other classes. Subsequently, we feed the enhanced features into the Prompt Generator to obtain a set of visual reference prompts. + +# 4.2. Prompt Generator + +The purpose of the Prompt Generator is to obtain a set of visual reference prompts embedding for the SAM mask decoder. As depicted in Figure 3, the process commences with the introduction of a set of learnable queries denoted as $Q \in \mathbf{R}^{N \times C}$ , where $N$ indicates the number of visual reference prompts. These queries first engage with the reference features $F_r'$ to obtain the category-specific information through cross-attention and self-attention layer: + +$$ +Q _ {r} ^ {\prime} = \operatorname {S e l f A t t n} _ {1} \left(\operatorname {C r o s s A t t n} _ {1} \left(Q, F _ {r} ^ {\prime}\right)\right) \tag {4} +$$ + +Here, we obtain a set of queries, $Q_r'$ , possessing knowledge about the object to be segmented. Subsequently, we employ cross-attention to interact these queries with target image feature to obtain the foreground information in target image. Following this, a self-attention layer is used to update the queries, generating a set of prompts $Q_t'$ that align with the representation of SAM: + +$$ +Q _ {t} ^ {\prime} = \operatorname {S e l f A t t n} _ {2} \left(\operatorname {C r o s s A t t n} _ {2} \left(Q _ {r} ^ {\prime}, F _ {t} ^ {\prime}\right)\right) \tag {5} +$$ + +The final $Q_{t}^{\prime}$ serve as the visual reference prompt embeddings for SAM, equipped with the capability to guide the segmentation of the foreground in the target image. By inputting this set of visual reference prompt embeddings into the mask decoder, the mask $M_{t}^{i}\in \mathbf{R}^{1\times H\times W}$ for category $i$ in the target image can be obtained. + +# 4.3. Loss Function + +We employ Binary Cross-Entropy (BCE) loss and Dice loss to supervise the learning of the Visual Reference Prompt + +Table 1. Compare with other foundation models. Results of one-shot semantic segmentation on COCO- $20^{i}$ . Gray indicates the model is trained by in-domain datasets. $\dagger$ indicates the method using SAM. + +
MethodsLabel typeF-0F-1F-2F-3Means
Painter [37]mask.31.235.333.532.433.1
SegGPT [38]56.357.458.951.756.1
PerSAM† [46]23.123.622.023.423.0
PerSAM-F† [46]22.324.023.424.123.5
Masker† [19]52.753.552.652.152.7
VRP-SAM†point.30.139.243.040.438.2
scribble.40.252.052.444.447.2
box.44.549.355.749.149.7
mask.48.155.860.051.653.9
+ +Encoder. The BCE loss ensures pixel-wise accuracy, while the Dice loss provides additional context for pixel-level segmentation. Therefore, the total loss of VRP-SAM is: + +$$ +\begin{array}{l} L _ {\mathrm {t o t a l}} = \underbrace {- \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ y _ {i} \log (p _ {i}) + (1 - y _ {i}) \log (1 - p _ {i}) \right]} _ {\mathrm {B C E L o s s}} \\ + \underbrace {1 - \frac {2 \sum_ {i = 1} ^ {N} \left(p _ {i} \cdot y _ {i}\right)}{\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \sum_ {i = 1} ^ {N} y _ {i} ^ {2}}} _ {\text {D i c e L o s s}} \tag {6} \\ \end{array} +$$ + +where $N$ represents the total number of pixels, $y_{i}$ is the ground truth label for pixel $i$ , and $p_i$ is the predicted probability of pixel $i$ belonging to the object. By combining these two losses, we comprehensively consider both accuracy and contextual effects, thus guiding the Visual Reference Prompt Encoder more effectively in generating precise segmentation results. + +# 5. Experiments + +# 5.1. Setting + +Datasets: To validate segmentation performance and generalization capability of VRP-SAM, we conducted extensive experiments following the few-shot setting [14, 31, 44] on COCO-20 $^i$ [24] and PASCAL-5 $^i$ [29] datasets. Specifically, we organized all classes from both datasets into 4 folds. For each fold, PASCAL-5 $^i$ [29] comprises 15 base classes for training and 5 novel classes for testing, while COCO-20 $^i$ [24] includes 60 training base classes and 20 testing novel classes. To assess performance of model, we randomly sampled 1000 reference-target pairs in each fold. In each fold, As the mentioned datasets lack labels for point, scribble, and box annotations, we followed the SEEM [48] to generate these annotation labels by simulating user inputs randomly based on the reference ground truth masks. + +Implementation details: In visual reference prompt encoder, we use VGG-16 [3] and ResNet-50 [9] as the im + +age encoder and initialize it with ImageNet [28] pre-trained weights. We employed the AdamW optimizer [20] along with a cosine learning rate decay strategy for training VRPSAM. Specifically, on the COCO- $20^{i}$ dataset, we conducted 50 epochs of training with an initial learning rate of 1e-4 and a batch size of 8. For the PASCAL- $5^{i}$ dataset, the model was trained for 100 epochs with an initial learning rate of 2e-4 and a batch size of 8. In VRP, the number of queries is set to 50 by default, and the input image size of all experiments needs to be adjusted to $512 \times 512$ . Following SEEM [48], during training, we obtain annotations for points, scribbles, and boxes based on mask annotations. We provide detailed descriptions of this process in the Appendix. To ensure a fair comparison, VRP-SAM is exclusively compared to previous works [16, 26, 37] using visual reference prompts based on the mean intersection over union (mIoU). + +# 5.2. Comparison with the State-of-the-art + +Comparison with other foundation models. Leveraging the foundation model enables few-shot segmentation. We compared various approaches utilizing Painter [37], SegGPT [38], and SAM as foundation models for few-shot segmentation, providing evaluation metrics on the COCO-20 $^i$ dataset. As shown in Table 1, VRP-SAM achieves $53.9\%$ mean mIoU without training on novel classes, achieving comparable with SegGPT. Note that the training data of SegGPT include all classes in COCO. Furthermore, our VRP-SAM outperforms other SAM-base methods including Matcher [19], which employs a DINOv2 [25] pretrained ViT-L [34] as image encoder. + +Comparison with few-shot methods. To validate the effectiveness of VRP-SAM, we compared it with state-of-the-art few-shot segmentors on the novel set of COCO- $20^{i}$ and PASCAL- $5^{i}$ datasets. To ensure a fair comparison, VRP-SAM is trained and tested separately in each fold, ensuring no overlap in classes between the training and test sets. The results in Table 2 demonstrate that VRP-SAM achieves state-of-the-art results on COCO $20^{i}$ and PASCAL- $5^{i}$ , with mIoU scores of 53.9 and 71.9, respectively. Notably, when using VGG-16 as the image encoder for VRP, VRP-SAM achieves mIoU scores of 48.0 and 68.7 on COCO- $20^{i}$ and PASCAL- $5^{i}$ datasets. It indicates that VRP-SAM achieves optimal performance on the novel set with only 1.6M learnable parameters, highlighting its powerful generalization capability. + +# 5.3. Comparison with Geometric Prompts + +In this paper, we design Visual Reference Prompts to represent target information. Different from the Geometric Prompts (GP) provided by SAM, our VRP is more flexible and robust. We conducted experiments comparing GP and VRP to validate the superiority of our method (see Table 3). + +Table 2. Performance of one-shot semantic segmentation on COCO- ${20}^{i}$ and PASCAL- ${5}^{i}$ . The red and blue colors respectively represent the optimal and suboptimal results. + +
MethodImage encoderLearnable paramsCOCO-20iPASCAL-5i
F-0F-1F-2F-3MeanF-0F-1F-2F-3Mean
PFENet [33]VGG-1610.4M35.438.136.834.736.356.968.254.552.458.0
BAM [14]4.9M36.447.143.341.742.163.270.866.157.564.4
HDMNet [26]4.2M40.750.648.244.045.964.871.467.756.465.1
VRP-SAM1.6M43.651.750.046.548.070.074.768.361.968.7
PFENet [33]ResNet-5010.4M36.538.634.533.835.861.769.555.456.360.8
HSNet [22]2.6M36.343.138.738.739.264.370.760.360.564.0
CyCTR [44]15.4M38.943.039.639.840.365.771.059.559.764.0
SSP [7]8.7M35.539.637.936.737.460.567.866.451.061.4
NTRENet [18]19.9M36.842.639.937.939.365.472.359.459.864.2
DPCN [17]-42.047.043.339.743.065.771.669.160.666.7
VAT [10]3.2M39.043.842.639.741.367.672.062.360.165.5
BAM [14]4.9M39.449.946.245.245.269.073.667.661.167.8
HDMNet [26]4.2M43.855.351.649.450.071.075.468.962.169.4
VRP-SAM1.6M48.155.860.051.653.973.978.370.665.071.9
DCAMASwin-B47.7M49.552.752.848.750.972.273.864.367.169.3
+ +Table 3. Comparison with geometric prompts. Geometric prompts are randomly sampled from the pseudo-mask. $\dagger$ indicates the carefully designed sampling strategy proposed in [19]. + +
MethodImage EncoderPromptsMean IoU
GP-SAMResNet-50box.19.7
point.21.7
box. + point.23.2
DINOv2box.31.3
point.36.6
box. + point.37.8
box. + point. † [19]52.7
VRP-SAMResNet-50point.38.4
scribble.47.3
box.49.7
mask.53.9
DINOv2-B/14mask.60.4
+ +In the GP-SAM experiments, we initially used an Image Encoder to generate a pseudo-mask for the target image following [33], and subsequently obtained points or bounding boxes from the pseudo-mask as geometric prompts. Experimental results consistently demonstrate the ongoing superiority of our VRP over the GP approach. We visualize the segmentation results of GP-SAM and our VRP-SAM in Figure 4. Visualization results indicate that the GP approach is prone to generating false-positive prompts, significantly impacting segmentation performance. In contrast, our VRP effectively avoids such issues. + +# 5.4. Generalization Evaluation + +Domain shift: Next, we assess the effectiveness of VRPSAM in a domain shift scenario, which necessitates signif + +![](images/bcd4cb3092239f913c25dc0beb955a87c28b8944125d4e25fefa2acc636beef1.jpg) + +![](images/3a0d83cd3fd737170c45a218170bf81f94e28c0a14dd2af163c23b80cad60d12.jpg) + +![](images/49c984490b6632e211a9d199f10b1b7321ac142d499abe111d699a47740d64d2.jpg) + +![](images/246eaed965df90310858d4cf1967a3a6e7f460df87435f1af853c772be1834d3.jpg) +Visual Reference + +![](images/2288c72273e4b217ab76c97dd49b7e950f62ef71d613960ee04e39fc1b411d77.jpg) + +![](images/fd1334a4f875164eeacc12e1e6f4ffb643bcb14ad493cd1d25a678f8c896023f.jpg) + +![](images/f837fc9d7de39914eae01792ad01bd68fad350c47af3191eab0a7c134ea314ac.jpg) + +![](images/14d799081ecd8daa077f5395ffff76aa2f3495feba55114608424e0d453ea8d1.jpg) +GP-SAM-RN50 + +![](images/d87dd1cd65efd2c715a7b6a14a93e7e2bea2730b863dff0b68c6f2a647481f31.jpg) + +![](images/8ab4534d999a15ae392d8fd7e3d56e417612e99c6db0af8b116a71a95160155a.jpg) + +![](images/f51aeb8de1305a24306bbb087480d778617e39b1b2ec5e4c675483ff9c75c7a4.jpg) + +![](images/9863e7291521b2a40cd51d5f602289c1c5e7583c0556573922b37afb26225bcc.jpg) +GP-SAM-DINOv2 + +![](images/222840ba44ce1080c1f40bdd2a15f0d8b071d015da2202b495587cc685e10132.jpg) + +![](images/0ab87febacbd1c910b62f682511d6376d21560beb44ce5fbc1e0623ae533e729.jpg) + +![](images/d8af1b9e16ae86437828b11c40198bbf1f5e11892a1189fa480e30992f8244a8.jpg) + +![](images/f29b352fed25b2d46515260f7b3151c14a8c046bf2f1bc0fae1089afbf9a0a40.jpg) +VRP-SAM-RN50 +Figure 4. The visualization results of VRP-SAM and GP-SAM. + +icant domain differences between training and testing sets. Following prior work [22, 32, 33], we trained on COCO- $20^{i}$ and tested on PASCAL- $5^{i}$ , where the classes in training set do not overlap with those in test set. The results in Table 4, representing the average across four folds, clearly demonstrate the superior performance of VRP-SAM in domain shift scenario. Notably, VRP-SAM with scribble annotations outperforms FP-Trans [45] by $2.8\%$ , while VRP-SAM with mask annotations surpasses DGPNet [11] by $5.8\%$ . This affirms the robust generalization capability of VRPSAM and its effectiveness in domain transfer scenarios. + +Visualization: To assess the generalization capability of VRP-SAM across diverse image styles, we curated a collection of target images from web, spanning various gen + +![](images/80652b85f5f7f83b00f5695e5d600e4dec217ca35099b5b37bd926294b0b86a6.jpg) +JIOd + +![](images/8b81ba2695a92d5437bc87437341b00615ac7e52e28bc315e638a7d8ac62d7ed.jpg) + +![](images/6dd71cd829e2a5a4bfa435ff0fe8c09bff84b7a2a7d6fdf159835edac1018225.jpg) + +![](images/4a897af2669cd10a89d0452c5ffdce5118bbaf97be580f7df978177f3a12b1b1.jpg) + +![](images/5f3c329154d485d4af9bb1f4e12566bc038ce20c73df509cdcbeda1c66db2948.jpg) + +![](images/27639975f24ff2b816b35af3ca2e62ff8266a3f3e6975aa46c16991db4e20b42.jpg) + +![](images/1fd91bc4d07435e0cfd07d65b7a5a09c16a055212e6eda90ccbb5eae5132576d.jpg) +S + +![](images/b61d50a7148940ddf68cedf883129879d89f3974b3be04cad5320869a33f6a9b.jpg) + +![](images/7a99809c44c45c18a4b57e4d836772b166f1059003f811ad36f86675f18fa32d.jpg) + +![](images/42af8c625ce8c59d0360b6b381a3fd89198a1415bffb6467ff94e1294a7d192f.jpg) + +![](images/2c3859036847fc6cc75094782a41f91a92fc5095e47049da6ba3be65dec2876b.jpg) + +![](images/9e464d08743027204ca2b7f7ab3ae33c6f4c8ce1580f3edb577f5a21c1a4fa0f.jpg) + +![](images/e8b7fc657345272017af5182b6eafe5846e20ab226ba50c0bf3d0be193a252b9.jpg) + +![](images/a442cb22db880e4eee1a97386ed0c728eb9c3facdc41c546fd62755176eb68e0.jpg) + +![](images/e8dbffd86720765f7e4682f4b5b7790a3d8be1406a1da46d2a0b7452beba61ce.jpg) + +![](images/4532c83c7bf26e7647ec2cafe03f9a3d13f7bce3e09af07116064478af3872a9.jpg) + +![](images/1f3dc69ecfeb8643dd5b93149f7815d24a81db00884253607dd439a7760c9921.jpg) + +![](images/06b324a7d31b1b11a80b04ca0fb60368552c3075388330ba3766a8dcbfe8d161.jpg) + +![](images/9e041444f742e32b85f737a51fedf0574c170b5a7e296c2dffdc4a1ed3ce7ff0.jpg) +VRP +Figure 5. Qualitative results of VRP-SAM across diverse image styles is presented. The target images were sourced from the internet. + +![](images/2b08169c45522b6f22eff1e42f3dd26e7fff2bb541e0d38ce5c48e1791e9bb41.jpg) +Image-1 + +![](images/38c2788b42e327737dec89980469bfb142bdbfe9772264927362bec3ea1cbd69.jpg) +Image-2 + +![](images/6a4017350ddf98c60aa16176e6a44a95491249c073531cf151254f5dc156839d.jpg) +Image-3 + +![](images/8aa470003623f34b70d56547105e222cf8591414cca5656ac2af956f5207d295.jpg) +Image-4 + +![](images/d7f5f00a4c2bd6bd2ff7157cfcb467fe8b2547563e9d55a1f23fc01c2f305e5c.jpg) +Image-5 + +res such as natural landscapes, artworks, and complex environments. Visual Reference Prompts (VRPs) were selected from the COCO dataset to guide the segmentation of these target images. The segmentation results on different image styles are presented in Figure 5. It demonstrate that VRP-SAM adeptly adapts to varied image styles, accurately delineating target objects. Notably, VRP-SAM excels in handling ink-style images, showcasing refined segmentation performance. This compellingly underscores the robustness and effectiveness of VRP-SAM when confronted with images of unknown styles. + +# 5.5. Ablation Study + +To validate the effectiveness of VRP-SAM, we conducted extensive ablation studies on PASCAL- $5^{i}$ . ResNet-50 was chosen as the image encoder for VRP to ensure experimental consistency. These experiments aimed to thoroughly investigate the performance of VRP-SAM under various conditions. + +Loss: To assess the impact of Binary Cross-Entropy (BCE) and Dice losses on VRP-SAM, experiments were conducted on the PASCAL- $5^{i}$ dataset. Results in Table 5 indicate comparable performance when using BCE or Dice loss individually. Notably, the optimal performance for + +Table 4. Evaluation (Mean IoU $(\%)$ ) under the domain shift from COCO-20 to PASCAL-5. + +
MethodImage encoderLabel typeMean
RPMM [40]ResNet-50mask.49.6
PFENet [33]61.1
RePRI [4]63.2
VAT-HM [23]65.1
HSNet [22]ResNet-101mask.64.1
DGPNet [11]70.1
FP-Trans [45]DeiT-B/16mask.69.7
VRP-SAMResNet-50point.63.5
scribble.72.5
box.72.3
mask.75.9
+ +VRP-SAM is achieved when both losses are combined. It emphasizes the importance of BCE and Dice losses in guiding VRP-SAM to produce more robust and accurate masks. BCE loss ensures precise pixel-level classification, while Dice loss facilitates accurate spatial localization of objects. Simultaneously employing BCE and Dice losses in VRP-SAM maximizes their complementary advantages. + +The number of query: In Figure 6, we conducted a comprehensive analysis of the impact of varying query quantities on VRP-SAM's performance. We observed + +Table 5. Ablation study on different loss function in VRP-SAM. + +
MethodLabel typeBce lossDice lossMeans
VRP-SAMmask.✓✗70.1
✗✓70.3
✓✓71.9
+ +![](images/b1d89d230e8993a88026585303bcab175c30f91996180c870bd4a2ff8d009941.jpg) +Figure 6. Ablation study on VRP-SAM with different query numbers. The x-axis shows the number of queries, and the y-axis represents model performance. + +a positive correlation between increased query numbers and improved segmentation quality. However, once the query count surpassed 50, the performance gain started to diminish. This phenomenon suggests that 50 queries provide ample effective guidance for the model, and further increases in query count do not yield significant improvements, leading to performance saturation. To maintain high performance while minimizing model learnable parameters, we set the query quantity in VRP-SAM to 50. This decision optimally balances guidance information and model efficiency. + +Initialization of query: In Table 6, we compared various query initialization strategies, such as random initialization, foreground prototype (FP), background prototype (BP), and a hybrid prototype with half foreground and half background (half-FP & half-BP). Surprisingly, random initialization outperformed all other strategies, showcasing superior performance. This unexpected outcome can be attributed to the intricate segmentation task and the dataset's object diversity. Random initialization enables the model to explore a broader range of states, avoiding potential local minima. In contrast, prototype-based initializations, whether foreground, background, or a combination, might introduce biases, limiting adaptability to diverse object characteristics. + +Number of VRPs: In our investigation of the influence of number of visual reference prompts on segmentation results, we explored utilization of few visual reference prompts. This approach involves sending several visual reference prompts to the VRP encoder, generating multiple prompt embeddings. These embeddings are concatenated and forwarded to the mask decoder, resulting in the final mask. The result in Table 7, we observed a substantial en + +Table 6. Ablation study of different query initialization methods on VRP-SAM. + +
MethodLabel typeImage encoderMean IoU
VRP-SAMmask.FP68.2
BP62.6
half-FP & half-BP67.4
random71.9
+ +Table 7. Ablation Study on Few Visual Reference prompts for VRP-SAM. + +
MethodLabel type1-VRP5-VRP
VRP-SAMpoint.62.964.1(+1.2)
scribble.66.868.4(+1.6)
box.69.470.5(+1.1)
mask.71.972.9(+1.0)
+ +hancement in segmentation performance with an increase in the number of visual reference prompts. The results emphasize positive influence of incorporating multiple visual reference prompts, enhancing the model's ability to accurately capture intricate details and nuances associated with target object. The diverse prompts contribute enriched contextual information, facilitating a more comprehensive understanding of the object's characteristics and leading to generation of precise and nuanced segmentation masks. + +# 6. Conclusion + +In this paper, we present VRP-SAM, an innovative extension of the SAM framework achieved through integration of a Visual Reference Prompt (VRP) encoder. This addition empowers SAM to leverage visual reference prompts for guided segmentation. The core methodology involves encoding annotated reference images through the VRP encoder, which then interacts with the target image to generate meaningful prompts for segmentation within the SAM framework. The seamless fusion of VRP encoder with SAM, resulting in VRP-SAM, enhances the model's generality and adaptability. VRP-SAM overcomes limitations posed by SAM's existing prompt formats, especially in complex scenarios and large datasets. The introduced visual reference prompts, including point, box, scribble, and mask annotations, offer a flexible solution, expanding the model's applicability. Extensive empirical studies conducted showcase VRP-SAM's state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Notably, VRP-SAM demonstrates robust generalization capabilities, excelling in segmentation tasks for novel objects and cross-domain scenarios. + +Acknowledgment This work was partially supported by the National Key Research and Development Program of China under Grant 2022ZD0118802 and the National Natural Science Foundation of China (Grant No. U20B2064 and U21B2043). + +# References + +[1] Mohsen Ahmadi, Ahmad Gholizadeh Lonbar, Abbas Sharifi, Ali Tarlani Beris, Mohammadsadegh Nouri, and Amir Sharifzadeh Javidi. Application of segment anything model for civil infrastructure defect assessment. arXiv preprint arXiv:2304.12600, 2023. +[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In Advances in Neural Information Processing Systems, pages 25005-25017, 2022. +[3] Yoshua Bengio and Yann LeCun. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. +[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021. +[5] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023. +[6] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, Wei Liu, et al. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023. +[7] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022. +[8] Iraklis Giannakis, Anshuman Bhardwaj, Lydia Sam, and Georgios Leontidis. Deep learning universal crater detection using segment anything model (sam). arXiv preprint arXiv:2304.07764, 2023. +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. +[10] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126, 2022. +[11] Joakim Johnander, Johan Edstedt, Michael Felsberg, Fahad Shahbaz Khan, and Martin Danelljan. Dense gaussian processes for few-shot segmentation. In European Conference on Computer Vision, pages 217-234, 2022. +[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. arXiv preprint arXiv:2306.01567, 2023. +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. +[14] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE conference on computer vision and pattern recognition, pages 8057-8067, 2022. + +[15] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021. +[16] Zechao Li, Yanpeng Sun, Liyan Zhang, and Jinhui Tang. Cnt-net: Context-based tandem network for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9904-9917, 2021. +[17] Jie Liu, Yanqi Bao, Guo-Sen Xie, Huan Xiong, Jan-Jakob Sonke, and Efstratios Gavves. Dynamic prototype convolution network for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2022. +[18] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022. +[19] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. +[21] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In IEEE International Conference on Computer Vision, pages 8741-8750, 2021. +[22] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In IEEE international conference on computer vision, pages 6941-6952, 2021. +[23] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Hm: Hybrid masking for few-shot segmentation. In European Conference on Computer Vision, pages 506-523, 2022. +[24] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In IEEE International Conference on Computer Vision, pages 622-631, 2019. +[25] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[26] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. +[27] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023. +[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, + +Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. +[29] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017. +[30] Yanpeng Sun and Zechao Li. Ssa: Semantic structure aware inference for weakly pixel-wise dense predictions without cost. arXiv preprint arXiv:2111.03392, 2021. +[31] Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, and Jingdong Wang. Singular value fine-tuning: Few-shot segmentation requires few-parameters fine-tuning. In Advances in Neural Information Processing Systems, pages 37484–37496, 2022. +[32] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. +[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. +[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. +[35] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, 2016. +[36] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In IEEE international conference on computer vision, pages 9197-9206, 2019. +[37] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. +[38] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In IEEE International Conference on Computer Vision, pages 1130-1140, 2023. +[39] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. +[40] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 763-778, 2020. +[41] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. +[42] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with seg- + +ment anything models. arXiv preprint arXiv:2306.04121, 2023. +[43] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. +[44] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. In Advances in Neural Information Processing Systems, pages 21984-21996, 2021. +[45] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. In Advances in Neural Information Processing Systems, pages 6575–6588, 2022. +[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. +[47] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. +[48] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023. \ No newline at end of file diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/images.zip b/2024/VRP-SAM_ SAM with Visual Reference Prompt/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d62c5e36a7938cef13aeef3064d002ebafad5ec --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c22e218aa2eaae028f693cc455d938bacb8a9c096820d9b9cbb7a1c2cb403b2d +size 865334 diff --git a/2024/VRP-SAM_ SAM with Visual Reference Prompt/layout.json b/2024/VRP-SAM_ SAM with Visual Reference Prompt/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42b7d514acb441604c310c2e8066d5455b60ddcb --- /dev/null +++ b/2024/VRP-SAM_ SAM with Visual Reference Prompt/layout.json @@ -0,0 +1,9848 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "text", + "content": "VRP-SAM: SAM with Visual Reference Prompt" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": "Yanpeng Sun" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Jiahui Chen" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Shan Zhang" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Xinyu Zhang" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Qiang Chen" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Gang Zhang" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Jingdong Wang" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "text", + "content": ", Zechao Li" + }, + { + "bbox": [ + 70, + 141, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "spans": [ + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "text", + "content": "Nanjing University of Science and Technology, " + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "text", + "content": "Baidu VIS, " + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "text", + "content": "Beihang University, " + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 115, + 178, + 477, + 223 + ], + "type": "text", + "content": "Australian National University {yanpeng.sun, zechao.li}@njust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 249, + 192, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 249, + 192, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 249, + 192, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 274, + 290, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 290, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 290, + 561 + ], + "type": "text", + "content": "In this paper, we propose a novel Visual Reference Prompt (VRP) encoder that empowers the Segment Anything Model (SAM) to utilize annotated reference images as prompts for segmentation, creating the VRP-SAM model. In essence, VRP-SAM can utilize annotated reference images to comprehend specific objects and perform segmentation of specific objects in target image. It is note that the VRP encoder can support a variety of annotation formats for reference images, including point, box, scribble, and mask. VRP-SAM achieves a breakthrough within the SAM framework by extending its versatility and applicability while preserving SAM's inherent strengths, thus enhancing user-friendliness. To enhance the generalization ability of VRP-SAM, the VRP encoder adopts a meta-learning strategy. To validate the effectiveness of VRP-SAM, we conducted extensive empirical studies on the Pascal and COCO datasets. Remarkably, VRP-SAM achieved state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Furthermore, VRP-SAM demonstrates strong generalization capabilities, allowing it to perform segmentation of unseen objects and enabling cross-domain segmentation. The source code and models will be available at https://github.com/syp2ysy/VRPSAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 583, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 602, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 687 + ], + "type": "text", + "content": "In recent, the Segment Anything Model (SAM) [13] has emerged as a foundational visual model for image segmentation. Trained on an extensive datasets comprising billions of labels, SAM has demonstrated remarkable versatility in the realm of universal segmentation. What sets SAM apart is its human-interactive design, allowing segmentation based on user-provided prompts, be they in the form of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 249, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 308, + 249, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 371 + ], + "type": "image", + "image_path": "279cd4ad8789425d21e414ca504cc98936cbc938093da7b046c4f32cfcb431d9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 546, + 411 + ], + "type": "text", + "content": "Figure 1. Comparison of SAM's built-in Point and Box prompt modes with Visual Reference Prompt when handling numerous images. The prompts are all provided by users." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 422, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 458 + ], + "type": "text", + "content": "points, bounding boxes, or coarse masks. This distinctive feature positions SAM as a robust tool that can be adaptable to various tasks and requirements [12, 48]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 674 + ], + "type": "text", + "content": "However, the existing prompt formats of SAM present significant challenges in practical applications, especially when dealing with complex scenes and numerous images. As shown in Figure 2(a), SAM relies on user-provided prompts (points, boxes, coarse mask) to segment objects in the target image, demanding users to possess a comprehensive understanding of the target objects. In real-world applications, especially in complex scenarios, the level of user familiarity with the target objects can significantly impact the effectiveness of providing specific prompts. Furthermore, variations in the position, size, and quantity of target objects across different images require custom prompts for each image. As illustrated in Figure 1 with the aim of segmenting 'bicycles', users need to customize different prompts for each image, significantly impacting efficiency of SAM. Therefore, we propose integrating visual reference prompts to overcome these limitations and enhance adaptability of SAM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": "Visual reference prompts is annotated reference images that delineate the objects users expect to segment. As shown in Figure 1, by simply providing a visual reference prompt" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 137, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 137, + 702 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 137, + 702 + ], + "type": "text", + "content": "*equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23565" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 283, + 144 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 283, + 144 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 283, + 144 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 283, + 144 + ], + "type": "image", + "image_path": "692023180000c933416f24d6830d1c26a4a0fdd99fb2f92e00cf025279a266c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 145, + 283, + 258 + ], + "blocks": [ + { + "bbox": [ + 52, + 145, + 283, + 258 + ], + "lines": [ + { + "bbox": [ + 52, + 145, + 283, + 258 + ], + "spans": [ + { + "bbox": [ + 52, + 145, + 283, + 258 + ], + "type": "image", + "image_path": "ef0ce84e843f6b8e08e4a28818863278170bc594756302651c4ae41d7f47f21a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 310 + ], + "type": "text", + "content": "Figure 2. A Comparison between SAM and VRP-SAM. VRP-SAM introduces a visual reference prompt encoder, accepting annotated reference images with point, scribble, box, and mask formats, offering a distinct enhancement over SAM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 319, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 488 + ], + "type": "text", + "content": "with bicycle, we can segment bicycle in different images without requiring users to provide specific prompts for each image. It significantly enhances the efficiency of SAM while reducing reliance on user familiarity with objects. To achieve this goal, some methods [19, 46] incorporate semantic correlation models [9, 25] to establish reference-target correlation1, obtaining pseudo-masks for the target objects. Following this, a sampling strategy is devised to extract a set of points and bounding boxes from the pseudomasks, serving as prompts for SAM segment the target image. These methods overlook false positives within the pseudo-mask and exhibit high sensitivity to hyperparameters. Consequently, it heavily relies on the quality of the pseudo-mask and has poor generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 488, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 681 + ], + "type": "text", + "content": "Toward this end, we propose a straightforward and effective Visual Reference Prompt (VRP) encoder using meta-learning technique, integrated with SAM to create VRPSAM. It leverages annotated reference images as prompts to segment similar semantic objects in the target image. As illustrated in Figure 2(b), the VRP encoder accepts inputs in various annotation formats, including points, scribbles, boxes, and masks. Specifically, the VRP encoder introduces a semantic-related model to encode reference and target images into the same space. Following meta-learning methods, we first extract prototypes of target objects from annotated information in reference images, enhancing the representation of the target objects in both images. Following meta-learning methods, the prototypes of target objects (users' markers) are first generated from annotated reference images, which aim to highlight such target instances" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "content": "in both reference and target images. Then, we introduce a set of learnable queries to extract the semantic cues of the target objects from attentive enhanced reference features. Then, these queries interact with target images, generating prompt embeddings usable by mask decoder to segment semantically specific objects in the target image. Building upon the foundation of SAM's inherent capabilities, VRPSAM enhances the model's visual reference segmentation prowess. The introduction of Visual Reference Prompts (VRP) not only diversifies the prompts, enabling the model to swiftly segment objects with identical semantics, but also incorporates a meta-learning mechanism that significantly boosts the generalization of model, particularly in dealing with novel objects and cross-domain scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "text", + "content": "To quantitatively assess the generalization capability of VRP-SAM, we follow the dataset configurations commonly used in few-shot segmentation and evaluate our VRP-SAM on Pascal- " + }, + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "text", + "content": " and COCO-20i datasets. Extensive experiments show that VRP-SAM overcomes the limitations of SAM in prompt format, enabling efficient visual reference segmentation. It is significantly better than sampling-based methods, achieving state-of-the-art results on Pascal- " + }, + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 240, + 546, + 431 + ], + "type": "text", + "content": " and COCO-20i datasets. Moreover, we present solid evidence of VRP-SAM's superior performance in handling unknown objects and cross-domain scenarios. Our experiments highlight that VRP-SAM achieves the rapid segmentation of a large number of images based on semantics. Furthermore, the incorporation of meta-learning principles significantly enhances its generalization, making it applicable across various scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 445, + 392, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 392, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 392, + 457 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 465, + 422, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 422, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 422, + 478 + ], + "type": "text", + "content": "2.1. Application of SAM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 664 + ], + "type": "text", + "content": "Segmentation Anything model (SAM) [13] is a recently introduced category-agnostic interactive segmentation model by Meta. It leverages user-guided instructions for segmentation, such as points, bounding boxes, and coarse masks. Currently, SAM has two primary application approaches. One involves using SAM's segmentation results as prior information to assist downstream tasks. For instance, Inpaint Anything (IA) [43] and Edit Everything [39] leverage SAM's segmentation results for image editing and restoration [42, 47] within the masked regions. Additionally, SEPL [5] employs SAM's segmentation results to enhance pseudo-labels in weakly supervised segmentation tasks [27, 30]. Furthermore, SAM segmentation results serve as prior information in tasks such as crack and volcano crater detection [1, 8]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The second involves guiding SAM's segmentation through various prompt combinations. For example, in the context of visual reference segmentation [31], Matcher [19] samples points and boxes from pseudo-masks, utilizing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "1SAM, being a category-agnostic model, encounters challenges in adequately capturing reference-target correlations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23566" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "SAM to refine these pseudo-masks. TAM [41] is applied in object tracking tasks, where it uses point prompts to initialize object masks and subsequently uses SAM to refine low-quality masks. Additionally, SAMAug [6] introduces a visual point enhancement method tailored for SAM, facilitating automatic image annotation. These examples illustrate SAM's effectiveness across different tasks. However, existing methods are constrained by SAM's existing prompt modalities, and they may struggle when confronted with complex objects and unfamiliar scenes. To break these limitations, we have designed a visual prompt encoder for SAM, expanding its applicability to a wider range of scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 234, + 211, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 211, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 211, + 247 + ], + "type": "text", + "content": "2.2. Visual reference segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 529 + ], + "type": "text", + "content": "Visual reference segmentation [31, 33, 48] aims to guide the segmentation of a target image using a reference image. The goal of this task is to utilize a semantically annotated reference image to instruct the segmentation of objects or regions in the target image that share the same semantics as those in the reference image. In current research, methods can be broadly categorized into two groups: prototype-based and feature-matching-based. Prototype-based methods, such as PFENet [33], PANet [36], and CWT [21], usually focus on distinguishing prototypes with different class-specific features. ASGNet [15], on the other hand, improves the segmentation performance by increasing the number of prototypes. Another feature-matching approach [22, 35] leverages the pixel-level correlations between reference and target images to significantly enhance segmentation performance, as demonstrated by methods like CyCTR [44] and HDMNet [26]. Moreover, modern large-scale vision models [2, 37, 48] have recognized visual reference segmentation as a primary task, given its indispensable role in handling complex objects and unknown scenes. However, it's important to note that SAM [13] does not possess the capability to perform this task, underscoring the necessity of introducing VPR-SAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 537, + 125, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 125, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 125, + 551 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 594 + ], + "type": "text", + "content": "In this section, we first review the architecture of Segment Anything Model. Then we formulate the problem setting in this paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 157, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 157, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 157, + 611 + ], + "type": "text", + "content": "3.1. SAM Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "SAM is an interactive segmentation model composed of an image encoder, a prompt encoder, and a mask decoder. Given a target image " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " and some geometric prompts, SAM first employs a Vision Transformer (ViT) [34] as image encoder to extract image features. Subsequently, the prompt encoder is utilized to generate the prompt embeddings derived from the user-provided points, boxes, or masks. Finally, the mask decoder integrates the image features and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "prompt embeddings, and generates a high-quality mask in a class-agnostic manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 97, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 169 + ], + "type": "text", + "content": "To the best of our knowledge, current research based on SAM still relies on geometric prompts for segmentation. Thus far, there have been no efforts to introduce new forms of prompts to enhance SAM. We design a novel visual reference prompt to extend the visual reference segmentation capabilities of SAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 178, + 430, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 430, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 430, + 190 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": "Visual reference segmentation aims to segment objects in target image that same semantic as the annotated object in reference image. Specifically, let " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": " represent the target image and " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": " represent the reference image. Given a reference image " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": " and its annotation " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "M_{r}^{i}" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": " denotes the category of the annotated object, we leverage this information to segment all regions in the target image belonging to category " + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 197, + 547, + 365 + ], + "type": "text", + "content": ". Depending on the annotation granularity, the reference image have four annotation formats: point, scribble, box, and mask. Point annotation involves providing specific points on the target, scribble requires outlining the target region with arbitrary curves, boxes correspond to bounding boxes around the target, and masks provide pixel-level annotations for the entire target." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 376, + 377, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 377, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 377, + 389 + ], + "type": "text", + "content": "4. VRP-SAM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 397, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 505 + ], + "type": "text", + "content": "VRP-SAM extends SAM to perform visual reference segmentation without compromising its original functionality. We propose a training-efficient visual reference prompt encoder that, firstly, accommodates various granularities of visual references and, secondly, directly encodes these visual references into prompt embeddings rather than geometric prompts. These prompt embeddings are then fed directly into the mask decoder of SAM, resulting in the generation of the target mask." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "type": "text", + "content": "As depicted in Figure 3, the VRP Encoder consists of feature augmenter and prompt generator. Next, we will delve into the details of VRP Encoder and the loss function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 421, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 421, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 421, + 563 + ], + "type": "text", + "content": "4.1. Feature Augmenter" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Inspired by meta-learning, the Feature Augmenter separately encodes reference annotations " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "M_r^i" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " into features of both the reference and target images. This process is designed to distinguish between foreground and background representations. To capture semantic correlations between reference and target images, we introduce a semantic-aware image encoder within the VRP encoder, encoding them into the same latent space. To prevent overfitting of the VRP encoder, we freeze the image encoder during the training phase. This ensures a balance between capturing semantic relevance and preventing excessive specialization of the VRP encoder." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23567" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 545, + 300 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 545, + 300 + ], + "type": "image", + "image_path": "1200482160451724d569485715c0f72ffa0d9b6becd63a62d4df9c425cdcf670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 306, + 546, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 546, + 340 + ], + "type": "text", + "content": "Figure 3. Proposed VRP-SAM framework. Our approach enables SAM to perform visual reference segmentation by extends a VRP encoder. It takes various granularities of visual references as inputs and encodes these visual references into prompt embeddings. Our VRP encoder consists of a feature augmenter and a prompt generator." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": "The Feature Augmenter is illustrated in Figure 3. Initially, leveraging a semantic-aware image encoder (e.g. ResNet-50), we encode " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " separately, resulting in the feature map " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "F_{r} \\in \\mathbf{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "F_{t} \\in \\mathbf{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": ". Subsequently, we extract the prototype feature " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " corresponding to class " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": " using the mask " + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "inline_equation", + "content": "M_{r}^{i}" + }, + { + "bbox": [ + 46, + 343, + 288, + 427 + ], + "type": "text", + "content": ". This process can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 435, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 435, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 102, + 435, + 287, + 449 + ], + "type": "interline_equation", + "content": "P _ {i} = \\operatorname {M a s k A v g P o o l} \\left(F _ {r}, M _ {r} ^ {i}\\right) \\tag {1}", + "image_path": "c099782242b3f3a753443a345cdb568cba891e052f5f5edce426e1d9a9e79a5b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "M_r^i" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " represents one of following annotation formats: point, scribble, box, or mask. To enhance the context information about class " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": ", we concatenate the prototype features and mask with " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "F_r" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "F_r" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " is concatenated with mask " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " is concatenated with pseudo-mask " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "m_i^{\\text{pseudo}}" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "m_i^{\\text{pseudo}}" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " is obtained using a common training-free approach, and detailed descriptions are provided in the Appendix. Subsequently, we employ a shared " + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 458, + 287, + 579 + ], + "type": "text", + "content": " convolution layer to reduce dimensionality of enhanced features. This process can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 586, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 586, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 90, + 586, + 287, + 601 + ], + "type": "interline_equation", + "content": "F _ {r} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {r}, P _ {i}, m _ {i}\\right)\\right) \\tag {2}", + "image_path": "5bbde16229e9e84765918e99bac6e287026943e13aabdca420147b4a7810e70d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 604, + 287, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 604, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 91, + 604, + 287, + 619 + ], + "type": "interline_equation", + "content": "F _ {t} ^ {\\prime} = \\operatorname {C o n v} \\left(\\operatorname {c o n c a t} \\left(F _ {t}, P _ {i}, m _ {i} ^ {\\text {p s e u d o}}\\right)\\right) \\tag {3}", + "image_path": "90730e7d6aad87691ba625aa50aeae3d1f8de29af9075006ec777c3ab12475d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": "Ultimately, we obtain enhanced image features " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_r' \\in \\mathbf{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_t' \\in \\mathbf{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": ", which have enhanced context information for the category " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": ". Thus, the enhanced features comprise foreground representations for class " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": " and background representations for the other classes. Subsequently, we feed the enhanced features into the Prompt Generator to obtain a set of visual reference prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 343, + 416, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 416, + 357 + ], + "type": "text", + "content": "4.2. Prompt Generator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "text", + "content": "The purpose of the Prompt Generator is to obtain a set of visual reference prompts embedding for the SAM mask decoder. As depicted in Figure 3, the process commences with the introduction of a set of learnable queries denoted as " + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbf{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "text", + "content": " indicates the number of visual reference prompts. These queries first engage with the reference features " + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "inline_equation", + "content": "F_r'" + }, + { + "bbox": [ + 304, + 361, + 546, + 457 + ], + "type": "text", + "content": " to obtain the category-specific information through cross-attention and self-attention layer:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 463, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 463, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 350, + 463, + 545, + 478 + ], + "type": "interline_equation", + "content": "Q _ {r} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {1} \\left(\\operatorname {C r o s s A t t n} _ {1} \\left(Q, F _ {r} ^ {\\prime}\\right)\\right) \\tag {4}", + "image_path": "a378536d33e69a32c85a6ad4b5ac45f31573bec653f132d7b53b6211fbb74ad9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": "Here, we obtain a set of queries, " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "Q_r'" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": ", possessing knowledge about the object to be segmented. Subsequently, we employ cross-attention to interact these queries with target image feature to obtain the foreground information in target image. Following this, a self-attention layer is used to update the queries, generating a set of prompts " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "Q_t'" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": " that align with the representation of SAM:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 572, + 545, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 572, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 348, + 572, + 545, + 588 + ], + "type": "interline_equation", + "content": "Q _ {t} ^ {\\prime} = \\operatorname {S e l f A t t n} _ {2} \\left(\\operatorname {C r o s s A t t n} _ {2} \\left(Q _ {r} ^ {\\prime}, F _ {t} ^ {\\prime}\\right)\\right) \\tag {5}", + "image_path": "b1c0247f4560bc133e6d63a1b9648812f312e1f01d7e233c3bcd14445d2522ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "content": "The final " + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "inline_equation", + "content": "Q_{t}^{\\prime}" + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "content": " serve as the visual reference prompt embeddings for SAM, equipped with the capability to guide the segmentation of the foreground in the target image. By inputting this set of visual reference prompt embeddings into the mask decoder, the mask " + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "inline_equation", + "content": "M_{t}^{i}\\in \\mathbf{R}^{1\\times H\\times W}" + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "content": " for category " + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 592, + 545, + 665 + ], + "type": "text", + "content": " in the target image can be obtained." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "type": "text", + "content": "4.3. Loss Function" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We employ Binary Cross-Entropy (BCE) loss and Dice loss to supervise the learning of the Visual Reference Prompt" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23568" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 118, + 287, + 234 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 1. Compare with other foundation models. Results of one-shot semantic segmentation on COCO- " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": ". Gray indicates the model is trained by in-domain datasets. " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " indicates the method using SAM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 118, + 287, + 234 + ], + "lines": [ + { + "bbox": [ + 50, + 118, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 50, + 118, + 287, + 234 + ], + "type": "table", + "html": "
MethodsLabel typeF-0F-1F-2F-3Means
Painter [37]mask.31.235.333.532.433.1
SegGPT [38]56.357.458.951.756.1
PerSAM† [46]23.123.622.023.423.0
PerSAM-F† [46]22.324.023.424.123.5
Masker† [19]52.753.552.652.152.7
VRP-SAM†point.30.139.243.040.438.2
scribble.40.252.052.444.447.2
box.44.549.355.749.149.7
mask.48.155.860.051.653.9
", + "image_path": "64e041d8ad3002fc5827f74ce1df680c355e73c01bc4cf17cc6ee01e2879acab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 237, + 287, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 273 + ], + "type": "text", + "content": "Encoder. The BCE loss ensures pixel-wise accuracy, while the Dice loss provides additional context for pixel-level segmentation. Therefore, the total loss of VRP-SAM is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 274, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 287, + 367 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {t o t a l}} = \\underbrace {- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log (p _ {i}) + (1 - y _ {i}) \\log (1 - p _ {i}) \\right]} _ {\\mathrm {B C E L o s s}} \\\\ + \\underbrace {1 - \\frac {2 \\sum_ {i = 1} ^ {N} \\left(p _ {i} \\cdot y _ {i}\\right)}{\\sum_ {i = 1} ^ {N} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {N} y _ {i} ^ {2}}} _ {\\text {D i c e L o s s}} \\tag {6} \\\\ \\end{array}", + "image_path": "bab5d03b34015efcae9f41b018be8aa08eaacaaf1a9b07bde1765d69ffc009fe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " represents the total number of pixels, " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " is the ground truth label for pixel " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " is the predicted probability of pixel " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " belonging to the object. By combining these two losses, we comprehensively consider both accuracy and contextual effects, thus guiding the Visual Reference Prompt Encoder more effectively in generating precise segmentation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 104, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 104, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 104, + 493 + ], + "type": "text", + "content": "5.1. Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": "Datasets: To validate segmentation performance and generalization capability of VRP-SAM, we conducted extensive experiments following the few-shot setting [14, 31, 44] on COCO-20" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " [24] and PASCAL-5" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " [29] datasets. Specifically, we organized all classes from both datasets into 4 folds. For each fold, PASCAL-5" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " [29] comprises 15 base classes for training and 5 novel classes for testing, while COCO-20" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 46, + 498, + 288, + 677 + ], + "type": "text", + "content": " [24] includes 60 training base classes and 20 testing novel classes. To assess performance of model, we randomly sampled 1000 reference-target pairs in each fold. In each fold, As the mentioned datasets lack labels for point, scribble, and box annotations, we followed the SEEM [48] to generate these annotation labels by simulating user inputs randomly based on the reference ground truth masks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Implementation details: In visual reference prompt encoder, we use VGG-16 [3] and ResNet-50 [9] as the im" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "content": "age encoder and initialize it with ImageNet [28] pre-trained weights. We employed the AdamW optimizer [20] along with a cosine learning rate decay strategy for training VRPSAM. Specifically, on the COCO- " + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "content": " dataset, we conducted 50 epochs of training with an initial learning rate of 1e-4 and a batch size of 8. For the PASCAL- " + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "content": " dataset, the model was trained for 100 epochs with an initial learning rate of 2e-4 and a batch size of 8. In VRP, the number of queries is set to 50 by default, and the input image size of all experiments needs to be adjusted to " + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "content": ". Following SEEM [48], during training, we obtain annotations for points, scribbles, and boxes based on mask annotations. We provide detailed descriptions of this process in the Appendix. To ensure a fair comparison, VRP-SAM is exclusively compared to previous works [16, 26, 37] using visual reference prompts based on the mean intersection over union (mIoU)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 285, + 503, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 503, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 503, + 298 + ], + "type": "text", + "content": "5.2. Comparison with the State-of-the-art" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "content": "Comparison with other foundation models. Leveraging the foundation model enables few-shot segmentation. We compared various approaches utilizing Painter [37], SegGPT [38], and SAM as foundation models for few-shot segmentation, providing evaluation metrics on the COCO-20" + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "content": " dataset. As shown in Table 1, VRP-SAM achieves " + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 448 + ], + "type": "text", + "content": " mean mIoU without training on novel classes, achieving comparable with SegGPT. Note that the training data of SegGPT include all classes in COCO. Furthermore, our VRP-SAM outperforms other SAM-base methods including Matcher [19], which employs a DINOv2 [25] pretrained ViT-L [34] as image encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": "Comparison with few-shot methods. To validate the effectiveness of VRP-SAM, we compared it with state-of-the-art few-shot segmentors on the novel set of COCO- " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " datasets. To ensure a fair comparison, VRP-SAM is trained and tested separately in each fold, ensuring no overlap in classes between the training and test sets. The results in Table 2 demonstrate that VRP-SAM achieves state-of-the-art results on COCO" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ", with mIoU scores of 53.9 and 71.9, respectively. Notably, when using VGG-16 as the image encoder for VRP, VRP-SAM achieves mIoU scores of 48.0 and 68.7 on COCO- " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " datasets. It indicates that VRP-SAM achieves optimal performance on the novel set with only 1.6M learnable parameters, highlighting its powerful generalization capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 634, + 504, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 634, + 504, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 634, + 504, + 647 + ], + "type": "text", + "content": "5.3. Comparison with Geometric Prompts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we design Visual Reference Prompts to represent target information. Different from the Geometric Prompts (GP) provided by SAM, our VRP is more flexible and robust. We conducted experiments comparing GP and VRP to validate the superiority of our method (see Table 3)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23569" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 98, + 547, + 288 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Performance of one-shot semantic segmentation on COCO- " + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "inline_equation", + "content": "{20}^{i}" + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "inline_equation", + "content": "{5}^{i}" + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": " . The red and blue colors respectively represent the optimal and suboptimal results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 98, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 49, + 98, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 98, + 547, + 288 + ], + "type": "table", + "html": "
MethodImage encoderLearnable paramsCOCO-20iPASCAL-5i
F-0F-1F-2F-3MeanF-0F-1F-2F-3Mean
PFENet [33]VGG-1610.4M35.438.136.834.736.356.968.254.552.458.0
BAM [14]4.9M36.447.143.341.742.163.270.866.157.564.4
HDMNet [26]4.2M40.750.648.244.045.964.871.467.756.465.1
VRP-SAM1.6M43.651.750.046.548.070.074.768.361.968.7
PFENet [33]ResNet-5010.4M36.538.634.533.835.861.769.555.456.360.8
HSNet [22]2.6M36.343.138.738.739.264.370.760.360.564.0
CyCTR [44]15.4M38.943.039.639.840.365.771.059.559.764.0
SSP [7]8.7M35.539.637.936.737.460.567.866.451.061.4
NTRENet [18]19.9M36.842.639.937.939.365.472.359.459.864.2
DPCN [17]-42.047.043.339.743.065.771.669.160.666.7
VAT [10]3.2M39.043.842.639.741.367.672.062.360.165.5
BAM [14]4.9M39.449.946.245.245.269.073.667.661.167.8
HDMNet [26]4.2M43.855.351.649.450.071.075.468.962.169.4
VRP-SAM1.6M48.155.860.051.653.973.978.370.665.071.9
DCAMASwin-B47.7M49.552.752.848.750.972.273.864.367.169.3
", + "image_path": "1ba7aa129b550ea44ffc57401ad97af1bb341afbfb39f814d41c98f1c61e4395.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 340, + 285, + 494 + ], + "blocks": [ + { + "bbox": [ + 47, + 297, + 288, + 331 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 288, + 331 + ], + "type": "text", + "content": "Table 3. Comparison with geometric prompts. Geometric prompts are randomly sampled from the pseudo-mask. " + }, + { + "bbox": [ + 47, + 297, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 297, + 288, + 331 + ], + "type": "text", + "content": " indicates the carefully designed sampling strategy proposed in [19]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 340, + 285, + 494 + ], + "lines": [ + { + "bbox": [ + 49, + 340, + 285, + 494 + ], + "spans": [ + { + "bbox": [ + 49, + 340, + 285, + 494 + ], + "type": "table", + "html": "
MethodImage EncoderPromptsMean IoU
GP-SAMResNet-50box.19.7
point.21.7
box. + point.23.2
DINOv2box.31.3
point.36.6
box. + point.37.8
box. + point. † [19]52.7
VRP-SAMResNet-50point.38.4
scribble.47.3
box.49.7
mask.53.9
DINOv2-B/14mask.60.4
", + "image_path": "57951279882aec9b3f90783f37cc2322b289f674b7ccf96bca57cfff9f77fa11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "content": "In the GP-SAM experiments, we initially used an Image Encoder to generate a pseudo-mask for the target image following [33], and subsequently obtained points or bounding boxes from the pseudo-mask as geometric prompts. Experimental results consistently demonstrate the ongoing superiority of our VRP over the GP approach. We visualize the segmentation results of GP-SAM and our VRP-SAM in Figure 4. Visualization results indicate that the GP approach is prone to generating false-positive prompts, significantly impacting segmentation performance. In contrast, our VRP effectively avoids such issues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 640, + 194, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 194, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 194, + 651 + ], + "type": "text", + "content": "5.4. Generalization Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 287, + 683 + ], + "type": "text", + "content": "Domain shift: Next, we assess the effectiveness of VRPSAM in a domain shift scenario, which necessitates signif" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 297, + 364, + 346 + ], + "blocks": [ + { + "bbox": [ + 310, + 297, + 364, + 346 + ], + "lines": [ + { + "bbox": [ + 310, + 297, + 364, + 346 + ], + "spans": [ + { + "bbox": [ + 310, + 297, + 364, + 346 + ], + "type": "image", + "image_path": "bcd4cb3092239f913c25dc0beb955a87c28b8944125d4e25fefa2acc636beef1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 311, + 346, + 364, + 400 + ], + "blocks": [ + { + "bbox": [ + 311, + 346, + 364, + 400 + ], + "lines": [ + { + "bbox": [ + 311, + 346, + 364, + 400 + ], + "spans": [ + { + "bbox": [ + 311, + 346, + 364, + 400 + ], + "type": "image", + "image_path": "3a0d83cd3fd737170c45a218170bf81f94e28c0a14dd2af163c23b80cad60d12.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 401, + 364, + 453 + ], + "blocks": [ + { + "bbox": [ + 311, + 401, + 364, + 453 + ], + "lines": [ + { + "bbox": [ + 311, + 401, + 364, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 401, + 364, + 453 + ], + "type": "image", + "image_path": "49c984490b6632e211a9d199f10b1b7321ac142d499abe111d699a47740d64d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 456, + 364, + 509 + ], + "blocks": [ + { + "bbox": [ + 311, + 456, + 364, + 509 + ], + "lines": [ + { + "bbox": [ + 311, + 456, + 364, + 509 + ], + "spans": [ + { + "bbox": [ + 311, + 456, + 364, + 509 + ], + "type": "image", + "image_path": "246eaed965df90310858d4cf1967a3a6e7f460df87435f1af853c772be1834d3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 510, + 361, + 517 + ], + "lines": [ + { + "bbox": [ + 316, + 510, + 361, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 361, + 517 + ], + "type": "text", + "content": "Visual Reference" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 369, + 297, + 421, + 346 + ], + "blocks": [ + { + "bbox": [ + 369, + 297, + 421, + 346 + ], + "lines": [ + { + "bbox": [ + 369, + 297, + 421, + 346 + ], + "spans": [ + { + "bbox": [ + 369, + 297, + 421, + 346 + ], + "type": "image", + "image_path": "2288c72273e4b217ab76c97dd49b7e950f62ef71d613960ee04e39fc1b411d77.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 369, + 346, + 421, + 399 + ], + "blocks": [ + { + "bbox": [ + 369, + 346, + 421, + 399 + ], + "lines": [ + { + "bbox": [ + 369, + 346, + 421, + 399 + ], + "spans": [ + { + "bbox": [ + 369, + 346, + 421, + 399 + ], + "type": "image", + "image_path": "fd1334a4f875164eeacc12e1e6f4ffb643bcb14ad493cd1d25a678f8c896023f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 369, + 401, + 421, + 453 + ], + "blocks": [ + { + "bbox": [ + 369, + 401, + 421, + 453 + ], + "lines": [ + { + "bbox": [ + 369, + 401, + 421, + 453 + ], + "spans": [ + { + "bbox": [ + 369, + 401, + 421, + 453 + ], + "type": "image", + "image_path": "f837fc9d7de39914eae01792ad01bd68fad350c47af3191eab0a7c134ea314ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 369, + 456, + 421, + 509 + ], + "blocks": [ + { + "bbox": [ + 369, + 456, + 421, + 509 + ], + "lines": [ + { + "bbox": [ + 369, + 456, + 421, + 509 + ], + "spans": [ + { + "bbox": [ + 369, + 456, + 421, + 509 + ], + "type": "image", + "image_path": "14d799081ecd8daa077f5395ffff76aa2f3495feba55114608424e0d453ea8d1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 510, + 418, + 517 + ], + "lines": [ + { + "bbox": [ + 374, + 510, + 418, + 517 + ], + "spans": [ + { + "bbox": [ + 374, + 510, + 418, + 517 + ], + "type": "text", + "content": "GP-SAM-RN50" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 424, + 297, + 476, + 346 + ], + "blocks": [ + { + "bbox": [ + 424, + 297, + 476, + 346 + ], + "lines": [ + { + "bbox": [ + 424, + 297, + 476, + 346 + ], + "spans": [ + { + "bbox": [ + 424, + 297, + 476, + 346 + ], + "type": "image", + "image_path": "d87dd1cd65efd2c715a7b6a14a93e7e2bea2730b863dff0b68c6f2a647481f31.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 424, + 346, + 476, + 399 + ], + "blocks": [ + { + "bbox": [ + 424, + 346, + 476, + 399 + ], + "lines": [ + { + "bbox": [ + 424, + 346, + 476, + 399 + ], + "spans": [ + { + "bbox": [ + 424, + 346, + 476, + 399 + ], + "type": "image", + "image_path": "8ab4534d999a15ae392d8fd7e3d56e417612e99c6db0af8b116a71a95160155a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 424, + 401, + 476, + 453 + ], + "blocks": [ + { + "bbox": [ + 424, + 401, + 476, + 453 + ], + "lines": [ + { + "bbox": [ + 424, + 401, + 476, + 453 + ], + "spans": [ + { + "bbox": [ + 424, + 401, + 476, + 453 + ], + "type": "image", + "image_path": "f51aeb8de1305a24306bbb087480d778617e39b1b2ec5e4c675483ff9c75c7a4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 424, + 456, + 476, + 509 + ], + "blocks": [ + { + "bbox": [ + 424, + 456, + 476, + 509 + ], + "lines": [ + { + "bbox": [ + 424, + 456, + 476, + 509 + ], + "spans": [ + { + "bbox": [ + 424, + 456, + 476, + 509 + ], + "type": "image", + "image_path": "9863e7291521b2a40cd51d5f602289c1c5e7583c0556573922b37afb26225bcc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 510, + 476, + 517 + ], + "lines": [ + { + "bbox": [ + 424, + 510, + 476, + 517 + ], + "spans": [ + { + "bbox": [ + 424, + 510, + 476, + 517 + ], + "type": "text", + "content": "GP-SAM-DINOv2" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 479, + 297, + 531, + 346 + ], + "blocks": [ + { + "bbox": [ + 479, + 297, + 531, + 346 + ], + "lines": [ + { + "bbox": [ + 479, + 297, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 479, + 297, + 531, + 346 + ], + "type": "image", + "image_path": "222840ba44ce1080c1f40bdd2a15f0d8b071d015da2202b495587cc685e10132.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 479, + 346, + 531, + 399 + ], + "blocks": [ + { + "bbox": [ + 479, + 346, + 531, + 399 + ], + "lines": [ + { + "bbox": [ + 479, + 346, + 531, + 399 + ], + "spans": [ + { + "bbox": [ + 479, + 346, + 531, + 399 + ], + "type": "image", + "image_path": "0ab87febacbd1c910b62f682511d6376d21560beb44ce5fbc1e0623ae533e729.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 479, + 401, + 531, + 453 + ], + "blocks": [ + { + "bbox": [ + 479, + 401, + 531, + 453 + ], + "lines": [ + { + "bbox": [ + 479, + 401, + 531, + 453 + ], + "spans": [ + { + "bbox": [ + 479, + 401, + 531, + 453 + ], + "type": "image", + "image_path": "d8af1b9e16ae86437828b11c40198bbf1f5e11892a1189fa480e30992f8244a8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 479, + 456, + 531, + 509 + ], + "blocks": [ + { + "bbox": [ + 479, + 456, + 531, + 509 + ], + "lines": [ + { + "bbox": [ + 479, + 456, + 531, + 509 + ], + "spans": [ + { + "bbox": [ + 479, + 456, + 531, + 509 + ], + "type": "image", + "image_path": "f29b352fed25b2d46515260f7b3151c14a8c046bf2f1bc0fae1089afbf9a0a40.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 510, + 529, + 517 + ], + "lines": [ + { + "bbox": [ + 481, + 510, + 529, + 517 + ], + "spans": [ + { + "bbox": [ + 481, + 510, + 529, + 517 + ], + "type": "text", + "content": "VRP-SAM-RN50" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 525, + 541, + 536 + ], + "lines": [ + { + "bbox": [ + 310, + 525, + 541, + 536 + ], + "spans": [ + { + "bbox": [ + 310, + 525, + 541, + 536 + ], + "type": "text", + "content": "Figure 4. The visualization results of VRP-SAM and GP-SAM." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": "icant domain differences between training and testing sets. Following prior work [22, 32, 33], we trained on COCO- " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": " and tested on PASCAL- " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": ", where the classes in training set do not overlap with those in test set. The results in Table 4, representing the average across four folds, clearly demonstrate the superior performance of VRP-SAM in domain shift scenario. Notably, VRP-SAM with scribble annotations outperforms FP-Trans [45] by " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": ", while VRP-SAM with mask annotations surpasses DGPNet [11] by " + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": ". This affirms the robust generalization capability of VRPSAM and its effectiveness in domain transfer scenarios." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Visualization: To assess the generalization capability of VRP-SAM across diverse image styles, we curated a collection of target images from web, spanning various gen" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "2For point prompts, we randomly sample 5 points from the pseudomask. For box prompt, we adopt the bounding box of the pseudo-mask." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23570" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 141, + 154 + ], + "blocks": [ + { + "bbox": [ + 50, + 103, + 61, + 127 + ], + "lines": [ + { + "bbox": [ + 50, + 103, + 61, + 127 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 61, + 127 + ], + "type": "text", + "content": "JIOd" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 76, + 141, + 154 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 141, + 154 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 141, + 154 + ], + "type": "image", + "image_path": "80652b85f5f7f83b00f5695e5d600e4dec217ca35099b5b37bd926294b0b86a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 76, + 222, + 154 + ], + "blocks": [ + { + "bbox": [ + 144, + 76, + 222, + 154 + ], + "lines": [ + { + "bbox": [ + 144, + 76, + 222, + 154 + ], + "spans": [ + { + "bbox": [ + 144, + 76, + 222, + 154 + ], + "type": "image", + "image_path": "8b81ba2695a92d5437bc87437341b00615ac7e52e28bc315e638a7d8ac62d7ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 224, + 76, + 302, + 154 + ], + "blocks": [ + { + "bbox": [ + 224, + 76, + 302, + 154 + ], + "lines": [ + { + "bbox": [ + 224, + 76, + 302, + 154 + ], + "spans": [ + { + "bbox": [ + 224, + 76, + 302, + 154 + ], + "type": "image", + "image_path": "6dd71cd829e2a5a4bfa435ff0fe8c09bff84b7a2a7d6fdf159835edac1018225.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 76, + 381, + 154 + ], + "blocks": [ + { + "bbox": [ + 303, + 76, + 381, + 154 + ], + "lines": [ + { + "bbox": [ + 303, + 76, + 381, + 154 + ], + "spans": [ + { + "bbox": [ + 303, + 76, + 381, + 154 + ], + "type": "image", + "image_path": "4a897af2669cd10a89d0452c5ffdce5118bbaf97be580f7df978177f3a12b1b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 382, + 76, + 460, + 154 + ], + "blocks": [ + { + "bbox": [ + 382, + 76, + 460, + 154 + ], + "lines": [ + { + "bbox": [ + 382, + 76, + 460, + 154 + ], + "spans": [ + { + "bbox": [ + 382, + 76, + 460, + 154 + ], + "type": "image", + "image_path": "5f3c329154d485d4af9bb1f4e12566bc038ce20c73df509cdcbeda1c66db2948.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 462, + 76, + 540, + 154 + ], + "blocks": [ + { + "bbox": [ + 462, + 76, + 540, + 154 + ], + "lines": [ + { + "bbox": [ + 462, + 76, + 540, + 154 + ], + "spans": [ + { + "bbox": [ + 462, + 76, + 540, + 154 + ], + "type": "image", + "image_path": "27639975f24ff2b816b35af3ca2e62ff8266a3f3e6975aa46c16991db4e20b42.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 156, + 141, + 234 + ], + "blocks": [ + { + "bbox": [ + 50, + 180, + 61, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 61, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 61, + 217 + ], + "type": "text", + "content": "S" + } + ] + } + ], + "index": 1, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 156, + 141, + 234 + ], + "lines": [ + { + "bbox": [ + 50, + 156, + 141, + 234 + ], + "spans": [ + { + "bbox": [ + 50, + 156, + 141, + 234 + ], + "type": "image", + "image_path": "1fd91bc4d07435e0cfd07d65b7a5a09c16a055212e6eda90ccbb5eae5132576d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 144, + 156, + 223, + 234 + ], + "blocks": [ + { + "bbox": [ + 144, + 156, + 223, + 234 + ], + "lines": [ + { + "bbox": [ + 144, + 156, + 223, + 234 + ], + "spans": [ + { + "bbox": [ + 144, + 156, + 223, + 234 + ], + "type": "image", + "image_path": "b61d50a7148940ddf68cedf883129879d89f3974b3be04cad5320869a33f6a9b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 156, + 302, + 234 + ], + "blocks": [ + { + "bbox": [ + 225, + 156, + 302, + 234 + ], + "lines": [ + { + "bbox": [ + 225, + 156, + 302, + 234 + ], + "spans": [ + { + "bbox": [ + 225, + 156, + 302, + 234 + ], + "type": "image", + "image_path": "7a99809c44c45c18a4b57e4d836772b166f1059003f811ad36f86675f18fa32d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 303, + 156, + 381, + 234 + ], + "blocks": [ + { + "bbox": [ + 303, + 156, + 381, + 234 + ], + "lines": [ + { + "bbox": [ + 303, + 156, + 381, + 234 + ], + "spans": [ + { + "bbox": [ + 303, + 156, + 381, + 234 + ], + "type": "image", + "image_path": "42af8c625ce8c59d0360b6b381a3fd89198a1415bffb6467ff94e1294a7d192f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 382, + 156, + 460, + 234 + ], + "blocks": [ + { + "bbox": [ + 382, + 156, + 460, + 234 + ], + "lines": [ + { + "bbox": [ + 382, + 156, + 460, + 234 + ], + "spans": [ + { + "bbox": [ + 382, + 156, + 460, + 234 + ], + "type": "image", + "image_path": "2c3859036847fc6cc75094782a41f91a92fc5095e47049da6ba3be65dec2876b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 462, + 156, + 539, + 234 + ], + "blocks": [ + { + "bbox": [ + 462, + 156, + 539, + 234 + ], + "lines": [ + { + "bbox": [ + 462, + 156, + 539, + 234 + ], + "spans": [ + { + "bbox": [ + 462, + 156, + 539, + 234 + ], + "type": "image", + "image_path": "9e464d08743027204ca2b7f7ab3ae33c6f4c8ce1580f3edb577f5a21c1a4fa0f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 50, + 235, + 141, + 312 + ], + "blocks": [ + { + "bbox": [ + 50, + 235, + 141, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 235, + 141, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 235, + 141, + 312 + ], + "type": "image", + "image_path": "e8b7fc657345272017af5182b6eafe5846e20ab226ba50c0bf3d0be193a252b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 144, + 235, + 223, + 312 + ], + "blocks": [ + { + "bbox": [ + 144, + 235, + 223, + 312 + ], + "lines": [ + { + "bbox": [ + 144, + 235, + 223, + 312 + ], + "spans": [ + { + "bbox": [ + 144, + 235, + 223, + 312 + ], + "type": "image", + "image_path": "a442cb22db880e4eee1a97386ed0c728eb9c3facdc41c546fd62755176eb68e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 225, + 235, + 302, + 312 + ], + "blocks": [ + { + "bbox": [ + 225, + 235, + 302, + 312 + ], + "lines": [ + { + "bbox": [ + 225, + 235, + 302, + 312 + ], + "spans": [ + { + "bbox": [ + 225, + 235, + 302, + 312 + ], + "type": "image", + "image_path": "e8dbffd86720765f7e4682f4b5b7790a3d8be1406a1da46d2a0b7452beba61ce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 303, + 235, + 381, + 312 + ], + "blocks": [ + { + "bbox": [ + 303, + 235, + 381, + 312 + ], + "lines": [ + { + "bbox": [ + 303, + 235, + 381, + 312 + ], + "spans": [ + { + "bbox": [ + 303, + 235, + 381, + 312 + ], + "type": "image", + "image_path": "4532c83c7bf26e7647ec2cafe03f9a3d13f7bce3e09af07116064478af3872a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 382, + 235, + 460, + 312 + ], + "blocks": [ + { + "bbox": [ + 382, + 235, + 460, + 312 + ], + "lines": [ + { + "bbox": [ + 382, + 235, + 460, + 312 + ], + "spans": [ + { + "bbox": [ + 382, + 235, + 460, + 312 + ], + "type": "image", + "image_path": "1f3dc69ecfeb8643dd5b93149f7815d24a81db00884253607dd439a7760c9921.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 462, + 235, + 539, + 312 + ], + "blocks": [ + { + "bbox": [ + 462, + 235, + 539, + 312 + ], + "lines": [ + { + "bbox": [ + 462, + 235, + 539, + 312 + ], + "spans": [ + { + "bbox": [ + 462, + 235, + 539, + 312 + ], + "type": "image", + "image_path": "06b324a7d31b1b11a80b04ca0fb60368552c3075388330ba3766a8dcbfe8d161.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 50, + 314, + 141, + 392 + ], + "blocks": [ + { + "bbox": [ + 50, + 314, + 141, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 314, + 141, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 314, + 141, + 392 + ], + "type": "image", + "image_path": "9e041444f742e32b85f737a51fedf0574c170b5a7e296c2dffdc4a1ed3ce7ff0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 393, + 112, + 402 + ], + "lines": [ + { + "bbox": [ + 92, + 393, + 112, + 402 + ], + "spans": [ + { + "bbox": [ + 92, + 393, + 112, + 402 + ], + "type": "text", + "content": "VRP" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 411, + 538, + 422 + ], + "lines": [ + { + "bbox": [ + 54, + 411, + 538, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 538, + 422 + ], + "type": "text", + "content": "Figure 5. Qualitative results of VRP-SAM across diverse image styles is presented. The target images were sourced from the internet." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 144, + 314, + 223, + 392 + ], + "blocks": [ + { + "bbox": [ + 144, + 314, + 223, + 392 + ], + "lines": [ + { + "bbox": [ + 144, + 314, + 223, + 392 + ], + "spans": [ + { + "bbox": [ + 144, + 314, + 223, + 392 + ], + "type": "image", + "image_path": "2b08169c45522b6f22eff1e42f3dd26e7fff2bb541e0d38ce5c48e1791e9bb41.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 393, + 200, + 404 + ], + "lines": [ + { + "bbox": [ + 166, + 393, + 200, + 404 + ], + "spans": [ + { + "bbox": [ + 166, + 393, + 200, + 404 + ], + "type": "text", + "content": "Image-1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 225, + 314, + 302, + 392 + ], + "blocks": [ + { + "bbox": [ + 225, + 314, + 302, + 392 + ], + "lines": [ + { + "bbox": [ + 225, + 314, + 302, + 392 + ], + "spans": [ + { + "bbox": [ + 225, + 314, + 302, + 392 + ], + "type": "image", + "image_path": "38c2788b42e327737dec89980469bfb142bdbfe9772264927362bec3ea1cbd69.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 393, + 280, + 403 + ], + "lines": [ + { + "bbox": [ + 244, + 393, + 280, + 403 + ], + "spans": [ + { + "bbox": [ + 244, + 393, + 280, + 403 + ], + "type": "text", + "content": "Image-2" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 303, + 314, + 381, + 392 + ], + "blocks": [ + { + "bbox": [ + 303, + 314, + 381, + 392 + ], + "lines": [ + { + "bbox": [ + 303, + 314, + 381, + 392 + ], + "spans": [ + { + "bbox": [ + 303, + 314, + 381, + 392 + ], + "type": "image", + "image_path": "6a4017350ddf98c60aa16176e6a44a95491249c073531cf151254f5dc156839d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 393, + 359, + 403 + ], + "lines": [ + { + "bbox": [ + 324, + 393, + 359, + 403 + ], + "spans": [ + { + "bbox": [ + 324, + 393, + 359, + 403 + ], + "type": "text", + "content": "Image-3" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 382, + 314, + 460, + 392 + ], + "blocks": [ + { + "bbox": [ + 382, + 314, + 460, + 392 + ], + "lines": [ + { + "bbox": [ + 382, + 314, + 460, + 392 + ], + "spans": [ + { + "bbox": [ + 382, + 314, + 460, + 392 + ], + "type": "image", + "image_path": "8aa470003623f34b70d56547105e222cf8591414cca5656ac2af956f5207d295.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 393, + 438, + 403 + ], + "lines": [ + { + "bbox": [ + 404, + 393, + 438, + 403 + ], + "spans": [ + { + "bbox": [ + 404, + 393, + 438, + 403 + ], + "type": "text", + "content": "Image-4" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 462, + 314, + 539, + 392 + ], + "blocks": [ + { + "bbox": [ + 462, + 314, + 539, + 392 + ], + "lines": [ + { + "bbox": [ + 462, + 314, + 539, + 392 + ], + "spans": [ + { + "bbox": [ + 462, + 314, + 539, + 392 + ], + "type": "image", + "image_path": "d7f5f00a4c2bd6bd2ff7157cfcb467fe8b2547563e9d55a1f23fc01c2f305e5c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 393, + 522, + 403 + ], + "lines": [ + { + "bbox": [ + 487, + 393, + 522, + 403 + ], + "spans": [ + { + "bbox": [ + 487, + 393, + 522, + 403 + ], + "type": "text", + "content": "Image-5" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 429, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 560 + ], + "type": "text", + "content": "res such as natural landscapes, artworks, and complex environments. Visual Reference Prompts (VRPs) were selected from the COCO dataset to guide the segmentation of these target images. The segmentation results on different image styles are presented in Figure 5. It demonstrate that VRP-SAM adeptly adapts to varied image styles, accurately delineating target objects. Notably, VRP-SAM excels in handling ink-style images, showcasing refined segmentation performance. This compellingly underscores the robustness and effectiveness of VRP-SAM when confronted with images of unknown styles." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 563, + 140, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 140, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 140, + 576 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": "To validate the effectiveness of VRP-SAM, we conducted extensive ablation studies on PASCAL- " + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": ". ResNet-50 was chosen as the image encoder for VRP to ensure experimental consistency. These experiments aimed to thoroughly investigate the performance of VRP-SAM under various conditions." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Loss: To assess the impact of Binary Cross-Entropy (BCE) and Dice losses on VRP-SAM, experiments were conducted on the PASCAL- " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " dataset. Results in Table 5 indicate comparable performance when using BCE or Dice loss individually. Notably, the optimal performance for" + } + ] + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 309, + 454, + 545, + 589 + ], + "blocks": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "text", + "content": "Table 4. Evaluation (Mean IoU " + }, + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 427, + 545, + 449 + ], + "type": "text", + "content": ") under the domain shift from COCO-20 to PASCAL-5." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 454, + 545, + 589 + ], + "lines": [ + { + "bbox": [ + 309, + 454, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 309, + 454, + 545, + 589 + ], + "type": "table", + "html": "
MethodImage encoderLabel typeMean
RPMM [40]ResNet-50mask.49.6
PFENet [33]61.1
RePRI [4]63.2
VAT-HM [23]65.1
HSNet [22]ResNet-101mask.64.1
DGPNet [11]70.1
FP-Trans [45]DeiT-B/16mask.69.7
VRP-SAMResNet-50point.63.5
scribble.72.5
box.72.3
mask.75.9
", + "image_path": "9c2c6c28c8e3d3435185dcd0dfe040d79011d904a9ea84b3e66d736cadc2f9d4.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 596, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 680 + ], + "type": "text", + "content": "VRP-SAM is achieved when both losses are combined. It emphasizes the importance of BCE and Dice losses in guiding VRP-SAM to produce more robust and accurate masks. BCE loss ensures precise pixel-level classification, while Dice loss facilitates accurate spatial localization of objects. Simultaneously employing BCE and Dice losses in VRP-SAM maximizes their complementary advantages." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 681, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 681, + 545, + 713 + ], + "type": "text", + "content": "The number of query: In Figure 6, we conducted a comprehensive analysis of the impact of varying query quantities on VRP-SAM's performance. We observed" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23571" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 85, + 287, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 81 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 81 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 81 + ], + "type": "text", + "content": "Table 5. Ablation study on different loss function in VRP-SAM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 85, + 287, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 85, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 85, + 287, + 137 + ], + "type": "table", + "html": "
MethodLabel typeBce lossDice lossMeans
VRP-SAMmask.✓✗70.1
✗✓70.3
✓✓71.9
", + "image_path": "227fac2ac9f4b2027bfc56f7b2f1400f29a66f484303c2cf95d5e4b722b68530.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 137, + 282, + 285 + ], + "blocks": [ + { + "bbox": [ + 50, + 137, + 282, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 137, + 282, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 137, + 282, + 285 + ], + "type": "image", + "image_path": "b1d89d230e8993a88026585303bcab175c30f91996180c870bd4a2ff8d009941.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 319 + ], + "type": "text", + "content": "Figure 6. Ablation study on VRP-SAM with different query numbers. The x-axis shows the number of queries, and the y-axis represents model performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 321, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 453 + ], + "type": "text", + "content": "a positive correlation between increased query numbers and improved segmentation quality. However, once the query count surpassed 50, the performance gain started to diminish. This phenomenon suggests that 50 queries provide ample effective guidance for the model, and further increases in query count do not yield significant improvements, leading to performance saturation. To maintain high performance while minimizing model learnable parameters, we set the query quantity in VRP-SAM to 50. This decision optimally balances guidance information and model efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 618 + ], + "type": "text", + "content": "Initialization of query: In Table 6, we compared various query initialization strategies, such as random initialization, foreground prototype (FP), background prototype (BP), and a hybrid prototype with half foreground and half background (half-FP & half-BP). Surprisingly, random initialization outperformed all other strategies, showcasing superior performance. This unexpected outcome can be attributed to the intricate segmentation task and the dataset's object diversity. Random initialization enables the model to explore a broader range of states, avoiding potential local minima. In contrast, prototype-based initializations, whether foreground, background, or a combination, might introduce biases, limiting adaptability to diverse object characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Number of VRPs: In our investigation of the influence of number of visual reference prompts on segmentation results, we explored utilization of few visual reference prompts. This approach involves sending several visual reference prompts to the VRP encoder, generating multiple prompt embeddings. These embeddings are concatenated and forwarded to the mask decoder, resulting in the final mask. The result in Table 7, we observed a substantial en" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 95, + 533, + 156 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 91 + ], + "type": "text", + "content": "Table 6. Ablation study of different query initialization methods on VRP-SAM." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 95, + 533, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 95, + 533, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 533, + 156 + ], + "type": "table", + "html": "
MethodLabel typeImage encoderMean IoU
VRP-SAMmask.FP68.2
BP62.6
half-FP & half-BP67.4
random71.9
", + "image_path": "7bc7f61c8f44e17fdd64c2c715fa8200f882cc917db77e91644d5370e4afc465.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 186, + 545, + 253 + ], + "blocks": [ + { + "bbox": [ + 305, + 160, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 160, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 545, + 182 + ], + "type": "text", + "content": "Table 7. Ablation Study on Few Visual Reference prompts for VRP-SAM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 186, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 309, + 186, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 309, + 186, + 545, + 253 + ], + "type": "table", + "html": "
MethodLabel type1-VRP5-VRP
VRP-SAMpoint.62.964.1(+1.2)
scribble.66.868.4(+1.6)
box.69.470.5(+1.1)
mask.71.972.9(+1.0)
", + "image_path": "cb0c7f11ea7557bf172608879da9254c25215c73fdaff7defa8f62ee134f227a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 366 + ], + "type": "text", + "content": "hancement in segmentation performance with an increase in the number of visual reference prompts. The results emphasize positive influence of incorporating multiple visual reference prompts, enhancing the model's ability to accurately capture intricate details and nuances associated with target object. The diverse prompts contribute enriched contextual information, facilitating a more comprehensive understanding of the object's characteristics and leading to generation of precise and nuanced segmentation masks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 380, + 378, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 378, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 378, + 393 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 652 + ], + "type": "text", + "content": "In this paper, we present VRP-SAM, an innovative extension of the SAM framework achieved through integration of a Visual Reference Prompt (VRP) encoder. This addition empowers SAM to leverage visual reference prompts for guided segmentation. The core methodology involves encoding annotated reference images through the VRP encoder, which then interacts with the target image to generate meaningful prompts for segmentation within the SAM framework. The seamless fusion of VRP encoder with SAM, resulting in VRP-SAM, enhances the model's generality and adaptability. VRP-SAM overcomes limitations posed by SAM's existing prompt formats, especially in complex scenarios and large datasets. The introduced visual reference prompts, including point, box, scribble, and mask annotations, offer a flexible solution, expanding the model's applicability. Extensive empirical studies conducted showcase VRP-SAM's state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Notably, VRP-SAM demonstrates robust generalization capabilities, excelling in segmentation tasks for novel objects and cross-domain scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgment This work was partially supported by the National Key Research and Development Program of China under Grant 2022ZD0118802 and the National Natural Science Foundation of China (Grant No. U20B2064 and U21B2043)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23572" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mohsen Ahmadi, Ahmad Gholizadeh Lonbar, Abbas Sharifi, Ali Tarlani Beris, Mohammadsadegh Nouri, and Amir Sharifzadeh Javidi. Application of segment anything model for civil infrastructure defect assessment. arXiv preprint arXiv:2304.12600, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In Advances in Neural Information Processing Systems, pages 25005-25017, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 226 + ], + "type": "text", + "content": "[3] Yoshua Bengio and Yann LeCun. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 226, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 288, + 282 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In IEEE Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "type": "text", + "content": "[6] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, Wei Liu, et al. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 407 + ], + "type": "text", + "content": "[7] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 288, + 452 + ], + "type": "text", + "content": "[8] Iraklis Giannakis, Anshuman Bhardwaj, Lydia Sam, and Georgios Leontidis. Deep learning universal crater detection using segment anything model (sam). arXiv preprint arXiv:2304.07764, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "type": "text", + "content": "[10] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Joakim Johnander, Johan Edstedt, Michael Felsberg, Fahad Shahbaz Khan, and Martin Danelljan. Dense gaussian processes for few-shot segmentation. In European Conference on Computer Vision, pages 217-234, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. arXiv preprint arXiv:2306.01567, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[14] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In IEEE conference on computer vision and pattern recognition, pages 8057-8067, 2022." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[15] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[16] Zechao Li, Yanpeng Sun, Liyan Zhang, and Jinhui Tang. Cnt-net: Context-based tandem network for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9904-9917, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "type": "text", + "content": "[17] Jie Liu, Yanqi Bao, Guo-Sen Xie, Huan Xiong, Jan-Jakob Sonke, and Efstratios Gavves. Dynamic prototype convolution network for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "text", + "content": "[18] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "type": "text", + "content": "[19] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 360 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 404 + ], + "type": "text", + "content": "[21] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In IEEE International Conference on Computer Vision, pages 8741-8750, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 546, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 447 + ], + "type": "text", + "content": "[22] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In IEEE international conference on computer vision, pages 6941-6952, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 546, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 502 + ], + "type": "text", + "content": "[23] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Hm: Hybrid masking for few-shot segmentation. In European Conference on Computer Vision, pages 506-523, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 503, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 546, + 536 + ], + "type": "text", + "content": "[24] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In IEEE International Conference on Computer Vision, pages 622-631, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "text", + "content": "[25] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "type": "text", + "content": "[26] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "type": "text", + "content": "[27] Lixiang Ru, Heliang Zheng, Yibing Zhan, and Bo Du. Token contrast for weakly-supervised semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "text", + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23573" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[29] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[30] Yanpeng Sun and Zechao Li. Ssa: Semantic structure aware inference for weakly pixel-wise dense predictions without cost. arXiv preprint arXiv:2111.03392, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[31] Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, and Jingdong Wang. Singular value fine-tuning: Few-shot segmentation requires few-parameters fine-tuning. In Advances in Neural Information Processing Systems, pages 37484–37496, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[32] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "text", + "content": "[35] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[36] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In IEEE international conference on computer vision, pages 9197-9206, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[37] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[38] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In IEEE International Conference on Computer Vision, pages 1130-1140, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[39] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[40] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 763-778, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[41] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[42] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with seg-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 353 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "ment anything models. arXiv preprint arXiv:2306.04121, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[43] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[44] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. In Advances in Neural Information Processing Systems, pages 21984-21996, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[45] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. In Advances in Neural Information Processing Systems, pages 6575–6588, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[46] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "text", + "content": "[47] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 319, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 353 + ], + "type": "text", + "content": "[48] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23574" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_content_list.json b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bcf0cd3aba12f5404278c4d7f97d3f87faf8eefb --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_content_list.json @@ -0,0 +1,1580 @@ +[ + { + "type": "text", + "text": "VRetouchEr: Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos", + "text_level": 1, + "bbox": [ + 78, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wen Xue1, Le Jiang1, Lianxin Xie1, Si Wu1,2,3*, Yong Xu1,2,3 and Hau San Wong4 \n1School of Computer Science and Engineering, South China University of Technology \n2Peng Cheng Laboratory 3PAZHOU LAB \n4Department of Computer Science, City University of Hong Kong", + "bbox": [ + 143, + 202, + 823, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{csxuewen, csjangle, cslianxin.xie}@mail.scut.edu.cn", + "bbox": [ + 250, + 277, + 718, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cswusi, yxu}@scut.edu.cn, cshswong@cityu.edu.hk", + "bbox": [ + 272, + 295, + 696, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face Video Retouching is a complex task that often requires labor-intensive manual editing. Conventional image retouching methods perform less satisfactorily in terms of generalization performance and stability when applied to videos without exploiting the correlation among frames. To address this issue, we propose a Video Retouching transformEr to remove facial imperfections in videos, which is referred to as VRetouchEr. Specifically, we estimate the apparent motion of imperfections between two consecutive frames, and the resulting displacement vectors are used to refine the imperfection map, which is synthesized from the current frame together with the corresponding encoder features. The flow-based imperfection refinement is critical for precise and stable retouching across frames. To leverage the temporal contextual information, we inject the refined imperfection map into each transformer block for multi-frame masked attention computation, such that we can capture the interdependence between the current frame and multiple reference frames. As a result, the imperfection regions can be replaced with normal skin with high fidelity, while at the same time keeping the other regions unchanged. Extensive experiments are performed to verify the superiority of VRetouchEr over state-of-the-art image retouching methods in terms of fidelity and stability.", + "bbox": [ + 75, + 376, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of digital media enhancement, face retouching in videos plays an important role in improving the facial appearance of persons in dynamic sequences. There are a wide range of applications, from professional video production to the burgeoning field of live-streaming and virtual conferencing. These applications require the meticulous removal of visual imperfections. On the other", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/489ea0340d441e997c7407c852d45dc1e6d9b4638677b9205ecfcb3f54ab1250.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 340, + 870, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9edbcf82fb4a4d69533d41fe3f218277e98e824a6d85a78e35046ad05256b4da.jpg", + "image_caption": [ + "Figure 1. An example to demonstrate the superiority of VRetouchEr over the main competing methods. (Upper row) The quantitative results of the methods in terms of PSNR. (Middle row) Imperfection detection results of BPFRe. (Bottom row) Imperfection detection results of VRetouchEr." + ], + "image_footnote": [], + "bbox": [ + 517, + 476, + 870, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hand, the enhancements should be imperceptible to preserve the subject's natural appearance.", + "bbox": [ + 496, + 671, + 890, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent video enhancement methods [23-25] fall short in the context of face video retouching due to their inability to accurately model and track the movement of facial imperfections over time. For instance, the optical flow technique utilized in ProPainter [51] lacks the necessary precision for facial flaw localization. The image-to-image translation methods [21, 33, 48] and GAN-based methods [17, 45, 47] are designed for static image retouching, and fall short in maintaining the performance and stability when addressing image sequences. The deficiency of paired training data further impedes the progress in this domain. Hence, there is a significant gap in current methodologies to perform face video retouching, underscoring the need for handling the", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9141", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal and spatial dynamics of facial imperfections in video streams.", + "bbox": [ + 78, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address the limitations of existing retouching methods by proposing a face Video Retouching transformEr with facial imperfection flow-based multiframe attention, which is referred to as VRetouchEr. The core idea of our framework is to incorporate an imperfection flow module to estimate the displacement vectors of imperfections between consecutive frames. Conditioned on the estimation, we can correct the imperfection detection results in each frame via adaptive modulation at specific pixel locations. It is worth noting that the resulting motion information plays an important role in localizing facial imperfection across frames in a stable and precise way. Further, we design a multi-frame masked attention mechanism to synthesize high-fidelity content in the imperfection regions. To achieve this, we allow long-distance interactions between the facial regions from different frames, since multiple frames provide richer information of normal skin characteristics. On the other hand, the estimated imperfection maps are used to weight the intermediate features, such that the imperfections are suppressed. As a result, the attention mechanism enables stable and precise retouching across frames. In summary, the main contribution of this work are as follows: (a) We propose VRetouchEr, a novel face video retouching approach that addresses the challenging of imperfection localization and removal over image sequences. (b) By estimating imperfection flow to correct the imperfection localization in each frame, the obtained spatial information becomes more stable and reliable, which is beneficial to stabilizing the retouching performance. (c) By performing multi-frame masked attention computation, VRetouchEr is able to leverage the contextual information from different frames and synthesize more precise retouching results than existing methods operating on single frame.", + "bbox": [ + 80, + 121, + 467, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 638, + 215, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-to-Image Translation", + "text_level": 1, + "bbox": [ + 80, + 665, + 326, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-image translation aims to learn a mapping across domains of visually distinguishable images, while preserving the content representation to a certain extent [19, 27, 29, 34]. Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) [14] have led to significant progress in this domain. As a representative GAN-based method, Pix2Pix [17] aimed at learning a mapping to minimize the pixel-wise discrepancy. Imposing cycle consistency regularization on the original and the reconstructed images, as in CycleGAN [52] and DiscGAN [19], is also observed to be effective. A multi-stage method MPRNet [48] leveraged the high-level contextual information and spatial details to improve synthesis results. On the other hand, GPEN [47] combined a U-shaped CNN", + "bbox": [ + 80, + 689, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with a GAN, in which the generative prior was useful for high-quality image generation. To perform effective translation among multiple domains, StarGAN [6,7] learnt a unified translation framework, which was conditional on the domain label. Additionally, the latent space of a pre-trained GAN was observed to possess semantic organization, which allowed semantic editing on images [8, 13, 18]. By leveraging GAN inversion methods [15, 32, 36] to project images back to the latent space, attribute manipulation was performed by imposing the attribute-associated transformations on the resulting latent vectors. To address the multiframe setting, burst image restoration methods [3, 43] utilized cross-frame cues to merge burst shot frames, which ultimately led to a significant improvement in restoration performance. In addition, BIPNet [11] distilled and aligned frame-wise features through an edge-focused alignment module, and then refined the resolution progressively.", + "bbox": [ + 503, + 90, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision Transformer", + "text_level": 1, + "bbox": [ + 503, + 359, + 684, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Vision Transformer (ViT) architecture [10, 44, 46] has emerged as a powerful framework, drawing inspiration from transformer research in natural language processing (NLP) [9, 37]. By incorporating the multi-head attention mechanism, positional encoding and position-wise feedforward networks, ViT was able to capture the relationships among input features. ViT achieved significant improvements across multiple vision tasks, such as object detection [4, 53], semantic segmentation [39, 50] and imaged-to-image translation [5, 16]. Building on ViT, the SwinTransformer [30] improved the original design by arranging the image data in a more organized manner, leading to better results in computer vision tasks. For image synthesis, VQGAN [12] combined the transformer's ability to capture complex relationships with the image generation capability of GANs, resulting in increased realism of images generated from text descriptions. RestoreFormer [42] learnt a dictionary in terms of key-value pairs from high-quality data to restore degraded images. Transformers have also shown to be effective for video inpainting [24, 51] and restoration [25, 26]. ProPainter [51] incorporated optical flow into an inpainting module, and a dual-domain feature propagation approach was adopted to synthesize the content in the masked regions.", + "bbox": [ + 503, + 383, + 890, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Face Retouching", + "text_level": 1, + "bbox": [ + 503, + 757, + 661, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face retouching is a specialized task in the domain of image translation, which focuses on the removal of imperfections from human faces in images or videos, while preserving facial attributes as much as possible. Traditional retouching approaches [1,2,22,28,38] typically adopt a global smoothing strategy for the entire face region, which may result in the loss of fine-grained details. AutoRetouch [33] adapted a GAN-based framework to face retouching, and", + "bbox": [ + 503, + 781, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9142", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f9c1011a3e6d1d795d1f199673e12724b30ebc7adb8d7e3da398392fd27d309e.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed VRetouchEr's structure. An Encoder-Transformer-Decoder architecture is specifically designed for Face Video Retouching. VRetouchEr leverages two key modules: the Flow-based Imperfection Refinement (FIR) module and the Multiframe Masked Attention (MMA) mechanism. FIR performs imperfection flow estimation, and further utilizes the displacement information to refines the imperfection localization on each frame. By injecting the imperfection information into the latent transformation, MMA leads to more accurate and consistent retouching outcomes throughout the frame sequence." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 433, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2127569e151ab03704d77eb29602211b702cd80445039bd1e56e83122e95160.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 89, + 890, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4fbde740d364e426a4fa86f1ef87ea83717e6fe70082d7447da50fecd17536b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 202, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the encoder-decoder-based generator is induced to synthesize clean face images in an adversarial training process. To apply more attention on the local regions, ABPN [21] incorporated context-aware local retouching layers together with an adaptive blend pyramid layer for coarse-to-fine retouching. Similarly, BPFRe [45] performed imperfection detections followed by a two-stage retouching process, in which the intermediate features of a U-net were injected as side information into a StyleGAN generator to perform conditional clean face image generation.", + "bbox": [ + 75, + 429, + 468, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are fundamental differences between the proposed VRetouchEr and the above existing methods. First, this work focuses on face retouching in videos, while the video enhancement methods, like ProPainter, fall shot in this task due to lack of imperfection localization. Second, the state-of-the-art model, BPFRe, is designed for single-image retouching, thus its performance is unstable when applied to image sequences. To address the limitations, VRetouchEr performs imperfection flow estimation to improve the localization in terms of both stability and precision. Further, we adopt a task-specific transformer architecture by designing multi-frame masked attention mechanism, such that the long-distance relationship between facial features can be effectively captured across frames.", + "bbox": [ + 75, + 580, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 806, + 210, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 830, + 187, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on the design of the proposed approach, VRetouchEr, for high-quality face retouching in videos. Given a raw image sequence $X$ and the cor", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "responding retouched one $Y$ , our goal is to precisely locate the imperfections in $X$ and replace with the content consistent with the surrounding normal skin. The key idea behind VRetouchEr is to leverage the temporal context information to stabilize face retouching over video frames. To achieve this, we adopt an encoder-transformer-decoder based architecture for VRetouchEr as illustrated in Figure 2. An encoder $E$ extracts features of the target frame $X_{t}$ and multiple reference frames $X_{r}$ in sequence $X$ , and a flow estimation network $S$ is incorporated to model the motions of imperfections between consecutive frames. The estimated imperfection flow $O$ is fed into a Flow-based Imperfection Localization (FIR) module to improve the estimation stability and precision over the sequence. The resulting imperfection map $M$ , together with extracted target feature $f_{t}$ and reference feature $f_{r}$ , are fed into the latent transformer denoted by $T$ to restore the clean face features. In each intermediate block, we can collect richer information of normal skin from the reference frames, and our Multi-frame Masked Attention (MMA) mechanism enables the model to concentrate on the most relevant parts. Finally, the decoder $G$ rebuilds each target frame from the transformed features.", + "bbox": [ + 496, + 429, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Flow-based Imperfection Refinement", + "text_level": 1, + "bbox": [ + 500, + 771, + 821, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The FIR module is specifically designed to enhance the accuracy of imperfection localization by leveraging flow information. This process involves imperfection movement estimation, imperfection localization, and imperfection map refinement. The refinement is crucial for stable retouching in each target frame. To achieve this, we define the imperfections are the differences between the manually retouched", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9143", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frames and raw frames, represented as $M_{gt}$ . We apply a pretrained SpyNet [31] on $M_{gt}$ to produce the ground truth imperfection flow, denoted as $O_{gt}$ .", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our designed FIR module consists of an imperfection localization network $N$ and learnable align factors $\\alpha, \\beta$ . Given frame $X^i$ and its consecutive frame $X^j$ , the flow estimation network $S$ is trained to predict an imperfection displacement map $O^{i \\to j}$ , which is required to approximate the corresponding ground truth imperfection flow as accurately as possible. We evaluate the imperfection flow prediction by measuring the degree of consistency as follows:", + "bbox": [ + 75, + 136, + 467, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO ^ {i \\rightarrow j} = S \\left(X ^ {i}, X ^ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 267, + 468, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f l o w} = \\mathbb {E} _ {X} \\left[ | O ^ {i \\rightarrow j} - O _ {g t} ^ {i \\rightarrow j} | _ {1} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 289, + 385, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By minimizing $\\mathcal{L}_{flow}$ , $S$ is encouraged to focus more on the displacement of imperfections instead of background and normal skin regions between consecutive frames.", + "bbox": [ + 75, + 318, + 468, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, the imperfection localization network $N$ is trained to detect imperfections from frame $X^i$ together with the corresponding encoder feature $f_r^i$ . We observe that the detection results on individual frames are unstable when there are significant changes in facial postures. To address this issue, we utilize the information from consecutive frames and the estimated imperfection flow $O^{i \\to j}$ to refine the detection results. Specifically, the imperfection flow $O^{i \\to j}$ will warp $M^i$ into a flow-based imperfection map $\\mathcal{M}^j$ . Let $\\mathcal{W}(\\cdot, \\cdot)$ denotes the warping operation, and we have:", + "bbox": [ + 75, + 363, + 468, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {i} = N (X ^ {i}, f _ {r} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 525, + 320, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {j} = N \\left(X ^ {j}, f _ {r} ^ {j}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 546, + 468, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} ^ {j} = \\mathcal {W} (O ^ {i \\rightarrow j}, M ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 566, + 351, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}^j$ denotes the warped imperfection map. Next, our designed align factors $\\alpha$ and $\\beta$ are employed to align the flow-based imperfection map $\\mathcal{M}^j$ with predicted imperfection map $M^j$ . We utilized adaptive convolution blocks [35], denoted as $\\theta (\\cdot ,\\cdot)$ , for an effective fusion of these features. The refinement operation is defined as follows:", + "bbox": [ + 75, + 592, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {a} ^ {j} = \\theta (\\alpha * \\mathcal {M} ^ {j} + \\beta , \\mathcal {M} ^ {j}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 693, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} ^ {j} = \\sigma \\left(\\theta \\left(M ^ {j}, \\mathcal {M} _ {a} ^ {j}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 715, + 336, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{M}^j$ represents the refined imperfection map, $\\mathcal{M}_a^j$ is the aligned map and $\\sigma$ represents the sigmoid function. All the refined imperfection maps will be stacked together to form the output, denoted as $M$ . The elements in $M$ are normalized in the range from 0 to 1, representing the refined estimation of the imperfections across frames.", + "bbox": [ + 75, + 746, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-frame Masked Attention Transformer", + "text_level": 1, + "bbox": [ + 76, + 847, + 452, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed MMA mechanism enables the latent transformer to modify the features $f_{t}$ in the imperfection regions", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by leveraging the features $f_{r}$ extracted from multiple frames as references. To achieve this, we use the obtained imperfection map $M$ to weight the features of target and reference frames, and perform cross-attention computation between them. Let the number of reference frames be $\\delta$ , the weighted target feature is used as query, and the weighted reference ones are used as keys and values, defined as follows:", + "bbox": [ + 496, + 90, + 890, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} = W _ {q} \\left(f _ {t} \\otimes M _ {t} + b _ {q}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 210, + 761, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {r} ^ {i} = W _ {k} \\left(f _ {r} ^ {i} \\otimes \\left(1 - M _ {r} ^ {i}\\right) + b _ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 228, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {r} ^ {i} = W _ {v} \\left(f _ {r} ^ {i} \\otimes (1 - M _ {r} ^ {i}) + b _ {v}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 250, + 808, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{t}$ and $M_{r}$ represent the refined imperfection map for target frame and reference frames correspondingly. Next, we perform the cross-attention computation to obtain the modification maps as follows:", + "bbox": [ + 496, + 273, + 890, + 334 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {f _ {t}} = \\operatorname {s o f t m a x} \\left(Q _ {t} \\cdot \\sum_ {i} ^ {\\delta} K _ {r} ^ {i} / \\sqrt {\\Lambda}\\right) \\cdot \\sum_ {i} ^ {\\delta} V _ {r} ^ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 347, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Lambda$ denotes the channel number of the features. In Eq.(4), we adopt the imperfection map $M$ as a soft mask to isolate the regions to be edited, $M$ and $(1 - M)$ represents the mask for imperfection regions and normal regions respectively. Based on this, we can realize the goal of replacing imperfection regions with normal skin by strengthening the interactions between the target features associated with imperfections and the reference features associated with normal skin. Furthermore, we use additive attention in Eq.(5) with linear complexity to aggregate the key and value vectors from multiple reference frame into global vectors, which is used to perform element-wise multiplication to induce temporal context information.", + "bbox": [ + 496, + 398, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that the background and normal skin in the target frame should be preserved properly, the original feature maps are weighted combined with the modification maps as follows:", + "bbox": [ + 496, + 595, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {t} = f _ {t} \\otimes (1 - M _ {t}) + \\Delta_ {f _ {t}} \\otimes M _ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 667, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{f}_t$ denotes the retouched feature, can be fed into the subsequent transformer block as input target feature for further refinement. In this process, the masked attention mechanism enables effective information exchange across frames, and the imperfections in the target frame are progressively suppressed and replaced with the content learnt from normal skin.", + "bbox": [ + 496, + 699, + 890, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 500, + 816, + 656, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training loss functions of VRetouchEr involves the following three aspects: imperfection flow estimation, imperfection localization and retouching evaluation. As illustrated in Subsection 3.2, we employ a pretrained SpyNet", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9144", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to produce the ground truth of imperfection flow estimation, and the network $S$ takes consecutive frames as input and predicts the flow as accurately as possible. In addition, the flow is used to improve the imperfection localization on each frame, and we evaluate the precision by measuring the discrepancy between the prediction $M$ and the ground truth imperfections $M_{gt}$ as follows:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m p} = \\mathbb {E} _ {X} \\| \\mathcal {I} (M) - M _ {g t} \\| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 204, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ denotes a learnable layer to align $M$ 's channel number with $M_{gt}$ . We optimize $N$ and $S$ by minimizing $\\mathcal{L}_{imp}$ together with the flow estimation loss $\\mathcal{L}_{flow}$ defined in Eq.(2), and the formulation is expressed as follows:", + "bbox": [ + 76, + 227, + 468, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {S, N} \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {i m p}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 296, + 468, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\hat{y} = G(\\hat{f}_t)$ denote the retouched frame synthesized by VRetouchEr. We perform retouching on every raw frame and combine the results together to get the retouched sequence $\\hat{Y}$ , which is evaluated by measuring the degree of consistency with the manually retouched frames $Y$ as follows:", + "bbox": [ + 75, + 328, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = \\mathbb {E} _ {X} \\left[ \\zeta \\| Y - \\widehat {Y} | _ {1} + \\| \\mathcal {V} (Y) - \\mathcal {V} (\\widehat {Y}) \\| _ {2} ^ {2} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 425, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\zeta$ denotes a weighting factor, and $\\mathcal{V}(\\cdot)$ represents the features extracted from a pre-trained VGG-19. Considering that high-fidelity content synthesis can benefit from adversarial training, we thus adopt the real-fake discrimination loss formulated as follows:", + "bbox": [ + 75, + 450, + 468, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {s y n t} = \\mathbb {E} _ {X} \\bigl [ \\log (D (\\hat {Y})) \\bigr ]\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 531, + 349, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {r e a l} = \\mathbb {E} _ {X} [ \\log (1 - D (X)) ] + \\mathbb {E} _ {Y} [ \\log (D (Y)) ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 550, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where discriminator $D$ takes the sequence as input, and is trained to distinguish the manually retouched frames from the synthesized ones. By integrating the loss functions of retouching evaluation and adversarial training, the optimization formulation of the constituent networks is expressed as follows:", + "bbox": [ + 75, + 575, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E T C} \\mathcal {L} _ {c o n} + \\mathcal {L} _ {a d v} ^ {s y n},\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 671, + 339, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE, I, G \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 688, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {D} \\mathcal {L} _ {a d v} ^ {r e a l}.\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 699, + 290, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The constituent networks of the proposed VRetouchEr are jointly optimized from scratch.", + "bbox": [ + 75, + 727, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 768, + 200, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we perform extensive experiments to assess the effectiveness of VRetouchEr in face video retouching. We first present the details of the datasets and the experimental setup. Next, we compare our VRetouchEr with previous state-of-the-art face retouching methods quantitatively and qualitatively. Finally, we provide insights by analyzing the design elements of VRetouchEr.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 90, + 707, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Due to the difficulty in collecting paired raw-retouched image sequences, we utilized the Flickr-Face-HQ-Retouching (FFHQR) dataset [33] to construct emulated video data. The training/validation/test data in FFHQR consists of $56\\mathrm{k} / 7\\mathrm{k} / 7\\mathrm{k}$ raw-retouched image pairs. For each image pair $\\{x,y\\}$ in FFHQR, we perform random cropping, flipping and translation on both $x$ and $y$ to generate a pair of image sequences $\\{X,Y\\}$ , and the resulting dataset is referred to FFHQR-Seq. In addition, we build another Manually Retouching Face Video dataset, referred to as MRFV, which contains 200 in-the-wild portrait videos with different types of facial imperfections, and each of them contains at least 500 frames. We employ multiple retouchers to manually retouch each frame of these videos, and evaluate the retouching performance of our VRetouchEr and the competing methods trained on FFHQR-Seq.", + "bbox": [ + 496, + 112, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. The images in both training and test sequences are resized to $512 \\times 512$ for a fair comparison with the existing face retouching methods. In the training process, the parameters of VRetouchEr are updated by the Adam optimizer [20] with the learning rate of $2 \\times 10^{-4}$ . There are a total of 400k training iterations, the batch size is set to 1, and the hyper-parameter $\\zeta$ in Eq.(9) is set to 10. We implement VRetouchEr by using PyTorch and train it on single NVIDIA GeForce GTX 3090 GPU.", + "bbox": [ + 496, + 369, + 890, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. In the experiments, we implement all the competing methods by using the open source codes. We adopt the Soft Intersection over Union Loss (Soft-IoU) to measure the correctness of imperfection localization. To measure the consistency between the synthesized image sequences and the manually retouched ones, we adopt the widely used metrics: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index Measure (SSIM) and the Learnt Perceptual Image Patch Similarity (LPIPS) [49]. Furthermore, we quantitatively assess the diversity and the degree of realism of synthesized image sequences in terms of Video Fréchet Inception Distance (VFID) [40].", + "bbox": [ + 496, + 506, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison to State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 696, + 781, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform a comprehensive comparative analysis between VRetouchEr and representative state-of-the-art methods, including a typical image translation method: Pix2PixHD [41], generic image restoration methods: MPR-Net [48] and RestoreFormer [42], a blind face restoration method: GPEN [47], video enhancement methods: BIP-Net [11] and ProPainter [51], and face retouching methods: AutoRetouch [33] and BPFRe [45]. We train BIP-Net, ProPainter and VRetouchEr on FFHQR-Seq, and the remaining methods on FFHQR. All the trained models are evaluated on the test data of FFHQR-Seq and MRFV. We quantitatively assess the retouch performance of the meth", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9145", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dff76043d0a875577c4b5b2f37f8af051cea56a1584aebee61ae36d072a9f0fc.jpg", + "image_caption": [ + "Figure 3. Visual comparison between VRetouchEr and the competing methods on representative frames of an in-the-wild videos." + ], + "image_footnote": [], + "bbox": [ + 122, + 93, + 846, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0981a464a5cfa44e197b48809a73ebfe4172391a1cac6211455321be4496014.jpg", + "image_caption": [ + "Figure 4. Representative high-quality retouching results. From top left to bottom right, the images are the raw frame and the retouching results of BPFRe, ProPainter and VRetouchEr, respectively." + ], + "image_footnote": [], + "bbox": [ + 117, + 340, + 424, + 578 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d2d7c0a17486317ea8fc30cbffc9c2594c230cdda59580229bd083abebd6e24b.jpg", + "table_caption": [ + "Table 1. Quantitative comparison between VRetouchEr and competing methods on FFHQR-Seq." + ], + "table_footnote": [], + "table_body": "
MethodsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs37.500.97040.033710.779
Pix2PixHD [41]37.630.97140.03039.252
AutoRetouch [33]38.270.97710.02618.941
MPRNet [48]38.330.97420.02678.231
GPEN [47]38.270.97390.02568.125
RestoreFormer [42]38.460.97330.02327.931
BIPNet [11]38.140.97110.02758.241
ProPainter [51]38.540.97680.02407.872
BPFRe [45]38.690.97740.02197.604
VRetouchEr39.750.98130.01696.375
", + "bbox": [ + 81, + 664, + 468, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods per frame, and report their average PSNR, SSIM, LPIPS and VFID scores in Tables 1-2. On FFHQR-Seq, the com", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7f120d226b53e52f94d061377a0e90e3427eb528b9c0fdc64eb7d064291ee659.jpg", + "table_caption": [ + "Table 2. Quantitative comparison between VRetouchEr and competing methods on MRFV." + ], + "table_footnote": [], + "table_body": "
MethodsMRFV
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs32.070.90540.089141.830
Pix2PixHD [41]33.510.91280.073938.403
AutoRetouch [33]35.030.92110.050231.244
MPRNet [48]35.820.93270.049129.893
GPEN [47]35.710.92810.046330.135
RestoreFormer [42]35.760.92990.051629.785
BIPNet [11]35.260.92310.059333.897
ProPainter [51]36.270.93840.045220.344
BPFRe [45]36.320.94270.040118.772
VRetouchEr37.630.95300.035710.368
", + "bbox": [ + 504, + 369, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "peting methods achieve similar retouching performance except Pix2PixHD, and our VRetouchEr outperforms them in terms of all the metrics. In particular, VRetouchEr achieves the PSNR score of 39.75, which is higher than that of the second best method: BPFRe, by $1.06\\mathrm{dB}$ . On MRFV, VRetouchEr also achieves superior retouching performance over the competing methods. The VFID score of VRetouchEr reaches 10.368, which is lower than that of BPFRe by a significant improvement of about 44.77 percentage points.", + "bbox": [ + 496, + 582, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 3, we perform the detailed visual comparison. We can make the following observations: Most of the competing methods fail to remove the marked acnes. BPFRe falls short in maintaining the stability of retouching in videos, since it is designed for single-image retouching. Although ProPainter is specifically designed for video enhancement, it fails to achieve satisfactory retouching results due to lack of imperfection localization. In contrast, our VRetouchEr is able to consistently deliver stable and high-quality retouching outcomes. In Figure 4, we provide visual examples to further compare the performance of our framework with the main competing methods in removing dense", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9146", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/768685824ca83c75b6e039eed6ef525f29b4a191b37078ee167cc0a4ab879c21.jpg", + "image_caption": [ + "Figure 5. Quantitative comparison in imperfection localization per frame. VRetoucher achieves a more precise and stable result than 'VRetoucher w/o FIR' and BPFRe." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 446, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d698e74db156da10dc7fb0d9db6f63e3798116c5781b36c77f302971418a49a.jpg", + "image_caption": [ + "Figure 6. Quantitative comparison in imperfection localization on MRFV dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 279, + 444, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "imperfections. These results highlight the VRetouchEr's capability of face retouching in videos even without any manually retouched videos as training data.", + "bbox": [ + 76, + 460, + 467, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 76, + 515, + 199, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further perform a user study to evaluate VRetouchEr and the competing methods in terms of human perception. The models trained on FFHQR/FFHQR-Seq are applied to a total of 40 raw videos randomly sampled from MRFV. There are 50 participants, each of which is presented with the synthesized videos and required to carefully compare and sort the retouching results, according to the criteria including overall visual quality, effectiveness of imperfection removal, naturalness, and consistency. Table 3 shows the ranking results, and we are able to draw the conclusion regarding the superior performance of VRetouchEr against the competing methods.", + "bbox": [ + 75, + 537, + 468, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Flow-based Imperfection Refinement", + "text_level": 1, + "bbox": [ + 76, + 729, + 398, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observed that our FIR module plays in an important role in facilitating face retouching in videos. Different from", + "bbox": [ + 76, + 752, + 467, + 782 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1bfe29063c0e2a525ff5e75ec4a192190deb7accbba992f3b0c9846b482f8476.jpg", + "table_caption": [ + "Table 3. The voting result $(\\%)$ of user study on MRFV." + ], + "table_footnote": [], + "table_body": "
MethodRank-1Rank-2Rank-3Rank-4Rank-5Rank-6
VRetouchEr88.4511.050.300.200.000.00
BPFRe [45]11.1563.1519.355.051.200.10
ProPainter [51]0.2518.2061.6016.303.150.50
MPRNet [48]0.056.0511.6559.256.4016.60
RestoreFormer [42]0.051.506.1011.2561.3019.80
BIPNet [11]0.050.051.007.9527.9563.00
", + "bbox": [ + 84, + 813, + 465, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f923883ed0edd53d967130d676c238c229a8a2f78de5df5531ab5c3eb8bf81cf.jpg", + "image_caption": [ + "Figure 7. Visual comparison in imperfection localization." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 867, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BPFRe detecting facial imperfections on each frame separately, we estimate imperfection flow between consecutive frames followed by imperfection refinement. The displacement information is useful for refining and stabilizing the imperfection maps of the video to be edited. In Figure 5, we quantitatively evaluate imperfection localization results on a video sequence in terms of Soft-IoU per-frame between the results and the ground truth imperfection maps $M_{gt}$ as defined in Sec. 3.2. In Figure 6, we perform the evaluation of imperfection localization between the results and $M_{gt}$ in terms of mean SSIM and mean Soft-IoU on MRFV dataset. When disabling FIR, the variant 'VRetouchEr w/o FIR' performs single-image imperfection localization, and has comparable performance with BPFRe. We can find that FIR leads to a significant and stable improvement for cross-frame imperfection localization. We further visualize the imperfection maps of BPFRe, 'VRetouchEr w/o FIR' and VRetouchEr in Figure 7, and observe that VRetouchEr can stably locate the most imperfections, and its results are more consistent with the ground truth.", + "bbox": [ + 496, + 287, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Impact of the Number of Reference Frames", + "text_level": 1, + "bbox": [ + 500, + 604, + 869, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The existing face retouching methods suffer from the problem of overlooking temporal context information in handling videos. As another important module of VRetouchEr, we consider that our designed MMA mechanism also contributes to the superior performance of VRetouchEr. To verify this, we investigate the impact of the number of reference frames $\\delta$ on the final retouching performance. In Figure 8, we plot the VFID scores of VRetouchEr on MRFV during training. For the case of $\\delta = 1$ , we perform single-image retouching on each frame. One can observe that increasing the value of $\\delta$ for MMA indeed leads to a significant improvement. In particular, VRetouchEr achieves the best retouching performance when $\\delta = 6$ . Beyond this value, VRetouchEr's performance becomes stable. In Figure 9, we visualize the retouching results to further verify this observation. This insight provides a valuable guidance for optimizing the performance of VRetouchEr in the face video retouching task.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9147", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd8342aa4ac1cff3c1828e35e6372dae4713f278ba94a75ecc0747995fa32a09.jpg", + "image_caption": [ + "Figure 8. Quantitative results of VRetouchEr with different $\\delta$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 467, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7d1125df38e4cca0c9f3c13def2b8bf059f55384a15a4809e44dc10c12d91f8.jpg", + "image_caption": [ + "Figure 9. The retouching results of VRetouchEr with different $\\delta$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 287, + 468, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 556, + 228, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The two proposed modules: FIR and MMA, distinguish the proposed VRetouchEr from the existing works. We perform a set of ablative experiments to investigate the impact of FIR and MMA on the retouching performance on FFHQR-Seq. Specifically, we build the Base model of VRetouchEr by disabling both FIR and MMA, and find that VRetouchEr significantly outperforms the base model as shown in Table 4. Additionally, we build two variants by disabling FIR and MMA, and the resulting models are referred to as 'VRetouchEr w/o FIR' and 'VRetouchEr w/o MMA', respectively. In 'VRetouchEr w/o FIR', the maps produced by the imperfection localization network $N$ are directly fed into the latent transformer. As analyzed in Sec.4.4, disabling FIR leads to unstable imperfection localization, which ultimately degrades the retouching performance on videos. As shown in Table 4, the PSNR score of 'VRetouchEr w/o FIR' is 38.74 dB, which is worse than that of VRetouchEr by 1.01 dB. When disabling MMA, we also observe a significant performance drop of about 40.83 percentage points in terms of LPIPS. Qualitative results in Figure 10 also prove the effectiveness of FIR and MMA.", + "bbox": [ + 75, + 583, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ddcacb34fb806dfab4e41899ce9812108ff4ffedd8a8c28747a5e5b18b50bf32.jpg", + "table_caption": [ + "Table 4. Quantitative results of ablative models." + ], + "table_footnote": [], + "table_body": "
VariantsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Base model38.130.97380.02648.671
VRetouchEr w/o MMA38.670.97580.02387.534
VRetouchEr w/o FIR38.740.97690.02077.312
VRetouchEr39.750.98130.01696.375
", + "bbox": [ + 504, + 99, + 893, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aaa3a5f66eb10092b82d621a4803456069bc5761967082c2b45b47226a3a7188.jpg", + "image_caption": [ + "Figure 10. Representative retouching results of ablative models." + ], + "image_footnote": [], + "bbox": [ + 500, + 215, + 890, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 436, + 617, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a latent retouching transformer for removing facial imperfections in videos. Different from the existing methods that focus on single-image retouching, we leverage the temporal context information to significantly facilitate face retouching in videos. Toward this end, we perform imperfection flow estimation to obtain the displacement information between consecutive frames, and further refine the imperfection localization by fusing motion-based and frame-based predictions. By injecting the imperfection maps as side information into a latent retouching transformer, we adopt the masked attention computation over multiple frames, such that the features of normal skin from different frames can be leveraged to substitute for that of imperfections in the target frame. Extensive experiments are performed to demonstrate the superior capability of the proposed approach over state-of-the-art face retouching methods in stabilizing the retouching performance over video frames.", + "bbox": [ + 496, + 462, + 890, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 748, + 678, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2020A1515010484, 2022A1515011160), and in part by TCL Science and Technology Innovation Fund (Project No. 20231752).", + "bbox": [ + 496, + 773, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9148", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kaoru Arakawa. Nonlinear digital filters for beautifying facial images in multimedia systems. In Proc. IEEE International Symposium on Curcuits and Systems, 2004. 2", + "[2] Nazre Batool and Rama Chellappa. Detection and inpainting of facial wrinkles using texture orientation fields and Markov random field modeling. IEEE Transactions on Image Processing, 23(9):3773-3788, 2014. 2", + "[3] Goutam Bhat, Martin Danelljan, Luc van Gool, and Radu Timofte. Deep burst super-resolution. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2021. 2", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. European Conference on Computer Vision, 2020. 2", + "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[6] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 2", + "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 2", + "[8] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 2", + "[9] Jocob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In arXiv:1801.04805, 2018. 2", + "[10] Alexy Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2", + "[11] Akshay Dudhane, Syed Waqas Zamir, Salman Khan1, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burst image restoration and enhancement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 2, 5, 6, 7", + "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[13] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive im" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age properties. In Proc. International Conference on Computer Vision, 2019. 2", + "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 2", + "[15] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 2", + "[16] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[17] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2", + "[18] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 2", + "[19] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 2", + "[20] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5", + "[21] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: Adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 3", + "[22] Tommer Leyvand, Daniel Cohen-Or, Gideon Dror, and Dani Lischinski. Data-driven enhancement of facial attractiveness. In Proc. ACM Conference on Special Interest Group on Computer Graphics and Interactive Techniques, 2008. 2", + "[23] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In Proc. European Conference on Computer Vision, 2022. 1", + "[24] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo1, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 2", + "[25] Jing Lin, Yuanhao Cai, Xiaowan Hu, Haoqian Wang, and Youliang Yan. Flow-guided sparse transformer for video deblurring. In Proc. International Conference on Machine Learning, 2022. 1, 2", + "[26] Jing Lin, Xiaowan Hu, Yuanhao Cai, and Haoqian Wang. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In Proc. International Conference on Machine Learning, 2022. 2", + "[27] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. EditGAN: high-precision semantic image editing. In Proc. Neural Information Processing Systems, 2021. 2" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9149", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Uri Lipowezky and Sarah Cahen. Automatic freckles detection and retouching. In Proc. IEEE Convention of Electrical and Electronics Engineers in Israel, 2008. 2", + "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2", + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. International Conference on Computer Vision, 2021. 2", + "[31] Anurag RanjanMichael and J. Black. Optical flow estimation using a spatial pyramid network. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2017. 4", + "[32] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGan encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2", + "[33] Alireza Shafaei, James J. Little, and Mark Schmidt. AutoRetouch: automatic professional face retouching. In Proc. IEEE Winter Conference on Applications of Computer Vision, 2021. 1, 2, 5, 6", + "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[35] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 4", + "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems, 2017. 2", + "[38] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. FabSoften: face beautification via dynamic skin smoothing, guided feathering and texture restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition Workshop, 2020. 2", + "[39] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: end-to-end panoptic segmentation with mask transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[40] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2018. 5", + "[41] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. RestoreFormer: high-quality blind face restoration from undegraded key-value pairs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 5, 6, 7", + "[43] Pengxu Wei, Yujing Sun, Xingbei Guo, Chang Liu, Guanbin Li, Jie Chen, Xiangyang Ji, and Liang Lin. Towards real-world burst image super-resolution: Benchmark and method. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2", + "[44] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvR: introducing convolutions to vision transformers. arXiv:2103.15808, 2021. 2", + "[45] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face retouching with limited paired data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 3, 5, 6, 7", + "[46] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[47] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. GAN prior embedded network for blind face restoration in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6", + "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Huang Yang, and Ling Shao. Multi-stage progressive image restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7", + "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 5", + "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2", + "[51] Shangchen Zhou, Chongyi Li, Kelvin C.K. Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 2, 5, 6, 7", + "[52] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 2", + "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 853 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9150", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_model.json b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5636e19d899b3c95659f7c2c6ea8a95d7940dfe2 --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "VRetouchEr: Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.825, + 0.277 + ], + "angle": 0, + "content": "Wen Xue1, Le Jiang1, Lianxin Xie1, Si Wu1,2,3*, Yong Xu1,2,3 and Hau San Wong4 \n1School of Computer Science and Engineering, South China University of Technology \n2Peng Cheng Laboratory 3PAZHOU LAB \n4Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.278, + 0.72, + 0.293 + ], + "angle": 0, + "content": "{csxuewen, csjangle, cslianxin.xie}@mail.scut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.296, + 0.697, + 0.31 + ], + "angle": 0, + "content": "{cswusi, yxu}@scut.edu.cn, cshswong@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.741 + ], + "angle": 0, + "content": "Face Video Retouching is a complex task that often requires labor-intensive manual editing. Conventional image retouching methods perform less satisfactorily in terms of generalization performance and stability when applied to videos without exploiting the correlation among frames. To address this issue, we propose a Video Retouching transformEr to remove facial imperfections in videos, which is referred to as VRetouchEr. Specifically, we estimate the apparent motion of imperfections between two consecutive frames, and the resulting displacement vectors are used to refine the imperfection map, which is synthesized from the current frame together with the corresponding encoder features. The flow-based imperfection refinement is critical for precise and stable retouching across frames. To leverage the temporal contextual information, we inject the refined imperfection map into each transformer block for multi-frame masked attention computation, such that we can capture the interdependence between the current frame and multiple reference frames. As a result, the imperfection regions can be replaced with normal skin with high fidelity, while at the same time keeping the other regions unchanged. Extensive experiments are performed to verify the superiority of VRetouchEr over state-of-the-art image retouching methods in terms of fidelity and stability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.769, + 0.21, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With the development of digital media enhancement, face retouching in videos plays an important role in improving the facial appearance of persons in dynamic sequences. There are a wide range of applications, from professional video production to the burgeoning field of live-streaming and virtual conferencing. These applications require the meticulous removal of visual imperfections. On the other" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.342, + 0.871, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.477, + 0.871, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Figure 1. An example to demonstrate the superiority of VRetouchEr over the main competing methods. (Upper row) The quantitative results of the methods in terms of PSNR. (Middle row) Imperfection detection results of BPFRe. (Bottom row) Imperfection detection results of VRetouchEr." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.703 + ], + "angle": 0, + "content": "hand, the enhancements should be imperceptible to preserve the subject's natural appearance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recent video enhancement methods [23-25] fall short in the context of face video retouching due to their inability to accurately model and track the movement of facial imperfections over time. For instance, the optical flow technique utilized in ProPainter [51] lacks the necessary precision for facial flaw localization. The image-to-image translation methods [21, 33, 48] and GAN-based methods [17, 45, 47] are designed for static image retouching, and fall short in maintaining the performance and stability when addressing image sequences. The deficiency of paired training data further impedes the progress in this domain. Hence, there is a significant gap in current methodologies to perform face video retouching, underscoring the need for handling the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9141" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "temporal and spatial dynamics of facial imperfections in video streams." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.122, + 0.468, + 0.621 + ], + "angle": 0, + "content": "In this work, we address the limitations of existing retouching methods by proposing a face Video Retouching transformEr with facial imperfection flow-based multiframe attention, which is referred to as VRetouchEr. The core idea of our framework is to incorporate an imperfection flow module to estimate the displacement vectors of imperfections between consecutive frames. Conditioned on the estimation, we can correct the imperfection detection results in each frame via adaptive modulation at specific pixel locations. It is worth noting that the resulting motion information plays an important role in localizing facial imperfection across frames in a stable and precise way. Further, we design a multi-frame masked attention mechanism to synthesize high-fidelity content in the imperfection regions. To achieve this, we allow long-distance interactions between the facial regions from different frames, since multiple frames provide richer information of normal skin characteristics. On the other hand, the estimated imperfection maps are used to weight the intermediate features, such that the imperfections are suppressed. As a result, the attention mechanism enables stable and precise retouching across frames. In summary, the main contribution of this work are as follows: (a) We propose VRetouchEr, a novel face video retouching approach that addresses the challenging of imperfection localization and removal over image sequences. (b) By estimating imperfection flow to correct the imperfection localization in each frame, the obtained spatial information becomes more stable and reliable, which is beneficial to stabilizing the retouching performance. (c) By performing multi-frame masked attention computation, VRetouchEr is able to leverage the contextual information from different frames and synthesize more precise retouching results than existing methods operating on single frame." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.64, + 0.217, + 0.654 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.666, + 0.328, + 0.681 + ], + "angle": 0, + "content": "2.1. Image-to-Image Translation" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.69, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Image-to-image translation aims to learn a mapping across domains of visually distinguishable images, while preserving the content representation to a certain extent [19, 27, 29, 34]. Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) [14] have led to significant progress in this domain. As a representative GAN-based method, Pix2Pix [17] aimed at learning a mapping to minimize the pixel-wise discrepancy. Imposing cycle consistency regularization on the original and the reconstructed images, as in CycleGAN [52] and DiscGAN [19], is also observed to be effective. A multi-stage method MPRNet [48] leveraged the high-level contextual information and spatial details to improve synthesis results. On the other hand, GPEN [47] combined a U-shaped CNN" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.348 + ], + "angle": 0, + "content": "with a GAN, in which the generative prior was useful for high-quality image generation. To perform effective translation among multiple domains, StarGAN [6,7] learnt a unified translation framework, which was conditional on the domain label. Additionally, the latent space of a pre-trained GAN was observed to possess semantic organization, which allowed semantic editing on images [8, 13, 18]. By leveraging GAN inversion methods [15, 32, 36] to project images back to the latent space, attribute manipulation was performed by imposing the attribute-associated transformations on the resulting latent vectors. To address the multiframe setting, burst image restoration methods [3, 43] utilized cross-frame cues to merge burst shot frames, which ultimately led to a significant improvement in restoration performance. In addition, BIPNet [11] distilled and aligned frame-wise features through an edge-focused alignment module, and then refined the resolution progressively." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.36, + 0.686, + 0.374 + ], + "angle": 0, + "content": "2.2. Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.384, + 0.891, + 0.744 + ], + "angle": 0, + "content": "The Vision Transformer (ViT) architecture [10, 44, 46] has emerged as a powerful framework, drawing inspiration from transformer research in natural language processing (NLP) [9, 37]. By incorporating the multi-head attention mechanism, positional encoding and position-wise feedforward networks, ViT was able to capture the relationships among input features. ViT achieved significant improvements across multiple vision tasks, such as object detection [4, 53], semantic segmentation [39, 50] and imaged-to-image translation [5, 16]. Building on ViT, the SwinTransformer [30] improved the original design by arranging the image data in a more organized manner, leading to better results in computer vision tasks. For image synthesis, VQGAN [12] combined the transformer's ability to capture complex relationships with the image generation capability of GANs, resulting in increased realism of images generated from text descriptions. RestoreFormer [42] learnt a dictionary in terms of key-value pairs from high-quality data to restore degraded images. Transformers have also shown to be effective for video inpainting [24, 51] and restoration [25, 26]. ProPainter [51] incorporated optical flow into an inpainting module, and a dual-domain feature propagation approach was adopted to synthesize the content in the masked regions." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.758, + 0.663, + 0.772 + ], + "angle": 0, + "content": "2.3. Face Retouching" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.782, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Face retouching is a specialized task in the domain of image translation, which focuses on the removal of imperfections from human faces in images or videos, while preserving facial attributes as much as possible. Traditional retouching approaches [1,2,22,28,38] typically adopt a global smoothing strategy for the entire face region, which may result in the loss of fine-grained details. AutoRetouch [33] adapted a GAN-based framework to face retouching, and" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.434, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.09, + 0.891, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.203, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.894, + 0.407 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed VRetouchEr's structure. An Encoder-Transformer-Decoder architecture is specifically designed for Face Video Retouching. VRetouchEr leverages two key modules: the Flow-based Imperfection Refinement (FIR) module and the Multiframe Masked Attention (MMA) mechanism. FIR performs imperfection flow estimation, and further utilizes the displacement information to refines the imperfection localization on each frame. By injecting the imperfection information into the latent transformation, MMA leads to more accurate and consistent retouching outcomes throughout the frame sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.581 + ], + "angle": 0, + "content": "the encoder-decoder-based generator is induced to synthesize clean face images in an adversarial training process. To apply more attention on the local regions, ABPN [21] incorporated context-aware local retouching layers together with an adaptive blend pyramid layer for coarse-to-fine retouching. Similarly, BPFRe [45] performed imperfection detections followed by a two-stage retouching process, in which the intermediate features of a U-net were injected as side information into a StyleGAN generator to perform conditional clean face image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.793 + ], + "angle": 0, + "content": "There are fundamental differences between the proposed VRetouchEr and the above existing methods. First, this work focuses on face retouching in videos, while the video enhancement methods, like ProPainter, fall shot in this task due to lack of imperfection localization. Second, the state-of-the-art model, BPFRe, is designed for single-image retouching, thus its performance is unstable when applied to image sequences. To address the limitations, VRetouchEr performs imperfection flow estimation to improve the localization in terms of both stability and precision. Further, we adopt a task-specific transformer architecture by designing multi-frame masked attention mechanism, such that the long-distance relationship between facial features can be effectively captured across frames." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.807, + 0.212, + 0.825 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.188, + 0.847 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we elaborate on the design of the proposed approach, VRetouchEr, for high-quality face retouching in videos. Given a raw image sequence \\(X\\) and the cor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.763 + ], + "angle": 0, + "content": "responding retouched one \\( Y \\), our goal is to precisely locate the imperfections in \\( X \\) and replace with the content consistent with the surrounding normal skin. The key idea behind VRetouchEr is to leverage the temporal context information to stabilize face retouching over video frames. To achieve this, we adopt an encoder-transformer-decoder based architecture for VRetouchEr as illustrated in Figure 2. An encoder \\( E \\) extracts features of the target frame \\( X_{t} \\) and multiple reference frames \\( X_{r} \\) in sequence \\( X \\), and a flow estimation network \\( S \\) is incorporated to model the motions of imperfections between consecutive frames. The estimated imperfection flow \\( O \\) is fed into a Flow-based Imperfection Localization (FIR) module to improve the estimation stability and precision over the sequence. The resulting imperfection map \\( M \\), together with extracted target feature \\( f_{t} \\) and reference feature \\( f_{r} \\), are fed into the latent transformer denoted by \\( T \\) to restore the clean face features. In each intermediate block, we can collect richer information of normal skin from the reference frames, and our Multi-frame Masked Attention (MMA) mechanism enables the model to concentrate on the most relevant parts. Finally, the decoder \\( G \\) rebuilds each target frame from the transformed features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.822, + 0.788 + ], + "angle": 0, + "content": "3.2. Flow-based Imperfection Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The FIR module is specifically designed to enhance the accuracy of imperfection localization by leveraging flow information. This process involves imperfection movement estimation, imperfection localization, and imperfection map refinement. The refinement is crucial for stable retouching in each target frame. To achieve this, we define the imperfections are the differences between the manually retouched" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "frames and raw frames, represented as \\( M_{gt} \\). We apply a pretrained SpyNet [31] on \\( M_{gt} \\) to produce the ground truth imperfection flow, denoted as \\( O_{gt} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Our designed FIR module consists of an imperfection localization network \\(N\\) and learnable align factors \\(\\alpha, \\beta\\). Given frame \\(X^i\\) and its consecutive frame \\(X^j\\), the flow estimation network \\(S\\) is trained to predict an imperfection displacement map \\(O^{i \\to j}\\), which is required to approximate the corresponding ground truth imperfection flow as accurately as possible. We evaluate the imperfection flow prediction by measuring the degree of consistency as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.268, + 0.47, + 0.294 + ], + "angle": 0, + "content": "\\[\nO ^ {i \\rightarrow j} = S \\left(X ^ {i}, X ^ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.29, + 0.386, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f l o w} = \\mathbb {E} _ {X} \\left[ | O ^ {i \\rightarrow j} - O _ {g t} ^ {i \\rightarrow j} | _ {1} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.469, + 0.363 + ], + "angle": 0, + "content": "By minimizing \\(\\mathcal{L}_{flow}\\), \\(S\\) is encouraged to focus more on the displacement of imperfections instead of background and normal skin regions between consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.469, + 0.527 + ], + "angle": 0, + "content": "On the other hand, the imperfection localization network \\(N\\) is trained to detect imperfections from frame \\(X^i\\) together with the corresponding encoder feature \\(f_r^i\\). We observe that the detection results on individual frames are unstable when there are significant changes in facial postures. To address this issue, we utilize the information from consecutive frames and the estimated imperfection flow \\(O^{i \\to j}\\) to refine the detection results. Specifically, the imperfection flow \\(O^{i \\to j}\\) will warp \\(M^i\\) into a flow-based imperfection map \\(\\mathcal{M}^j\\). Let \\(\\mathcal{W}(\\cdot, \\cdot)\\) denotes the warping operation, and we have:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.526, + 0.321, + 0.544 + ], + "angle": 0, + "content": "\\[\nM ^ {i} = N (X ^ {i}, f _ {r} ^ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.547, + 0.469, + 0.565 + ], + "angle": 0, + "content": "\\[\nM ^ {j} = N \\left(X ^ {j}, f _ {r} ^ {j}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.567, + 0.352, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} ^ {j} = \\mathcal {W} (O ^ {i \\rightarrow j}, M ^ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.683 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}^j\\) denotes the warped imperfection map. Next, our designed align factors \\(\\alpha\\) and \\(\\beta\\) are employed to align the flow-based imperfection map \\(\\mathcal{M}^j\\) with predicted imperfection map \\(M^j\\). We utilized adaptive convolution blocks [35], denoted as \\(\\theta (\\cdot ,\\cdot)\\), for an effective fusion of these features. The refinement operation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.694, + 0.47, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {a} ^ {j} = \\theta (\\alpha * \\mathcal {M} ^ {j} + \\beta , \\mathcal {M} ^ {j}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.716, + 0.338, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\hat {M} ^ {j} = \\sigma \\left(\\theta \\left(M ^ {j}, \\mathcal {M} _ {a} ^ {j}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.838 + ], + "angle": 0, + "content": "where \\(\\hat{M}^j\\) represents the refined imperfection map, \\(\\mathcal{M}_a^j\\) is the aligned map and \\(\\sigma\\) represents the sigmoid function. All the refined imperfection maps will be stacked together to form the output, denoted as \\(M\\). The elements in \\(M\\) are normalized in the range from 0 to 1, representing the refined estimation of the imperfections across frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.453, + 0.863 + ], + "angle": 0, + "content": "3.3. Multi-frame Masked Attention Transformer" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our proposed MMA mechanism enables the latent transformer to modify the features \\( f_{t} \\) in the imperfection regions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.209 + ], + "angle": 0, + "content": "by leveraging the features \\( f_{r} \\) extracted from multiple frames as references. To achieve this, we use the obtained imperfection map \\( M \\) to weight the features of target and reference frames, and perform cross-attention computation between them. Let the number of reference frames be \\( \\delta \\), the weighted target feature is used as query, and the weighted reference ones are used as keys and values, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.211, + 0.762, + 0.227 + ], + "angle": 0, + "content": "\\[\nQ _ {t} = W _ {q} \\left(f _ {t} \\otimes M _ {t} + b _ {q}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.229, + 0.891, + 0.248 + ], + "angle": 0, + "content": "\\[\nK _ {r} ^ {i} = W _ {k} \\left(f _ {r} ^ {i} \\otimes \\left(1 - M _ {r} ^ {i}\\right) + b _ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.251, + 0.81, + 0.269 + ], + "angle": 0, + "content": "\\[\nV _ {r} ^ {i} = W _ {v} \\left(f _ {r} ^ {i} \\otimes (1 - M _ {r} ^ {i}) + b _ {v}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.335 + ], + "angle": 0, + "content": "where \\(M_{t}\\) and \\(M_{r}\\) represent the refined imperfection map for target frame and reference frames correspondingly. Next, we perform the cross-attention computation to obtain the modification maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.348, + 0.892, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {f _ {t}} = \\operatorname {s o f t m a x} \\left(Q _ {t} \\cdot \\sum_ {i} ^ {\\delta} K _ {r} ^ {i} / \\sqrt {\\Lambda}\\right) \\cdot \\sum_ {i} ^ {\\delta} V _ {r} ^ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.595 + ], + "angle": 0, + "content": "where \\(\\Lambda\\) denotes the channel number of the features. In Eq.(4), we adopt the imperfection map \\(M\\) as a soft mask to isolate the regions to be edited, \\(M\\) and \\((1 - M)\\) represents the mask for imperfection regions and normal regions respectively. Based on this, we can realize the goal of replacing imperfection regions with normal skin by strengthening the interactions between the target features associated with imperfections and the reference features associated with normal skin. Furthermore, we use additive attention in Eq.(5) with linear complexity to aggregate the key and value vectors from multiple reference frame into global vectors, which is used to perform element-wise multiplication to induce temporal context information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Considering that the background and normal skin in the target frame should be preserved properly, the original feature maps are weighted combined with the modification maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.668, + 0.891, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {t} = f _ {t} \\otimes (1 - M _ {t}) + \\Delta_ {f _ {t}} \\otimes M _ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.805 + ], + "angle": 0, + "content": "where \\(\\hat{f}_t\\) denotes the retouched feature, can be fed into the subsequent transformer block as input target feature for further refinement. In this process, the masked attention mechanism enables effective information exchange across frames, and the imperfections in the target frame are progressively suppressed and replaced with the content learnt from normal skin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.658, + 0.833 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The training loss functions of VRetouchEr involves the following three aspects: imperfection flow estimation, imperfection localization and retouching evaluation. As illustrated in Subsection 3.2, we employ a pretrained SpyNet" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9144" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "to produce the ground truth of imperfection flow estimation, and the network \\( S \\) takes consecutive frames as input and predicts the flow as accurately as possible. In addition, the flow is used to improve the imperfection localization on each frame, and we evaluate the precision by measuring the discrepancy between the prediction \\( M \\) and the ground truth imperfections \\( M_{gt} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.205, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m p} = \\mathbb {E} _ {X} \\| \\mathcal {I} (M) - M _ {g t} \\| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.228, + 0.47, + 0.289 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) denotes a learnable layer to align \\(M\\)'s channel number with \\(M_{gt}\\). We optimize \\(N\\) and \\(S\\) by minimizing \\(\\mathcal{L}_{imp}\\) together with the flow estimation loss \\(\\mathcal{L}_{flow}\\) defined in Eq.(2), and the formulation is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.297, + 0.47, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\min _ {S, N} \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {i m p}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.419 + ], + "angle": 0, + "content": "Let \\(\\hat{y} = G(\\hat{f}_t)\\) denote the retouched frame synthesized by VRetouchEr. We perform retouching on every raw frame and combine the results together to get the retouched sequence \\(\\hat{Y}\\), which is evaluated by measuring the degree of consistency with the manually retouched frames \\(Y\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.426, + 0.469, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = \\mathbb {E} _ {X} \\left[ \\zeta \\| Y - \\widehat {Y} | _ {1} + \\| \\mathcal {V} (Y) - \\mathcal {V} (\\widehat {Y}) \\| _ {2} ^ {2} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.526 + ], + "angle": 0, + "content": "where \\(\\zeta\\) denotes a weighting factor, and \\(\\mathcal{V}(\\cdot)\\) represents the features extracted from a pre-trained VGG-19. Considering that high-fidelity content synthesis can benefit from adversarial training, we thus adopt the real-fake discrimination loss formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.532, + 0.35, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {s y n t} = \\mathbb {E} _ {X} \\bigl [ \\log (D (\\hat {Y})) \\bigr ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.551, + 0.47, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {r e a l} = \\mathbb {E} _ {X} [ \\log (1 - D (X)) ] + \\mathbb {E} _ {Y} [ \\log (D (Y)) ], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.666 + ], + "angle": 0, + "content": "where discriminator \\( D \\) takes the sequence as input, and is trained to distinguish the manually retouched frames from the synthesized ones. By integrating the loss functions of retouching evaluation and adversarial training, the optimization formulation of the constituent networks is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.672, + 0.341, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\min _ {E T C} \\mathcal {L} _ {c o n} + \\mathcal {L} _ {a d v} ^ {s y n},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.689, + 0.469, + 0.703 + ], + "angle": 0, + "content": "\\[\nE, I, G \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.7, + 0.291, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\max _ {D} \\mathcal {L} _ {a d v} ^ {r e a l}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.469, + 0.759 + ], + "angle": 0, + "content": "The constituent networks of the proposed VRetouchEr are jointly optimized from scratch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.202, + 0.787 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we perform extensive experiments to assess the effectiveness of VRetouchEr in face video retouching. We first present the details of the datasets and the experimental setup. Next, we compare our VRetouchEr with previous state-of-the-art face retouching methods quantitatively and qualitatively. Finally, we provide insights by analyzing the design elements of VRetouchEr." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.708, + 0.108 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Datasets. Due to the difficulty in collecting paired raw-retouched image sequences, we utilized the Flickr-Face-HQ-Retouching (FFHQR) dataset [33] to construct emulated video data. The training/validation/test data in FFHQR consists of \\(56\\mathrm{k} / 7\\mathrm{k} / 7\\mathrm{k}\\) raw-retouched image pairs. For each image pair \\(\\{x,y\\}\\) in FFHQR, we perform random cropping, flipping and translation on both \\(x\\) and \\(y\\) to generate a pair of image sequences \\(\\{X,Y\\}\\), and the resulting dataset is referred to FFHQR-Seq. In addition, we build another Manually Retouching Face Video dataset, referred to as MRFV, which contains 200 in-the-wild portrait videos with different types of facial imperfections, and each of them contains at least 500 frames. We employ multiple retouchers to manually retouch each frame of these videos, and evaluate the retouching performance of our VRetouchEr and the competing methods trained on FFHQR-Seq." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Training Details. The images in both training and test sequences are resized to \\(512 \\times 512\\) for a fair comparison with the existing face retouching methods. In the training process, the parameters of VRetouchEr are updated by the Adam optimizer [20] with the learning rate of \\(2 \\times 10^{-4}\\). There are a total of 400k training iterations, the batch size is set to 1, and the hyper-parameter \\(\\zeta\\) in Eq.(9) is set to 10. We implement VRetouchEr by using PyTorch and train it on single NVIDIA GeForce GTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Evaluation Protocols. In the experiments, we implement all the competing methods by using the open source codes. We adopt the Soft Intersection over Union Loss (Soft-IoU) to measure the correctness of imperfection localization. To measure the consistency between the synthesized image sequences and the manually retouched ones, we adopt the widely used metrics: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index Measure (SSIM) and the Learnt Perceptual Image Patch Similarity (LPIPS) [49]. Furthermore, we quantitatively assess the diversity and the degree of realism of synthesized image sequences in terms of Video Fréchet Inception Distance (VFID) [40]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.782, + 0.713 + ], + "angle": 0, + "content": "4.2. Comparison to State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We perform a comprehensive comparative analysis between VRetouchEr and representative state-of-the-art methods, including a typical image translation method: Pix2PixHD [41], generic image restoration methods: MPR-Net [48] and RestoreFormer [42], a blind face restoration method: GPEN [47], video enhancement methods: BIP-Net [11] and ProPainter [51], and face retouching methods: AutoRetouch [33] and BPFRe [45]. We train BIP-Net, ProPainter and VRetouchEr on FFHQR-Seq, and the remaining methods on FFHQR. All the trained models are evaluated on the test data of FFHQR-Seq and MRFV. We quantitatively assess the retouch performance of the meth" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.094, + 0.847, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.316, + 0.866, + 0.331 + ], + "angle": 0, + "content": "Figure 3. Visual comparison between VRetouchEr and the competing methods on representative frames of an in-the-wild videos." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.342, + 0.426, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.628 + ], + "angle": 0, + "content": "Figure 4. Representative high-quality retouching results. From top left to bottom right, the images are the raw frame and the retouching results of BPFRe, ProPainter and VRetouchEr, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.636, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between VRetouchEr and competing methods on FFHQR-Seq." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.665, + 0.47, + 0.862 + ], + "angle": 0, + "content": "
MethodsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs37.500.97040.033710.779
Pix2PixHD [41]37.630.97140.03039.252
AutoRetouch [33]38.270.97710.02618.941
MPRNet [48]38.330.97420.02678.231
GPEN [47]38.270.97390.02568.125
RestoreFormer [42]38.460.97330.02327.931
BIPNet [11]38.140.97110.02758.241
ProPainter [51]38.540.97680.02407.872
BPFRe [45]38.690.97740.02197.604
VRetouchEr39.750.98130.01696.375
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "ods per frame, and report their average PSNR, SSIM, LPIPS and VFID scores in Tables 1-2. On FFHQR-Seq, the com" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.341, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between VRetouchEr and competing methods on MRFV." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.371, + 0.892, + 0.567 + ], + "angle": 0, + "content": "
MethodsMRFV
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs32.070.90540.089141.830
Pix2PixHD [41]33.510.91280.073938.403
AutoRetouch [33]35.030.92110.050231.244
MPRNet [48]35.820.93270.049129.893
GPEN [47]35.710.92810.046330.135
RestoreFormer [42]35.760.92990.051629.785
BIPNet [11]35.260.92310.059333.897
ProPainter [51]36.270.93840.045220.344
BPFRe [45]36.320.94270.040118.772
VRetouchEr37.630.95300.035710.368
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.719 + ], + "angle": 0, + "content": "peting methods achieve similar retouching performance except Pix2PixHD, and our VRetouchEr outperforms them in terms of all the metrics. In particular, VRetouchEr achieves the PSNR score of 39.75, which is higher than that of the second best method: BPFRe, by \\(1.06\\mathrm{dB}\\). On MRFV, VRetouchEr also achieves superior retouching performance over the competing methods. The VFID score of VRetouchEr reaches 10.368, which is lower than that of BPFRe by a significant improvement of about 44.77 percentage points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In Figure 3, we perform the detailed visual comparison. We can make the following observations: Most of the competing methods fail to remove the marked acnes. BPFRe falls short in maintaining the stability of retouching in videos, since it is designed for single-image retouching. Although ProPainter is specifically designed for video enhancement, it fails to achieve satisfactory retouching results due to lack of imperfection localization. In contrast, our VRetouchEr is able to consistently deliver stable and high-quality retouching outcomes. In Figure 4, we provide visual examples to further compare the performance of our framework with the main competing methods in removing dense" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.447, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.231, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 5. Quantitative comparison in imperfection localization per frame. VRetoucher achieves a more precise and stable result than 'VRetoucher w/o FIR' and BPFRe." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.28, + 0.446, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.422, + 0.468, + 0.449 + ], + "angle": 0, + "content": "Figure 6. Quantitative comparison in imperfection localization on MRFV dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.468, + 0.507 + ], + "angle": 0, + "content": "imperfections. These results highlight the VRetouchEr's capability of face retouching in videos even without any manually retouched videos as training data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.516, + 0.2, + 0.533 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.721 + ], + "angle": 0, + "content": "We further perform a user study to evaluate VRetouchEr and the competing methods in terms of human perception. The models trained on FFHQR/FFHQR-Seq are applied to a total of 40 raw videos randomly sampled from MRFV. There are 50 participants, each of which is presented with the synthesized videos and required to carefully compare and sort the retouching results, according to the criteria including overall visual quality, effectiveness of imperfection removal, naturalness, and consistency. Table 3 shows the ranking results, and we are able to draw the conclusion regarding the superior performance of VRetouchEr against the competing methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.399, + 0.746 + ], + "angle": 0, + "content": "4.4. Flow-based Imperfection Refinement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.468, + 0.784 + ], + "angle": 0, + "content": "We observed that our FIR module plays in an important role in facilitating face retouching in videos. Different from" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.799, + 0.436, + 0.812 + ], + "angle": 0, + "content": "Table 3. The voting result \\((\\%)\\) of user study on MRFV." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.814, + 0.466, + 0.899 + ], + "angle": 0, + "content": "
MethodRank-1Rank-2Rank-3Rank-4Rank-5Rank-6
VRetouchEr88.4511.050.300.200.000.00
BPFRe [45]11.1563.1519.355.051.200.10
ProPainter [51]0.2518.2061.6016.303.150.50
MPRNet [48]0.056.0511.6559.256.4016.60
RestoreFormer [42]0.051.506.1011.2561.3019.80
BIPNet [11]0.050.051.007.9527.9563.00
" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.092, + 0.869, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.261, + 0.866, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Visual comparison in imperfection localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.591 + ], + "angle": 0, + "content": "BPFRe detecting facial imperfections on each frame separately, we estimate imperfection flow between consecutive frames followed by imperfection refinement. The displacement information is useful for refining and stabilizing the imperfection maps of the video to be edited. In Figure 5, we quantitatively evaluate imperfection localization results on a video sequence in terms of Soft-IoU per-frame between the results and the ground truth imperfection maps \\( M_{gt} \\) as defined in Sec. 3.2. In Figure 6, we perform the evaluation of imperfection localization between the results and \\( M_{gt} \\) in terms of mean SSIM and mean Soft-IoU on MRFV dataset. When disabling FIR, the variant 'VRetouchEr w/o FIR' performs single-image imperfection localization, and has comparable performance with BPFRe. We can find that FIR leads to a significant and stable improvement for cross-frame imperfection localization. We further visualize the imperfection maps of BPFRe, 'VRetouchEr w/o FIR' and VRetouchEr in Figure 7, and observe that VRetouchEr can stably locate the most imperfections, and its results are more consistent with the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.87, + 0.621 + ], + "angle": 0, + "content": "4.5. Impact of the Number of Reference Frames" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The existing face retouching methods suffer from the problem of overlooking temporal context information in handling videos. As another important module of VRetouchEr, we consider that our designed MMA mechanism also contributes to the superior performance of VRetouchEr. To verify this, we investigate the impact of the number of reference frames \\(\\delta\\) on the final retouching performance. In Figure 8, we plot the VFID scores of VRetouchEr on MRFV during training. For the case of \\(\\delta = 1\\), we perform single-image retouching on each frame. One can observe that increasing the value of \\(\\delta\\) for MMA indeed leads to a significant improvement. In particular, VRetouchEr achieves the best retouching performance when \\(\\delta = 6\\). Beyond this value, VRetouchEr's performance becomes stable. In Figure 9, we visualize the retouching results to further verify this observation. This insight provides a valuable guidance for optimizing the performance of VRetouchEr in the face video retouching task." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.468, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.26, + 0.457, + 0.273 + ], + "angle": 0, + "content": "Figure 8. Quantitative results of VRetouchEr with different \\(\\delta\\)." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.289, + 0.47, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.521, + 0.466, + 0.535 + ], + "angle": 0, + "content": "Figure 9. The retouching results of VRetouchEr with different \\(\\delta\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.23, + 0.574 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The two proposed modules: FIR and MMA, distinguish the proposed VRetouchEr from the existing works. We perform a set of ablative experiments to investigate the impact of FIR and MMA on the retouching performance on FFHQR-Seq. Specifically, we build the Base model of VRetouchEr by disabling both FIR and MMA, and find that VRetouchEr significantly outperforms the base model as shown in Table 4. Additionally, we build two variants by disabling FIR and MMA, and the resulting models are referred to as 'VRetouchEr w/o FIR' and 'VRetouchEr w/o MMA', respectively. In 'VRetouchEr w/o FIR', the maps produced by the imperfection localization network \\( N \\) are directly fed into the latent transformer. As analyzed in Sec.4.4, disabling FIR leads to unstable imperfection localization, which ultimately degrades the retouching performance on videos. As shown in Table 4, the PSNR score of 'VRetouchEr w/o FIR' is 38.74 dB, which is worse than that of VRetouchEr by 1.01 dB. When disabling MMA, we also observe a significant performance drop of about 40.83 percentage points in terms of LPIPS. Qualitative results in Figure 10 also prove the effectiveness of FIR and MMA." + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.076, + 0.838, + 0.089 + ], + "angle": 0, + "content": "Table 4. Quantitative results of ablative models." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.1, + 0.894, + 0.204 + ], + "angle": 0, + "content": "
VariantsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Base model38.130.97380.02648.671
VRetouchEr w/o MMA38.670.97580.02387.534
VRetouchEr w/o FIR38.740.97690.02077.312
VRetouchEr39.750.98130.01696.375
" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.216, + 0.892, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.398, + 0.885, + 0.412 + ], + "angle": 0, + "content": "Figure 10. Representative retouching results of ablative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.618, + 0.453 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.734 + ], + "angle": 0, + "content": "In this work, we propose a latent retouching transformer for removing facial imperfections in videos. Different from the existing methods that focus on single-image retouching, we leverage the temporal context information to significantly facilitate face retouching in videos. Toward this end, we perform imperfection flow estimation to obtain the displacement information between consecutive frames, and further refine the imperfection localization by fusing motion-based and frame-based predictions. By injecting the imperfection maps as side information into a latent retouching transformer, we adopt the masked attention computation over multiple frames, such that the features of normal skin from different frames can be leveraged to substitute for that of imperfections in the target frame. Extensive experiments are performed to demonstrate the superior capability of the proposed approach over state-of-the-art face retouching methods in stabilizing the retouching performance over video frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.749, + 0.679, + 0.766 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.774, + 0.892, + 0.895 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2020A1515010484, 2022A1515011160), and in part by TCL Science and Technology Innovation Fund (Project No. 20231752)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9148" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.158 + ], + "angle": 0, + "content": "[1] Kaoru Arakawa. Nonlinear digital filters for beautifying facial images in multimedia systems. In Proc. IEEE International Symposium on Curcuits and Systems, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[2] Nazre Batool and Rama Chellappa. Detection and inpainting of facial wrinkles using texture orientation fields and Markov random field modeling. IEEE Transactions on Image Processing, 23(9):3773-3788, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.271 + ], + "angle": 0, + "content": "[3] Goutam Bhat, Martin Danelljan, Luc van Gool, and Radu Timofte. Deep burst super-resolution. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[6] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[8] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[9] Jocob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In arXiv:1801.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[10] Alexy Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[11] Akshay Dudhane, Syed Waqas Zamir, Salman Khan1, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burst image restoration and enhancement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive im" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "age properties. In Proc. International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[15] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[19] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[20] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: Adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Tommer Leyvand, Daniel Cohen-Or, Gideon Dror, and Dani Lischinski. Data-driven enhancement of facial attractiveness. In Proc. ACM Conference on Special Interest Group on Computer Graphics and Interactive Techniques, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[23] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In Proc. European Conference on Computer Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo1, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Jing Lin, Yuanhao Cai, Xiaowan Hu, Haoqian Wang, and Youliang Yan. Flow-guided sparse transformer for video deblurring. In Proc. International Conference on Machine Learning, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Jing Lin, Xiaowan Hu, Yuanhao Cai, and Haoqian Wang. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In Proc. International Conference on Machine Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. EditGAN: high-precision semantic image editing. In Proc. Neural Information Processing Systems, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9149" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[28] Uri Lipowezky and Sarah Cahen. Automatic freckles detection and retouching. In Proc. IEEE Convention of Electrical and Electronics Engineers in Israel, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[31] Anurag RanjanMichael and J. Black. Optical flow estimation using a spatial pyramid network. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[32] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGan encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[33] Alireza Shafaei, James J. Little, and Mark Schmidt. AutoRetouch: automatic professional face retouching. In Proc. IEEE Winter Conference on Applications of Computer Vision, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[35] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.52, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.577, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.634, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[38] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. FabSoften: face beautification via dynamic skin smoothing, guided feathering and texture restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition Workshop, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[39] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: end-to-end panoptic segmentation with mask transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[40] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[42] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. RestoreFormer: high-quality blind face restoration from undegraded key-value pairs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[43] Pengxu Wei, Yujing Sun, Xingbei Guo, Chang Liu, Guanbin Li, Jie Chen, Xiangyang Ji, and Liang Lin. Towards real-world burst image super-resolution: Benchmark and method. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvR: introducing convolutions to vision transformers. arXiv:2103.15808, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[45] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face retouching with limited paired data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[46] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[47] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. GAN prior embedded network for blind face restoration in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Huang Yang, and Ling Shao. Multi-stage progressive image restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[51] Shangchen Zhou, Chongyi Li, Kelvin C.K. Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[52] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.892, + 0.854 + ], + "angle": 0, + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9150" + } + ] +] \ No newline at end of file diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_origin.pdf b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6b48455cd0af2f7801ef6766bb15db0f179d435 --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/8a896d01-f26d-44df-a67e-710260f2b006_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7d2816a021d855e521ef4fa76bcfb60e469ccb60cfeee28726872bc103816a9 +size 1252461 diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/full.md b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9ef2cb63d04d0e3dc90568b5e7165a1bb966104 --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/full.md @@ -0,0 +1,342 @@ +# VRetouchEr: Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos + +Wen Xue1, Le Jiang1, Lianxin Xie1, Si Wu1,2,3*, Yong Xu1,2,3 and Hau San Wong4 +1School of Computer Science and Engineering, South China University of Technology +2Peng Cheng Laboratory 3PAZHOU LAB +4Department of Computer Science, City University of Hong Kong + +{csxuewen, csjangle, cslianxin.xie}@mail.scut.edu.cn + +{cswusi, yxu}@scut.edu.cn, cshswong@cityu.edu.hk + +# Abstract + +Face Video Retouching is a complex task that often requires labor-intensive manual editing. Conventional image retouching methods perform less satisfactorily in terms of generalization performance and stability when applied to videos without exploiting the correlation among frames. To address this issue, we propose a Video Retouching transformEr to remove facial imperfections in videos, which is referred to as VRetouchEr. Specifically, we estimate the apparent motion of imperfections between two consecutive frames, and the resulting displacement vectors are used to refine the imperfection map, which is synthesized from the current frame together with the corresponding encoder features. The flow-based imperfection refinement is critical for precise and stable retouching across frames. To leverage the temporal contextual information, we inject the refined imperfection map into each transformer block for multi-frame masked attention computation, such that we can capture the interdependence between the current frame and multiple reference frames. As a result, the imperfection regions can be replaced with normal skin with high fidelity, while at the same time keeping the other regions unchanged. Extensive experiments are performed to verify the superiority of VRetouchEr over state-of-the-art image retouching methods in terms of fidelity and stability. + +# 1. Introduction + +With the development of digital media enhancement, face retouching in videos plays an important role in improving the facial appearance of persons in dynamic sequences. There are a wide range of applications, from professional video production to the burgeoning field of live-streaming and virtual conferencing. These applications require the meticulous removal of visual imperfections. On the other + +![](images/489ea0340d441e997c7407c852d45dc1e6d9b4638677b9205ecfcb3f54ab1250.jpg) + +![](images/9edbcf82fb4a4d69533d41fe3f218277e98e824a6d85a78e35046ad05256b4da.jpg) +Figure 1. An example to demonstrate the superiority of VRetouchEr over the main competing methods. (Upper row) The quantitative results of the methods in terms of PSNR. (Middle row) Imperfection detection results of BPFRe. (Bottom row) Imperfection detection results of VRetouchEr. + +hand, the enhancements should be imperceptible to preserve the subject's natural appearance. + +Recent video enhancement methods [23-25] fall short in the context of face video retouching due to their inability to accurately model and track the movement of facial imperfections over time. For instance, the optical flow technique utilized in ProPainter [51] lacks the necessary precision for facial flaw localization. The image-to-image translation methods [21, 33, 48] and GAN-based methods [17, 45, 47] are designed for static image retouching, and fall short in maintaining the performance and stability when addressing image sequences. The deficiency of paired training data further impedes the progress in this domain. Hence, there is a significant gap in current methodologies to perform face video retouching, underscoring the need for handling the + +temporal and spatial dynamics of facial imperfections in video streams. + +In this work, we address the limitations of existing retouching methods by proposing a face Video Retouching transformEr with facial imperfection flow-based multiframe attention, which is referred to as VRetouchEr. The core idea of our framework is to incorporate an imperfection flow module to estimate the displacement vectors of imperfections between consecutive frames. Conditioned on the estimation, we can correct the imperfection detection results in each frame via adaptive modulation at specific pixel locations. It is worth noting that the resulting motion information plays an important role in localizing facial imperfection across frames in a stable and precise way. Further, we design a multi-frame masked attention mechanism to synthesize high-fidelity content in the imperfection regions. To achieve this, we allow long-distance interactions between the facial regions from different frames, since multiple frames provide richer information of normal skin characteristics. On the other hand, the estimated imperfection maps are used to weight the intermediate features, such that the imperfections are suppressed. As a result, the attention mechanism enables stable and precise retouching across frames. In summary, the main contribution of this work are as follows: (a) We propose VRetouchEr, a novel face video retouching approach that addresses the challenging of imperfection localization and removal over image sequences. (b) By estimating imperfection flow to correct the imperfection localization in each frame, the obtained spatial information becomes more stable and reliable, which is beneficial to stabilizing the retouching performance. (c) By performing multi-frame masked attention computation, VRetouchEr is able to leverage the contextual information from different frames and synthesize more precise retouching results than existing methods operating on single frame. + +# 2. Related Work + +# 2.1. Image-to-Image Translation + +Image-to-image translation aims to learn a mapping across domains of visually distinguishable images, while preserving the content representation to a certain extent [19, 27, 29, 34]. Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) [14] have led to significant progress in this domain. As a representative GAN-based method, Pix2Pix [17] aimed at learning a mapping to minimize the pixel-wise discrepancy. Imposing cycle consistency regularization on the original and the reconstructed images, as in CycleGAN [52] and DiscGAN [19], is also observed to be effective. A multi-stage method MPRNet [48] leveraged the high-level contextual information and spatial details to improve synthesis results. On the other hand, GPEN [47] combined a U-shaped CNN + +with a GAN, in which the generative prior was useful for high-quality image generation. To perform effective translation among multiple domains, StarGAN [6,7] learnt a unified translation framework, which was conditional on the domain label. Additionally, the latent space of a pre-trained GAN was observed to possess semantic organization, which allowed semantic editing on images [8, 13, 18]. By leveraging GAN inversion methods [15, 32, 36] to project images back to the latent space, attribute manipulation was performed by imposing the attribute-associated transformations on the resulting latent vectors. To address the multiframe setting, burst image restoration methods [3, 43] utilized cross-frame cues to merge burst shot frames, which ultimately led to a significant improvement in restoration performance. In addition, BIPNet [11] distilled and aligned frame-wise features through an edge-focused alignment module, and then refined the resolution progressively. + +# 2.2. Vision Transformer + +The Vision Transformer (ViT) architecture [10, 44, 46] has emerged as a powerful framework, drawing inspiration from transformer research in natural language processing (NLP) [9, 37]. By incorporating the multi-head attention mechanism, positional encoding and position-wise feedforward networks, ViT was able to capture the relationships among input features. ViT achieved significant improvements across multiple vision tasks, such as object detection [4, 53], semantic segmentation [39, 50] and imaged-to-image translation [5, 16]. Building on ViT, the SwinTransformer [30] improved the original design by arranging the image data in a more organized manner, leading to better results in computer vision tasks. For image synthesis, VQGAN [12] combined the transformer's ability to capture complex relationships with the image generation capability of GANs, resulting in increased realism of images generated from text descriptions. RestoreFormer [42] learnt a dictionary in terms of key-value pairs from high-quality data to restore degraded images. Transformers have also shown to be effective for video inpainting [24, 51] and restoration [25, 26]. ProPainter [51] incorporated optical flow into an inpainting module, and a dual-domain feature propagation approach was adopted to synthesize the content in the masked regions. + +# 2.3. Face Retouching + +Face retouching is a specialized task in the domain of image translation, which focuses on the removal of imperfections from human faces in images or videos, while preserving facial attributes as much as possible. Traditional retouching approaches [1,2,22,28,38] typically adopt a global smoothing strategy for the entire face region, which may result in the loss of fine-grained details. AutoRetouch [33] adapted a GAN-based framework to face retouching, and + +![](images/f9c1011a3e6d1d795d1f199673e12724b30ebc7adb8d7e3da398392fd27d309e.jpg) +Figure 2. An overview of the proposed VRetouchEr's structure. An Encoder-Transformer-Decoder architecture is specifically designed for Face Video Retouching. VRetouchEr leverages two key modules: the Flow-based Imperfection Refinement (FIR) module and the Multiframe Masked Attention (MMA) mechanism. FIR performs imperfection flow estimation, and further utilizes the displacement information to refines the imperfection localization on each frame. By injecting the imperfection information into the latent transformation, MMA leads to more accurate and consistent retouching outcomes throughout the frame sequence. + +![](images/d2127569e151ab03704d77eb29602211b702cd80445039bd1e56e83122e95160.jpg) + +![](images/4fbde740d364e426a4fa86f1ef87ea83717e6fe70082d7447da50fecd17536b0.jpg) + +the encoder-decoder-based generator is induced to synthesize clean face images in an adversarial training process. To apply more attention on the local regions, ABPN [21] incorporated context-aware local retouching layers together with an adaptive blend pyramid layer for coarse-to-fine retouching. Similarly, BPFRe [45] performed imperfection detections followed by a two-stage retouching process, in which the intermediate features of a U-net were injected as side information into a StyleGAN generator to perform conditional clean face image generation. + +There are fundamental differences between the proposed VRetouchEr and the above existing methods. First, this work focuses on face retouching in videos, while the video enhancement methods, like ProPainter, fall shot in this task due to lack of imperfection localization. Second, the state-of-the-art model, BPFRe, is designed for single-image retouching, thus its performance is unstable when applied to image sequences. To address the limitations, VRetouchEr performs imperfection flow estimation to improve the localization in terms of both stability and precision. Further, we adopt a task-specific transformer architecture by designing multi-frame masked attention mechanism, such that the long-distance relationship between facial features can be effectively captured across frames. + +# 3. Methodology + +# 3.1. Overview + +In this section, we elaborate on the design of the proposed approach, VRetouchEr, for high-quality face retouching in videos. Given a raw image sequence $X$ and the cor + +responding retouched one $Y$ , our goal is to precisely locate the imperfections in $X$ and replace with the content consistent with the surrounding normal skin. The key idea behind VRetouchEr is to leverage the temporal context information to stabilize face retouching over video frames. To achieve this, we adopt an encoder-transformer-decoder based architecture for VRetouchEr as illustrated in Figure 2. An encoder $E$ extracts features of the target frame $X_{t}$ and multiple reference frames $X_{r}$ in sequence $X$ , and a flow estimation network $S$ is incorporated to model the motions of imperfections between consecutive frames. The estimated imperfection flow $O$ is fed into a Flow-based Imperfection Localization (FIR) module to improve the estimation stability and precision over the sequence. The resulting imperfection map $M$ , together with extracted target feature $f_{t}$ and reference feature $f_{r}$ , are fed into the latent transformer denoted by $T$ to restore the clean face features. In each intermediate block, we can collect richer information of normal skin from the reference frames, and our Multi-frame Masked Attention (MMA) mechanism enables the model to concentrate on the most relevant parts. Finally, the decoder $G$ rebuilds each target frame from the transformed features. + +# 3.2. Flow-based Imperfection Refinement + +The FIR module is specifically designed to enhance the accuracy of imperfection localization by leveraging flow information. This process involves imperfection movement estimation, imperfection localization, and imperfection map refinement. The refinement is crucial for stable retouching in each target frame. To achieve this, we define the imperfections are the differences between the manually retouched + +frames and raw frames, represented as $M_{gt}$ . We apply a pretrained SpyNet [31] on $M_{gt}$ to produce the ground truth imperfection flow, denoted as $O_{gt}$ . + +Our designed FIR module consists of an imperfection localization network $N$ and learnable align factors $\alpha, \beta$ . Given frame $X^i$ and its consecutive frame $X^j$ , the flow estimation network $S$ is trained to predict an imperfection displacement map $O^{i \to j}$ , which is required to approximate the corresponding ground truth imperfection flow as accurately as possible. We evaluate the imperfection flow prediction by measuring the degree of consistency as follows: + +$$ +O ^ {i \rightarrow j} = S \left(X ^ {i}, X ^ {j}\right), \tag {1} +$$ + +$$ +\mathcal {L} _ {f l o w} = \mathbb {E} _ {X} \left[ | O ^ {i \rightarrow j} - O _ {g t} ^ {i \rightarrow j} | _ {1} \right], +$$ + +By minimizing $\mathcal{L}_{flow}$ , $S$ is encouraged to focus more on the displacement of imperfections instead of background and normal skin regions between consecutive frames. + +On the other hand, the imperfection localization network $N$ is trained to detect imperfections from frame $X^i$ together with the corresponding encoder feature $f_r^i$ . We observe that the detection results on individual frames are unstable when there are significant changes in facial postures. To address this issue, we utilize the information from consecutive frames and the estimated imperfection flow $O^{i \to j}$ to refine the detection results. Specifically, the imperfection flow $O^{i \to j}$ will warp $M^i$ into a flow-based imperfection map $\mathcal{M}^j$ . Let $\mathcal{W}(\cdot, \cdot)$ denotes the warping operation, and we have: + +$$ +M ^ {i} = N (X ^ {i}, f _ {r} ^ {i}), +$$ + +$$ +M ^ {j} = N \left(X ^ {j}, f _ {r} ^ {j}\right), \tag {2} +$$ + +$$ +\mathcal {M} ^ {j} = \mathcal {W} (O ^ {i \rightarrow j}, M ^ {i}), +$$ + +where $\mathcal{M}^j$ denotes the warped imperfection map. Next, our designed align factors $\alpha$ and $\beta$ are employed to align the flow-based imperfection map $\mathcal{M}^j$ with predicted imperfection map $M^j$ . We utilized adaptive convolution blocks [35], denoted as $\theta (\cdot ,\cdot)$ , for an effective fusion of these features. The refinement operation is defined as follows: + +$$ +\mathcal {M} _ {a} ^ {j} = \theta (\alpha * \mathcal {M} ^ {j} + \beta , \mathcal {M} ^ {j}), \tag {3} +$$ + +$$ +\hat {M} ^ {j} = \sigma \left(\theta \left(M ^ {j}, \mathcal {M} _ {a} ^ {j}\right)\right), +$$ + +where $\hat{M}^j$ represents the refined imperfection map, $\mathcal{M}_a^j$ is the aligned map and $\sigma$ represents the sigmoid function. All the refined imperfection maps will be stacked together to form the output, denoted as $M$ . The elements in $M$ are normalized in the range from 0 to 1, representing the refined estimation of the imperfections across frames. + +# 3.3. Multi-frame Masked Attention Transformer + +Our proposed MMA mechanism enables the latent transformer to modify the features $f_{t}$ in the imperfection regions + +by leveraging the features $f_{r}$ extracted from multiple frames as references. To achieve this, we use the obtained imperfection map $M$ to weight the features of target and reference frames, and perform cross-attention computation between them. Let the number of reference frames be $\delta$ , the weighted target feature is used as query, and the weighted reference ones are used as keys and values, defined as follows: + +$$ +Q _ {t} = W _ {q} \left(f _ {t} \otimes M _ {t} + b _ {q}\right), +$$ + +$$ +K _ {r} ^ {i} = W _ {k} \left(f _ {r} ^ {i} \otimes \left(1 - M _ {r} ^ {i}\right) + b _ {k}\right), \tag {4} +$$ + +$$ +V _ {r} ^ {i} = W _ {v} \left(f _ {r} ^ {i} \otimes (1 - M _ {r} ^ {i}) + b _ {v}\right), +$$ + +where $M_{t}$ and $M_{r}$ represent the refined imperfection map for target frame and reference frames correspondingly. Next, we perform the cross-attention computation to obtain the modification maps as follows: + +$$ +\Delta_ {f _ {t}} = \operatorname {s o f t m a x} \left(Q _ {t} \cdot \sum_ {i} ^ {\delta} K _ {r} ^ {i} / \sqrt {\Lambda}\right) \cdot \sum_ {i} ^ {\delta} V _ {r} ^ {i}, \tag {5} +$$ + +where $\Lambda$ denotes the channel number of the features. In Eq.(4), we adopt the imperfection map $M$ as a soft mask to isolate the regions to be edited, $M$ and $(1 - M)$ represents the mask for imperfection regions and normal regions respectively. Based on this, we can realize the goal of replacing imperfection regions with normal skin by strengthening the interactions between the target features associated with imperfections and the reference features associated with normal skin. Furthermore, we use additive attention in Eq.(5) with linear complexity to aggregate the key and value vectors from multiple reference frame into global vectors, which is used to perform element-wise multiplication to induce temporal context information. + +Considering that the background and normal skin in the target frame should be preserved properly, the original feature maps are weighted combined with the modification maps as follows: + +$$ +\hat {f} _ {t} = f _ {t} \otimes (1 - M _ {t}) + \Delta_ {f _ {t}} \otimes M _ {t}, \tag {6} +$$ + +where $\hat{f}_t$ denotes the retouched feature, can be fed into the subsequent transformer block as input target feature for further refinement. In this process, the masked attention mechanism enables effective information exchange across frames, and the imperfections in the target frame are progressively suppressed and replaced with the content learnt from normal skin. + +# 3.4. Model Training + +The training loss functions of VRetouchEr involves the following three aspects: imperfection flow estimation, imperfection localization and retouching evaluation. As illustrated in Subsection 3.2, we employ a pretrained SpyNet + +to produce the ground truth of imperfection flow estimation, and the network $S$ takes consecutive frames as input and predicts the flow as accurately as possible. In addition, the flow is used to improve the imperfection localization on each frame, and we evaluate the precision by measuring the discrepancy between the prediction $M$ and the ground truth imperfections $M_{gt}$ as follows: + +$$ +\mathcal {L} _ {i m p} = \mathbb {E} _ {X} \| \mathcal {I} (M) - M _ {g t} \| _ {1}, \tag {7} +$$ + +where $\mathcal{I}$ denotes a learnable layer to align $M$ 's channel number with $M_{gt}$ . We optimize $N$ and $S$ by minimizing $\mathcal{L}_{imp}$ together with the flow estimation loss $\mathcal{L}_{flow}$ defined in Eq.(2), and the formulation is expressed as follows: + +$$ +\min _ {S, N} \mathcal {L} _ {\text {f l o w}} + \mathcal {L} _ {\text {i m p}}. \tag {8} +$$ + +Let $\hat{y} = G(\hat{f}_t)$ denote the retouched frame synthesized by VRetouchEr. We perform retouching on every raw frame and combine the results together to get the retouched sequence $\hat{Y}$ , which is evaluated by measuring the degree of consistency with the manually retouched frames $Y$ as follows: + +$$ +\mathcal {L} _ {c o n} = \mathbb {E} _ {X} \left[ \zeta \| Y - \widehat {Y} | _ {1} + \| \mathcal {V} (Y) - \mathcal {V} (\widehat {Y}) \| _ {2} ^ {2} \right], \tag {9} +$$ + +where $\zeta$ denotes a weighting factor, and $\mathcal{V}(\cdot)$ represents the features extracted from a pre-trained VGG-19. Considering that high-fidelity content synthesis can benefit from adversarial training, we thus adopt the real-fake discrimination loss formulated as follows: + +$$ +\mathcal {L} _ {a d v} ^ {s y n t} = \mathbb {E} _ {X} \bigl [ \log (D (\hat {Y})) \bigr ] +$$ + +$$ +\mathcal {L} _ {a d v} ^ {r e a l} = \mathbb {E} _ {X} [ \log (1 - D (X)) ] + \mathbb {E} _ {Y} [ \log (D (Y)) ], \tag {10} +$$ + +where discriminator $D$ takes the sequence as input, and is trained to distinguish the manually retouched frames from the synthesized ones. By integrating the loss functions of retouching evaluation and adversarial training, the optimization formulation of the constituent networks is expressed as follows: + +$$ +\min _ {E T C} \mathcal {L} _ {c o n} + \mathcal {L} _ {a d v} ^ {s y n}, +$$ + +$$ +E, I, G \tag {11} +$$ + +$$ +\max _ {D} \mathcal {L} _ {a d v} ^ {r e a l}. +$$ + +The constituent networks of the proposed VRetouchEr are jointly optimized from scratch. + +# 4. Experiment + +In this section, we perform extensive experiments to assess the effectiveness of VRetouchEr in face video retouching. We first present the details of the datasets and the experimental setup. Next, we compare our VRetouchEr with previous state-of-the-art face retouching methods quantitatively and qualitatively. Finally, we provide insights by analyzing the design elements of VRetouchEr. + +# 4.1. Experimental Settings + +Datasets. Due to the difficulty in collecting paired raw-retouched image sequences, we utilized the Flickr-Face-HQ-Retouching (FFHQR) dataset [33] to construct emulated video data. The training/validation/test data in FFHQR consists of $56\mathrm{k} / 7\mathrm{k} / 7\mathrm{k}$ raw-retouched image pairs. For each image pair $\{x,y\}$ in FFHQR, we perform random cropping, flipping and translation on both $x$ and $y$ to generate a pair of image sequences $\{X,Y\}$ , and the resulting dataset is referred to FFHQR-Seq. In addition, we build another Manually Retouching Face Video dataset, referred to as MRFV, which contains 200 in-the-wild portrait videos with different types of facial imperfections, and each of them contains at least 500 frames. We employ multiple retouchers to manually retouch each frame of these videos, and evaluate the retouching performance of our VRetouchEr and the competing methods trained on FFHQR-Seq. + +Training Details. The images in both training and test sequences are resized to $512 \times 512$ for a fair comparison with the existing face retouching methods. In the training process, the parameters of VRetouchEr are updated by the Adam optimizer [20] with the learning rate of $2 \times 10^{-4}$ . There are a total of 400k training iterations, the batch size is set to 1, and the hyper-parameter $\zeta$ in Eq.(9) is set to 10. We implement VRetouchEr by using PyTorch and train it on single NVIDIA GeForce GTX 3090 GPU. + +Evaluation Protocols. In the experiments, we implement all the competing methods by using the open source codes. We adopt the Soft Intersection over Union Loss (Soft-IoU) to measure the correctness of imperfection localization. To measure the consistency between the synthesized image sequences and the manually retouched ones, we adopt the widely used metrics: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index Measure (SSIM) and the Learnt Perceptual Image Patch Similarity (LPIPS) [49]. Furthermore, we quantitatively assess the diversity and the degree of realism of synthesized image sequences in terms of Video Fréchet Inception Distance (VFID) [40]. + +# 4.2. Comparison to State-of-the-arts + +We perform a comprehensive comparative analysis between VRetouchEr and representative state-of-the-art methods, including a typical image translation method: Pix2PixHD [41], generic image restoration methods: MPR-Net [48] and RestoreFormer [42], a blind face restoration method: GPEN [47], video enhancement methods: BIP-Net [11] and ProPainter [51], and face retouching methods: AutoRetouch [33] and BPFRe [45]. We train BIP-Net, ProPainter and VRetouchEr on FFHQR-Seq, and the remaining methods on FFHQR. All the trained models are evaluated on the test data of FFHQR-Seq and MRFV. We quantitatively assess the retouch performance of the meth + +![](images/dff76043d0a875577c4b5b2f37f8af051cea56a1584aebee61ae36d072a9f0fc.jpg) +Figure 3. Visual comparison between VRetouchEr and the competing methods on representative frames of an in-the-wild videos. + +![](images/f0981a464a5cfa44e197b48809a73ebfe4172391a1cac6211455321be4496014.jpg) +Figure 4. Representative high-quality retouching results. From top left to bottom right, the images are the raw frame and the retouching results of BPFRe, ProPainter and VRetouchEr, respectively. + +Table 1. Quantitative comparison between VRetouchEr and competing methods on FFHQR-Seq. + +
MethodsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs37.500.97040.033710.779
Pix2PixHD [41]37.630.97140.03039.252
AutoRetouch [33]38.270.97710.02618.941
MPRNet [48]38.330.97420.02678.231
GPEN [47]38.270.97390.02568.125
RestoreFormer [42]38.460.97330.02327.931
BIPNet [11]38.140.97110.02758.241
ProPainter [51]38.540.97680.02407.872
BPFRe [45]38.690.97740.02197.604
VRetouchEr39.750.98130.01696.375
+ +ods per frame, and report their average PSNR, SSIM, LPIPS and VFID scores in Tables 1-2. On FFHQR-Seq, the com + +Table 2. Quantitative comparison between VRetouchEr and competing methods on MRFV. + +
MethodsMRFV
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs32.070.90540.089141.830
Pix2PixHD [41]33.510.91280.073938.403
AutoRetouch [33]35.030.92110.050231.244
MPRNet [48]35.820.93270.049129.893
GPEN [47]35.710.92810.046330.135
RestoreFormer [42]35.760.92990.051629.785
BIPNet [11]35.260.92310.059333.897
ProPainter [51]36.270.93840.045220.344
BPFRe [45]36.320.94270.040118.772
VRetouchEr37.630.95300.035710.368
+ +peting methods achieve similar retouching performance except Pix2PixHD, and our VRetouchEr outperforms them in terms of all the metrics. In particular, VRetouchEr achieves the PSNR score of 39.75, which is higher than that of the second best method: BPFRe, by $1.06\mathrm{dB}$ . On MRFV, VRetouchEr also achieves superior retouching performance over the competing methods. The VFID score of VRetouchEr reaches 10.368, which is lower than that of BPFRe by a significant improvement of about 44.77 percentage points. + +In Figure 3, we perform the detailed visual comparison. We can make the following observations: Most of the competing methods fail to remove the marked acnes. BPFRe falls short in maintaining the stability of retouching in videos, since it is designed for single-image retouching. Although ProPainter is specifically designed for video enhancement, it fails to achieve satisfactory retouching results due to lack of imperfection localization. In contrast, our VRetouchEr is able to consistently deliver stable and high-quality retouching outcomes. In Figure 4, we provide visual examples to further compare the performance of our framework with the main competing methods in removing dense + +![](images/768685824ca83c75b6e039eed6ef525f29b4a191b37078ee167cc0a4ab879c21.jpg) +Figure 5. Quantitative comparison in imperfection localization per frame. VRetoucher achieves a more precise and stable result than 'VRetoucher w/o FIR' and BPFRe. + +![](images/6d698e74db156da10dc7fb0d9db6f63e3798116c5781b36c77f302971418a49a.jpg) +Figure 6. Quantitative comparison in imperfection localization on MRFV dataset. + +imperfections. These results highlight the VRetouchEr's capability of face retouching in videos even without any manually retouched videos as training data. + +# 4.3. User Study + +We further perform a user study to evaluate VRetouchEr and the competing methods in terms of human perception. The models trained on FFHQR/FFHQR-Seq are applied to a total of 40 raw videos randomly sampled from MRFV. There are 50 participants, each of which is presented with the synthesized videos and required to carefully compare and sort the retouching results, according to the criteria including overall visual quality, effectiveness of imperfection removal, naturalness, and consistency. Table 3 shows the ranking results, and we are able to draw the conclusion regarding the superior performance of VRetouchEr against the competing methods. + +# 4.4. Flow-based Imperfection Refinement + +We observed that our FIR module plays in an important role in facilitating face retouching in videos. Different from + +Table 3. The voting result $(\%)$ of user study on MRFV. + +
MethodRank-1Rank-2Rank-3Rank-4Rank-5Rank-6
VRetouchEr88.4511.050.300.200.000.00
BPFRe [45]11.1563.1519.355.051.200.10
ProPainter [51]0.2518.2061.6016.303.150.50
MPRNet [48]0.056.0511.6559.256.4016.60
RestoreFormer [42]0.051.506.1011.2561.3019.80
BIPNet [11]0.050.051.007.9527.9563.00
+ +![](images/f923883ed0edd53d967130d676c238c229a8a2f78de5df5531ab5c3eb8bf81cf.jpg) +Figure 7. Visual comparison in imperfection localization. + +BPFRe detecting facial imperfections on each frame separately, we estimate imperfection flow between consecutive frames followed by imperfection refinement. The displacement information is useful for refining and stabilizing the imperfection maps of the video to be edited. In Figure 5, we quantitatively evaluate imperfection localization results on a video sequence in terms of Soft-IoU per-frame between the results and the ground truth imperfection maps $M_{gt}$ as defined in Sec. 3.2. In Figure 6, we perform the evaluation of imperfection localization between the results and $M_{gt}$ in terms of mean SSIM and mean Soft-IoU on MRFV dataset. When disabling FIR, the variant 'VRetouchEr w/o FIR' performs single-image imperfection localization, and has comparable performance with BPFRe. We can find that FIR leads to a significant and stable improvement for cross-frame imperfection localization. We further visualize the imperfection maps of BPFRe, 'VRetouchEr w/o FIR' and VRetouchEr in Figure 7, and observe that VRetouchEr can stably locate the most imperfections, and its results are more consistent with the ground truth. + +# 4.5. Impact of the Number of Reference Frames + +The existing face retouching methods suffer from the problem of overlooking temporal context information in handling videos. As another important module of VRetouchEr, we consider that our designed MMA mechanism also contributes to the superior performance of VRetouchEr. To verify this, we investigate the impact of the number of reference frames $\delta$ on the final retouching performance. In Figure 8, we plot the VFID scores of VRetouchEr on MRFV during training. For the case of $\delta = 1$ , we perform single-image retouching on each frame. One can observe that increasing the value of $\delta$ for MMA indeed leads to a significant improvement. In particular, VRetouchEr achieves the best retouching performance when $\delta = 6$ . Beyond this value, VRetouchEr's performance becomes stable. In Figure 9, we visualize the retouching results to further verify this observation. This insight provides a valuable guidance for optimizing the performance of VRetouchEr in the face video retouching task. + +![](images/fd8342aa4ac1cff3c1828e35e6372dae4713f278ba94a75ecc0747995fa32a09.jpg) +Figure 8. Quantitative results of VRetouchEr with different $\delta$ . + +![](images/e7d1125df38e4cca0c9f3c13def2b8bf059f55384a15a4809e44dc10c12d91f8.jpg) +Figure 9. The retouching results of VRetouchEr with different $\delta$ . + +# 4.6. Ablation Study + +The two proposed modules: FIR and MMA, distinguish the proposed VRetouchEr from the existing works. We perform a set of ablative experiments to investigate the impact of FIR and MMA on the retouching performance on FFHQR-Seq. Specifically, we build the Base model of VRetouchEr by disabling both FIR and MMA, and find that VRetouchEr significantly outperforms the base model as shown in Table 4. Additionally, we build two variants by disabling FIR and MMA, and the resulting models are referred to as 'VRetouchEr w/o FIR' and 'VRetouchEr w/o MMA', respectively. In 'VRetouchEr w/o FIR', the maps produced by the imperfection localization network $N$ are directly fed into the latent transformer. As analyzed in Sec.4.4, disabling FIR leads to unstable imperfection localization, which ultimately degrades the retouching performance on videos. As shown in Table 4, the PSNR score of 'VRetouchEr w/o FIR' is 38.74 dB, which is worse than that of VRetouchEr by 1.01 dB. When disabling MMA, we also observe a significant performance drop of about 40.83 percentage points in terms of LPIPS. Qualitative results in Figure 10 also prove the effectiveness of FIR and MMA. + +Table 4. Quantitative results of ablative models. + +
VariantsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Base model38.130.97380.02648.671
VRetouchEr w/o MMA38.670.97580.02387.534
VRetouchEr w/o FIR38.740.97690.02077.312
VRetouchEr39.750.98130.01696.375
+ +![](images/aaa3a5f66eb10092b82d621a4803456069bc5761967082c2b45b47226a3a7188.jpg) +Figure 10. Representative retouching results of ablative models. + +# 5. Conclusion + +In this work, we propose a latent retouching transformer for removing facial imperfections in videos. Different from the existing methods that focus on single-image retouching, we leverage the temporal context information to significantly facilitate face retouching in videos. Toward this end, we perform imperfection flow estimation to obtain the displacement information between consecutive frames, and further refine the imperfection localization by fusing motion-based and frame-based predictions. By injecting the imperfection maps as side information into a latent retouching transformer, we adopt the masked attention computation over multiple frames, such that the features of normal skin from different frames can be leveraged to substitute for that of imperfections in the target frame. Extensive experiments are performed to demonstrate the superior capability of the proposed approach over state-of-the-art face retouching methods in stabilizing the retouching performance over video frames. + +# 6. Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2020A1515010484, 2022A1515011160), and in part by TCL Science and Technology Innovation Fund (Project No. 20231752). + +# References + +[1] Kaoru Arakawa. Nonlinear digital filters for beautifying facial images in multimedia systems. In Proc. IEEE International Symposium on Curcuits and Systems, 2004. 2 +[2] Nazre Batool and Rama Chellappa. Detection and inpainting of facial wrinkles using texture orientation fields and Markov random field modeling. IEEE Transactions on Image Processing, 23(9):3773-3788, 2014. 2 +[3] Goutam Bhat, Martin Danelljan, Luc van Gool, and Radu Timofte. Deep burst super-resolution. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2021. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. European Conference on Computer Vision, 2020. 2 +[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[6] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 2 +[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 2 +[8] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 2 +[9] Jocob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In arXiv:1801.04805, 2018. 2 +[10] Alexy Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2 +[11] Akshay Dudhane, Syed Waqas Zamir, Salman Khan1, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burst image restoration and enhancement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 2, 5, 6, 7 +[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[13] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive im + +age properties. In Proc. International Conference on Computer Vision, 2019. 2 +[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 2 +[15] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 2 +[16] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[17] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2 +[18] Ali Jahanian, Lucy Chai, and Phillip Isola. On the "steerability" of generative adversarial networks. In arXiv:1907.07171, 2019. 2 +[19] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 2 +[20] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5 +[21] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: Adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 3 +[22] Tommer Leyvand, Daniel Cohen-Or, Gideon Dror, and Dani Lischinski. Data-driven enhancement of facial attractiveness. In Proc. ACM Conference on Special Interest Group on Computer Graphics and Interactive Techniques, 2008. 2 +[23] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In Proc. European Conference on Computer Vision, 2022. 1 +[24] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo1, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 2 +[25] Jing Lin, Yuanhao Cai, Xiaowan Hu, Haoqian Wang, and Youliang Yan. Flow-guided sparse transformer for video deblurring. In Proc. International Conference on Machine Learning, 2022. 1, 2 +[26] Jing Lin, Xiaowan Hu, Yuanhao Cai, and Haoqian Wang. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In Proc. International Conference on Machine Learning, 2022. 2 +[27] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. EditGAN: high-precision semantic image editing. In Proc. Neural Information Processing Systems, 2021. 2 + +[28] Uri Lipowezky and Sarah Cahen. Automatic freckles detection and retouching. In Proc. IEEE Convention of Electrical and Electronics Engineers in Israel, 2008. 2 +[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2 +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. International Conference on Computer Vision, 2021. 2 +[31] Anurag RanjanMichael and J. Black. Optical flow estimation using a spatial pyramid network. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2017. 4 +[32] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGan encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2 +[33] Alireza Shafaei, James J. Little, and Mark Schmidt. AutoRetouch: automatic professional face retouching. In Proc. IEEE Winter Conference on Applications of Computer Vision, 2021. 1, 2, 5, 6 +[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[35] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 4 +[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems, 2017. 2 +[38] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. FabSoften: face beautification via dynamic skin smoothing, guided feathering and texture restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition Workshop, 2020. 2 +[39] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: end-to-end panoptic segmentation with mask transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[40] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2018. 5 +[41] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 5, 6 + +[42] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. RestoreFormer: high-quality blind face restoration from undegraded key-value pairs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 5, 6, 7 +[43] Pengxu Wei, Yujing Sun, Xingbei Guo, Chang Liu, Guanbin Li, Jie Chen, Xiangyang Ji, and Liang Lin. Towards real-world burst image super-resolution: Benchmark and method. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2 +[44] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvR: introducing convolutions to vision transformers. arXiv:2103.15808, 2021. 2 +[45] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face retouching with limited paired data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 3, 5, 6, 7 +[46] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[47] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. GAN prior embedded network for blind face restoration in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6 +[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Huang Yang, and Ling Shao. Multi-stage progressive image restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 5 +[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2 +[51] Shangchen Zhou, Chongyi Li, Kelvin C.K. Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 2, 5, 6, 7 +[52] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 2 +[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 2 \ No newline at end of file diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/images.zip b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46f8c1a41fa16b8351079d733884e8ed8d65dea6 --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93a9c7c4aa233c6e49df1318934ca5c22b497517cb6a9a14751f098a9bb4d52d +size 747301 diff --git a/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/layout.json b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9d0b698c501da4bcece2fd003529982077fc4ad4 --- /dev/null +++ b/2024/VRetouchEr_ Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos/layout.json @@ -0,0 +1,8536 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "type": "text", + "content": "VRetouchEr: Learning Cross-frame Feature Interdependence with Imperfection Flow for Face Retouching in Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 504, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 504, + 219 + ], + "type": "text", + "content": "Wen Xue1, Le Jiang1, Lianxin Xie1, Si Wu1,2,3*, Yong Xu1,2,3 and Hau San Wong4 \n1School of Computer Science and Engineering, South China University of Technology \n2Peng Cheng Laboratory 3PAZHOU LAB \n4Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 220, + 440, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 220, + 440, + 232 + ], + "spans": [ + { + "bbox": [ + 153, + 220, + 440, + 232 + ], + "type": "text", + "content": "{csxuewen, csjangle, cslianxin.xie}@mail.scut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 234, + 426, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 234, + 426, + 245 + ], + "spans": [ + { + "bbox": [ + 167, + 234, + 426, + 245 + ], + "type": "text", + "content": "{cswusi, yxu}@scut.edu.cn, cshswong@cityu.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 298, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 586 + ], + "type": "text", + "content": "Face Video Retouching is a complex task that often requires labor-intensive manual editing. Conventional image retouching methods perform less satisfactorily in terms of generalization performance and stability when applied to videos without exploiting the correlation among frames. To address this issue, we propose a Video Retouching transformEr to remove facial imperfections in videos, which is referred to as VRetouchEr. Specifically, we estimate the apparent motion of imperfections between two consecutive frames, and the resulting displacement vectors are used to refine the imperfection map, which is synthesized from the current frame together with the corresponding encoder features. The flow-based imperfection refinement is critical for precise and stable retouching across frames. To leverage the temporal contextual information, we inject the refined imperfection map into each transformer block for multi-frame masked attention computation, such that we can capture the interdependence between the current frame and multiple reference frames. As a result, the imperfection regions can be replaced with normal skin with high fidelity, while at the same time keeping the other regions unchanged. Extensive experiments are performed to verify the superiority of VRetouchEr over state-of-the-art image retouching methods in terms of fidelity and stability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "With the development of digital media enhancement, face retouching in videos plays an important role in improving the facial appearance of persons in dynamic sequences. There are a wide range of applications, from professional video production to the burgeoning field of live-streaming and virtual conferencing. These applications require the meticulous removal of visual imperfections. On the other" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 270, + 533, + 377 + ], + "blocks": [ + { + "bbox": [ + 317, + 270, + 533, + 377 + ], + "lines": [ + { + "bbox": [ + 317, + 270, + 533, + 377 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 533, + 377 + ], + "type": "image", + "image_path": "489ea0340d441e997c7407c852d45dc1e6d9b4638677b9205ecfcb3f54ab1250.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 377, + 533, + 454 + ], + "blocks": [ + { + "bbox": [ + 317, + 377, + 533, + 454 + ], + "lines": [ + { + "bbox": [ + 317, + 377, + 533, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 533, + 454 + ], + "type": "image", + "image_path": "9edbcf82fb4a4d69533d41fe3f218277e98e824a6d85a78e35046ad05256b4da.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 459, + 547, + 515 + ], + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 515 + ], + "type": "text", + "content": "Figure 1. An example to demonstrate the superiority of VRetouchEr over the main competing methods. (Upper row) The quantitative results of the methods in terms of PSNR. (Middle row) Imperfection detection results of BPFRe. (Bottom row) Imperfection detection results of VRetouchEr." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 532, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 556 + ], + "type": "text", + "content": "hand, the enhancements should be imperceptible to preserve the subject's natural appearance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Recent video enhancement methods [23-25] fall short in the context of face video retouching due to their inability to accurately model and track the movement of facial imperfections over time. For instance, the optical flow technique utilized in ProPainter [51] lacks the necessary precision for facial flaw localization. The image-to-image translation methods [21, 33, 48] and GAN-based methods [17, 45, 47] are designed for static image retouching, and fall short in maintaining the performance and stability when addressing image sequences. The deficiency of paired training data further impedes the progress in this domain. Hence, there is a significant gap in current methodologies to perform face video retouching, underscoring the need for handling the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "9141" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "temporal and spatial dynamics of facial imperfections in video streams." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 96, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 96, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 286, + 491 + ], + "type": "text", + "content": "In this work, we address the limitations of existing retouching methods by proposing a face Video Retouching transformEr with facial imperfection flow-based multiframe attention, which is referred to as VRetouchEr. The core idea of our framework is to incorporate an imperfection flow module to estimate the displacement vectors of imperfections between consecutive frames. Conditioned on the estimation, we can correct the imperfection detection results in each frame via adaptive modulation at specific pixel locations. It is worth noting that the resulting motion information plays an important role in localizing facial imperfection across frames in a stable and precise way. Further, we design a multi-frame masked attention mechanism to synthesize high-fidelity content in the imperfection regions. To achieve this, we allow long-distance interactions between the facial regions from different frames, since multiple frames provide richer information of normal skin characteristics. On the other hand, the estimated imperfection maps are used to weight the intermediate features, such that the imperfections are suppressed. As a result, the attention mechanism enables stable and precise retouching across frames. In summary, the main contribution of this work are as follows: (a) We propose VRetouchEr, a novel face video retouching approach that addresses the challenging of imperfection localization and removal over image sequences. (b) By estimating imperfection flow to correct the imperfection localization in each frame, the obtained spatial information becomes more stable and reliable, which is beneficial to stabilizing the retouching performance. (c) By performing multi-frame masked attention computation, VRetouchEr is able to leverage the contextual information from different frames and synthesize more precise retouching results than existing methods operating on single frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 506, + 132, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 506, + 132, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 506, + 132, + 517 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 527, + 200, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 527, + 200, + 539 + ], + "spans": [ + { + "bbox": [ + 49, + 527, + 200, + 539 + ], + "type": "text", + "content": "2.1. Image-to-Image Translation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 546, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 546, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 546, + 286, + 712 + ], + "type": "text", + "content": "Image-to-image translation aims to learn a mapping across domains of visually distinguishable images, while preserving the content representation to a certain extent [19, 27, 29, 34]. Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) [14] have led to significant progress in this domain. As a representative GAN-based method, Pix2Pix [17] aimed at learning a mapping to minimize the pixel-wise discrepancy. Imposing cycle consistency regularization on the original and the reconstructed images, as in CycleGAN [52] and DiscGAN [19], is also observed to be effective. A multi-stage method MPRNet [48] leveraged the high-level contextual information and spatial details to improve synthesis results. On the other hand, GPEN [47] combined a U-shaped CNN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 275 + ], + "type": "text", + "content": "with a GAN, in which the generative prior was useful for high-quality image generation. To perform effective translation among multiple domains, StarGAN [6,7] learnt a unified translation framework, which was conditional on the domain label. Additionally, the latent space of a pre-trained GAN was observed to possess semantic organization, which allowed semantic editing on images [8, 13, 18]. By leveraging GAN inversion methods [15, 32, 36] to project images back to the latent space, attribute manipulation was performed by imposing the attribute-associated transformations on the resulting latent vectors. To address the multiframe setting, burst image restoration methods [3, 43] utilized cross-frame cues to merge burst shot frames, which ultimately led to a significant improvement in restoration performance. In addition, BIPNet [11] distilled and aligned frame-wise features through an edge-focused alignment module, and then refined the resolution progressively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 285, + 419, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 419, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 419, + 296 + ], + "type": "text", + "content": "2.2. Vision Transformer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 304, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 589 + ], + "type": "text", + "content": "The Vision Transformer (ViT) architecture [10, 44, 46] has emerged as a powerful framework, drawing inspiration from transformer research in natural language processing (NLP) [9, 37]. By incorporating the multi-head attention mechanism, positional encoding and position-wise feedforward networks, ViT was able to capture the relationships among input features. ViT achieved significant improvements across multiple vision tasks, such as object detection [4, 53], semantic segmentation [39, 50] and imaged-to-image translation [5, 16]. Building on ViT, the SwinTransformer [30] improved the original design by arranging the image data in a more organized manner, leading to better results in computer vision tasks. For image synthesis, VQGAN [12] combined the transformer's ability to capture complex relationships with the image generation capability of GANs, resulting in increased realism of images generated from text descriptions. RestoreFormer [42] learnt a dictionary in terms of key-value pairs from high-quality data to restore degraded images. Transformers have also shown to be effective for video inpainting [24, 51] and restoration [25, 26]. ProPainter [51] incorporated optical flow into an inpainting module, and a dual-domain feature propagation approach was adopted to synthesize the content in the masked regions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 600, + 405, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 405, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 405, + 611 + ], + "type": "text", + "content": "2.3. Face Retouching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 619, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 619, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 619, + 545, + 712 + ], + "type": "text", + "content": "Face retouching is a specialized task in the domain of image translation, which focuses on the removal of imperfections from human faces in images or videos, while preserving facial attributes as much as possible. Traditional retouching approaches [1,2,22,28,38] typically adopt a global smoothing strategy for the entire face region, which may result in the loss of fine-grained details. AutoRetouch [33] adapted a GAN-based framework to face retouching, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "9142" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 265, + 257 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 265, + 257 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 265, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 265, + 257 + ], + "type": "image", + "image_path": "f9c1011a3e6d1d795d1f199673e12724b30ebc7adb8d7e3da398392fd27d309e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 322 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed VRetouchEr's structure. An Encoder-Transformer-Decoder architecture is specifically designed for Face Video Retouching. VRetouchEr leverages two key modules: the Flow-based Imperfection Refinement (FIR) module and the Multiframe Masked Attention (MMA) mechanism. FIR performs imperfection flow estimation, and further utilizes the displacement information to refines the imperfection localization on each frame. By injecting the imperfection information into the latent transformation, MMA leads to more accurate and consistent retouching outcomes throughout the frame sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 273, + 71, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 273, + 71, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 273, + 71, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 273, + 71, + 545, + 159 + ], + "type": "image", + "image_path": "d2127569e151ab03704d77eb29602211b702cd80445039bd1e56e83122e95160.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 274, + 160, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 274, + 160, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 274, + 160, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 274, + 160, + 545, + 257 + ], + "type": "image", + "image_path": "4fbde740d364e426a4fa86f1ef87ea83717e6fe70082d7447da50fecd17536b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "content": "the encoder-decoder-based generator is induced to synthesize clean face images in an adversarial training process. To apply more attention on the local regions, ABPN [21] incorporated context-aware local retouching layers together with an adaptive blend pyramid layer for coarse-to-fine retouching. Similarly, BPFRe [45] performed imperfection detections followed by a two-stage retouching process, in which the intermediate features of a U-net were injected as side information into a StyleGAN generator to perform conditional clean face image generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 628 + ], + "type": "text", + "content": "There are fundamental differences between the proposed VRetouchEr and the above existing methods. First, this work focuses on face retouching in videos, while the video enhancement methods, like ProPainter, fall shot in this task due to lack of imperfection localization. Second, the state-of-the-art model, BPFRe, is designed for single-image retouching, thus its performance is unstable when applied to image sequences. To address the limitations, VRetouchEr performs imperfection flow estimation to improve the localization in terms of both stability and precision. Further, we adopt a task-specific transformer architecture by designing multi-frame masked attention mechanism, such that the long-distance relationship between facial features can be effectively captured across frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 129, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 129, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 129, + 653 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 115, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 115, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 115, + 670 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "In this section, we elaborate on the design of the proposed approach, VRetouchEr, for high-quality face retouching in videos. Given a raw image sequence " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " and the cor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": "responding retouched one " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": ", our goal is to precisely locate the imperfections in " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " and replace with the content consistent with the surrounding normal skin. The key idea behind VRetouchEr is to leverage the temporal context information to stabilize face retouching over video frames. To achieve this, we adopt an encoder-transformer-decoder based architecture for VRetouchEr as illustrated in Figure 2. An encoder " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " extracts features of the target frame " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " and multiple reference frames " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " in sequence " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": ", and a flow estimation network " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " is incorporated to model the motions of imperfections between consecutive frames. The estimated imperfection flow " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " is fed into a Flow-based Imperfection Localization (FIR) module to improve the estimation stability and precision over the sequence. The resulting imperfection map " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": ", together with extracted target feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " and reference feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": ", are fed into the latent transformer denoted by " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " to restore the clean face features. In each intermediate block, we can collect richer information of normal skin from the reference frames, and our Multi-frame Masked Attention (MMA) mechanism enables the model to concentrate on the most relevant parts. Finally, the decoder " + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": " rebuilds each target frame from the transformed features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 611, + 503, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 503, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 503, + 624 + ], + "type": "text", + "content": "3.2. Flow-based Imperfection Refinement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The FIR module is specifically designed to enhance the accuracy of imperfection localization by leveraging flow information. This process involves imperfection movement estimation, imperfection localization, and imperfection map refinement. The refinement is crucial for stable retouching in each target frame. To achieve this, we define the imperfections are the differences between the manually retouched" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9143" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "frames and raw frames, represented as " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ". We apply a pretrained SpyNet [31] on " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " to produce the ground truth imperfection flow, denoted as " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "O_{gt}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": "Our designed FIR module consists of an imperfection localization network " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": " and learnable align factors " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": ". Given frame " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "X^i" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": " and its consecutive frame " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "X^j" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": ", the flow estimation network " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": " is trained to predict an imperfection displacement map " + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "inline_equation", + "content": "O^{i \\to j}" + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": ", which is required to approximate the corresponding ground truth imperfection flow as accurately as possible. We evaluate the imperfection flow prediction by measuring the degree of consistency as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 212, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 212, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 97, + 212, + 287, + 232 + ], + "type": "interline_equation", + "content": "O ^ {i \\rightarrow j} = S \\left(X ^ {i}, X ^ {j}\\right), \\tag {1}", + "image_path": "d48572ef1119d03cc91d5fa5ac55d63a72b4425cffea6751fc49c80deedf92c8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 229, + 236, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 229, + 236, + 244 + ], + "spans": [ + { + "bbox": [ + 99, + 229, + 236, + 244 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f l o w} = \\mathbb {E} _ {X} \\left[ | O ^ {i \\rightarrow j} - O _ {g t} ^ {i \\rightarrow j} | _ {1} \\right],", + "image_path": "d06b797c801e8bbd1d6184ba121aab09050abada0e0662f7512ddfc14fd64376.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "text", + "content": "By minimizing " + }, + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 252, + 287, + 287 + ], + "type": "text", + "content": " is encouraged to focus more on the displacement of imperfections instead of background and normal skin regions between consecutive frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": "On the other hand, the imperfection localization network " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " is trained to detect imperfections from frame " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "X^i" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " together with the corresponding encoder feature " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "f_r^i" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": ". We observe that the detection results on individual frames are unstable when there are significant changes in facial postures. To address this issue, we utilize the information from consecutive frames and the estimated imperfection flow " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "O^{i \\to j}" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " to refine the detection results. Specifically, the imperfection flow " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "O^{i \\to j}" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " will warp " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "M^i" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " into a flow-based imperfection map " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^j" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 288, + 287, + 417 + ], + "type": "text", + "content": " denotes the warping operation, and we have:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 416, + 196, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 416, + 196, + 430 + ], + "spans": [ + { + "bbox": [ + 120, + 416, + 196, + 430 + ], + "type": "interline_equation", + "content": "M ^ {i} = N (X ^ {i}, f _ {r} ^ {i}),", + "image_path": "e7af51e527a59b995a3fa0ad4b9ae9929c4fa874db23811d34e24b72e13b6869.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 433, + 287, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 433, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 121, + 433, + 287, + 447 + ], + "type": "interline_equation", + "content": "M ^ {j} = N \\left(X ^ {j}, f _ {r} ^ {j}\\right), \\tag {2}", + "image_path": "d7570639aa95290839b66652f9cdbb9bf133f6a9682a2400a2f4b9dc7a273c64.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 449, + 215, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 449, + 215, + 462 + ], + "spans": [ + { + "bbox": [ + 119, + 449, + 215, + 462 + ], + "type": "interline_equation", + "content": "\\mathcal {M} ^ {j} = \\mathcal {W} (O ^ {i \\rightarrow j}, M ^ {i}),", + "image_path": "77224dd40d6fde8f538fcfd936909f98b506117de9a5c84b92652b099a739ddd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^j" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " denotes the warped imperfection map. Next, our designed align factors " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " are employed to align the flow-based imperfection map " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^j" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " with predicted imperfection map " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "M^j" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": ". We utilized adaptive convolution blocks [35], denoted as " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\theta (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": ", for an effective fusion of these features. The refinement operation is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 549, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 549, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 107, + 549, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {a} ^ {j} = \\theta (\\alpha * \\mathcal {M} ^ {j} + \\beta , \\mathcal {M} ^ {j}), \\tag {3}", + "image_path": "c851acfd519ee9578eaef4622ad7fe26f81ad213865f63a1c23e6c59df0eb67c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 108, + 567, + 206, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 567, + 206, + 581 + ], + "spans": [ + { + "bbox": [ + 108, + 567, + 206, + 581 + ], + "type": "interline_equation", + "content": "\\hat {M} ^ {j} = \\sigma \\left(\\theta \\left(M ^ {j}, \\mathcal {M} _ {a} ^ {j}\\right)\\right),", + "image_path": "80c2c43c1abd52d98b2c45305880276e0d7ba3819e1593085260801c08d7ae76.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\hat{M}^j" + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": " represents the refined imperfection map, " + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_a^j" + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": " is the aligned map and " + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": " represents the sigmoid function. All the refined imperfection maps will be stacked together to form the output, denoted as " + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": ". The elements in " + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 591, + 287, + 663 + ], + "type": "text", + "content": " are normalized in the range from 0 to 1, representing the refined estimation of the imperfections across frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 671, + 277, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 277, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 277, + 683 + ], + "type": "text", + "content": "3.3. Multi-frame Masked Attention Transformer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Our proposed MMA mechanism enables the latent transformer to modify the features " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " in the imperfection regions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "text", + "content": "by leveraging the features " + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "text", + "content": " extracted from multiple frames as references. To achieve this, we use the obtained imperfection map " + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "text", + "content": " to weight the features of target and reference frames, and perform cross-attention computation between them. Let the number of reference frames be " + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 72, + 545, + 165 + ], + "type": "text", + "content": ", the weighted target feature is used as query, and the weighted reference ones are used as keys and values, defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 167, + 466, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 167, + 466, + 179 + ], + "spans": [ + { + "bbox": [ + 356, + 167, + 466, + 179 + ], + "type": "interline_equation", + "content": "Q _ {t} = W _ {q} \\left(f _ {t} \\otimes M _ {t} + b _ {q}\\right),", + "image_path": "206b11b73fdb0eea07cd6b48092cece87aaa9c257691d5d63b4c1d9f0c56f230.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 181, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 181, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 356, + 181, + 545, + 196 + ], + "type": "interline_equation", + "content": "K _ {r} ^ {i} = W _ {k} \\left(f _ {r} ^ {i} \\otimes \\left(1 - M _ {r} ^ {i}\\right) + b _ {k}\\right), \\tag {4}", + "image_path": "3db75097cf348c1395d9c4a137fc8e305bb29760d68500fca41957bf1663889f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 198, + 495, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 198, + 495, + 213 + ], + "spans": [ + { + "bbox": [ + 358, + 198, + 495, + 213 + ], + "type": "interline_equation", + "content": "V _ {r} ^ {i} = W _ {v} \\left(f _ {r} ^ {i} \\otimes (1 - M _ {r} ^ {i}) + b _ {v}\\right),", + "image_path": "79f20f978d28440e32f0a1e437d6c368230c46b0da9de85f9124b98d30310934.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 304, + 217, + 545, + 265 + ], + "type": "text", + "content": " represent the refined imperfection map for target frame and reference frames correspondingly. Next, we perform the cross-attention computation to obtain the modification maps as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 333, + 275, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 275, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 333, + 275, + 545, + 308 + ], + "type": "interline_equation", + "content": "\\Delta_ {f _ {t}} = \\operatorname {s o f t m a x} \\left(Q _ {t} \\cdot \\sum_ {i} ^ {\\delta} K _ {r} ^ {i} / \\sqrt {\\Lambda}\\right) \\cdot \\sum_ {i} ^ {\\delta} V _ {r} ^ {i}, \\tag {5}", + "image_path": "826f167bf3262e1b404beaf308d864fbf6a9c11c8824f6dd1a65ce007876eef7.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "content": " denotes the channel number of the features. In Eq.(4), we adopt the imperfection map " + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "content": " as a soft mask to isolate the regions to be edited, " + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "inline_equation", + "content": "(1 - M)" + }, + { + "bbox": [ + 304, + 316, + 545, + 471 + ], + "type": "text", + "content": " represents the mask for imperfection regions and normal regions respectively. Based on this, we can realize the goal of replacing imperfection regions with normal skin by strengthening the interactions between the target features associated with imperfections and the reference features associated with normal skin. Furthermore, we use additive attention in Eq.(5) with linear complexity to aggregate the key and value vectors from multiple reference frame into global vectors, which is used to perform element-wise multiplication to induce temporal context information." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": "Considering that the background and normal skin in the target frame should be preserved properly, the original feature maps are weighted combined with the modification maps as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 355, + 529, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 529, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 355, + 529, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {t} = f _ {t} \\otimes (1 - M _ {t}) + \\Delta_ {f _ {t}} \\otimes M _ {t}, \\tag {6}", + "image_path": "14960057fe51afc32e43e5c26ce181d5d2a24f6083b5df021c3ac7e36c97f879.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 554, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 637 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 554, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\hat{f}_t" + }, + { + "bbox": [ + 304, + 554, + 545, + 637 + ], + "type": "text", + "content": " denotes the retouched feature, can be fed into the subsequent transformer block as input target feature for further refinement. In this process, the masked attention mechanism enables effective information exchange across frames, and the imperfections in the target frame are progressively suppressed and replaced with the content learnt from normal skin." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 647, + 402, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 402, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 402, + 659 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The training loss functions of VRetouchEr involves the following three aspects: imperfection flow estimation, imperfection localization and retouching evaluation. As illustrated in Subsection 3.2, we employ a pretrained SpyNet" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9144" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "to produce the ground truth of imperfection flow estimation, and the network " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " takes consecutive frames as input and predicts the flow as accurately as possible. In addition, the flow is used to improve the imperfection localization on each frame, and we evaluate the precision by measuring the discrepancy between the prediction " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and the ground truth imperfections " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 162, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 162, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 287, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m p} = \\mathbb {E} _ {X} \\| \\mathcal {I} (M) - M _ {g t} \\| _ {1}, \\tag {7}", + "image_path": "9cfe97390bc2ccb2beecf984ee1ab06a5e609a74d6584d2118bc133d2373ae5d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": " denotes a learnable layer to align " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": "'s channel number with " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": ". We optimize " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{imp}" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": " together with the flow estimation loss " + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 47, + 180, + 287, + 228 + ], + "type": "text", + "content": " defined in Eq.(2), and the formulation is expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 235, + 287, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 235, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 124, + 235, + 287, + 253 + ], + "type": "interline_equation", + "content": "\\min _ {S, N} \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {i m p}}. \\tag {8}", + "image_path": "572a56a03acdb497aaccebdb1560f6ecb4998abdf91f25bc113987fc85dbc2e4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{y} = G(\\hat{f}_t)" + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": " denote the retouched frame synthesized by VRetouchEr. We perform retouching on every raw frame and combine the results together to get the retouched sequence " + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": ", which is evaluated by measuring the degree of consistency with the manually retouched frames " + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 260, + 287, + 331 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 337, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 337, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 71, + 337, + 287, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = \\mathbb {E} _ {X} \\left[ \\zeta \\| Y - \\widehat {Y} | _ {1} + \\| \\mathcal {V} (Y) - \\mathcal {V} (\\widehat {Y}) \\| _ {2} ^ {2} \\right], \\tag {9}", + "image_path": "dbc541c0c90aa578ffd29a6eeec16227c1708907d528abf9de16fe8b9e7fe845.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "text", + "content": " denotes a weighting factor, and " + }, + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot)" + }, + { + "bbox": [ + 46, + 357, + 287, + 416 + ], + "type": "text", + "content": " represents the features extracted from a pre-trained VGG-19. Considering that high-fidelity content synthesis can benefit from adversarial training, we thus adopt the real-fake discrimination loss formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 421, + 214, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 421, + 214, + 436 + ], + "spans": [ + { + "bbox": [ + 110, + 421, + 214, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {s y n t} = \\mathbb {E} _ {X} \\bigl [ \\log (D (\\hat {Y})) \\bigr ]", + "image_path": "9fbdb651a453a406fd903dcd422301aab62cd25248aca7a2dba822ee8db410c5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 436, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 436, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 436, + 287, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {r e a l} = \\mathbb {E} _ {X} [ \\log (1 - D (X)) ] + \\mathbb {E} _ {Y} [ \\log (D (Y)) ], \\tag {10}", + "image_path": "958e626c01ca93f27a30be7f92e1b5623c31dbc27004296c572e87e15f242720.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 456, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 527 + ], + "type": "text", + "content": "where discriminator " + }, + { + "bbox": [ + 46, + 456, + 287, + 527 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 456, + 287, + 527 + ], + "type": "text", + "content": " takes the sequence as input, and is trained to distinguish the manually retouched frames from the synthesized ones. By integrating the loss functions of retouching evaluation and adversarial training, the optimization formulation of the constituent networks is expressed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 532, + 208, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 532, + 208, + 546 + ], + "spans": [ + { + "bbox": [ + 126, + 532, + 208, + 546 + ], + "type": "interline_equation", + "content": "\\min _ {E T C} \\mathcal {L} _ {c o n} + \\mathcal {L} _ {a d v} ^ {s y n},", + "image_path": "28701c23135abd633d779581cb92ef3fe3e866d39f582d33df8af1b1ce330574.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 129, + 545, + 287, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 545, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 129, + 545, + 287, + 556 + ], + "type": "interline_equation", + "content": "E, I, G \\tag {11}", + "image_path": "98be776cc4bf606c35025fc52fb4161b45a3a2ed1d1c07f15a60f0de8ea36b23.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 554, + 178, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 554, + 178, + 571 + ], + "spans": [ + { + "bbox": [ + 129, + 554, + 178, + 571 + ], + "type": "interline_equation", + "content": "\\max _ {D} \\mathcal {L} _ {a d v} ^ {r e a l}.", + "image_path": "30707367775ef8e4afa46baee9bd235a8fd063e650c1b3b9e86fc2031f2f7e9d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 576, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 601 + ], + "type": "text", + "content": "The constituent networks of the proposed VRetouchEr are jointly optimized from scratch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 609, + 123, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 123, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 123, + 623 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "In this section, we perform extensive experiments to assess the effectiveness of VRetouchEr in face video retouching. We first present the details of the datasets and the experimental setup. Next, we compare our VRetouchEr with previous state-of-the-art face retouching methods quantitatively and qualitatively. Finally, we provide insights by analyzing the design elements of VRetouchEr." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 433, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 433, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 433, + 85 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": "Datasets. Due to the difficulty in collecting paired raw-retouched image sequences, we utilized the Flickr-Face-HQ-Retouching (FFHQR) dataset [33] to construct emulated video data. The training/validation/test data in FFHQR consists of " + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "inline_equation", + "content": "56\\mathrm{k} / 7\\mathrm{k} / 7\\mathrm{k}" + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": " raw-retouched image pairs. For each image pair " + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\{x,y\\}" + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": " in FFHQR, we perform random cropping, flipping and translation on both " + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": " to generate a pair of image sequences " + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\{X,Y\\}" + }, + { + "bbox": [ + 304, + 89, + 545, + 293 + ], + "type": "text", + "content": ", and the resulting dataset is referred to FFHQR-Seq. In addition, we build another Manually Retouching Face Video dataset, referred to as MRFV, which contains 200 in-the-wild portrait videos with different types of facial imperfections, and each of them contains at least 500 frames. We employ multiple retouchers to manually retouch each frame of these videos, and evaluate the retouching performance of our VRetouchEr and the competing methods trained on FFHQR-Seq." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "text", + "content": "Training Details. The images in both training and test sequences are resized to " + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "text", + "content": " for a fair comparison with the existing face retouching methods. In the training process, the parameters of VRetouchEr are updated by the Adam optimizer [20] with the learning rate of " + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "text", + "content": ". There are a total of 400k training iterations, the batch size is set to 1, and the hyper-parameter " + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 304, + 293, + 545, + 400 + ], + "type": "text", + "content": " in Eq.(9) is set to 10. We implement VRetouchEr by using PyTorch and train it on single NVIDIA GeForce GTX 3090 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "content": "Evaluation Protocols. In the experiments, we implement all the competing methods by using the open source codes. We adopt the Soft Intersection over Union Loss (Soft-IoU) to measure the correctness of imperfection localization. To measure the consistency between the synthesized image sequences and the manually retouched ones, we adopt the widely used metrics: Peak Signal-to-Noise Ratio (PSNR), Structure Similarity Index Measure (SSIM) and the Learnt Perceptual Image Patch Similarity (LPIPS) [49]. Furthermore, we quantitatively assess the diversity and the degree of realism of synthesized image sequences in terms of Video Fréchet Inception Distance (VFID) [40]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 552, + 478, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 478, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 478, + 564 + ], + "type": "text", + "content": "4.2. Comparison to State-of-the-arts" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We perform a comprehensive comparative analysis between VRetouchEr and representative state-of-the-art methods, including a typical image translation method: Pix2PixHD [41], generic image restoration methods: MPR-Net [48] and RestoreFormer [42], a blind face restoration method: GPEN [47], video enhancement methods: BIP-Net [11] and ProPainter [51], and face retouching methods: AutoRetouch [33] and BPFRe [45]. We train BIP-Net, ProPainter and VRetouchEr on FFHQR-Seq, and the remaining methods on FFHQR. All the trained models are evaluated on the test data of FFHQR-Seq and MRFV. We quantitatively assess the retouch performance of the meth" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9145" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 74, + 518, + 248 + ], + "blocks": [ + { + "bbox": [ + 75, + 74, + 518, + 248 + ], + "lines": [ + { + "bbox": [ + 75, + 74, + 518, + 248 + ], + "spans": [ + { + "bbox": [ + 75, + 74, + 518, + 248 + ], + "type": "image", + "image_path": "dff76043d0a875577c4b5b2f37f8af051cea56a1584aebee61ae36d072a9f0fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 250, + 529, + 262 + ], + "lines": [ + { + "bbox": [ + 63, + 250, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 63, + 250, + 529, + 262 + ], + "type": "text", + "content": "Figure 3. Visual comparison between VRetouchEr and the competing methods on representative frames of an in-the-wild videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 270, + 260, + 458 + ], + "blocks": [ + { + "bbox": [ + 72, + 270, + 260, + 458 + ], + "lines": [ + { + "bbox": [ + 72, + 270, + 260, + 458 + ], + "spans": [ + { + "bbox": [ + 72, + 270, + 260, + 458 + ], + "type": "image", + "image_path": "f0981a464a5cfa44e197b48809a73ebfe4172391a1cac6211455321be4496014.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 463, + 288, + 497 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 497 + ], + "type": "text", + "content": "Figure 4. Representative high-quality retouching results. From top left to bottom right, the images are the raw frame and the retouching results of BPFRe, ProPainter and VRetouchEr, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 526, + 287, + 682 + ], + "blocks": [ + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between VRetouchEr and competing methods on FFHQR-Seq." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 526, + 287, + 682 + ], + "lines": [ + { + "bbox": [ + 50, + 526, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 50, + 526, + 287, + 682 + ], + "type": "table", + "html": "
MethodsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs37.500.97040.033710.779
Pix2PixHD [41]37.630.97140.03039.252
AutoRetouch [33]38.270.97710.02618.941
MPRNet [48]38.330.97420.02678.231
GPEN [47]38.270.97390.02568.125
RestoreFormer [42]38.460.97330.02327.931
BIPNet [11]38.140.97110.02758.241
ProPainter [51]38.540.97680.02407.872
BPFRe [45]38.690.97740.02197.604
VRetouchEr39.750.98130.01696.375
", + "image_path": "d2d7c0a17486317ea8fc30cbffc9c2594c230cdda59580229bd083abebd6e24b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "ods per frame, and report their average PSNR, SSIM, LPIPS and VFID scores in Tables 1-2. On FFHQR-Seq, the com" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 293, + 545, + 449 + ], + "blocks": [ + { + "bbox": [ + 306, + 270, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 292 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between VRetouchEr and competing methods on MRFV." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 293, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 309, + 293, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 309, + 293, + 545, + 449 + ], + "type": "table", + "html": "
MethodsMRFV
PSNR↑SSIM↑LPIPS ↓VFID↓
Raw pairs32.070.90540.089141.830
Pix2PixHD [41]33.510.91280.073938.403
AutoRetouch [33]35.030.92110.050231.244
MPRNet [48]35.820.93270.049129.893
GPEN [47]35.710.92810.046330.135
RestoreFormer [42]35.760.92990.051629.785
BIPNet [11]35.260.92310.059333.897
ProPainter [51]36.270.93840.045220.344
BPFRe [45]36.320.94270.040118.772
VRetouchEr37.630.95300.035710.368
", + "image_path": "7f120d226b53e52f94d061377a0e90e3427eb528b9c0fdc64eb7d064291ee659.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "content": "peting methods achieve similar retouching performance except Pix2PixHD, and our VRetouchEr outperforms them in terms of all the metrics. In particular, VRetouchEr achieves the PSNR score of 39.75, which is higher than that of the second best method: BPFRe, by " + }, + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "inline_equation", + "content": "1.06\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "content": ". On MRFV, VRetouchEr also achieves superior retouching performance over the competing methods. The VFID score of VRetouchEr reaches 10.368, which is lower than that of BPFRe by a significant improvement of about 44.77 percentage points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "In Figure 3, we perform the detailed visual comparison. We can make the following observations: Most of the competing methods fail to remove the marked acnes. BPFRe falls short in maintaining the stability of retouching in videos, since it is designed for single-image retouching. Although ProPainter is specifically designed for video enhancement, it fails to achieve satisfactory retouching results due to lack of imperfection localization. In contrast, our VRetouchEr is able to consistently deliver stable and high-quality retouching outcomes. In Figure 4, we provide visual examples to further compare the performance of our framework with the main competing methods in removing dense" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9146" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 273, + 180 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 273, + 180 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 273, + 180 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 273, + 180 + ], + "type": "image", + "image_path": "768685824ca83c75b6e039eed6ef525f29b4a191b37078ee167cc0a4ab879c21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 182, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 287, + 216 + ], + "type": "text", + "content": "Figure 5. Quantitative comparison in imperfection localization per frame. VRetoucher achieves a more precise and stable result than 'VRetoucher w/o FIR' and BPFRe." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 221, + 272, + 331 + ], + "blocks": [ + { + "bbox": [ + 58, + 221, + 272, + 331 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 272, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 272, + 331 + ], + "type": "image", + "image_path": "6d698e74db156da10dc7fb0d9db6f63e3798116c5781b36c77f302971418a49a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 334, + 286, + 355 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 286, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 286, + 355 + ], + "type": "text", + "content": "Figure 6. Quantitative comparison in imperfection localization on MRFV dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 286, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 286, + 401 + ], + "type": "text", + "content": "imperfections. These results highlight the VRetouchEr's capability of face retouching in videos even without any manually retouched videos as training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 408, + 122, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 122, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 122, + 422 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": "We further perform a user study to evaluate VRetouchEr and the competing methods in terms of human perception. The models trained on FFHQR/FFHQR-Seq are applied to a total of 40 raw videos randomly sampled from MRFV. There are 50 participants, each of which is presented with the synthesized videos and required to carefully compare and sort the retouching results, according to the criteria including overall visual quality, effectiveness of imperfection removal, naturalness, and consistency. Table 3 shows the ranking results, and we are able to draw the conclusion regarding the superior performance of VRetouchEr against the competing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 244, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 244, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 244, + 590 + ], + "type": "text", + "content": "4.4. Flow-based Imperfection Refinement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 596, + 286, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 286, + 620 + ], + "type": "text", + "content": "We observed that our FIR module plays in an important role in facilitating face retouching in videos. Different from" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 52, + 644, + 285, + 712 + ], + "blocks": [ + { + "bbox": [ + 67, + 632, + 266, + 643 + ], + "lines": [ + { + "bbox": [ + 67, + 632, + 266, + 643 + ], + "spans": [ + { + "bbox": [ + 67, + 632, + 266, + 643 + ], + "type": "text", + "content": "Table 3. The voting result " + }, + { + "bbox": [ + 67, + 632, + 266, + 643 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 67, + 632, + 266, + 643 + ], + "type": "text", + "content": " of user study on MRFV." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 644, + 285, + 712 + ], + "lines": [ + { + "bbox": [ + 52, + 644, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 52, + 644, + 285, + 712 + ], + "type": "table", + "html": "
MethodRank-1Rank-2Rank-3Rank-4Rank-5Rank-6
VRetouchEr88.4511.050.300.200.000.00
BPFRe [45]11.1563.1519.355.051.200.10
ProPainter [51]0.2518.2061.6016.303.150.50
MPRNet [48]0.056.0511.6559.256.4016.60
RestoreFormer [42]0.051.506.1011.2561.3019.80
BIPNet [11]0.050.051.007.9527.9563.00
", + "image_path": "1bfe29063c0e2a525ff5e75ec4a192190deb7accbba992f3b0c9846b482f8476.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 531, + 203 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 531, + 203 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 531, + 203 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 531, + 203 + ], + "type": "image", + "image_path": "f923883ed0edd53d967130d676c238c229a8a2f78de5df5531ab5c3eb8bf81cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 206, + 529, + 217 + ], + "lines": [ + { + "bbox": [ + 321, + 206, + 529, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 529, + 217 + ], + "type": "text", + "content": "Figure 7. Visual comparison in imperfection localization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "text", + "content": "BPFRe detecting facial imperfections on each frame separately, we estimate imperfection flow between consecutive frames followed by imperfection refinement. The displacement information is useful for refining and stabilizing the imperfection maps of the video to be edited. In Figure 5, we quantitatively evaluate imperfection localization results on a video sequence in terms of Soft-IoU per-frame between the results and the ground truth imperfection maps " + }, + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "text", + "content": " as defined in Sec. 3.2. In Figure 6, we perform the evaluation of imperfection localization between the results and " + }, + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 304, + 228, + 545, + 468 + ], + "type": "text", + "content": " in terms of mean SSIM and mean Soft-IoU on MRFV dataset. When disabling FIR, the variant 'VRetouchEr w/o FIR' performs single-image imperfection localization, and has comparable performance with BPFRe. We can find that FIR leads to a significant and stable improvement for cross-frame imperfection localization. We further visualize the imperfection maps of BPFRe, 'VRetouchEr w/o FIR' and VRetouchEr in Figure 7, and observe that VRetouchEr can stably locate the most imperfections, and its results are more consistent with the ground truth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 479, + 532, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 532, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 532, + 491 + ], + "type": "text", + "content": "4.5. Impact of the Number of Reference Frames" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "The existing face retouching methods suffer from the problem of overlooking temporal context information in handling videos. As another important module of VRetouchEr, we consider that our designed MMA mechanism also contributes to the superior performance of VRetouchEr. To verify this, we investigate the impact of the number of reference frames " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " on the final retouching performance. In Figure 8, we plot the VFID scores of VRetouchEr on MRFV during training. For the case of " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 1" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", we perform single-image retouching on each frame. One can observe that increasing the value of " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " for MMA indeed leads to a significant improvement. In particular, VRetouchEr achieves the best retouching performance when " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 6" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". Beyond this value, VRetouchEr's performance becomes stable. In Figure 9, we visualize the retouching results to further verify this observation. This insight provides a valuable guidance for optimizing the performance of VRetouchEr in the face video retouching task." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9147" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 286, + 201 + ], + "type": "image", + "image_path": "fd8342aa4ac1cff3c1828e35e6372dae4713f278ba94a75ecc0747995fa32a09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 279, + 216 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 279, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 279, + 216 + ], + "type": "text", + "content": "Figure 8. Quantitative results of VRetouchEr with different " + }, + { + "bbox": [ + 54, + 205, + 279, + 216 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 205, + 279, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 228, + 287, + 410 + ], + "blocks": [ + { + "bbox": [ + 47, + 228, + 287, + 410 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 410 + ], + "type": "image", + "image_path": "e7d1125df38e4cca0c9f3c13def2b8bf059f55384a15a4809e44dc10c12d91f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 412, + 285, + 423 + ], + "lines": [ + { + "bbox": [ + 49, + 412, + 285, + 423 + ], + "spans": [ + { + "bbox": [ + 49, + 412, + 285, + 423 + ], + "type": "text", + "content": "Figure 9. The retouching results of VRetouchEr with different " + }, + { + "bbox": [ + 49, + 412, + 285, + 423 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 49, + 412, + 285, + 423 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 441, + 140, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 140, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 140, + 454 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": "The two proposed modules: FIR and MMA, distinguish the proposed VRetouchEr from the existing works. We perform a set of ablative experiments to investigate the impact of FIR and MMA on the retouching performance on FFHQR-Seq. Specifically, we build the Base model of VRetouchEr by disabling both FIR and MMA, and find that VRetouchEr significantly outperforms the base model as shown in Table 4. Additionally, we build two variants by disabling FIR and MMA, and the resulting models are referred to as 'VRetouchEr w/o FIR' and 'VRetouchEr w/o MMA', respectively. In 'VRetouchEr w/o FIR', the maps produced by the imperfection localization network " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " are directly fed into the latent transformer. As analyzed in Sec.4.4, disabling FIR leads to unstable imperfection localization, which ultimately degrades the retouching performance on videos. As shown in Table 4, the PSNR score of 'VRetouchEr w/o FIR' is 38.74 dB, which is worse than that of VRetouchEr by 1.01 dB. When disabling MMA, we also observe a significant performance drop of about 40.83 percentage points in terms of LPIPS. Qualitative results in Figure 10 also prove the effectiveness of FIR and MMA." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 79, + 547, + 161 + ], + "blocks": [ + { + "bbox": [ + 338, + 60, + 512, + 70 + ], + "lines": [ + { + "bbox": [ + 338, + 60, + 512, + 70 + ], + "spans": [ + { + "bbox": [ + 338, + 60, + 512, + 70 + ], + "type": "text", + "content": "Table 4. Quantitative results of ablative models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 79, + 547, + 161 + ], + "lines": [ + { + "bbox": [ + 309, + 79, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 309, + 79, + 547, + 161 + ], + "type": "table", + "html": "
VariantsFFHQR-Seq
PSNR↑SSIM↑LPIPS ↓VFID↓
Base model38.130.97380.02648.671
VRetouchEr w/o MMA38.670.97580.02387.534
VRetouchEr w/o FIR38.740.97690.02077.312
VRetouchEr39.750.98130.01696.375
", + "image_path": "ddcacb34fb806dfab4e41899ce9812108ff4ffedd8a8c28747a5e5b18b50bf32.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 171, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 306, + 171, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 311 + ], + "type": "image", + "image_path": "aaa3a5f66eb10092b82d621a4803456069bc5761967082c2b45b47226a3a7188.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 315, + 541, + 326 + ], + "lines": [ + { + "bbox": [ + 309, + 315, + 541, + 326 + ], + "spans": [ + { + "bbox": [ + 309, + 315, + 541, + 326 + ], + "type": "text", + "content": "Figure 10. Representative retouching results of ablative models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 346, + 378, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 378, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 378, + 358 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 581 + ], + "type": "text", + "content": "In this work, we propose a latent retouching transformer for removing facial imperfections in videos. Different from the existing methods that focus on single-image retouching, we leverage the temporal context information to significantly facilitate face retouching in videos. Toward this end, we perform imperfection flow estimation to obtain the displacement information between consecutive frames, and further refine the imperfection localization by fusing motion-based and frame-based predictions. By injecting the imperfection maps as side information into a latent retouching transformer, we adopt the masked attention computation over multiple frames, such that the features of normal skin from different frames can be leveraged to substitute for that of imperfections in the target frame. Extensive experiments are performed to demonstrate the superior capability of the proposed approach over state-of-the-art face retouching methods in stabilizing the retouching performance over video frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 593, + 415, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 593, + 415, + 606 + ], + "spans": [ + { + "bbox": [ + 306, + 593, + 415, + 606 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 613, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 708 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2020A1515010484, 2022A1515011160), and in part by TCL Science and Technology Innovation Fund (Project No. 20231752)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9148" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Kaoru Arakawa. Nonlinear digital filters for beautifying facial images in multimedia systems. In Proc. IEEE International Symposium on Curcuits and Systems, 2004. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "type": "text", + "content": "[2] Nazre Batool and Rama Chellappa. Detection and inpainting of facial wrinkles using texture orientation fields and Markov random field modeling. IEEE Transactions on Image Processing, 23(9):3773-3788, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "text", + "content": "[3] Goutam Bhat, Martin Danelljan, Luc van Gool, and Radu Timofte. Deep burst super-resolution. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 317 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "text", + "content": "[6] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "text", + "content": "[8] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 508 + ], + "type": "text", + "content": "[9] Jocob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In arXiv:1801.04805, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 586 + ], + "type": "text", + "content": "[10] Alexy Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "text", + "content": "[11] Akshay Dudhane, Syed Waqas Zamir, Salman Khan1, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burst image restoration and enhancement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive im" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "age properties. In Proc. International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[15] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[16] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[17] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 304 + ], + "type": "text", + "content": "[18] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"steerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[19] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[20] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "text", + "content": "[21] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: Adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 481 + ], + "type": "text", + "content": "[22] Tommer Leyvand, Daniel Cohen-Or, Gideon Dror, and Dani Lischinski. Data-driven enhancement of facial attractiveness. In Proc. ACM Conference on Special Interest Group on Computer Graphics and Interactive Techniques, 2008. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 536 + ], + "type": "text", + "content": "[23] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In Proc. European Conference on Computer Vision, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[24] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo1, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[25] Jing Lin, Yuanhao Cai, Xiaowan Hu, Haoqian Wang, and Youliang Yan. Flow-guided sparse transformer for video deblurring. In Proc. International Conference on Machine Learning, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[26] Jing Lin, Xiaowan Hu, Yuanhao Cai, and Haoqian Wang. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In Proc. International Conference on Machine Learning, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. EditGAN: high-precision semantic image editing. In Proc. Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9149" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Uri Lipowezky and Sarah Cahen. Automatic freckles detection and retouching. In Proc. IEEE Convention of Electrical and Electronics Engineers in Israel, 2008. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "type": "text", + "content": "[31] Anurag RanjanMichael and J. Black. Optical flow estimation using a spatial pyramid network. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 275 + ], + "type": "text", + "content": "[32] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGan encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 287, + 319 + ], + "type": "text", + "content": "[33] Alireza Shafaei, James J. Little, and Mark Schmidt. AutoRetouch: automatic professional face retouching. In Proc. IEEE Winter Conference on Applications of Computer Vision, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 287, + 365 + ], + "type": "text", + "content": "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 287, + 409 + ], + "type": "text", + "content": "[35] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "text", + "content": "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 456, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 500 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 502, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 502, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 502, + 287, + 555 + ], + "type": "text", + "content": "[38] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. FabSoften: face beautification via dynamic skin smoothing, guided feathering and texture restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition Workshop, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 610 + ], + "type": "text", + "content": "[39] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: end-to-end panoptic segmentation with mask transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 657 + ], + "type": "text", + "content": "[40] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 713 + ], + "type": "text", + "content": "[41] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 676 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[42] Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. RestoreFormer: high-quality blind face restoration from undegraded key-value pairs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[43] Pengxu Wei, Yujing Sun, Xingbei Guo, Chang Liu, Guanbin Li, Jie Chen, Xiangyang Ji, and Liang Lin. Towards real-world burst image super-resolution: Benchmark and method. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 217 + ], + "type": "text", + "content": "[44] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvR: introducing convolutions to vision transformers. arXiv:2103.15808, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 272 + ], + "type": "text", + "content": "[45] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face retouching with limited paired data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "text", + "content": "[46] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "text", + "content": "[47] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. GAN prior embedded network for blind face restoration in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 418 + ], + "type": "text", + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Huang Yang, and Ling Shao. Multi-stage progressive image restoration. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "type": "text", + "content": "[50] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "text", + "content": "[51] Shangchen Zhou, Chongyi Li, Kelvin C.K. Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 587, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 631 + ], + "type": "text", + "content": "[52] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 632, + 545, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 676 + ], + "type": "text", + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9150" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_content_list.json b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dc37730669a66cc268f28764a11160124a476a67 --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "VSCode: General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning", + "text_level": 1, + "bbox": [ + 143, + 128, + 826, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyang Luo $^{1}$ Nian Liu $^{2,*}$ Wangbo Zhao $^{3}$ Xuguang Yang $^{1}$ Dingwen Zhang $^{1}$ Deng-Ping Fan $^{5,6}$ Fahad Khan $^{2,4}$ Junwei Han $^{1,7,*}$", + "bbox": [ + 183, + 202, + 825, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwestern Polytechnical University $^{2}$ Mohamed bin Zayed University of Artificial Intelligence $^{3}$ National University of Singapore $^{4}$ CVL, Linköping University $^{5}$ TBI Center, CS, Nankai University $^{6}$ Nankai International Advanced Research Institute (SHENZHEN FUTIAN) $^{7}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 84, + 239, + 915, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD. Source code has been available at https://github.com/Sssssuperior/VSCode.", + "bbox": [ + 75, + 376, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 659, + 209, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual salient object detection (SOD) and camouflaged object detection (COD) are two interconnected yet unique tasks. The goal of SOD is to identify prominent objects within an image that significantly contrast with their surroundings [5], which can be used to promote segmentation [46, 111], detection [94], and Part-Object Relational visual saliency [57, 58]. While COD focuses on identifying objects concealed within their environment. These objects intentionally blend in by sharing structural or textural similarities with their surroundings [15]. Despite the seemingly different definitions of SOD and COD, they both belong to the realm of binary segmentation and share some vital fundamental similarities, such as", + "bbox": [ + 75, + 684, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d9dcc601fd71069722e03084033f4d8f3f14eb25d1c0a16179f54b2847a67cec.jpg", + "image_caption": [ + "Figure 1. Relationship of SOD, COD, and multimodal tasks. Each specific task is seen as a combination of two dimensions, i.e. domain (RGB/Depth/Thermal/Flow) and task (SOD/COD)." + ], + "image_footnote": [], + "bbox": [ + 501, + 342, + 893, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objectness and structuredness.", + "bbox": [ + 500, + 566, + 700, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To cater to various scenarios, both SOD and COD have given rise to several sub-tasks with different modalities, including RGB SOD [83, 108], RGB COD [27, 66, 101], RGB-D SOD [42, 71], RGB-D COD [92], and RGB-T SOD [80, 107]. By leveraging optical flow maps, Video SOD (VSOD) [44, 91] and VCOD [10, 36] tasks can also be seen as a combination of two modalities. The relationship of SOD, COD, and multimodal tasks is shown in Figure 1, where each specific task can be considered as a combination of two dimensions, i.e. domain and task. Although these multimodal tasks differ in the complementary cues they employ, these modalities share some key commonalities. For instance, depth, thermal, and optical flow maps often show obvious objectness as in RGB images.", + "bbox": [ + 496, + 580, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although previous CNN-based [10, 61, 71, 77, 86, 109] and transformer-based [54, 116] approaches have effectively addressed these tasks and achieved favorable results, they usually rely on meticulously designed models to tackle each task individually. Designing models specifically for individual tasks can be problematic since the training data of one task is typically limited. Task-specific specialist models", + "bbox": [ + 496, + 794, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Nian Liu (liunian228@gmail.com) and Junwei Han (junweihan2010@gmail.com)", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17169", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90588295a26c4c21545a4a8b1af0048b5eef90aa5c2d19338e6c1063d0e4abba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 88, + 416, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1787677a2f090fef175c757d803891dbd6a1c216ffcc84e119d4ff67a546af24.jpg", + "image_caption": [ + "Figure 2. Overall architecture of our VSCode model. We use VST [54] as the foundation model to acquire commonalities among multimodal SOD and COD tasks. For each task, we integrate 2D prompts to aggregate peculiarities along the domain dimension and the task dimension, including four domain-specific prompts and two task-specific prompts." + ], + "image_footnote": [], + "bbox": [ + 80, + 160, + 467, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "may be overly adapted to a particular task and overfitted to specific training data distribution, which ultimately sacrifices generalization ability and results in suboptimal performance. One solution may be using more data, however, being costly and time-consuming for data annotation. To this end, joint learning a generalist model emerges as a more promising option, as it allows for the maximum use of all data and the effective learning of the commonalities of all tasks, hence significantly reducing the risk of overfitting and enhancing the generalization capability [34, 55]. However, joint learning multiple tasks is not straightforward. On one hand, simultaneously handling both commonalities and peculiarities of all tasks poses a significant challenge as the incompatibility among different tasks easily leads to a decline in performance with simple joint training [40]. On the other hand, it usually introduces additional complexity, computational costs, and parameters.", + "bbox": [ + 75, + 353, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a general Visual Salient and Camouflaged object detection (VSCode) model which encapsulates both commonalities and peculiarities of different tasks with a simple but effective design, as illustrated in Figure 2. On one hand, we adopt VST [54] as the shared segmentation foundation model to assimilate commonalities of different tasks by leveraging its simple and pure-transformer-based architecture. On the other hand, inspired by the recent emergence of the parameter-efficient prompting technique [31, 63, 112], we propose 2D prompts to capture task peculiarities. Specifically, we decompose these peculiarities along the domain dimension and the task dimension, and consequently design domain-specific prompts and task-specific prompts to comprehend the differences among diverse domains and tasks, respectively. These 2D prompts can effectively disentangle domain and task peculiarities, making our model easily adaptable by combining them to tackle specific tasks and even unseen ones. Furthermore, we present a prompt discrimination loss to encourage the 2D prompts", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to focus on acquiring adequate peculiarities and enable the foundational model to concentrate on commonality learning.", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we train our VSCode model on four SOD tasks and two COD tasks, demonstrating its effectiveness against state-of-the-art methods. What's more, we carry out evaluations on a reserved task and reveal remarkable zero-shot generalization ability of our model, which has never been explored in previous works. The main contributions in this work can be summarized as follows:", + "bbox": [ + 496, + 123, + 893, + 228 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present VSCode, the first generalist model for multimodal SOD and COD tasks.", + "- We propose to use a foundation segmentation model to aggregate commonalities and introduce 2D prompts to learn peculiarities along the domain and task dimensions, respectively.", + "- A prompt discrimination loss is proposed to effectively enhance the learning of peculiarities and commonalities for 2D prompts and the foundation model, respectively.", + "- Our VSCode model surpasses all existing state-of-the-art models across all tasks on 26 datasets and showcases its ability to generalize to unseen tasks, further emphasizing the superiority of our approach." + ], + "bbox": [ + 517, + 232, + 893, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 440, + 640, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Learning Based SOD and COD", + "text_level": 1, + "bbox": [ + 500, + 463, + 818, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SOD. Previous RGB SOD works delved into attention-based [9, 19, 49, 71, 108], multi-level fusion-based [20, 24, 67, 83, 102, 113], recurrent-model-based [8, 12, 48, 60, 82], and multi-task-based methods [29, 72, 84, 90, 104, 106, 110]. In the case of RGB-D SOD, some models [9, 42, 43, 52, 53, 71, 109] leveraged various attention mechanisms to incorporate depth cues into RGB features. With regard to RGB-T SOD, recent studies also introduced attention-based methods [77, 80] and multi-level fusion [76, 107] to excavate the relationship between RGB and thermal features. Regarding the VSOD task, some works [18, 21, 30, 74, 86] mined spatial-temporal and appearance cues. More recently, there was a growing trend where various research [28, 44, 51, 73] endeavored to incorporate optical flow for combining motion cues with appearance details. Consistent with recent studies, we treat optical flow as a form of modality information and view VSOD as a multimodal SOD task.", + "bbox": [ + 496, + 486, + 893, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "COD. Currently, COD has RGB COD, RGB-D COD, and VCOD tasks. RGB COD methods can be broadly categorized as multi-task-based approaches [61, 101], multi-input-based approaches [66, 114], and refinement-based approaches [15, 32]. The RGB-D COD task was initially introduced in [92], where depth inference models are adapted for object segmentation. For VCOD, prior studies segmented the moving camouflaged objects via dense optical flow [3, 4] or well-designed models [10, 36]. For a more comprehensive literature review, please refer to [16].", + "bbox": [ + 496, + 750, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17170", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1bc501445f115cb0c0cd9229560adb5ba2902e8b4b1afffc5ded6241d82cfb87.jpg", + "image_caption": [ + "Figure 3. Overall framework of our proposed VSCode model with 2D prompt learning. Based on the VST [54] foundation model, we insert the respective domain-specific prompts and task-specific prompts in the attention windows in the Swin transformer [59] encoder layers to learn domain and task-specific encoder features. The converter is used for multimodal feature fusion. Within the transformer decoder layers, task-specific prompts are appended to image feature tokens to perform task-specific decoding. We also provide detailed structures of an encoder layer $(i = 0)$ and a decoder layer $(j = 0)$ ." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt in Computer Vision", + "text_level": 1, + "bbox": [ + 76, + 348, + 328, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt was initially introduced in the field of NLP [6] and has been successfully integrated into computer vision tasks [25]. VPT [31] introduced a small amount of trainable parameters as prompts in the input space. ViPT [115] put forth the idea of modality-complementary prompts for task-oriented multi-modal tracking. Prior research has primarily focused on specific tasks, such as classification or tracking. In this paper, we propose to use 2D prompts for assembling different multimodal tasks and enable zero-shot generalization on unseen tasks, which has not been explored before.", + "bbox": [ + 75, + 367, + 470, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generalist Segmentation Architecture", + "text_level": 1, + "bbox": [ + 76, + 520, + 405, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, several generalist frameworks have emerged for a range of segmentation tasks using a variety of prompts [1]. On one hand, X-Decoder [117] utilized generic nonsemantic queries and semantic queries to decode different pixel-level and token-level outputs. UNINEXT [95] introduced three types of prompts, namely category names, language expressions, and reference annotations. On the other hand, Painter [87] and SegGPT [88] leveraged image-mask pairs from the same task as prompts. Unlike the approaches mentioned above, which mainly concentrate on task differences, our VSCode dissects unique characteristics based on both domain and task dimensions, leading to a more versatile design.", + "bbox": [ + 75, + 537, + 470, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of SOD and COD, EVP [56] introduced adaptors into the encoder and trained each task individually for various foreground segmentation tasks. Different from them, we consider not only multiple tasks but also multiple modalities and we train all tasks simultaneously.", + "bbox": [ + 75, + 734, + 470, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 818, + 210, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose VSCode with the aim of jointly training SOD and COD tasks in an efficient and effective way. We allow VST [54] to incorporate commonalities (Section 3.1), and utilize 2D prompts, which comprise domain", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "specific (Section 3.2) and task-specific prompts (Section 3.3), to encapsulate peculiarities. To accurately disentangle domain and task peculiarities in 2D prompts and encourage commonality learning in VST, we introduce a prompt discrimination loss (Section 3.5). Figure 3 shows the overall architecture of our proposed VSCode.", + "bbox": [ + 496, + 348, + 893, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Foundation Model", + "text_level": 1, + "bbox": [ + 498, + 443, + 679, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve a more comprehensive integration of commonities from SOD and COD tasks, we select VST [54] as our fundamental model. VST was originally proposed for RGB and RGB-D SOD and comprises three primary components, i.e. a transformer encoder, a transformer converter, and a multi-task transformer decoder. It initially employs the transformer encoder to capture long-range dependencies within the image features $\\pmb{f}_i^E \\in \\mathbb{R}^{l_i \\times c_i}$ , where $i \\in [0,1,2,3]$ indicates the index of blocks in the encoder, $l_i$ and $c_i$ mean the length of the patch sequence and the channel number of $\\pmb{f}_i^E$ . Subsequently, the transformer converter integrates the complement between RGB and depth features via cross-attention for RGB-D SOD or uses self-attention for RGB SOD. In the decoder, which is composed of a sequence of self-attention layers, VST predicts saliency maps and boundary maps simultaneously via a saliency token, a boundary token, and decoder features $\\pmb{f}_j^D \\in \\mathbb{R}^{l_j \\times d}$ , where $j$ corresponds the index of blocks in the decoder. Here $j \\in [2,1,0]$ for descending order and $d = 384$ . Due to the simple and pure-transformer-based architecture, VST can be easily used for other multimodal tasks and COD tasks without the need for model redesign. As a result, it emerges as a superior choice for constructing a generalist model for general multimodal SOD and COD.", + "bbox": [ + 496, + 462, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In pursuit of improved outcomes and a more suitable structure, we introduce modifications to VST. First, we select Swin transformer [59] as our backbone due to its efficiency and high performance. Second, to maintain a unified structure for both RGB tasks and other multimodal tasks,", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we utilize the RGB converter in VST, which comprises standard transformer layers. For multimodal tasks, we simply concatenate the supplementary modality's features with the RGB features along the channel dimension and employ a multilayer perceptron (MLP) to project them from $2d$ channels to $d$ channels. For RGB tasks, no alterations are made. Third, we incorporate certain extensions from $\\mathrm{VST}++$ [50], specifically including the token-supervised prediction loss.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Domain-specific Prompt", + "text_level": 1, + "bbox": [ + 76, + 219, + 302, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within the encoder, lower layers are dedicated to extracting low-level features, encompassing edges, colors, and textures, which exhibit distinct characteristics in various domains [100]. For instance, depth maps are typically rendered in grayscale, while thermal maps present a broader color spectrum. Higher layers, on the other hand, capture semantic information from modality features, which is crucial for all tasks. Consequently, we introduce domain-specific prompts $\\pmb{p}_i^d$ at each block $i$ in the encoder and design four kinds of domain-specific prompts for RGB, depth, thermal, and optical flow, respectively, to highlight the disparities among domains, as shown in Figure 3.", + "bbox": [ + 75, + 238, + 472, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the image features $\\pmb{f}_i^E$ from a specific block in the Swin transformer encoder, we use window-attention [59] and partition the feature $\\pmb{f}_i^E$ into window features $\\pmb{f}_{i - w}^{E}\\in \\mathbb{R}^{l_i / M^2\\times M^2\\times c_i}$ , where $M$ represents the window size and $l_{i} / M^{2}$ is the number of windows. Then, we replicate the prompts $\\pmb{p}_i^d\\in \\mathbb{R}^{N_i\\times c_i}$ for each window and obtain $\\pmb{p}_i^{d'}\\in \\mathbb{R}^{l_i / M^2\\times N_i\\times c_i}$ , where $N_{i}$ represents the number of learnable prompt tokens. Next, we append them to the patch feature tokens in each window and perform self-attention within each window, which can be defined as", + "bbox": [ + 75, + 420, + 472, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i + 1} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\text {S W / W - M S A} \\left(\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right]\\right)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 585, + 468, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where W-MSA and SW-MSA are multi-head self-attention modules with regular and shifted windowing configurations, respectively. Here we omit the residual connection [23], and layer normalization [2]. Next, we segment $\\pmb{p}_{i+1}^{d'}$ from each window and calculate the average of them to obtain $\\pmb{p}_{i+1}^d$ , and then reassemble the output window feature $\\pmb{f}_{i-w}^E$ to $\\pmb{f}_{i+1}^E$ for the next block.", + "bbox": [ + 75, + 631, + 472, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Task-specific Prompt", + "text_level": 1, + "bbox": [ + 76, + 744, + 276, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior research [39] has traditionally regarded SOD and COD as opposing tasks, emphasizing the disparities between the features extracted by the SOD encoder and the COD encoder as much as possible. However, we believe that SOD and COD share significant commonalities in their features, such as low-level cues, high-level objectness, and spatial structuredness. As a result, we introduce task-specific prompts to learn the peculiarities while retaining the primary stream parameters shared to capture commonalities. We add the", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6af400023f52890cbabb02097cd0d57b3ff39542b90839c1003789f2a2480639.jpg", + "image_caption": [ + "Figure 4. Illustration of the influence of using different task prompts." + ], + "image_footnote": [], + "bbox": [ + 503, + 85, + 893, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "task-specific prompts in both VST encoder and decoder, and the overall impact of adding these prompts is illustrated in Figure 4.", + "bbox": [ + 498, + 256, + 893, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder. Although the encoder primarily focuses on domain-specific features with domain prompts, semantic features still play a pivotal role in distinguishing SOD and COD tasks. Semantic features from the encoder typically emphasize the most relevant region for a particular task and allocate more attention accordingly. In the case of the SOD task, the foreground region receives greater attention, whereas for the COD task, the background usually gains large importance since objects are typically concealed within it. Hence, it is essential to incorporate task-specific prompts to encourage learning task-related features in the encoder. Otherwise, we risk initially activating the wrong objects before the decoding process. Following the pattern of domain-specific prompts, we introduce task-specific prompts $\\pmb{p}_i^{te} \\in \\mathbb{R}^{N_i \\times c_i}$ in each encoder block and use them in the same way as how domain-specific prompts are used.", + "bbox": [ + 496, + 301, + 893, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder. Camouflaged objects typically exhibit more intricate and detailed boundaries compared to salient objects. This complexity arises because concealed objects often share color or textual similarities with their surroundings, resulting in imperceptible boundaries. Therefore, solely introducing task-specific prompts in the encoder may not be adequate, as camouflaged objects require a more refined process within the decoder. We incorporate task-specific prompts in the decoder to allocate distinct attention for reconstructing both the boundary and object regions based on the features extracted by the encoder. In contrast, previous research [39] has not adequately explored the differences between these two tasks in the decoder, as they typically use a single decoder to handle both.", + "bbox": [ + 496, + 542, + 893, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding task-specific prompts in the decoder, we simply append learnable prompts $\\pmb{p}_{j+1}^{td} \\in \\mathbb{R}^{N \\times d}$ to the decoder feature tokens $\\pmb{f}_{j+1}^{D}$ from a specific block $j+1$ in the decoder. Then, we apply the self-attention as follows:", + "bbox": [ + 498, + 755, + 895, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j} ^ {t d} \\\\ \\boldsymbol {f} _ {j} ^ {D} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\operatorname {M S A} (\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j + 1} ^ {t d} \\\\ \\boldsymbol {f} _ {j + 1} ^ {D} \\end{array} \\right])) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 825, + 893, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MSA denotes the multi-head self-attention. Here we omit the saliency and boundary tokens in the VST decoder", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17172", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6888e4c12a8ae976cf966a2359425a17f932a17732c20915eeba0a3d07603d39.jpg", + "image_caption": [ + "Figure 5. Correlation of prompt pairs at each encoder block." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 472, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for conciseness. Please note that our task-specific prompts differ from saliency and boundary tokens since we do not introduce any supervision for them.", + "bbox": [ + 76, + 296, + 468, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prompts Layout and Discussion", + "text_level": 1, + "bbox": [ + 76, + 343, + 357, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To incorporate the aforementioned prompts within the encoder-decoder architecture, inspired by VPT [31], we offer two prompt inserting versions. In the deep version, new prompts are introduced at the start of each transformer block, whereas the shallow version involves proposing prompts at first and updating them across all blocks. To unveil the specific relationship among different domains and tasks at varying network depths, we employ the deep version for both domain and task-specific prompts within the encoder. Based on VST's design, which introduces a saliency and a boundary token at the beginning of the decoder, we use the shallow version for task-specific prompts in the decoder.", + "bbox": [ + 75, + 362, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculate the correlations of different domain and task prompt pairs at different blocks in Figure 5. It is evident that depth, thermal, and optical flow exhibit relatively strong correlations in low-level features, as all of them usually show obvious low-level contrast between target objects and backgrounds in terms of color or luminance. However, at higher levels, most domains exhibit lower correlations, highlighting the distinctions among them. Additionally, as for task-specific prompts, it is clear that SOD prompts and COD prompts exhibit more shared knowledge in the lower layers. As we progress to higher layers, the correlation decreases, indicating that high-level features gradually learn unrelated information. This observation urges us to implement the deep version of domain-specific prompts and task-specific prompts in the encoder in our final design, as different blocks acquire distinct knowledge. Moreover, the gradually decreased correlation values along with the increase of the network depth encourage us to use a progressively larger number of prompt tokens, as lower correlation means larger peculiarities and hence requires more parameters to learn.", + "bbox": [ + 75, + 544, + 472, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 849, + 223, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The design principle of our model is to use 2D prompts for encompassing peculiarities while integrating commonalities", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into the foundation model. However, this is not straightforward for freely learned prompts. As shown in Figure 5, they still suggest certain correlations. This indicates that the learned prompts are entangled, risking the model's capacity to differentiate among various domains and tasks and resulting in suboptimal optimization. Hence, we propose a prompt discrimination loss to minimize the correlation among the prompts of the same type, guaranteeing that each prompt acquires unique domain or task knowledge. Specifically, we aggregate prompts of the same domain/task into a single embedding and then perform discrimination. First, we average the input prompt tokens of each same prompt type at each block and use linear projections to align the channel numbers to $d$ . Subsequently, for each type of prompt, we concatenate the averaged prompts of different blocks, and use MLP to obtain the overall domain-specific prompt $p_{l}^{d_{all}}$ and task-specific encoder prompt $p_{k}^{te_{all}}$ :", + "bbox": [ + 496, + 90, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {l} ^ {d _ {a l l}} = \\operatorname {M L P} \\left[ \\mathrm {L A} \\left(\\boldsymbol {p} _ {0} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {1} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {2} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {3} ^ {d}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 357, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} ^ {t e _ {a l l}} = \\mathrm {M L P} [ \\mathrm {L A} (\\boldsymbol {p} _ {0} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {1} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {2} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {3} ^ {t e}) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 380, + 864, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{L}$ and $\\mathrm{A}$ represent the linear and average operation, respectively, with $l \\in \\{\\text{depth}, \\text{thermal}, \\text{flow}, \\text{rgb}\\}$ , and $k \\in \\{SOD, COD\\}$ . Since the task-specific prompts in the decoder are shallow, we simply average them.", + "bbox": [ + 500, + 406, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterward, we calculate the cosine similarity between prompt pairs, resulting in eight types of cosine similarity results $\\mathcal{CS}_m$ . Here $m$ means the combination of domains/tasks, namely $\\{RD, RT, RF, DF, DT, TF\\}$ for domain-aggregated prompts and $\\{SC_{EN}, SC_{DE}\\}$ for task-aggregated prompts in the encoder and decoder, respectively. Finally, we minimize the correlation within these prompt pairs and define our prompt discrimination loss as", + "bbox": [ + 496, + 465, + 893, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} = \\sum_ {m} \\ln (1 + | \\mathcal {C S} _ {m} |), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 595, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is further combined with the segmentation losses and boundary losses [54] to train our model.", + "bbox": [ + 496, + 636, + 890, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 672, + 624, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 693, + 785, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For RGB SOD, we evaluate our proposed model using six commonly used benchmark datasets, i.e. DUTS [81], ECSSD [97], HKU-IS [45], PASCAL-S [47], DUT-O [99], and SOD [62]. For RGB-D SOD, we use six large benchmark datasets, including STERE [64], NJUD [33], NLPR [69], DUTLF-Depth [71], SIP [17], and ReDWeb-S [53]. In terms of RGB-T SOD, we consider three public datasets: VT821 [79], VT1000 [78], and VT5000 [77]. For VSOD, we employ six widely used benchmark datasets: DAVIS [70], FBMS [65], ViSal [85], SegV2 [41], DAVSOD-Easy, and DAVSOD-Normal [18]. Regarding RGB COD, three extensive benchmark datasets are considered, including", + "bbox": [ + 496, + 715, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17173", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/453549e6085f2f85241e63222a7a89fe215e898a265beb9a299dda5628ea6e41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsParams (M)RGB SOD DUTS[81] Sm↑Fm↑Em↑RGB-D SOD NJUD[33] Sm↑Fm↑Em↑RGB-T SOD VT5000[77] Sm↑Fm↑Em↑VSOD SegV2[41] Sm↑Fm↑Em↑RGB COD CAMO[37] Sm↑Fm↑Em↑VCOD CAD[3] Sm↑Fm↑Em↑
.900.885.940.927.928.958.900.863.938.896.870.952.793.751.871.686.522.787
ST323.46*.898.884.941.924.922.954.903.886.942.930.911.972------
GT54.06.902.890.945.931.932.962.909.877.947.931.917.975------
GT+p^d54.06.902.890.945.931.931.961.906.892.946.925.910.970.804.776.876.759.639.808
GT+p^d+pt54.09.909.899.948.935.938.965.912.882.950.943.930.984.811.782.884.736.614.797
GT+p^d+pt+L_dis54.09.908.896.947.932.932.960.909.878.947.933.907.966.800.770.872.743.611.798
w/o p^te54.07.908.896.947.932.929.959.904.872.941.940.919.975.799.770.875.740.599.814
w/o p^td54.08.902.889.943.929.929.959.904.872.941.940.919.975.799.770.875.740.599.814
", + "bbox": [ + 81, + 75, + 883, + 204 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d025a612dd0f3e9162e090e9a2626cbc8b4dea50eb343fbd5a91cd8f5c6c67c.jpg", + "table_caption": [ + "Table 1. Ablation studies of our VSCode on the Swin-T [59] backbone with $224 \\times 224$ image size. We conduct evaluations on one representative dataset for each task. \"ST\" indicates special training, \"GT\" means general training, $p^d$ represents domain-specific prompts, and $p^t$ is task-specific prompts, which consists of $p^{te}$ in the encoder and $p^{td}$ in the decoder. $\\mathcal{L}_{dis}$ is our prompt discrimination loss. The best results under each setting are labeled in bold." + ], + "table_footnote": [], + "table_body": "
SettingsParams(M)RGB SOD DUTS[81] Sm↑ Fm↑ Em↑RGB-D SOD NJUD[33] Sm↑ Fm↑ Em↑RGB-T SOD VT5000[77] Sm↑ Fm↑ Em↑VSOD SegV2[41] Sm↑ Fm↑ Em↑RGB COD CAMO[37] Sm↑ Fm↑ Em↑VCOD CAD[3] Sm↑ Fm↑ Em↑
shallow or deep version for domain-specific prompts
shallow54.83.900.887.943.926.926.955.905.873.944.931.917.972------
deep54.06.902.890.945.931.932.962.909.877.947.931.917.975------
shallow or deep version for task-specific prompts in the encoder
shallow54.45.902.888.943.928.928.958.902.870.940.927.905.964.793.763.866.747.616.798
deep54.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
shallow or deep version for task-specific prompts in the decoder
deep54.10.903.891.945.930.932.960.905.888.943.922.903.966.802.774.881.738.605.801
shallow54.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
number of domain-specific prompts at four blocks
1,1,1,154.06.902.890.945.931.932.962.909.877.947.931.917.975------
5,5,5,554.08.903.890.945.931.935.961.902.869.940.918.887.954------
number of task-specific prompts in the encoder at four blocks
5,5,5,554.07.903.893.947.928.930.959.903.870.940.931.918.975.795.766.866.739.600.799
1,1,5,1054.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
number of task-specific prompts in the decoder
554.08.904.890.946.929.931.957.904.890.943.931.911.969.807.782.881.746.626.805
1054.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
1554.09.903.889.944.929.933.956.904.890.942.932.913.974.798.771.875.743.621.791
", + "bbox": [ + 81, + 272, + 883, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation studies of different designs of prompt layout.", + "bbox": [ + 287, + 547, + 679, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COD10K [15], CAMO [37], and NC4K [61]. For VCOD, we utilize two widely accepted benchmark datasets: CAD [3] and MoCA-Mask [10]. To ensure a consistent evaluation across all SOD and COD tasks, we employ three commonly used evaluation metrics to assess model performance: structure-measure $S_{m}$ [13], maximum enhanced-alignment measure $E_{m}$ [14], and maximum F-measure $F_{m}$ .", + "bbox": [ + 75, + 571, + 470, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 684, + 294, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on prior research [10, 22, 51, 54, 77], we employ the following datasets to train our model concurrently: the training set of DUTS for RGB SOD, the training sets of NJUD, NLPR, and DUTLF-Depth for RGB-D SOD, the training set of VT5000 for RGB-T SOD, the training sets of DAVIS and DAVSOD for VSOD, the training sets of COD10K and CAMO for RGB COD, and the training set of MoCA-Mask for VCOD. To ensure a fair comparison with previous works [22, 38, 56, 76, 116], we resize each", + "bbox": [ + 75, + 704, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image to $384 \\times 384$ pixels and then randomly crop them to $352 \\times 352$ image regions for training. Our training process employs the Adam optimizer [35] with an initial learning rate of 0.0001, which is reduced by a factor of 10 at half and three-quarters of the total training steps. We conduct a total of 150,000 training steps using a 3090 GPU. We mix the above six tasks in each training iteration with two samples for each task, leading to a total batch size of 12.", + "bbox": [ + 496, + 571, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 699, + 653, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture Design. To demonstrate the efficacy of various components in our VSCode model, we report the quantitative results in Table 1. We start by performing special training (ST) on each task individually and then conduct general training (GT) on all SOD tasks. Please note that here we do not consider COD tasks since no task prompt is used. We observe improved performance on RGB-T SOD and VSOD, demonstrating the significant benefit of shared knowledge in different tasks, especially for those with limited training data diversity. However, the results of RGB SOD and RGB-D SOD do not show a significant increase. Our hypothesis is that amalgamating the training of multi-", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "*The parameters for our specialized training methods amount to 53.61M for the RGB task and 54.06M for the multimodal task, resulting in a total of 323.46M parameters for all six tasks.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17174", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bbf2b19ecabc1f06ed69a6a79a750767db0e8c2185db22effb881c26c6948d96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams (M)DUTS[81]ECSSD[97]HKU-IS[45]PASCAL-S[47]DUT-O[99]SOD[62]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
VST[54]44.48.896.877.939.932.944.964.928.937.968.873.850.900.850.800.888.854.866.902
ICON-R[116]33.09.890.876.931.928.943.960.920.931.960.862.844.888.845.799.884.848.861.899
VST-T++ [50]53.60.901.887.943.937.949.968.930.939.968.878.855.901.853.804.892.853.866.899
MENet[89]27.83.905.895.943.927.938.956.927.939.965.871.848.892.850.792.879.841.847.884
VSCode-T54.09.917.910.954.945.957.971.935.946.970.878.852.900.869.830.910.863.879.908
EVP[56]64.52†.917.910.956.936.949.965.935.945.971.880.859.902.864.822.902.854.873.901
VSCode-S74.72†.926.922.960.949.959.974.940.951.974.887.864.904.877.840.912.870.882.910
", + "bbox": [ + 78, + 82, + 887, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eaebda4b6eb4a1cd848ed3da69fe5665815ae2ae8bcbdca8b7497c8b0017fe21.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of our VSCode with other 5 SOTA RGB SOD methods on six benchmark datasets. “-R”, “-T” and “-S” mean the ResNet50 [23], Swin-T, and Swin-S[59] backbones, respectively. “-” indicates the code is not available. The best performance under all settings is bolded, and the best results under each setting are labeled in bold." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)NJUD [33]NLPR[69]DUTLF-Depth[71]ReDWeb-S[53]STERE[64]SIP[17]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
CMINet[103]188.12.929.934.957.932.922.963.912.913.938.725.726.800.918.916.951.899.910.939
VST[54]53.83.922.920.951.932.920.962.943.948.969.759.763.826.913.907.951.904.915.944
VST-T++ [50]100.51.928.929.958.933.921.964.944.948.969.756.757.819.916.911.950.903.914.944
SPSN[38]----.923.912.960------.907.902.945.892.900.936
CAVER[68]55.79.920.924.953.929.921.964.931.939.962.730.724.802.914.911.951.893.906.934
VSCode-T54.09.941.945.967.938.930.966.952.959.974.766.771.831.928.926.957.917.936.955
VSCode-S74.72.944.949.970.941.932.968.960.967.980.777.776.829.931.928.958.924.942.958
", + "bbox": [ + 78, + 252, + 887, + 368 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a067c8d6d5169fa4f4b0555425b7052941a9d0e80f1ed94f3d2e2ee0a3944155.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of our VSCode with other 5 SOTA RGB-D SOD methods on six benchmark datasets." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)VT821[79]VT1000[78]VT5000[77]
SmFmEmSmFmEmSmFmEm
MIDD[76]52.43.871.847.916.916.904.956.868.834.919
TNet[11]87.41.899.885.936.929.921.965.895.864.936
CGMDR[7]-.894.872.932.931.927.966.896.877.939
VST-T++ [50]100.51.894.861.923.941.931.972.895.854.933
CAVER[68]55.79.891.874.933.936.927.970.892.857.935
VSCode-T54.09.921.906.951.949.944.981.918.892.954
VSCode-S74.72.926.910.954.952.947.981.925.900.959
", + "bbox": [ + 78, + 391, + 465, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of our VSCode with other 5 SOTA RGB-T SOD methods on three benchmark datasets.", + "bbox": [ + 76, + 508, + 468, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modal images within a shared model might prevent further optimization on those well-learned tasks. Based on this, we introduce domain-specific prompts $p^d$ , resulting in substantial improvements across all datasets, which demonstrates the efficacy of domain-specific prompts in consolidating peculiarities within their respective domains. Subsequently, we introduce task-specific prompts $p^t$ in the encoder-decoder architecture, enabling the capability to handle COD tasks. This brings slightly improved performance on some SOD tasks, however, significantly improves the performance on all COD tasks compared with the ST baseline, which probably owes to the well-learned commonalities from different tasks. Moreover, the incorporation of the prompt discrimination loss $\\mathcal{L}_{dis}$ leads to improved performance on most tasks, reaffirming its effectiveness in disentangling peculiarities.", + "bbox": [ + 75, + 546, + 468, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the effectiveness of the task-specific prompts in the encoder and decoder, we remove them individually, resulting in performance decrease. This indicates that using task prompts in both encoder and decoder is necessary. We also observe our 2D prompts only bring around 0.03M parameters, which makes our model much more parameter-efficient than the traditional special training scheme.", + "bbox": [ + 75, + 772, + 470, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt Location. Following VPT [31], we design other forms of prompt layout based on Section 3.4. Table 2 reveals that employing the shallow version of task-specific prompts in the decoder, the deep version of domain-specific prompts and task-specific prompts in the encoder yields the best results. One plausible rationale is that each block aggregates distinct-level features within the encoder, thus it is better to propose unique prompts for each block. In our decoder, we follow VST and used skip connection to fuse decoder features with encoder features, which have already utilized deep task prompts for distinction. Hence, using more task prompts in the decoder may not be essential, and the shallow version seems to be a more fitting choice.", + "bbox": [ + 500, + 398, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt Length. We perform experiments with varying lengths for three kinds of prompts. As shown in Table 2, for domain-specific prompts, using one prompt token at each block achieves better performance than using more tokens. This suggests that it's possible to effectively capture domain distinctions using only a small number of prompts, which matches the observed relatively large correlation within domain prompts in Figure 5. Regarding task-specific prompts within the encoder, a prompt layout of 1,1,5,10 tokens at four blocks is found to be optimal on COD tasks, highlighting the importance of high-level semantic features over low-level features in distinguishing between SOD and COD tasks. This observation matches Figure 5 as well in which the correlations of SC in deep blocks are smaller than those in shallow blocks. Regarding the number of task-specific prompts in the decoder, performance starts to decline when it exceeds", + "bbox": [ + 496, + 607, + 893, + 848 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "trast to EVP, which uses task-specific training. Therefore, comparing the parameters of our model with EVP may not be completely fair owing to the differences in training strategies and backbone utilization.", + "bbox": [ + 500, + 862, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Please note that our model shares parameters across six tasks, in con", + "bbox": [ + 94, + 886, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17175", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0d68f81de213051044524c6ffe6a79f5fd72328ad35ef1c535fa054c2029f634.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams (M)DAVIS [70]FBMS[65]ViSal[85]SegV2[41]DAVSOD-Easy[18]DAVSOD-Normal[18]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
DCFNet[105]71.66.914.899.970.883.853.910.952.953.990.903.870.953.729.612.781.708.601.791
FSNet[28]102.30.922.909.972.875.867.918---.849.773.920.760.637.796.732.623.789
CoSTFormer[51]-.923.906.978------.874.813.943.779.667.819.730.614.777
UFO[75]55.92.918.906.978.858.868.911.926.917.969.888.850.951.747.626.799.711.605.773
VSCode-T54.09.930.913.970.891.880.923.952.954.989.943.937.984.792.696.831.738.631.797
VSCode-S74.72.936.922.973.905.902.939.955.957.991.946.937.984.800.710.835.758.666.815
", + "bbox": [ + 78, + 82, + 887, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/10c1e0f9a220444b78e1b9ebb0c4143b52787202b8ac15ac54990ff0150a0c0d.jpg", + "table_caption": [ + "Table 6. Quantitative comparison of our VSCode with other 4 SOTA VSOD methods on six benchmark datasets." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)COD10K[15]NC4K[61]CAMO[37]
SmFmEmSmFmEmSmFmEm
MGL[101]63.60.814.738.890---.776.741.842
UJSC[39]121.63.817.750.902.856.835.920.803.775.867
SegMar[32]56.21.833.755.907.841.827.907.816.803.884
FEDER[22]44.13.822.768.905.847.833.915.802.789.873
VSCode-T54.09.847.795.925.874.853.930.838.821.909
EVP[56]64.52.845.794.926.874.855.933.849.833.918
VSCode-S74.72.869.827.942.891.878.944.873.861.938
", + "bbox": [ + 80, + 208, + 464, + 323 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9edac1cadb11f0759f1ce79df6a3ad62811ee07b14a5be26da15d522a7a3da1f.jpg", + "table_caption": [ + "Table 7. Quantitative comparison of our VSCode with other 5 SOTA RGB COD methods on three benchmark datasets." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)CAD [3]MoCA-Mask[10]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑
PNS-Net[26]26.87.671.473.787.514.068.599
RCRNet[96]53.79.664.405.786.559.170.593
MG[98]-.608.378.673.500.138.514
SLT-Net[10]164.68.715.542.823.624.327.768
VSCode-T54.09.757.659.808.650.339.787
VSCode-S74.72.790.680.853.665.386.796
", + "bbox": [ + 80, + 359, + 464, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "10. This emphasizes that blindly increasing the number of prompts doesn't guarantee improved performance.", + "bbox": [ + 76, + 503, + 468, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 537, + 444, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to space limitation, we only report the performance comparison of our methods against other most highly-performed state-of-the-art methods, including 4 specialist RGB SOD models [50, 54, 89, 116], 5 specialist RGB-D SOD models [38, 50, 54, 68, 103], 5 specialist RGB-T SOD models [7, 11, 50, 68, 76], 3 specialist VSOD models [28, 51, 105], 4 specialist RGB COD models [22, 32, 39, 101], and 4 specialist VCOD models [10, 26, 96, 98]. Two generalist models [56, 75] are also reported. To ensure a relatively fair comparison with EVP [56], which utilizes SegFormer-B4 [93] as their backbone (64.1M parameters), we switch our backbone to Swin-S [59] as it has a similar number of parameters (50M). As shown in Table 3, Table 4, Table 5, Table 6, Table 7, and Table 8, our VSCode significantly outperforms all specialist methods and two generalist models across all six tasks, underscoring the effectiveness of our specially designed 2D prompts and prompt discrimination loss. The supplementary material displays visual comparison results among the top-performing models.", + "bbox": [ + 76, + 556, + 468, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis of Generalization Ability", + "text_level": 1, + "bbox": [ + 76, + 849, + 372, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Previous generalist research [56, 75] primarily concentrated on assessing the effectiveness of models in training tasks, ne", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3dfc0b6a5ad6f098dff2033d829a44d8b9e3176b2b531c6f1a772b5142dacaeb.jpg", + "table_caption": [ + "Table 8. Quantitative comparison of our VSCode with other 4 SOTA VCOD methods on two benchmark datasets." + ], + "table_footnote": [], + "table_body": "
SummaryCOD10K[15]NC4K[61]CAMO[37]
MethodTaskSmFmEmSmFmEmSmFmEm
PopNet[92]RGB-D.851.802.919.861.843.919.808.792.874
VSCode-TZSRGB-D.882.849.945.902.894.950.885.885.945
VSCode-TRGB.847.795.925.874.853.930.838.821.909
", + "bbox": [ + 500, + 210, + 890, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Comparison with the SOTA RGB-D COD method on three benchmark datasets. \"ZS\" indicates zero-shot.", + "bbox": [ + 500, + 277, + 890, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "glecting their capacity for generalizing to novel tasks. Therefore, we employ the RGB-D COD task, which is not used in training, to further investigate the zero-shot generalization capabilities of our model. Specifically, we utilize our well-trained model and combine depth and COD prompts to tackle the RGB-D COD task. As shown in Table 9, our VSCode model significantly outperforms the state-of-the-art specialist model PopNet [92], although ours works in a pure zero-shot way. This demonstrates the superior zero-shot generalization ability of our proposed method. We also present the results of our model using only RGB information, which yields considerably lower performance compared to zero-shot RGB-D results. This validates that our zero-shot performance is not reliant on the utilization of seen RGB COD information but on the effectiveness of consolidating domain- and task-specific knowledge, which allows for the straightforward combination of various domain- and task-specific prompts for unseen tasks.", + "bbox": [ + 496, + 320, + 893, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 609, + 617, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present VSCode, a novel generalist and parameter-efficient model that tackles general multimodal SOD and COD tasks. Concretely, we use a foundation model to assimilate commonalities and 2D prompts to learn domain and task peculiarities. Furthermore, a prompt discrimination loss is introduced to help effectively disentangle specific knowledge and learn better shared knowledge. Our experiments demonstrate the effectiveness of VSCode on six training tasks and one unseen task.", + "bbox": [ + 496, + 630, + 893, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 782, + 656, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08; the National Natural Science Foundation of China under Grants 62136007, U20B2065, 62036005, 62322605; the Key Research and Development Program of Jiangsu Province under Grant BE2021093; the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center Project under Grant 21KT008; and by the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009.", + "bbox": [ + 496, + 803, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17176", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 3", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[3] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449. Springer, 2016. 2, 6, 8", + "[4] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV, pages 0–0, 2018. 2", + "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVMJ, 5:117-150, 2019. 1", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 3", + "[7] Gang Chen, Feng Shao, Xiongli Chai, Hangwei Chen, Qiuping Jiang, Xiangchao Meng, and Yo-Sung Ho. Cgmdrnet: Cross-guided modality difference reduction network for rgb-t salient object detection. IEEE TCSVT, 32(9):6308-6323, 2022. 7, 8", + "[8] Shuhan Chen and Yun Fu. Progressively guided alternate refinement network for rgb-d salient object detection. In ECCV, pages 520-538, 2020. 2", + "[9] Zuyao Chen, Runmin Cong, Qianqian Xu, and Qingming Huang. Dpanet: Depth potentiality-aware gated attention network for rgb-d salient object detection. IEEE TIP, 2020. 2", + "[10] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 6, 8", + "[11] Runmin Cong, Kepu Zhang, Chen Zhang, Feng Zheng, Yao Zhao, Qingming Huang, and Sam Kwong. Does thermal really always matter for rgb-t salient object detection? IEEE TMM, 2022. 7, 8", + "[12] Zijun Deng, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Jing Qin, Guoqiang Han, and Pheng-Ann Heng. R3net: Recurrent residual refinement network for saliency detection. In IJCAI, pages 684–690, 2018. 2", + "[13] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6", + "[14] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In IJCAI, pages 698-704, 2018. 6", + "[15] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 2, 6, 8", + "[16] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng," + ], + "bbox": [ + 86, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. VI, 1(1):16, 2023. 2", + "[17] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2020. 5, 7", + "[18] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In CVPR, pages 8554-8564, 2019. 2, 5, 8", + "[19] Deng-Ping Fan, Yingjie Zhai, Ali Borji, Jufeng Yang, and Ling Shao. Bbs-net: Rgb-d salient object detection with a bifurcated backbone strategy network. In ECCV, pages 275–292, 2020. 2", + "[20] Chaowei Fang, Haibin Tian, Dingwen Zhang, Qiang Zhang, Jungong Han, and Junwei Han. Densely nested top-down flows for salient object detection. SCIS, 65(8):182103, 2022. 2", + "[21] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In AAAI, volume 34, pages 10869–10876, 2020. 2", + "[22] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 6, 8", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4, 7", + "[24] Q Hou, MM Cheng, X Hu, A Borji, Z Tu, and PHS Torr. Deeply supervised salient object detection with short connections. IEEE TPAMI, 41(4):815-828, 2018. 2", + "[25] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. MIR, 20(2):249-262, 2023. 3", + "[26] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In MICCAI, pages 142–152. Springer, 2021. 8", + "[27] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 20(1):92-108, 2023. 1", + "[28] Ge-Peng Ji, Deng-Ping Fan, Keren Fu, Zhe Wu, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In CVMJ, pages 155-175, 2023. 2, 8", + "[29] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate rgb-d salient object detection via collaborative learning. In ECCV, pages 52-69, 2020. 2", + "[30] Yuzhu Ji, Hajunj Zhang, Zequn Jie, Lin Ma, and QM Jonathan Wu. Casnet: A cross-attention siamese network for video salient object detection. IEEE TNNLS, 32(6):2676–2690, 2020. 2", + "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 2, 3, 5, 7", + "[32] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-" + ], + "bbox": [ + 509, + 92, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17177", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4722, 2022. 2, 8", + "[33] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In ICIP, pages 1115-1119, 2014. 5, 6, 7", + "[34] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 2", + "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[36] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2", + "[37] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 6, 8", + "[38] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, pages 630-647. Springer, 2022. 6, 7, 8", + "[39] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In CVPR, pages 10071-10081, 2021. 4, 8", + "[40] Aixuan Li, Jing Zhang, Yunqiu Lv, Tong Zhang, Yiran Zhong, Mingyi He, and Yuchao Dai. Joint salient object detection and camouflaged object detection via uncertainty-aware learning. arXiv preprint arXiv:2307.04651, 2023. 2", + "[41] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 5, 6, 8", + "[42] Gongyang Li, Zhi Liu, and Haibin Ling. Icnet: Information conversion network for rgb-d based salient object detection. IEEE TIP, 29:4873-4884, 2020. 1, 2", + "[43] Gongyang Li, Zhi Liu, Linwei Ye, Yang Wang, and Haibin Ling. Cross-modal weighting network for rgb-d salient object detection. In ECCV, pages 665-681, 2020. 2", + "[44] Guanbin Li, Yuan Xie, Tianhao Wei, Keze Wang, and Liang Lin. Flow guided recurrent neural encoder for video salient object detection. In CVPR, pages 3243-3252, 2018. 1, 2", + "[45] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 5, 7", + "[46] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In CVPR, pages 15485-15494, 2023. 1", + "[47] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 7", + "[48] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, pages 678-686, 2016. 2", + "[49] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet:" + ], + "bbox": [ + 86, + 92, + 470, + 887 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 2", + "[50] Nian Liu, Ziyang Luo, Ni Zhang, and Junwei Han. Vst++: Efficient and stronger visual saliency transformer. arXiv preprint arXiv:2310.11725, 2023. 4, 7, 8", + "[51] Nian Liu, Kepan Nan, Wangbo Zhao, Xiwen Yao, and Junwei Han. Learning complementary spatial-temporal transformer for video salient object detection. IEEE TNNLS, 2023. 2, 6, 8", + "[52] Nian Liu, Ni Zhang, and Junwei Han. Learning selective self-mutual attention for rgb-d saliency detection. In CVPR, pages 13756-13765, 2020. 2", + "[53] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI, 44(12):9026-9042, 2021. 2, 5, 7", + "[54] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, pages 4722-4732, October 2021. 1, 2, 3, 5, 6, 7, 8", + "[55] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, pages 1871-1880, 2019. 2", + "[56] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for universal foreground segmentations. arXiv preprint arXiv:2305.18476, 2023. 3, 6, 7, 8", + "[57] Yi Liu, Dingwen Zhang, Nian Liu, Shoukun Xu, and Jungong Han. Disentangled capsule routing for fast part-object relational saliency. IEEE TIP, 31:6719-6732, 2022. 1", + "[58] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. TPAMI, 44(7):3688-3704, 2021. 1", + "[59] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 4, 6, 7, 8", + "[60] Zhengyi Liu, Song Shi, Quntao Duan, Wei Zhang, and Peng Zhao. Salient object detection for rgb-d image by single stream recurrent convolution neural network. *IJON*, 363:46–57, 2019. 2", + "[61] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 1, 2, 6, 8", + "[62] Vida Movahedi and James H Elder. Design and perceptual validation of performance measures for salient object segmentation. In CVPR, pages 49-56, 2010. 5, 7", + "[63] Xing Nie, Bolin Ni, Jianlong Chang, Gaomeng Meng, Chunlei Huo, Zhaoxiang Zhang, Shiming Xiang, Qi Tian, and Chunhong Pan. Pro-tuning: Unified prompt tuning for vision tasks. arXiv preprint arXiv:2207.14381, 2022. 2", + "[64] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In CVPR, pages 454-461, 2012. 5, 7", + "[65] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE TPAMI, 36(6):1187-1200, 2013. 5, 8", + "[66] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages" + ], + "bbox": [ + 509, + 92, + 893, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17178", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2160-2170, 2022. 1, 2", + "[67] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 2", + "[68] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Hutchuan Lu. Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE TIP, 32:892-904, 2023. 7, 8", + "[69] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In ECCV, pages 92-109, 2014. 5, 7", + "[70] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5, 8", + "[71] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 1, 2, 5, 7", + "[72] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 2", + "[73] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In ECCV, pages 212-228. Springer, 2020. 2", + "[74] Hongmei Song, Wenguan Wang, Sanyuan Zhao, Jianbing Shen, and Kin-Man Lam. Pyramid dilated deeper convlstm for video salient object detection. In ECCV, pages 715-731, 2018. 2", + "[75] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, Hanjing Su, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE TMM, 2023. 8", + "[76] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE TIP, 30:5678-5691, 2021. 2, 6, 7, 8", + "[77] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A large-scale dataset and benchmark. IEEE TMM, 2022. 1, 2, 5, 6, 7", + "[78] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE TMM, 22(1):160-173, 2019. 5, 7", + "[79] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. Rgb-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In IJIG, pages 359-369. Springer, 2018. 5, 7", + "[80] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2021. 1, 2", + "[81] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 5, 6, 7", + "[82] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang," + ], + "bbox": [ + 86, + 92, + 470, + 888 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Xiang Ruan. Salient object detection with recurrent fully convolutional networks. IEEE TPAMI, 41(7):1734-1746, 2018. 2", + "[83] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 1, 2", + "[84] Wenguan Wang, Jianbing Shen, Xingping Dong, and Ali Borji. Salient object detection driven by fixation prediction. In CVPR, pages 1711-1720, 2018. 2", + "[85] Wenguan Wang, Jianbing Shen, and Ling Shao. Consistent video saliency using local gradient flow optimization and global refinement. IEEE TIP, 24(11):4185-4196, 2015. 5, 8", + "[86] Wenguan Wang, Jianbing Shen, and Ling Shao. Video salient object detection via fully convolutional networks. IEEE TIP, 27(1):38-49, 2017. 1, 2", + "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, pages 6830-6839, 2023. 3", + "[88] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 3", + "[89] Yi Wang, Ruili Wang, Xin Fan, Tianzhu Wang, and Xi-angjian He. Pixels, regions, and objects: Multiple enhancement for salient object detection. In CVPR, pages 10031–10040, 2023. 7, 8", + "[90] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, pages 13025-13034, 2020. 2", + "[91] Lina Wei, Shanshan Zhao, Omar Farouk Bourahla, Xi Li, Fei Wu, Yueting Zhuang, Junwei Han, and Mingliang Xu. End-to-end video saliency detection via a deep contextual spatiotemporal network. IEEE TNNLS, 32(4):1691-1702, 2020. 1", + "[92] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In ICCV, pages 1032-1042, 2023. 1, 2, 8", + "[93] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 8", + "[94] Jin Xie, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Yanwei Pang, Ling Shao, and Mubarak Shah. Count-and similarity-aware r-cnn for pedestrian detection. In ECCV, pages 88-104. Springer, 2020. 1", + "[95] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, pages 15325-15336, 2023. 3", + "[96] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 8", + "[97] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 7", + "[98] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman," + ], + "bbox": [ + 509, + 92, + 893, + 888 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17179", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 8", + "[99] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173, 2013. 5, 7", + "[100] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 4", + "[101] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 1, 2, 8", + "[102] Dingwen Zhang, Junwei Han, Yu Zhang, and Dong Xu. Synthesizing supervision for learning deep saliency network without human annotation. IEEE TPAMI, 42(7):1755-1769, 2019. 2", + "[103] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. Rgb-d saliency detection via cascaded mutual information minimization. In ICCV, 2021. 7, 8", + "[104] Lu Zhang, Jianming Zhang, Zhe Lin, Huchuan Lu, and You He. Capsal: Leveraging captioning to boost semantics for salient object detection. In CVPR, pages 6024–6033, 2019. 2", + "[105] Miao Zhang, Jie Liu, Yifei Wang, Yongri Piao, Shunyu Yao, Wei Ji, Jingjing Li, Huchuan Lu, and Zhongxuan Luo. Dynamic context-sensitive filtering network for video salient object detection. In ICCV, pages 1553-1563, 2021. 8", + "[106] Miao Zhang, Weisong Ren, Yongri Piao, Zhengkun Rong, and Huchuan Lu. Select, supplement and focus for rgb-d saliency detection. In CVPR, pages 3472-3481, 2020. 2", + "[107] Qiang Zhang, Nianchang Huang, Lin Yao, Dingwen Zhang, Caifeng Shan, and Jungong Han. Rgb-t salient object detection via fusing multi-level cnn features. IEEE TIP, 29:3321-3335, 2019. 1, 2", + "[108] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 1, 2", + "[109] Jia-Xing Zhao, Yang Cao, Deng-Ping Fan, Ming-Ming Cheng, Xuan-Yi Li, and Le Zhang. Contrast prior and fluid pyramid integration for rgbd salient object detection. In CVPR, pages 3927-3936, 2019. 1, 2", + "[110] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet:edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 2", + "[111] Wangbo Zhao, Kepan Nan, Songyang Zhang, Kai Chen, Dahua Lin, and Yang You. Learning referring video object segmentation from weak annotation. arXiv preprint arXiv:2308.02162, 2023. 1", + "[112] Wangbo Zhao, Jiasheng Tang, Yizeng Han, Yibing Song, Kai Wang, Gao Huang, Fan Wang, and Yang You. Dynamic tuning towards parameter and inference efficiency for vit adaptation. arXiv preprint arXiv:2403.11808, 2024. 2", + "[113] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51, 2020. 2", + "[114] Dehua Zheng, Xiaochen Zheng, Laurence T Yang, Yuan" + ], + "bbox": [ + 80, + 90, + 470, + 888 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gao, Chenlu Zhu, and Yiheng Ruan. Mffn: Multi-view feature fusion network for camouflaged object detection. In WACV, pages 6232-6242, 2023. 2", + "[115] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, pages 9516-9526, 2023. 3", + "[116] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 1, 6, 7, 8", + "[117] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, pages 15116-15127, 2023. 3" + ], + "bbox": [ + 501, + 90, + 890, + 268 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17180", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_model.json b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_model.json new file mode 100644 index 0000000000000000000000000000000000000000..929b345f6acccf0e60ab64edb8b5e98564b87a15 --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_model.json @@ -0,0 +1,2930 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.13, + 0.828, + 0.177 + ], + "angle": 0, + "content": "VSCode: General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.203, + 0.826, + 0.239 + ], + "angle": 0, + "content": "Ziyang Luo\\(^{1}\\) Nian Liu\\(^{2,*}\\) Wangbo Zhao\\(^{3}\\) Xuguang Yang\\(^{1}\\) Dingwen Zhang\\(^{1}\\) Deng-Ping Fan\\(^{5,6}\\) Fahad Khan\\(^{2,4}\\) Junwei Han\\(^{1,7,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.24, + 0.916, + 0.311 + ], + "angle": 0, + "content": "\\(^{1}\\)Northwestern Polytechnical University \\(^{2}\\)Mohamed bin Zayed University of Artificial Intelligence \\(^{3}\\)National University of Singapore \\(^{4}\\)CVL, Linköping University \\(^{5}\\)TBI Center, CS, Nankai University \\(^{6}\\)Nankai International Advanced Research Institute (SHENZHEN FUTIAN) \\(^{7}\\)Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.633 + ], + "angle": 0, + "content": "Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD. Source code has been available at https://github.com/Sssssuperior/VSCode." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.66, + 0.21, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.472, + 0.866 + ], + "angle": 0, + "content": "Visual salient object detection (SOD) and camouflaged object detection (COD) are two interconnected yet unique tasks. The goal of SOD is to identify prominent objects within an image that significantly contrast with their surroundings [5], which can be used to promote segmentation [46, 111], detection [94], and Part-Object Relational visual saliency [57, 58]. While COD focuses on identifying objects concealed within their environment. These objects intentionally blend in by sharing structural or textural similarities with their surroundings [15]. Despite the seemingly different definitions of SOD and COD, they both belong to the realm of binary segmentation and share some vital fundamental similarities, such as" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.343, + 0.895, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.513, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Relationship of SOD, COD, and multimodal tasks. Each specific task is seen as a combination of two dimensions, i.e. domain (RGB/Depth/Thermal/Flow) and task (SOD/COD)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.567, + 0.702, + 0.581 + ], + "angle": 0, + "content": "objectness and structuredness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.794 + ], + "angle": 0, + "content": "To cater to various scenarios, both SOD and COD have given rise to several sub-tasks with different modalities, including RGB SOD [83, 108], RGB COD [27, 66, 101], RGB-D SOD [42, 71], RGB-D COD [92], and RGB-T SOD [80, 107]. By leveraging optical flow maps, Video SOD (VSOD) [44, 91] and VCOD [10, 36] tasks can also be seen as a combination of two modalities. The relationship of SOD, COD, and multimodal tasks is shown in Figure 1, where each specific task can be considered as a combination of two dimensions, i.e. domain and task. Although these multimodal tasks differ in the complementary cues they employ, these modalities share some key commonalities. For instance, depth, thermal, and optical flow maps often show obvious objectness as in RGB images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Although previous CNN-based [10, 61, 71, 77, 86, 109] and transformer-based [54, 116] approaches have effectively addressed these tasks and achieved favorable results, they usually rely on meticulously designed models to tackle each task individually. Designing models specifically for individual tasks can be problematic since the training data of one task is typically limited. Task-specific specialist models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Nian Liu (liunian228@gmail.com) and Junwei Han (junweihan2010@gmail.com)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.089, + 0.418, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.161, + 0.468, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.333 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of our VSCode model. We use VST [54] as the foundation model to acquire commonalities among multimodal SOD and COD tasks. For each task, we integrate 2D prompts to aggregate peculiarities along the domain dimension and the task dimension, including four domain-specific prompts and two task-specific prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.473, + 0.61 + ], + "angle": 0, + "content": "may be overly adapted to a particular task and overfitted to specific training data distribution, which ultimately sacrifices generalization ability and results in suboptimal performance. One solution may be using more data, however, being costly and time-consuming for data annotation. To this end, joint learning a generalist model emerges as a more promising option, as it allows for the maximum use of all data and the effective learning of the commonalities of all tasks, hence significantly reducing the risk of overfitting and enhancing the generalization capability [34, 55]. However, joint learning multiple tasks is not straightforward. On one hand, simultaneously handling both commonalities and peculiarities of all tasks poses a significant challenge as the incompatibility among different tasks easily leads to a decline in performance with simple joint training [40]. On the other hand, it usually introduces additional complexity, computational costs, and parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present a general Visual Salient and Camouflaged object detection (VSCode) model which encapsulates both commonalities and peculiarities of different tasks with a simple but effective design, as illustrated in Figure 2. On one hand, we adopt VST [54] as the shared segmentation foundation model to assimilate commonalities of different tasks by leveraging its simple and pure-transformer-based architecture. On the other hand, inspired by the recent emergence of the parameter-efficient prompting technique [31, 63, 112], we propose 2D prompts to capture task peculiarities. Specifically, we decompose these peculiarities along the domain dimension and the task dimension, and consequently design domain-specific prompts and task-specific prompts to comprehend the differences among diverse domains and tasks, respectively. These 2D prompts can effectively disentangle domain and task peculiarities, making our model easily adaptable by combining them to tackle specific tasks and even unseen ones. Furthermore, we present a prompt discrimination loss to encourage the 2D prompts" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.122 + ], + "angle": 0, + "content": "to focus on acquiring adequate peculiarities and enable the foundational model to concentrate on commonality learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.895, + 0.229 + ], + "angle": 0, + "content": "Finally, we train our VSCode model on four SOD tasks and two COD tasks, demonstrating its effectiveness against state-of-the-art methods. What's more, we carry out evaluations on a reserved task and reveal remarkable zero-shot generalization ability of our model, which has never been explored in previous works. The main contributions in this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.233, + 0.892, + 0.262 + ], + "angle": 0, + "content": "- We present VSCode, the first generalist model for multimodal SOD and COD tasks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.264, + 0.894, + 0.323 + ], + "angle": 0, + "content": "- We propose to use a foundation segmentation model to aggregate commonalities and introduce 2D prompts to learn peculiarities along the domain and task dimensions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.324, + 0.892, + 0.369 + ], + "angle": 0, + "content": "- A prompt discrimination loss is proposed to effectively enhance the learning of peculiarities and commonalities for 2D prompts and the foundation model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.37, + 0.894, + 0.429 + ], + "angle": 0, + "content": "- Our VSCode model surpasses all existing state-of-the-art models across all tasks on 26 datasets and showcases its ability to generalize to unseen tasks, further emphasizing the superiority of our approach." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.233, + 0.894, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.642, + 0.456 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.82, + 0.48 + ], + "angle": 0, + "content": "2.1. Deep Learning Based SOD and COD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.743 + ], + "angle": 0, + "content": "SOD. Previous RGB SOD works delved into attention-based [9, 19, 49, 71, 108], multi-level fusion-based [20, 24, 67, 83, 102, 113], recurrent-model-based [8, 12, 48, 60, 82], and multi-task-based methods [29, 72, 84, 90, 104, 106, 110]. In the case of RGB-D SOD, some models [9, 42, 43, 52, 53, 71, 109] leveraged various attention mechanisms to incorporate depth cues into RGB features. With regard to RGB-T SOD, recent studies also introduced attention-based methods [77, 80] and multi-level fusion [76, 107] to excavate the relationship between RGB and thermal features. Regarding the VSOD task, some works [18, 21, 30, 74, 86] mined spatial-temporal and appearance cues. More recently, there was a growing trend where various research [28, 44, 51, 73] endeavored to incorporate optical flow for combining motion cues with appearance details. Consistent with recent studies, we treat optical flow as a form of modality information and view VSOD as a multimodal SOD task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.901 + ], + "angle": 0, + "content": "COD. Currently, COD has RGB COD, RGB-D COD, and VCOD tasks. RGB COD methods can be broadly categorized as multi-task-based approaches [61, 101], multi-input-based approaches [66, 114], and refinement-based approaches [15, 32]. The RGB-D COD task was initially introduced in [92], where depth inference models are adapted for object segmentation. For VCOD, prior studies segmented the moving camouflaged objects via dense optical flow [3, 4] or well-designed models [10, 36]. For a more comprehensive literature review, please refer to [16]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Figure 3. Overall framework of our proposed VSCode model with 2D prompt learning. Based on the VST [54] foundation model, we insert the respective domain-specific prompts and task-specific prompts in the attention windows in the Swin transformer [59] encoder layers to learn domain and task-specific encoder features. The converter is used for multimodal feature fusion. Within the transformer decoder layers, task-specific prompts are appended to image feature tokens to perform task-specific decoding. We also provide detailed structures of an encoder layer \\((i = 0)\\) and a decoder layer \\((j = 0)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.349, + 0.329, + 0.365 + ], + "angle": 0, + "content": "2.2. Prompt in Computer Vision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Prompt was initially introduced in the field of NLP [6] and has been successfully integrated into computer vision tasks [25]. VPT [31] introduced a small amount of trainable parameters as prompts in the input space. ViPT [115] put forth the idea of modality-complementary prompts for task-oriented multi-modal tracking. Prior research has primarily focused on specific tasks, such as classification or tracking. In this paper, we propose to use 2D prompts for assembling different multimodal tasks and enable zero-shot generalization on unseen tasks, which has not been explored before." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.521, + 0.406, + 0.537 + ], + "angle": 0, + "content": "2.3. Generalist Segmentation Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Recently, several generalist frameworks have emerged for a range of segmentation tasks using a variety of prompts [1]. On one hand, X-Decoder [117] utilized generic nonsemantic queries and semantic queries to decode different pixel-level and token-level outputs. UNINEXT [95] introduced three types of prompts, namely category names, language expressions, and reference annotations. On the other hand, Painter [87] and SegGPT [88] leveraged image-mask pairs from the same task as prompts. Unlike the approaches mentioned above, which mainly concentrate on task differences, our VSCode dissects unique characteristics based on both domain and task dimensions, leading to a more versatile design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.472, + 0.812 + ], + "angle": 0, + "content": "In the field of SOD and COD, EVP [56] introduced adaptors into the encoder and trained each task individually for various foreground segmentation tasks. Different from them, we consider not only multiple tasks but also multiple modalities and we train all tasks simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.819, + 0.212, + 0.836 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose VSCode with the aim of jointly training SOD and COD tasks in an efficient and effective way. We allow VST [54] to incorporate commonalities (Section 3.1), and utilize 2D prompts, which comprise domain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.895, + 0.44 + ], + "angle": 0, + "content": "specific (Section 3.2) and task-specific prompts (Section 3.3), to encapsulate peculiarities. To accurately disentangle domain and task peculiarities in 2D prompts and encourage commonality learning in VST, we introduce a prompt discrimination loss (Section 3.5). Figure 3 shows the overall architecture of our proposed VSCode." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.444, + 0.68, + 0.458 + ], + "angle": 0, + "content": "3.1. Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.825 + ], + "angle": 0, + "content": "To achieve a more comprehensive integration of commonities from SOD and COD tasks, we select VST [54] as our fundamental model. VST was originally proposed for RGB and RGB-D SOD and comprises three primary components, i.e. a transformer encoder, a transformer converter, and a multi-task transformer decoder. It initially employs the transformer encoder to capture long-range dependencies within the image features \\( \\pmb{f}_i^E \\in \\mathbb{R}^{l_i \\times c_i} \\), where \\( i \\in [0,1,2,3] \\) indicates the index of blocks in the encoder, \\( l_i \\) and \\( c_i \\) mean the length of the patch sequence and the channel number of \\( \\pmb{f}_i^E \\). Subsequently, the transformer converter integrates the complement between RGB and depth features via cross-attention for RGB-D SOD or uses self-attention for RGB SOD. In the decoder, which is composed of a sequence of self-attention layers, VST predicts saliency maps and boundary maps simultaneously via a saliency token, a boundary token, and decoder features \\( \\pmb{f}_j^D \\in \\mathbb{R}^{l_j \\times d} \\), where \\( j \\) corresponds the index of blocks in the decoder. Here \\( j \\in [2,1,0] \\) for descending order and \\( d = 384 \\). Due to the simple and pure-transformer-based architecture, VST can be easily used for other multimodal tasks and COD tasks without the need for model redesign. As a result, it emerges as a superior choice for constructing a generalist model for general multimodal SOD and COD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.9 + ], + "angle": 0, + "content": "In pursuit of improved outcomes and a more suitable structure, we introduce modifications to VST. First, we select Swin transformer [59] as our backbone due to its efficiency and high performance. Second, to maintain a unified structure for both RGB tasks and other multimodal tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "we utilize the RGB converter in VST, which comprises standard transformer layers. For multimodal tasks, we simply concatenate the supplementary modality's features with the RGB features along the channel dimension and employ a multilayer perceptron (MLP) to project them from \\(2d\\) channels to \\(d\\) channels. For RGB tasks, no alterations are made. Third, we incorporate certain extensions from \\(\\mathrm{VST}++\\) [50], specifically including the token-supervised prediction loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.22, + 0.303, + 0.236 + ], + "angle": 0, + "content": "3.2. Domain-specific Prompt" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.473, + 0.42 + ], + "angle": 0, + "content": "Within the encoder, lower layers are dedicated to extracting low-level features, encompassing edges, colors, and textures, which exhibit distinct characteristics in various domains [100]. For instance, depth maps are typically rendered in grayscale, while thermal maps present a broader color spectrum. Higher layers, on the other hand, capture semantic information from modality features, which is crucial for all tasks. Consequently, we introduce domain-specific prompts \\( \\pmb{p}_i^d \\) at each block \\( i \\) in the encoder and design four kinds of domain-specific prompts for RGB, depth, thermal, and optical flow, respectively, to highlight the disparities among domains, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Given the image features \\( \\pmb{f}_i^E \\) from a specific block in the Swin transformer encoder, we use window-attention [59] and partition the feature \\( \\pmb{f}_i^E \\) into window features \\( \\pmb{f}_{i - w}^{E}\\in \\mathbb{R}^{l_i / M^2\\times M^2\\times c_i} \\), where \\( M \\) represents the window size and \\( l_{i} / M^{2} \\) is the number of windows. Then, we replicate the prompts \\( \\pmb{p}_i^d\\in \\mathbb{R}^{N_i\\times c_i} \\) for each window and obtain \\( \\pmb{p}_i^{d'}\\in \\mathbb{R}^{l_i / M^2\\times N_i\\times c_i} \\), where \\( N_{i} \\) represents the number of learnable prompt tokens. Next, we append them to the patch feature tokens in each window and perform self-attention within each window, which can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.587, + 0.47, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i + 1} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\text {S W / W - M S A} \\left(\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right]\\right)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.473, + 0.738 + ], + "angle": 0, + "content": "where W-MSA and SW-MSA are multi-head self-attention modules with regular and shifted windowing configurations, respectively. Here we omit the residual connection [23], and layer normalization [2]. Next, we segment \\( \\pmb{p}_{i+1}^{d'} \\) from each window and calculate the average of them to obtain \\( \\pmb{p}_{i+1}^d \\), and then reassemble the output window feature \\( \\pmb{f}_{i-w}^E \\) to \\( \\pmb{f}_{i+1}^E \\) for the next block." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.745, + 0.277, + 0.761 + ], + "angle": 0, + "content": "3.3. Task-specific Prompt" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Prior research [39] has traditionally regarded SOD and COD as opposing tasks, emphasizing the disparities between the features extracted by the SOD encoder and the COD encoder as much as possible. However, we believe that SOD and COD share significant commonalities in their features, such as low-level cues, high-level objectness, and spatial structuredness. As a result, we introduce task-specific prompts to learn the peculiarities while retaining the primary stream parameters shared to capture commonalities. We add the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.087, + 0.895, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.222, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Figure 4. Illustration of the influence of using different task prompts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.894, + 0.302 + ], + "angle": 0, + "content": "task-specific prompts in both VST encoder and decoder, and the overall impact of adding these prompts is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.544 + ], + "angle": 0, + "content": "Encoder. Although the encoder primarily focuses on domain-specific features with domain prompts, semantic features still play a pivotal role in distinguishing SOD and COD tasks. Semantic features from the encoder typically emphasize the most relevant region for a particular task and allocate more attention accordingly. In the case of the SOD task, the foreground region receives greater attention, whereas for the COD task, the background usually gains large importance since objects are typically concealed within it. Hence, it is essential to incorporate task-specific prompts to encourage learning task-related features in the encoder. Otherwise, we risk initially activating the wrong objects before the decoding process. Following the pattern of domain-specific prompts, we introduce task-specific prompts \\( \\pmb{p}_i^{te} \\in \\mathbb{R}^{N_i \\times c_i} \\) in each encoder block and use them in the same way as how domain-specific prompts are used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.895, + 0.753 + ], + "angle": 0, + "content": "Decoder. Camouflaged objects typically exhibit more intricate and detailed boundaries compared to salient objects. This complexity arises because concealed objects often share color or textual similarities with their surroundings, resulting in imperceptible boundaries. Therefore, solely introducing task-specific prompts in the encoder may not be adequate, as camouflaged objects require a more refined process within the decoder. We incorporate task-specific prompts in the decoder to allocate distinct attention for reconstructing both the boundary and object regions based on the features extracted by the encoder. In contrast, previous research [39] has not adequately explored the differences between these two tasks in the decoder, as they typically use a single decoder to handle both." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.756, + 0.897, + 0.819 + ], + "angle": 0, + "content": "Regarding task-specific prompts in the decoder, we simply append learnable prompts \\( \\pmb{p}_{j+1}^{td} \\in \\mathbb{R}^{N \\times d} \\) to the decoder feature tokens \\( \\pmb{f}_{j+1}^{D} \\) from a specific block \\( j+1 \\) in the decoder. Then, we apply the self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.827, + 0.895, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j} ^ {t d} \\\\ \\boldsymbol {f} _ {j} ^ {D} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\operatorname {M S A} (\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j + 1} ^ {t d} \\\\ \\boldsymbol {f} _ {j + 1} ^ {D} \\end{array} \\right])) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where MSA denotes the multi-head self-attention. Here we omit the saliency and boundary tokens in the VST decoder" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.277, + 0.465, + 0.292 + ], + "angle": 0, + "content": "Figure 5. Correlation of prompt pairs at each encoder block." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.47, + 0.342 + ], + "angle": 0, + "content": "for conciseness. Please note that our task-specific prompts differ from saliency and boundary tokens since we do not introduce any supervision for them." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.344, + 0.358, + 0.359 + ], + "angle": 0, + "content": "3.4. Prompts Layout and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.544 + ], + "angle": 0, + "content": "To incorporate the aforementioned prompts within the encoder-decoder architecture, inspired by VPT [31], we offer two prompt inserting versions. In the deep version, new prompts are introduced at the start of each transformer block, whereas the shallow version involves proposing prompts at first and updating them across all blocks. To unveil the specific relationship among different domains and tasks at varying network depths, we employ the deep version for both domain and task-specific prompts within the encoder. Based on VST's design, which introduces a saliency and a boundary token at the beginning of the decoder, we use the shallow version for task-specific prompts in the decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.473, + 0.847 + ], + "angle": 0, + "content": "We calculate the correlations of different domain and task prompt pairs at different blocks in Figure 5. It is evident that depth, thermal, and optical flow exhibit relatively strong correlations in low-level features, as all of them usually show obvious low-level contrast between target objects and backgrounds in terms of color or luminance. However, at higher levels, most domains exhibit lower correlations, highlighting the distinctions among them. Additionally, as for task-specific prompts, it is clear that SOD prompts and COD prompts exhibit more shared knowledge in the lower layers. As we progress to higher layers, the correlation decreases, indicating that high-level features gradually learn unrelated information. This observation urges us to implement the deep version of domain-specific prompts and task-specific prompts in the encoder in our final design, as different blocks acquire distinct knowledge. Moreover, the gradually decreased correlation values along with the increase of the network depth encourage us to use a progressively larger number of prompt tokens, as lower correlation means larger peculiarities and hence requires more parameters to learn." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.851, + 0.224, + 0.865 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The design principle of our model is to use 2D prompts for encompassing peculiarities while integrating commonalities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.35 + ], + "angle": 0, + "content": "into the foundation model. However, this is not straightforward for freely learned prompts. As shown in Figure 5, they still suggest certain correlations. This indicates that the learned prompts are entangled, risking the model's capacity to differentiate among various domains and tasks and resulting in suboptimal optimization. Hence, we propose a prompt discrimination loss to minimize the correlation among the prompts of the same type, guaranteeing that each prompt acquires unique domain or task knowledge. Specifically, we aggregate prompts of the same domain/task into a single embedding and then perform discrimination. First, we average the input prompt tokens of each same prompt type at each block and use linear projections to align the channel numbers to \\(d\\). Subsequently, for each type of prompt, we concatenate the averaged prompts of different blocks, and use MLP to obtain the overall domain-specific prompt \\(p_{l}^{d_{all}}\\) and task-specific encoder prompt \\(p_{k}^{te_{all}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.358, + 0.892, + 0.385 + ], + "angle": 0, + "content": "\\[\np _ {l} ^ {d _ {a l l}} = \\operatorname {M L P} \\left[ \\mathrm {L A} \\left(\\boldsymbol {p} _ {0} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {1} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {2} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {3} ^ {d}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.381, + 0.866, + 0.399 + ], + "angle": 0, + "content": "\\[\np _ {k} ^ {t e _ {a l l}} = \\mathrm {M L P} [ \\mathrm {L A} (\\boldsymbol {p} _ {0} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {1} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {2} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {3} ^ {t e}) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.407, + 0.892, + 0.467 + ], + "angle": 0, + "content": "where \\( \\mathrm{L} \\) and \\( \\mathrm{A} \\) represent the linear and average operation, respectively, with \\( l \\in \\{\\text{depth}, \\text{thermal}, \\text{flow}, \\text{rgb}\\} \\), and \\( k \\in \\{SOD, COD\\} \\). Since the task-specific prompts in the decoder are shallow, we simply average them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.589 + ], + "angle": 0, + "content": "Afterward, we calculate the cosine similarity between prompt pairs, resulting in eight types of cosine similarity results \\(\\mathcal{CS}_m\\). Here \\(m\\) means the combination of domains/tasks, namely \\(\\{RD, RT, RF, DF, DT, TF\\}\\) for domain-aggregated prompts and \\(\\{SC_{EN}, SC_{DE}\\}\\) for task-aggregated prompts in the encoder and decoder, respectively. Finally, we minimize the correlation within these prompt pairs and define our prompt discrimination loss as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.597, + 0.892, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} = \\sum_ {m} \\ln (1 + | \\mathcal {C S} _ {m} |), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.667 + ], + "angle": 0, + "content": "which is further combined with the segmentation losses and boundary losses [54] to train our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.625, + 0.689 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.787, + 0.709 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For RGB SOD, we evaluate our proposed model using six commonly used benchmark datasets, i.e. DUTS [81], ECSSD [97], HKU-IS [45], PASCAL-S [47], DUT-O [99], and SOD [62]. For RGB-D SOD, we use six large benchmark datasets, including STERE [64], NJUD [33], NLPR [69], DUTLF-Depth [71], SIP [17], and ReDWeb-S [53]. In terms of RGB-T SOD, we consider three public datasets: VT821 [79], VT1000 [78], and VT5000 [77]. For VSOD, we employ six widely used benchmark datasets: DAVIS [70], FBMS [65], ViSal [85], SegV2 [41], DAVSOD-Easy, and DAVSOD-Normal [18]. Regarding RGB COD, three extensive benchmark datasets are considered, including" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17173" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.076, + 0.885, + 0.205 + ], + "angle": 0, + "content": "
SettingsParams (M)RGB SOD DUTS[81] Sm↑Fm↑Em↑RGB-D SOD NJUD[33] Sm↑Fm↑Em↑RGB-T SOD VT5000[77] Sm↑Fm↑Em↑VSOD SegV2[41] Sm↑Fm↑Em↑RGB COD CAMO[37] Sm↑Fm↑Em↑VCOD CAD[3] Sm↑Fm↑Em↑
.900.885.940.927.928.958.900.863.938.896.870.952.793.751.871.686.522.787
ST323.46*.898.884.941.924.922.954.903.886.942.930.911.972------
GT54.06.902.890.945.931.932.962.909.877.947.931.917.975------
GT+p^d54.06.902.890.945.931.931.961.906.892.946.925.910.970.804.776.876.759.639.808
GT+p^d+pt54.09.909.899.948.935.938.965.912.882.950.943.930.984.811.782.884.736.614.797
GT+p^d+pt+L_dis54.09.908.896.947.932.932.960.909.878.947.933.907.966.800.770.872.743.611.798
w/o p^te54.07.908.896.947.932.929.959.904.872.941.940.919.975.799.770.875.740.599.814
w/o p^td54.08.902.889.943.929.929.959.904.872.941.940.919.975.799.770.875.740.599.814
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.207, + 0.894, + 0.263 + ], + "angle": 0, + "content": "Table 1. Ablation studies of our VSCode on the Swin-T [59] backbone with \\(224 \\times 224\\) image size. We conduct evaluations on one representative dataset for each task. \"ST\" indicates special training, \"GT\" means general training, \\(p^d\\) represents domain-specific prompts, and \\(p^t\\) is task-specific prompts, which consists of \\(p^{te}\\) in the encoder and \\(p^{td}\\) in the decoder. \\(\\mathcal{L}_{dis}\\) is our prompt discrimination loss. The best results under each setting are labeled in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.273, + 0.885, + 0.549 + ], + "angle": 0, + "content": "
SettingsParams(M)RGB SOD DUTS[81] Sm↑ Fm↑ Em↑RGB-D SOD NJUD[33] Sm↑ Fm↑ Em↑RGB-T SOD VT5000[77] Sm↑ Fm↑ Em↑VSOD SegV2[41] Sm↑ Fm↑ Em↑RGB COD CAMO[37] Sm↑ Fm↑ Em↑VCOD CAD[3] Sm↑ Fm↑ Em↑
shallow or deep version for domain-specific prompts
shallow54.83.900.887.943.926.926.955.905.873.944.931.917.972------
deep54.06.902.890.945.931.932.962.909.877.947.931.917.975------
shallow or deep version for task-specific prompts in the encoder
shallow54.45.902.888.943.928.928.958.902.870.940.927.905.964.793.763.866.747.616.798
deep54.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
shallow or deep version for task-specific prompts in the decoder
deep54.10.903.891.945.930.932.960.905.888.943.922.903.966.802.774.881.738.605.801
shallow54.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
number of domain-specific prompts at four blocks
1,1,1,154.06.902.890.945.931.932.962.909.877.947.931.917.975------
5,5,5,554.08.903.890.945.931.935.961.902.869.940.918.887.954------
number of task-specific prompts in the encoder at four blocks
5,5,5,554.07.903.893.947.928.930.959.903.870.940.931.918.975.795.766.866.739.600.799
1,1,5,1054.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
number of task-specific prompts in the decoder
554.08.904.890.946.929.931.957.904.890.943.931.911.969.807.782.881.746.626.805
1054.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
1554.09.903.889.944.929.933.956.904.890.942.932.913.974.798.771.875.743.621.791
" + }, + { + "type": "table_caption", + "bbox": [ + 0.288, + 0.549, + 0.681, + 0.561 + ], + "angle": 0, + "content": "Table 2. Ablation studies of different designs of prompt layout." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.472, + 0.679 + ], + "angle": 0, + "content": "COD10K [15], CAMO [37], and NC4K [61]. For VCOD, we utilize two widely accepted benchmark datasets: CAD [3] and MoCA-Mask [10]. To ensure a consistent evaluation across all SOD and COD tasks, we employ three commonly used evaluation metrics to assess model performance: structure-measure \\(S_{m}\\) [13], maximum enhanced-alignment measure \\(E_{m}\\) [14], and maximum F-measure \\(F_{m}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.295, + 0.7 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Building on prior research [10, 22, 51, 54, 77], we employ the following datasets to train our model concurrently: the training set of DUTS for RGB SOD, the training sets of NJUD, NLPR, and DUTLF-Depth for RGB-D SOD, the training set of VT5000 for RGB-T SOD, the training sets of DAVIS and DAVSOD for VSOD, the training sets of COD10K and CAMO for RGB COD, and the training set of MoCA-Mask for VCOD. To ensure a fair comparison with previous works [22, 38, 56, 76, 116], we resize each" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": "*The parameters for our specialized training methods amount to 53.61M for the RGB task and 54.06M for the multimodal task, resulting in a total of 323.46M parameters for all six tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.693 + ], + "angle": 0, + "content": "image to \\(384 \\times 384\\) pixels and then randomly crop them to \\(352 \\times 352\\) image regions for training. Our training process employs the Adam optimizer [35] with an initial learning rate of 0.0001, which is reduced by a factor of 10 at half and three-quarters of the total training steps. We conduct a total of 150,000 training steps using a 3090 GPU. We mix the above six tasks in each training iteration with two samples for each task, leading to a total batch size of 12." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.7, + 0.655, + 0.716 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Architecture Design. To demonstrate the efficacy of various components in our VSCode model, we report the quantitative results in Table 1. We start by performing special training (ST) on each task individually and then conduct general training (GT) on all SOD tasks. Please note that here we do not consider COD tasks since no task prompt is used. We observe improved performance on RGB-T SOD and VSOD, demonstrating the significant benefit of shared knowledge in different tasks, especially for those with limited training data diversity. However, the results of RGB SOD and RGB-D SOD do not show a significant increase. Our hypothesis is that amalgamating the training of multi-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17174" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.083, + 0.888, + 0.199 + ], + "angle": 0, + "content": "
MethodParams (M)DUTS[81]ECSSD[97]HKU-IS[45]PASCAL-S[47]DUT-O[99]SOD[62]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
VST[54]44.48.896.877.939.932.944.964.928.937.968.873.850.900.850.800.888.854.866.902
ICON-R[116]33.09.890.876.931.928.943.960.920.931.960.862.844.888.845.799.884.848.861.899
VST-T++ [50]53.60.901.887.943.937.949.968.930.939.968.878.855.901.853.804.892.853.866.899
MENet[89]27.83.905.895.943.927.938.956.927.939.965.871.848.892.850.792.879.841.847.884
VSCode-T54.09.917.910.954.945.957.971.935.946.970.878.852.900.869.830.910.863.879.908
EVP[56]64.52†.917.910.956.936.949.965.935.945.971.880.859.902.864.822.902.854.873.901
VSCode-S74.72†.926.922.960.949.959.974.940.951.974.887.864.904.877.840.912.870.882.910
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.2, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of our VSCode with other 5 SOTA RGB SOD methods on six benchmark datasets. “-R”, “-T” and “-S” mean the ResNet50 [23], Swin-T, and Swin-S[59] backbones, respectively. “-” indicates the code is not available. The best performance under all settings is bolded, and the best results under each setting are labeled in bold." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.253, + 0.888, + 0.369 + ], + "angle": 0, + "content": "
MethodParams (M)NJUD [33]NLPR[69]DUTLF-Depth[71]ReDWeb-S[53]STERE[64]SIP[17]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
CMINet[103]188.12.929.934.957.932.922.963.912.913.938.725.726.800.918.916.951.899.910.939
VST[54]53.83.922.920.951.932.920.962.943.948.969.759.763.826.913.907.951.904.915.944
VST-T++ [50]100.51.928.929.958.933.921.964.944.948.969.756.757.819.916.911.950.903.914.944
SPSN[38]----.923.912.960------.907.902.945.892.900.936
CAVER[68]55.79.920.924.953.929.921.964.931.939.962.730.724.802.914.911.951.893.906.934
VSCode-T54.09.941.945.967.938.930.966.952.959.974.766.771.831.928.926.957.917.936.955
VSCode-S74.72.944.949.970.941.932.968.960.967.980.777.776.829.931.928.958.924.942.958
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.37, + 0.856, + 0.383 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of our VSCode with other 5 SOTA RGB-D SOD methods on six benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.392, + 0.467, + 0.508 + ], + "angle": 0, + "content": "
MethodParams (M)VT821[79]VT1000[78]VT5000[77]
SmFmEmSmFmEmSmFmEm
MIDD[76]52.43.871.847.916.916.904.956.868.834.919
TNet[11]87.41.899.885.936.929.921.965.895.864.936
CGMDR[7]-.894.872.932.931.927.966.896.877.939
VST-T++ [50]100.51.894.861.923.941.931.972.895.854.933
CAVER[68]55.79.891.874.933.936.927.970.892.857.935
VSCode-T54.09.921.906.951.949.944.981.918.892.954
VSCode-S74.72.926.910.954.952.947.981.925.900.959
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.509, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of our VSCode with other 5 SOTA RGB-T SOD methods on three benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.773 + ], + "angle": 0, + "content": "modal images within a shared model might prevent further optimization on those well-learned tasks. Based on this, we introduce domain-specific prompts \\( p^d \\), resulting in substantial improvements across all datasets, which demonstrates the efficacy of domain-specific prompts in consolidating peculiarities within their respective domains. Subsequently, we introduce task-specific prompts \\( p^t \\) in the encoder-decoder architecture, enabling the capability to handle COD tasks. This brings slightly improved performance on some SOD tasks, however, significantly improves the performance on all COD tasks compared with the ST baseline, which probably owes to the well-learned commonalities from different tasks. Moreover, the incorporation of the prompt discrimination loss \\( \\mathcal{L}_{dis} \\) leads to improved performance on most tasks, reaffirming its effectiveness in disentangling peculiarities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.471, + 0.879 + ], + "angle": 0, + "content": "To further evaluate the effectiveness of the task-specific prompts in the encoder and decoder, we remove them individually, resulting in performance decrease. This indicates that using task prompts in both encoder and decoder is necessary. We also observe our 2D prompts only bring around 0.03M parameters, which makes our model much more parameter-efficient than the traditional special training scheme." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.4, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Prompt Location. Following VPT [31], we design other forms of prompt layout based on Section 3.4. Table 2 reveals that employing the shallow version of task-specific prompts in the decoder, the deep version of domain-specific prompts and task-specific prompts in the encoder yields the best results. One plausible rationale is that each block aggregates distinct-level features within the encoder, thus it is better to propose unique prompts for each block. In our decoder, we follow VST and used skip connection to fuse decoder features with encoder features, which have already utilized deep task prompts for distinction. Hence, using more task prompts in the decoder may not be essential, and the shallow version seems to be a more fitting choice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.894, + 0.849 + ], + "angle": 0, + "content": "Prompt Length. We perform experiments with varying lengths for three kinds of prompts. As shown in Table 2, for domain-specific prompts, using one prompt token at each block achieves better performance than using more tokens. This suggests that it's possible to effectively capture domain distinctions using only a small number of prompts, which matches the observed relatively large correlation within domain prompts in Figure 5. Regarding task-specific prompts within the encoder, a prompt layout of 1,1,5,10 tokens at four blocks is found to be optimal on COD tasks, highlighting the importance of high-level semantic features over low-level features in distinguishing between SOD and COD tasks. This observation matches Figure 5 as well in which the correlations of SC in deep blocks are smaller than those in shallow blocks. Regarding the number of task-specific prompts in the decoder, performance starts to decline when it exceeds" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "trast to EVP, which uses task-specific training. Therefore, comparing the parameters of our model with EVP may not be completely fair owing to the differences in training strategies and backbone utilization." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.471, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Please note that our model shares parameters across six tasks, in con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17175" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.083, + 0.888, + 0.187 + ], + "angle": 0, + "content": "
MethodParams (M)DAVIS [70]FBMS[65]ViSal[85]SegV2[41]DAVSOD-Easy[18]DAVSOD-Normal[18]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
DCFNet[105]71.66.914.899.970.883.853.910.952.953.990.903.870.953.729.612.781.708.601.791
FSNet[28]102.30.922.909.972.875.867.918---.849.773.920.760.637.796.732.623.789
CoSTFormer[51]-.923.906.978------.874.813.943.779.667.819.730.614.777
UFO[75]55.92.918.906.978.858.868.911.926.917.969.888.850.951.747.626.799.711.605.773
VSCode-T54.09.930.913.970.891.880.923.952.954.989.943.937.984.792.696.831.738.631.797
VSCode-S74.72.936.922.973.905.902.939.955.957.991.946.937.984.800.710.835.758.666.815
" + }, + { + "type": "table_caption", + "bbox": [ + 0.133, + 0.188, + 0.835, + 0.201 + ], + "angle": 0, + "content": "Table 6. Quantitative comparison of our VSCode with other 4 SOTA VSOD methods on six benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.209, + 0.465, + 0.324 + ], + "angle": 0, + "content": "
MethodParams (M)COD10K[15]NC4K[61]CAMO[37]
SmFmEmSmFmEmSmFmEm
MGL[101]63.60.814.738.890---.776.741.842
UJSC[39]121.63.817.750.902.856.835.920.803.775.867
SegMar[32]56.21.833.755.907.841.827.907.816.803.884
FEDER[22]44.13.822.768.905.847.833.915.802.789.873
VSCode-T54.09.847.795.925.874.853.930.838.821.909
EVP[56]64.52.845.794.926.874.855.933.849.833.918
VSCode-S74.72.869.827.942.891.878.944.873.861.938
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.326, + 0.468, + 0.352 + ], + "angle": 0, + "content": "Table 7. Quantitative comparison of our VSCode with other 5 SOTA RGB COD methods on three benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.361, + 0.465, + 0.465 + ], + "angle": 0, + "content": "
MethodParams (M)CAD [3]MoCA-Mask[10]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑
PNS-Net[26]26.87.671.473.787.514.068.599
RCRNet[96]53.79.664.405.786.559.170.593
MG[98]-.608.378.673.500.138.514
SLT-Net[10]164.68.715.542.823.624.327.768
VSCode-T54.09.757.659.808.650.339.787
VSCode-S74.72.790.680.853.665.386.796
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.467, + 0.469, + 0.493 + ], + "angle": 0, + "content": "Table 8. Quantitative comparison of our VSCode with other 4 SOTA VCOD methods on two benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.505, + 0.469, + 0.535 + ], + "angle": 0, + "content": "10. This emphasizes that blindly increasing the number of prompts doesn't guarantee improved performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.538, + 0.445, + 0.553 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.47, + 0.845 + ], + "angle": 0, + "content": "Due to space limitation, we only report the performance comparison of our methods against other most highly-performed state-of-the-art methods, including 4 specialist RGB SOD models [50, 54, 89, 116], 5 specialist RGB-D SOD models [38, 50, 54, 68, 103], 5 specialist RGB-T SOD models [7, 11, 50, 68, 76], 3 specialist VSOD models [28, 51, 105], 4 specialist RGB COD models [22, 32, 39, 101], and 4 specialist VCOD models [10, 26, 96, 98]. Two generalist models [56, 75] are also reported. To ensure a relatively fair comparison with EVP [56], which utilizes SegFormer-B4 [93] as their backbone (64.1M parameters), we switch our backbone to Swin-S [59] as it has a similar number of parameters (50M). As shown in Table 3, Table 4, Table 5, Table 6, Table 7, and Table 8, our VSCode significantly outperforms all specialist methods and two generalist models across all six tasks, underscoring the effectiveness of our specially designed 2D prompts and prompt discrimination loss. The supplementary material displays visual comparison results among the top-performing models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.851, + 0.373, + 0.867 + ], + "angle": 0, + "content": "4.5. Analysis of Generalization Ability" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Previous generalist research [56, 75] primarily concentrated on assessing the effectiveness of models in training tasks, ne" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.211, + 0.891, + 0.277 + ], + "angle": 0, + "content": "
SummaryCOD10K[15]NC4K[61]CAMO[37]
MethodTaskSmFmEmSmFmEmSmFmEm
PopNet[92]RGB-D.851.802.919.861.843.919.808.792.874
VSCode-TZSRGB-D.882.849.945.902.894.950.885.885.945
VSCode-TRGB.847.795.925.874.853.930.838.821.909
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.279, + 0.891, + 0.306 + ], + "angle": 0, + "content": "Table 9. Comparison with the SOTA RGB-D COD method on three benchmark datasets. \"ZS\" indicates zero-shot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.593 + ], + "angle": 0, + "content": "glecting their capacity for generalizing to novel tasks. Therefore, we employ the RGB-D COD task, which is not used in training, to further investigate the zero-shot generalization capabilities of our model. Specifically, we utilize our well-trained model and combine depth and COD prompts to tackle the RGB-D COD task. As shown in Table 9, our VSCode model significantly outperforms the state-of-the-art specialist model PopNet [92], although ours works in a pure zero-shot way. This demonstrates the superior zero-shot generalization ability of our proposed method. We also present the results of our model using only RGB information, which yields considerably lower performance compared to zero-shot RGB-D results. This validates that our zero-shot performance is not reliant on the utilization of seen RGB COD information but on the effectiveness of consolidating domain- and task-specific knowledge, which allows for the straightforward combination of various domain- and task-specific prompts for unseen tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.618, + 0.625 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.894, + 0.767 + ], + "angle": 0, + "content": "In this paper, we present VSCode, a novel generalist and parameter-efficient model that tackles general multimodal SOD and COD tasks. Concretely, we use a foundation model to assimilate commonalities and 2D prompts to learn domain and task peculiarities. Furthermore, a prompt discrimination loss is introduced to help effectively disentangle specific knowledge and learn better shared knowledge. Our experiments demonstrate the effectiveness of VSCode on six training tasks and one unseen task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.658, + 0.801 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.894, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08; the National Natural Science Foundation of China under Grants 62136007, U20B2065, 62036005, 62322605; the Key Research and Development Program of Jiangsu Province under Grant BE2021093; the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center Project under Grant 21KT008; and by the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17176" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.184, + 0.472, + 0.223 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[3] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449. Springer, 2016. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[4] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV, pages 0–0, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVMJ, 5:117-150, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.472, + 0.495 + ], + "angle": 0, + "content": "[7] Gang Chen, Feng Shao, Xiongli Chai, Hangwei Chen, Qiuping Jiang, Xiangchao Meng, and Yo-Sung Ho. Cgmdrnet: Cross-guided modality difference reduction network for rgb-t salient object detection. IEEE TCSVT, 32(9):6308-6323, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.495, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[8] Shuhan Chen and Yun Fu. Progressively guided alternate refinement network for rgb-d salient object detection. In ECCV, pages 520-538, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[9] Zuyao Chen, Runmin Cong, Qianqian Xu, and Qingming Huang. Dpanet: Depth potentiality-aware gated attention network for rgb-d salient object detection. IEEE TIP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[10] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.472, + 0.697 + ], + "angle": 0, + "content": "[11] Runmin Cong, Kepu Zhang, Chen Zhang, Feng Zheng, Yao Zhao, Qingming Huang, and Sam Kwong. Does thermal really always matter for rgb-t salient object detection? IEEE TMM, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.697, + 0.472, + 0.751 + ], + "angle": 0, + "content": "[12] Zijun Deng, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Jing Qin, Guoqiang Han, and Pheng-Ann Heng. R3net: Recurrent residual refinement network for saliency detection. In IJCAI, pages 684–690, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.752, + 0.472, + 0.792 + ], + "angle": 0, + "content": "[13] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.792, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[14] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In IJCAI, pages 698-704, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.472, + 0.886 + ], + "angle": 0, + "content": "[15] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.886, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[16] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng," + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. VI, 1(1):16, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.12, + 0.894, + 0.174 + ], + "angle": 0, + "content": "[17] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.174, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[18] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In CVPR, pages 8554-8564, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.214, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[19] Deng-Ping Fan, Yingjie Zhai, Ali Borji, Jufeng Yang, and Ling Shao. Bbs-net: Rgb-d salient object detection with a bifurcated backbone strategy network. In ECCV, pages 275–292, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.268, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[20] Chaowei Fang, Haibin Tian, Dingwen Zhang, Qiang Zhang, Jungong Han, and Junwei Han. Densely nested top-down flows for salient object detection. SCIS, 65(8):182103, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.322, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[21] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In AAAI, volume 34, pages 10869–10876, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[22] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.431, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.471, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[24] Q Hou, MM Cheng, X Hu, A Borji, Z Tu, and PHS Torr. Deeply supervised salient object detection with short connections. IEEE TPAMI, 41(4):815-828, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.512, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[25] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. MIR, 20(2):249-262, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.551, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[26] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In MICCAI, pages 142–152. Springer, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.606, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[27] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 20(1):92-108, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.659, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[28] Ge-Peng Ji, Deng-Ping Fan, Keren Fu, Zhe Wu, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In CVMJ, pages 155-175, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.7, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[29] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate rgb-d salient object detection via collaborative learning. In ECCV, pages 52-69, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.74, + 0.892, + 0.794 + ], + "angle": 0, + "content": "[30] Yuzhu Ji, Hajunj Zhang, Zequn Jie, Lin Ma, and QM Jonathan Wu. Casnet: A cross-attention siamese network for video salient object detection. IEEE TNNLS, 32(6):2676–2690, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.794, + 0.892, + 0.848 + ], + "angle": 0, + "content": "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.848, + 0.892, + 0.89 + ], + "angle": 0, + "content": "[32] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17177" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.093, + 0.22, + 0.105 + ], + "angle": 0, + "content": "4722, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.106, + 0.472, + 0.159 + ], + "angle": 0, + "content": "[33] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In ICIP, pages 1115-1119, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.16, + 0.472, + 0.201 + ], + "angle": 0, + "content": "[34] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.201, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.241, + 0.472, + 0.281 + ], + "angle": 0, + "content": "[36] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.281, + 0.472, + 0.333 + ], + "angle": 0, + "content": "[37] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.335, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[38] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, pages 630-647. Springer, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.39, + 0.472, + 0.442 + ], + "angle": 0, + "content": "[39] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In CVPR, pages 10071-10081, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.443, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[40] Aixuan Li, Jing Zhang, Yunqiu Lv, Tong Zhang, Yiran Zhong, Mingyi He, and Yuchao Dai. Joint salient object detection and camouflaged object detection via uncertainty-aware learning. arXiv preprint arXiv:2307.04651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.511, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[41] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.565, + 0.472, + 0.605 + ], + "angle": 0, + "content": "[42] Gongyang Li, Zhi Liu, and Haibin Ling. Icnet: Information conversion network for rgb-d based salient object detection. IEEE TIP, 29:4873-4884, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.606, + 0.472, + 0.646 + ], + "angle": 0, + "content": "[43] Gongyang Li, Zhi Liu, Linwei Ye, Yang Wang, and Haibin Ling. Cross-modal weighting network for rgb-d salient object detection. In ECCV, pages 665-681, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.646, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[44] Guanbin Li, Yuan Xie, Tianhao Wei, Keze Wang, and Liang Lin. Flow guided recurrent neural encoder for video salient object detection. In CVPR, pages 3243-3252, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.686, + 0.472, + 0.724 + ], + "angle": 0, + "content": "[45] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.726, + 0.472, + 0.793 + ], + "angle": 0, + "content": "[46] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In CVPR, pages 15485-15494, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.794, + 0.472, + 0.835 + ], + "angle": 0, + "content": "[47] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.835, + 0.472, + 0.875 + ], + "angle": 0, + "content": "[48] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, pages 678-686, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.875, + 0.472, + 0.888 + ], + "angle": 0, + "content": "[49] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet:" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.12, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[50] Nian Liu, Ziyang Luo, Ni Zhang, and Junwei Han. Vst++: Efficient and stronger visual saliency transformer. arXiv preprint arXiv:2310.11725, 2023. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.16, + 0.894, + 0.214 + ], + "angle": 0, + "content": "[51] Nian Liu, Kepan Nan, Wangbo Zhao, Xiwen Yao, and Junwei Han. Learning complementary spatial-temporal transformer for video salient object detection. IEEE TNNLS, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.215, + 0.894, + 0.255 + ], + "angle": 0, + "content": "[52] Nian Liu, Ni Zhang, and Junwei Han. Learning selective self-mutual attention for rgb-d saliency detection. In CVPR, pages 13756-13765, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.255, + 0.894, + 0.295 + ], + "angle": 0, + "content": "[53] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI, 44(12):9026-9042, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.295, + 0.894, + 0.335 + ], + "angle": 0, + "content": "[54] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, pages 4722-4732, October 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.335, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[55] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, pages 1871-1880, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.376, + 0.894, + 0.429 + ], + "angle": 0, + "content": "[56] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for universal foreground segmentations. arXiv preprint arXiv:2305.18476, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.43, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[57] Yi Liu, Dingwen Zhang, Nian Liu, Shoukun Xu, and Jungong Han. Disentangled capsule routing for fast part-object relational saliency. IEEE TIP, 31:6719-6732, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[58] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. TPAMI, 44(7):3688-3704, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.51, + 0.894, + 0.564 + ], + "angle": 0, + "content": "[59] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.565, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[60] Zhengyi Liu, Song Shi, Quntao Duan, Wei Zhang, and Peng Zhao. Salient object detection for rgb-d image by single stream recurrent convolution neural network. *IJON*, 363:46–57, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.618, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[61] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.672, + 0.894, + 0.713 + ], + "angle": 0, + "content": "[62] Vida Movahedi and James H Elder. Design and perceptual validation of performance measures for salient object segmentation. In CVPR, pages 49-56, 2010. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.713, + 0.894, + 0.767 + ], + "angle": 0, + "content": "[63] Xing Nie, Bolin Ni, Jianlong Chang, Gaomeng Meng, Chunlei Huo, Zhaoxiang Zhang, Shiming Xiang, Qi Tian, and Chunhong Pan. Pro-tuning: Unified prompt tuning for vision tasks. arXiv preprint arXiv:2207.14381, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.767, + 0.894, + 0.806 + ], + "angle": 0, + "content": "[64] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In CVPR, pages 454-461, 2012. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.806, + 0.894, + 0.847 + ], + "angle": 0, + "content": "[65] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE TPAMI, 36(6):1187-1200, 2013. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.847, + 0.894, + 0.888 + ], + "angle": 0, + "content": "[66] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17178" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.093, + 0.258, + 0.105 + ], + "angle": 0, + "content": "2160-2170, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.106, + 0.472, + 0.146 + ], + "angle": 0, + "content": "[67] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.147, + 0.471, + 0.187 + ], + "angle": 0, + "content": "[68] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Hutchuan Lu. Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE TIP, 32:892-904, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.187, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[69] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In ECCV, pages 92-109, 2014. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.227, + 0.471, + 0.281 + ], + "angle": 0, + "content": "[70] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.281, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[71] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.471, + 0.388 + ], + "angle": 0, + "content": "[72] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.389, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[73] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In ECCV, pages 212-228. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.443, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[74] Hongmei Song, Wenguan Wang, Sanyuan Zhao, Jianbing Shen, and Kin-Man Lam. Pyramid dilated deeper convlstm for video salient object detection. In ECCV, pages 715-731, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.497, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[75] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, Hanjing Su, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE TMM, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.565, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[76] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE TIP, 30:5678-5691, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.606, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[77] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A large-scale dataset and benchmark. IEEE TMM, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.659, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[78] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE TMM, 22(1):160-173, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.714, + 0.471, + 0.768 + ], + "angle": 0, + "content": "[79] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. Rgb-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In IJIG, pages 359-369. Springer, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.768, + 0.471, + 0.822 + ], + "angle": 0, + "content": "[80] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.822, + 0.471, + 0.876 + ], + "angle": 0, + "content": "[81] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.876, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[82] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang," + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "and Xiang Ruan. Salient object detection with recurrent fully convolutional networks. IEEE TPAMI, 41(7):1734-1746, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.134, + 0.894, + 0.187 + ], + "angle": 0, + "content": "[83] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.187, + 0.894, + 0.228 + ], + "angle": 0, + "content": "[84] Wenguan Wang, Jianbing Shen, Xingping Dong, and Ali Borji. Salient object detection driven by fixation prediction. In CVPR, pages 1711-1720, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.228, + 0.894, + 0.269 + ], + "angle": 0, + "content": "[85] Wenguan Wang, Jianbing Shen, and Ling Shao. Consistent video saliency using local gradient flow optimization and global refinement. IEEE TIP, 24(11):4185-4196, 2015. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.269, + 0.894, + 0.308 + ], + "angle": 0, + "content": "[86] Wenguan Wang, Jianbing Shen, and Ling Shao. Video salient object detection via fully convolutional networks. IEEE TIP, 27(1):38-49, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.309, + 0.894, + 0.361 + ], + "angle": 0, + "content": "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, pages 6830-6839, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.362, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[88] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.416, + 0.894, + 0.47 + ], + "angle": 0, + "content": "[89] Yi Wang, Ruili Wang, Xin Fan, Tianzhu Wang, and Xi-angjian He. Pixels, regions, and objects: Multiple enhancement for salient object detection. In CVPR, pages 10031–10040, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.47, + 0.894, + 0.512 + ], + "angle": 0, + "content": "[90] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, pages 13025-13034, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.512, + 0.894, + 0.579 + ], + "angle": 0, + "content": "[91] Lina Wei, Shanshan Zhao, Omar Farouk Bourahla, Xi Li, Fei Wu, Yueting Zhuang, Junwei Han, and Mingliang Xu. End-to-end video saliency detection via a deep contextual spatiotemporal network. IEEE TNNLS, 32(4):1691-1702, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.579, + 0.894, + 0.633 + ], + "angle": 0, + "content": "[92] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In ICCV, pages 1032-1042, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.633, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[93] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.687, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[94] Jin Xie, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Yanwei Pang, Ling Shao, and Mubarak Shah. Count-and similarity-aware r-cnn for pedestrian detection. In ECCV, pages 88-104. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.742, + 0.894, + 0.795 + ], + "angle": 0, + "content": "[95] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, pages 15325-15336, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.795, + 0.894, + 0.849 + ], + "angle": 0, + "content": "[96] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.849, + 0.894, + 0.876 + ], + "angle": 0, + "content": "[97] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.876, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[98] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman," + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17179" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.12, + 0.471, + 0.16 + ], + "angle": 0, + "content": "[99] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173, 2013. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.16, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[100] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.2, + 0.471, + 0.254 + ], + "angle": 0, + "content": "[101] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.254, + 0.471, + 0.307 + ], + "angle": 0, + "content": "[102] Dingwen Zhang, Junwei Han, Yu Zhang, and Dong Xu. Synthesizing supervision for learning deep saliency network without human annotation. IEEE TPAMI, 42(7):1755-1769, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.308, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[103] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. Rgb-d saliency detection via cascaded mutual information minimization. In ICCV, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[104] Lu Zhang, Jianming Zhang, Zhe Lin, Huchuan Lu, and You He. Capsal: Leveraging captioning to boost semantics for salient object detection. In CVPR, pages 6024–6033, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.416, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[105] Miao Zhang, Jie Liu, Yifei Wang, Yongri Piao, Shunyu Yao, Wei Ji, Jingjing Li, Huchuan Lu, and Zhongxuan Luo. Dynamic context-sensitive filtering network for video salient object detection. In ICCV, pages 1553-1563, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.471, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[106] Miao Zhang, Weisong Ren, Yongri Piao, Zhengkun Rong, and Huchuan Lu. Select, supplement and focus for rgb-d saliency detection. In CVPR, pages 3472-3481, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[107] Qiang Zhang, Nianchang Huang, Lin Yao, Dingwen Zhang, Caifeng Shan, and Jungong Han. Rgb-t salient object detection via fusing multi-level cnn features. IEEE TIP, 29:3321-3335, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.565, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[108] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.619, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[109] Jia-Xing Zhao, Yang Cao, Deng-Ping Fan, Ming-Ming Cheng, Xuan-Yi Li, and Le Zhang. Contrast prior and fluid pyramid integration for rgbd salient object detection. In CVPR, pages 3927-3936, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.471, + 0.727 + ], + "angle": 0, + "content": "[110] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet:edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.727, + 0.471, + 0.781 + ], + "angle": 0, + "content": "[111] Wangbo Zhao, Kepan Nan, Songyang Zhang, Kai Chen, Dahua Lin, and Yang You. Learning referring video object segmentation from weak annotation. arXiv preprint arXiv:2308.02162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.781, + 0.471, + 0.835 + ], + "angle": 0, + "content": "[112] Wangbo Zhao, Jiasheng Tang, Yizeng Han, Yibing Song, Kai Wang, Gao Huang, Fan Wang, and Yang You. Dynamic tuning towards parameter and inference efficiency for vit adaptation. arXiv preprint arXiv:2403.11808, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.835, + 0.471, + 0.876 + ], + "angle": 0, + "content": "[113] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.876, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[114] Dehua Zheng, Xiaochen Zheng, Laurence T Yang, Yuan" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.092, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Gao, Chenlu Zhu, and Yiheng Ruan. Mffn: Multi-view feature fusion network for camouflaged object detection. In WACV, pages 6232-6242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[115] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, pages 9516-9526, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[116] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.214, + 0.892, + 0.27 + ], + "angle": 0, + "content": "[117] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, pages 15116-15127, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17180" + } + ] +] \ No newline at end of file diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_origin.pdf b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8df992233221409121b43fde19a28bb36f5a062 --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/745b90c6-8ea5-4303-a0f1-3914d657b814_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683ed52566fca61f78dd81a0e5cd0c052c03ae6d1ceb76edf504497eb9deb9ae +size 2707773 diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/full.md b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f276e9a4ec5b264cffd6d2c288b7a20a738befe --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/full.md @@ -0,0 +1,359 @@ +# VSCode: General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning + +Ziyang Luo $^{1}$ Nian Liu $^{2,*}$ Wangbo Zhao $^{3}$ Xuguang Yang $^{1}$ Dingwen Zhang $^{1}$ Deng-Ping Fan $^{5,6}$ Fahad Khan $^{2,4}$ Junwei Han $^{1,7,*}$ + +$^{1}$ Northwestern Polytechnical University $^{2}$ Mohamed bin Zayed University of Artificial Intelligence $^{3}$ National University of Singapore $^{4}$ CVL, Linköping University $^{5}$ TBI Center, CS, Nankai University $^{6}$ Nankai International Advanced Research Institute (SHENZHEN FUTIAN) $^{7}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +# Abstract + +Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD. Source code has been available at https://github.com/Sssssuperior/VSCode. + +# 1. Introduction + +Visual salient object detection (SOD) and camouflaged object detection (COD) are two interconnected yet unique tasks. The goal of SOD is to identify prominent objects within an image that significantly contrast with their surroundings [5], which can be used to promote segmentation [46, 111], detection [94], and Part-Object Relational visual saliency [57, 58]. While COD focuses on identifying objects concealed within their environment. These objects intentionally blend in by sharing structural or textural similarities with their surroundings [15]. Despite the seemingly different definitions of SOD and COD, they both belong to the realm of binary segmentation and share some vital fundamental similarities, such as + +![](images/d9dcc601fd71069722e03084033f4d8f3f14eb25d1c0a16179f54b2847a67cec.jpg) +Figure 1. Relationship of SOD, COD, and multimodal tasks. Each specific task is seen as a combination of two dimensions, i.e. domain (RGB/Depth/Thermal/Flow) and task (SOD/COD). + +objectness and structuredness. + +To cater to various scenarios, both SOD and COD have given rise to several sub-tasks with different modalities, including RGB SOD [83, 108], RGB COD [27, 66, 101], RGB-D SOD [42, 71], RGB-D COD [92], and RGB-T SOD [80, 107]. By leveraging optical flow maps, Video SOD (VSOD) [44, 91] and VCOD [10, 36] tasks can also be seen as a combination of two modalities. The relationship of SOD, COD, and multimodal tasks is shown in Figure 1, where each specific task can be considered as a combination of two dimensions, i.e. domain and task. Although these multimodal tasks differ in the complementary cues they employ, these modalities share some key commonalities. For instance, depth, thermal, and optical flow maps often show obvious objectness as in RGB images. + +Although previous CNN-based [10, 61, 71, 77, 86, 109] and transformer-based [54, 116] approaches have effectively addressed these tasks and achieved favorable results, they usually rely on meticulously designed models to tackle each task individually. Designing models specifically for individual tasks can be problematic since the training data of one task is typically limited. Task-specific specialist models + +![](images/90588295a26c4c21545a4a8b1af0048b5eef90aa5c2d19338e6c1063d0e4abba.jpg) + +![](images/1787677a2f090fef175c757d803891dbd6a1c216ffcc84e119d4ff67a546af24.jpg) +Figure 2. Overall architecture of our VSCode model. We use VST [54] as the foundation model to acquire commonalities among multimodal SOD and COD tasks. For each task, we integrate 2D prompts to aggregate peculiarities along the domain dimension and the task dimension, including four domain-specific prompts and two task-specific prompts. + +may be overly adapted to a particular task and overfitted to specific training data distribution, which ultimately sacrifices generalization ability and results in suboptimal performance. One solution may be using more data, however, being costly and time-consuming for data annotation. To this end, joint learning a generalist model emerges as a more promising option, as it allows for the maximum use of all data and the effective learning of the commonalities of all tasks, hence significantly reducing the risk of overfitting and enhancing the generalization capability [34, 55]. However, joint learning multiple tasks is not straightforward. On one hand, simultaneously handling both commonalities and peculiarities of all tasks poses a significant challenge as the incompatibility among different tasks easily leads to a decline in performance with simple joint training [40]. On the other hand, it usually introduces additional complexity, computational costs, and parameters. + +In this paper, we present a general Visual Salient and Camouflaged object detection (VSCode) model which encapsulates both commonalities and peculiarities of different tasks with a simple but effective design, as illustrated in Figure 2. On one hand, we adopt VST [54] as the shared segmentation foundation model to assimilate commonalities of different tasks by leveraging its simple and pure-transformer-based architecture. On the other hand, inspired by the recent emergence of the parameter-efficient prompting technique [31, 63, 112], we propose 2D prompts to capture task peculiarities. Specifically, we decompose these peculiarities along the domain dimension and the task dimension, and consequently design domain-specific prompts and task-specific prompts to comprehend the differences among diverse domains and tasks, respectively. These 2D prompts can effectively disentangle domain and task peculiarities, making our model easily adaptable by combining them to tackle specific tasks and even unseen ones. Furthermore, we present a prompt discrimination loss to encourage the 2D prompts + +to focus on acquiring adequate peculiarities and enable the foundational model to concentrate on commonality learning. + +Finally, we train our VSCode model on four SOD tasks and two COD tasks, demonstrating its effectiveness against state-of-the-art methods. What's more, we carry out evaluations on a reserved task and reveal remarkable zero-shot generalization ability of our model, which has never been explored in previous works. The main contributions in this work can be summarized as follows: + +- We present VSCode, the first generalist model for multimodal SOD and COD tasks. +- We propose to use a foundation segmentation model to aggregate commonalities and introduce 2D prompts to learn peculiarities along the domain and task dimensions, respectively. +- A prompt discrimination loss is proposed to effectively enhance the learning of peculiarities and commonalities for 2D prompts and the foundation model, respectively. +- Our VSCode model surpasses all existing state-of-the-art models across all tasks on 26 datasets and showcases its ability to generalize to unseen tasks, further emphasizing the superiority of our approach. + +# 2. Related Work + +# 2.1. Deep Learning Based SOD and COD + +SOD. Previous RGB SOD works delved into attention-based [9, 19, 49, 71, 108], multi-level fusion-based [20, 24, 67, 83, 102, 113], recurrent-model-based [8, 12, 48, 60, 82], and multi-task-based methods [29, 72, 84, 90, 104, 106, 110]. In the case of RGB-D SOD, some models [9, 42, 43, 52, 53, 71, 109] leveraged various attention mechanisms to incorporate depth cues into RGB features. With regard to RGB-T SOD, recent studies also introduced attention-based methods [77, 80] and multi-level fusion [76, 107] to excavate the relationship between RGB and thermal features. Regarding the VSOD task, some works [18, 21, 30, 74, 86] mined spatial-temporal and appearance cues. More recently, there was a growing trend where various research [28, 44, 51, 73] endeavored to incorporate optical flow for combining motion cues with appearance details. Consistent with recent studies, we treat optical flow as a form of modality information and view VSOD as a multimodal SOD task. + +COD. Currently, COD has RGB COD, RGB-D COD, and VCOD tasks. RGB COD methods can be broadly categorized as multi-task-based approaches [61, 101], multi-input-based approaches [66, 114], and refinement-based approaches [15, 32]. The RGB-D COD task was initially introduced in [92], where depth inference models are adapted for object segmentation. For VCOD, prior studies segmented the moving camouflaged objects via dense optical flow [3, 4] or well-designed models [10, 36]. For a more comprehensive literature review, please refer to [16]. + +![](images/1bc501445f115cb0c0cd9229560adb5ba2902e8b4b1afffc5ded6241d82cfb87.jpg) +Figure 3. Overall framework of our proposed VSCode model with 2D prompt learning. Based on the VST [54] foundation model, we insert the respective domain-specific prompts and task-specific prompts in the attention windows in the Swin transformer [59] encoder layers to learn domain and task-specific encoder features. The converter is used for multimodal feature fusion. Within the transformer decoder layers, task-specific prompts are appended to image feature tokens to perform task-specific decoding. We also provide detailed structures of an encoder layer $(i = 0)$ and a decoder layer $(j = 0)$ . + +# 2.2. Prompt in Computer Vision + +Prompt was initially introduced in the field of NLP [6] and has been successfully integrated into computer vision tasks [25]. VPT [31] introduced a small amount of trainable parameters as prompts in the input space. ViPT [115] put forth the idea of modality-complementary prompts for task-oriented multi-modal tracking. Prior research has primarily focused on specific tasks, such as classification or tracking. In this paper, we propose to use 2D prompts for assembling different multimodal tasks and enable zero-shot generalization on unseen tasks, which has not been explored before. + +# 2.3. Generalist Segmentation Architecture + +Recently, several generalist frameworks have emerged for a range of segmentation tasks using a variety of prompts [1]. On one hand, X-Decoder [117] utilized generic nonsemantic queries and semantic queries to decode different pixel-level and token-level outputs. UNINEXT [95] introduced three types of prompts, namely category names, language expressions, and reference annotations. On the other hand, Painter [87] and SegGPT [88] leveraged image-mask pairs from the same task as prompts. Unlike the approaches mentioned above, which mainly concentrate on task differences, our VSCode dissects unique characteristics based on both domain and task dimensions, leading to a more versatile design. + +In the field of SOD and COD, EVP [56] introduced adaptors into the encoder and trained each task individually for various foreground segmentation tasks. Different from them, we consider not only multiple tasks but also multiple modalities and we train all tasks simultaneously. + +# 3. Methodology + +In this work, we propose VSCode with the aim of jointly training SOD and COD tasks in an efficient and effective way. We allow VST [54] to incorporate commonalities (Section 3.1), and utilize 2D prompts, which comprise domain + +specific (Section 3.2) and task-specific prompts (Section 3.3), to encapsulate peculiarities. To accurately disentangle domain and task peculiarities in 2D prompts and encourage commonality learning in VST, we introduce a prompt discrimination loss (Section 3.5). Figure 3 shows the overall architecture of our proposed VSCode. + +# 3.1. Foundation Model + +To achieve a more comprehensive integration of commonities from SOD and COD tasks, we select VST [54] as our fundamental model. VST was originally proposed for RGB and RGB-D SOD and comprises three primary components, i.e. a transformer encoder, a transformer converter, and a multi-task transformer decoder. It initially employs the transformer encoder to capture long-range dependencies within the image features $\pmb{f}_i^E \in \mathbb{R}^{l_i \times c_i}$ , where $i \in [0,1,2,3]$ indicates the index of blocks in the encoder, $l_i$ and $c_i$ mean the length of the patch sequence and the channel number of $\pmb{f}_i^E$ . Subsequently, the transformer converter integrates the complement between RGB and depth features via cross-attention for RGB-D SOD or uses self-attention for RGB SOD. In the decoder, which is composed of a sequence of self-attention layers, VST predicts saliency maps and boundary maps simultaneously via a saliency token, a boundary token, and decoder features $\pmb{f}_j^D \in \mathbb{R}^{l_j \times d}$ , where $j$ corresponds the index of blocks in the decoder. Here $j \in [2,1,0]$ for descending order and $d = 384$ . Due to the simple and pure-transformer-based architecture, VST can be easily used for other multimodal tasks and COD tasks without the need for model redesign. As a result, it emerges as a superior choice for constructing a generalist model for general multimodal SOD and COD. + +In pursuit of improved outcomes and a more suitable structure, we introduce modifications to VST. First, we select Swin transformer [59] as our backbone due to its efficiency and high performance. Second, to maintain a unified structure for both RGB tasks and other multimodal tasks, + +we utilize the RGB converter in VST, which comprises standard transformer layers. For multimodal tasks, we simply concatenate the supplementary modality's features with the RGB features along the channel dimension and employ a multilayer perceptron (MLP) to project them from $2d$ channels to $d$ channels. For RGB tasks, no alterations are made. Third, we incorporate certain extensions from $\mathrm{VST}++$ [50], specifically including the token-supervised prediction loss. + +# 3.2. Domain-specific Prompt + +Within the encoder, lower layers are dedicated to extracting low-level features, encompassing edges, colors, and textures, which exhibit distinct characteristics in various domains [100]. For instance, depth maps are typically rendered in grayscale, while thermal maps present a broader color spectrum. Higher layers, on the other hand, capture semantic information from modality features, which is crucial for all tasks. Consequently, we introduce domain-specific prompts $\pmb{p}_i^d$ at each block $i$ in the encoder and design four kinds of domain-specific prompts for RGB, depth, thermal, and optical flow, respectively, to highlight the disparities among domains, as shown in Figure 3. + +Given the image features $\pmb{f}_i^E$ from a specific block in the Swin transformer encoder, we use window-attention [59] and partition the feature $\pmb{f}_i^E$ into window features $\pmb{f}_{i - w}^{E}\in \mathbb{R}^{l_i / M^2\times M^2\times c_i}$ , where $M$ represents the window size and $l_{i} / M^{2}$ is the number of windows. Then, we replicate the prompts $\pmb{p}_i^d\in \mathbb{R}^{N_i\times c_i}$ for each window and obtain $\pmb{p}_i^{d'}\in \mathbb{R}^{l_i / M^2\times N_i\times c_i}$ , where $N_{i}$ represents the number of learnable prompt tokens. Next, we append them to the patch feature tokens in each window and perform self-attention within each window, which can be defined as + +$$ +\left[ \begin{array}{l} \boldsymbol {p} _ {i + 1} ^ {d ^ {\prime}} \\ \boldsymbol {f} _ {i - w} ^ {E} \end{array} \right] \leftarrow \operatorname {M L P} (\text {S W / W - M S A} \left(\left[ \begin{array}{l} \boldsymbol {p} _ {i} ^ {d ^ {\prime}} \\ \boldsymbol {f} _ {i - w} ^ {E} \end{array} \right]\right)), \tag {1} +$$ + +where W-MSA and SW-MSA are multi-head self-attention modules with regular and shifted windowing configurations, respectively. Here we omit the residual connection [23], and layer normalization [2]. Next, we segment $\pmb{p}_{i+1}^{d'}$ from each window and calculate the average of them to obtain $\pmb{p}_{i+1}^d$ , and then reassemble the output window feature $\pmb{f}_{i-w}^E$ to $\pmb{f}_{i+1}^E$ for the next block. + +# 3.3. Task-specific Prompt + +Prior research [39] has traditionally regarded SOD and COD as opposing tasks, emphasizing the disparities between the features extracted by the SOD encoder and the COD encoder as much as possible. However, we believe that SOD and COD share significant commonalities in their features, such as low-level cues, high-level objectness, and spatial structuredness. As a result, we introduce task-specific prompts to learn the peculiarities while retaining the primary stream parameters shared to capture commonalities. We add the + +![](images/6af400023f52890cbabb02097cd0d57b3ff39542b90839c1003789f2a2480639.jpg) +Figure 4. Illustration of the influence of using different task prompts. + +task-specific prompts in both VST encoder and decoder, and the overall impact of adding these prompts is illustrated in Figure 4. + +Encoder. Although the encoder primarily focuses on domain-specific features with domain prompts, semantic features still play a pivotal role in distinguishing SOD and COD tasks. Semantic features from the encoder typically emphasize the most relevant region for a particular task and allocate more attention accordingly. In the case of the SOD task, the foreground region receives greater attention, whereas for the COD task, the background usually gains large importance since objects are typically concealed within it. Hence, it is essential to incorporate task-specific prompts to encourage learning task-related features in the encoder. Otherwise, we risk initially activating the wrong objects before the decoding process. Following the pattern of domain-specific prompts, we introduce task-specific prompts $\pmb{p}_i^{te} \in \mathbb{R}^{N_i \times c_i}$ in each encoder block and use them in the same way as how domain-specific prompts are used. + +Decoder. Camouflaged objects typically exhibit more intricate and detailed boundaries compared to salient objects. This complexity arises because concealed objects often share color or textual similarities with their surroundings, resulting in imperceptible boundaries. Therefore, solely introducing task-specific prompts in the encoder may not be adequate, as camouflaged objects require a more refined process within the decoder. We incorporate task-specific prompts in the decoder to allocate distinct attention for reconstructing both the boundary and object regions based on the features extracted by the encoder. In contrast, previous research [39] has not adequately explored the differences between these two tasks in the decoder, as they typically use a single decoder to handle both. + +Regarding task-specific prompts in the decoder, we simply append learnable prompts $\pmb{p}_{j+1}^{td} \in \mathbb{R}^{N \times d}$ to the decoder feature tokens $\pmb{f}_{j+1}^{D}$ from a specific block $j+1$ in the decoder. Then, we apply the self-attention as follows: + +$$ +\left[ \begin{array}{l} \boldsymbol {p} _ {j} ^ {t d} \\ \boldsymbol {f} _ {j} ^ {D} \end{array} \right] \leftarrow \operatorname {M L P} (\operatorname {M S A} (\left[ \begin{array}{l} \boldsymbol {p} _ {j + 1} ^ {t d} \\ \boldsymbol {f} _ {j + 1} ^ {D} \end{array} \right])) \tag {2} +$$ + +where MSA denotes the multi-head self-attention. Here we omit the saliency and boundary tokens in the VST decoder + +![](images/6888e4c12a8ae976cf966a2359425a17f932a17732c20915eeba0a3d07603d39.jpg) +Figure 5. Correlation of prompt pairs at each encoder block. + +for conciseness. Please note that our task-specific prompts differ from saliency and boundary tokens since we do not introduce any supervision for them. + +# 3.4. Prompts Layout and Discussion + +To incorporate the aforementioned prompts within the encoder-decoder architecture, inspired by VPT [31], we offer two prompt inserting versions. In the deep version, new prompts are introduced at the start of each transformer block, whereas the shallow version involves proposing prompts at first and updating them across all blocks. To unveil the specific relationship among different domains and tasks at varying network depths, we employ the deep version for both domain and task-specific prompts within the encoder. Based on VST's design, which introduces a saliency and a boundary token at the beginning of the decoder, we use the shallow version for task-specific prompts in the decoder. + +We calculate the correlations of different domain and task prompt pairs at different blocks in Figure 5. It is evident that depth, thermal, and optical flow exhibit relatively strong correlations in low-level features, as all of them usually show obvious low-level contrast between target objects and backgrounds in terms of color or luminance. However, at higher levels, most domains exhibit lower correlations, highlighting the distinctions among them. Additionally, as for task-specific prompts, it is clear that SOD prompts and COD prompts exhibit more shared knowledge in the lower layers. As we progress to higher layers, the correlation decreases, indicating that high-level features gradually learn unrelated information. This observation urges us to implement the deep version of domain-specific prompts and task-specific prompts in the encoder in our final design, as different blocks acquire distinct knowledge. Moreover, the gradually decreased correlation values along with the increase of the network depth encourage us to use a progressively larger number of prompt tokens, as lower correlation means larger peculiarities and hence requires more parameters to learn. + +# 3.5. Loss Function + +The design principle of our model is to use 2D prompts for encompassing peculiarities while integrating commonalities + +into the foundation model. However, this is not straightforward for freely learned prompts. As shown in Figure 5, they still suggest certain correlations. This indicates that the learned prompts are entangled, risking the model's capacity to differentiate among various domains and tasks and resulting in suboptimal optimization. Hence, we propose a prompt discrimination loss to minimize the correlation among the prompts of the same type, guaranteeing that each prompt acquires unique domain or task knowledge. Specifically, we aggregate prompts of the same domain/task into a single embedding and then perform discrimination. First, we average the input prompt tokens of each same prompt type at each block and use linear projections to align the channel numbers to $d$ . Subsequently, for each type of prompt, we concatenate the averaged prompts of different blocks, and use MLP to obtain the overall domain-specific prompt $p_{l}^{d_{all}}$ and task-specific encoder prompt $p_{k}^{te_{all}}$ : + +$$ +p _ {l} ^ {d _ {a l l}} = \operatorname {M L P} \left[ \mathrm {L A} \left(\boldsymbol {p} _ {0} ^ {d}\right); \mathrm {L A} \left(\boldsymbol {p} _ {1} ^ {d}\right); \mathrm {L A} \left(\boldsymbol {p} _ {2} ^ {d}\right); \mathrm {L A} \left(\boldsymbol {p} _ {3} ^ {d}\right) \right], \tag {3} +$$ + +$$ +p _ {k} ^ {t e _ {a l l}} = \mathrm {M L P} [ \mathrm {L A} (\boldsymbol {p} _ {0} ^ {t e}); \mathrm {L A} (\boldsymbol {p} _ {1} ^ {t e}); \mathrm {L A} (\boldsymbol {p} _ {2} ^ {t e}); \mathrm {L A} (\boldsymbol {p} _ {3} ^ {t e}) ], +$$ + +where $\mathrm{L}$ and $\mathrm{A}$ represent the linear and average operation, respectively, with $l \in \{\text{depth}, \text{thermal}, \text{flow}, \text{rgb}\}$ , and $k \in \{SOD, COD\}$ . Since the task-specific prompts in the decoder are shallow, we simply average them. + +Afterward, we calculate the cosine similarity between prompt pairs, resulting in eight types of cosine similarity results $\mathcal{CS}_m$ . Here $m$ means the combination of domains/tasks, namely $\{RD, RT, RF, DF, DT, TF\}$ for domain-aggregated prompts and $\{SC_{EN}, SC_{DE}\}$ for task-aggregated prompts in the encoder and decoder, respectively. Finally, we minimize the correlation within these prompt pairs and define our prompt discrimination loss as + +$$ +\mathcal {L} _ {d i s} = \sum_ {m} \ln (1 + | \mathcal {C S} _ {m} |), \tag {4} +$$ + +which is further combined with the segmentation losses and boundary losses [54] to train our model. + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +For RGB SOD, we evaluate our proposed model using six commonly used benchmark datasets, i.e. DUTS [81], ECSSD [97], HKU-IS [45], PASCAL-S [47], DUT-O [99], and SOD [62]. For RGB-D SOD, we use six large benchmark datasets, including STERE [64], NJUD [33], NLPR [69], DUTLF-Depth [71], SIP [17], and ReDWeb-S [53]. In terms of RGB-T SOD, we consider three public datasets: VT821 [79], VT1000 [78], and VT5000 [77]. For VSOD, we employ six widely used benchmark datasets: DAVIS [70], FBMS [65], ViSal [85], SegV2 [41], DAVSOD-Easy, and DAVSOD-Normal [18]. Regarding RGB COD, three extensive benchmark datasets are considered, including + +
SettingsParams (M)RGB SOD DUTS[81] Sm↑Fm↑Em↑RGB-D SOD NJUD[33] Sm↑Fm↑Em↑RGB-T SOD VT5000[77] Sm↑Fm↑Em↑VSOD SegV2[41] Sm↑Fm↑Em↑RGB COD CAMO[37] Sm↑Fm↑Em↑VCOD CAD[3] Sm↑Fm↑Em↑
.900.885.940.927.928.958.900.863.938.896.870.952.793.751.871.686.522.787
ST323.46*.898.884.941.924.922.954.903.886.942.930.911.972------
GT54.06.902.890.945.931.932.962.909.877.947.931.917.975------
GT+p^d54.06.902.890.945.931.931.961.906.892.946.925.910.970.804.776.876.759.639.808
GT+p^d+pt54.09.909.899.948.935.938.965.912.882.950.943.930.984.811.782.884.736.614.797
GT+p^d+pt+L_dis54.09.908.896.947.932.932.960.909.878.947.933.907.966.800.770.872.743.611.798
w/o p^te54.07.908.896.947.932.929.959.904.872.941.940.919.975.799.770.875.740.599.814
w/o p^td54.08.902.889.943.929.929.959.904.872.941.940.919.975.799.770.875.740.599.814
+ +Table 1. Ablation studies of our VSCode on the Swin-T [59] backbone with $224 \times 224$ image size. We conduct evaluations on one representative dataset for each task. "ST" indicates special training, "GT" means general training, $p^d$ represents domain-specific prompts, and $p^t$ is task-specific prompts, which consists of $p^{te}$ in the encoder and $p^{td}$ in the decoder. $\mathcal{L}_{dis}$ is our prompt discrimination loss. The best results under each setting are labeled in bold. + +
SettingsParams(M)RGB SOD DUTS[81] Sm↑ Fm↑ Em↑RGB-D SOD NJUD[33] Sm↑ Fm↑ Em↑RGB-T SOD VT5000[77] Sm↑ Fm↑ Em↑VSOD SegV2[41] Sm↑ Fm↑ Em↑RGB COD CAMO[37] Sm↑ Fm↑ Em↑VCOD CAD[3] Sm↑ Fm↑ Em↑
shallow or deep version for domain-specific prompts
shallow54.83.900.887.943.926.926.955.905.873.944.931.917.972------
deep54.06.902.890.945.931.932.962.909.877.947.931.917.975------
shallow or deep version for task-specific prompts in the encoder
shallow54.45.902.888.943.928.928.958.902.870.940.927.905.964.793.763.866.747.616.798
deep54.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
shallow or deep version for task-specific prompts in the decoder
deep54.10.903.891.945.930.932.960.905.888.943.922.903.966.802.774.881.738.605.801
shallow54.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
number of domain-specific prompts at four blocks
1,1,1,154.06.902.890.945.931.932.962.909.877.947.931.917.975------
5,5,5,554.08.903.890.945.931.935.961.902.869.940.918.887.954------
number of task-specific prompts in the encoder at four blocks
5,5,5,554.07.903.893.947.928.930.959.903.870.940.931.918.975.795.766.866.739.600.799
1,1,5,1054.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
number of task-specific prompts in the decoder
554.08.904.890.946.929.931.957.904.890.943.931.911.969.807.782.881.746.626.805
1054.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
1554.09.903.889.944.929.933.956.904.890.942.932.913.974.798.771.875.743.621.791
+ +Table 2. Ablation studies of different designs of prompt layout. + +COD10K [15], CAMO [37], and NC4K [61]. For VCOD, we utilize two widely accepted benchmark datasets: CAD [3] and MoCA-Mask [10]. To ensure a consistent evaluation across all SOD and COD tasks, we employ three commonly used evaluation metrics to assess model performance: structure-measure $S_{m}$ [13], maximum enhanced-alignment measure $E_{m}$ [14], and maximum F-measure $F_{m}$ . + +# 4.2. Implementation Details + +Building on prior research [10, 22, 51, 54, 77], we employ the following datasets to train our model concurrently: the training set of DUTS for RGB SOD, the training sets of NJUD, NLPR, and DUTLF-Depth for RGB-D SOD, the training set of VT5000 for RGB-T SOD, the training sets of DAVIS and DAVSOD for VSOD, the training sets of COD10K and CAMO for RGB COD, and the training set of MoCA-Mask for VCOD. To ensure a fair comparison with previous works [22, 38, 56, 76, 116], we resize each + +image to $384 \times 384$ pixels and then randomly crop them to $352 \times 352$ image regions for training. Our training process employs the Adam optimizer [35] with an initial learning rate of 0.0001, which is reduced by a factor of 10 at half and three-quarters of the total training steps. We conduct a total of 150,000 training steps using a 3090 GPU. We mix the above six tasks in each training iteration with two samples for each task, leading to a total batch size of 12. + +# 4.3. Ablation Study + +Architecture Design. To demonstrate the efficacy of various components in our VSCode model, we report the quantitative results in Table 1. We start by performing special training (ST) on each task individually and then conduct general training (GT) on all SOD tasks. Please note that here we do not consider COD tasks since no task prompt is used. We observe improved performance on RGB-T SOD and VSOD, demonstrating the significant benefit of shared knowledge in different tasks, especially for those with limited training data diversity. However, the results of RGB SOD and RGB-D SOD do not show a significant increase. Our hypothesis is that amalgamating the training of multi- + +
MethodParams (M)DUTS[81]ECSSD[97]HKU-IS[45]PASCAL-S[47]DUT-O[99]SOD[62]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
VST[54]44.48.896.877.939.932.944.964.928.937.968.873.850.900.850.800.888.854.866.902
ICON-R[116]33.09.890.876.931.928.943.960.920.931.960.862.844.888.845.799.884.848.861.899
VST-T++ [50]53.60.901.887.943.937.949.968.930.939.968.878.855.901.853.804.892.853.866.899
MENet[89]27.83.905.895.943.927.938.956.927.939.965.871.848.892.850.792.879.841.847.884
VSCode-T54.09.917.910.954.945.957.971.935.946.970.878.852.900.869.830.910.863.879.908
EVP[56]64.52†.917.910.956.936.949.965.935.945.971.880.859.902.864.822.902.854.873.901
VSCode-S74.72†.926.922.960.949.959.974.940.951.974.887.864.904.877.840.912.870.882.910
+ +Table 3. Quantitative comparison of our VSCode with other 5 SOTA RGB SOD methods on six benchmark datasets. “-R”, “-T” and “-S” mean the ResNet50 [23], Swin-T, and Swin-S[59] backbones, respectively. “-” indicates the code is not available. The best performance under all settings is bolded, and the best results under each setting are labeled in bold. + +
MethodParams (M)NJUD [33]NLPR[69]DUTLF-Depth[71]ReDWeb-S[53]STERE[64]SIP[17]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
CMINet[103]188.12.929.934.957.932.922.963.912.913.938.725.726.800.918.916.951.899.910.939
VST[54]53.83.922.920.951.932.920.962.943.948.969.759.763.826.913.907.951.904.915.944
VST-T++ [50]100.51.928.929.958.933.921.964.944.948.969.756.757.819.916.911.950.903.914.944
SPSN[38]----.923.912.960------.907.902.945.892.900.936
CAVER[68]55.79.920.924.953.929.921.964.931.939.962.730.724.802.914.911.951.893.906.934
VSCode-T54.09.941.945.967.938.930.966.952.959.974.766.771.831.928.926.957.917.936.955
VSCode-S74.72.944.949.970.941.932.968.960.967.980.777.776.829.931.928.958.924.942.958
+ +Table 4. Quantitative comparison of our VSCode with other 5 SOTA RGB-D SOD methods on six benchmark datasets. + +
MethodParams (M)VT821[79]VT1000[78]VT5000[77]
SmFmEmSmFmEmSmFmEm
MIDD[76]52.43.871.847.916.916.904.956.868.834.919
TNet[11]87.41.899.885.936.929.921.965.895.864.936
CGMDR[7]-.894.872.932.931.927.966.896.877.939
VST-T++ [50]100.51.894.861.923.941.931.972.895.854.933
CAVER[68]55.79.891.874.933.936.927.970.892.857.935
VSCode-T54.09.921.906.951.949.944.981.918.892.954
VSCode-S74.72.926.910.954.952.947.981.925.900.959
+ +Table 5. Quantitative comparison of our VSCode with other 5 SOTA RGB-T SOD methods on three benchmark datasets. + +modal images within a shared model might prevent further optimization on those well-learned tasks. Based on this, we introduce domain-specific prompts $p^d$ , resulting in substantial improvements across all datasets, which demonstrates the efficacy of domain-specific prompts in consolidating peculiarities within their respective domains. Subsequently, we introduce task-specific prompts $p^t$ in the encoder-decoder architecture, enabling the capability to handle COD tasks. This brings slightly improved performance on some SOD tasks, however, significantly improves the performance on all COD tasks compared with the ST baseline, which probably owes to the well-learned commonalities from different tasks. Moreover, the incorporation of the prompt discrimination loss $\mathcal{L}_{dis}$ leads to improved performance on most tasks, reaffirming its effectiveness in disentangling peculiarities. + +To further evaluate the effectiveness of the task-specific prompts in the encoder and decoder, we remove them individually, resulting in performance decrease. This indicates that using task prompts in both encoder and decoder is necessary. We also observe our 2D prompts only bring around 0.03M parameters, which makes our model much more parameter-efficient than the traditional special training scheme. + +Prompt Location. Following VPT [31], we design other forms of prompt layout based on Section 3.4. Table 2 reveals that employing the shallow version of task-specific prompts in the decoder, the deep version of domain-specific prompts and task-specific prompts in the encoder yields the best results. One plausible rationale is that each block aggregates distinct-level features within the encoder, thus it is better to propose unique prompts for each block. In our decoder, we follow VST and used skip connection to fuse decoder features with encoder features, which have already utilized deep task prompts for distinction. Hence, using more task prompts in the decoder may not be essential, and the shallow version seems to be a more fitting choice. + +Prompt Length. We perform experiments with varying lengths for three kinds of prompts. As shown in Table 2, for domain-specific prompts, using one prompt token at each block achieves better performance than using more tokens. This suggests that it's possible to effectively capture domain distinctions using only a small number of prompts, which matches the observed relatively large correlation within domain prompts in Figure 5. Regarding task-specific prompts within the encoder, a prompt layout of 1,1,5,10 tokens at four blocks is found to be optimal on COD tasks, highlighting the importance of high-level semantic features over low-level features in distinguishing between SOD and COD tasks. This observation matches Figure 5 as well in which the correlations of SC in deep blocks are smaller than those in shallow blocks. Regarding the number of task-specific prompts in the decoder, performance starts to decline when it exceeds + +
MethodParams (M)DAVIS [70]FBMS[65]ViSal[85]SegV2[41]DAVSOD-Easy[18]DAVSOD-Normal[18]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
DCFNet[105]71.66.914.899.970.883.853.910.952.953.990.903.870.953.729.612.781.708.601.791
FSNet[28]102.30.922.909.972.875.867.918---.849.773.920.760.637.796.732.623.789
CoSTFormer[51]-.923.906.978------.874.813.943.779.667.819.730.614.777
UFO[75]55.92.918.906.978.858.868.911.926.917.969.888.850.951.747.626.799.711.605.773
VSCode-T54.09.930.913.970.891.880.923.952.954.989.943.937.984.792.696.831.738.631.797
VSCode-S74.72.936.922.973.905.902.939.955.957.991.946.937.984.800.710.835.758.666.815
+ +Table 6. Quantitative comparison of our VSCode with other 4 SOTA VSOD methods on six benchmark datasets. + +
MethodParams (M)COD10K[15]NC4K[61]CAMO[37]
SmFmEmSmFmEmSmFmEm
MGL[101]63.60.814.738.890---.776.741.842
UJSC[39]121.63.817.750.902.856.835.920.803.775.867
SegMar[32]56.21.833.755.907.841.827.907.816.803.884
FEDER[22]44.13.822.768.905.847.833.915.802.789.873
VSCode-T54.09.847.795.925.874.853.930.838.821.909
EVP[56]64.52.845.794.926.874.855.933.849.833.918
VSCode-S74.72.869.827.942.891.878.944.873.861.938
+ +Table 7. Quantitative comparison of our VSCode with other 5 SOTA RGB COD methods on three benchmark datasets. + +
MethodParams (M)CAD [3]MoCA-Mask[10]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑
PNS-Net[26]26.87.671.473.787.514.068.599
RCRNet[96]53.79.664.405.786.559.170.593
MG[98]-.608.378.673.500.138.514
SLT-Net[10]164.68.715.542.823.624.327.768
VSCode-T54.09.757.659.808.650.339.787
VSCode-S74.72.790.680.853.665.386.796
+ +10. This emphasizes that blindly increasing the number of prompts doesn't guarantee improved performance. + +# 4.4. Comparison with State-of-the-Art Methods + +Due to space limitation, we only report the performance comparison of our methods against other most highly-performed state-of-the-art methods, including 4 specialist RGB SOD models [50, 54, 89, 116], 5 specialist RGB-D SOD models [38, 50, 54, 68, 103], 5 specialist RGB-T SOD models [7, 11, 50, 68, 76], 3 specialist VSOD models [28, 51, 105], 4 specialist RGB COD models [22, 32, 39, 101], and 4 specialist VCOD models [10, 26, 96, 98]. Two generalist models [56, 75] are also reported. To ensure a relatively fair comparison with EVP [56], which utilizes SegFormer-B4 [93] as their backbone (64.1M parameters), we switch our backbone to Swin-S [59] as it has a similar number of parameters (50M). As shown in Table 3, Table 4, Table 5, Table 6, Table 7, and Table 8, our VSCode significantly outperforms all specialist methods and two generalist models across all six tasks, underscoring the effectiveness of our specially designed 2D prompts and prompt discrimination loss. The supplementary material displays visual comparison results among the top-performing models. + +# 4.5. Analysis of Generalization Ability + +Previous generalist research [56, 75] primarily concentrated on assessing the effectiveness of models in training tasks, ne + +Table 8. Quantitative comparison of our VSCode with other 4 SOTA VCOD methods on two benchmark datasets. + +
SummaryCOD10K[15]NC4K[61]CAMO[37]
MethodTaskSmFmEmSmFmEmSmFmEm
PopNet[92]RGB-D.851.802.919.861.843.919.808.792.874
VSCode-TZSRGB-D.882.849.945.902.894.950.885.885.945
VSCode-TRGB.847.795.925.874.853.930.838.821.909
+ +Table 9. Comparison with the SOTA RGB-D COD method on three benchmark datasets. "ZS" indicates zero-shot. + +glecting their capacity for generalizing to novel tasks. Therefore, we employ the RGB-D COD task, which is not used in training, to further investigate the zero-shot generalization capabilities of our model. Specifically, we utilize our well-trained model and combine depth and COD prompts to tackle the RGB-D COD task. As shown in Table 9, our VSCode model significantly outperforms the state-of-the-art specialist model PopNet [92], although ours works in a pure zero-shot way. This demonstrates the superior zero-shot generalization ability of our proposed method. We also present the results of our model using only RGB information, which yields considerably lower performance compared to zero-shot RGB-D results. This validates that our zero-shot performance is not reliant on the utilization of seen RGB COD information but on the effectiveness of consolidating domain- and task-specific knowledge, which allows for the straightforward combination of various domain- and task-specific prompts for unseen tasks. + +# 5. Conclusion + +In this paper, we present VSCode, a novel generalist and parameter-efficient model that tackles general multimodal SOD and COD tasks. Concretely, we use a foundation model to assimilate commonalities and 2D prompts to learn domain and task peculiarities. Furthermore, a prompt discrimination loss is introduced to help effectively disentangle specific knowledge and learn better shared knowledge. Our experiments demonstrate the effectiveness of VSCode on six training tasks and one unseen task. + +# Acknowledgments + +This work was supported in part by the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08; the National Natural Science Foundation of China under Grants 62136007, U20B2065, 62036005, 62322605; the Key Research and Development Program of Jiangsu Province under Grant BE2021093; the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center Project under Grant 21KT008; and by the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009. + +# References + +[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 3 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[3] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449. Springer, 2016. 2, 6, 8 +[4] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV, pages 0–0, 2018. 2 +[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVMJ, 5:117-150, 2019. 1 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 3 +[7] Gang Chen, Feng Shao, Xiongli Chai, Hangwei Chen, Qiuping Jiang, Xiangchao Meng, and Yo-Sung Ho. Cgmdrnet: Cross-guided modality difference reduction network for rgb-t salient object detection. IEEE TCSVT, 32(9):6308-6323, 2022. 7, 8 +[8] Shuhan Chen and Yun Fu. Progressively guided alternate refinement network for rgb-d salient object detection. In ECCV, pages 520-538, 2020. 2 +[9] Zuyao Chen, Runmin Cong, Qianqian Xu, and Qingming Huang. Dpanet: Depth potentiality-aware gated attention network for rgb-d salient object detection. IEEE TIP, 2020. 2 +[10] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 6, 8 +[11] Runmin Cong, Kepu Zhang, Chen Zhang, Feng Zheng, Yao Zhao, Qingming Huang, and Sam Kwong. Does thermal really always matter for rgb-t salient object detection? IEEE TMM, 2022. 7, 8 +[12] Zijun Deng, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Jing Qin, Guoqiang Han, and Pheng-Ann Heng. R3net: Recurrent residual refinement network for saliency detection. In IJCAI, pages 684–690, 2018. 2 +[13] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6 +[14] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In IJCAI, pages 698-704, 2018. 6 +[15] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 2, 6, 8 +[16] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, + +Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. VI, 1(1):16, 2023. 2 +[17] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2020. 5, 7 +[18] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In CVPR, pages 8554-8564, 2019. 2, 5, 8 +[19] Deng-Ping Fan, Yingjie Zhai, Ali Borji, Jufeng Yang, and Ling Shao. Bbs-net: Rgb-d salient object detection with a bifurcated backbone strategy network. In ECCV, pages 275–292, 2020. 2 +[20] Chaowei Fang, Haibin Tian, Dingwen Zhang, Qiang Zhang, Jungong Han, and Junwei Han. Densely nested top-down flows for salient object detection. SCIS, 65(8):182103, 2022. 2 +[21] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In AAAI, volume 34, pages 10869–10876, 2020. 2 +[22] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 6, 8 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4, 7 +[24] Q Hou, MM Cheng, X Hu, A Borji, Z Tu, and PHS Torr. Deeply supervised salient object detection with short connections. IEEE TPAMI, 41(4):815-828, 2018. 2 +[25] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. MIR, 20(2):249-262, 2023. 3 +[26] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In MICCAI, pages 142–152. Springer, 2021. 8 +[27] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 20(1):92-108, 2023. 1 +[28] Ge-Peng Ji, Deng-Ping Fan, Keren Fu, Zhe Wu, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In CVMJ, pages 155-175, 2023. 2, 8 +[29] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate rgb-d salient object detection via collaborative learning. In ECCV, pages 52-69, 2020. 2 +[30] Yuzhu Ji, Hajunj Zhang, Zequn Jie, Lin Ma, and QM Jonathan Wu. Casnet: A cross-attention siamese network for video salient object detection. IEEE TNNLS, 32(6):2676–2690, 2020. 2 +[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 2, 3, 5, 7 +[32] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713- + +4722, 2022. 2, 8 +[33] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In ICIP, pages 1115-1119, 2014. 5, 6, 7 +[34] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 2 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[36] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2 +[37] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 6, 8 +[38] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, pages 630-647. Springer, 2022. 6, 7, 8 +[39] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In CVPR, pages 10071-10081, 2021. 4, 8 +[40] Aixuan Li, Jing Zhang, Yunqiu Lv, Tong Zhang, Yiran Zhong, Mingyi He, and Yuchao Dai. Joint salient object detection and camouflaged object detection via uncertainty-aware learning. arXiv preprint arXiv:2307.04651, 2023. 2 +[41] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 5, 6, 8 +[42] Gongyang Li, Zhi Liu, and Haibin Ling. Icnet: Information conversion network for rgb-d based salient object detection. IEEE TIP, 29:4873-4884, 2020. 1, 2 +[43] Gongyang Li, Zhi Liu, Linwei Ye, Yang Wang, and Haibin Ling. Cross-modal weighting network for rgb-d salient object detection. In ECCV, pages 665-681, 2020. 2 +[44] Guanbin Li, Yuan Xie, Tianhao Wei, Keze Wang, and Liang Lin. Flow guided recurrent neural encoder for video salient object detection. In CVPR, pages 3243-3252, 2018. 1, 2 +[45] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 5, 7 +[46] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In CVPR, pages 15485-15494, 2023. 1 +[47] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 7 +[48] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, pages 678-686, 2016. 2 +[49] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: + +Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 2 +[50] Nian Liu, Ziyang Luo, Ni Zhang, and Junwei Han. Vst++: Efficient and stronger visual saliency transformer. arXiv preprint arXiv:2310.11725, 2023. 4, 7, 8 +[51] Nian Liu, Kepan Nan, Wangbo Zhao, Xiwen Yao, and Junwei Han. Learning complementary spatial-temporal transformer for video salient object detection. IEEE TNNLS, 2023. 2, 6, 8 +[52] Nian Liu, Ni Zhang, and Junwei Han. Learning selective self-mutual attention for rgb-d saliency detection. In CVPR, pages 13756-13765, 2020. 2 +[53] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI, 44(12):9026-9042, 2021. 2, 5, 7 +[54] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, pages 4722-4732, October 2021. 1, 2, 3, 5, 6, 7, 8 +[55] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, pages 1871-1880, 2019. 2 +[56] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for universal foreground segmentations. arXiv preprint arXiv:2305.18476, 2023. 3, 6, 7, 8 +[57] Yi Liu, Dingwen Zhang, Nian Liu, Shoukun Xu, and Jungong Han. Disentangled capsule routing for fast part-object relational saliency. IEEE TIP, 31:6719-6732, 2022. 1 +[58] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. TPAMI, 44(7):3688-3704, 2021. 1 +[59] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 4, 6, 7, 8 +[60] Zhengyi Liu, Song Shi, Quntao Duan, Wei Zhang, and Peng Zhao. Salient object detection for rgb-d image by single stream recurrent convolution neural network. *IJON*, 363:46–57, 2019. 2 +[61] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 1, 2, 6, 8 +[62] Vida Movahedi and James H Elder. Design and perceptual validation of performance measures for salient object segmentation. In CVPR, pages 49-56, 2010. 5, 7 +[63] Xing Nie, Bolin Ni, Jianlong Chang, Gaomeng Meng, Chunlei Huo, Zhaoxiang Zhang, Shiming Xiang, Qi Tian, and Chunhong Pan. Pro-tuning: Unified prompt tuning for vision tasks. arXiv preprint arXiv:2207.14381, 2022. 2 +[64] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In CVPR, pages 454-461, 2012. 5, 7 +[65] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE TPAMI, 36(6):1187-1200, 2013. 5, 8 +[66] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages + +2160-2170, 2022. 1, 2 +[67] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 2 +[68] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Hutchuan Lu. Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE TIP, 32:892-904, 2023. 7, 8 +[69] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In ECCV, pages 92-109, 2014. 5, 7 +[70] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5, 8 +[71] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 1, 2, 5, 7 +[72] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 2 +[73] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In ECCV, pages 212-228. Springer, 2020. 2 +[74] Hongmei Song, Wenguan Wang, Sanyuan Zhao, Jianbing Shen, and Kin-Man Lam. Pyramid dilated deeper convlstm for video salient object detection. In ECCV, pages 715-731, 2018. 2 +[75] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, Hanjing Su, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE TMM, 2023. 8 +[76] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE TIP, 30:5678-5691, 2021. 2, 6, 7, 8 +[77] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A large-scale dataset and benchmark. IEEE TMM, 2022. 1, 2, 5, 6, 7 +[78] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE TMM, 22(1):160-173, 2019. 5, 7 +[79] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. Rgb-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In IJIG, pages 359-369. Springer, 2018. 5, 7 +[80] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2021. 1, 2 +[81] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 5, 6, 7 +[82] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang, + +and Xiang Ruan. Salient object detection with recurrent fully convolutional networks. IEEE TPAMI, 41(7):1734-1746, 2018. 2 +[83] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 1, 2 +[84] Wenguan Wang, Jianbing Shen, Xingping Dong, and Ali Borji. Salient object detection driven by fixation prediction. In CVPR, pages 1711-1720, 2018. 2 +[85] Wenguan Wang, Jianbing Shen, and Ling Shao. Consistent video saliency using local gradient flow optimization and global refinement. IEEE TIP, 24(11):4185-4196, 2015. 5, 8 +[86] Wenguan Wang, Jianbing Shen, and Ling Shao. Video salient object detection via fully convolutional networks. IEEE TIP, 27(1):38-49, 2017. 1, 2 +[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, pages 6830-6839, 2023. 3 +[88] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 3 +[89] Yi Wang, Ruili Wang, Xin Fan, Tianzhu Wang, and Xi-angjian He. Pixels, regions, and objects: Multiple enhancement for salient object detection. In CVPR, pages 10031–10040, 2023. 7, 8 +[90] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, pages 13025-13034, 2020. 2 +[91] Lina Wei, Shanshan Zhao, Omar Farouk Bourahla, Xi Li, Fei Wu, Yueting Zhuang, Junwei Han, and Mingliang Xu. End-to-end video saliency detection via a deep contextual spatiotemporal network. IEEE TNNLS, 32(4):1691-1702, 2020. 1 +[92] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In ICCV, pages 1032-1042, 2023. 1, 2, 8 +[93] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 8 +[94] Jin Xie, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Yanwei Pang, Ling Shao, and Mubarak Shah. Count-and similarity-aware r-cnn for pedestrian detection. In ECCV, pages 88-104. Springer, 2020. 1 +[95] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, pages 15325-15336, 2023. 3 +[96] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 8 +[97] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 7 +[98] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, + +and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 8 +[99] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173, 2013. 5, 7 +[100] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 4 +[101] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 1, 2, 8 +[102] Dingwen Zhang, Junwei Han, Yu Zhang, and Dong Xu. Synthesizing supervision for learning deep saliency network without human annotation. IEEE TPAMI, 42(7):1755-1769, 2019. 2 +[103] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. Rgb-d saliency detection via cascaded mutual information minimization. In ICCV, 2021. 7, 8 +[104] Lu Zhang, Jianming Zhang, Zhe Lin, Huchuan Lu, and You He. Capsal: Leveraging captioning to boost semantics for salient object detection. In CVPR, pages 6024–6033, 2019. 2 +[105] Miao Zhang, Jie Liu, Yifei Wang, Yongri Piao, Shunyu Yao, Wei Ji, Jingjing Li, Huchuan Lu, and Zhongxuan Luo. Dynamic context-sensitive filtering network for video salient object detection. In ICCV, pages 1553-1563, 2021. 8 +[106] Miao Zhang, Weisong Ren, Yongri Piao, Zhengkun Rong, and Huchuan Lu. Select, supplement and focus for rgb-d saliency detection. In CVPR, pages 3472-3481, 2020. 2 +[107] Qiang Zhang, Nianchang Huang, Lin Yao, Dingwen Zhang, Caifeng Shan, and Jungong Han. Rgb-t salient object detection via fusing multi-level cnn features. IEEE TIP, 29:3321-3335, 2019. 1, 2 +[108] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 1, 2 +[109] Jia-Xing Zhao, Yang Cao, Deng-Ping Fan, Ming-Ming Cheng, Xuan-Yi Li, and Le Zhang. Contrast prior and fluid pyramid integration for rgbd salient object detection. In CVPR, pages 3927-3936, 2019. 1, 2 +[110] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet:edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 2 +[111] Wangbo Zhao, Kepan Nan, Songyang Zhang, Kai Chen, Dahua Lin, and Yang You. Learning referring video object segmentation from weak annotation. arXiv preprint arXiv:2308.02162, 2023. 1 +[112] Wangbo Zhao, Jiasheng Tang, Yizeng Han, Yibing Song, Kai Wang, Gao Huang, Fan Wang, and Yang You. Dynamic tuning towards parameter and inference efficiency for vit adaptation. arXiv preprint arXiv:2403.11808, 2024. 2 +[113] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51, 2020. 2 +[114] Dehua Zheng, Xiaochen Zheng, Laurence T Yang, Yuan + +Gao, Chenlu Zhu, and Yiheng Ruan. Mffn: Multi-view feature fusion network for camouflaged object detection. In WACV, pages 6232-6242, 2023. 2 +[115] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, pages 9516-9526, 2023. 3 +[116] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 1, 6, 7, 8 +[117] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, pages 15116-15127, 2023. 3 \ No newline at end of file diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/images.zip b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..145f70e1b2c92db8da68f49dc692a0ae7b3a82f8 --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0604a348ce8f481c569899396a0c97884b8a30e0c44df149d0ce70510fc31a59 +size 865492 diff --git a/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/layout.json b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d5f22ec5d9177de081015b8c8fe78cb72ed55d23 --- /dev/null +++ b/2024/VSCode_ General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning/layout.json @@ -0,0 +1,10303 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "type": "text", + "content": "VSCode: General Visual Salient and Camouflaged Object Detection with 2D Prompt Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": "Ziyang Luo" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Nian Liu" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Wangbo Zhao" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Xuguang Yang" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Dingwen Zhang" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Deng-Ping Fan" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Fahad Khan" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "text", + "content": " Junwei Han" + }, + { + "bbox": [ + 112, + 160, + 505, + 189 + ], + "type": "inline_equation", + "content": "^{1,7,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "spans": [ + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "Northwestern Polytechnical University " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "Mohamed bin Zayed University of Artificial Intelligence " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "CVL, Linköping University " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "TBI Center, CS, Nankai University " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "Nankai International Advanced Research Institute (SHENZHEN FUTIAN) " + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 52, + 190, + 560, + 246 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 298, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 501 + ], + "type": "text", + "content": "Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD. Source code has been available at https://github.com/Sssssuperior/VSCode." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 542, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 288, + 685 + ], + "type": "text", + "content": "Visual salient object detection (SOD) and camouflaged object detection (COD) are two interconnected yet unique tasks. The goal of SOD is to identify prominent objects within an image that significantly contrast with their surroundings [5], which can be used to promote segmentation [46, 111], detection [94], and Part-Object Relational visual saliency [57, 58]. While COD focuses on identifying objects concealed within their environment. These objects intentionally blend in by sharing structural or textural similarities with their surroundings [15]. Despite the seemingly different definitions of SOD and COD, they both belong to the realm of binary segmentation and share some vital fundamental similarities, such as" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 271, + 547, + 403 + ], + "blocks": [ + { + "bbox": [ + 307, + 271, + 547, + 403 + ], + "lines": [ + { + "bbox": [ + 307, + 271, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 547, + 403 + ], + "type": "image", + "image_path": "d9dcc601fd71069722e03084033f4d8f3f14eb25d1c0a16179f54b2847a67cec.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 406, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 406, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Relationship of SOD, COD, and multimodal tasks. Each specific task is seen as a combination of two dimensions, i.e. domain (RGB/Depth/Thermal/Flow) and task (SOD/COD)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 449, + 429, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 429, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 429, + 460 + ], + "type": "text", + "content": "objectness and structuredness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": "To cater to various scenarios, both SOD and COD have given rise to several sub-tasks with different modalities, including RGB SOD [83, 108], RGB COD [27, 66, 101], RGB-D SOD [42, 71], RGB-D COD [92], and RGB-T SOD [80, 107]. By leveraging optical flow maps, Video SOD (VSOD) [44, 91] and VCOD [10, 36] tasks can also be seen as a combination of two modalities. The relationship of SOD, COD, and multimodal tasks is shown in Figure 1, where each specific task can be considered as a combination of two dimensions, i.e. domain and task. Although these multimodal tasks differ in the complementary cues they employ, these modalities share some key commonalities. For instance, depth, thermal, and optical flow maps often show obvious objectness as in RGB images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": "Although previous CNN-based [10, 61, 71, 77, 86, 109] and transformer-based [54, 116] approaches have effectively addressed these tasks and achieved favorable results, they usually rely on meticulously designed models to tackle each task individually. Designing models specifically for individual tasks can be problematic since the training data of one task is typically limited. Task-specific specialist models" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Corresponding author: Nian Liu (liunian228@gmail.com) and Junwei Han (junweihan2010@gmail.com)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17169" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 70, + 255, + 125 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 255, + 125 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 255, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 255, + 125 + ], + "type": "image", + "image_path": "90588295a26c4c21545a4a8b1af0048b5eef90aa5c2d19338e6c1063d0e4abba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 127, + 286, + 194 + ], + "blocks": [ + { + "bbox": [ + 49, + 127, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 127, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 127, + 286, + 194 + ], + "type": "image", + "image_path": "1787677a2f090fef175c757d803891dbd6a1c216ffcc84e119d4ff67a546af24.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "text", + "content": "Figure 2. Overall architecture of our VSCode model. We use VST [54] as the foundation model to acquire commonalities among multimodal SOD and COD tasks. For each task, we integrate 2D prompts to aggregate peculiarities along the domain dimension and the task dimension, including four domain-specific prompts and two task-specific prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 280, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 289, + 483 + ], + "type": "text", + "content": "may be overly adapted to a particular task and overfitted to specific training data distribution, which ultimately sacrifices generalization ability and results in suboptimal performance. One solution may be using more data, however, being costly and time-consuming for data annotation. To this end, joint learning a generalist model emerges as a more promising option, as it allows for the maximum use of all data and the effective learning of the commonalities of all tasks, hence significantly reducing the risk of overfitting and enhancing the generalization capability [34, 55]. However, joint learning multiple tasks is not straightforward. On one hand, simultaneously handling both commonalities and peculiarities of all tasks poses a significant challenge as the incompatibility among different tasks easily leads to a decline in performance with simple joint training [40]. On the other hand, it usually introduces additional complexity, computational costs, and parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "In this paper, we present a general Visual Salient and Camouflaged object detection (VSCode) model which encapsulates both commonalities and peculiarities of different tasks with a simple but effective design, as illustrated in Figure 2. On one hand, we adopt VST [54] as the shared segmentation foundation model to assimilate commonalities of different tasks by leveraging its simple and pure-transformer-based architecture. On the other hand, inspired by the recent emergence of the parameter-efficient prompting technique [31, 63, 112], we propose 2D prompts to capture task peculiarities. Specifically, we decompose these peculiarities along the domain dimension and the task dimension, and consequently design domain-specific prompts and task-specific prompts to comprehend the differences among diverse domains and tasks, respectively. These 2D prompts can effectively disentangle domain and task peculiarities, making our model easily adaptable by combining them to tackle specific tasks and even unseen ones. Furthermore, we present a prompt discrimination loss to encourage the 2D prompts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "to focus on acquiring adequate peculiarities and enable the foundational model to concentrate on commonality learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 98, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 547, + 181 + ], + "type": "text", + "content": "Finally, we train our VSCode model on four SOD tasks and two COD tasks, demonstrating its effectiveness against state-of-the-art methods. What's more, we carry out evaluations on a reserved task and reveal remarkable zero-shot generalization ability of our model, which has never been explored in previous works. The main contributions in this work can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 184, + 547, + 339 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 184, + 545, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 545, + 207 + ], + "type": "text", + "content": "- We present VSCode, the first generalist model for multimodal SOD and COD tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 209, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 547, + 255 + ], + "type": "text", + "content": "- We propose to use a foundation segmentation model to aggregate commonalities and introduce 2D prompts to learn peculiarities along the domain and task dimensions, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 256, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 256, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 545, + 292 + ], + "type": "text", + "content": "- A prompt discrimination loss is proposed to effectively enhance the learning of peculiarities and commonalities for 2D prompts and the foundation model, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 293, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 293, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 547, + 339 + ], + "type": "text", + "content": "- Our VSCode model surpasses all existing state-of-the-art models across all tasks on 26 datasets and showcases its ability to generalize to unseen tasks, further emphasizing the superiority of our approach." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 392, + 361 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 367, + 501, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 501, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 501, + 380 + ], + "type": "text", + "content": "2.1. Deep Learning Based SOD and COD" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 385, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 588 + ], + "type": "text", + "content": "SOD. Previous RGB SOD works delved into attention-based [9, 19, 49, 71, 108], multi-level fusion-based [20, 24, 67, 83, 102, 113], recurrent-model-based [8, 12, 48, 60, 82], and multi-task-based methods [29, 72, 84, 90, 104, 106, 110]. In the case of RGB-D SOD, some models [9, 42, 43, 52, 53, 71, 109] leveraged various attention mechanisms to incorporate depth cues into RGB features. With regard to RGB-T SOD, recent studies also introduced attention-based methods [77, 80] and multi-level fusion [76, 107] to excavate the relationship between RGB and thermal features. Regarding the VSOD task, some works [18, 21, 30, 74, 86] mined spatial-temporal and appearance cues. More recently, there was a growing trend where various research [28, 44, 51, 73] endeavored to incorporate optical flow for combining motion cues with appearance details. Consistent with recent studies, we treat optical flow as a form of modality information and view VSOD as a multimodal SOD task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "content": "COD. Currently, COD has RGB COD, RGB-D COD, and VCOD tasks. RGB COD methods can be broadly categorized as multi-task-based approaches [61, 101], multi-input-based approaches [66, 114], and refinement-based approaches [15, 32]. The RGB-D COD task was initially introduced in [92], where depth inference models are adapted for object segmentation. For VCOD, prior studies segmented the moving camouflaged objects via dense optical flow [3, 4] or well-designed models [10, 36]. For a more comprehensive literature review, please refer to [16]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17170" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 208 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 208 + ], + "type": "image", + "image_path": "1bc501445f115cb0c0cd9229560adb5ba2902e8b4b1afffc5ded6241d82cfb87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "type": "text", + "content": "Figure 3. Overall framework of our proposed VSCode model with 2D prompt learning. Based on the VST [54] foundation model, we insert the respective domain-specific prompts and task-specific prompts in the attention windows in the Swin transformer [59] encoder layers to learn domain and task-specific encoder features. The converter is used for multimodal feature fusion. Within the transformer decoder layers, task-specific prompts are appended to image feature tokens to perform task-specific decoding. We also provide detailed structures of an encoder layer " + }, + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "type": "inline_equation", + "content": "(i = 0)" + }, + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "type": "text", + "content": " and a decoder layer " + }, + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "type": "inline_equation", + "content": "(j = 0)" + }, + { + "bbox": [ + 46, + 211, + 547, + 268 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 276, + 201, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 201, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 201, + 289 + ], + "type": "text", + "content": "2.2. Prompt in Computer Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 411 + ], + "type": "text", + "content": "Prompt was initially introduced in the field of NLP [6] and has been successfully integrated into computer vision tasks [25]. VPT [31] introduced a small amount of trainable parameters as prompts in the input space. ViPT [115] put forth the idea of modality-complementary prompts for task-oriented multi-modal tracking. Prior research has primarily focused on specific tasks, such as classification or tracking. In this paper, we propose to use 2D prompts for assembling different multimodal tasks and enable zero-shot generalization on unseen tasks, which has not been explored before." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 412, + 248, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 248, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 248, + 425 + ], + "type": "text", + "content": "2.3. Generalist Segmentation Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 581 + ], + "type": "text", + "content": "Recently, several generalist frameworks have emerged for a range of segmentation tasks using a variety of prompts [1]. On one hand, X-Decoder [117] utilized generic nonsemantic queries and semantic queries to decode different pixel-level and token-level outputs. UNINEXT [95] introduced three types of prompts, namely category names, language expressions, and reference annotations. On the other hand, Painter [87] and SegGPT [88] leveraged image-mask pairs from the same task as prompts. Unlike the approaches mentioned above, which mainly concentrate on task differences, our VSCode dissects unique characteristics based on both domain and task dimensions, leading to a more versatile design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 643 + ], + "type": "text", + "content": "In the field of SOD and COD, EVP [56] introduced adaptors into the encoder and trained each task individually for various foreground segmentation tasks. Different from them, we consider not only multiple tasks but also multiple modalities and we train all tasks simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 648, + 129, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 129, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 129, + 662 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In this work, we propose VSCode with the aim of jointly training SOD and COD tasks in an efficient and effective way. We allow VST [54] to incorporate commonalities (Section 3.1), and utilize 2D prompts, which comprise domain" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 276, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 547, + 348 + ], + "type": "text", + "content": "specific (Section 3.2) and task-specific prompts (Section 3.3), to encapsulate peculiarities. To accurately disentangle domain and task peculiarities in 2D prompts and encourage commonality learning in VST, we introduce a prompt discrimination loss (Section 3.5). Figure 3 shows the overall architecture of our proposed VSCode." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 351, + 416, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 416, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 416, + 362 + ], + "type": "text", + "content": "3.1. Foundation Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": "To achieve a more comprehensive integration of commonities from SOD and COD tasks, we select VST [54] as our fundamental model. VST was originally proposed for RGB and RGB-D SOD and comprises three primary components, i.e. a transformer encoder, a transformer converter, and a multi-task transformer decoder. It initially employs the transformer encoder to capture long-range dependencies within the image features " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^E \\in \\mathbb{R}^{l_i \\times c_i}" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "i \\in [0,1,2,3]" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": " indicates the index of blocks in the encoder, " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": " mean the length of the patch sequence and the channel number of " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^E" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": ". Subsequently, the transformer converter integrates the complement between RGB and depth features via cross-attention for RGB-D SOD or uses self-attention for RGB SOD. In the decoder, which is composed of a sequence of self-attention layers, VST predicts saliency maps and boundary maps simultaneously via a saliency token, a boundary token, and decoder features " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{f}_j^D \\in \\mathbb{R}^{l_j \\times d}" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": " corresponds the index of blocks in the decoder. Here " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "j \\in [2,1,0]" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": " for descending order and " + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "inline_equation", + "content": "d = 384" + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": ". Due to the simple and pure-transformer-based architecture, VST can be easily used for other multimodal tasks and COD tasks without the need for model redesign. As a result, it emerges as a superior choice for constructing a generalist model for general multimodal SOD and COD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "In pursuit of improved outcomes and a more suitable structure, we introduce modifications to VST. First, we select Swin transformer [59] as our backbone due to its efficiency and high performance. Second, to maintain a unified structure for both RGB tasks and other multimodal tasks," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17171" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "we utilize the RGB converter in VST, which comprises standard transformer layers. For multimodal tasks, we simply concatenate the supplementary modality's features with the RGB features along the channel dimension and employ a multilayer perceptron (MLP) to project them from " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "2d" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " channels to " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " channels. For RGB tasks, no alterations are made. Third, we incorporate certain extensions from " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{VST}++" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " [50], specifically including the token-supervised prediction loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 174, + 185, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 185, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 185, + 186 + ], + "type": "text", + "content": "3.2. Domain-specific Prompt" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "text", + "content": "Within the encoder, lower layers are dedicated to extracting low-level features, encompassing edges, colors, and textures, which exhibit distinct characteristics in various domains [100]. For instance, depth maps are typically rendered in grayscale, while thermal maps present a broader color spectrum. Higher layers, on the other hand, capture semantic information from modality features, which is crucial for all tasks. Consequently, we introduce domain-specific prompts " + }, + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^d" + }, + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "text", + "content": " at each block " + }, + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 189, + 289, + 332 + ], + "type": "text", + "content": " in the encoder and design four kinds of domain-specific prompts for RGB, depth, thermal, and optical flow, respectively, to highlight the disparities among domains, as shown in Figure 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": "Given the image features " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^E" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " from a specific block in the Swin transformer encoder, we use window-attention [59] and partition the feature " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^E" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " into window features " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i - w}^{E}\\in \\mathbb{R}^{l_i / M^2\\times M^2\\times c_i}" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " represents the window size and " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "l_{i} / M^{2}" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " is the number of windows. Then, we replicate the prompts " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^d\\in \\mathbb{R}^{N_i\\times c_i}" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " for each window and obtain " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{d'}\\in \\mathbb{R}^{l_i / M^2\\times N_i\\times c_i}" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 46, + 333, + 289, + 456 + ], + "type": "text", + "content": " represents the number of learnable prompt tokens. Next, we append them to the patch feature tokens in each window and perform self-attention within each window, which can be defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 464, + 287, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 464, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 85, + 464, + 287, + 492 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i + 1} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\text {S W / W - M S A} \\left(\\left[ \\begin{array}{l} \\boldsymbol {p} _ {i} ^ {d ^ {\\prime}} \\\\ \\boldsymbol {f} _ {i - w} ^ {E} \\end{array} \\right]\\right)), \\tag {1}", + "image_path": "8ccd993d69026d14769871ec519b2a2f2ff1eb926cfa68c064ac24a55bf6d3d7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "content": "where W-MSA and SW-MSA are multi-head self-attention modules with regular and shifted windowing configurations, respectively. Here we omit the residual connection [23], and layer normalization [2]. Next, we segment " + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{i+1}^{d'}" + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "content": " from each window and calculate the average of them to obtain " + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{i+1}^d" + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "content": ", and then reassemble the output window feature " + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i-w}^E" + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i+1}^E" + }, + { + "bbox": [ + 46, + 500, + 289, + 584 + ], + "type": "text", + "content": " for the next block." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 169, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 169, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 169, + 602 + ], + "type": "text", + "content": "3.3. Task-specific Prompt" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Prior research [39] has traditionally regarded SOD and COD as opposing tasks, emphasizing the disparities between the features extracted by the SOD encoder and the COD encoder as much as possible. However, we believe that SOD and COD share significant commonalities in their features, such as low-level cues, high-level objectness, and spatial structuredness. As a result, we introduce task-specific prompts to learn the peculiarities while retaining the primary stream parameters shared to capture commonalities. We add the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 68, + 547, + 176 + ], + "blocks": [ + { + "bbox": [ + 308, + 68, + 547, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 68, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 68, + 547, + 176 + ], + "type": "image", + "image_path": "6af400023f52890cbabb02097cd0d57b3ff39542b90839c1003789f2a2480639.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 198 + ], + "type": "text", + "content": "Figure 4. Illustration of the influence of using different task prompts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 203, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 547, + 239 + ], + "type": "text", + "content": "task-specific prompts in both VST encoder and decoder, and the overall impact of adding these prompts is illustrated in Figure 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 239, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 430 + ], + "type": "text", + "content": "Encoder. Although the encoder primarily focuses on domain-specific features with domain prompts, semantic features still play a pivotal role in distinguishing SOD and COD tasks. Semantic features from the encoder typically emphasize the most relevant region for a particular task and allocate more attention accordingly. In the case of the SOD task, the foreground region receives greater attention, whereas for the COD task, the background usually gains large importance since objects are typically concealed within it. Hence, it is essential to incorporate task-specific prompts to encourage learning task-related features in the encoder. Otherwise, we risk initially activating the wrong objects before the decoding process. Following the pattern of domain-specific prompts, we introduce task-specific prompts " + }, + { + "bbox": [ + 304, + 239, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{te} \\in \\mathbb{R}^{N_i \\times c_i}" + }, + { + "bbox": [ + 304, + 239, + 547, + 430 + ], + "type": "text", + "content": " in each encoder block and use them in the same way as how domain-specific prompts are used." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 430, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 596 + ], + "type": "text", + "content": "Decoder. Camouflaged objects typically exhibit more intricate and detailed boundaries compared to salient objects. This complexity arises because concealed objects often share color or textual similarities with their surroundings, resulting in imperceptible boundaries. Therefore, solely introducing task-specific prompts in the encoder may not be adequate, as camouflaged objects require a more refined process within the decoder. We incorporate task-specific prompts in the decoder to allocate distinct attention for reconstructing both the boundary and object regions based on the features extracted by the encoder. In contrast, previous research [39] has not adequately explored the differences between these two tasks in the decoder, as they typically use a single decoder to handle both." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "text", + "content": "Regarding task-specific prompts in the decoder, we simply append learnable prompts " + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{j+1}^{td} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "text", + "content": " to the decoder feature tokens " + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{j+1}^{D}" + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "text", + "content": " from a specific block " + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "inline_equation", + "content": "j+1" + }, + { + "bbox": [ + 305, + 598, + 548, + 648 + ], + "type": "text", + "content": " in the decoder. Then, we apply the self-attention as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 654, + 547, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 654, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 359, + 654, + 547, + 682 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j} ^ {t d} \\\\ \\boldsymbol {f} _ {j} ^ {D} \\end{array} \\right] \\leftarrow \\operatorname {M L P} (\\operatorname {M S A} (\\left[ \\begin{array}{l} \\boldsymbol {p} _ {j + 1} ^ {t d} \\\\ \\boldsymbol {f} _ {j + 1} ^ {D} \\end{array} \\right])) \\tag {2}", + "image_path": "f7b6bbc7c2a8b937596401cb167707cb18dafefb2081e2862fa7d13bbcd9beb7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "where MSA denotes the multi-head self-attention. Here we omit the saliency and boundary tokens in the VST decoder" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17172" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 289, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 217 + ], + "type": "image", + "image_path": "6888e4c12a8ae976cf966a2359425a17f932a17732c20915eeba0a3d07603d39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 219, + 284, + 231 + ], + "lines": [ + { + "bbox": [ + 49, + 219, + 284, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 284, + 231 + ], + "type": "text", + "content": "Figure 5. Correlation of prompt pairs at each encoder block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 235, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 287, + 270 + ], + "type": "text", + "content": "for conciseness. Please note that our task-specific prompts differ from saliency and boundary tokens since we do not introduce any supervision for them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 272, + 219, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 219, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 219, + 284 + ], + "type": "text", + "content": "3.4. Prompts Layout and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 287, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 430 + ], + "type": "text", + "content": "To incorporate the aforementioned prompts within the encoder-decoder architecture, inspired by VPT [31], we offer two prompt inserting versions. In the deep version, new prompts are introduced at the start of each transformer block, whereas the shallow version involves proposing prompts at first and updating them across all blocks. To unveil the specific relationship among different domains and tasks at varying network depths, we employ the deep version for both domain and task-specific prompts within the encoder. Based on VST's design, which introduces a saliency and a boundary token at the beginning of the decoder, we use the shallow version for task-specific prompts in the decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 431, + 289, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 289, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 289, + 670 + ], + "type": "text", + "content": "We calculate the correlations of different domain and task prompt pairs at different blocks in Figure 5. It is evident that depth, thermal, and optical flow exhibit relatively strong correlations in low-level features, as all of them usually show obvious low-level contrast between target objects and backgrounds in terms of color or luminance. However, at higher levels, most domains exhibit lower correlations, highlighting the distinctions among them. Additionally, as for task-specific prompts, it is clear that SOD prompts and COD prompts exhibit more shared knowledge in the lower layers. As we progress to higher layers, the correlation decreases, indicating that high-level features gradually learn unrelated information. This observation urges us to implement the deep version of domain-specific prompts and task-specific prompts in the encoder in our final design, as different blocks acquire distinct knowledge. Moreover, the gradually decreased correlation values along with the increase of the network depth encourage us to use a progressively larger number of prompt tokens, as lower correlation means larger peculiarities and hence requires more parameters to learn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 673, + 137, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 137, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 137, + 685 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The design principle of our model is to use 2D prompts for encompassing peculiarities while integrating commonalities" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": "into the foundation model. However, this is not straightforward for freely learned prompts. As shown in Figure 5, they still suggest certain correlations. This indicates that the learned prompts are entangled, risking the model's capacity to differentiate among various domains and tasks and resulting in suboptimal optimization. Hence, we propose a prompt discrimination loss to minimize the correlation among the prompts of the same type, guaranteeing that each prompt acquires unique domain or task knowledge. Specifically, we aggregate prompts of the same domain/task into a single embedding and then perform discrimination. First, we average the input prompt tokens of each same prompt type at each block and use linear projections to align the channel numbers to " + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": ". Subsequently, for each type of prompt, we concatenate the averaged prompts of different blocks, and use MLP to obtain the overall domain-specific prompt " + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "inline_equation", + "content": "p_{l}^{d_{all}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": " and task-specific encoder prompt " + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "inline_equation", + "content": "p_{k}^{te_{all}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 283, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 283, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 326, + 283, + 545, + 304 + ], + "type": "interline_equation", + "content": "p _ {l} ^ {d _ {a l l}} = \\operatorname {M L P} \\left[ \\mathrm {L A} \\left(\\boldsymbol {p} _ {0} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {1} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {2} ^ {d}\\right); \\mathrm {L A} \\left(\\boldsymbol {p} _ {3} ^ {d}\\right) \\right], \\tag {3}", + "image_path": "0bbfacf80f796cddef9e00202a53a06d15d45f6177ddcf4a1df46062a6ec2a17.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 301, + 529, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 301, + 529, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 301, + 529, + 316 + ], + "type": "interline_equation", + "content": "p _ {k} ^ {t e _ {a l l}} = \\mathrm {M L P} [ \\mathrm {L A} (\\boldsymbol {p} _ {0} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {1} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {2} ^ {t e}); \\mathrm {L A} (\\boldsymbol {p} _ {3} ^ {t e}) ],", + "image_path": "2718f27267cbd7d0856de31deb33b3b0808154bfa30df7e3d416e05c20268a24.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{L}" + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{A}" + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "content": " represent the linear and average operation, respectively, with " + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "inline_equation", + "content": "l \\in \\{\\text{depth}, \\text{thermal}, \\text{flow}, \\text{rgb}\\}" + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "inline_equation", + "content": "k \\in \\{SOD, COD\\}" + }, + { + "bbox": [ + 306, + 322, + 545, + 369 + ], + "type": "text", + "content": ". Since the task-specific prompts in the decoder are shallow, we simply average them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "content": "Afterward, we calculate the cosine similarity between prompt pairs, resulting in eight types of cosine similarity results " + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{CS}_m" + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "content": " means the combination of domains/tasks, namely " + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\{RD, RT, RF, DF, DT, TF\\}" + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "content": " for domain-aggregated prompts and " + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\{SC_{EN}, SC_{DE}\\}" + }, + { + "bbox": [ + 304, + 369, + 547, + 466 + ], + "type": "text", + "content": " for task-aggregated prompts in the encoder and decoder, respectively. Finally, we minimize the correlation within these prompt pairs and define our prompt discrimination loss as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 369, + 472, + 545, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 472, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 369, + 472, + 545, + 497 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} = \\sum_ {m} \\ln (1 + | \\mathcal {C S} _ {m} |), \\tag {4}", + "image_path": "7be2b716b71f701a1e6681d3c85069923140dd7ef6349b1175f6b12336ca8eed.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 504, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 528 + ], + "type": "text", + "content": "which is further combined with the segmentation losses and boundary losses [54] to train our model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 533, + 382, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 382, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 382, + 545 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 549, + 481, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 481, + 561 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 567, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 713 + ], + "type": "text", + "content": "For RGB SOD, we evaluate our proposed model using six commonly used benchmark datasets, i.e. DUTS [81], ECSSD [97], HKU-IS [45], PASCAL-S [47], DUT-O [99], and SOD [62]. For RGB-D SOD, we use six large benchmark datasets, including STERE [64], NJUD [33], NLPR [69], DUTLF-Depth [71], SIP [17], and ReDWeb-S [53]. In terms of RGB-T SOD, we consider three public datasets: VT821 [79], VT1000 [78], and VT5000 [77]. For VSOD, we employ six widely used benchmark datasets: DAVIS [70], FBMS [65], ViSal [85], SegV2 [41], DAVSOD-Easy, and DAVSOD-Normal [18]. Regarding RGB COD, three extensive benchmark datasets are considered, including" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17173" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 60, + 541, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 60, + 541, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 60, + 541, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 60, + 541, + 162 + ], + "type": "table", + "html": "
SettingsParams (M)RGB SOD DUTS[81] Sm↑Fm↑Em↑RGB-D SOD NJUD[33] Sm↑Fm↑Em↑RGB-T SOD VT5000[77] Sm↑Fm↑Em↑VSOD SegV2[41] Sm↑Fm↑Em↑RGB COD CAMO[37] Sm↑Fm↑Em↑VCOD CAD[3] Sm↑Fm↑Em↑
.900.885.940.927.928.958.900.863.938.896.870.952.793.751.871.686.522.787
ST323.46*.898.884.941.924.922.954.903.886.942.930.911.972------
GT54.06.902.890.945.931.932.962.909.877.947.931.917.975------
GT+p^d54.06.902.890.945.931.931.961.906.892.946.925.910.970.804.776.876.759.639.808
GT+p^d+pt54.09.909.899.948.935.938.965.912.882.950.943.930.984.811.782.884.736.614.797
GT+p^d+pt+L_dis54.09.908.896.947.932.932.960.909.878.947.933.907.966.800.770.872.743.611.798
w/o p^te54.07.908.896.947.932.929.959.904.872.941.940.919.975.799.770.875.740.599.814
w/o p^td54.08.902.889.943.929.929.959.904.872.941.940.919.975.799.770.875.740.599.814
", + "image_path": "453549e6085f2f85241e63222a7a89fe215e898a265beb9a299dda5628ea6e41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 216, + 541, + 434 + ], + "blocks": [ + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": "Table 1. Ablation studies of our VSCode on the Swin-T [59] backbone with " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " image size. We conduct evaluations on one representative dataset for each task. \"ST\" indicates special training, \"GT\" means general training, " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "p^d" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " represents domain-specific prompts, and " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "p^t" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " is task-specific prompts, which consists of " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "p^{te}" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " in the encoder and " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "p^{td}" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " in the decoder. " + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 47, + 163, + 547, + 208 + ], + "type": "text", + "content": " is our prompt discrimination loss. The best results under each setting are labeled in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 216, + 541, + 434 + ], + "lines": [ + { + "bbox": [ + 50, + 216, + 541, + 434 + ], + "spans": [ + { + "bbox": [ + 50, + 216, + 541, + 434 + ], + "type": "table", + "html": "
SettingsParams(M)RGB SOD DUTS[81] Sm↑ Fm↑ Em↑RGB-D SOD NJUD[33] Sm↑ Fm↑ Em↑RGB-T SOD VT5000[77] Sm↑ Fm↑ Em↑VSOD SegV2[41] Sm↑ Fm↑ Em↑RGB COD CAMO[37] Sm↑ Fm↑ Em↑VCOD CAD[3] Sm↑ Fm↑ Em↑
shallow or deep version for domain-specific prompts
shallow54.83.900.887.943.926.926.955.905.873.944.931.917.972------
deep54.06.902.890.945.931.932.962.909.877.947.931.917.975------
shallow or deep version for task-specific prompts in the encoder
shallow54.45.902.888.943.928.928.958.902.870.940.927.905.964.793.763.866.747.616.798
deep54.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
shallow or deep version for task-specific prompts in the decoder
deep54.10.903.891.945.930.932.960.905.888.943.922.903.966.802.774.881.738.605.801
shallow54.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
number of domain-specific prompts at four blocks
1,1,1,154.06.902.890.945.931.932.962.909.877.947.931.917.975------
5,5,5,554.08.903.890.945.931.935.961.902.869.940.918.887.954------
number of task-specific prompts in the encoder at four blocks
5,5,5,554.07.903.893.947.928.930.959.903.870.940.931.918.975.795.766.866.739.600.799
1,1,5,1054.08.903.890.944.934.934.962.905.874.943.924.903.960.804.772.881.759.651.831
number of task-specific prompts in the decoder
554.08.904.890.946.929.931.957.904.890.943.931.911.969.807.782.881.746.626.805
1054.09.904.892.945.930.931.961.906.892.946.925.910.970.804.776.876.759.639.808
1554.09.903.889.944.929.933.956.904.890.942.932.913.974.798.771.875.743.621.791
", + "image_path": "2d025a612dd0f3e9162e090e9a2626cbc8b4dea50eb343fbd5a91cd8f5c6c67c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 434, + 416, + 444 + ], + "lines": [ + { + "bbox": [ + 176, + 434, + 416, + 444 + ], + "spans": [ + { + "bbox": [ + 176, + 434, + 416, + 444 + ], + "type": "text", + "content": "Table 2. Ablation studies of different designs of prompt layout." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "text", + "content": "COD10K [15], CAMO [37], and NC4K [61]. For VCOD, we utilize two widely accepted benchmark datasets: CAD [3] and MoCA-Mask [10]. To ensure a consistent evaluation across all SOD and COD tasks, we employ three commonly used evaluation metrics to assess model performance: structure-measure " + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "text", + "content": " [13], maximum enhanced-alignment measure " + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "text", + "content": " [14], and maximum F-measure " + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 46, + 453, + 288, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 542, + 180, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 180, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 180, + 554 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 676 + ], + "type": "text", + "content": "Building on prior research [10, 22, 51, 54, 77], we employ the following datasets to train our model concurrently: the training set of DUTS for RGB SOD, the training sets of NJUD, NLPR, and DUTLF-Depth for RGB-D SOD, the training set of VT5000 for RGB-T SOD, the training sets of DAVIS and DAVSOD for VSOD, the training sets of COD10K and CAMO for RGB COD, and the training set of MoCA-Mask for VCOD. To ensure a fair comparison with previous works [22, 38, 56, 76, 116], we resize each" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "content": "image to " + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "content": " pixels and then randomly crop them to " + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 304, + 453, + 545, + 548 + ], + "type": "text", + "content": " image regions for training. Our training process employs the Adam optimizer [35] with an initial learning rate of 0.0001, which is reduced by a factor of 10 at half and three-quarters of the total training steps. We conduct a total of 150,000 training steps using a 3090 GPU. We mix the above six tasks in each training iteration with two samples for each task, leading to a total batch size of 12." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 554, + 400, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 400, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 400, + 567 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Architecture Design. To demonstrate the efficacy of various components in our VSCode model, we report the quantitative results in Table 1. We start by performing special training (ST) on each task individually and then conduct general training (GT) on all SOD tasks. Please note that here we do not consider COD tasks since no task prompt is used. We observe improved performance on RGB-T SOD and VSOD, demonstrating the significant benefit of shared knowledge in different tasks, especially for those with limited training data diversity. However, the results of RGB SOD and RGB-D SOD do not show a significant increase. Our hypothesis is that amalgamating the training of multi-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "*The parameters for our specialized training methods amount to 53.61M for the RGB task and 54.06M for the multimodal task, resulting in a total of 323.46M parameters for all six tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17174" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 65, + 543, + 157 + ], + "blocks": [ + { + "bbox": [ + 48, + 65, + 543, + 157 + ], + "lines": [ + { + "bbox": [ + 48, + 65, + 543, + 157 + ], + "spans": [ + { + "bbox": [ + 48, + 65, + 543, + 157 + ], + "type": "table", + "html": "
MethodParams (M)DUTS[81]ECSSD[97]HKU-IS[45]PASCAL-S[47]DUT-O[99]SOD[62]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
VST[54]44.48.896.877.939.932.944.964.928.937.968.873.850.900.850.800.888.854.866.902
ICON-R[116]33.09.890.876.931.928.943.960.920.931.960.862.844.888.845.799.884.848.861.899
VST-T++ [50]53.60.901.887.943.937.949.968.930.939.968.878.855.901.853.804.892.853.866.899
MENet[89]27.83.905.895.943.927.938.956.927.939.965.871.848.892.850.792.879.841.847.884
VSCode-T54.09.917.910.954.945.957.971.935.946.970.878.852.900.869.830.910.863.879.908
EVP[56]64.52†.917.910.956.936.949.965.935.945.971.880.859.902.864.822.902.854.873.901
VSCode-S74.72†.926.922.960.949.959.974.940.951.974.887.864.904.877.840.912.870.882.910
", + "image_path": "bbf2b19ecabc1f06ed69a6a79a750767db0e8c2185db22effb881c26c6948d96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 200, + 543, + 292 + ], + "blocks": [ + { + "bbox": [ + 46, + 158, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 545, + 191 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of our VSCode with other 5 SOTA RGB SOD methods on six benchmark datasets. “-R”, “-T” and “-S” mean the ResNet50 [23], Swin-T, and Swin-S[59] backbones, respectively. “-” indicates the code is not available. The best performance under all settings is bolded, and the best results under each setting are labeled in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 200, + 543, + 292 + ], + "lines": [ + { + "bbox": [ + 48, + 200, + 543, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 543, + 292 + ], + "type": "table", + "html": "
MethodParams (M)NJUD [33]NLPR[69]DUTLF-Depth[71]ReDWeb-S[53]STERE[64]SIP[17]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
CMINet[103]188.12.929.934.957.932.922.963.912.913.938.725.726.800.918.916.951.899.910.939
VST[54]53.83.922.920.951.932.920.962.943.948.969.759.763.826.913.907.951.904.915.944
VST-T++ [50]100.51.928.929.958.933.921.964.944.948.969.756.757.819.916.911.950.903.914.944
SPSN[38]----.923.912.960------.907.902.945.892.900.936
CAVER[68]55.79.920.924.953.929.921.964.931.939.962.730.724.802.914.911.951.893.906.934
VSCode-T54.09.941.945.967.938.930.966.952.959.974.766.771.831.928.926.957.917.936.955
VSCode-S74.72.944.949.970.941.932.968.960.967.980.777.776.829.931.928.958.924.942.958
", + "image_path": "eaebda4b6eb4a1cd848ed3da69fe5665815ae2ae8bcbdca8b7497c8b0017fe21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 310, + 285, + 402 + ], + "blocks": [ + { + "bbox": [ + 69, + 293, + 523, + 303 + ], + "lines": [ + { + "bbox": [ + 69, + 293, + 523, + 303 + ], + "spans": [ + { + "bbox": [ + 69, + 293, + 523, + 303 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of our VSCode with other 5 SOTA RGB-D SOD methods on six benchmark datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 310, + 285, + 402 + ], + "lines": [ + { + "bbox": [ + 48, + 310, + 285, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 285, + 402 + ], + "type": "table", + "html": "
MethodParams (M)VT821[79]VT1000[78]VT5000[77]
SmFmEmSmFmEmSmFmEm
MIDD[76]52.43.871.847.916.916.904.956.868.834.919
TNet[11]87.41.899.885.936.929.921.965.895.864.936
CGMDR[7]-.894.872.932.931.927.966.896.877.939
VST-T++ [50]100.51.894.861.923.941.931.972.895.854.933
CAVER[68]55.79.891.874.933.936.927.970.892.857.935
VSCode-T54.09.921.906.951.949.944.981.918.892.954
VSCode-S74.72.926.910.954.952.947.981.925.900.959
", + "image_path": "a067c8d6d5169fa4f4b0555425b7052941a9d0e80f1ed94f3d2e2ee0a3944155.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 287, + 424 + ], + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 424 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of our VSCode with other 5 SOTA RGB-T SOD methods on three benchmark datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "text", + "content": "modal images within a shared model might prevent further optimization on those well-learned tasks. Based on this, we introduce domain-specific prompts " + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "inline_equation", + "content": "p^d" + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "text", + "content": ", resulting in substantial improvements across all datasets, which demonstrates the efficacy of domain-specific prompts in consolidating peculiarities within their respective domains. Subsequently, we introduce task-specific prompts " + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "inline_equation", + "content": "p^t" + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "text", + "content": " in the encoder-decoder architecture, enabling the capability to handle COD tasks. This brings slightly improved performance on some SOD tasks, however, significantly improves the performance on all COD tasks compared with the ST baseline, which probably owes to the well-learned commonalities from different tasks. Moreover, the incorporation of the prompt discrimination loss " + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 46, + 433, + 287, + 612 + ], + "type": "text", + "content": " leads to improved performance on most tasks, reaffirming its effectiveness in disentangling peculiarities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "type": "text", + "content": "To further evaluate the effectiveness of the task-specific prompts in the encoder and decoder, we remove them individually, resulting in performance decrease. This indicates that using task prompts in both encoder and decoder is necessary. We also observe our 2D prompts only bring around 0.03M parameters, which makes our model much more parameter-efficient than the traditional special training scheme." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 316, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 472 + ], + "type": "text", + "content": "Prompt Location. Following VPT [31], we design other forms of prompt layout based on Section 3.4. Table 2 reveals that employing the shallow version of task-specific prompts in the decoder, the deep version of domain-specific prompts and task-specific prompts in the encoder yields the best results. One plausible rationale is that each block aggregates distinct-level features within the encoder, thus it is better to propose unique prompts for each block. In our decoder, we follow VST and used skip connection to fuse decoder features with encoder features, which have already utilized deep task prompts for distinction. Hence, using more task prompts in the decoder may not be essential, and the shallow version seems to be a more fitting choice." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 481, + 547, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 547, + 672 + ], + "type": "text", + "content": "Prompt Length. We perform experiments with varying lengths for three kinds of prompts. As shown in Table 2, for domain-specific prompts, using one prompt token at each block achieves better performance than using more tokens. This suggests that it's possible to effectively capture domain distinctions using only a small number of prompts, which matches the observed relatively large correlation within domain prompts in Figure 5. Regarding task-specific prompts within the encoder, a prompt layout of 1,1,5,10 tokens at four blocks is found to be optimal on COD tasks, highlighting the importance of high-level semantic features over low-level features in distinguishing between SOD and COD tasks. This observation matches Figure 5 as well in which the correlations of SC in deep blocks are smaller than those in shallow blocks. Regarding the number of task-specific prompts in the decoder, performance starts to decline when it exceeds" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "text", + "content": "trast to EVP, which uses task-specific training. Therefore, comparing the parameters of our model with EVP may not be completely fair owing to the differences in training strategies and backbone utilization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 702, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 702, + 288, + 712 + ], + "type": "text", + "content": " Please note that our model shares parameters across six tasks, in con" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17175" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 65, + 543, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 65, + 543, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 65, + 543, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 65, + 543, + 148 + ], + "type": "table", + "html": "
MethodParams (M)DAVIS [70]FBMS[65]ViSal[85]SegV2[41]DAVSOD-Easy[18]DAVSOD-Normal[18]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑Sm↑Fm↑Em↑
DCFNet[105]71.66.914.899.970.883.853.910.952.953.990.903.870.953.729.612.781.708.601.791
FSNet[28]102.30.922.909.972.875.867.918---.849.773.920.760.637.796.732.623.789
CoSTFormer[51]-.923.906.978------.874.813.943.779.667.819.730.614.777
UFO[75]55.92.918.906.978.858.868.911.926.917.969.888.850.951.747.626.799.711.605.773
VSCode-T54.09.930.913.970.891.880.923.952.954.989.943.937.984.792.696.831.738.631.797
VSCode-S74.72.936.922.973.905.902.939.955.957.991.946.937.984.800.710.835.758.666.815
", + "image_path": "0d68f81de213051044524c6ffe6a79f5fd72328ad35ef1c535fa054c2029f634.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 165, + 284, + 256 + ], + "blocks": [ + { + "bbox": [ + 81, + 148, + 511, + 159 + ], + "lines": [ + { + "bbox": [ + 81, + 148, + 511, + 159 + ], + "spans": [ + { + "bbox": [ + 81, + 148, + 511, + 159 + ], + "type": "text", + "content": "Table 6. Quantitative comparison of our VSCode with other 4 SOTA VSOD methods on six benchmark datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 165, + 284, + 256 + ], + "lines": [ + { + "bbox": [ + 49, + 165, + 284, + 256 + ], + "spans": [ + { + "bbox": [ + 49, + 165, + 284, + 256 + ], + "type": "table", + "html": "
MethodParams (M)COD10K[15]NC4K[61]CAMO[37]
SmFmEmSmFmEmSmFmEm
MGL[101]63.60.814.738.890---.776.741.842
UJSC[39]121.63.817.750.902.856.835.920.803.775.867
SegMar[32]56.21.833.755.907.841.827.907.816.803.884
FEDER[22]44.13.822.768.905.847.833.915.802.789.873
VSCode-T54.09.847.795.925.874.853.930.838.821.909
EVP[56]64.52.845.794.926.874.855.933.849.833.918
VSCode-S74.72.869.827.942.891.878.944.873.861.938
", + "image_path": "10c1e0f9a220444b78e1b9ebb0c4143b52787202b8ac15ac54990ff0150a0c0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 285, + 284, + 368 + ], + "blocks": [ + { + "bbox": [ + 48, + 258, + 286, + 278 + ], + "lines": [ + { + "bbox": [ + 48, + 258, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 286, + 278 + ], + "type": "text", + "content": "Table 7. Quantitative comparison of our VSCode with other 5 SOTA RGB COD methods on three benchmark datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 285, + 284, + 368 + ], + "lines": [ + { + "bbox": [ + 49, + 285, + 284, + 368 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 284, + 368 + ], + "type": "table", + "html": "
MethodParams (M)CAD [3]MoCA-Mask[10]
Sm↑Fm↑Em↑Sm↑Fm↑Em↑
PNS-Net[26]26.87.671.473.787.514.068.599
RCRNet[96]53.79.664.405.786.559.170.593
MG[98]-.608.378.673.500.138.514
SLT-Net[10]164.68.715.542.823.624.327.768
VSCode-T54.09.757.659.808.650.339.787
VSCode-S74.72.790.680.853.665.386.796
", + "image_path": "9edac1cadb11f0759f1ce79df6a3ad62811ee07b14a5be26da15d522a7a3da1f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 399, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 423 + ], + "type": "text", + "content": "10. This emphasizes that blindly increasing the number of prompts doesn't guarantee improved performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 426, + 272, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 272, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 272, + 437 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 441, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 669 + ], + "type": "text", + "content": "Due to space limitation, we only report the performance comparison of our methods against other most highly-performed state-of-the-art methods, including 4 specialist RGB SOD models [50, 54, 89, 116], 5 specialist RGB-D SOD models [38, 50, 54, 68, 103], 5 specialist RGB-T SOD models [7, 11, 50, 68, 76], 3 specialist VSOD models [28, 51, 105], 4 specialist RGB COD models [22, 32, 39, 101], and 4 specialist VCOD models [10, 26, 96, 98]. Two generalist models [56, 75] are also reported. To ensure a relatively fair comparison with EVP [56], which utilizes SegFormer-B4 [93] as their backbone (64.1M parameters), we switch our backbone to Swin-S [59] as it has a similar number of parameters (50M). As shown in Table 3, Table 4, Table 5, Table 6, Table 7, and Table 8, our VSCode significantly outperforms all specialist methods and two generalist models across all six tasks, underscoring the effectiveness of our specially designed 2D prompts and prompt discrimination loss. The supplementary material displays visual comparison results among the top-performing models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 228, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 228, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 228, + 686 + ], + "type": "text", + "content": "4.5. Analysis of Generalization Ability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Previous generalist research [56, 75] primarily concentrated on assessing the effectiveness of models in training tasks, ne" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 167, + 545, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 369, + 287, + 390 + ], + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 390 + ], + "type": "text", + "content": "Table 8. Quantitative comparison of our VSCode with other 4 SOTA VCOD methods on two benchmark datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 167, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 306, + 167, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 545, + 219 + ], + "type": "table", + "html": "
SummaryCOD10K[15]NC4K[61]CAMO[37]
MethodTaskSmFmEmSmFmEmSmFmEm
PopNet[92]RGB-D.851.802.919.861.843.919.808.792.874
VSCode-TZSRGB-D.882.849.945.902.894.950.885.885.945
VSCode-TRGB.847.795.925.874.853.930.838.821.909
", + "image_path": "3dfc0b6a5ad6f098dff2033d829a44d8b9e3176b2b531c6f1a772b5142dacaeb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 220, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 242 + ], + "type": "text", + "content": "Table 9. Comparison with the SOTA RGB-D COD method on three benchmark datasets. \"ZS\" indicates zero-shot." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 254, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 469 + ], + "type": "text", + "content": "glecting their capacity for generalizing to novel tasks. Therefore, we employ the RGB-D COD task, which is not used in training, to further investigate the zero-shot generalization capabilities of our model. Specifically, we utilize our well-trained model and combine depth and COD prompts to tackle the RGB-D COD task. As shown in Table 9, our VSCode model significantly outperforms the state-of-the-art specialist model PopNet [92], although ours works in a pure zero-shot way. This demonstrates the superior zero-shot generalization ability of our proposed method. We also present the results of our model using only RGB information, which yields considerably lower performance compared to zero-shot RGB-D results. This validates that our zero-shot performance is not reliant on the utilization of seen RGB COD information but on the effectiveness of consolidating domain- and task-specific knowledge, which allows for the straightforward combination of various domain- and task-specific prompts for unseen tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 483, + 378, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 378, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 378, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 499, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 547, + 607 + ], + "type": "text", + "content": "In this paper, we present VSCode, a novel generalist and parameter-efficient model that tackles general multimodal SOD and COD tasks. Concretely, we use a foundation model to assimilate commonalities and 2D prompts to learn domain and task peculiarities. Furthermore, a prompt discrimination loss is introduced to help effectively disentangle specific knowledge and learn better shared knowledge. Our experiments demonstrate the effectiveness of VSCode on six training tasks and one unseen task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 620, + 402, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 402, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 402, + 634 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 636, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 547, + 712 + ], + "type": "text", + "content": "This work was supported in part by the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08; the National Natural Science Foundation of China under Grants 62136007, U20B2065, 62036005, 62322605; the Key Research and Development Program of Jiangsu Province under Grant BE2021093; the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center Project under Grant 21KT008; and by the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17176" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 145, + 288, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 288, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 288, + 176 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 288, + 219 + ], + "type": "text", + "content": "[3] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449. Springer, 2016. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 288, + 252 + ], + "type": "text", + "content": "[4] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In ECCV, pages 0–0, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 252, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 288, + 283 + ], + "type": "text", + "content": "[5] Ali Borji, Ming-Ming Cheng, Qibin Hou, Huaizu Jiang, and Jia Li. Salient object detection: A survey. CVMJ, 5:117-150, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 288, + 337 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33:1877–1901, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 337, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 288, + 392 + ], + "type": "text", + "content": "[7] Gang Chen, Feng Shao, Xiongli Chai, Hangwei Chen, Qiuping Jiang, Xiangchao Meng, and Yo-Sung Ho. Cgmdrnet: Cross-guided modality difference reduction network for rgb-t salient object detection. IEEE TCSVT, 32(9):6308-6323, 2022. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 392, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 288, + 423 + ], + "type": "text", + "content": "[8] Shuhan Chen and Yun Fu. Progressively guided alternate refinement network for rgb-d salient object detection. In ECCV, pages 520-538, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 423, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 288, + 464 + ], + "type": "text", + "content": "[9] Zuyao Chen, Runmin Cong, Qianqian Xu, and Qingming Huang. Dpanet: Depth potentiality-aware gated attention network for rgb-d salient object detection. IEEE TIP, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 466, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 509 + ], + "type": "text", + "content": "[10] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "text", + "content": "[11] Runmin Cong, Kepu Zhang, Chen Zhang, Feng Zheng, Yao Zhao, Qingming Huang, and Sam Kwong. Does thermal really always matter for rgb-t salient object detection? IEEE TMM, 2022. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 552, + 288, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 552, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 53, + 552, + 288, + 594 + ], + "type": "text", + "content": "[12] Zijun Deng, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Jing Qin, Guoqiang Han, and Pheng-Ann Heng. R3net: Recurrent residual refinement network for saliency detection. In IJCAI, pages 684–690, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 595, + 288, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 595, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 53, + 595, + 288, + 627 + ], + "type": "text", + "content": "[13] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 627, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 627, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 627, + 288, + 669 + ], + "type": "text", + "content": "[14] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In IJCAI, pages 698-704, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 670, + 288, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 701 + ], + "type": "text", + "content": "[15] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 701, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 701, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 701, + 288, + 712 + ], + "type": "text", + "content": "[16] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. VI, 1(1):16, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 95, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 547, + 137 + ], + "type": "text", + "content": "[17] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2020. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 137, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 137, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 312, + 137, + 545, + 169 + ], + "type": "text", + "content": "[18] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In CVPR, pages 8554-8564, 2019. 2, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "type": "text", + "content": "[19] Deng-Ping Fan, Yingjie Zhai, Ali Borji, Jufeng Yang, and Ling Shao. Bbs-net: Rgb-d salient object detection with a bifurcated backbone strategy network. In ECCV, pages 275–292, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 212, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 212, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 545, + 255 + ], + "type": "text", + "content": "[20] Chaowei Fang, Haibin Tian, Dingwen Zhang, Qiang Zhang, Jungong Han, and Junwei Han. Densely nested top-down flows for salient object detection. SCIS, 65(8):182103, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 255, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 255, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 312, + 255, + 545, + 298 + ], + "type": "text", + "content": "[21] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In AAAI, volume 34, pages 10869–10876, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 341 + ], + "type": "text", + "content": "[22] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In CVPR, pages 22046-22055, 2023. 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 341, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 341, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 341, + 545, + 373 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 545, + 405 + ], + "type": "text", + "content": "[24] Q Hou, MM Cheng, X Hu, A Borji, Z Tu, and PHS Torr. Deeply supervised salient object detection with short connections. IEEE TPAMI, 41(4):815-828, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 405, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 545, + 436 + ], + "type": "text", + "content": "[25] Guyue Hu, Bin He, and Hanwang Zhang. Compositional prompting video-language models to understand procedure in instructional videos. MIR, 20(2):249-262, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 436, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 436, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 312, + 436, + 545, + 479 + ], + "type": "text", + "content": "[26] Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively normalized self-attention network for video polyp segmentation. In MICCAI, pages 142–152. Springer, 2021. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 545, + 521 + ], + "type": "text", + "content": "[27] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 20(1):92-108, 2023. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 521, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 545, + 555 + ], + "type": "text", + "content": "[28] Ge-Peng Ji, Deng-Ping Fan, Keren Fu, Zhe Wu, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In CVMJ, pages 155-175, 2023. 2, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 312, + 554, + 545, + 586 + ], + "type": "text", + "content": "[29] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate rgb-d salient object detection via collaborative learning. In ECCV, pages 52-69, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 586, + 545, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 586, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 586, + 545, + 628 + ], + "type": "text", + "content": "[30] Yuzhu Ji, Hajunj Zhang, Zequn Jie, Lin Ma, and QM Jonathan Wu. Casnet: A cross-attention siamese network for video salient object detection. IEEE TNNLS, 32(6):2676–2690, 2020. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 628, + 545, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 628, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 312, + 628, + 545, + 671 + ], + "type": "text", + "content": "[31] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 671, + 545, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 671, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 671, + 545, + 704 + ], + "type": "text", + "content": "[32] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17177" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 72, + 73, + 134, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 134, + 83 + ], + "type": "text", + "content": "4722, 2022. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 83, + 288, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 83, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 288, + 125 + ], + "type": "text", + "content": "[33] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In ICIP, pages 1115-1119, 2014. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "text", + "content": "[34] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 190 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 190, + 288, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 288, + 222 + ], + "type": "text", + "content": "[36] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In ACCV, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "type": "text", + "content": "[37] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 265, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 265, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 265, + 288, + 308 + ], + "type": "text", + "content": "[38] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, pages 630-647. Springer, 2022. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "type": "text", + "content": "[39] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In CVPR, pages 10071-10081, 2021. 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 403 + ], + "type": "text", + "content": "[40] Aixuan Li, Jing Zhang, Yunqiu Lv, Tong Zhang, Yiran Zhong, Mingyi He, and Yuchao Dai. Joint salient object detection and camouflaged object detection via uncertainty-aware learning. arXiv preprint arXiv:2307.04651, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "type": "text", + "content": "[41] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In ICCV, pages 2192-2199, 2013. 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[42] Gongyang Li, Zhi Liu, and Haibin Ling. Icnet: Information conversion network for rgb-d based salient object detection. IEEE TIP, 29:4873-4884, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 479, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 511 + ], + "type": "text", + "content": "[43] Gongyang Li, Zhi Liu, Linwei Ye, Yang Wang, and Haibin Ling. Cross-modal weighting network for rgb-d salient object detection. In ECCV, pages 665-681, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 511, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 288, + 543 + ], + "type": "text", + "content": "[44] Guanbin Li, Yuan Xie, Tianhao Wei, Keze Wang, and Liang Lin. Flow guided recurrent neural encoder for video salient object detection. In CVPR, pages 3243-3252, 2018. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 543, + 288, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 543, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 53, + 543, + 288, + 573 + ], + "type": "text", + "content": "[45] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 574, + 288, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 574, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 53, + 574, + 288, + 628 + ], + "type": "text", + "content": "[46] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In CVPR, pages 15485-15494, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 628, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 628, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 628, + 288, + 661 + ], + "type": "text", + "content": "[47] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 661, + 288, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 661, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 661, + 288, + 693 + ], + "type": "text", + "content": "[48] Nian Liu and Junwei Han. Dhsnet: Deep hierarchical saliency network for salient object detection. In CVPR, pages 678-686, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 693, + 288, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 693, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 53, + 693, + 288, + 703 + ], + "type": "text", + "content": "[49] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 703 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 331, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 547, + 95 + ], + "type": "text", + "content": "Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 95, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 547, + 126 + ], + "type": "text", + "content": "[50] Nian Liu, Ziyang Luo, Ni Zhang, and Junwei Han. Vst++: Efficient and stronger visual saliency transformer. arXiv preprint arXiv:2310.11725, 2023. 4, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 126, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 126, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 547, + 169 + ], + "type": "text", + "content": "[51] Nian Liu, Kepan Nan, Wangbo Zhao, Xiwen Yao, and Junwei Han. Learning complementary spatial-temporal transformer for video salient object detection. IEEE TNNLS, 2023. 2, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 170, + 547, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 170, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 312, + 170, + 547, + 201 + ], + "type": "text", + "content": "[52] Nian Liu, Ni Zhang, and Junwei Han. Learning selective self-mutual attention for rgb-d saliency detection. In CVPR, pages 13756-13765, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 201, + 547, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 201, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 312, + 201, + 547, + 233 + ], + "type": "text", + "content": "[53] Nian Liu, Ni Zhang, Ling Shao, and Junwei Han. Learning selective mutual attention and contrast for rgb-d saliency detection. IEEE TPAMI, 44(12):9026-9042, 2021. 2, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 233, + 547, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 233, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 312, + 233, + 547, + 265 + ], + "type": "text", + "content": "[54] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In ICCV, pages 4722-4732, October 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 265, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 265, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 547, + 297 + ], + "type": "text", + "content": "[55] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, pages 1871-1880, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 297, + 547, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 297, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 547, + 339 + ], + "type": "text", + "content": "[56] Weihuang Liu, Xi Shen, Chi-Man Pun, and Xiaodong Cun. Explicit visual prompting for universal foreground segmentations. arXiv preprint arXiv:2305.18476, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 340, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 340, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 340, + 547, + 371 + ], + "type": "text", + "content": "[57] Yi Liu, Dingwen Zhang, Nian Liu, Shoukun Xu, and Jungong Han. Disentangled capsule routing for fast part-object relational saliency. IEEE TIP, 31:6719-6732, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 372, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 547, + 403 + ], + "type": "text", + "content": "[58] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. TPAMI, 44(7):3688-3704, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 403, + 547, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 403, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 547, + 446 + ], + "type": "text", + "content": "[59] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 447, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 447, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 312, + 447, + 547, + 488 + ], + "type": "text", + "content": "[60] Zhengyi Liu, Song Shi, Quntao Duan, Wei Zhang, and Peng Zhao. Salient object detection for rgb-d image by single stream recurrent convolution neural network. *IJON*, 363:46–57, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 489, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 489, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 312, + 489, + 547, + 532 + ], + "type": "text", + "content": "[61] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 532, + 547, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 547, + 564 + ], + "type": "text", + "content": "[62] Vida Movahedi and James H Elder. Design and perceptual validation of performance measures for salient object segmentation. In CVPR, pages 49-56, 2010. 5, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 564, + 547, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 564, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 312, + 564, + 547, + 607 + ], + "type": "text", + "content": "[63] Xing Nie, Bolin Ni, Jianlong Chang, Gaomeng Meng, Chunlei Huo, Zhaoxiang Zhang, Shiming Xiang, Qi Tian, and Chunhong Pan. Pro-tuning: Unified prompt tuning for vision tasks. arXiv preprint arXiv:2207.14381, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 607, + 547, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 607, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 312, + 607, + 547, + 638 + ], + "type": "text", + "content": "[64] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In CVPR, pages 454-461, 2012. 5, 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 638, + 547, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 638, + 547, + 670 + ], + "spans": [ + { + "bbox": [ + 312, + 638, + 547, + 670 + ], + "type": "text", + "content": "[65] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE TPAMI, 36(6):1187-1200, 2013. 5, 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 312, + 670, + 547, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 547, + 703 + ], + "type": "text", + "content": "[66] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17178" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 72, + 73, + 157, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 157, + 83 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 157, + 83 + ], + "type": "text", + "content": "2160-2170, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 83, + 288, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 83, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 288, + 115 + ], + "type": "text", + "content": "[67] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "type": "text", + "content": "[68] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Hutchuan Lu. Caver: Cross-modal view-mixed transformer for bi-modal salient object detection. IEEE TIP, 32:892-904, 2023. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "text", + "content": "[69] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In ECCV, pages 92-109, 2014. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 179, + 288, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 222 + ], + "type": "text", + "content": "[70] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 288, + 263 + ], + "type": "text", + "content": "[71] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 1, 2, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 264, + 288, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 307 + ], + "type": "text", + "content": "[72] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 350 + ], + "type": "text", + "content": "[73] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In ECCV, pages 212-228. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "text", + "content": "[74] Hongmei Song, Wenguan Wang, Sanyuan Zhao, Jianbing Shen, and Kin-Man Lam. Pyramid dilated deeper convlstm for video salient object detection. In ECCV, pages 715-731, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "type": "text", + "content": "[75] Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, Hanjing Su, and Qingyao Wu. A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE TMM, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[76] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE TIP, 30:5678-5691, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "type": "text", + "content": "[77] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A large-scale dataset and benchmark. IEEE TMM, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 521, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 564 + ], + "type": "text", + "content": "[78] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE TMM, 22(1):160-173, 2019. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 565, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 565, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 53, + 565, + 288, + 608 + ], + "type": "text", + "content": "[79] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. Rgb-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In IJIG, pages 359-369. Springer, 2018. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 608, + 288, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 608, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 53, + 608, + 288, + 651 + ], + "type": "text", + "content": "[80] Jie Wang, Kechen Song, Yanqi Bao, Liming Huang, and Yunhui Yan. Cgfnet: Cross-guided fusion network for rgb-t salient object detection. IEEE TCSVT, 32(5):2949-2961, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 651, + 288, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 651, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 651, + 288, + 693 + ], + "type": "text", + "content": "[81] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 693, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 693, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 53, + 693, + 288, + 704 + ], + "type": "text", + "content": "[82] Linzhao Wang, Lijun Wang, Huchuan Lu, Pingping Zhang," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "text", + "content": "and Xiang Ruan. Salient object detection with recurrent fully convolutional networks. IEEE TPAMI, 41(7):1734-1746, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 106, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 547, + 148 + ], + "type": "text", + "content": "[83] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 148, + 547, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 547, + 180 + ], + "type": "text", + "content": "[84] Wenguan Wang, Jianbing Shen, Xingping Dong, and Ali Borji. Salient object detection driven by fixation prediction. In CVPR, pages 1711-1720, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 180, + 547, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 180, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 180, + 547, + 213 + ], + "type": "text", + "content": "[85] Wenguan Wang, Jianbing Shen, and Ling Shao. Consistent video saliency using local gradient flow optimization and global refinement. IEEE TIP, 24(11):4185-4196, 2015. 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 213, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 547, + 243 + ], + "type": "text", + "content": "[86] Wenguan Wang, Jianbing Shen, and Ling Shao. Video salient object detection via fully convolutional networks. IEEE TIP, 27(1):38-49, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 244, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 244, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 244, + 547, + 285 + ], + "type": "text", + "content": "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, pages 6830-6839, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 286, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 547, + 329 + ], + "type": "text", + "content": "[88] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 329, + 547, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 329, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 312, + 329, + 547, + 372 + ], + "type": "text", + "content": "[89] Yi Wang, Ruili Wang, Xin Fan, Tianzhu Wang, and Xi-angjian He. Pixels, regions, and objects: Multiple enhancement for salient object detection. In CVPR, pages 10031–10040, 2023. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 372, + 547, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 547, + 405 + ], + "type": "text", + "content": "[90] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In CVPR, pages 13025-13034, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 405, + 547, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 547, + 458 + ], + "type": "text", + "content": "[91] Lina Wei, Shanshan Zhao, Omar Farouk Bourahla, Xi Li, Fei Wu, Yueting Zhuang, Junwei Han, and Mingliang Xu. End-to-end video saliency detection via a deep contextual spatiotemporal network. IEEE TNNLS, 32(4):1691-1702, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 458, + 547, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 458, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 312, + 458, + 547, + 501 + ], + "type": "text", + "content": "[92] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In ICCV, pages 1032-1042, 2023. 1, 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 501, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 501, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 501, + 547, + 544 + ], + "type": "text", + "content": "[93] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 544, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 547, + 587 + ], + "type": "text", + "content": "[94] Jin Xie, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Yanwei Pang, Ling Shao, and Mubarak Shah. Count-and similarity-aware r-cnn for pedestrian detection. In ECCV, pages 88-104. Springer, 2020. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 587, + 547, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 587, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 587, + 547, + 629 + ], + "type": "text", + "content": "[95] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, pages 15325-15336, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 629, + 547, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 547, + 672 + ], + "type": "text", + "content": "[96] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, pages 7284-7293, 2019. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 672, + 547, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 672, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 312, + 672, + 547, + 693 + ], + "type": "text", + "content": "[97] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 693, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 693, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 693, + 547, + 704 + ], + "type": "text", + "content": "[98] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman," + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17179" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "and Weidi Xie. Self-supervised video object segmentation by motion grouping. In ICCV, pages 7177-7188, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 288, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 288, + 126 + ], + "type": "text", + "content": "[99] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173, 2013. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 126, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 126, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 49, + 126, + 288, + 158 + ], + "type": "text", + "content": "[100] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 288, + 201 + ], + "type": "text", + "content": "[101] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "type": "text", + "content": "[102] Dingwen Zhang, Junwei Han, Yu Zhang, and Dong Xu. Synthesizing supervision for learning deep saliency network without human annotation. IEEE TPAMI, 42(7):1755-1769, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "type": "text", + "content": "[103] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. Rgb-d saliency detection via cascaded mutual information minimization. In ICCV, 2021. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 286, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 288, + 328 + ], + "type": "text", + "content": "[104] Lu Zhang, Jianming Zhang, Zhe Lin, Huchuan Lu, and You He. Capsal: Leveraging captioning to boost semantics for salient object detection. In CVPR, pages 6024–6033, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 329, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 329, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 329, + 288, + 373 + ], + "type": "text", + "content": "[105] Miao Zhang, Jie Liu, Yifei Wang, Yongri Piao, Shunyu Yao, Wei Ji, Jingjing Li, Huchuan Lu, and Zhongxuan Luo. Dynamic context-sensitive filtering network for video salient object detection. In ICCV, pages 1553-1563, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 373, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 373, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 373, + 288, + 404 + ], + "type": "text", + "content": "[106] Miao Zhang, Weisong Ren, Yongri Piao, Zhengkun Rong, and Huchuan Lu. Select, supplement and focus for rgb-d saliency detection. In CVPR, pages 3472-3481, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 404, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 288, + 446 + ], + "type": "text", + "content": "[107] Qiang Zhang, Nianchang Huang, Lin Yao, Dingwen Zhang, Caifeng Shan, and Jungong Han. Rgb-t salient object detection via fusing multi-level cnn features. IEEE TIP, 29:3321-3335, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 447, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 447, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 49, + 447, + 288, + 490 + ], + "type": "text", + "content": "[108] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 490, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 490, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 490, + 288, + 533 + ], + "type": "text", + "content": "[109] Jia-Xing Zhao, Yang Cao, Deng-Ping Fan, Ming-Ming Cheng, Xuan-Yi Li, and Le Zhang. Contrast prior and fluid pyramid integration for rgbd salient object detection. In CVPR, pages 3927-3936, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 533, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 288, + 575 + ], + "type": "text", + "content": "[110] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet:edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 575, + 288, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 575, + 288, + 618 + ], + "spans": [ + { + "bbox": [ + 49, + 575, + 288, + 618 + ], + "type": "text", + "content": "[111] Wangbo Zhao, Kepan Nan, Songyang Zhang, Kai Chen, Dahua Lin, and Yang You. Learning referring video object segmentation from weak annotation. arXiv preprint arXiv:2308.02162, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 618, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 618, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 49, + 618, + 288, + 661 + ], + "type": "text", + "content": "[112] Wangbo Zhao, Jiasheng Tang, Yizeng Han, Yibing Song, Kai Wang, Gao Huang, Fan Wang, and Yang You. Dynamic tuning towards parameter and inference efficiency for vit adaptation. arXiv preprint arXiv:2403.11808, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 661, + 288, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 661, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 49, + 661, + 288, + 693 + ], + "type": "text", + "content": "[113] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 693, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 693, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 49, + 693, + 288, + 704 + ], + "type": "text", + "content": "[114] Dehua Zheng, Xiaochen Zheng, Laurence T Yang, Yuan" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 213 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "text", + "content": "Gao, Chenlu Zhu, and Yiheng Ruan. Mffn: Multi-view feature fusion network for camouflaged object detection. In WACV, pages 6232-6242, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "text", + "content": "[115] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, pages 9516-9526, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 137, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 169 + ], + "type": "text", + "content": "[116] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 1, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 169, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 545, + 213 + ], + "type": "text", + "content": "[117] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, pages 15116-15127, 2023. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17180" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_content_list.json b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..142ea065640ae83177acde29dc133ede1950783a --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection", + "text_level": 1, + "bbox": [ + 197, + 130, + 772, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihua Liu* Tokyo Institute of Technology zliu@ok.sc.e.titech.ac.jp", + "bbox": [ + 129, + 203, + 369, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hiroki Sakuma* T2 Inc. sakuma.h@t2.auto", + "bbox": [ + 413, + 203, + 557, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masatoshi Okutomi Tokyo Institute of Technology mxo@sc.e.titech.ac.jp", + "bbox": [ + 599, + 203, + 839, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D object detection poses a significant challenge in 3D scene understanding due to its inherently ill-posed nature in monocular depth estimation. Existing methods heavily rely on supervised learning using abundant 3D labels, typically obtained through expensive and labor-intensive annotation on LiDAR point clouds. To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each instance as a signed distance field (SDF) and render its silhouette as an instance mask through our proposed instance-aware volumetric silhouette rendering. To directly optimize the 3D bounding boxes through rendering, we decompose the SDF of each instance into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This mechanism enables us to optimize the 3D bounding boxes in an end-to-end manner by comparing the rendered instance masks with the ground truth instance masks. The optimized 3D bounding boxes serve as effective training data for 3D object detection. We conduct extensive experiments on the KITTI-360 dataset, demonstrating that our method outperforms the existing weakly supervised 3D object detection methods. The code is available at https://github.com/skmhrk1209/VSRD.", + "bbox": [ + 75, + 323, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 790, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection is one of the most critical components in the perception system for autonomous driving. With the recent success of deep learning in computer vision, numerous 3D object detection methods have been proposed to", + "bbox": [ + 75, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a81a6ed70f7d663c5e24906782fde59b6f6c3aa7016ac521e90998dd2c9f2a4.jpg", + "image_caption": [ + "Figure 1. Illustration of our proposed weakly supervised 3D object detection framework, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 890, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "directly regress 3D bounding boxes from a LiDAR point cloud, multi-view images, or a monocular image. Among them, monocular 3D object detection is the most challenging in principle due to its inherently ill-posed nature in monocular depth estimation. Therefore, existing methods [22, 25, 28, 34, 35] heavily rely on supervised learning using abundant 3D labels manually annotated on LiDAR point clouds. This annotation cost is extremely high, posing a significant barrier to deploying 3D object detectors into autonomous driving systems.", + "bbox": [ + 496, + 489, + 893, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. As illustrated in Fig. 1, VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each 3D bounding box as a signed distance field (SDF) and render its silhouette as an instance mask through volumetric rendering. Comparing the rendered instance masks with the ground truth instance masks enables us to optimize the 3D bounding boxes directly. We introduce two novel mechanisms in this auto-labeling stage. The first is the instance-aware volumetric silhouette rendering that integrates instance labels along a ray to render the silhouette of each instance rather", + "bbox": [ + 496, + 643, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. The order was determined by a coin flip.", + "bbox": [ + 94, + 886, + 421, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17354", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "than the entire scene. This mechanism enables the silhouette of each instance to be rendered while considering geometric relationships among instances, such as occlusion. The second is the SDF decomposition, whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This decomposition models the spatial gap between the surfaces of each instance and the 3D bounding box, enabling more accurate silhouette rendering and providing more reliable feedback signals during optimization.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The 3D bounding boxes optimized by the proposed auto-labeling can serve as pseudo labels for training 3D object detectors. However, dynamic objects and inaccurate camera poses lead to low-quality pseudo labels, which negatively impact the training of 3D object detectors. Therefore, we propose a simple but effective algorithm to assign a confidence score representing the label quality to each pseudo label. We demonstrate that using these confidence scores as per-instance loss weights boosts the performance of the 3D object detectors trained on the pseudo labels.", + "bbox": [ + 75, + 243, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 395, + 434, + 409 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a weakly supervised 3D object detection framework consisting of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors.", + "- We propose a novel instance-aware volumetric silhouette rendering method whereby the silhouette of each instance can be rendered as an instance mask.", + "- We propose a novel SDF decomposition whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF).", + "- We propose a simple but effective confidence assignment algorithm to incorporate the quality of each pseudo label into the training of 3D object detectors.", + "- Extensive experiments on the KITTI-360 dataset demonstrate that our method outperforms the existing weakly supervised 3D object detection methods." + ], + "bbox": [ + 76, + 411, + 468, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 654, + 218, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 679, + 357, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D detection is a challenging task due to limited 3D information from monocular imagery. Deep3Box [19] pioneered this area by regressing relatively stable 3D object properties and combining these estimates with geometric constraints provided by the 2D bounding box. SMOKE [16] estimates 3D bounding boxes by combining keypoint estimates with regressed 3D box parameters. FCOS3D [28] employs a fully convolutional single-stage detector, transforming 7-DoF 3D targets to the image domain and decoupling them as 2D and 3D attributes. To enhance the performance of monocular 3D object detection for truncated objects, MonoFlex [35] explicitly decouples truncated objects and adaptively combines multiple approaches for ob", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ject depth estimation. M3D-RPN [4] and MonoDETR [34] also explored the usage of depth cues to improve monocular 3D object detection. The former designed depth-aware convolutional layers that enable location-specific feature extraction and consequently improved 3D scene understanding, while the latter modified the vanilla Transformer [7] to be depth-aware to guide the whole detection process by contextual depth cues. Despite these advances, the reliance on expensive and labor-intensive manual annotation on LiDAR point clouds remains a significant limitation.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Weakly Supervised 3D Object Detection", + "text_level": 1, + "bbox": [ + 498, + 252, + 844, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many weakly supervised methods have been proposed to mitigate the high annotation cost in 3D object detection. WS3D [17] introduced a LiDAR-based two-stage pipeline where cylindrical object proposals are first generated under weak supervision and then refined using a few labeled object instances. VS3D [24] introduced an unsupervised 3D proposal module that generates object proposals by leveraging normalized point cloud densities. WeakM3D [23] introduced a weakly supervised monocular 3D object detection method that leverages the 3D alignment loss between each predicted 3D bounding box and corresponding RoI LiDAR points. Furthermore, it introduced a method to estimate the orientation from RoI LiDAR points based on its statistics. Recently, WeakMono3D [26] leverages projection loss with multi-view and direction consistency, achieving a weakly supervised monocular object detection that relies only on 2D supervision. However, its reliance on 2D direction annotations restricts its applicability to large-scale datasets. Zakharov et al. [33] proposed an auto-labeling pipeline that integrates an SDF-based differentiable shape renderer and normalized object coordinate spaces (NOCS). However, additional training on synthetic data is still required for shape and coordinate estimation. In contrast, our method purely relies on 2D supervision, eliminating the necessity of synthetic data or 3D supervision.", + "bbox": [ + 496, + 276, + 893, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. 3D Object Detection with Neural Fields", + "text_level": 1, + "bbox": [ + 498, + 664, + 839, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [18] introduced a new perspective for implicit learning of 3D geometry from posed multi-view images by volume rendering. Building upon the vanilla NeRF, subsequent research has focused on enhancing novel view synthesis [1-3, 21] or accelerating volume rendering [6, 8, 20]. NeuS [27] introduced a novel volume rendering scheme to learn a neural SDF representation by introducing a density distribution induced by the SDF. Similarly, VolSDF [32] defined the volume density function as Laplace's cumulative distribution function applied to an SDF representation. Many works [12, 30, 31] recently have attempted to utilize neural fields for 3D object detection. Notably, NeRF-RPN [12] demonstrated that the 3D bounding boxes of objects in NeRF can be directly regressed with", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17355", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/779429e0437a8ab701f8bc8482c50fccecc89a93d4c4e9aea1bf27decc31d3d1.jpg", + "image_caption": [ + "Figure 2. Illustration of the pipeline of our proposed multi-view 3D auto-labeling. We represent the surface of each instance as an SDF and decompose it into the SDF of a 3D bounding box and the residual distance field (RDF), which is learned via a hypernetwork. The composed instance SDF is used to render the silhouette of the instance through our proposed instance-aware volumetric silhouette rendering. All the 3D bounding boxes are optimized based on the loss between the rendered and ground truth instance masks." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 890, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "out rendering. NeRF-Det [30] connects the detection and NeRF branches through a shared MLP, enabling an efficient adaptation of NeRF to detection and yielding geometry-aware volumetric representations. MonoNeRD [31] models scenes with SDFs and renders RGB images and depth maps through volume rendering to obtain intermediate 3D representations for detection. However, these approaches still rely on ground truth 3D labels for supervision. In contrast, we propose the first volume rendering-based weakly supervised 3D object detection framework that relies on multiview geometry and 2D supervision without any 3D supervision, such as 3D bounding boxes or LiDAR point clouds.", + "bbox": [ + 75, + 523, + 472, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 727, + 169, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-View 3D Auto-Labeling", + "text_level": 1, + "bbox": [ + 76, + 753, + 344, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our proposed multi-view 3D auto-labeling is illustrated in Fig. 2. In Sec. 3.1.1, we review the SDF-based volume rendering formulation, which is the foundation of our method. In Sec. 3.1.2, we define the optimization problem for the multi-view 3D auto-labeling. In Sec. 3.1.3, we introduce the SDF for cuboids, which is used to optimize 3D bounding boxes through rendering. In Sec. 3.1.4, we introduce a novel neural field named residual distance", + "bbox": [ + 75, + 779, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "field to fill the spatial gap between the surfaces of each instance and the 3D bounding box. In Sec. 3.1.5, we introduce instance-aware volumetric silhouette rendering to render the silhouette of each instance based on its SDF. In Sec. 3.1.6, we introduce loss functions using instance masks and 2D bounding boxes as weak supervision.", + "bbox": [ + 496, + 523, + 893, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 633, + 650, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SDF-based Volumetric Rendering NeRF [18] represents a 3D scene with neural density and color fields. Given a camera position $\\pmb{o} \\in \\mathbb{R}^3$ and a ray direction $\\pmb{d} \\in \\mathbb{R}^3$ emitted from a pixel, the volume rendering scheme integrates the colors of sampled points along the ray as follows:", + "bbox": [ + 496, + 657, + 893, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {C}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) \\mathrm {d} t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 743, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw (t) = \\exp (- \\int_ {0} ^ {t} \\sigma (\\boldsymbol {r} (u)) \\mathrm {d} u) \\sigma (\\boldsymbol {r} (t)), \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 777, + 890, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{C} (\\pmb {o},\\pmb {d})\\in \\mathbb{R}^3$ denotes the rendered color of the pixel, $\\boldsymbol {r}(t) = \\boldsymbol {o} + t\\boldsymbol{d}$ denotes the ray, $c(\\pmb {p},\\pmb {d})$ denotes the color at the position $\\pmb{p}$ and view direction $\\pmb{d}$ , and $\\sigma (\\pmb {p})$ denotes the volume density at the position $\\pmb{p}$ . Since the density field cannot represent the surfaces explicitly, novel SDF-", + "bbox": [ + 496, + 823, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17356", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based volume rendering formulation has been introduced in NeuS [27], where a surface represented by an SDF is reinterpreted as a participating medium represented by a density field, enabling the surface to be rendered through volume rendering. In this formulation, the weight $w(t)$ in Eq. (2) is re-written by introducing opaque density $\\rho(t)$ as follows:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw (t) = \\exp \\left(- \\int_ {0} ^ {t} \\rho (u) \\mathrm {d} u\\right) \\rho (t), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 208, + 468, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (t) = \\max \\left(- \\frac {\\frac {\\mathrm {d} \\Phi}{\\mathrm {d} t} \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}{\\Phi \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}, 0\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 244, + 468, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{F}} (\\cdot)$ denotes the SDF for the entire scene and $\\Phi (\\cdot)$ denotes the Sigmoid function. Our proposed instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5 is based on the same SDF-based weight formulation as Eq. (3) but integrates instance labels instead of colors along a ray to render instance masks.", + "bbox": [ + 75, + 297, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 417, + 264, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a monocular video consisting of posed frames, each frame annotated with instance masks, our goal is to optimize the 3D bounding box frame by frame without 3D supervision. More specifically, for each target frame $t$ in the video, we sample multiple source frames $\\mathcal{S}$ and optimize the $N$ 3D bounding boxes in the target frame using the instance masks of the source frames as weak supervision, where $N$ denotes the number of instances in the target frame. Please refer to the supplementary material for how to sample $\\mathcal{S}$ . We parameterize the $n$ -th 3D bounding box $\\hat{B}_n \\in \\mathbb{R}^{8 \\times 3}$ in the target frame with a dimension $\\hat{d}_n \\in \\mathbb{R}_+^3$ , location $\\hat{\\ell}_n \\in \\mathbb{R}^3$ , and orientation $\\hat{\\theta}_n \\in \\mathbb{R}$ , which is the rotation angle in the bird's-eye-view. In addition to these parameters for each bounding box, we prepare a learnable instance embedding $z_n \\in \\mathbb{R}^D$ for each instance and a shared hypernetwork parameterized by $\\psi$ for the residual distance field introduced in Sec. 3.1.4. We stack each parameter group into a single tensor over all the instances, yielding dimensions $\\hat{D} \\in \\mathbb{R}_+^{N \\times 3}$ , locations $\\hat{L} \\in \\mathbb{R}^{N \\times 3}$ , orientations $\\hat{\\Theta} \\in \\mathbb{R}^{N \\times 1}$ , and instance embeddings $Z \\in \\mathbb{R}^{N \\times D}$ . Given the loss function $\\mathcal{L}$ explained in Sec. 3.1.6, we optimize $\\hat{D}$ , $\\hat{L}$ , $\\hat{\\Theta}$ , $Z$ , and $\\psi$ via stochastic gradient descent as follows:", + "bbox": [ + 75, + 445, + 472, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { * } \\boldsymbol { D } , { } ^ { * } \\boldsymbol { L } , { } ^ { * } \\boldsymbol { \\Theta } , { } ^ { * } \\boldsymbol { Z } , { } ^ { * } \\boldsymbol { \\psi } = \\operatorname* { a r g m i n } _ { \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } } \\mathcal { L } ( \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } ) . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 811, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The optimized 3D bounding boxes ${}^{*}\\pmb{B}$ decoded from ${}^{*}\\pmb{D}$ , ${}^{*}\\pmb{L}$ , and ${}^{*}\\Theta$ can be used as pseudo labels for training 3D object detectors, as explained in Sec. 3.2.", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 3D Bounding Box Represented as an SDF", + "text_level": 1, + "bbox": [ + 500, + 90, + 843, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the 3D bounding boxes in each target frame by solving Eq. (5) through rendering, we represent the surface of each 3D bounding box as a signed distance field (SDF), which is one of the most common surface representations. An SDF is represented as a function $\\mathcal{F}:\\mathbb{R}^3\\to \\mathbb{R}$ that maps a spatial position $\\pmb {p}\\in \\mathbb{R}^3$ to its signed distance to the closest point on the surface, indicating that the zero-level set $\\{\\pmb {p}\\in \\mathbb{R}^3\\mid \\mathcal{F}(\\pmb {p}) = 0\\}$ represents the surface itself. The SDF for a cuboid parameterized by a dimension $d\\in \\mathbb{R}_+^3$ location $\\ell \\in \\mathbb{R}^3$ , and orientation $\\pmb {R}\\in \\mathrm{SO}(3)$ can be derived theoretically and is denoted by $\\mathcal{B}(\\cdot ;d,\\ell ,\\pmb {R})$ . Please refer to the supplementary material for how to derive this formula.", + "bbox": [ + 496, + 114, + 893, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.4 Residual Distance Field", + "text_level": 1, + "bbox": [ + 500, + 313, + 718, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, the shape of each instance is not a cuboid. Therefore, if we leverage only the cuboid SDF introduced in Sec. 3.1.3 to render the surface of each instance, it is no longer possible to render an accurate silhouette due to the spatial gap between the surfaces of the instance and the 3D bounding box, leading to unreliable feedback signals during optimization. Therefore, we propose a novel neural field named residual distance field (RDF), which models the residual between the signed distances to the surfaces of the instance and the 3D bounding box. Let $\\mathcal{F}_n(\\cdot)$ be the true SDF of the surface bounded by the $n$ -th 3D bounding box $\\hat{\\pmb{B}}_n$ whose SDF is given by $\\hat{\\mathcal{B}}_n(\\cdot) = \\mathcal{B}(\\cdot; \\hat{\\pmb{d}}_n, \\hat{\\ell}_n, \\pmb{R}_y(\\theta_n))$ , where $\\pmb{R}_y(\\theta)$ denotes the rotation matrix around the $y$ -axis by an angle $\\theta$ . For any point $\\pmb{p} \\in \\mathbb{R}^3$ , we define the RDF as $\\hat{\\mathcal{R}}_n(\\pmb{p}) := \\mathcal{F}_n(\\pmb{p}) - \\hat{\\mathcal{B}}_n(\\pmb{p})$ . Here, based on the definition of a 3D bounding box that it encloses the corresponding instance, $\\forall \\pmb{p} \\in \\mathbb{R}^3: \\hat{\\mathcal{R}}_n(\\pmb{p}) \\geq 0$ is required. One straightforward way to model the RDF for each instance is to train $N$ individual networks. However, as objects belonging to the same semantic class often have similar geometric shapes, we train a single hypernetwork [10] that regresses the weights of the neural RDF directly from an instance embedding instead. Given the $n$ -th instance embedding $\\pmb{z}_n \\in \\mathbb{R}^D$ , where $D$ denotes the number of dimensions, the $n$ -th neural RDF is given by:", + "bbox": [ + 496, + 337, + 893, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {R}} _ {n} (\\boldsymbol {p}) = \\sigma \\left(\\mathcal {G} (\\boldsymbol {p}; \\phi_ {n})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 720, + 890, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {n} = \\mathcal {H} \\left(\\boldsymbol {z} _ {n}; \\psi\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 741, + 890, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}(\\cdot ;\\phi_n)$ denotes the $n$ -th neural network parameterized by $\\phi_{n}$ $\\mathcal{H}(\\cdot ;\\psi)$ denotes the shared hypernetwork parameterized by $\\psi$ , and $\\sigma (\\cdot)$ denotes the Softplus function to force the residual distance to be positive based on the definition of a bounding box. For the instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5, we employ $\\hat{\\mathcal{F}}_n(\\pmb {p}) = \\hat{\\mathcal{B}}_n(\\pmb {p}) + \\hat{\\mathcal{R}}_n(\\pmb {p})$ as the SDF of each instance that represents an arbitrary surface bounded by the 3D bounding box $\\hat{B}_n$ for more accurate silhouette rendering.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17357", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/55548fd88ddc25c3267c6a6fcc528680246f8ded7476167037ed917c72946d02.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed instance-aware volumetric silhouette rendering. The instance labels are averaged for each sampled point along a ray based on the signed distance to each instance. The averaged instance labels are integrated along the ray based on the SDF-based volume rendering formulation [27]." + ], + "image_footnote": [], + "bbox": [ + 78, + 93, + 467, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.5 Instance-Aware Volumetric Silhouette Rendering", + "text_level": 1, + "bbox": [ + 76, + 349, + 465, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section details our approach to optimizing 3D bounding boxes through volumetric rendering. The core idea is to render instance masks and compare them with ground truth instance masks. To achieve this, we propose a novel SDF-based instance-aware volumetric silhouette rendering, where instance labels instead of colors are integrated along a ray based on the same SDF-based volume rendering formulation as Eq. (3) as follows:", + "bbox": [ + 75, + 373, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {S}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {s} (\\boldsymbol {r} (t)) d t, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 505, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} (\\boldsymbol {p}) = \\sum_ {n = 1} ^ {N} \\operatorname {s o f t m i n} \\left(\\left[ \\hat {\\mathcal {F}} _ {n} (\\boldsymbol {p}) \\right] _ {n = 1} ^ {N}\\right) _ {n} \\cdot \\boldsymbol {y} _ {n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 541, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\hat{S} (\\pmb {o},\\pmb {d})\\in [0,1]^N$ denotes the rendered soft instance label, $s(\\pmb {p})\\in [0,1]^N$ denotes the weighted average instance label at the position $\\pmb{p}$ indicating how relatively close the position $\\pmb{p}$ is to each instance, and $\\pmb {y}_n\\in \\{0,1\\} ^N$ denotes the one-hot instance label of the $n$ -th instance. $\\hat{\\mathcal{F}}_n(\\cdot)$ denotes the SDF of the $n$ -th instance introduced in Sec. 3.1.4. $w(t)$ in Eq. (8) is derived from Eq. (3). To compute Eq. (3), we model the entire scene as the union of all the object surfaces, i.e., $\\hat{\\mathcal{F}} (\\pmb {p}) = \\min (\\hat{\\mathcal{F}}_1(\\pmb {p}),\\dots ,\\hat{\\mathcal{F}}_N(\\pmb {p}))$ . This mechanism enables us to render instance masks considering the geometric relationships among instances, such as occlusions.", + "bbox": [ + 75, + 595, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.6 Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 785, + 235, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimize the parameters of the 3D bounding boxes $\\hat{\\Omega} = \\{\\hat{D},\\hat{L},\\hat{\\Theta}\\}$ in the target frame along with the instance embeddings $Z$ and parameter $\\psi$ of the hypernetwork $\\mathcal{H}(\\cdot ;\\psi)$ . The final loss $\\mathcal{L}$ is defined as the combination of the multi-view projection loss $\\mathcal{L}_{\\mathrm{proj}}$ , multi-view silhouette loss $\\mathcal{L}_{\\mathrm{sh}}$ , and Eikonal regularization $\\mathcal{L}_{\\mathrm{reg}}$ [27] as follows:", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\hat {\\Omega}, Z, \\psi) = \\lambda_ {\\mathrm {p r o j}} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\Omega}) + \\lambda_ {\\mathrm {s l h}} \\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\Omega}, Z, \\psi) \\\\ + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 113, + 890, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{proj}}$ , $\\lambda_{\\mathrm{slh}}$ , and $\\lambda_{\\mathrm{reg}}$ denote the loss weights. However, although the multi-view projection loss $\\mathcal{L}_{\\mathrm{proj}}$ and multi-view silhouette loss $\\mathcal{L}_{\\mathrm{slh}}$ are based on 2D supervision, the 3D-2D correspondence to define the losses is not obvious. Therefore, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm [14]. Each pair-wise matching cost is defined as the projection loss introduced in this section, but it is computed for not all the source frames but only the target frame. We assume that the ground truth 2D bounding boxes and instance masks have already been reordered based on the optimal permutation for simplicity.", + "bbox": [ + 498, + 162, + 890, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-View Projection Loss Although introducing the residual distance field enables us to model an arbitrary surface bounded by a 3D bounding box, it introduces another problem: the 3D bounding box can grow indefinitely without any constraint. To address this problem, we employ the ground truth 2D bounding box as a constraint to keep the 3D bounding box tight. The multi-view projection loss $\\mathcal{L}_{\\mathrm{proj}}$ is defined as the average distance between the projected and ground truth 2D bounding boxes as follows:", + "bbox": [ + 498, + 361, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\boldsymbol {\\Omega}}) = \\alpha \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\| \\hat {\\boldsymbol {B}} _ {i n} ^ {2 \\mathrm {D}} - \\boldsymbol {B} _ {i n} ^ {2 \\mathrm {D}} \\| _ {\\mathrm {H}} \\\\ - \\beta \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\mathrm {D I o U} \\left(\\hat {\\boldsymbol {B}} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}\\right), \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 505, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_{\\mathrm{H}}$ denotes the Huber loss, $\\mathrm{DIOU}(\\cdot ,\\cdot)$ denotes the Distance-IoU [36], and $\\alpha ,\\beta$ denote balancing coefficients. The projected 2D bounding box $\\hat{B}_{in}^{2\\mathrm{D}}$ is defined as the rectangle with minimal area enclosing the projected vertices $\\hat{\\pmb{V}}_{in}^{2\\mathrm{D}}\\propto \\hat{\\pmb{B}}_n\\pmb {E}_i^T\\pmb {K}_i^T$ where $\\pmb {E}_i$ and $\\pmb {K}_i$ denote the extrinsic and intrinsic matrices for frame $i$ , respectively.", + "bbox": [ + 498, + 599, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-View Silhouette Loss The multi-view silhouette loss is defined as the cross entropy between the rendered and ground truth instance masks. Since the spatial gap between the surfaces of each instance and the 3D bounding box is modeled by the residual distance field introduced in Sec. 3.1.4, the 3D bounding box constrained to tightly fit the ground truth 2D bounding box by the multi-view projection loss $\\mathcal{L}_{\\mathrm{proj}}$ is further refined by the multi-view silhouette loss $\\mathcal{L}_{\\mathrm{sh}}$ , which is given by:", + "bbox": [ + 498, + 709, + 890, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi) = \\sum_ {i \\in S} \\sum_ {j = 1} ^ {R _ {i}} \\operatorname {C E} \\left(\\hat {\\boldsymbol {S}} \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {d} _ {i j}\\right), \\boldsymbol {S} _ {i j}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 854, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17358", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14332f1df406c5da578b10d25d2c580561d7770afa185dadc569b2809c83d25b.jpg", + "image_caption": [ + "(a) Static Scene" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 271, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cea7b8334a9ec2e7bace0dafbaa265073467806aa144829af9b5184f9199cbfa.jpg", + "image_caption": [ + "(b) Dynamic Scene", + "Figure 4. Comparison of the confidence scores in static and dynamic scenes. It can be seen that the confidence scores are lower for dynamic, occluded, or truncated objects, indicating less influence on the subsequent training of 3D object detectors." + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 467, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\boldsymbol{o}_i \\in \\mathbb{R}^3$ , $\\boldsymbol{d}_{ij} \\in \\mathbb{R}^3$ , and $S_{ij} \\in \\{0,1\\}^N$ denote the camera position, ray direction, and ground truth instance label at the $j$ -th sampled pixel in frame $i$ , respectively. $\\hat{S}(\\boldsymbol{o}_i, \\boldsymbol{d}_{ij})$ denotes the rendered instance label based on Eq. (8). $\\mathrm{CE}(\\cdot, \\cdot)$ denotes the cross entropy loss. $R_i$ denotes the number of rays sampled for frame $i$ . Please refer to the supplementary material for how to determine $R_i$ .", + "bbox": [ + 75, + 263, + 467, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Training of 3D Object Detectors", + "text_level": 1, + "bbox": [ + 76, + 380, + 359, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the 3D bounding boxes are optimized by the proposed multi-view 3D auto-labeling, they can serve as pseudo labels for training 3D object detectors. To incorporate the quality of each pseudo label into the training of 3D object detectors, we introduce a simple but effective confidence assignment method in Sec. 3.2.1 and confidence-based weighted loss for bounding box regression in Sec. 3.2.2.", + "bbox": [ + 75, + 402, + 468, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.1 Confidence Assignment", + "text_level": 1, + "bbox": [ + 76, + 530, + 294, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the 3D bounding boxes optimized by the proposed auto-labeling are not reliable for dynamic, occluded, or truncated objects. Therefore, we propose a confidence assignment method based on the multi-view projection loss. First, for each target frame $t$ , we identify a set of source frames $\\mathcal{I}$ such that all the instances in the target frame are visible from every frame in the set. Next, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm. The cost matrix $Q \\in \\mathbb{R}^{N \\times N}$ is defined as the pairwise IoUs between the projected and ground truth 2D bounding boxes averaged over all the source frames, as follows:", + "bbox": [ + 75, + 553, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Q} _ {n m} = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} 1 - \\operatorname {I o U} \\left(^ {*} \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i m} ^ {\\mathrm {2 D}}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 760, + 468, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where ${}^{*}B_{in}^{2D}$ denotes the $n$ -th projected 2D bounding box in frame $i$ , which is obtained by projecting the $n$ -th optimized 3D bounding box ${}^{*}B_{n}$ onto frame $i$ , and ${B_{im}^{2D}}$ denotes the $m$ -th ground truth 2D bounding box in frame $i$ . Once an optimal permutation matrix $\\pmb{P} \\in \\{0,1\\}^{N \\times N}$ is obtained, the confidence scores ${}^{*}\\pmb{C} \\in [0,1]^N$ for the optimized 3D", + "bbox": [ + 75, + 808, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": " bounding boxes $^* B$ are given as follows:", + "bbox": [ + 500, + 90, + 771, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n^ {*} C = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} \\operatorname {I o U} \\left(^ {*} B _ {i} ^ {\\mathrm {2 D}}, P B _ {i} ^ {\\mathrm {2 D}}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 111, + 890, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2 Confidence-based Weighted Loss", + "text_level": 1, + "bbox": [ + 500, + 162, + 782, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The 3D bounding boxes optimized by the proposed multiview 3D auto-labeling serve as pseudo labels for most 3D object detectors, which leverage 3D bounding boxes and semantic class labels as supervision, without any modification of the architectures, loss functions, and training procedures, except for the confidence-based loss weighting. We incorporate the confidence scores computed by Eq. (14) into only the regression loss. Given a regression loss function $\\mathcal{L}_{\\mathrm{box}}$ its confidence-based weighted version $\\tilde{\\mathcal{L}}_{\\mathrm{box}}$ is given by:", + "bbox": [ + 496, + 186, + 890, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {L}} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}}, ^ {*} \\boldsymbol {B}, ^ {*} \\boldsymbol {C}\\right) = \\sum_ {m = 1} ^ {M} ^ {*} \\boldsymbol {C} _ {\\pi (m)} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}} _ {m}, ^ {*} \\boldsymbol {B} _ {\\pi (m)}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 329, + 890, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{B}$ denotes the predicted 3D bounding boxes, $^* B$ denotes the 3D bounding boxes optimized by the proposed auto-labeling, and $^* C$ denotes the corresponding confidence scores. $M$ denotes the number of positive anchors, and $\\pi (\\cdot)$ denotes the label assigner that maps the index of an anchor to that of the matched ground truth.", + "bbox": [ + 496, + 378, + 890, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 481, + 630, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 506, + 594, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the KITTI-360 [15] dataset for our experiments, splitting it into training (43,855 images), validation (1,173 images), and test sets (2,531 images). We follow the same evaluation protocol as the KITTI dataset [9]. However, as occlusion and truncation labels are not available for the KITTI-360 dataset, unlike the KITTI dataset, we consider only two difficulty levels, namely Easy and Hard, based on whether the height of each ground truth 2D bounding box is greater than 40 and 25 in pixels, respectively. Following prior works, we evaluate our method on only category Car.", + "bbox": [ + 496, + 527, + 890, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 686, + 715, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Multi-View 3D Auto-Labeling", + "text_level": 1, + "bbox": [ + 500, + 710, + 761, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We sample 16 source frames for each target frame. We sample 1000 rays across all the source frames, i.e., $\\sum_{i\\in S}R_i = 1000$ , at each iteration based on the ground truth instance masks. Please refer to the supplementary material for more details. We employ the same hierarchical volume sampling as NeRF [18] and sample 100 query points for both coarse and fine sampling. Unlike NeRF, both coarse and fine samples are drawn from the single scene SDF rather than two distinct ones. We set the number of dimensions of each instance embedding as $D = 256$ . The neural RDF $\\mathcal{G}$ and hypernetwork $\\mathcal{H}$ are implemented as MLPs with four hidden", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17359", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d44ab2bbdb4a32ea061c5929f3193b182db1dbc4db9dde4397929c37664caaf2.jpg", + "table_caption": [ + "Table 1. Ablation study on the KITTI-360 training set to verify the effectiveness of each component in our proposed multi-view 3D auto-labeling. Just one randomly selected sequence is used." + ], + "table_footnote": [], + "table_body": "
ComponentsAPBEV/AP3D@0.3APBEV/AP3D@0.5@0.5
LprojLslhRDFEasyHardEasyHard
✓60.77/54.8863.99/57.6637.38/23.3337.44/24.82
✓✓63.84/56.1160.86/56.8741.73/26.8439.22/25.58
✓✓✓73.84/66.6473.22/66.3246.35/31.1143.07/30.16
", + "bbox": [ + 81, + 142, + 468, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ddc646b7a2b13fac8eea7923281c0c3b2357ca160c6c2b9e2f34dfe1001ec31e.jpg", + "table_caption": [ + "Table 2. Ablation study on the KITTI-360 test set to verify the effectiveness of the confidence scores against monocular 3D object detection. S-WeakM3D is used as a monocular 3D object detector." + ], + "table_footnote": [], + "table_body": "
Conf.APBEV/AP3D@0.3APBEV/AP3D@0.5
EasyHardEasyHard
✓38.67/31.6231.25/23.8810.95/8.255.37/4.41
51.09/42.9441.27/33.9719.50/11.9114.39/8.46
", + "bbox": [ + 83, + 294, + 460, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "layers, each of which has 256 and 16 channels, respectively. We use the Adam optimizer [13] and the learning rates are decayed exponentially from $1e^{-2}$ , $1e^{-3}$ , and $1e^{-4}$ to $1e^{-4}$ , $1e^{-5}$ , and $1e^{-6}$ for the box parameters $\\hat{\\Omega}$ , instance embeddings $\\mathbf{Z}$ , and parameter $\\psi$ of the hypernetwork $\\mathcal{H}(\\cdot; \\psi)$ over 3000 iterations, respectively. For the loss weights, we set $\\alpha = 1.0$ , $\\beta = 0.1$ , $\\lambda_{\\mathrm{proj}} = 1.0$ , $\\lambda_{\\mathrm{slh}} = 1.0$ , and $\\lambda_{\\mathrm{reg}} = 0.01$ .", + "bbox": [ + 75, + 385, + 468, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 513, + 352, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare our method and Autolabels [33] with WeakM3D [23], we modify the architecture of WeakM3D so that it can be trained in a supervised manner using pseudo labels. More specifically, we train dimension and confidence heads in addition to the existing location and orientation heads using the same supervised loss as MonoDIS [25]. We call this model S-WeakM3D. For both WeakM3D and S-WeakM3D, ground truth 2D bounding boxes and instance masks are used for RoIAlign [11] and generating Li-DAR points on each object surface during training, respectively. During inference, we employ Cascade Mask R-CNN [5] with InternImage-XL [29] as the off-the-shelf 2D detector to provide 2D bounding boxes for RoIAlign. Please refer to the supplementary material for more details.", + "bbox": [ + 75, + 537, + 470, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 761, + 230, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Multi-View 3D Auto-Labeling", + "text_level": 1, + "bbox": [ + 76, + 784, + 339, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to demonstrate the effectiveness of each component in our proposed multi-view 3D auto-labeling. As can be seen from Tab. 1, the multi-view projection loss $\\mathcal{L}_{\\mathrm{proj}}$ serves as a practical baseline on its own. The multi-view silhouette loss $\\mathcal{L}_{\\mathrm{slh}}$ further boosts the quality of the pseudo labels a little bit, while the spatial gap", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a7253053f0383cadc7a90dadfc27599a495c7d713ad3310eccc1bff1b4252d5.jpg", + "image_caption": [ + "Figure 5. Visualization results of the optimized 3D bounding boxes (1st row) and rendered instance masks (2nd row). We assign a unique color to each instance, and each pixel is colored as the weighted summation based on the rendered soft instance label." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1532242ef4a982981301b465e11a391d504016c7014e4bd9aa1194064badcd34.jpg", + "table_caption": [ + "Table 3. Evaluation results of our proposed multi-view 3D auto-labeling on the KITTI-360 training set compared with the LiDAR-based monocular 3D auto-labeling proposed in Autolabels [33]. *Reproduced with the official code." + ], + "table_footnote": [], + "table_body": "
Conf.MethodAPBEV/AP3D@@0.3APBEV/AP3D@0.5
EasyHardEasyHard
≥0.0Autolabels*71.24/15.0967.33/11.4251.85/4.6546.10/2.92
VSRD75.03/68.5372.11/65.6447.12/35.2543.91/32.64
≥0.8Autolabels*75.56/17.1872.23/12.1958.64/5.7752.99/3.75
VSRD84.54/80.2581.66/77.3758.57/45.7655.09/44.17
", + "bbox": [ + 504, + 364, + 887, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "between the surfaces of each instance and the 3D bounding box limits further improvements. However, the residual distance field (RDF) can significantly boost the quality of the pseudo labels by addressing this problem. The systematic improvements with the inclusion of each component indicate their effectiveness, resulting in a precise auto-labeling system for weakly supervised 3D object detection.", + "bbox": [ + 496, + 481, + 890, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Confidence Assignment", + "text_level": 1, + "bbox": [ + 498, + 607, + 718, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the unreliability of the 3D bounding boxes optimized by the proposed auto-labeling for dynamic, occluded, or truncated objects, we conduct an ablation study to demonstrate the effectiveness of our proposed confidence assignment. Tab. 2 highlights substantial enhancement in detection performance when using the confidence-incorporated pseudo labels compared with the baseline, demonstrating the effectiveness of the proposed confidence assignment.", + "bbox": [ + 496, + 631, + 890, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Evaluation Results", + "text_level": 1, + "bbox": [ + 500, + 777, + 679, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Multi-View 3D Auto-Labeling", + "text_level": 1, + "bbox": [ + 500, + 800, + 761, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Autolabels [33] employs a similar two-stage framework consisting of auto-labeling and subsequent training of 3D object detectors using the pseudo labels. We compare our method with Autolabels to evaluate the quality of the pseudo labels. As our pseudo labels are unreliable for dy", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17360", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6c476d812d5937b2856fd75cad2d143244fecf4fd9a0eedcd2d2e0f42236f9c.jpg", + "table_caption": [ + "Table 4. Evaluation results of monocular 3D object detection on the KITTI-360 test set. *Reproduced with the official code. †CAD models are used as extra data. ${}^{ \\ddagger }M\\left( D\\right)$ indicates that detection model $D$ is employed for model-agnostic method $M$ ." + ], + "table_footnote": [], + "table_body": "
MethodWeak SupervisionFull SupervisionAPBEV/AP3D@0.3APBEV/AP3D@0.5
LiDARMasks3D BoxesEasyHardEasyHard
WeakM3D* [23]✓✓29.89/21.2524.01/15.348.10/2.962.96/2.01
Autolabels*†‡ [33] (S-WeakM3D)✓✓48.16/12.9237.34/9.9420.18/4.6914.33/2.79
VSRD‡ (S-WeakM3D)✓51.09/42.9441.28/33.7819.51/11.9114.39/8.46
VSRD‡ (MonoFlex)✓54.40/48.1645.67/40.0429.18/18.5322.31/13.60
VSRD‡ (MonoDETR)✓58.40/50.8650.61/43.4529.07/21.7722.83/16.46
MonoFlex* [35]✓69.70/67.0759.86/57.2650.82/43.1141.78/34.43
MonoDETR* [34]✓63.07/60.4954.04/50.0347.21/41.0136.05/30.38
", + "bbox": [ + 80, + 128, + 893, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5b083b3d8164479d1464f507946e5de2a174b31d744a9c44180819da41751c05.jpg", + "table_caption": [ + "Table 5. Evaluation results of semi-supervised monocular 3D object detection on the KITTI validation set." + ], + "table_footnote": [], + "table_body": "
MethodRatioAPBEV/AP3D@0.7
EasyModerateHard
MonoDETR [34]1.0037.99/29.3626.76/20.6423.02/17.30
0.000.002/0.0010.004/0.0010.005/0.002
0.2531.72/21.7622.32/15.4318.86/12.55
VSRD (MonoDETR)0.5043.44/31.0531.54/21.4827.17/17.93
0.7542.58/32.9531.08/24.6827.19/21.38
", + "bbox": [ + 81, + 333, + 465, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "namic, occluded, or truncated objects, we group the dataset based on whether the average confidence score for each image is greater than a certain threshold. The evaluation results with the confidence thresholds of 0.0 and 0.8 are shown in Tab. 3. Our method exhibits superior performance in terms of $\\mathrm{AP}_{\\mathrm{BEV}}@0.3$ and $\\mathrm{AP}_{\\mathrm{3D}}$ . For $\\mathrm{AP}_{\\mathrm{BEV}}@0.5$ , our method exhibits slightly lower performance for the confidence threshold of 0.0. However, by raising the confidence threshold to 0.8, our method demonstrates superior performance, indicating that our method can generate more high-quality pseudo labels while abandoning low-quality pseudo labels. The optimized 3D bounding boxes and rendered instance masks are visualized in Fig. 5.", + "bbox": [ + 75, + 463, + 468, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 681, + 354, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The pseudo labels generated by the proposed auto-labeling serve as 3D supervision. We further investigate their applicability using the existing monocular 3D object detectors.", + "bbox": [ + 75, + 707, + 468, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weakly Supervised Setting Tab. 4 shows the evaluation results of our method compared with the existing weakly supervised and fully supervised methods. Our method demonstrates a significant superiority over WeakM3D [23] across all the metrics while eliminating the need for LiDAR points for 3D supervision. Moreover, the detector trained on the pseudo labels generated by the proposed auto-labeling outperforms that trained on the pseudo labels generated by Autolabels [33]. Furthermore, employing more sophisticated", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "monocular 3D object detectors such as MonoFlex [35] and MonoDETR [34] further improves detection performance, demonstrating the broad versatility of our method, which is not limited to a specific detection model. It is noteworthy that the detectors trained on the pseudo labels generated by our method demonstrate competitive performance compared with those trained in a fully supervised manner.", + "bbox": [ + 496, + 297, + 890, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semi-Supervised Setting The essential advantage of our method is that it avoids costly 3D annotations, making more data available for training. Therefore, we investigate a realistic scenario where a detector pre-trained on a large amount of unlabeled data from a source domain is fine-tuned on a small amount of labeled data from a target domain. We select the KITTI-360 and KITTI datasets as source and target domains, respectively. Tab. 5 shows the performance of the detector pre-trained on the KITTI-360 dataset in a weakly supervised manner with the proposed auto-labeling and then fine-tuned on a subset of the KITTI dataset in a supervised manner. The zero-shot performance is quite low due to the characteristic that monocular depth estimation is greatly affected by the differences in camera parameters, but the performance of the detector fine-tuned on only $50\\%$ of the labeled data significantly outperforms that trained on the whole data from scratch, highlighting the broad applicability of our method.", + "bbox": [ + 496, + 409, + 892, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel weakly supervised 3D object detection framework named VSRD, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. Our method demonstrates superior performance compared with the existing weakly supervised 3D object detection methods. Moreover, it exhibits remarkable scalability using partially labeled data for semi-supervised learning. Our proposed method allows leveraging abundant 2D annotations to enhance 3D object detection without explicit 3D supervision, providing a promising avenue for further advancements in the field.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17361", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022.", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2", + "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2", + "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 7", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6", + "[10] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[12] Benran Hu, Junkai Huang, Yichen Liu, Yu-Wing Tai, and Chi-Keung Tang. Nerf-rpn: A general framework for object detection in nerfs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23528–23538, 2023. 2", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[15] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 6", + "[16] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 2", + "[17] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In European Conference on computer vision, pages 515-531. Springer, 2020. 2", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 6", + "[19] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7074-7082, 2017. 2", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[21] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2", + "[22] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 1", + "[23] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. arXiv preprint arXiv:2203.08332, 2022. 2, 7, 8", + "[24] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2", + "[25] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 1, 7", + "[26] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17362", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d object detection using multi-view projection and direction consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17492, 2023. 2", + "[27] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4, 5", + "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 1, 2", + "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14408-14419, 2023. 7", + "[30] Chenfeng Xu, Bichen Wu, Ji Hou, Sam Tsai, Ruilong Li, Jialiang Wang, Wei Zhan, Zijian He, Peter Vajda, Kurt Keutzer, et al. Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23320-23330, 2023. 2, 3", + "[31] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2, 3", + "[32] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2", + "[33] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 2, 7, 8", + "[34] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 1, 2, 8", + "[35] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 8", + "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, pages 12993-13000, 2020. 5" + ], + "bbox": [ + 78, + 90, + 470, + 809 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_model.json b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fde48ee901221a38e022538d42584b6a58ee9d42 --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.773, + 0.177 + ], + "angle": 0, + "content": "VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.204, + 0.37, + 0.258 + ], + "angle": 0, + "content": "Zihua Liu* Tokyo Institute of Technology zliu@ok.sc.e.titech.ac.jp" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.204, + 0.558, + 0.256 + ], + "angle": 0, + "content": "Hiroki Sakuma* T2 Inc. sakuma.h@t2.auto" + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.204, + 0.841, + 0.258 + ], + "angle": 0, + "content": "Masatoshi Okutomi Tokyo Institute of Technology mxo@sc.e.titech.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.765 + ], + "angle": 0, + "content": "Monocular 3D object detection poses a significant challenge in 3D scene understanding due to its inherently ill-posed nature in monocular depth estimation. Existing methods heavily rely on supervised learning using abundant 3D labels, typically obtained through expensive and labor-intensive annotation on LiDAR point clouds. To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each instance as a signed distance field (SDF) and render its silhouette as an instance mask through our proposed instance-aware volumetric silhouette rendering. To directly optimize the 3D bounding boxes through rendering, we decompose the SDF of each instance into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This mechanism enables us to optimize the 3D bounding boxes in an end-to-end manner by comparing the rendered instance masks with the ground truth instance masks. The optimized 3D bounding boxes serve as effective training data for 3D object detection. We conduct extensive experiments on the KITTI-360 dataset, demonstrating that our method outperforms the existing weakly supervised 3D object detection methods. The code is available at https://github.com/skmhrk1209/VSRD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.21, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "3D object detection is one of the most critical components in the perception system for autonomous driving. With the recent success of deep learning in computer vision, numerous 3D object detection methods have been proposed to" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.406, + 0.895, + 0.464 + ], + "angle": 0, + "content": "Figure 1. Illustration of our proposed weakly supervised 3D object detection framework, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.643 + ], + "angle": 0, + "content": "directly regress 3D bounding boxes from a LiDAR point cloud, multi-view images, or a monocular image. Among them, monocular 3D object detection is the most challenging in principle due to its inherently ill-posed nature in monocular depth estimation. Therefore, existing methods [22, 25, 28, 34, 35] heavily rely on supervised learning using abundant 3D labels manually annotated on LiDAR point clouds. This annotation cost is extremely high, posing a significant barrier to deploying 3D object detectors into autonomous driving systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.896, + 0.902 + ], + "angle": 0, + "content": "To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. As illustrated in Fig. 1, VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each 3D bounding box as a signed distance field (SDF) and render its silhouette as an instance mask through volumetric rendering. Comparing the rendered instance masks with the ground truth instance masks enables us to optimize the 3D bounding boxes directly. We introduce two novel mechanisms in this auto-labeling stage. The first is the instance-aware volumetric silhouette rendering that integrates instance labels along a ray to render the silhouette of each instance rather" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.422, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. The order was determined by a coin flip." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17354" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "than the entire scene. This mechanism enables the silhouette of each instance to be rendered while considering geometric relationships among instances, such as occlusion. The second is the SDF decomposition, whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This decomposition models the spatial gap between the surfaces of each instance and the 3D bounding box, enabling more accurate silhouette rendering and providing more reliable feedback signals during optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.395 + ], + "angle": 0, + "content": "The 3D bounding boxes optimized by the proposed auto-labeling can serve as pseudo labels for training 3D object detectors. However, dynamic objects and inaccurate camera poses lead to low-quality pseudo labels, which negatively impact the training of 3D object detectors. Therefore, we propose a simple but effective algorithm to assign a confidence score representing the label quality to each pseudo label. We demonstrate that using these confidence scores as per-instance loss weights boosts the performance of the 3D object detectors trained on the pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.396, + 0.435, + 0.41 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.412, + 0.47, + 0.457 + ], + "angle": 0, + "content": "- We propose a weakly supervised 3D object detection framework consisting of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.457, + 0.47, + 0.5 + ], + "angle": 0, + "content": "- We propose a novel instance-aware volumetric silhouette rendering method whereby the silhouette of each instance can be rendered as an instance mask." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.502, + 0.47, + 0.547 + ], + "angle": 0, + "content": "- We propose a novel SDF decomposition whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.548, + 0.47, + 0.592 + ], + "angle": 0, + "content": "- We propose a simple but effective confidence assignment algorithm to incorporate the quality of each pseudo label into the training of 3D object detectors." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.593, + 0.468, + 0.638 + ], + "angle": 0, + "content": "- Extensive experiments on the KITTI-360 dataset demonstrate that our method outperforms the existing weakly supervised 3D object detection methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.412, + 0.47, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.22, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.358, + 0.696 + ], + "angle": 0, + "content": "2.1. Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Monocular 3D detection is a challenging task due to limited 3D information from monocular imagery. Deep3Box [19] pioneered this area by regressing relatively stable 3D object properties and combining these estimates with geometric constraints provided by the 2D bounding box. SMOKE [16] estimates 3D bounding boxes by combining keypoint estimates with regressed 3D box parameters. FCOS3D [28] employs a fully convolutional single-stage detector, transforming 7-DoF 3D targets to the image domain and decoupling them as 2D and 3D attributes. To enhance the performance of monocular 3D object detection for truncated objects, MonoFlex [35] explicitly decouples truncated objects and adaptively combines multiple approaches for ob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "ject depth estimation. M3D-RPN [4] and MonoDETR [34] also explored the usage of depth cues to improve monocular 3D object detection. The former designed depth-aware convolutional layers that enable location-specific feature extraction and consequently improved 3D scene understanding, while the latter modified the vanilla Transformer [7] to be depth-aware to guide the whole detection process by contextual depth cues. Despite these advances, the reliance on expensive and labor-intensive manual annotation on LiDAR point clouds remains a significant limitation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.253, + 0.845, + 0.27 + ], + "angle": 0, + "content": "2.2. Weakly Supervised 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Many weakly supervised methods have been proposed to mitigate the high annotation cost in 3D object detection. WS3D [17] introduced a LiDAR-based two-stage pipeline where cylindrical object proposals are first generated under weak supervision and then refined using a few labeled object instances. VS3D [24] introduced an unsupervised 3D proposal module that generates object proposals by leveraging normalized point cloud densities. WeakM3D [23] introduced a weakly supervised monocular 3D object detection method that leverages the 3D alignment loss between each predicted 3D bounding box and corresponding RoI LiDAR points. Furthermore, it introduced a method to estimate the orientation from RoI LiDAR points based on its statistics. Recently, WeakMono3D [26] leverages projection loss with multi-view and direction consistency, achieving a weakly supervised monocular object detection that relies only on 2D supervision. However, its reliance on 2D direction annotations restricts its applicability to large-scale datasets. Zakharov et al. [33] proposed an auto-labeling pipeline that integrates an SDF-based differentiable shape renderer and normalized object coordinate spaces (NOCS). However, additional training on synthetic data is still required for shape and coordinate estimation. In contrast, our method purely relies on 2D supervision, eliminating the necessity of synthetic data or 3D supervision." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.84, + 0.682 + ], + "angle": 0, + "content": "2.3. 3D Object Detection with Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [18] introduced a new perspective for implicit learning of 3D geometry from posed multi-view images by volume rendering. Building upon the vanilla NeRF, subsequent research has focused on enhancing novel view synthesis [1-3, 21] or accelerating volume rendering [6, 8, 20]. NeuS [27] introduced a novel volume rendering scheme to learn a neural SDF representation by introducing a density distribution induced by the SDF. Similarly, VolSDF [32] defined the volume density function as Laplace's cumulative distribution function applied to an SDF representation. Many works [12, 30, 31] recently have attempted to utilize neural fields for 3D object detection. Notably, NeRF-RPN [12] demonstrated that the 3D bounding boxes of objects in NeRF can be directly regressed with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.891, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.895, + 0.501 + ], + "angle": 0, + "content": "Figure 2. Illustration of the pipeline of our proposed multi-view 3D auto-labeling. We represent the surface of each instance as an SDF and decompose it into the SDF of a 3D bounding box and the residual distance field (RDF), which is learned via a hypernetwork. The composed instance SDF is used to render the silhouette of the instance through our proposed instance-aware volumetric silhouette rendering. All the 3D bounding boxes are optimized based on the loss between the rendered and ground truth instance masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.473, + 0.707 + ], + "angle": 0, + "content": "out rendering. NeRF-Det [30] connects the detection and NeRF branches through a shared MLP, enabling an efficient adaptation of NeRF to detection and yielding geometry-aware volumetric representations. MonoNeRD [31] models scenes with SDFs and renders RGB images and depth maps through volume rendering to obtain intermediate 3D representations for detection. However, these approaches still rely on ground truth 3D labels for supervision. In contrast, we propose the first volume rendering-based weakly supervised 3D object detection framework that relies on multiview geometry and 2D supervision without any 3D supervision, such as 3D bounding boxes or LiDAR point clouds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.728, + 0.17, + 0.743 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.345, + 0.772 + ], + "angle": 0, + "content": "3.1. Multi-View 3D Auto-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The pipeline of our proposed multi-view 3D auto-labeling is illustrated in Fig. 2. In Sec. 3.1.1, we review the SDF-based volume rendering formulation, which is the foundation of our method. In Sec. 3.1.2, we define the optimization problem for the multi-view 3D auto-labeling. In Sec. 3.1.3, we introduce the SDF for cuboids, which is used to optimize 3D bounding boxes through rendering. In Sec. 3.1.4, we introduce a novel neural field named residual distance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.895, + 0.618 + ], + "angle": 0, + "content": "field to fill the spatial gap between the surfaces of each instance and the 3D bounding box. In Sec. 3.1.5, we introduce instance-aware volumetric silhouette rendering to render the silhouette of each instance based on its SDF. In Sec. 3.1.6, we introduce loss functions using instance masks and 2D bounding boxes as weak supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.651, + 0.648 + ], + "angle": 0, + "content": "3.1.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.894, + 0.735 + ], + "angle": 0, + "content": "SDF-based Volumetric Rendering NeRF [18] represents a 3D scene with neural density and color fields. Given a camera position \\( \\pmb{o} \\in \\mathbb{R}^3 \\) and a ray direction \\( \\pmb{d} \\in \\mathbb{R}^3 \\) emitted from a pixel, the volume rendering scheme integrates the colors of sampled points along the ray as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.744, + 0.892, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {C}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) \\mathrm {d} t, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.779, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\nw (t) = \\exp (- \\int_ {0} ^ {t} \\sigma (\\boldsymbol {r} (u)) \\mathrm {d} u) \\sigma (\\boldsymbol {r} (t)), \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{C} (\\pmb {o},\\pmb {d})\\in \\mathbb{R}^3\\) denotes the rendered color of the pixel, \\(\\boldsymbol {r}(t) = \\boldsymbol {o} + t\\boldsymbol{d}\\) denotes the ray, \\(c(\\pmb {p},\\pmb {d})\\) denotes the color at the position \\(\\pmb{p}\\) and view direction \\(\\pmb{d}\\), and \\(\\sigma (\\pmb {p})\\) denotes the volume density at the position \\(\\pmb{p}\\). Since the density field cannot represent the surfaces explicitly, novel SDF-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "based volume rendering formulation has been introduced in NeuS [27], where a surface represented by an SDF is reinterpreted as a participating medium represented by a density field, enabling the surface to be rendered through volume rendering. In this formulation, the weight \\( w(t) \\) in Eq. (2) is re-written by introducing opaque density \\( \\rho(t) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.209, + 0.47, + 0.244 + ], + "angle": 0, + "content": "\\[\nw (t) = \\exp \\left(- \\int_ {0} ^ {t} \\rho (u) \\mathrm {d} u\\right) \\rho (t), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.246, + 0.469, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\rho (t) = \\max \\left(- \\frac {\\frac {\\mathrm {d} \\Phi}{\\mathrm {d} t} \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}{\\Phi \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}, 0\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.47, + 0.39 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{F}} (\\cdot)\\) denotes the SDF for the entire scene and \\(\\Phi (\\cdot)\\) denotes the Sigmoid function. Our proposed instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5 is based on the same SDF-based weight formulation as Eq. (3) but integrates instance labels instead of colors along a ray to render instance masks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.419, + 0.265, + 0.433 + ], + "angle": 0, + "content": "3.1.2 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.784 + ], + "angle": 0, + "content": "Given a monocular video consisting of posed frames, each frame annotated with instance masks, our goal is to optimize the 3D bounding box frame by frame without 3D supervision. More specifically, for each target frame \\(t\\) in the video, we sample multiple source frames \\(\\mathcal{S}\\) and optimize the \\(N\\) 3D bounding boxes in the target frame using the instance masks of the source frames as weak supervision, where \\(N\\) denotes the number of instances in the target frame. Please refer to the supplementary material for how to sample \\(\\mathcal{S}\\). We parameterize the \\(n\\)-th 3D bounding box \\(\\hat{B}_n \\in \\mathbb{R}^{8 \\times 3}\\) in the target frame with a dimension \\(\\hat{d}_n \\in \\mathbb{R}_+^3\\), location \\(\\hat{\\ell}_n \\in \\mathbb{R}^3\\), and orientation \\(\\hat{\\theta}_n \\in \\mathbb{R}\\), which is the rotation angle in the bird's-eye-view. In addition to these parameters for each bounding box, we prepare a learnable instance embedding \\(z_n \\in \\mathbb{R}^D\\) for each instance and a shared hypernetwork parameterized by \\(\\psi\\) for the residual distance field introduced in Sec. 3.1.4. We stack each parameter group into a single tensor over all the instances, yielding dimensions \\(\\hat{D} \\in \\mathbb{R}_+^{N \\times 3}\\), locations \\(\\hat{L} \\in \\mathbb{R}^{N \\times 3}\\), orientations \\(\\hat{\\Theta} \\in \\mathbb{R}^{N \\times 1}\\), and instance embeddings \\(Z \\in \\mathbb{R}^{N \\times D}\\). Given the loss function \\(\\mathcal{L}\\) explained in Sec. 3.1.6, we optimize \\(\\hat{D}\\), \\(\\hat{L}\\), \\(\\hat{\\Theta}\\), \\(Z\\), and \\(\\psi\\) via stochastic gradient descent as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.812, + 0.47, + 0.842 + ], + "angle": 0, + "content": "\\[\n{ } ^ { * } \\boldsymbol { D } , { } ^ { * } \\boldsymbol { L } , { } ^ { * } \\boldsymbol { \\Theta } , { } ^ { * } \\boldsymbol { Z } , { } ^ { * } \\boldsymbol { \\psi } = \\operatorname* { a r g m i n } _ { \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } } \\mathcal { L } ( \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } ) . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The optimized 3D bounding boxes \\({}^{*}\\pmb{B}\\) decoded from \\({}^{*}\\pmb{D}\\), \\({}^{*}\\pmb{L}\\), and \\({}^{*}\\Theta\\) can be used as pseudo labels for training 3D object detectors, as explained in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.844, + 0.107 + ], + "angle": 0, + "content": "3.1.3 3D Bounding Box Represented as an SDF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.298 + ], + "angle": 0, + "content": "To optimize the 3D bounding boxes in each target frame by solving Eq. (5) through rendering, we represent the surface of each 3D bounding box as a signed distance field (SDF), which is one of the most common surface representations. An SDF is represented as a function \\(\\mathcal{F}:\\mathbb{R}^3\\to \\mathbb{R}\\) that maps a spatial position \\(\\pmb {p}\\in \\mathbb{R}^3\\) to its signed distance to the closest point on the surface, indicating that the zero-level set \\(\\{\\pmb {p}\\in \\mathbb{R}^3\\mid \\mathcal{F}(\\pmb {p}) = 0\\}\\) represents the surface itself. The SDF for a cuboid parameterized by a dimension \\(d\\in \\mathbb{R}_+^3\\) location \\(\\ell \\in \\mathbb{R}^3\\), and orientation \\(\\pmb {R}\\in \\mathrm{SO}(3)\\) can be derived theoretically and is denoted by \\(\\mathcal{B}(\\cdot ;d,\\ell ,\\pmb {R})\\). Please refer to the supplementary material for how to derive this formula." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.719, + 0.328 + ], + "angle": 0, + "content": "3.1.4 Residual Distance Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.716 + ], + "angle": 0, + "content": "In general, the shape of each instance is not a cuboid. Therefore, if we leverage only the cuboid SDF introduced in Sec. 3.1.3 to render the surface of each instance, it is no longer possible to render an accurate silhouette due to the spatial gap between the surfaces of the instance and the 3D bounding box, leading to unreliable feedback signals during optimization. Therefore, we propose a novel neural field named residual distance field (RDF), which models the residual between the signed distances to the surfaces of the instance and the 3D bounding box. Let \\(\\mathcal{F}_n(\\cdot)\\) be the true SDF of the surface bounded by the \\(n\\)-th 3D bounding box \\(\\hat{\\pmb{B}}_n\\) whose SDF is given by \\(\\hat{\\mathcal{B}}_n(\\cdot) = \\mathcal{B}(\\cdot; \\hat{\\pmb{d}}_n, \\hat{\\ell}_n, \\pmb{R}_y(\\theta_n))\\), where \\(\\pmb{R}_y(\\theta)\\) denotes the rotation matrix around the \\(y\\)-axis by an angle \\(\\theta\\). For any point \\(\\pmb{p} \\in \\mathbb{R}^3\\), we define the RDF as \\(\\hat{\\mathcal{R}}_n(\\pmb{p}) := \\mathcal{F}_n(\\pmb{p}) - \\hat{\\mathcal{B}}_n(\\pmb{p})\\). Here, based on the definition of a 3D bounding box that it encloses the corresponding instance, \\(\\forall \\pmb{p} \\in \\mathbb{R}^3: \\hat{\\mathcal{R}}_n(\\pmb{p}) \\geq 0\\) is required. One straightforward way to model the RDF for each instance is to train \\(N\\) individual networks. However, as objects belonging to the same semantic class often have similar geometric shapes, we train a single hypernetwork [10] that regresses the weights of the neural RDF directly from an instance embedding instead. Given the \\(n\\)-th instance embedding \\(\\pmb{z}_n \\in \\mathbb{R}^D\\), where \\(D\\) denotes the number of dimensions, the \\(n\\)-th neural RDF is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.721, + 0.891, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {R}} _ {n} (\\boldsymbol {p}) = \\sigma \\left(\\mathcal {G} (\\boldsymbol {p}; \\phi_ {n})\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.742, + 0.892, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\phi_ {n} = \\mathcal {H} \\left(\\boldsymbol {z} _ {n}; \\psi\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}(\\cdot ;\\phi_n)\\) denotes the \\(n\\) -th neural network parameterized by \\(\\phi_{n}\\) \\(\\mathcal{H}(\\cdot ;\\psi)\\) denotes the shared hypernetwork parameterized by \\(\\psi\\) , and \\(\\sigma (\\cdot)\\) denotes the Softplus function to force the residual distance to be positive based on the definition of a bounding box. For the instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5, we employ \\(\\hat{\\mathcal{F}}_n(\\pmb {p}) = \\hat{\\mathcal{B}}_n(\\pmb {p}) + \\hat{\\mathcal{R}}_n(\\pmb {p})\\) as the SDF of each instance that represents an arbitrary surface bounded by the 3D bounding box \\(\\hat{B}_n\\) for more accurate silhouette rendering." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.094, + 0.468, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.471, + 0.322 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed instance-aware volumetric silhouette rendering. The instance labels are averaged for each sampled point along a ray based on the signed distance to each instance. The averaged instance labels are integrated along the ray based on the SDF-based volume rendering formulation [27]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.467, + 0.365 + ], + "angle": 0, + "content": "3.1.5 Instance-Aware Volumetric Silhouette Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.47, + 0.495 + ], + "angle": 0, + "content": "This section details our approach to optimizing 3D bounding boxes through volumetric rendering. The core idea is to render instance masks and compare them with ground truth instance masks. To achieve this, we propose a novel SDF-based instance-aware volumetric silhouette rendering, where instance labels instead of colors are integrated along a ray based on the same SDF-based volume rendering formulation as Eq. (3) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.506, + 0.47, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {S}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {s} (\\boldsymbol {r} (t)) d t, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.542, + 0.469, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} (\\boldsymbol {p}) = \\sum_ {n = 1} ^ {N} \\operatorname {s o f t m i n} \\left(\\left[ \\hat {\\mathcal {F}} _ {n} (\\boldsymbol {p}) \\right] _ {n = 1} ^ {N}\\right) _ {n} \\cdot \\boldsymbol {y} _ {n}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.765 + ], + "angle": 0, + "content": "where, \\(\\hat{S} (\\pmb {o},\\pmb {d})\\in [0,1]^N\\) denotes the rendered soft instance label, \\(s(\\pmb {p})\\in [0,1]^N\\) denotes the weighted average instance label at the position \\(\\pmb{p}\\) indicating how relatively close the position \\(\\pmb{p}\\) is to each instance, and \\(\\pmb {y}_n\\in \\{0,1\\} ^N\\) denotes the one-hot instance label of the \\(n\\) -th instance. \\(\\hat{\\mathcal{F}}_n(\\cdot)\\) denotes the SDF of the \\(n\\) -th instance introduced in Sec. 3.1.4. \\(w(t)\\) in Eq. (8) is derived from Eq. (3). To compute Eq. (3), we model the entire scene as the union of all the object surfaces, i.e., \\(\\hat{\\mathcal{F}} (\\pmb {p}) = \\min (\\hat{\\mathcal{F}}_1(\\pmb {p}),\\dots ,\\hat{\\mathcal{F}}_N(\\pmb {p}))\\) . This mechanism enables us to render instance masks considering the geometric relationships among instances, such as occlusions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.236, + 0.8 + ], + "angle": 0, + "content": "3.1.6 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We optimize the parameters of the 3D bounding boxes \\(\\hat{\\Omega} = \\{\\hat{D},\\hat{L},\\hat{\\Theta}\\}\\) in the target frame along with the instance embeddings \\(Z\\) and parameter \\(\\psi\\) of the hypernetwork \\(\\mathcal{H}(\\cdot ;\\psi)\\). The final loss \\(\\mathcal{L}\\) is defined as the combination of the multi-view projection loss \\(\\mathcal{L}_{\\mathrm{proj}}\\), multi-view silhouette loss \\(\\mathcal{L}_{\\mathrm{sh}}\\), and Eikonal regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\) [27] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.114, + 0.891, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\hat {\\Omega}, Z, \\psi) = \\lambda_ {\\mathrm {p r o j}} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\Omega}) + \\lambda_ {\\mathrm {s l h}} \\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\Omega}, Z, \\psi) \\\\ + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.345 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{proj}}\\), \\(\\lambda_{\\mathrm{slh}}\\), and \\(\\lambda_{\\mathrm{reg}}\\) denote the loss weights. However, although the multi-view projection loss \\(\\mathcal{L}_{\\mathrm{proj}}\\) and multi-view silhouette loss \\(\\mathcal{L}_{\\mathrm{slh}}\\) are based on 2D supervision, the 3D-2D correspondence to define the losses is not obvious. Therefore, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm [14]. Each pair-wise matching cost is defined as the projection loss introduced in this section, but it is computed for not all the source frames but only the target frame. We assume that the ground truth 2D bounding boxes and instance masks have already been reordered based on the optimal permutation for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Multi-View Projection Loss Although introducing the residual distance field enables us to model an arbitrary surface bounded by a 3D bounding box, it introduces another problem: the 3D bounding box can grow indefinitely without any constraint. To address this problem, we employ the ground truth 2D bounding box as a constraint to keep the 3D bounding box tight. The multi-view projection loss \\(\\mathcal{L}_{\\mathrm{proj}}\\) is defined as the average distance between the projected and ground truth 2D bounding boxes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.506, + 0.891, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\boldsymbol {\\Omega}}) = \\alpha \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\| \\hat {\\boldsymbol {B}} _ {i n} ^ {2 \\mathrm {D}} - \\boldsymbol {B} _ {i n} ^ {2 \\mathrm {D}} \\| _ {\\mathrm {H}} \\\\ - \\beta \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\mathrm {D I o U} \\left(\\hat {\\boldsymbol {B}} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}\\right), \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.694 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_{\\mathrm{H}}\\) denotes the Huber loss, \\(\\mathrm{DIOU}(\\cdot ,\\cdot)\\) denotes the Distance-IoU [36], and \\(\\alpha ,\\beta\\) denote balancing coefficients. The projected 2D bounding box \\(\\hat{B}_{in}^{2\\mathrm{D}}\\) is defined as the rectangle with minimal area enclosing the projected vertices \\(\\hat{\\pmb{V}}_{in}^{2\\mathrm{D}}\\propto \\hat{\\pmb{B}}_n\\pmb {E}_i^T\\pmb {K}_i^T\\) where \\(\\pmb {E}_i\\) and \\(\\pmb {K}_i\\) denote the extrinsic and intrinsic matrices for frame \\(i\\) , respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.892, + 0.847 + ], + "angle": 0, + "content": "Multi-View Silhouette Loss The multi-view silhouette loss is defined as the cross entropy between the rendered and ground truth instance masks. Since the spatial gap between the surfaces of each instance and the 3D bounding box is modeled by the residual distance field introduced in Sec. 3.1.4, the 3D bounding box constrained to tightly fit the ground truth 2D bounding box by the multi-view projection loss \\(\\mathcal{L}_{\\mathrm{proj}}\\) is further refined by the multi-view silhouette loss \\(\\mathcal{L}_{\\mathrm{sh}}\\), which is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.855, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi) = \\sum_ {i \\in S} \\sum_ {j = 1} ^ {R _ {i}} \\operatorname {C E} \\left(\\hat {\\boldsymbol {S}} \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {d} _ {i j}\\right), \\boldsymbol {S} _ {i j}\\right), \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.272, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.167, + 0.218, + 0.177 + ], + "angle": 0, + "content": "(a) Static Scene" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.468, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.167, + 0.42, + 0.178 + ], + "angle": 0, + "content": "(b) Dynamic Scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.182, + 0.47, + 0.238 + ], + "angle": 0, + "content": "Figure 4. Comparison of the confidence scores in static and dynamic scenes. It can be seen that the confidence scores are lower for dynamic, occluded, or truncated objects, indicating less influence on the subsequent training of 3D object detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.468, + 0.371 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{o}_i \\in \\mathbb{R}^3 \\), \\( \\boldsymbol{d}_{ij} \\in \\mathbb{R}^3 \\), and \\( S_{ij} \\in \\{0,1\\}^N \\) denote the camera position, ray direction, and ground truth instance label at the \\( j \\)-th sampled pixel in frame \\( i \\), respectively. \\( \\hat{S}(\\boldsymbol{o}_i, \\boldsymbol{d}_{ij}) \\) denotes the rendered instance label based on Eq. (8). \\( \\mathrm{CE}(\\cdot, \\cdot) \\) denotes the cross entropy loss. \\( R_i \\) denotes the number of rays sampled for frame \\( i \\). Please refer to the supplementary material for how to determine \\( R_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.36, + 0.397 + ], + "angle": 0, + "content": "3.2. Training of 3D Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.469, + 0.511 + ], + "angle": 0, + "content": "Once the 3D bounding boxes are optimized by the proposed multi-view 3D auto-labeling, they can serve as pseudo labels for training 3D object detectors. To incorporate the quality of each pseudo label into the training of 3D object detectors, we introduce a simple but effective confidence assignment method in Sec. 3.2.1 and confidence-based weighted loss for bounding box regression in Sec. 3.2.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.295, + 0.546 + ], + "angle": 0, + "content": "3.2.1 Confidence Assignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.469, + 0.75 + ], + "angle": 0, + "content": "As shown in Fig. 4, the 3D bounding boxes optimized by the proposed auto-labeling are not reliable for dynamic, occluded, or truncated objects. Therefore, we propose a confidence assignment method based on the multi-view projection loss. First, for each target frame \\( t \\), we identify a set of source frames \\( \\mathcal{I} \\) such that all the instances in the target frame are visible from every frame in the set. Next, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm. The cost matrix \\( Q \\in \\mathbb{R}^{N \\times N} \\) is defined as the pairwise IoUs between the projected and ground truth 2D bounding boxes averaged over all the source frames, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.761, + 0.469, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Q} _ {n m} = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} 1 - \\operatorname {I o U} \\left(^ {*} \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i m} ^ {\\mathrm {2 D}}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.469, + 0.903 + ], + "angle": 0, + "content": "where \\({}^{*}B_{in}^{2D}\\) denotes the \\(n\\)-th projected 2D bounding box in frame \\(i\\), which is obtained by projecting the \\(n\\)-th optimized 3D bounding box \\({}^{*}B_{n}\\) onto frame \\(i\\), and \\({B_{im}^{2D}}\\) denotes the \\(m\\)-th ground truth 2D bounding box in frame \\(i\\). Once an optimal permutation matrix \\(\\pmb{P} \\in \\{0,1\\}^{N \\times N}\\) is obtained, the confidence scores \\({}^{*}\\pmb{C} \\in [0,1]^N\\) for the optimized 3D" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.772, + 0.107 + ], + "angle": 0, + "content": " bounding boxes \\(^* B\\) are given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.112, + 0.892, + 0.149 + ], + "angle": 0, + "content": "\\[\n^ {*} C = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} \\operatorname {I o U} \\left(^ {*} B _ {i} ^ {\\mathrm {2 D}}, P B _ {i} ^ {\\mathrm {2 D}}\\right). \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.164, + 0.783, + 0.179 + ], + "angle": 0, + "content": "3.2.2 Confidence-based Weighted Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.892, + 0.325 + ], + "angle": 0, + "content": "The 3D bounding boxes optimized by the proposed multiview 3D auto-labeling serve as pseudo labels for most 3D object detectors, which leverage 3D bounding boxes and semantic class labels as supervision, without any modification of the architectures, loss functions, and training procedures, except for the confidence-based loss weighting. We incorporate the confidence scores computed by Eq. (14) into only the regression loss. Given a regression loss function \\(\\mathcal{L}_{\\mathrm{box}}\\) its confidence-based weighted version \\(\\tilde{\\mathcal{L}}_{\\mathrm{box}}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {L}} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}}, ^ {*} \\boldsymbol {B}, ^ {*} \\boldsymbol {C}\\right) = \\sum_ {m = 1} ^ {M} ^ {*} \\boldsymbol {C} _ {\\pi (m)} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}} _ {m}, ^ {*} \\boldsymbol {B} _ {\\pi (m)}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.471 + ], + "angle": 0, + "content": "where \\(\\hat{B}\\) denotes the predicted 3D bounding boxes, \\(^* B\\) denotes the 3D bounding boxes optimized by the proposed auto-labeling, and \\(^* C\\) denotes the corresponding confidence scores. \\(M\\) denotes the number of positive anchors, and \\(\\pi (\\cdot)\\) denotes the label assigner that maps the index of an anchor to that of the matched ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.632, + 0.499 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.596, + 0.521 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.681 + ], + "angle": 0, + "content": "We use the KITTI-360 [15] dataset for our experiments, splitting it into training (43,855 images), validation (1,173 images), and test sets (2,531 images). We follow the same evaluation protocol as the KITTI dataset [9]. However, as occlusion and truncation labels are not available for the KITTI-360 dataset, unlike the KITTI dataset, we consider only two difficulty levels, namely Easy and Hard, based on whether the height of each ground truth 2D bounding box is greater than 40 and 25 in pixels, respectively. Following prior works, we evaluate our method on only category Car." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.717, + 0.704 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.763, + 0.727 + ], + "angle": 0, + "content": "4.2.1 Multi-View 3D Auto-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We sample 16 source frames for each target frame. We sample 1000 rays across all the source frames, i.e., \\(\\sum_{i\\in S}R_i = 1000\\), at each iteration based on the ground truth instance masks. Please refer to the supplementary material for more details. We employ the same hierarchical volume sampling as NeRF [18] and sample 100 query points for both coarse and fine sampling. Unlike NeRF, both coarse and fine samples are drawn from the single scene SDF rather than two distinct ones. We set the number of dimensions of each instance embedding as \\(D = 256\\). The neural RDF \\(\\mathcal{G}\\) and hypernetwork \\(\\mathcal{H}\\) are implemented as MLPs with four hidden" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17359" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.132 + ], + "angle": 0, + "content": "Table 1. Ablation study on the KITTI-360 training set to verify the effectiveness of each component in our proposed multi-view 3D auto-labeling. Just one randomly selected sequence is used." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.143, + 0.47, + 0.227 + ], + "angle": 0, + "content": "
ComponentsAPBEV/AP3D@0.3APBEV/AP3D@0.5@0.5
LprojLslhRDFEasyHardEasyHard
✓60.77/54.8863.99/57.6637.38/23.3337.44/24.82
✓✓63.84/56.1160.86/56.8741.73/26.8439.22/25.58
✓✓✓73.84/66.6473.22/66.3246.35/31.1143.07/30.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.242, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Table 2. Ablation study on the KITTI-360 test set to verify the effectiveness of the confidence scores against monocular 3D object detection. S-WeakM3D is used as a monocular 3D object detector." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.295, + 0.462, + 0.368 + ], + "angle": 0, + "content": "
Conf.APBEV/AP3D@0.3APBEV/AP3D@0.5
EasyHardEasyHard
✓38.67/31.6231.25/23.8810.95/8.255.37/4.41
51.09/42.9441.27/33.9719.50/11.9114.39/8.46
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.493 + ], + "angle": 0, + "content": "layers, each of which has 256 and 16 channels, respectively. We use the Adam optimizer [13] and the learning rates are decayed exponentially from \\(1e^{-2}\\), \\(1e^{-3}\\), and \\(1e^{-4}\\) to \\(1e^{-4}\\), \\(1e^{-5}\\), and \\(1e^{-6}\\) for the box parameters \\(\\hat{\\Omega}\\), instance embeddings \\(\\mathbf{Z}\\), and parameter \\(\\psi\\) of the hypernetwork \\(\\mathcal{H}(\\cdot; \\psi)\\) over 3000 iterations, respectively. For the loss weights, we set \\(\\alpha = 1.0\\), \\(\\beta = 0.1\\), \\(\\lambda_{\\mathrm{proj}} = 1.0\\), \\(\\lambda_{\\mathrm{slh}} = 1.0\\), and \\(\\lambda_{\\mathrm{reg}} = 0.01\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.354, + 0.53 + ], + "angle": 0, + "content": "4.2.2 Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.751 + ], + "angle": 0, + "content": "To compare our method and Autolabels [33] with WeakM3D [23], we modify the architecture of WeakM3D so that it can be trained in a supervised manner using pseudo labels. More specifically, we train dimension and confidence heads in addition to the existing location and orientation heads using the same supervised loss as MonoDIS [25]. We call this model S-WeakM3D. For both WeakM3D and S-WeakM3D, ground truth 2D bounding boxes and instance masks are used for RoIAlign [11] and generating Li-DAR points on each object surface during training, respectively. During inference, we employ Cascade Mask R-CNN [5] with InternImage-XL [29] as the off-the-shelf 2D detector to provide 2D bounding boxes for RoIAlign. Please refer to the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.231, + 0.779 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.34, + 0.801 + ], + "angle": 0, + "content": "4.3.1 Multi-View 3D Auto-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We conduct an ablation study to demonstrate the effectiveness of each component in our proposed multi-view 3D auto-labeling. As can be seen from Tab. 1, the multi-view projection loss \\(\\mathcal{L}_{\\mathrm{proj}}\\) serves as a practical baseline on its own. The multi-view silhouette loss \\(\\mathcal{L}_{\\mathrm{slh}}\\) further boosts the quality of the pseudo labels a little bit, while the spatial gap" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.088, + 0.891, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.894, + 0.285 + ], + "angle": 0, + "content": "Figure 5. Visualization results of the optimized 3D bounding boxes (1st row) and rendered instance masks (2nd row). We assign a unique color to each instance, and each pixel is colored as the weighted summation based on the rendered soft instance label." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.299, + 0.893, + 0.356 + ], + "angle": 0, + "content": "Table 3. Evaluation results of our proposed multi-view 3D auto-labeling on the KITTI-360 training set compared with the LiDAR-based monocular 3D auto-labeling proposed in Autolabels [33]. *Reproduced with the official code." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.366, + 0.888, + 0.464 + ], + "angle": 0, + "content": "
Conf.MethodAPBEV/AP3D@@0.3APBEV/AP3D@0.5
EasyHardEasyHard
≥0.0Autolabels*71.24/15.0967.33/11.4251.85/4.6546.10/2.92
VSRD75.03/68.5372.11/65.6447.12/35.2543.91/32.64
≥0.8Autolabels*75.56/17.1872.23/12.1958.64/5.7752.99/3.75
VSRD84.54/80.2581.66/77.3758.57/45.7655.09/44.17
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.59 + ], + "angle": 0, + "content": "between the surfaces of each instance and the 3D bounding box limits further improvements. However, the residual distance field (RDF) can significantly boost the quality of the pseudo labels by addressing this problem. The systematic improvements with the inclusion of each component indicate their effectiveness, resulting in a precise auto-labeling system for weakly supervised 3D object detection." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.608, + 0.719, + 0.623 + ], + "angle": 0, + "content": "4.3.2 Confidence Assignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.754 + ], + "angle": 0, + "content": "Due to the unreliability of the 3D bounding boxes optimized by the proposed auto-labeling for dynamic, occluded, or truncated objects, we conduct an ablation study to demonstrate the effectiveness of our proposed confidence assignment. Tab. 2 highlights substantial enhancement in detection performance when using the confidence-incorporated pseudo labels compared with the baseline, demonstrating the effectiveness of the proposed confidence assignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.681, + 0.793 + ], + "angle": 0, + "content": "4.4. Evaluation Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.763, + 0.817 + ], + "angle": 0, + "content": "4.4.1 Multi-View 3D Auto-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Autolabels [33] employs a similar two-stage framework consisting of auto-labeling and subsequent training of 3D object detectors using the pseudo labels. We compare our method with Autolabels to evaluate the quality of the pseudo labels. As our pseudo labels are unreliable for dy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17360" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4. Evaluation results of monocular 3D object detection on the KITTI-360 test set. *Reproduced with the official code. †CAD models are used as extra data. \\( {}^{ \\ddagger }M\\left( D\\right) \\) indicates that detection model \\( D \\) is employed for model-agnostic method \\( M \\) ." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.129, + 0.895, + 0.279 + ], + "angle": 0, + "content": "
MethodWeak SupervisionFull SupervisionAPBEV/AP3D@0.3APBEV/AP3D@0.5
LiDARMasks3D BoxesEasyHardEasyHard
WeakM3D* [23]✓✓29.89/21.2524.01/15.348.10/2.962.96/2.01
Autolabels*†‡ [33] (S-WeakM3D)✓✓48.16/12.9237.34/9.9420.18/4.6914.33/2.79
VSRD‡ (S-WeakM3D)✓51.09/42.9441.28/33.7819.51/11.9114.39/8.46
VSRD‡ (MonoFlex)✓54.40/48.1645.67/40.0429.18/18.5322.31/13.60
VSRD‡ (MonoDETR)✓58.40/50.8650.61/43.4529.07/21.7722.83/16.46
MonoFlex* [35]✓69.70/67.0759.86/57.2650.82/43.1141.78/34.43
MonoDETR* [34]✓63.07/60.4954.04/50.0347.21/41.0136.05/30.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.295, + 0.468, + 0.324 + ], + "angle": 0, + "content": "Table 5. Evaluation results of semi-supervised monocular 3D object detection on the KITTI validation set." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.334, + 0.466, + 0.45 + ], + "angle": 0, + "content": "
MethodRatioAPBEV/AP3D@0.7
EasyModerateHard
MonoDETR [34]1.0037.99/29.3626.76/20.6423.02/17.30
0.000.002/0.0010.004/0.0010.005/0.002
0.2531.72/21.7622.32/15.4318.86/12.55
VSRD (MonoDETR)0.5043.44/31.0531.54/21.4827.17/17.93
0.7542.58/32.9531.08/24.6827.19/21.38
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.66 + ], + "angle": 0, + "content": "namic, occluded, or truncated objects, we group the dataset based on whether the average confidence score for each image is greater than a certain threshold. The evaluation results with the confidence thresholds of 0.0 and 0.8 are shown in Tab. 3. Our method exhibits superior performance in terms of \\(\\mathrm{AP}_{\\mathrm{BEV}}@0.3\\) and \\(\\mathrm{AP}_{\\mathrm{3D}}\\). For \\(\\mathrm{AP}_{\\mathrm{BEV}}@0.5\\), our method exhibits slightly lower performance for the confidence threshold of 0.0. However, by raising the confidence threshold to 0.8, our method demonstrates superior performance, indicating that our method can generate more high-quality pseudo labels while abandoning low-quality pseudo labels. The optimized 3D bounding boxes and rendered instance masks are visualized in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.683, + 0.355, + 0.699 + ], + "angle": 0, + "content": "4.4.2 Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.754 + ], + "angle": 0, + "content": "The pseudo labels generated by the proposed auto-labeling serve as 3D supervision. We further investigate their applicability using the existing monocular 3D object detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Weakly Supervised Setting Tab. 4 shows the evaluation results of our method compared with the existing weakly supervised and fully supervised methods. Our method demonstrates a significant superiority over WeakM3D [23] across all the metrics while eliminating the need for LiDAR points for 3D supervision. Moreover, the detector trained on the pseudo labels generated by the proposed auto-labeling outperforms that trained on the pseudo labels generated by Autolabels [33]. Furthermore, employing more sophisticated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.405 + ], + "angle": 0, + "content": "monocular 3D object detectors such as MonoFlex [35] and MonoDETR [34] further improves detection performance, demonstrating the broad versatility of our method, which is not limited to a specific detection model. It is noteworthy that the detectors trained on the pseudo labels generated by our method demonstrate competitive performance compared with those trained in a fully supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.682 + ], + "angle": 0, + "content": "Semi-Supervised Setting The essential advantage of our method is that it avoids costly 3D annotations, making more data available for training. Therefore, we investigate a realistic scenario where a detector pre-trained on a large amount of unlabeled data from a source domain is fine-tuned on a small amount of labeled data from a target domain. We select the KITTI-360 and KITTI datasets as source and target domains, respectively. Tab. 5 shows the performance of the detector pre-trained on the KITTI-360 dataset in a weakly supervised manner with the proposed auto-labeling and then fine-tuned on a subset of the KITTI dataset in a supervised manner. The zero-shot performance is quite low due to the characteristic that monocular depth estimation is greatly affected by the differences in camera parameters, but the performance of the detector fine-tuned on only \\(50\\%\\) of the labeled data significantly outperforms that trained on the whole data from scratch, highlighting the broad applicability of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel weakly supervised 3D object detection framework named VSRD, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. Our method demonstrates superior performance compared with the existing weakly supervised 3D object detection methods. Moreover, it exhibits remarkable scalability using partially labeled data for semi-supervised learning. Our proposed method allows leveraging abundant 2D annotations to enhance 3D object detection without explicit 3D supervision, providing a promising avenue for further advancements in the field." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17361" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.498 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.657, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.715, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[10] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[12] Benran Hu, Junkai Huang, Yichen Liu, Yu-Wing Tai, and Chi-Keung Tang. Nerf-rpn: A general framework for object detection in nerfs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23528–23538, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[14] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[15] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[17] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In European Conference on computer vision, pages 515-531. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[19] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7074-7082, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[21] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[22] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[23] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. arXiv preprint arXiv:2203.08332, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[24] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17362" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "3d object detection using multi-view projection and direction consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17492, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14408-14419, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.43 + ], + "angle": 0, + "content": "[30] Chenfeng Xu, Bichen Wu, Ji Hou, Sam Tsai, Ruilong Li, Jialiang Wang, Wei Zhan, Zijian He, Peter Vajda, Kurt Keutzer, et al. Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23320-23330, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[31] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[32] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[33] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.471, + 0.684 + ], + "angle": 0, + "content": "[34] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[35] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.471, + 0.81 + ], + "angle": 0, + "content": "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, pages 12993-13000, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17363" + } + ] +] \ No newline at end of file diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_origin.pdf b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b76891c88edb6877457cde9d71a27dc61b64045 --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/af94e918-61b9-42a3-a56a-ca638b10a2e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:496a9c93aee139fa01dc1e78d368ef816659c92e0656752e8fc18563c8df8165 +size 1090032 diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/full.md b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83a991c1c351d60a712aab61d326e121213535ad --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/full.md @@ -0,0 +1,314 @@ +# VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection + +Zihua Liu* Tokyo Institute of Technology zliu@ok.sc.e.titech.ac.jp + +Hiroki Sakuma* T2 Inc. sakuma.h@t2.auto + +Masatoshi Okutomi Tokyo Institute of Technology mxo@sc.e.titech.ac.jp + +# Abstract + +Monocular 3D object detection poses a significant challenge in 3D scene understanding due to its inherently ill-posed nature in monocular depth estimation. Existing methods heavily rely on supervised learning using abundant 3D labels, typically obtained through expensive and labor-intensive annotation on LiDAR point clouds. To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each instance as a signed distance field (SDF) and render its silhouette as an instance mask through our proposed instance-aware volumetric silhouette rendering. To directly optimize the 3D bounding boxes through rendering, we decompose the SDF of each instance into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This mechanism enables us to optimize the 3D bounding boxes in an end-to-end manner by comparing the rendered instance masks with the ground truth instance masks. The optimized 3D bounding boxes serve as effective training data for 3D object detection. We conduct extensive experiments on the KITTI-360 dataset, demonstrating that our method outperforms the existing weakly supervised 3D object detection methods. The code is available at https://github.com/skmhrk1209/VSRD. + +# 1. Introduction + +3D object detection is one of the most critical components in the perception system for autonomous driving. With the recent success of deep learning in computer vision, numerous 3D object detection methods have been proposed to + +![](images/6a81a6ed70f7d663c5e24906782fde59b6f6c3aa7016ac521e90998dd2c9f2a4.jpg) +Figure 1. Illustration of our proposed weakly supervised 3D object detection framework, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. + +directly regress 3D bounding boxes from a LiDAR point cloud, multi-view images, or a monocular image. Among them, monocular 3D object detection is the most challenging in principle due to its inherently ill-posed nature in monocular depth estimation. Therefore, existing methods [22, 25, 28, 34, 35] heavily rely on supervised learning using abundant 3D labels manually annotated on LiDAR point clouds. This annotation cost is extremely high, posing a significant barrier to deploying 3D object detectors into autonomous driving systems. + +To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. As illustrated in Fig. 1, VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each 3D bounding box as a signed distance field (SDF) and render its silhouette as an instance mask through volumetric rendering. Comparing the rendered instance masks with the ground truth instance masks enables us to optimize the 3D bounding boxes directly. We introduce two novel mechanisms in this auto-labeling stage. The first is the instance-aware volumetric silhouette rendering that integrates instance labels along a ray to render the silhouette of each instance rather + +than the entire scene. This mechanism enables the silhouette of each instance to be rendered while considering geometric relationships among instances, such as occlusion. The second is the SDF decomposition, whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This decomposition models the spatial gap between the surfaces of each instance and the 3D bounding box, enabling more accurate silhouette rendering and providing more reliable feedback signals during optimization. + +The 3D bounding boxes optimized by the proposed auto-labeling can serve as pseudo labels for training 3D object detectors. However, dynamic objects and inaccurate camera poses lead to low-quality pseudo labels, which negatively impact the training of 3D object detectors. Therefore, we propose a simple but effective algorithm to assign a confidence score representing the label quality to each pseudo label. We demonstrate that using these confidence scores as per-instance loss weights boosts the performance of the 3D object detectors trained on the pseudo labels. + +In summary, our main contributions are as follows: + +- We propose a weakly supervised 3D object detection framework consisting of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors. +- We propose a novel instance-aware volumetric silhouette rendering method whereby the silhouette of each instance can be rendered as an instance mask. +- We propose a novel SDF decomposition whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF). +- We propose a simple but effective confidence assignment algorithm to incorporate the quality of each pseudo label into the training of 3D object detectors. +- Extensive experiments on the KITTI-360 dataset demonstrate that our method outperforms the existing weakly supervised 3D object detection methods. + +# 2. Related Work + +# 2.1. Monocular 3D Object Detection + +Monocular 3D detection is a challenging task due to limited 3D information from monocular imagery. Deep3Box [19] pioneered this area by regressing relatively stable 3D object properties and combining these estimates with geometric constraints provided by the 2D bounding box. SMOKE [16] estimates 3D bounding boxes by combining keypoint estimates with regressed 3D box parameters. FCOS3D [28] employs a fully convolutional single-stage detector, transforming 7-DoF 3D targets to the image domain and decoupling them as 2D and 3D attributes. To enhance the performance of monocular 3D object detection for truncated objects, MonoFlex [35] explicitly decouples truncated objects and adaptively combines multiple approaches for ob + +ject depth estimation. M3D-RPN [4] and MonoDETR [34] also explored the usage of depth cues to improve monocular 3D object detection. The former designed depth-aware convolutional layers that enable location-specific feature extraction and consequently improved 3D scene understanding, while the latter modified the vanilla Transformer [7] to be depth-aware to guide the whole detection process by contextual depth cues. Despite these advances, the reliance on expensive and labor-intensive manual annotation on LiDAR point clouds remains a significant limitation. + +# 2.2. Weakly Supervised 3D Object Detection + +Many weakly supervised methods have been proposed to mitigate the high annotation cost in 3D object detection. WS3D [17] introduced a LiDAR-based two-stage pipeline where cylindrical object proposals are first generated under weak supervision and then refined using a few labeled object instances. VS3D [24] introduced an unsupervised 3D proposal module that generates object proposals by leveraging normalized point cloud densities. WeakM3D [23] introduced a weakly supervised monocular 3D object detection method that leverages the 3D alignment loss between each predicted 3D bounding box and corresponding RoI LiDAR points. Furthermore, it introduced a method to estimate the orientation from RoI LiDAR points based on its statistics. Recently, WeakMono3D [26] leverages projection loss with multi-view and direction consistency, achieving a weakly supervised monocular object detection that relies only on 2D supervision. However, its reliance on 2D direction annotations restricts its applicability to large-scale datasets. Zakharov et al. [33] proposed an auto-labeling pipeline that integrates an SDF-based differentiable shape renderer and normalized object coordinate spaces (NOCS). However, additional training on synthetic data is still required for shape and coordinate estimation. In contrast, our method purely relies on 2D supervision, eliminating the necessity of synthetic data or 3D supervision. + +# 2.3. 3D Object Detection with Neural Fields + +Neural Radiance Fields (NeRF) [18] introduced a new perspective for implicit learning of 3D geometry from posed multi-view images by volume rendering. Building upon the vanilla NeRF, subsequent research has focused on enhancing novel view synthesis [1-3, 21] or accelerating volume rendering [6, 8, 20]. NeuS [27] introduced a novel volume rendering scheme to learn a neural SDF representation by introducing a density distribution induced by the SDF. Similarly, VolSDF [32] defined the volume density function as Laplace's cumulative distribution function applied to an SDF representation. Many works [12, 30, 31] recently have attempted to utilize neural fields for 3D object detection. Notably, NeRF-RPN [12] demonstrated that the 3D bounding boxes of objects in NeRF can be directly regressed with + +![](images/779429e0437a8ab701f8bc8482c50fccecc89a93d4c4e9aea1bf27decc31d3d1.jpg) +Figure 2. Illustration of the pipeline of our proposed multi-view 3D auto-labeling. We represent the surface of each instance as an SDF and decompose it into the SDF of a 3D bounding box and the residual distance field (RDF), which is learned via a hypernetwork. The composed instance SDF is used to render the silhouette of the instance through our proposed instance-aware volumetric silhouette rendering. All the 3D bounding boxes are optimized based on the loss between the rendered and ground truth instance masks. + +out rendering. NeRF-Det [30] connects the detection and NeRF branches through a shared MLP, enabling an efficient adaptation of NeRF to detection and yielding geometry-aware volumetric representations. MonoNeRD [31] models scenes with SDFs and renders RGB images and depth maps through volume rendering to obtain intermediate 3D representations for detection. However, these approaches still rely on ground truth 3D labels for supervision. In contrast, we propose the first volume rendering-based weakly supervised 3D object detection framework that relies on multiview geometry and 2D supervision without any 3D supervision, such as 3D bounding boxes or LiDAR point clouds. + +# 3. Method + +# 3.1. Multi-View 3D Auto-Labeling + +The pipeline of our proposed multi-view 3D auto-labeling is illustrated in Fig. 2. In Sec. 3.1.1, we review the SDF-based volume rendering formulation, which is the foundation of our method. In Sec. 3.1.2, we define the optimization problem for the multi-view 3D auto-labeling. In Sec. 3.1.3, we introduce the SDF for cuboids, which is used to optimize 3D bounding boxes through rendering. In Sec. 3.1.4, we introduce a novel neural field named residual distance + +field to fill the spatial gap between the surfaces of each instance and the 3D bounding box. In Sec. 3.1.5, we introduce instance-aware volumetric silhouette rendering to render the silhouette of each instance based on its SDF. In Sec. 3.1.6, we introduce loss functions using instance masks and 2D bounding boxes as weak supervision. + +# 3.1.1 Preliminaries + +SDF-based Volumetric Rendering NeRF [18] represents a 3D scene with neural density and color fields. Given a camera position $\pmb{o} \in \mathbb{R}^3$ and a ray direction $\pmb{d} \in \mathbb{R}^3$ emitted from a pixel, the volume rendering scheme integrates the colors of sampled points along the ray as follows: + +$$ +\hat {\boldsymbol {C}} (\boldsymbol {o}, \boldsymbol {d}) = \int_ {0} ^ {\infty} w (t) \boldsymbol {c} (\boldsymbol {r} (t), \boldsymbol {d}) \mathrm {d} t, \tag {1} +$$ + +$$ +w (t) = \exp (- \int_ {0} ^ {t} \sigma (\boldsymbol {r} (u)) \mathrm {d} u) \sigma (\boldsymbol {r} (t)), \qquad (2) +$$ + +where $\hat{C} (\pmb {o},\pmb {d})\in \mathbb{R}^3$ denotes the rendered color of the pixel, $\boldsymbol {r}(t) = \boldsymbol {o} + t\boldsymbol{d}$ denotes the ray, $c(\pmb {p},\pmb {d})$ denotes the color at the position $\pmb{p}$ and view direction $\pmb{d}$ , and $\sigma (\pmb {p})$ denotes the volume density at the position $\pmb{p}$ . Since the density field cannot represent the surfaces explicitly, novel SDF- + +based volume rendering formulation has been introduced in NeuS [27], where a surface represented by an SDF is reinterpreted as a participating medium represented by a density field, enabling the surface to be rendered through volume rendering. In this formulation, the weight $w(t)$ in Eq. (2) is re-written by introducing opaque density $\rho(t)$ as follows: + +$$ +w (t) = \exp \left(- \int_ {0} ^ {t} \rho (u) \mathrm {d} u\right) \rho (t), \tag {3} +$$ + +$$ +\rho (t) = \max \left(- \frac {\frac {\mathrm {d} \Phi}{\mathrm {d} t} \left(\hat {\mathcal {F}} (\boldsymbol {r} (t))\right)}{\Phi \left(\hat {\mathcal {F}} (\boldsymbol {r} (t))\right)}, 0\right), \tag {4} +$$ + +where $\hat{\mathcal{F}} (\cdot)$ denotes the SDF for the entire scene and $\Phi (\cdot)$ denotes the Sigmoid function. Our proposed instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5 is based on the same SDF-based weight formulation as Eq. (3) but integrates instance labels instead of colors along a ray to render instance masks. + +# 3.1.2 Problem Definition + +Given a monocular video consisting of posed frames, each frame annotated with instance masks, our goal is to optimize the 3D bounding box frame by frame without 3D supervision. More specifically, for each target frame $t$ in the video, we sample multiple source frames $\mathcal{S}$ and optimize the $N$ 3D bounding boxes in the target frame using the instance masks of the source frames as weak supervision, where $N$ denotes the number of instances in the target frame. Please refer to the supplementary material for how to sample $\mathcal{S}$ . We parameterize the $n$ -th 3D bounding box $\hat{B}_n \in \mathbb{R}^{8 \times 3}$ in the target frame with a dimension $\hat{d}_n \in \mathbb{R}_+^3$ , location $\hat{\ell}_n \in \mathbb{R}^3$ , and orientation $\hat{\theta}_n \in \mathbb{R}$ , which is the rotation angle in the bird's-eye-view. In addition to these parameters for each bounding box, we prepare a learnable instance embedding $z_n \in \mathbb{R}^D$ for each instance and a shared hypernetwork parameterized by $\psi$ for the residual distance field introduced in Sec. 3.1.4. We stack each parameter group into a single tensor over all the instances, yielding dimensions $\hat{D} \in \mathbb{R}_+^{N \times 3}$ , locations $\hat{L} \in \mathbb{R}^{N \times 3}$ , orientations $\hat{\Theta} \in \mathbb{R}^{N \times 1}$ , and instance embeddings $Z \in \mathbb{R}^{N \times D}$ . Given the loss function $\mathcal{L}$ explained in Sec. 3.1.6, we optimize $\hat{D}$ , $\hat{L}$ , $\hat{\Theta}$ , $Z$ , and $\psi$ via stochastic gradient descent as follows: + +$$ +{ } ^ { * } \boldsymbol { D } , { } ^ { * } \boldsymbol { L } , { } ^ { * } \boldsymbol { \Theta } , { } ^ { * } \boldsymbol { Z } , { } ^ { * } \boldsymbol { \psi } = \operatorname* { a r g m i n } _ { \hat { \boldsymbol { D } } , \hat { \boldsymbol { L } } , \hat { \boldsymbol { \Theta } } , \boldsymbol { Z } , \boldsymbol { \psi } } \mathcal { L } ( \hat { \boldsymbol { D } } , \hat { \boldsymbol { L } } , \hat { \boldsymbol { \Theta } } , \boldsymbol { Z } , \boldsymbol { \psi } ) . \tag {5} +$$ + +The optimized 3D bounding boxes ${}^{*}\pmb{B}$ decoded from ${}^{*}\pmb{D}$ , ${}^{*}\pmb{L}$ , and ${}^{*}\Theta$ can be used as pseudo labels for training 3D object detectors, as explained in Sec. 3.2. + +# 3.1.3 3D Bounding Box Represented as an SDF + +To optimize the 3D bounding boxes in each target frame by solving Eq. (5) through rendering, we represent the surface of each 3D bounding box as a signed distance field (SDF), which is one of the most common surface representations. An SDF is represented as a function $\mathcal{F}:\mathbb{R}^3\to \mathbb{R}$ that maps a spatial position $\pmb {p}\in \mathbb{R}^3$ to its signed distance to the closest point on the surface, indicating that the zero-level set $\{\pmb {p}\in \mathbb{R}^3\mid \mathcal{F}(\pmb {p}) = 0\}$ represents the surface itself. The SDF for a cuboid parameterized by a dimension $d\in \mathbb{R}_+^3$ location $\ell \in \mathbb{R}^3$ , and orientation $\pmb {R}\in \mathrm{SO}(3)$ can be derived theoretically and is denoted by $\mathcal{B}(\cdot ;d,\ell ,\pmb {R})$ . Please refer to the supplementary material for how to derive this formula. + +# 3.1.4 Residual Distance Field + +In general, the shape of each instance is not a cuboid. Therefore, if we leverage only the cuboid SDF introduced in Sec. 3.1.3 to render the surface of each instance, it is no longer possible to render an accurate silhouette due to the spatial gap between the surfaces of the instance and the 3D bounding box, leading to unreliable feedback signals during optimization. Therefore, we propose a novel neural field named residual distance field (RDF), which models the residual between the signed distances to the surfaces of the instance and the 3D bounding box. Let $\mathcal{F}_n(\cdot)$ be the true SDF of the surface bounded by the $n$ -th 3D bounding box $\hat{\pmb{B}}_n$ whose SDF is given by $\hat{\mathcal{B}}_n(\cdot) = \mathcal{B}(\cdot; \hat{\pmb{d}}_n, \hat{\ell}_n, \pmb{R}_y(\theta_n))$ , where $\pmb{R}_y(\theta)$ denotes the rotation matrix around the $y$ -axis by an angle $\theta$ . For any point $\pmb{p} \in \mathbb{R}^3$ , we define the RDF as $\hat{\mathcal{R}}_n(\pmb{p}) := \mathcal{F}_n(\pmb{p}) - \hat{\mathcal{B}}_n(\pmb{p})$ . Here, based on the definition of a 3D bounding box that it encloses the corresponding instance, $\forall \pmb{p} \in \mathbb{R}^3: \hat{\mathcal{R}}_n(\pmb{p}) \geq 0$ is required. One straightforward way to model the RDF for each instance is to train $N$ individual networks. However, as objects belonging to the same semantic class often have similar geometric shapes, we train a single hypernetwork [10] that regresses the weights of the neural RDF directly from an instance embedding instead. Given the $n$ -th instance embedding $\pmb{z}_n \in \mathbb{R}^D$ , where $D$ denotes the number of dimensions, the $n$ -th neural RDF is given by: + +$$ +\hat {\mathcal {R}} _ {n} (\boldsymbol {p}) = \sigma \left(\mathcal {G} (\boldsymbol {p}; \phi_ {n})\right), \tag {6} +$$ + +$$ +\phi_ {n} = \mathcal {H} \left(\boldsymbol {z} _ {n}; \psi\right), \tag {7} +$$ + +where $\mathcal{G}(\cdot ;\phi_n)$ denotes the $n$ -th neural network parameterized by $\phi_{n}$ $\mathcal{H}(\cdot ;\psi)$ denotes the shared hypernetwork parameterized by $\psi$ , and $\sigma (\cdot)$ denotes the Softplus function to force the residual distance to be positive based on the definition of a bounding box. For the instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5, we employ $\hat{\mathcal{F}}_n(\pmb {p}) = \hat{\mathcal{B}}_n(\pmb {p}) + \hat{\mathcal{R}}_n(\pmb {p})$ as the SDF of each instance that represents an arbitrary surface bounded by the 3D bounding box $\hat{B}_n$ for more accurate silhouette rendering. + +![](images/55548fd88ddc25c3267c6a6fcc528680246f8ded7476167037ed917c72946d02.jpg) +Figure 3. Illustration of our proposed instance-aware volumetric silhouette rendering. The instance labels are averaged for each sampled point along a ray based on the signed distance to each instance. The averaged instance labels are integrated along the ray based on the SDF-based volume rendering formulation [27]. + +# 3.1.5 Instance-Aware Volumetric Silhouette Rendering + +This section details our approach to optimizing 3D bounding boxes through volumetric rendering. The core idea is to render instance masks and compare them with ground truth instance masks. To achieve this, we propose a novel SDF-based instance-aware volumetric silhouette rendering, where instance labels instead of colors are integrated along a ray based on the same SDF-based volume rendering formulation as Eq. (3) as follows: + +$$ +\hat {\boldsymbol {S}} (\boldsymbol {o}, \boldsymbol {d}) = \int_ {0} ^ {\infty} w (t) \boldsymbol {s} (\boldsymbol {r} (t)) d t, \tag {8} +$$ + +$$ +\boldsymbol {s} (\boldsymbol {p}) = \sum_ {n = 1} ^ {N} \operatorname {s o f t m i n} \left(\left[ \hat {\mathcal {F}} _ {n} (\boldsymbol {p}) \right] _ {n = 1} ^ {N}\right) _ {n} \cdot \boldsymbol {y} _ {n}, \tag {9} +$$ + +where, $\hat{S} (\pmb {o},\pmb {d})\in [0,1]^N$ denotes the rendered soft instance label, $s(\pmb {p})\in [0,1]^N$ denotes the weighted average instance label at the position $\pmb{p}$ indicating how relatively close the position $\pmb{p}$ is to each instance, and $\pmb {y}_n\in \{0,1\} ^N$ denotes the one-hot instance label of the $n$ -th instance. $\hat{\mathcal{F}}_n(\cdot)$ denotes the SDF of the $n$ -th instance introduced in Sec. 3.1.4. $w(t)$ in Eq. (8) is derived from Eq. (3). To compute Eq. (3), we model the entire scene as the union of all the object surfaces, i.e., $\hat{\mathcal{F}} (\pmb {p}) = \min (\hat{\mathcal{F}}_1(\pmb {p}),\dots ,\hat{\mathcal{F}}_N(\pmb {p}))$ . This mechanism enables us to render instance masks considering the geometric relationships among instances, such as occlusions. + +# 3.1.6 Loss Functions + +We optimize the parameters of the 3D bounding boxes $\hat{\Omega} = \{\hat{D},\hat{L},\hat{\Theta}\}$ in the target frame along with the instance embeddings $Z$ and parameter $\psi$ of the hypernetwork $\mathcal{H}(\cdot ;\psi)$ . The final loss $\mathcal{L}$ is defined as the combination of the multi-view projection loss $\mathcal{L}_{\mathrm{proj}}$ , multi-view silhouette loss $\mathcal{L}_{\mathrm{sh}}$ , and Eikonal regularization $\mathcal{L}_{\mathrm{reg}}$ [27] as follows: + +$$ +\begin{array}{l} \mathcal {L} (\hat {\Omega}, Z, \psi) = \lambda_ {\mathrm {p r o j}} \mathcal {L} _ {\mathrm {p r o j}} (\hat {\Omega}) + \lambda_ {\mathrm {s l h}} \mathcal {L} _ {\mathrm {s l h}} (\hat {\Omega}, Z, \psi) \\ + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} (\hat {\boldsymbol {\Omega}}, \boldsymbol {Z}, \psi), \tag {10} \\ \end{array} +$$ + +where $\lambda_{\mathrm{proj}}$ , $\lambda_{\mathrm{slh}}$ , and $\lambda_{\mathrm{reg}}$ denote the loss weights. However, although the multi-view projection loss $\mathcal{L}_{\mathrm{proj}}$ and multi-view silhouette loss $\mathcal{L}_{\mathrm{slh}}$ are based on 2D supervision, the 3D-2D correspondence to define the losses is not obvious. Therefore, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm [14]. Each pair-wise matching cost is defined as the projection loss introduced in this section, but it is computed for not all the source frames but only the target frame. We assume that the ground truth 2D bounding boxes and instance masks have already been reordered based on the optimal permutation for simplicity. + +Multi-View Projection Loss Although introducing the residual distance field enables us to model an arbitrary surface bounded by a 3D bounding box, it introduces another problem: the 3D bounding box can grow indefinitely without any constraint. To address this problem, we employ the ground truth 2D bounding box as a constraint to keep the 3D bounding box tight. The multi-view projection loss $\mathcal{L}_{\mathrm{proj}}$ is defined as the average distance between the projected and ground truth 2D bounding boxes as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {p r o j}} (\hat {\boldsymbol {\Omega}}) = \alpha \sum_ {i \in \mathcal {S}} \sum_ {n = 1} ^ {N} \| \hat {\boldsymbol {B}} _ {i n} ^ {2 \mathrm {D}} - \boldsymbol {B} _ {i n} ^ {2 \mathrm {D}} \| _ {\mathrm {H}} \\ - \beta \sum_ {i \in \mathcal {S}} \sum_ {n = 1} ^ {N} \mathrm {D I o U} \left(\hat {\boldsymbol {B}} _ {i n} ^ {\mathrm {2 D}}, \boldsymbol {B} _ {i n} ^ {\mathrm {2 D}}\right), \tag {11} \\ \end{array} +$$ + +where $\| \cdot \|_{\mathrm{H}}$ denotes the Huber loss, $\mathrm{DIOU}(\cdot ,\cdot)$ denotes the Distance-IoU [36], and $\alpha ,\beta$ denote balancing coefficients. The projected 2D bounding box $\hat{B}_{in}^{2\mathrm{D}}$ is defined as the rectangle with minimal area enclosing the projected vertices $\hat{\pmb{V}}_{in}^{2\mathrm{D}}\propto \hat{\pmb{B}}_n\pmb {E}_i^T\pmb {K}_i^T$ where $\pmb {E}_i$ and $\pmb {K}_i$ denote the extrinsic and intrinsic matrices for frame $i$ , respectively. + +Multi-View Silhouette Loss The multi-view silhouette loss is defined as the cross entropy between the rendered and ground truth instance masks. Since the spatial gap between the surfaces of each instance and the 3D bounding box is modeled by the residual distance field introduced in Sec. 3.1.4, the 3D bounding box constrained to tightly fit the ground truth 2D bounding box by the multi-view projection loss $\mathcal{L}_{\mathrm{proj}}$ is further refined by the multi-view silhouette loss $\mathcal{L}_{\mathrm{sh}}$ , which is given by: + +$$ +\mathcal {L} _ {\mathrm {s l h}} (\hat {\boldsymbol {\Omega}}, \boldsymbol {Z}, \psi) = \sum_ {i \in S} \sum_ {j = 1} ^ {R _ {i}} \operatorname {C E} \left(\hat {\boldsymbol {S}} \left(\boldsymbol {o} _ {i}, \boldsymbol {d} _ {i j}\right), \boldsymbol {S} _ {i j}\right), \tag {12} +$$ + +![](images/14332f1df406c5da578b10d25d2c580561d7770afa185dadc569b2809c83d25b.jpg) +(a) Static Scene + +![](images/cea7b8334a9ec2e7bace0dafbaa265073467806aa144829af9b5184f9199cbfa.jpg) +(b) Dynamic Scene +Figure 4. Comparison of the confidence scores in static and dynamic scenes. It can be seen that the confidence scores are lower for dynamic, occluded, or truncated objects, indicating less influence on the subsequent training of 3D object detectors. + +where $\boldsymbol{o}_i \in \mathbb{R}^3$ , $\boldsymbol{d}_{ij} \in \mathbb{R}^3$ , and $S_{ij} \in \{0,1\}^N$ denote the camera position, ray direction, and ground truth instance label at the $j$ -th sampled pixel in frame $i$ , respectively. $\hat{S}(\boldsymbol{o}_i, \boldsymbol{d}_{ij})$ denotes the rendered instance label based on Eq. (8). $\mathrm{CE}(\cdot, \cdot)$ denotes the cross entropy loss. $R_i$ denotes the number of rays sampled for frame $i$ . Please refer to the supplementary material for how to determine $R_i$ . + +# 3.2. Training of 3D Object Detectors + +Once the 3D bounding boxes are optimized by the proposed multi-view 3D auto-labeling, they can serve as pseudo labels for training 3D object detectors. To incorporate the quality of each pseudo label into the training of 3D object detectors, we introduce a simple but effective confidence assignment method in Sec. 3.2.1 and confidence-based weighted loss for bounding box regression in Sec. 3.2.2. + +# 3.2.1 Confidence Assignment + +As shown in Fig. 4, the 3D bounding boxes optimized by the proposed auto-labeling are not reliable for dynamic, occluded, or truncated objects. Therefore, we propose a confidence assignment method based on the multi-view projection loss. First, for each target frame $t$ , we identify a set of source frames $\mathcal{I}$ such that all the instances in the target frame are visible from every frame in the set. Next, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm. The cost matrix $Q \in \mathbb{R}^{N \times N}$ is defined as the pairwise IoUs between the projected and ground truth 2D bounding boxes averaged over all the source frames, as follows: + +$$ +\boldsymbol {Q} _ {n m} = \frac {1}{| \mathcal {I} |} \sum_ {i \in \mathcal {I}} 1 - \operatorname {I o U} \left(^ {*} \boldsymbol {B} _ {i n} ^ {\mathrm {2 D}}, \boldsymbol {B} _ {i m} ^ {\mathrm {2 D}}\right), \tag {13} +$$ + +where ${}^{*}B_{in}^{2D}$ denotes the $n$ -th projected 2D bounding box in frame $i$ , which is obtained by projecting the $n$ -th optimized 3D bounding box ${}^{*}B_{n}$ onto frame $i$ , and ${B_{im}^{2D}}$ denotes the $m$ -th ground truth 2D bounding box in frame $i$ . Once an optimal permutation matrix $\pmb{P} \in \{0,1\}^{N \times N}$ is obtained, the confidence scores ${}^{*}\pmb{C} \in [0,1]^N$ for the optimized 3D + +bounding boxes $^* B$ are given as follows: + +$$ +^ {*} C = \frac {1}{| \mathcal {I} |} \sum_ {i \in \mathcal {I}} \operatorname {I o U} \left(^ {*} B _ {i} ^ {\mathrm {2 D}}, P B _ {i} ^ {\mathrm {2 D}}\right). \tag {14} +$$ + +# 3.2.2 Confidence-based Weighted Loss + +The 3D bounding boxes optimized by the proposed multiview 3D auto-labeling serve as pseudo labels for most 3D object detectors, which leverage 3D bounding boxes and semantic class labels as supervision, without any modification of the architectures, loss functions, and training procedures, except for the confidence-based loss weighting. We incorporate the confidence scores computed by Eq. (14) into only the regression loss. Given a regression loss function $\mathcal{L}_{\mathrm{box}}$ its confidence-based weighted version $\tilde{\mathcal{L}}_{\mathrm{box}}$ is given by: + +$$ +\tilde {\mathcal {L}} _ {\mathrm {b o x}} \left(\hat {\boldsymbol {B}}, ^ {*} \boldsymbol {B}, ^ {*} \boldsymbol {C}\right) = \sum_ {m = 1} ^ {M} ^ {*} \boldsymbol {C} _ {\pi (m)} \mathcal {L} _ {\mathrm {b o x}} \left(\hat {\boldsymbol {B}} _ {m}, ^ {*} \boldsymbol {B} _ {\pi (m)}\right), \tag {15} +$$ + +where $\hat{B}$ denotes the predicted 3D bounding boxes, $^* B$ denotes the 3D bounding boxes optimized by the proposed auto-labeling, and $^* C$ denotes the corresponding confidence scores. $M$ denotes the number of positive anchors, and $\pi (\cdot)$ denotes the label assigner that maps the index of an anchor to that of the matched ground truth. + +# 4. Experiments + +# 4.1. Dataset + +We use the KITTI-360 [15] dataset for our experiments, splitting it into training (43,855 images), validation (1,173 images), and test sets (2,531 images). We follow the same evaluation protocol as the KITTI dataset [9]. However, as occlusion and truncation labels are not available for the KITTI-360 dataset, unlike the KITTI dataset, we consider only two difficulty levels, namely Easy and Hard, based on whether the height of each ground truth 2D bounding box is greater than 40 and 25 in pixels, respectively. Following prior works, we evaluate our method on only category Car. + +# 4.2. Implementation Details + +# 4.2.1 Multi-View 3D Auto-Labeling + +We sample 16 source frames for each target frame. We sample 1000 rays across all the source frames, i.e., $\sum_{i\in S}R_i = 1000$ , at each iteration based on the ground truth instance masks. Please refer to the supplementary material for more details. We employ the same hierarchical volume sampling as NeRF [18] and sample 100 query points for both coarse and fine sampling. Unlike NeRF, both coarse and fine samples are drawn from the single scene SDF rather than two distinct ones. We set the number of dimensions of each instance embedding as $D = 256$ . The neural RDF $\mathcal{G}$ and hypernetwork $\mathcal{H}$ are implemented as MLPs with four hidden + +Table 1. Ablation study on the KITTI-360 training set to verify the effectiveness of each component in our proposed multi-view 3D auto-labeling. Just one randomly selected sequence is used. + +
ComponentsAPBEV/AP3D@0.3APBEV/AP3D@0.5@0.5
LprojLslhRDFEasyHardEasyHard
✓60.77/54.8863.99/57.6637.38/23.3337.44/24.82
✓✓63.84/56.1160.86/56.8741.73/26.8439.22/25.58
✓✓✓73.84/66.6473.22/66.3246.35/31.1143.07/30.16
+ +Table 2. Ablation study on the KITTI-360 test set to verify the effectiveness of the confidence scores against monocular 3D object detection. S-WeakM3D is used as a monocular 3D object detector. + +
Conf.APBEV/AP3D@0.3APBEV/AP3D@0.5
EasyHardEasyHard
✓38.67/31.6231.25/23.8810.95/8.255.37/4.41
51.09/42.9441.27/33.9719.50/11.9114.39/8.46
+ +layers, each of which has 256 and 16 channels, respectively. We use the Adam optimizer [13] and the learning rates are decayed exponentially from $1e^{-2}$ , $1e^{-3}$ , and $1e^{-4}$ to $1e^{-4}$ , $1e^{-5}$ , and $1e^{-6}$ for the box parameters $\hat{\Omega}$ , instance embeddings $\mathbf{Z}$ , and parameter $\psi$ of the hypernetwork $\mathcal{H}(\cdot; \psi)$ over 3000 iterations, respectively. For the loss weights, we set $\alpha = 1.0$ , $\beta = 0.1$ , $\lambda_{\mathrm{proj}} = 1.0$ , $\lambda_{\mathrm{slh}} = 1.0$ , and $\lambda_{\mathrm{reg}} = 0.01$ . + +# 4.2.2 Monocular 3D Object Detection + +To compare our method and Autolabels [33] with WeakM3D [23], we modify the architecture of WeakM3D so that it can be trained in a supervised manner using pseudo labels. More specifically, we train dimension and confidence heads in addition to the existing location and orientation heads using the same supervised loss as MonoDIS [25]. We call this model S-WeakM3D. For both WeakM3D and S-WeakM3D, ground truth 2D bounding boxes and instance masks are used for RoIAlign [11] and generating Li-DAR points on each object surface during training, respectively. During inference, we employ Cascade Mask R-CNN [5] with InternImage-XL [29] as the off-the-shelf 2D detector to provide 2D bounding boxes for RoIAlign. Please refer to the supplementary material for more details. + +# 4.3. Ablation Study + +# 4.3.1 Multi-View 3D Auto-Labeling + +We conduct an ablation study to demonstrate the effectiveness of each component in our proposed multi-view 3D auto-labeling. As can be seen from Tab. 1, the multi-view projection loss $\mathcal{L}_{\mathrm{proj}}$ serves as a practical baseline on its own. The multi-view silhouette loss $\mathcal{L}_{\mathrm{slh}}$ further boosts the quality of the pseudo labels a little bit, while the spatial gap + +![](images/6a7253053f0383cadc7a90dadfc27599a495c7d713ad3310eccc1bff1b4252d5.jpg) +Figure 5. Visualization results of the optimized 3D bounding boxes (1st row) and rendered instance masks (2nd row). We assign a unique color to each instance, and each pixel is colored as the weighted summation based on the rendered soft instance label. + +Table 3. Evaluation results of our proposed multi-view 3D auto-labeling on the KITTI-360 training set compared with the LiDAR-based monocular 3D auto-labeling proposed in Autolabels [33]. *Reproduced with the official code. + +
Conf.MethodAPBEV/AP3D@@0.3APBEV/AP3D@0.5
EasyHardEasyHard
≥0.0Autolabels*71.24/15.0967.33/11.4251.85/4.6546.10/2.92
VSRD75.03/68.5372.11/65.6447.12/35.2543.91/32.64
≥0.8Autolabels*75.56/17.1872.23/12.1958.64/5.7752.99/3.75
VSRD84.54/80.2581.66/77.3758.57/45.7655.09/44.17
+ +between the surfaces of each instance and the 3D bounding box limits further improvements. However, the residual distance field (RDF) can significantly boost the quality of the pseudo labels by addressing this problem. The systematic improvements with the inclusion of each component indicate their effectiveness, resulting in a precise auto-labeling system for weakly supervised 3D object detection. + +# 4.3.2 Confidence Assignment + +Due to the unreliability of the 3D bounding boxes optimized by the proposed auto-labeling for dynamic, occluded, or truncated objects, we conduct an ablation study to demonstrate the effectiveness of our proposed confidence assignment. Tab. 2 highlights substantial enhancement in detection performance when using the confidence-incorporated pseudo labels compared with the baseline, demonstrating the effectiveness of the proposed confidence assignment. + +# 4.4. Evaluation Results + +# 4.4.1 Multi-View 3D Auto-Labeling + +Autolabels [33] employs a similar two-stage framework consisting of auto-labeling and subsequent training of 3D object detectors using the pseudo labels. We compare our method with Autolabels to evaluate the quality of the pseudo labels. As our pseudo labels are unreliable for dy + +Table 4. Evaluation results of monocular 3D object detection on the KITTI-360 test set. *Reproduced with the official code. †CAD models are used as extra data. ${}^{ \ddagger }M\left( D\right)$ indicates that detection model $D$ is employed for model-agnostic method $M$ . + +
MethodWeak SupervisionFull SupervisionAPBEV/AP3D@0.3APBEV/AP3D@0.5
LiDARMasks3D BoxesEasyHardEasyHard
WeakM3D* [23]✓✓29.89/21.2524.01/15.348.10/2.962.96/2.01
Autolabels*†‡ [33] (S-WeakM3D)✓✓48.16/12.9237.34/9.9420.18/4.6914.33/2.79
VSRD‡ (S-WeakM3D)✓51.09/42.9441.28/33.7819.51/11.9114.39/8.46
VSRD‡ (MonoFlex)✓54.40/48.1645.67/40.0429.18/18.5322.31/13.60
VSRD‡ (MonoDETR)✓58.40/50.8650.61/43.4529.07/21.7722.83/16.46
MonoFlex* [35]✓69.70/67.0759.86/57.2650.82/43.1141.78/34.43
MonoDETR* [34]✓63.07/60.4954.04/50.0347.21/41.0136.05/30.38
+ +Table 5. Evaluation results of semi-supervised monocular 3D object detection on the KITTI validation set. + +
MethodRatioAPBEV/AP3D@0.7
EasyModerateHard
MonoDETR [34]1.0037.99/29.3626.76/20.6423.02/17.30
0.000.002/0.0010.004/0.0010.005/0.002
0.2531.72/21.7622.32/15.4318.86/12.55
VSRD (MonoDETR)0.5043.44/31.0531.54/21.4827.17/17.93
0.7542.58/32.9531.08/24.6827.19/21.38
+ +namic, occluded, or truncated objects, we group the dataset based on whether the average confidence score for each image is greater than a certain threshold. The evaluation results with the confidence thresholds of 0.0 and 0.8 are shown in Tab. 3. Our method exhibits superior performance in terms of $\mathrm{AP}_{\mathrm{BEV}}@0.3$ and $\mathrm{AP}_{\mathrm{3D}}$ . For $\mathrm{AP}_{\mathrm{BEV}}@0.5$ , our method exhibits slightly lower performance for the confidence threshold of 0.0. However, by raising the confidence threshold to 0.8, our method demonstrates superior performance, indicating that our method can generate more high-quality pseudo labels while abandoning low-quality pseudo labels. The optimized 3D bounding boxes and rendered instance masks are visualized in Fig. 5. + +# 4.4.2 Monocular 3D Object Detection + +The pseudo labels generated by the proposed auto-labeling serve as 3D supervision. We further investigate their applicability using the existing monocular 3D object detectors. + +Weakly Supervised Setting Tab. 4 shows the evaluation results of our method compared with the existing weakly supervised and fully supervised methods. Our method demonstrates a significant superiority over WeakM3D [23] across all the metrics while eliminating the need for LiDAR points for 3D supervision. Moreover, the detector trained on the pseudo labels generated by the proposed auto-labeling outperforms that trained on the pseudo labels generated by Autolabels [33]. Furthermore, employing more sophisticated + +monocular 3D object detectors such as MonoFlex [35] and MonoDETR [34] further improves detection performance, demonstrating the broad versatility of our method, which is not limited to a specific detection model. It is noteworthy that the detectors trained on the pseudo labels generated by our method demonstrate competitive performance compared with those trained in a fully supervised manner. + +Semi-Supervised Setting The essential advantage of our method is that it avoids costly 3D annotations, making more data available for training. Therefore, we investigate a realistic scenario where a detector pre-trained on a large amount of unlabeled data from a source domain is fine-tuned on a small amount of labeled data from a target domain. We select the KITTI-360 and KITTI datasets as source and target domains, respectively. Tab. 5 shows the performance of the detector pre-trained on the KITTI-360 dataset in a weakly supervised manner with the proposed auto-labeling and then fine-tuned on a subset of the KITTI dataset in a supervised manner. The zero-shot performance is quite low due to the characteristic that monocular depth estimation is greatly affected by the differences in camera parameters, but the performance of the detector fine-tuned on only $50\%$ of the labeled data significantly outperforms that trained on the whole data from scratch, highlighting the broad applicability of our method. + +# 5. Conclusion + +In this paper, we propose a novel weakly supervised 3D object detection framework named VSRD, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. Our method demonstrates superior performance compared with the existing weakly supervised 3D object detection methods. Moreover, it exhibits remarkable scalability using partially labeled data for semi-supervised learning. Our proposed method allows leveraging abundant 2D annotations to enhance 3D object detection without explicit 3D supervision, providing a promising avenue for further advancements in the field. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2 +[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2 +[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 7 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6 +[10] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[12] Benran Hu, Junkai Huang, Yichen Liu, Yu-Wing Tai, and Chi-Keung Tang. Nerf-rpn: A general framework for object detection in nerfs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23528–23538, 2023. 2 +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 + +[14] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[15] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 6 +[16] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 2 +[17] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In European Conference on computer vision, pages 515-531. Springer, 2020. 2 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 6 +[19] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7074-7082, 2017. 2 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[21] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2 +[22] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 1 +[23] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. arXiv preprint arXiv:2203.08332, 2022. 2, 7, 8 +[24] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2 +[25] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 1, 7 +[26] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular + +3d object detection using multi-view projection and direction consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17492, 2023. 2 +[27] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4, 5 +[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 1, 2 +[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14408-14419, 2023. 7 +[30] Chenfeng Xu, Bichen Wu, Ji Hou, Sam Tsai, Ruilong Li, Jialiang Wang, Wei Zhan, Zijian He, Peter Vajda, Kurt Keutzer, et al. Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23320-23330, 2023. 2, 3 +[31] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2, 3 +[32] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2 +[33] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 2, 7, 8 +[34] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 1, 2, 8 +[35] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 8 +[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, pages 12993-13000, 2020. 5 \ No newline at end of file diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/images.zip b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9f8cc58f2ff073743f13cdb5983dc9843a4e021 --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffbd77263a338ccccfa06c6e86c603555f69d724e97ac6ed31546109bc40496d +size 516176 diff --git a/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/layout.json b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..181e7db3ac59f974b2002b34bd56b41d2facbb53 --- /dev/null +++ b/2024/VSRD_ Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection/layout.json @@ -0,0 +1,9448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 473, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 473, + 140 + ], + "type": "text", + "content": "VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 161, + 226, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 161, + 226, + 204 + ], + "spans": [ + { + "bbox": [ + 79, + 161, + 226, + 204 + ], + "type": "text", + "content": "Zihua Liu* Tokyo Institute of Technology zliu@ok.sc.e.titech.ac.jp" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 161, + 341, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 161, + 341, + 202 + ], + "spans": [ + { + "bbox": [ + 253, + 161, + 341, + 202 + ], + "type": "text", + "content": "Hiroki Sakuma* T2 Inc. sakuma.h@t2.auto" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 367, + 161, + 514, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 161, + 514, + 204 + ], + "spans": [ + { + "bbox": [ + 367, + 161, + 514, + 204 + ], + "type": "text", + "content": "Masatoshi Okutomi Tokyo Institute of Technology mxo@sc.e.titech.ac.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "content": "Monocular 3D object detection poses a significant challenge in 3D scene understanding due to its inherently ill-posed nature in monocular depth estimation. Existing methods heavily rely on supervised learning using abundant 3D labels, typically obtained through expensive and labor-intensive annotation on LiDAR point clouds. To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each instance as a signed distance field (SDF) and render its silhouette as an instance mask through our proposed instance-aware volumetric silhouette rendering. To directly optimize the 3D bounding boxes through rendering, we decompose the SDF of each instance into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This mechanism enables us to optimize the 3D bounding boxes in an end-to-end manner by comparing the rendered instance masks with the ground truth instance masks. The optimized 3D bounding boxes serve as effective training data for 3D object detection. We conduct extensive experiments on the KITTI-360 dataset, demonstrating that our method outperforms the existing weakly supervised 3D object detection methods. The code is available at https://github.com/skmhrk1209/VSRD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": "3D object detection is one of the most critical components in the perception system for autonomous driving. With the recent success of deep learning in computer vision, numerous 3D object detection methods have been proposed to" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 313 + ], + "type": "image", + "image_path": "6a81a6ed70f7d663c5e24906782fde59b6f6c3aa7016ac521e90998dd2c9f2a4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 321, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 305, + 321, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 547, + 367 + ], + "type": "text", + "content": "Figure 1. Illustration of our proposed weakly supervised 3D object detection framework, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 509 + ], + "type": "text", + "content": "directly regress 3D bounding boxes from a LiDAR point cloud, multi-view images, or a monocular image. Among them, monocular 3D object detection is the most challenging in principle due to its inherently ill-posed nature in monocular depth estimation. Therefore, existing methods [22, 25, 28, 34, 35] heavily rely on supervised learning using abundant 3D labels manually annotated on LiDAR point clouds. This annotation cost is extremely high, posing a significant barrier to deploying 3D object detectors into autonomous driving systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "content": "To tackle this problem, we propose a novel weakly supervised 3D object detection framework named VSRD (Volumetric Silhouette Rendering for Detection) to train 3D object detectors without any 3D supervision but only weak 2D supervision. As illustrated in Fig. 1, VSRD consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. In the auto-labeling stage, we represent the surface of each 3D bounding box as a signed distance field (SDF) and render its silhouette as an instance mask through volumetric rendering. Comparing the rendered instance masks with the ground truth instance masks enables us to optimize the 3D bounding boxes directly. We introduce two novel mechanisms in this auto-labeling stage. The first is the instance-aware volumetric silhouette rendering that integrates instance labels along a ray to render the silhouette of each instance rather" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 258, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 258, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 258, + 714 + ], + "type": "text", + "content": "* Equal contribution. The order was determined by a coin flip." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17354" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "than the entire scene. This mechanism enables the silhouette of each instance to be rendered while considering geometric relationships among instances, such as occlusion. The second is the SDF decomposition, whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF) that represents the residual from the cuboid. This decomposition models the spatial gap between the surfaces of each instance and the 3D bounding box, enabling more accurate silhouette rendering and providing more reliable feedback signals during optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 312 + ], + "type": "text", + "content": "The 3D bounding boxes optimized by the proposed auto-labeling can serve as pseudo labels for training 3D object detectors. However, dynamic objects and inaccurate camera poses lead to low-quality pseudo labels, which negatively impact the training of 3D object detectors. Therefore, we propose a simple but effective algorithm to assign a confidence score representing the label quality to each pseudo label. We demonstrate that using these confidence scores as per-instance loss weights boosts the performance of the 3D object detectors trained on the pseudo labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 313, + 266, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 313, + 266, + 324 + ], + "spans": [ + { + "bbox": [ + 59, + 313, + 266, + 324 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 326, + 287, + 505 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 326, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 287, + 361 + ], + "type": "text", + "content": "- We propose a weakly supervised 3D object detection framework consisting of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 361, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 396 + ], + "type": "text", + "content": "- We propose a novel instance-aware volumetric silhouette rendering method whereby the silhouette of each instance can be rendered as an instance mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 397, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 287, + 433 + ], + "type": "text", + "content": "- We propose a novel SDF decomposition whereby the SDF of each instance is decomposed into the SDF of a cuboid and the residual distance field (RDF)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 434, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 468 + ], + "type": "text", + "content": "- We propose a simple but effective confidence assignment algorithm to incorporate the quality of each pseudo label into the training of 3D object detectors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 469, + 286, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 286, + 505 + ], + "type": "text", + "content": "- Extensive experiments on the KITTI-360 dataset demonstrate that our method outperforms the existing weakly supervised 3D object detection methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 518, + 134, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 134, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 134, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 538, + 219, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 219, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 219, + 551 + ], + "type": "text", + "content": "2.1. Monocular 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "Monocular 3D detection is a challenging task due to limited 3D information from monocular imagery. Deep3Box [19] pioneered this area by regressing relatively stable 3D object properties and combining these estimates with geometric constraints provided by the 2D bounding box. SMOKE [16] estimates 3D bounding boxes by combining keypoint estimates with regressed 3D box parameters. FCOS3D [28] employs a fully convolutional single-stage detector, transforming 7-DoF 3D targets to the image domain and decoupling them as 2D and 3D attributes. To enhance the performance of monocular 3D object detection for truncated objects, MonoFlex [35] explicitly decouples truncated objects and adaptively combines multiple approaches for ob" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "ject depth estimation. M3D-RPN [4] and MonoDETR [34] also explored the usage of depth cues to improve monocular 3D object detection. The former designed depth-aware convolutional layers that enable location-specific feature extraction and consequently improved 3D scene understanding, while the latter modified the vanilla Transformer [7] to be depth-aware to guide the whole detection process by contextual depth cues. Despite these advances, the reliance on expensive and labor-intensive manual annotation on LiDAR point clouds remains a significant limitation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 200, + 517, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 517, + 213 + ], + "type": "text", + "content": "2.2. Weakly Supervised 3D Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 219, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 519 + ], + "type": "text", + "content": "Many weakly supervised methods have been proposed to mitigate the high annotation cost in 3D object detection. WS3D [17] introduced a LiDAR-based two-stage pipeline where cylindrical object proposals are first generated under weak supervision and then refined using a few labeled object instances. VS3D [24] introduced an unsupervised 3D proposal module that generates object proposals by leveraging normalized point cloud densities. WeakM3D [23] introduced a weakly supervised monocular 3D object detection method that leverages the 3D alignment loss between each predicted 3D bounding box and corresponding RoI LiDAR points. Furthermore, it introduced a method to estimate the orientation from RoI LiDAR points based on its statistics. Recently, WeakMono3D [26] leverages projection loss with multi-view and direction consistency, achieving a weakly supervised monocular object detection that relies only on 2D supervision. However, its reliance on 2D direction annotations restricts its applicability to large-scale datasets. Zakharov et al. [33] proposed an auto-labeling pipeline that integrates an SDF-based differentiable shape renderer and normalized object coordinate spaces (NOCS). However, additional training on synthetic data is still required for shape and coordinate estimation. In contrast, our method purely relies on 2D supervision, eliminating the necessity of synthetic data or 3D supervision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 526, + 514, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 514, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 514, + 540 + ], + "type": "text", + "content": "2.3. 3D Object Detection with Neural Fields" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [18] introduced a new perspective for implicit learning of 3D geometry from posed multi-view images by volume rendering. Building upon the vanilla NeRF, subsequent research has focused on enhancing novel view synthesis [1-3, 21] or accelerating volume rendering [6, 8, 20]. NeuS [27] introduced a novel volume rendering scheme to learn a neural SDF representation by introducing a density distribution induced by the SDF. Similarly, VolSDF [32] defined the volume density function as Laplace's cumulative distribution function applied to an SDF representation. Many works [12, 30, 31] recently have attempted to utilize neural fields for 3D object detection. Notably, NeRF-RPN [12] demonstrated that the 3D bounding boxes of objects in NeRF can be directly regressed with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17355" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 545, + 343 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 343 + ], + "type": "image", + "image_path": "779429e0437a8ab701f8bc8482c50fccecc89a93d4c4e9aea1bf27decc31d3d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 547, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 547, + 396 + ], + "type": "text", + "content": "Figure 2. Illustration of the pipeline of our proposed multi-view 3D auto-labeling. We represent the surface of each instance as an SDF and decompose it into the SDF of a 3D bounding box and the residual distance field (RDF), which is learned via a hypernetwork. The composed instance SDF is used to render the silhouette of the instance through our proposed instance-aware volumetric silhouette rendering. All the 3D bounding boxes are optimized based on the loss between the rendered and ground truth instance masks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 415, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 559 + ], + "type": "text", + "content": "out rendering. NeRF-Det [30] connects the detection and NeRF branches through a shared MLP, enabling an efficient adaptation of NeRF to detection and yielding geometry-aware volumetric representations. MonoNeRD [31] models scenes with SDFs and renders RGB images and depth maps through volume rendering to obtain intermediate 3D representations for detection. However, these approaches still rely on ground truth 3D labels for supervision. In contrast, we propose the first volume rendering-based weakly supervised 3D object detection framework that relies on multiview geometry and 2D supervision without any 3D supervision, such as 3D bounding boxes or LiDAR point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 576, + 104, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 104, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 104, + 588 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 211, + 611 + ], + "type": "text", + "content": "3.1. Multi-View 3D Auto-Labeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": "The pipeline of our proposed multi-view 3D auto-labeling is illustrated in Fig. 2. In Sec. 3.1.1, we review the SDF-based volume rendering formulation, which is the foundation of our method. In Sec. 3.1.2, we define the optimization problem for the multi-view 3D auto-labeling. In Sec. 3.1.3, we introduce the SDF for cuboids, which is used to optimize 3D bounding boxes through rendering. In Sec. 3.1.4, we introduce a novel neural field named residual distance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 415, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 489 + ], + "type": "text", + "content": "field to fill the spatial gap between the surfaces of each instance and the 3D bounding box. In Sec. 3.1.5, we introduce instance-aware volumetric silhouette rendering to render the silhouette of each instance based on its SDF. In Sec. 3.1.6, we introduce loss functions using instance masks and 2D bounding boxes as weak supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 502, + 398, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 398, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 398, + 513 + ], + "type": "text", + "content": "3.1.1 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "text", + "content": "SDF-based Volumetric Rendering NeRF [18] represents a 3D scene with neural density and color fields. Given a camera position " + }, + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "text", + "content": " and a ray direction " + }, + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{d} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 521, + 547, + 582 + ], + "type": "text", + "content": " emitted from a pixel, the volume rendering scheme integrates the colors of sampled points along the ray as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 589, + 545, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 589, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 336, + 589, + 545, + 615 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {C}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) \\mathrm {d} t, \\tag {1}", + "image_path": "d5ad81ef661a954ec466031d1bb33e4fb76cf5700f448159fb6bbbe236bc688f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 616, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 616, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 351, + 616, + 545, + 644 + ], + "type": "interline_equation", + "content": "w (t) = \\exp (- \\int_ {0} ^ {t} \\sigma (\\boldsymbol {r} (u)) \\mathrm {d} u) \\sigma (\\boldsymbol {r} (t)), \\qquad (2)", + "image_path": "b07d0395cb79f43258f0acf6d47c61d80cf0442c364302387ff44db67b162747.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{C} (\\pmb {o},\\pmb {d})\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": " denotes the rendered color of the pixel, " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol {r}(t) = \\boldsymbol {o} + t\\boldsymbol{d}" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": " denotes the ray, " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "c(\\pmb {p},\\pmb {d})" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": " denotes the color at the position " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sigma (\\pmb {p})" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": " denotes the volume density at the position " + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 652, + 547, + 715 + ], + "type": "text", + "content": ". Since the density field cannot represent the surfaces explicitly, novel SDF-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17356" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "based volume rendering formulation has been introduced in NeuS [27], where a surface represented by an SDF is reinterpreted as a participating medium represented by a density field, enabling the surface to be rendered through volume rendering. In this formulation, the weight " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " in Eq. (2) is re-written by introducing opaque density " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 165, + 287, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 165, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 97, + 165, + 287, + 193 + ], + "type": "interline_equation", + "content": "w (t) = \\exp \\left(- \\int_ {0} ^ {t} \\rho (u) \\mathrm {d} u\\right) \\rho (t), \\tag {3}", + "image_path": "64c452875cc58d1fb6ed70bb48b5e1c59ebaa1046293c55b16f6999f17e6d1e5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 194, + 287, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 194, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 100, + 194, + 287, + 224 + ], + "type": "interline_equation", + "content": "\\rho (t) = \\max \\left(- \\frac {\\frac {\\mathrm {d} \\Phi}{\\mathrm {d} t} \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}{\\Phi \\left(\\hat {\\mathcal {F}} (\\boldsymbol {r} (t))\\right)}, 0\\right), \\tag {4}", + "image_path": "eaa7237ef1de25b7418247e4743a1b2a5397eace42c64e6de2f217e55076a7ac.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}} (\\cdot)" + }, + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "text", + "content": " denotes the SDF for the entire scene and " + }, + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\Phi (\\cdot)" + }, + { + "bbox": [ + 46, + 236, + 287, + 308 + ], + "type": "text", + "content": " denotes the Sigmoid function. Our proposed instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5 is based on the same SDF-based weight formulation as Eq. (3) but integrates instance labels instead of colors along a ray to render instance masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 331, + 162, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 162, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 162, + 342 + ], + "type": "text", + "content": "3.1.2 Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": "Given a monocular video consisting of posed frames, each frame annotated with instance masks, our goal is to optimize the 3D bounding box frame by frame without 3D supervision. More specifically, for each target frame " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " in the video, we sample multiple source frames " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " and optimize the " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " 3D bounding boxes in the target frame using the instance masks of the source frames as weak supervision, where " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " denotes the number of instances in the target frame. Please refer to the supplementary material for how to sample " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ". We parameterize the " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": "-th 3D bounding box " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{B}_n \\in \\mathbb{R}^{8 \\times 3}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " in the target frame with a dimension " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{d}_n \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", location " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{\\ell}_n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", and orientation " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_n \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", which is the rotation angle in the bird's-eye-view. In addition to these parameters for each bounding box, we prepare a learnable instance embedding " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "z_n \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " for each instance and a shared hypernetwork parameterized by " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " for the residual distance field introduced in Sec. 3.1.4. We stack each parameter group into a single tensor over all the instances, yielding dimensions " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{D} \\in \\mathbb{R}_+^{N \\times 3}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", locations " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{L} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", orientations " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", and instance embeddings " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ". Given the loss function " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " explained in Sec. 3.1.6, we optimize " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 353, + 289, + 620 + ], + "type": "text", + "content": " via stochastic gradient descent as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 643, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 287, + 666 + ], + "type": "interline_equation", + "content": "{ } ^ { * } \\boldsymbol { D } , { } ^ { * } \\boldsymbol { L } , { } ^ { * } \\boldsymbol { \\Theta } , { } ^ { * } \\boldsymbol { Z } , { } ^ { * } \\boldsymbol { \\psi } = \\operatorname* { a r g m i n } _ { \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } } \\mathcal { L } ( \\hat { \\boldsymbol { D } } , \\hat { \\boldsymbol { L } } , \\hat { \\boldsymbol { \\Theta } } , \\boldsymbol { Z } , \\boldsymbol { \\psi } ) . \\tag {5}", + "image_path": "ad83fc240ee347ab7412dec978d4eaebba5f59fba6d9098bfe10cefa8b33583d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "The optimized 3D bounding boxes " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "{}^{*}\\pmb{B}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " decoded from " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "{}^{*}\\pmb{D}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "{}^{*}\\pmb{L}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "{}^{*}\\Theta" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " can be used as pseudo labels for training 3D object detectors, as explained in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 516, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 516, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 516, + 84 + ], + "type": "text", + "content": "3.1.3 3D Bounding Box Represented as an SDF" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": "To optimize the 3D bounding boxes in each target frame by solving Eq. (5) through rendering, we represent the surface of each 3D bounding box as a signed distance field (SDF), which is one of the most common surface representations. An SDF is represented as a function " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:\\mathbb{R}^3\\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": " that maps a spatial position " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\pmb {p}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": " to its signed distance to the closest point on the surface, indicating that the zero-level set " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\{\\pmb {p}\\in \\mathbb{R}^3\\mid \\mathcal{F}(\\pmb {p}) = 0\\}" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": " represents the surface itself. The SDF for a cuboid parameterized by a dimension " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "d\\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": " location " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": ", and orientation " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\pmb {R}\\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": " can be derived theoretically and is denoted by " + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\cdot ;d,\\ell ,\\pmb {R})" + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": ". Please refer to the supplementary material for how to derive this formula." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 248, + 440, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 440, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 440, + 259 + ], + "type": "text", + "content": "3.1.4 Residual Distance Field" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": "In general, the shape of each instance is not a cuboid. Therefore, if we leverage only the cuboid SDF introduced in Sec. 3.1.3 to render the surface of each instance, it is no longer possible to render an accurate silhouette due to the spatial gap between the surfaces of the instance and the 3D bounding box, leading to unreliable feedback signals during optimization. Therefore, we propose a novel neural field named residual distance field (RDF), which models the residual between the signed distances to the surfaces of the instance and the 3D bounding box. Let " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_n(\\cdot)" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " be the true SDF of the surface bounded by the " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": "-th 3D bounding box " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{B}}_n" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " whose SDF is given by " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}_n(\\cdot) = \\mathcal{B}(\\cdot; \\hat{\\pmb{d}}_n, \\hat{\\ell}_n, \\pmb{R}_y(\\theta_n))" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{R}_y(\\theta)" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " denotes the rotation matrix around the " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": "-axis by an angle " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": ". For any point " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": ", we define the RDF as " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}_n(\\pmb{p}) := \\mathcal{F}_n(\\pmb{p}) - \\hat{\\mathcal{B}}_n(\\pmb{p})" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": ". Here, based on the definition of a 3D bounding box that it encloses the corresponding instance, " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\forall \\pmb{p} \\in \\mathbb{R}^3: \\hat{\\mathcal{R}}_n(\\pmb{p}) \\geq 0" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " is required. One straightforward way to model the RDF for each instance is to train " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " individual networks. However, as objects belonging to the same semantic class often have similar geometric shapes, we train a single hypernetwork [10] that regresses the weights of the neural RDF directly from an instance embedding instead. Given the " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": "-th instance embedding " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{z}_n \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": " denotes the number of dimensions, the " + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 267, + 547, + 567 + ], + "type": "text", + "content": "-th neural RDF is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 571, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 571, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 375, + 571, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {R}} _ {n} (\\boldsymbol {p}) = \\sigma \\left(\\mathcal {G} (\\boldsymbol {p}; \\phi_ {n})\\right), \\tag {6}", + "image_path": "338cb87f66af626ff391bb9b4e1ee4536c00aaf0dbc419393bf3a561c37ae7e0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 402, + 587, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 587, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 402, + 587, + 545, + 600 + ], + "type": "interline_equation", + "content": "\\phi_ {n} = \\mathcal {H} \\left(\\boldsymbol {z} _ {n}; \\psi\\right), \\tag {7}", + "image_path": "afedc3b56b80da2223213c6b4d99f6f2379689d8df7e2db997b2c8edc33f618b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot ;\\phi_n)" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " -th neural network parameterized by " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{n}" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot ;\\psi)" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " denotes the shared hypernetwork parameterized by " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " denotes the Softplus function to force the residual distance to be positive based on the definition of a bounding box. For the instance-aware volumetric silhouette rendering introduced in Sec. 3.1.5, we employ " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_n(\\pmb {p}) = \\hat{\\mathcal{B}}_n(\\pmb {p}) + \\hat{\\mathcal{R}}_n(\\pmb {p})" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " as the SDF of each instance that represents an arbitrary surface bounded by the 3D bounding box " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{B}_n" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " for more accurate silhouette rendering." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17357" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 74, + 286, + 190 + ], + "blocks": [ + { + "bbox": [ + 48, + 74, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 48, + 74, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 74, + 286, + 190 + ], + "type": "image", + "image_path": "55548fd88ddc25c3267c6a6fcc528680246f8ded7476167037ed917c72946d02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 288, + 255 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed instance-aware volumetric silhouette rendering. The instance labels are averaged for each sampled point along a ray based on the signed distance to each instance. The averaged instance labels are integrated along the ray based on the SDF-based volume rendering formulation [27]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 285, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 285, + 289 + ], + "type": "text", + "content": "3.1.5 Instance-Aware Volumetric Silhouette Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 392 + ], + "type": "text", + "content": "This section details our approach to optimizing 3D bounding boxes through volumetric rendering. The core idea is to render instance masks and compare them with ground truth instance masks. To achieve this, we propose a novel SDF-based instance-aware volumetric silhouette rendering, where instance labels instead of colors are integrated along a ray based on the same SDF-based volume rendering formulation as Eq. (3) as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 400, + 287, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 400, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 77, + 400, + 287, + 427 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {S}} (\\boldsymbol {o}, \\boldsymbol {d}) = \\int_ {0} ^ {\\infty} w (t) \\boldsymbol {s} (\\boldsymbol {r} (t)) d t, \\tag {8}", + "image_path": "1fac6ed1f293ef732f01d7db84aec7f8955be5a1659ebb385845ca806ee75a0c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 429, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 429, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 429, + 287, + 460 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} (\\boldsymbol {p}) = \\sum_ {n = 1} ^ {N} \\operatorname {s o f t m i n} \\left(\\left[ \\hat {\\mathcal {F}} _ {n} (\\boldsymbol {p}) \\right] _ {n = 1} ^ {N}\\right) _ {n} \\cdot \\boldsymbol {y} _ {n}, \\tag {9}", + "image_path": "0db3829a861b3136499b64198ca34e3e23c77ca2cd315fcfec2c50ec574cceed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{S} (\\pmb {o},\\pmb {d})\\in [0,1]^N" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " denotes the rendered soft instance label, " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "s(\\pmb {p})\\in [0,1]^N" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " denotes the weighted average instance label at the position " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " indicating how relatively close the position " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " is to each instance, and " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb {y}_n\\in \\{0,1\\} ^N" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " denotes the one-hot instance label of the " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " -th instance. " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_n(\\cdot)" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " denotes the SDF of the " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " -th instance introduced in Sec. 3.1.4. " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " in Eq. (8) is derived from Eq. (3). To compute Eq. (3), we model the entire scene as the union of all the object surfaces, i.e., " + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}} (\\pmb {p}) = \\min (\\hat{\\mathcal{F}}_1(\\pmb {p}),\\dots ,\\hat{\\mathcal{F}}_N(\\pmb {p}))" + }, + { + "bbox": [ + 46, + 472, + 287, + 605 + ], + "type": "text", + "content": " . This mechanism enables us to render instance masks considering the geometric relationships among instances, such as occlusions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 144, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 144, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 144, + 633 + ], + "type": "text", + "content": "3.1.6 Loss Functions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "We optimize the parameters of the 3D bounding boxes " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\Omega} = \\{\\hat{D},\\hat{L},\\hat{\\Theta}\\}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " in the target frame along with the instance embeddings " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " and parameter " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " of the hypernetwork " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot ;\\psi)" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ". The final loss " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " is defined as the combination of the multi-view projection loss " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ", multi-view silhouette loss " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sh}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ", and Eikonal regularization " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " [27] as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 90, + 545, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 90, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 326, + 90, + 545, + 122 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\hat {\\Omega}, Z, \\psi) = \\lambda_ {\\mathrm {p r o j}} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\Omega}) + \\lambda_ {\\mathrm {s l h}} \\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\Omega}, Z, \\psi) \\\\ + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi), \\tag {10} \\\\ \\end{array}", + "image_path": "9e4f0a727ee65788427f5493421b976a16f9f0f860fc614d24275b7b766a6a63.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{slh}}" + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}}" + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": " denote the loss weights. However, although the multi-view projection loss " + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": " and multi-view silhouette loss " + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{slh}}" + }, + { + "bbox": [ + 305, + 129, + 545, + 273 + ], + "type": "text", + "content": " are based on 2D supervision, the 3D-2D correspondence to define the losses is not obvious. Therefore, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm [14]. Each pair-wise matching cost is defined as the projection loss introduced in this section, but it is computed for not all the source frames but only the target frame. We assume that the ground truth 2D bounding boxes and instance masks have already been reordered based on the optimal permutation for simplicity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 286, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 394 + ], + "type": "text", + "content": "Multi-View Projection Loss Although introducing the residual distance field enables us to model an arbitrary surface bounded by a 3D bounding box, it introduces another problem: the 3D bounding box can grow indefinitely without any constraint. To address this problem, we employ the ground truth 2D bounding box as a constraint to keep the 3D bounding box tight. The multi-view projection loss " + }, + { + "bbox": [ + 305, + 286, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 286, + 545, + 394 + ], + "type": "text", + "content": " is defined as the average distance between the projected and ground truth 2D bounding boxes as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 400, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 400, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 341, + 400, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p r o j}} (\\hat {\\boldsymbol {\\Omega}}) = \\alpha \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\| \\hat {\\boldsymbol {B}} _ {i n} ^ {2 \\mathrm {D}} - \\boldsymbol {B} _ {i n} ^ {2 \\mathrm {D}} \\| _ {\\mathrm {H}} \\\\ - \\beta \\sum_ {i \\in \\mathcal {S}} \\sum_ {n = 1} ^ {N} \\mathrm {D I o U} \\left(\\hat {\\boldsymbol {B}} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}\\right), \\tag {11} \\\\ \\end{array}", + "image_path": "33190076dba20c017dd099b04d86a36793b90ec6a07af8b7e95c724ec24f5c2b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\mathrm{H}}" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " denotes the Huber loss, " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{DIOU}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " denotes the Distance-IoU [36], and " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " denote balancing coefficients. The projected 2D bounding box " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{B}_{in}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " is defined as the rectangle with minimal area enclosing the projected vertices " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{V}}_{in}^{2\\mathrm{D}}\\propto \\hat{\\pmb{B}}_n\\pmb {E}_i^T\\pmb {K}_i^T" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\pmb {E}_i" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\pmb {K}_i" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " denote the extrinsic and intrinsic matrices for frame " + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 475, + 545, + 549 + ], + "type": "text", + "content": " , respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "text", + "content": "Multi-View Silhouette Loss The multi-view silhouette loss is defined as the cross entropy between the rendered and ground truth instance masks. Since the spatial gap between the surfaces of each instance and the 3D bounding box is modeled by the residual distance field introduced in Sec. 3.1.4, the 3D bounding box constrained to tightly fit the ground truth 2D bounding box by the multi-view projection loss " + }, + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "text", + "content": " is further refined by the multi-view silhouette loss " + }, + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sh}}" + }, + { + "bbox": [ + 305, + 562, + 545, + 670 + ], + "type": "text", + "content": ", which is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 677, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 677, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 321, + 677, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s l h}} (\\hat {\\boldsymbol {\\Omega}}, \\boldsymbol {Z}, \\psi) = \\sum_ {i \\in S} \\sum_ {j = 1} ^ {R _ {i}} \\operatorname {C E} \\left(\\hat {\\boldsymbol {S}} \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {d} _ {i j}\\right), \\boldsymbol {S} _ {i j}\\right), \\tag {12}", + "image_path": "c49476476c04a063ef79fadf93cc8d71af10992d3c430448f544e7237e0df0f7.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17358" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 166, + 129 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 166, + 129 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 166, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 166, + 129 + ], + "type": "image", + "image_path": "14332f1df406c5da578b10d25d2c580561d7770afa185dadc569b2809c83d25b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 132, + 133, + 140 + ], + "lines": [ + { + "bbox": [ + 81, + 132, + 133, + 140 + ], + "spans": [ + { + "bbox": [ + 81, + 132, + 133, + 140 + ], + "type": "text", + "content": "(a) Static Scene" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 286, + 129 + ], + "type": "image", + "image_path": "cea7b8334a9ec2e7bace0dafbaa265073467806aa144829af9b5184f9199cbfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 132, + 257, + 140 + ], + "lines": [ + { + "bbox": [ + 194, + 132, + 257, + 140 + ], + "spans": [ + { + "bbox": [ + 194, + 132, + 257, + 140 + ], + "type": "text", + "content": "(b) Dynamic Scene" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 144, + 287, + 188 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 188 + ], + "type": "text", + "content": "Figure 4. Comparison of the confidence scores in static and dynamic scenes. It can be seen that the confidence scores are lower for dynamic, occluded, or truncated objects, indicating less influence on the subsequent training of 3D object detectors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\boldsymbol{o}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\boldsymbol{d}_{ij} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "S_{ij} \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": " denote the camera position, ray direction, and ground truth instance label at the " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": "-th sampled pixel in frame " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\hat{S}(\\boldsymbol{o}_i, \\boldsymbol{d}_{ij})" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": " denotes the rendered instance label based on Eq. (8). " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": " denotes the cross entropy loss. " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": " denotes the number of rays sampled for frame " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": ". Please refer to the supplementary material for how to determine " + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 46, + 209, + 286, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 301, + 220, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 220, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 220, + 314 + ], + "type": "text", + "content": "3.2. Training of 3D Object Detectors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 319, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 404 + ], + "type": "text", + "content": "Once the 3D bounding boxes are optimized by the proposed multi-view 3D auto-labeling, they can serve as pseudo labels for training 3D object detectors. To incorporate the quality of each pseudo label into the training of 3D object detectors, we introduce a simple but effective confidence assignment method in Sec. 3.2.1 and confidence-based weighted loss for bounding box regression in Sec. 3.2.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 420, + 180, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 180, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 180, + 432 + ], + "type": "text", + "content": "3.2.1 Confidence Assignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": "As shown in Fig. 4, the 3D bounding boxes optimized by the proposed auto-labeling are not reliable for dynamic, occluded, or truncated objects. Therefore, we propose a confidence assignment method based on the multi-view projection loss. First, for each target frame " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": ", we identify a set of source frames " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": " such that all the instances in the target frame are visible from every frame in the set. Next, we find an optimal bipartite matching between the optimized 3D bounding boxes and ground truth 2D bounding boxes by the Hungarian algorithm. The cost matrix " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": " is defined as the pairwise IoUs between the projected and ground truth 2D bounding boxes averaged over all the source frames, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 602, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 602, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 86, + 602, + 287, + 631 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Q} _ {n m} = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} 1 - \\operatorname {I o U} \\left(^ {*} \\boldsymbol {B} _ {i n} ^ {\\mathrm {2 D}}, \\boldsymbol {B} _ {i m} ^ {\\mathrm {2 D}}\\right), \\tag {13}", + "image_path": "3274dd3029ffb9b4e6ce1f65102ea4f4b5d0e7fd195aa1b1742593ae2f05b6c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "{}^{*}B_{in}^{2D}" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": "-th projected 2D bounding box in frame " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": ", which is obtained by projecting the " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": "-th optimized 3D bounding box " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "{}^{*}B_{n}" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": " onto frame " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "{B_{im}^{2D}}" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": "-th ground truth 2D bounding box in frame " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": ". Once an optimal permutation matrix " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{P} \\in \\{0,1\\}^{N \\times N}" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": " is obtained, the confidence scores " + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "inline_equation", + "content": "{}^{*}\\pmb{C} \\in [0,1]^N" + }, + { + "bbox": [ + 46, + 640, + 287, + 715 + ], + "type": "text", + "content": " for the optimized 3D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "text", + "content": " bounding boxes " + }, + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "inline_equation", + "content": "^* B" + }, + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "text", + "content": " are given as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 88, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 88, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 354, + 88, + 545, + 118 + ], + "type": "interline_equation", + "content": "^ {*} C = \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} \\operatorname {I o U} \\left(^ {*} B _ {i} ^ {\\mathrm {2 D}}, P B _ {i} ^ {\\mathrm {2 D}}\\right). \\tag {14}", + "image_path": "293fdf7357b51547bad814770370416e891fd5be70c65ab0ce251e99ba7e508e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 129, + 479, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 479, + 141 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 479, + 141 + ], + "type": "text", + "content": "3.2.2 Confidence-based Weighted Loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "text", + "content": "The 3D bounding boxes optimized by the proposed multiview 3D auto-labeling serve as pseudo labels for most 3D object detectors, which leverage 3D bounding boxes and semantic class labels as supervision, without any modification of the architectures, loss functions, and training procedures, except for the confidence-based loss weighting. We incorporate the confidence scores computed by Eq. (14) into only the regression loss. Given a regression loss function " + }, + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "text", + "content": " its confidence-based weighted version " + }, + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 148, + 545, + 257 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 261, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {L}} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}}, ^ {*} \\boldsymbol {B}, ^ {*} \\boldsymbol {C}\\right) = \\sum_ {m = 1} ^ {M} ^ {*} \\boldsymbol {C} _ {\\pi (m)} \\mathcal {L} _ {\\mathrm {b o x}} \\left(\\hat {\\boldsymbol {B}} _ {m}, ^ {*} \\boldsymbol {B} _ {\\pi (m)}\\right), \\tag {15}", + "image_path": "025aa4d2906aa62e4ec7c19bf1d436d4db3e297eefc91d7edc9e717bd693af34.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": " denotes the predicted 3D bounding boxes, " + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "inline_equation", + "content": "^* B" + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": " denotes the 3D bounding boxes optimized by the proposed auto-labeling, and " + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "inline_equation", + "content": "^* C" + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": " denotes the corresponding confidence scores. " + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": " denotes the number of positive anchors, and " + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 304, + 300, + 545, + 373 + ], + "type": "text", + "content": " denotes the label assigner that maps the index of an anchor to that of the matched ground truth." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 381, + 386, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 386, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 386, + 395 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 401, + 364, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 364, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 364, + 412 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 418, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 539 + ], + "type": "text", + "content": "We use the KITTI-360 [15] dataset for our experiments, splitting it into training (43,855 images), validation (1,173 images), and test sets (2,531 images). We follow the same evaluation protocol as the KITTI dataset [9]. However, as occlusion and truncation labels are not available for the KITTI-360 dataset, unlike the KITTI dataset, we consider only two difficulty levels, namely Easy and Hard, based on whether the height of each ground truth 2D bounding box is greater than 40 and 25 in pixels, respectively. Following prior works, we evaluate our method on only category Car." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 544, + 438, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 438, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 438, + 557 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 563, + 466, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 466, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 466, + 575 + ], + "type": "text", + "content": "4.2.1 Multi-View 3D Auto-Labeling" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We sample 16 source frames for each target frame. We sample 1000 rays across all the source frames, i.e., " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sum_{i\\in S}R_i = 1000" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", at each iteration based on the ground truth instance masks. Please refer to the supplementary material for more details. We employ the same hierarchical volume sampling as NeRF [18] and sample 100 query points for both coarse and fine sampling. Unlike NeRF, both coarse and fine samples are drawn from the single scene SDF rather than two distinct ones. We set the number of dimensions of each instance embedding as " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "D = 256" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". The neural RDF " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and hypernetwork " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " are implemented as MLPs with four hidden" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17359" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 113, + 287, + 179 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 104 + ], + "type": "text", + "content": "Table 1. Ablation study on the KITTI-360 training set to verify the effectiveness of each component in our proposed multi-view 3D auto-labeling. Just one randomly selected sequence is used." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 113, + 287, + 179 + ], + "lines": [ + { + "bbox": [ + 50, + 113, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 50, + 113, + 287, + 179 + ], + "type": "table", + "html": "
ComponentsAPBEV/AP3D@0.3APBEV/AP3D@0.5@0.5
LprojLslhRDFEasyHardEasyHard
✓60.77/54.8863.99/57.6637.38/23.3337.44/24.82
✓✓63.84/56.1160.86/56.8741.73/26.8439.22/25.58
✓✓✓73.84/66.6473.22/66.3246.35/31.1143.07/30.16
", + "image_path": "d44ab2bbdb4a32ea061c5929f3193b182db1dbc4db9dde4397929c37664caaf2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 233, + 282, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 191, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 47, + 191, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 287, + 226 + ], + "type": "text", + "content": "Table 2. Ablation study on the KITTI-360 test set to verify the effectiveness of the confidence scores against monocular 3D object detection. S-WeakM3D is used as a monocular 3D object detector." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 233, + 282, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 233, + 282, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 233, + 282, + 291 + ], + "type": "table", + "html": "
Conf.APBEV/AP3D@0.3APBEV/AP3D@0.5
EasyHardEasyHard
✓38.67/31.6231.25/23.8810.95/8.255.37/4.41
51.09/42.9441.27/33.9719.50/11.9114.39/8.46
", + "image_path": "ddc646b7a2b13fac8eea7923281c0c3b2357ca160c6c2b9e2f34dfe1001ec31e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": "layers, each of which has 256 and 16 channels, respectively. We use the Adam optimizer [13] and the learning rates are decayed exponentially from " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " for the box parameters " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\Omega}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", instance embeddings " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", and parameter " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " of the hypernetwork " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot; \\psi)" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " over 3000 iterations, respectively. For the loss weights, we set " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{proj}} = 1.0" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{slh}} = 1.0" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}} = 0.01" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 216, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 216, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 216, + 419 + ], + "type": "text", + "content": "4.2.2 Monocular 3D Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "content": "To compare our method and Autolabels [33] with WeakM3D [23], we modify the architecture of WeakM3D so that it can be trained in a supervised manner using pseudo labels. More specifically, we train dimension and confidence heads in addition to the existing location and orientation heads using the same supervised loss as MonoDIS [25]. We call this model S-WeakM3D. For both WeakM3D and S-WeakM3D, ground truth 2D bounding boxes and instance masks are used for RoIAlign [11] and generating Li-DAR points on each object surface during training, respectively. During inference, we employ Cascade Mask R-CNN [5] with InternImage-XL [29] as the off-the-shelf 2D detector to provide 2D bounding boxes for RoIAlign. Please refer to the supplementary material for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 603, + 141, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 141, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 141, + 616 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 208, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 208, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 208, + 634 + ], + "type": "text", + "content": "4.3.1 Multi-View 3D Auto-Labeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "We conduct an ablation study to demonstrate the effectiveness of each component in our proposed multi-view 3D auto-labeling. As can be seen from Tab. 1, the multi-view projection loss " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " serves as a practical baseline on its own. The multi-view silhouette loss " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{slh}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " further boosts the quality of the pseudo labels a little bit, while the spatial gap" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 173 + ], + "type": "image", + "image_path": "6a7253053f0383cadc7a90dadfc27599a495c7d713ad3310eccc1bff1b4252d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 225 + ], + "type": "text", + "content": "Figure 5. Visualization results of the optimized 3D bounding boxes (1st row) and rendered instance masks (2nd row). We assign a unique color to each instance, and each pixel is colored as the weighted summation based on the rendered soft instance label." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 289, + 543, + 367 + ], + "blocks": [ + { + "bbox": [ + 304, + 236, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 236, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 546, + 281 + ], + "type": "text", + "content": "Table 3. Evaluation results of our proposed multi-view 3D auto-labeling on the KITTI-360 training set compared with the LiDAR-based monocular 3D auto-labeling proposed in Autolabels [33]. *Reproduced with the official code." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 289, + 543, + 367 + ], + "lines": [ + { + "bbox": [ + 309, + 289, + 543, + 367 + ], + "spans": [ + { + "bbox": [ + 309, + 289, + 543, + 367 + ], + "type": "table", + "html": "
Conf.MethodAPBEV/AP3D@@0.3APBEV/AP3D@0.5
EasyHardEasyHard
≥0.0Autolabels*71.24/15.0967.33/11.4251.85/4.6546.10/2.92
VSRD75.03/68.5372.11/65.6447.12/35.2543.91/32.64
≥0.8Autolabels*75.56/17.1872.23/12.1958.64/5.7752.99/3.75
VSRD84.54/80.2581.66/77.3758.57/45.7655.09/44.17
", + "image_path": "1532242ef4a982981301b465e11a391d504016c7014e4bd9aa1194064badcd34.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 381, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 467 + ], + "type": "text", + "content": "between the surfaces of each instance and the 3D bounding box limits further improvements. However, the residual distance field (RDF) can significantly boost the quality of the pseudo labels by addressing this problem. The systematic improvements with the inclusion of each component indicate their effectiveness, resulting in a precise auto-labeling system for weakly supervised 3D object detection." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 481, + 440, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 440, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 440, + 493 + ], + "type": "text", + "content": "4.3.2 Confidence Assignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 500, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 597 + ], + "type": "text", + "content": "Due to the unreliability of the 3D bounding boxes optimized by the proposed auto-labeling for dynamic, occluded, or truncated objects, we conduct an ablation study to demonstrate the effectiveness of our proposed confidence assignment. Tab. 2 highlights substantial enhancement in detection performance when using the confidence-incorporated pseudo labels compared with the baseline, demonstrating the effectiveness of the proposed confidence assignment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 616, + 416, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 416, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 416, + 628 + ], + "type": "text", + "content": "4.4. Evaluation Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 634, + 466, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 466, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 466, + 647 + ], + "type": "text", + "content": "4.4.1 Multi-View 3D Auto-Labeling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Autolabels [33] employs a similar two-stage framework consisting of auto-labeling and subsequent training of 3D object detectors using the pseudo labels. We compare our method with Autolabels to evaluate the quality of the pseudo labels. As our pseudo labels are unreliable for dy" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17360" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 4. Evaluation results of monocular 3D object detection on the KITTI-360 test set. *Reproduced with the official code. †CAD models are used as extra data. " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }M\\left( D\\right)" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " indicates that detection model " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " is employed for model-agnostic method " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 547, + 220 + ], + "type": "table", + "html": "
MethodWeak SupervisionFull SupervisionAPBEV/AP3D@0.3APBEV/AP3D@0.5
LiDARMasks3D BoxesEasyHardEasyHard
WeakM3D* [23]✓✓29.89/21.2524.01/15.348.10/2.962.96/2.01
Autolabels*†‡ [33] (S-WeakM3D)✓✓48.16/12.9237.34/9.9420.18/4.6914.33/2.79
VSRD‡ (S-WeakM3D)✓51.09/42.9441.28/33.7819.51/11.9114.39/8.46
VSRD‡ (MonoFlex)✓54.40/48.1645.67/40.0429.18/18.5322.31/13.60
VSRD‡ (MonoDETR)✓58.40/50.8650.61/43.4529.07/21.7722.83/16.46
MonoFlex* [35]✓69.70/67.0759.86/57.2650.82/43.1141.78/34.43
MonoDETR* [34]✓63.07/60.4954.04/50.0347.21/41.0136.05/30.38
", + "image_path": "c6c476d812d5937b2856fd75cad2d143244fecf4fd9a0eedcd2d2e0f42236f9c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 264, + 285, + 356 + ], + "blocks": [ + { + "bbox": [ + 47, + 233, + 286, + 256 + ], + "lines": [ + { + "bbox": [ + 47, + 233, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 286, + 256 + ], + "type": "text", + "content": "Table 5. Evaluation results of semi-supervised monocular 3D object detection on the KITTI validation set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 264, + 285, + 356 + ], + "lines": [ + { + "bbox": [ + 50, + 264, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 50, + 264, + 285, + 356 + ], + "type": "table", + "html": "
MethodRatioAPBEV/AP3D@0.7
EasyModerateHard
MonoDETR [34]1.0037.99/29.3626.76/20.6423.02/17.30
0.000.002/0.0010.004/0.0010.005/0.002
0.2531.72/21.7622.32/15.4318.86/12.55
VSRD (MonoDETR)0.5043.44/31.0531.54/21.4827.17/17.93
0.7542.58/32.9531.08/24.6827.19/21.38
", + "image_path": "5b083b3d8164479d1464f507946e5de2a174b31d744a9c44180819da41751c05.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "text", + "content": "namic, occluded, or truncated objects, we group the dataset based on whether the average confidence score for each image is greater than a certain threshold. The evaluation results with the confidence thresholds of 0.0 and 0.8 are shown in Tab. 3. Our method exhibits superior performance in terms of " + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}}@0.3" + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{BEV}}@0.5" + }, + { + "bbox": [ + 46, + 367, + 287, + 522 + ], + "type": "text", + "content": ", our method exhibits slightly lower performance for the confidence threshold of 0.0. However, by raising the confidence threshold to 0.8, our method demonstrates superior performance, indicating that our method can generate more high-quality pseudo labels while abandoning low-quality pseudo labels. The optimized 3D bounding boxes and rendered instance masks are visualized in Fig. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 217, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 217, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 217, + 553 + ], + "type": "text", + "content": "4.4.2 Monocular 3D Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "content": "The pseudo labels generated by the proposed auto-labeling serve as 3D supervision. We further investigate their applicability using the existing monocular 3D object detectors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Weakly Supervised Setting Tab. 4 shows the evaluation results of our method compared with the existing weakly supervised and fully supervised methods. Our method demonstrates a significant superiority over WeakM3D [23] across all the metrics while eliminating the need for LiDAR points for 3D supervision. Moreover, the detector trained on the pseudo labels generated by the proposed auto-labeling outperforms that trained on the pseudo labels generated by Autolabels [33]. Furthermore, employing more sophisticated" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 320 + ], + "type": "text", + "content": "monocular 3D object detectors such as MonoFlex [35] and MonoDETR [34] further improves detection performance, demonstrating the broad versatility of our method, which is not limited to a specific detection model. It is noteworthy that the detectors trained on the pseudo labels generated by our method demonstrate competitive performance compared with those trained in a fully supervised manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 540 + ], + "type": "text", + "content": "Semi-Supervised Setting The essential advantage of our method is that it avoids costly 3D annotations, making more data available for training. Therefore, we investigate a realistic scenario where a detector pre-trained on a large amount of unlabeled data from a source domain is fine-tuned on a small amount of labeled data from a target domain. We select the KITTI-360 and KITTI datasets as source and target domains, respectively. Tab. 5 shows the performance of the detector pre-trained on the KITTI-360 dataset in a weakly supervised manner with the proposed auto-labeling and then fine-tuned on a subset of the KITTI dataset in a supervised manner. The zero-shot performance is quite low due to the characteristic that monocular depth estimation is greatly affected by the differences in camera parameters, but the performance of the detector fine-tuned on only " + }, + { + "bbox": [ + 304, + 324, + 546, + 540 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 324, + 546, + 540 + ], + "type": "text", + "content": " of the labeled data significantly outperforms that trained on the whole data from scratch, highlighting the broad applicability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel weakly supervised 3D object detection framework named VSRD, which consists of multi-view 3D auto-labeling and subsequent training of monocular 3D object detectors using the pseudo labels generated in the auto-labeling stage. Our method demonstrates superior performance compared with the existing weakly supervised 3D object detection methods. Moreover, it exhibits remarkable scalability using partially labeled data for semi-supervised learning. Our proposed method allows leveraging abundant 2D annotations to enhance 3D object detection without explicit 3D supervision, providing a promising avenue for further advancements in the field." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17361" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 304 + ], + "type": "text", + "content": "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 349 + ], + "type": "text", + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 461 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 518 + ], + "type": "text", + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 520, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 287, + 564 + ], + "type": "text", + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 286, + 588 + ], + "type": "text", + "content": "[10] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[12] Benran Hu, Junkai Huang, Yichen Liu, Yu-Wing Tai, and Chi-Keung Tang. Nerf-rpn: A general framework for object detection in nerfs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23528–23538, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[14] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 151 + ], + "type": "text", + "content": "[15] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[16] Zechen Liu, Zizhang Wu, and Roland Toth. Smoke: Single-stage monocular 3d object detection via keypoint estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 996-997, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "text", + "content": "[17] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In European Conference on computer vision, pages 515-531. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[19] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3d bounding box estimation using deep learning and geometry. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7074-7082, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "type": "text", + "content": "[21] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "text", + "content": "[22] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "text", + "content": "[23] Liang Peng, Senbo Yan, Boxi Wu, Zheng Yang, Xiaofei He, and Deng Cai. Weakm3d: Towards weakly supervised monocular 3d object detection. arXiv preprint arXiv:2203.08332, 2022. 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[24] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 1, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Runzhou Tao, Wencheng Han, Zhongying Qiu, Chengzhong Xu, and Jianbing Shen. Weakly supervised monocular" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17362" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 641 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "3d object detection using multi-view projection and direction consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17492, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[27] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "type": "text", + "content": "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14408-14419, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 340 + ], + "type": "text", + "content": "[30] Chenfeng Xu, Bichen Wu, Ji Hou, Sam Tsai, Ruilong Li, Jialiang Wang, Wei Zhan, Zijian He, Peter Vajda, Kurt Keutzer, et al. Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23320-23330, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "type": "text", + "content": "[31] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 397, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 288, + 430 + ], + "type": "text", + "content": "[32] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "type": "text", + "content": "[33] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of sdf shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12224-12233, 2020. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 541 + ], + "type": "text", + "content": "[34] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 1, 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "text", + "content": "[35] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 588, + 288, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 288, + 641 + ], + "type": "text", + "content": "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, pages 12993-13000, 2020. 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17363" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_content_list.json b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f5abc3a424a2b848a4e1c54e58e0562ece185a3a --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift", + "text_level": 1, + "bbox": [ + 107, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Leyuan Liu $^{1}$ , Yuhan Li $^{1}$ , Yunqi Gao $^{1}$ , Changxin Gao $^{2}$ , Yuanyuan Liu $^{3}$ , Jingying Chen $^{1,*}$", + "$^{1}$ National Engineering Research Center for E-Learning, Central China Normal University $^{2}$ School of Artificial Intelligence and Automation, Huazhong University of Science and Technology $^{3}$ School of Computer Science, China University of Geosciences (Wuhan) {lyliu,chenjy} $@$ mail.ccnu.edu.cn, {lyh985812,gaoyunqi} $@$ mails.ccnu.edu.cn, cgao@hust.edu.cn, liuyy@cug.edu.cn" + ], + "bbox": [ + 88, + 179, + 880, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/419cbacb89d2b0a9ca89e8ff5af4c7ab12497b193c85adaa0ba1d2e3c5cdfcc3.jpg", + "image_caption": [ + "Figure 1. Using two stages of vertex shift, VS first stretches the estimated SMPL-X into a coarse human model and then refines it into a detailed clothed human model. Results on both benchmarks (1st row) and in-the-wild images (2nd row) show that VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothes. To comprehensively inspect results from all perspectives, we rotate the results through a full circle and show them from 8 views." + ], + "image_footnote": [], + "bbox": [ + 78, + 284, + 895, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 555, + 313, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various applications require high-fidelity and artifact-free 3D human reconstructions. However, current implicit function-based methods inevitably produce artifacts while existing deformation methods are difficult to reconstruct high-fidelity humans wearing loose clothing. In this paper, we propose a two-stage deformation method named Vertex Shift (VS) for reconstructing clothed 3D humans from single images. Specifically, VS first stretches the estimated SMPL-X mesh into a coarse 3D human model using shift fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network embedded with implicit-function-learned features. This \"stretch-refine\" strategy addresses large deformations required for reconstructing loose clothing and delicate deformations for recovering intricate and detailed surfaces, achieving high-fidelity reconstructions that faithfully convey the pose, clothing, and surface details from the input images. The graph convolutional network's ability to exploit neighborhood vertices coupled with the advantages inherited from the deformation methods ensure VS rarely produces artifacts like distortions and non-human shapes", + "bbox": [ + 73, + 583, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and never produces artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothing. Experimental results on three benchmarks and two in-the-wild datasets demonstrate that VS significantly outperforms current state-of-the-art methods. The code and models of VS are available for research purposes at https://github.com/starVisionTeam/VS.", + "bbox": [ + 496, + 556, + 893, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 723, + 632, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human avatars will be an essential medium for connecting the physical and digital worlds. Recent advancements in 3D human reconstruction [7, 40] have made it convenient to create human avatars even from single phone-taken images, but the reconstructions are not yet perfect. An ideal reconstruction should be high-fidelity and artifact-free. More specifically, it should faithfully convey the pose, clothing, and surface details of the human in the image while being free from artifacts like holes, broken or missing parts, disembodied limbs, distortions, and non-human shapes.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10498", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, encouraging reconstructions of humans with diverse clothing and poses have been yielded by implicit-function-based (IF-based) methods [14, 16, 36, 37]. They utilize deep neural networks to estimate occupancy fields [9, 31] or signed distance fields [32], from which detailed 3D human meshes can be reconstructed via MarchingCubes [29]. However, these IF-based methods often suffer from producing severe artifacts, as they do not exploit any knowledge about the human body structure. To address this issue, follow-up methods [3, 26, 48, 52, 54] utilize parametric body models [28, 34] to regularize the free-form implicit functions. While this regularization alleviates artifacts under the scenario of complex poses, it restricts the flexibility to loose clothing. The most recent work, ECON [49], has improved the quality of reconstructions to a new level, but it still produces reconstructions with wrong thickness, inaccurate poses, and missing parts in many cases. This indicates that more efforts are still needed to enhance the fidelity of the reconstructions and to eliminate artifacts. Besides IF-based methods, there is another line of works that use deformation methods, which naturally do not produce artifacts like holes and broken parts, for 3D object reconstruction [12, 13, 45]. Despite several works [22, 39, 47, 55] have applied deformation methods to clothed 3D human reconstruction, none of them has successfully reconstructed humans with loose clothing from single images. On the one hand, large deformations are required to fit loose clothing; on the other hand, recovering surface details needs to shift the mesh's vertices delicately.", + "bbox": [ + 76, + 90, + 472, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose VS, which stands for vertex shift, to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS is a two-stage deformation method that uses a \"stretch-refine\" strategy to reconcile the contradiction between large deformations required for reconstructing loose clothing and delicate deformations for recovering details. Specifically, VS first stretches the estimated SMPL-X mesh [34] into a coarse 3D human model using displacement fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network (GCN) embedded with implicit-function-learned features. In the stretching stage, VS shifts all vertices on the SMPL-X mesh to property positions where the coarse model can well align with the human in the input image. This enables VS to deal with loose clothing and correct pose/shape errors in the estimated SMPL-X. In the refining stage, the GCN driven by implicit-function-learned features can delicately deform the coarse model to recover the intricate and detailed surfaces of clothed humans. Moreover, each vertex on the graph model is constrained by its neighborhoods, which takes human body priors into account to suppress artifacts like distortions and non-human shapes. During the whole deforming process, VS only shifts vertices without altering edges,", + "bbox": [ + 76, + 537, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guaranteeing that it consistently produces holistic 3D human meshes without artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans on both benchmark datasets and in-the-wild images even under scenarios of challenging poses and loose clothing (See Fig.1).", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 517, + 181, + 821, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose a two-stage deformation method that uses a \"stretch-refine\" strategy for clothed 3D human reconstruction, contributing to reconciling the contradiction between large deformations for reconstructing loose clothing and delicate formations for recovering surface details.", + "(2) We introduce displacement fields inferred from normal maps for stretching the coarse model to align well with the input image, allowing our deformation method to handle loose clothing and correct inaccurate pose estimates.", + "(3) We combine implicit-function-learned features with a graph convolutional network, making VS not only recover surface details but also suppress artifacts." + ], + "bbox": [ + 496, + 196, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 392, + 640, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Body-agnostic methods. Given the remarkable expressive capability of deep implicit functions [8-10, 31, 32] for arbitrary 3D object surfaces, mainstream body-agnostic methods [1, 14, 23, 36, 37] employ them to represent clothed 3D humans. PIFu [36] proposes pixel-aligned image features for training the implicit function. PIFuHD [37] further enhances the geometric details by introducing a multi-level architecture that leverages both low- and high-resolution images. While PIFu and PIFuHD can recover intricate local shapes of 3D humans, they often suffer from artifacts such as disembodied limbs and non-human shapes, especially in the case of complex or novel poses. To reduce artifacts, Geo-PIFu [14] predicts a voxel-based coarse geometry model to regularize the deep implicit function. However, for humans with novel poses that are not included in the training data, Geo-PIFu often predicts poor coarse geometry models. Besides geometry, Stereo-PIFu [16] and Self-portrait [23] utilize depth information to regularize implicit functions, while PHORHUM [1] considers shading and lighting information to infer better geometric details. Since body-agnostic methods do not exploit the human body structure, which is a strong prior for 3D human reconstruction, such methods often struggle with artifacts.", + "bbox": [ + 496, + 417, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Body-aware methods. Body-aware methods [3, 5, 11, 15, 18, 19, 24-27, 42, 44, 48, 49, 52-54] utilize 3D parametric body models such as SMPL [28] and SMPL-X [34] or 2D joint points [6] as additional priors for clothed 3D human reconstruction. DeepHuman [53] fuses a semantic volume generated from the SMPL model with image features. However, due to an over-reliance on SMPL, DeepHuman is only able to reconstruct humans in tight clothing. PAMIR [54], HEI-Human [26], IP-Net [3], and DeepMult", + "bbox": [ + 496, + 765, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10499", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3be71975415f71a92397ddfddc356489c52a4cf95da5995070bfa1105602f762.jpg", + "image_caption": [ + "Figure 2. Overview of VS. VS employs a \"stretch-refine\" strategy to stepwise deform the SMPL-X $(\\mathcal{M}_o)$ into a coarse human model $(\\mathcal{M}_c)$ and a detailed human model $(\\mathcal{M}_d)$ using StretchVS and RefineVS modules, respectively. (1) First, two displacement fields $(S = \\{\\mathcal{S}_f,\\mathcal{S}_b\\})$ are inferred by warping the body normals $(\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})$ into clothing normal maps $(\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})$ via a CNN (ShiftEstNet, $\\mathcal{G}_s$ ). Then, StretchVS shifts vertices of the SMPL-X to from the coarse model using the displacement fields. (2) Taking the coarse model as input, RefineVS employs a GCN (DeformNet, $\\mathcal{G}_{\\mathrm{m}}$ ) embedded with implicit-function-learned features $(\\Psi)$ to infer vertex locations of the detailed model." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iCap [52] use voxelized SMPL to regularize free-form implicit functions. Nevertheless, this regularization restricts these methods' ability to reconstruct humans in loose clothing. Moreover, these methods are sensitive to the accuracy of SMPL estimates, and inaccurate SMPLs may introduce severe artifacts. ARCH [19], ARCH++ [15], FITE [25], and CAR [24] first learn the 3D shape of a human in the canonical space and then transform the learned shape into the posed space using SMPL. ICON [48] employs the SMPL model to guide both the detailed normals estimator and the visibility-aware implicit surface regressor. Recently, ECON [49] first utilizes BINI [4] to recover 2.5D front and back surfaces based on estimated normal maps, and then employs SMPL-X as a \"canvas\" to stitch together the two surfaces and \"inpaint\" the missing geometry. Although encouraging improvements have been achieved by these body-aware methods, they still inevitably produce artifacts and cannot yield high-fidelity reconstructions in many cases.", + "bbox": [ + 75, + 395, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformation methods for 3D object reconstruction. Our work is also inspired by the deformation-based 3D object reconstruction methods [13, 21, 22, 39, 45]. Pix2Mesh [45] employs a graph-based neural network to progressively deform ellipsoid meshes into 3D object models. Point2Mesh [13] iteratively shrinks and wraps a convex hull of the input point cloud to a watertight mesh. GCMR [21] uses a Graph-CNN to directly regress the 3D location of the down-sampled SMPL vertices. MGT-Net [22] utilizes a vertex-based deformation representation to model human shapes and employs cascaded multiscale graph transformation networks to generate topology-consistent 3D human models. Although holistic 3D models can be yielded, these methods are only suitable for reconstructing humans wearing tight clothing.", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 393, + 589, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of VS is depicted in Fig. 2. Given an input image, VS first estimates an SMPL-X mesh and reorganizes it as a watertight 2-manifold graph (Sec. 3.1), then stretches the SMPL-X mesh into a coarse human mesh using displacement fields inferred from visual cues (Sec. 3.2), and finally refines the coarse human mesh by a GCN to form the final detailed human mesh (Sec. 3.3).", + "bbox": [ + 496, + 419, + 890, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Representation", + "text_level": 1, + "bbox": [ + 500, + 532, + 705, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In VS, 3D meshes including the SMPL-X $(\\mathcal{M}_o)$ , the coarse and detailed clothed human models $(\\mathcal{M}_c$ and $\\mathcal{M}_d)$ are represented by graphs:", + "bbox": [ + 496, + 555, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {*} = \\left\\langle \\mathcal {V} _ {*}, \\mathcal {E} \\right\\rangle \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 606, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $* = \\{o, c, d\\}$ , $\\mathcal{V}_* = \\{\\pmb{v}_*,^1, \\pmb{v}_*,^2, \\dots\\}$ are 3D vertices, and $\\mathcal{E} = \\{e^1, e^2, \\dots\\}$ are edges. Particularly, we reorganize the vertices and edges to form watertight 2-manifold triangular meshes using [17]. The manifold and watertight properties ensure that our meshes have consistent surfaces without any holes, broken parts, or disembodied limbs, which are often observed in meshes reconstructed by IF-based methods. As mentioned above, VS stepwise deforms the SMPL-X model into coarse and detailed human models. In order to preserve the manifold and watertight properties of meshes, VS only shifts vertices without altering edges during the deforming process.", + "bbox": [ + 496, + 628, + 892, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original SMPL-X[34] only has 10,475 vertices and 20,908 faces, and meshes deformed from it are challenging to accurately represent clothed 3D humans, especially when loose clothing is involved. To address this issue, we use the method proposed in [17] to increase the vertices and faces of the SMPL-X mesh.", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10500", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Vertex Shift for Coarse Model Stretching", + "text_level": 1, + "bbox": [ + 76, + 90, + 429, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The StretchVS module stretches the SMPL-X mesh $(\\mathcal{M}_o)$ into the coarse human mesh $(\\mathcal{M}_c)$ using displacement fields $(S = \\{\\mathcal{S}_f, \\mathcal{S}_b\\}$ inferred from visual cues:", + "bbox": [ + 76, + 112, + 468, + 159 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S h i f t} \\left(\\mathcal {M} _ {o}, \\mathcal {S}\\right)\\rightarrow \\mathcal {M} _ {c} \\Leftrightarrow \\mathcal {V} _ {\\mathrm {o}} + \\mathcal {V} _ {\\mathrm {s}} \\rightarrow \\mathcal {V} _ {\\mathrm {c}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 166, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{V}_{\\mathrm{o}}$ and $\\mathcal{V}_{\\mathrm{c}}$ are respectively the vertices of $\\mathcal{M}_o$ and $\\mathcal{M}_c$ , and $\\mathcal{V}_{\\mathrm{s}} = \\{\\pmb{v}_{s}^{1},\\pmb{v}_{s}^{2},\\dots \\}$ is a set of shift vectors derived from $\\mathcal{S}$ . For each vertex $\\pmb{v}_s^i$ in $\\mathcal{V}_s$ :", + "bbox": [ + 76, + 188, + 468, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\langle \\mathcal {S} _ {f} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {i f} \\boldsymbol {v} _ {o} ^ {i} \\text {i s v i s i b l e} \\\\ \\left\\langle \\mathcal {S} _ {b} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {e l s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 241, + 468, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where and $\\pi (\\cdot)$ represents the orthogonal transformation that projects a 3D vertex onto the 2D displacement field plane. As described in Equ. 3, StretchVS shifts vertices on X-Y planes while maintaining the depth of SMPL-X, since it is challenging to estimate high-precision depth from 2D images using concise algorithms. Vertex shifts in the Z-direction will be handled by the RefineVS module later.", + "bbox": [ + 75, + 287, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key to StretchVS is inferring the displacement fields from visual cues. Pioneering works like PIFuHD [37], ICON [48] and ECON [49] have already demonstrated the effectiveness of normals in clothed 3D human reconstruction. Encouraged by these works, we choose normals as the visual cues for inferring the displacement fields.", + "bbox": [ + 75, + 393, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normals generation. We employ PyMAF-X [51] to estimate the SMPL-X, and then render two body normal maps $(\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})$ from the SMPL-X respectively from the front and back views. To estimate clothing normal maps $(\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})$ from the input image, the approach presented in ECON [49] is used.", + "bbox": [ + 76, + 484, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Displacement field inferring. In our method, the displacement fields are inferred by establishing the pixel-wise correspondences [41] between the body normal maps $(\\mathcal{R})$ and the clothing normal maps $(\\mathcal{N})$ . Once $\\mathcal{S}$ is inferred, $\\mathcal{R}$ can be warped into $\\mathcal{N}$ :", + "bbox": [ + 75, + 575, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {N}} _ {*} = \\mathcal {W} \\left(\\mathcal {R} _ {*}, \\mathcal {S} _ {*}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 656, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $* = \\{f, b\\}$ , and $\\mathcal{W}(\\cdot)$ is the dense image warp function [2]. For simplicity, we omit the subscripts 'f' and 'b' for the front and back views in the subsequent formulas.", + "bbox": [ + 75, + 680, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design a 5-layer lightweight CNN, named ShiftEst-Net $(\\mathcal{G}_s)$ , to infer the displacement fields $(S)$ :", + "bbox": [ + 76, + 726, + 468, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {s} (\\mathcal {R}, \\mathcal {N}) \\rightarrow \\mathcal {S} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 763, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the body normals and the clothing normals lay in different domains, there is no direct photometric evidence for establishing the dense correspondences between them. Inspired by [56] and [41], we propose a \"cycle wrapping\" strategy. To this end, ShiftEstNet is designed to have the capacity of inverse warping:", + "bbox": [ + 75, + 787, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {s} (\\mathcal {N}, \\mathcal {R}) \\rightarrow \\tilde {\\mathcal {S}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 883, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{S}$ is the inferred displacement fields for inverse warping:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {R}} = \\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 119, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a cycle, we can warp body normal maps into the clothing normal maps and then warp them back to body normal maps, and vice versa:", + "bbox": [ + 498, + 143, + 890, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {\\mathcal {R}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {R}, \\mathcal {S}), \\tilde {\\mathcal {S}}\\right) \\\\ \\hat {\\mathcal {N}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}), \\mathcal {S}\\right) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 196, + 890, + 246 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, ShiftEstNet can be optimized in an self-supervised manner. The objective function consists of three items:", + "bbox": [ + 498, + 253, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {S}, \\hat {\\mathcal {S}}} \\lambda_ {\\mathrm {w}} \\mathcal {L} _ {\\mathrm {w}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 299, + 890, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{w}}$ is a warping loss:", + "bbox": [ + 500, + 330, + 687, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {w}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 354, + 890, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{c}}$ is a cycle consistent loss:", + "bbox": [ + 500, + 382, + 689, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} _ {c} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} _ {c} \\right\\| _ {2} ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 407, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{\\mathrm{s}}$ is a smoothness loss that encourages the estimated displacement fields to be locally smooth [38]:", + "bbox": [ + 498, + 435, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}} = \\left\\| \\mathcal {N} _ {g} - \\hat {\\mathcal {N}} _ {g} \\right\\| _ {2} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 474, + 890, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {N}} _ {g} = \\mathcal {W} \\left(\\mathcal {W} \\left(\\mathcal {N} _ {g}, \\tilde {\\mathcal {S}}\\right), \\mathcal {S}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 502, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_g$ is an uniform-distributed random normal map.", + "bbox": [ + 500, + 531, + 874, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Vertex Shift for Detailed Model Refining", + "text_level": 1, + "bbox": [ + 500, + 555, + 844, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Driven by embedded features $(\\Psi)$ , the RefineVS module deforms the coarse human mesh $(\\mathcal{M}_c)$ into the detailed human mesh $(\\mathcal{M}_d)$ via a GCN named DeformNet $(\\mathcal{G}_{\\mathrm{m}})$ :", + "bbox": [ + 498, + 578, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {\\mathrm {m}} \\left(\\mathcal {M} _ {c}; \\Psi\\right)\\rightarrow \\mathcal {M} _ {d} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 635, + 890, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network structure. DeformNet is designed using the mesh convolution operation proposed in MeshCNN [12]. Recall that $\\mathcal{M}_c$ is a watertight 2-manifold mesh with triangular faces, which guarantees that each edge in the mesh is exactly adjacent to four other edges. Leveraging this invariant, we apply convolution operations to edges rather than vertices. The convolution for an edge $e_0$ and its four adjacent edges $\\{e_1,e_2,e_3,e_4\\}$ is defined as:", + "bbox": [ + 498, + 661, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {k = 0} ^ {4} \\boldsymbol {\\mu} _ {k} \\cdot \\boldsymbol {\\psi} _ {e _ {k}} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 792, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\pmb{\\mu}_0, \\dots, \\pmb{\\mu}_4]$ are convolutional kernels, and $\\psi_{e_k}$ is the feature embedded on edge $e_k$ .", + "bbox": [ + 498, + 844, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\psi} _ {e _ {k}} = \\boldsymbol {\\psi} _ {v _ {i}} \\oplus \\boldsymbol {\\psi} _ {v _ {j}} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 885, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{i}$ and $v_{j}$ are the two vertices of edge $e_{k}$ , and $\\oplus$ denotes the concatenate operator.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DeformNet is designed as a fully convolutional network, so that intermediate layers and the output layer (i.e., the detailed mesh $\\mathcal{M}_d$ ) after mesh convolution operations can maintain the same structure as the input mesh $\\mathcal{M}_c$ . In our implementation, DeformNet consists of 11 mesh convolutional layers, each followed by a LeakyReLU activation. Particularly, the number of channels of the convolution kernel before the output layer is set to 6, allowing us to use the 6-dimensional convolutional result to represent the positions of the two vertices of an edge. Note that a vertex is shared by an uncertain number of edges, which means that DeformNet predicts multiple new positions for each vertex. We take the mean of the multiple positions as the final positions of the vertex.", + "bbox": [ + 75, + 121, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Features. From the seminal PIFu [36] to the current ECON [49], it has been demonstrated that \"pixel-aligned\" features trained using deep implicit functions achieve encouraging results in clothed 3D human reconstruction. In our method, we employ an encoder (denoted as $\\mathcal{G}_e$ ) with a similar structure to PIFuHD [37] to extract visual features from the input image $(\\mathcal{I})$ and the estimated normal maps $(\\{\\mathcal{N}_f,\\mathcal{N}_b\\})$ :", + "bbox": [ + 76, + 332, + 468, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {e} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right)\\rightarrow \\Psi \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 454, + 468, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the encoder can be retrained, we have observed that using an off-the-shelf encoder is a more convenient and feasible solution. Thereby, we use the pre-trained encoder from PIFuHD [37]. Experiments show that DeformNet is not sensitive to feature encoders, and replacing the encoder with that of other methods like PIFu [36] and PaMIR [54] does not significantly decrease the quality of reconstructions (see \"ablation study\" in Sec. 4.4). In addition to visual features extracted by the encoder, we also embed the coordinate $(\\pmb{v})$ and normal $(\\mathcal{N}_v)$ of each vertex onto the vertex:", + "bbox": [ + 75, + 472, + 468, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\psi} _ {\\boldsymbol {v}} = \\left\\langle \\Psi (\\pi (\\boldsymbol {v})), \\mathcal {N} _ {\\boldsymbol {v}}, \\boldsymbol {v} \\right\\rangle \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 630, + 468, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi (\\pmb {v})$ denotes the orthogonal projection that maps a 3D vertex $\\pmb{v}$ onto the 2D feature plane.", + "bbox": [ + 76, + 652, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Losses. The coarse model $\\mathcal{M}_c$ input into DeformNet already has an approximate geometric topology, but it is not accurate enough and lacks local details. In addition, we expect not to introduce \"bulge\" artifacts when deforming $\\mathcal{M}_c$ into $\\mathcal{M}_d$ . Thus, DeformNet faces three tasks: (1) improving geometry accuracy, (2) recovering local details, and (3) preventing \"bulge\" artifacts. To these ends, the loss for training DeformNet consists of a geometric error term $(\\mathcal{L}_{\\mathrm{d}})$ , a local detail error term $(\\mathcal{L}_{\\mathrm{n}})$ , and a regularization term $(\\mathcal{L}_{\\mathrm{e}})$ :", + "bbox": [ + 75, + 683, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 825, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{\\mathrm{d}}$ is described by Chamfer distance:", + "bbox": [ + 76, + 847, + 366, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d}} = \\sum_ {\\hat {\\boldsymbol {p}} \\in \\hat {\\mathcal {P}}} \\min _ {\\boldsymbol {p} \\in \\mathcal {P}} \\| \\hat {\\boldsymbol {p}} - \\boldsymbol {p} \\| _ {2} + \\sum_ {\\boldsymbol {q} \\in \\mathcal {P}} \\min _ {\\hat {\\boldsymbol {q}} \\in \\hat {\\mathcal {P}}} \\| \\hat {\\boldsymbol {q}} - \\boldsymbol {q} \\| _ {2} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 867, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{P}}$ and $\\mathcal{P}$ are 3D point sets respectively sampled from the predicted ground-truth meshes. Note that, we sample uniformly distributed 3D points using the \"triangle point picking\" approach [46] when computing $\\mathcal{L}_{\\mathrm{d}}$ , since the original vertices of the predicted and ground-truth meshes are usually not distributed uniformly on their surfaces. $\\mathcal{L}_{\\mathrm{n}}$ is described using the cosin distance of normals:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n}} = \\sum_ {\\boldsymbol {p}, \\hat {\\boldsymbol {p}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {p}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}} \\| _ {2}} + \\sum_ {\\boldsymbol {q}, \\hat {\\boldsymbol {q}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {q}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {q}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}} \\| _ {2}} \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 200, + 890, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\pmb{p},\\hat{\\pmb{p}})$ and $(\\pmb{q},\\hat{\\pmb{q}})$ are bi-directional matched closest point pairs, and $\\pmb{n}$ represents the normal vector. The regularization term $\\mathcal{L}_{\\mathrm{e}}$ penalizes long edges:", + "bbox": [ + 496, + 244, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e}} = \\sum_ {e \\in \\mathcal {E}} \\| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\| _ {2} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 296, + 890, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{v}_i$ and $\\pmb{v}_j$ are the two vertices of edge $e$ . $\\mathcal{L}_{\\mathrm{e}}$ not only prevents \"bulges\" artifacts, but also helps convergence.", + "bbox": [ + 496, + 334, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 376, + 705, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VS is compared with five state-of-the-art (SOTA) single-image clothed 3D human reconstruction methods with publicly released code: PIFu [36], PIFuHD [37], PaMIR [54], ICON [48], and ECON [49]. We opt for the $\\mathrm{ECON_{EX}}$ version of ECON [49], as it is reported to have superior and stable performance over the $\\mathrm{ECON_{IF}}$ version [49]. All methods employ the post-processing technique used in ICON. For a fair comparison, we use the re-implemented PIFu and PaMIR provided by Xiu et al. [48], which have uniformized network settings and inputs with ICON. As such, all methods involved in our experiments fairly take an RGB image and the estimated front/back normal maps as input.", + "bbox": [ + 496, + 401, + 890, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 590, + 601, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training dataset. As with ECON [49], VS is trained on the THuman2.0 dataset [50], which contains 525 3D scans and SMPL-X models of 525 subjects with lab-acted poses and daily clothes. In our experiments, we split THuman2.0 into a training set and a testing set at a ratio of 20:1 by subjects. To augment the training data, we render 18 RGB images for each 3D scan by rotating a virtual camera w.r.t the frontal view at an interval of $20^{\\circ}$ . In the end, 9K pairs (RGB image, 3D scan) are obtained for training.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing datasets. Our VS is extensively evaluated on five datasets: the testing set of THuman 2.0 [50], CAPE [30], RenderPeople [35], SHHQ [20], and VcgPeople. Humans in these testing datasets exhibit a variety of poses and clothes, many of which are novel in the training data. For THuman2.0, CAPE, and RenderPeople, which have ground-truth 3D scans, we render 6 testing images from each scan at random viewpoints. Since SHHQ and VcgPeople just contain in-the-wild images, we only conduct qualitative evaluations on them.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10502", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/43d23c8ff2959153c96ecec15f93e0266e7ab9cbb7fc821304ece82a9fee9e41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPublicationsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]ICCV'191.7601.9040.05000.24082.9672.7380.04490.24092.7812.8570.05900.2773
†PIFuHD [37]CVPR'203.0883.1130.08910.36634.7143.8230.05550.27963.3113.31180.08460.3541
PaMIR [54]TPAMI'221.0641.1850.04380.19271.7721.4040.03370.16761.5801.6590.04860.2088
ICON [48]CVPR'220.9470.9250.04220.17611.1331.0960.03110.14311.2651.2510.04310.1871
ECON [49]CVPR'230.9060.8450.03790.18910.9370.9210.03350.16441.2851.0790.04170.1973
VSOurs0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
", + "bbox": [ + 161, + 88, + 810, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 239, + 297, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Chamfer distance $(\\varepsilon_{cd}$ , unit: cm), point-to-surface distance $(\\varepsilon_{p2s}$ , unit: cm), cosine distance $(\\varepsilon_{cos} \\in [0,1])$ , and L2 normal error $(\\varepsilon_{l2} \\in [0,1])$ are employed as quantitative metrics. The first two metrics mainly assess large geometric errors, while the latter two tend to measure local detail differences between reconstructions and ground-truth models. Following ECON [49], when reporting cosine distance and L2 normal error, four normal maps are rendered respectively for the reconstructed and ground-truth models at the angles of $\\{0^o, 90^o, 180^o, 270^o\\}$ .", + "bbox": [ + 75, + 262, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. As illustrated in Table 1, VS demonstrates superior performance on all three benchmark datasets. Especially, our VS exceeds all other methods by a large margin on the two metrics that represent large geometric errors (i.e., $\\varepsilon_{cd}$ and $\\varepsilon_{p2s}$ ). In terms of local surface details (i.e., $\\varepsilon_{cos}$ and $\\varepsilon_{L2}$ ), VS performs on par with ECON and much better than PIFu, PIFuHD, PaMIR, and ICON.", + "bbox": [ + 75, + 412, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 529, + 287, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual results. Fig. 4 shows the reconstructions by applying our VS on in-the-wild images. It can be seen that VS recovers high-fidelity clothed 3D humans even in challenging poses and complex loose clothes. Note that the poses, clothes, and accessories in these images are out of the distribution of the training data. These results reveal that VS not only has high reconstruction accuracy on challenging images but also has strong generalization capabilities.", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual comparison. Given that ECON [49] is the current SOTA method for single-image clothed 3D human reconstruction, we first compare our VS with it solely. Fig. 3 shows the reconstructions by ECON and our VS. It can be seen that VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. Fig. 5 demonstrates visual comparisons with all five SOTA methods on in-the-wild images. ECON and VS demonstrate a substantial advantage over PIFu [36], PIFuHD [37], PaMIR [54], and ICON [48].", + "bbox": [ + 75, + 674, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perceptual study. We also conduct a perceptual study on the in-the-wild SHHQ and VcgPeople datasets. To this end, we designed a perceptual evaluation tool (see Fig. 11 in SupMat) where reconstructions produced by different meth", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2081b4adff711e99ed34b9e08f7bc5813cea47e4c2d96e1afa27f088b25471fa.jpg", + "image_caption": [ + "Figure 3. Comparison of VS with ECON on in-the-wild images. VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. More comparisons are shown in Video2 in SupMat." + ], + "image_footnote": [], + "bbox": [ + 529, + 234, + 864, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods are blindly placed on one page. We recruited 25 volunteers to participate in the perceptual study, and randomly sampled 50 images from each of the two datasets for each participant. Participants were requested to select the best reconstruction for each input image after rotating the reconstructions and carefully inspecting them from all viewpoints. The average percentages of preferences from these participants are illustrated in Table 2. The majority of participants perceived that our VS achieves the best reconstructions on both datasets. It's worth noting that ECON and VS exhibit greater superiorities over other methods on the more challenging VcgPeople dataset.", + "bbox": [ + 496, + 508, + 892, + 690 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4dca442b15a1fa097347e88623de112915bf3353b106bd7be138b70212bb607a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison against other SOTAs on three benchmarks. Ground-truth SMPLs are used for methods that need SMPL(x) guidance. † Official model trained on Renderpeople." + ], + "table_footnote": [], + "table_body": "
DatasetsPIFuPIFuHDPaMIRICONECONVS
SHHQ3.64.311.13.919.058.1
VcgPeople0.20.15.65.024.764.4
", + "bbox": [ + 522, + 700, + 869, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Average percentages (\\%) of participants' preferences. VS is significantly preferred over other methods.", + "bbox": [ + 498, + 750, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 801, + 653, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the two vertex shift states. To assess the effectiveness of the StretchVS and RefineVS modules in VS, we implemented two variants of VS, each exclusively incorporating either the StretchVS module or the RefineVS module. The quantitative results achieved by the complete", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce4d7c47b41e61d6afc3a0589ff67f6503a04b0ad6749bb4ce69fbb28f3e33ed.jpg", + "image_caption": [ + "Figure 4. Results produced by VS on in-the-wild images. Each result is shown from three views. VS recovers high-fidelity and artifact-less clothed 3D humans even in challenging poses (P) and loose clothes (C). More results are shown in Fig.12 and Video1 in SupMat." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 872, + 291 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/684c39efe8b82876456e5e195f9dc34fb1ea06acb37b5c9db2103a86a9f95927.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VersionsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
VS0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
StretchVS0.9790.9150.06450.24690.9350.9070.04040.17431.2091.0400.06270.2443
RefineVS0.6690.5980.03910.16790.7210.7190.02710.12931.2630.9970.05290.2138
", + "bbox": [ + 161, + 342, + 807, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results achieved by three versions of VS.", + "bbox": [ + 297, + 417, + 669, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VS and the two variants are presented in Table 3. It is evident from these results that: (1) StretchVS exhibits a significant performance decrease on all three datasets, especially in terms of local details. (2) RefineVS performs consistently on THuman2.0 and CAPE without significant degradation, while it exhibits a noticeable decline on RenderPeople. This is because RenderPeople contains many out-of-distribution loose clothes, and it is difficult for the RefineVS module to deform an SMPL-X mesh to fit loose clothes without the assistance of the StretchVS module. The visual results shown in Fig. 6 are mutually corroborated with the quantitative results illustrated in Table 3.", + "bbox": [ + 75, + 458, + 470, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to SMPL-X estimates. Most body-aware methods are known to be sensitive to SMPL-(X) estimates, while SMPL-(X) estimation remains an unresolved issue. In this experiment, we use both accurate and inaccurate SMPL-X estimates to initiate VS. As shown in Fig. 7, VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates. Since StretchVS treats SMPL-X just as a \"seed\" and stretches it to align with visual cues, it has the \"error-correcting\" ability to shift vertices on inaccurate SMPL-X estimates to proper positions.", + "bbox": [ + 496, + 458, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flexibility to feature encoders. To validate that VS is insensitive to IF-based feature encoders, we evaluated VS", + "bbox": [ + 498, + 612, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c5cbfa23ef8d740a8138bc74716a0eb529558f22197610441183f767423d4bf.jpg", + "image_caption": [ + "Figure 5. Visual comparison with other SOTA methods on in-the-wild images. More results are shown in Fig. $16\\sim 19$ in SupMat." + ], + "image_footnote": [], + "bbox": [ + 117, + 660, + 851, + 871 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10504", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8cb52039b278bd51f4671b3b97221d6ccb44d5dc69b30d79e009cc5e43ba79b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EncodersTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]0.7660.6880.04110.17900.7350.7360.02720.12911.1040.8840.04650.1939
PaMIR [54]0.8380.7330.04020.17680.8610.7990.02660.13101.1820.9330.04280.1879
PIFuHD [37]0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
", + "bbox": [ + 142, + 88, + 828, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative evaluation of VS with different feature encoders.", + "bbox": [ + 274, + 164, + 692, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/119374a63e0ddbe4089eb4681bf6b6511380d82c00ace9bc1ab260c9f0630a5f.jpg", + "image_caption": [ + "Figure 6. Results produced by VS variants w/wo StretchVS and RefineVS modules. StretchVS handles large deformations like loose clothing, while RefineVS recovers surface details. For humans who don't wear loose clothing, RefineVS can work alone." + ], + "image_footnote": [], + "bbox": [ + 109, + 200, + 439, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f996322bece70b52b8cfd11cc5d8b73a324e67a14099762e01223caf22f562c5.jpg", + "image_caption": [ + "Figure 7. Robustness to SMPL-X estimates. We use both accurate (odd rows) and inaccurate (even rows) SMPL-X estimates to initiate VS. VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates." + ], + "image_footnote": [], + "bbox": [ + 109, + 420, + 437, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with pre-trained encoders from PIFu [36], PaMIR [54], and PIFuHD [37]. As shown in Table 4, the employment of different feature encoders does not result in significant discrepancies in the quantitative metrics achieved by VS.", + "bbox": [ + 75, + 722, + 470, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 799, + 194, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose VS to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS first stretches the estimated SMPL-X model into a coarse 3D human model using displacement fields inferred from visual cues, and then refines the coarse 3D human model into a de", + "bbox": [ + 75, + 824, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7862f9bd0f67eed98ff37b7663fd7b342715d5e9a5ad0f36874141a57bedb442.jpg", + "image_caption": [ + "Figure 8. Failure cases. VS may fail to handle extreme clothing, layer confusion, and large deformation in the Z-axis." + ], + "image_footnote": [], + "bbox": [ + 513, + 202, + 879, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tailed 3D human model via a GCN embedded with implicit-function-learned features. This two-stage deformation strategy reconciles the contradiction between large deformations for reconstructing loose clothing and delicate vertex shifts for recovering intricate and detailed surfaces. Extensive experiments on five datasets demonstrate that VS can reconstruct high-fidelity and artifact-less clothed 3D humans and achieves SOTA performance. VS confirms that deformation methods can reconstruct high-quality clothed 3D humans with complex poses and loose clothing, and even have advantages over IF-based methods in eliminating artifacts. We hope that VS can provide a new perspective for 3D human reconstruction and bring us one step closer to creating \"perfect\" human avatars for real-world applications.", + "bbox": [ + 496, + 356, + 890, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future work. Although VS achieves impressive results, it still has some limitations, as shown in Fig. 8. (A) Extreme clothing. VS may face challenges in deforming vertices on the legs of SMPL-X into extreme dresses. (B) Layer confusion. Currently, VS cannot distinguish the layers of the human body and clothing. (C) Wrong depth. VS fails in a couple of images exhibiting large deformation in the Z-axis (say, belts drifting towards the camera). In future work, we plan to separate a clothed human into the body and clothing parts and reconstruct them separately. As the clothing styles can be identified from the images [43] and the corresponding preset models can be provided [33], we hypothesize that using VS to deform preset clothing and body models separately can reconstruct extreme clothing and mitigate layer confusion.", + "bbox": [ + 496, + 569, + 892, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 808, + 669, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (62077026), and the Fundamental Research Funds for the Central Universities (CCNU22QN012).", + "bbox": [ + 496, + 834, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1506-1515, 2022. 2", + "[2] The TensorFlow Authors. Dense image warp. www.tensorflow.org/addons/api_docs/python/tfa/image/dense_image_warp, 2023.4", + "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Eur. Conf. Comput. Vis. (ECCV), page 311-329, 2020. 2", + "[4] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Eur. Conf. Comput. Vis. (ECCV), pages 552 - 567, 2022. 3", + "[5] Yukang Cao, Kai Han, and Kwan-Yee K. Wong. Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4647-4657, 2023. 2", + "[6] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7291-7299, 2017. 2", + "[7] Lu Chen, Sida Peng, and Xiaowei Zhou. Towards efficient and photorealistic 3d human reconstruction: A brief survey. Visual Informatics, 5(4):11-19, 2021. 1", + "[8] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 18522-18531, 2022. 2", + "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5932-5941, 2019. 2", + "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Int. Conf. Mach. Learn. (ICML), pages 3789-3799, 2020. 2", + "[11] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12869-12879, 2023. 2", + "[12] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes-Or, Shachar Fleishman, and Daniel. MeshCNN: A network with an edge. ACM Trans. Graph. (TOG), 38(4):90:1–90:12, 2019. 2, 4", + "[13] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph. (TOG), 39(4):126:1-126:12, 2020. 2, 3, 1", + "[14] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-PIFu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 2304–2314, 2020. 2", + "[15] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\\mathrm{ARCH}++$ : Animation-ready clothed human re" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "construction revisited. In Int. Conf. Comput. Vis. (ICCV), pages 11046-11056, 2021. 2, 3", + "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. StereoPIFu: Depth aware clothed human digitization via stereo vision. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 119-128, 2021. 2", + "[17] Jingwei Huang, Hao Su, and Leonidas J. Guibas. Robust watertight manifold surface generation method for shapenet models. ArXiv, 2018. 3, 1", + "[18] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In Int. Conf. 3D Vis.(3DV), 2024. 2, 1", + "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH:animatable reconstruction of clothed humans. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3093-3102, 2020. 2, 3", + "[20] Fu Jianglin, Li Shikai, Jiang Yuming, Lin Kwan-Yee, Qian Chen, Loy Chen Change, Wu Wayne, and Liu Ziwei. Stylegan-human: A data-centric odyssey of human generation. In Eur. Conf. Comput. Vis. (ECCV), pages 1-19, 2022. 5", + "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4501-4510, 2019. 3", + "[22] Kun Li, Wen Hao, Feng Qiao, Yuxiang Zhang, Xiongzheng Li, Jing Huang, Cunkuan Yuan, Yu-Kun Lai, and Yebin Liu. Image-guided human reconstruction via multi-scale graph transformation networks. IEEE Trans. Image Process. (TIP), 30:5239-5251, 2021. 2, 3", + "[23] Zhe Li, Tao Yu, Chuanyu Pan, Zerong Zheng, and Yebin Liu. Robust 3d self-portraits in seconds. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1344–1353, 2020. 2", + "[24] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8662–8672, 2023. 2, 3", + "[25] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In Eur. Conf. Comput. Vis. (ECCV), page 210-228, 2022. 3", + "[26] Leyuan Liu, Jianchi Sun, Yunqi Gao, and Jingying Chen. HEI-Human: A hybrid explicit and implicit method for single-view 3d clothed human reconstruction. In Chinese Conf. Pattern Recog. Comput. Vis. (PRCV), pages 251-262, 2021. 2", + "[27] Leyuan Liu, Yunqi Gao, Jianchi Sun, and Jingying Chen. Single-image clothed 3d human reconstruction guided by a well-aligned parametric body model. Multimedia Systems, page 1579-1592, 2023. 2", + "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 34(6):1-16, 2015. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10506", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] William Lorensen and Harvey Cline. Marching Cubes: A high resolution 3d surface construction algorithm. ACM Trans. Graph. (TOG), 21(4):163-169, 1987. 2", + "[30] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3d people in generative clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6469-6478, 2020. 5, 1", + "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d reconstruction in function space. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4455-4465, 2019. 2", + "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 165-174, 2019. 2", + "[33] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7365-7375, 2020. 8", + "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10975-10985, 2019. 2, 3", + "[35] RenderPeople. www.renderpeople.com, 2018. 5, 1", + "[36] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis. (ICCV), pages 2304-2314, 2019. 2, 5, 6, 8, 1", + "[37] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 84-93, 2020. 2, 4, 5, 6, 8, 1", + "[38] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I-Chao Chang, Yubo Fan, and Yan Xu. Unsupervised end-to-end learning for deformable medical image registrations. ArXiv, 2018. 4", + "[39] Qingyang Tan, Lin Gao, Yu-Kun Lai, Jie Yang, and Shihong Xia1. Mesh-based autoencoders for localized deformation component analysis. In AAAI, page 2452-2459, 2018. 2, 3", + "[40] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3D human mesh from monocular images: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), pages 15406 - 15425, 2023. 1", + "[41] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Int. Conf. Comput. Vis. (ICCV), pages 10346-10356, 2021. 4", + "[42] Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: Volumetric inference of 3d human body shapes. In Eur. Conf. Comput. Vis. (ECCV), pages 20-36, 2018. 2", + "[43] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-YOLOv4: Scaling cross stage" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "partial network. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13029-13038, 2021. 8", + "[44] Junying Wang, Jae Shin Yoon, Tuanfeng Y. Wang, Krishna Kumar Singh, and Ulrich Neumann. Complete 3d human reconstruction from a single incomplete image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8748-8758, 2023. 2", + "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In Eur. Conf. Comput. Vis. (ECCV), pages 52-67, 2018. 2, 3", + "[46] Eric W. Weissstein. Triangle point picking. mathworld. wolfram.com/TrianglePointPicking.html, 2023.5", + "[47] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Int. Conf. 3D Vis.(3DV), pages 322-332, 2020. 2", + "[48] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit clothed humans obtained from normals. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13296-13306, 2022. 2, 3, 4, 5, 6, 1", + "[49] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit clothed humans optimized via normal integration. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 512-523, 2023. 2, 3, 4, 5, 6, 1", + "[50] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5746-5756, 2021. 5, 1", + "[51] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: towards well-aligned full-body model regression from monocular images. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 45(10):12287 - 12303, 2023. 4", + "[52] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. DeepMultiCap: Performance capture of multiple characters using sparse multiview cameras. In Int. Conf. Comput. Vis. (ICCV), pages 6219-6229, 2021. 2, 3", + "[53] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3d human reconstruction from a single image. In Int. Conf. Comput. Vis. (ICCV), pages 7739-7749, 2019. 2", + "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 44(6):3170 - 3184, 2022. 2, 5, 6, 8, 1", + "[55] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4491-4500, 2019. 2", + "[56] Jun-yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis. (ICCV), pages 2223-2232, 2017. 4" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10507", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_model.json b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ab05e008348ed8d6aafe4ce888d0350a2570d2 --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.18, + 0.832, + 0.199 + ], + "angle": 0, + "content": "Leyuan Liu\\(^{1}\\), Yuhan Li\\(^{1}\\), Yunqi Gao\\(^{1}\\), Changxin Gao\\(^{2}\\), Yuanyuan Liu\\(^{3}\\), Jingying Chen\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.199, + 0.881, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\) National Engineering Research Center for E-Learning, Central China Normal University \\(^{2}\\) School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \\(^{3}\\) School of Computer Science, China University of Geosciences (Wuhan) {lyliu,chenjy} \\(@\\) mail.ccnu.edu.cn, {lyh985812,gaoyunqi} \\(@\\) mails.ccnu.edu.cn, cgao@hust.edu.cn, liuyy@cug.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.18, + 0.881, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.285, + 0.896, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.466, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 1. Using two stages of vertex shift, VS first stretches the estimated SMPL-X into a coarse human model and then refines it into a detailed clothed human model. Results on both benchmarks (1st row) and in-the-wild images (2nd row) show that VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothes. To comprehensively inspect results from all perspectives, we rotate the results through a full circle and show them from 8 views." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.556, + 0.314, + 0.572 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.584, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Various applications require high-fidelity and artifact-free 3D human reconstructions. However, current implicit function-based methods inevitably produce artifacts while existing deformation methods are difficult to reconstruct high-fidelity humans wearing loose clothing. In this paper, we propose a two-stage deformation method named Vertex Shift (VS) for reconstructing clothed 3D humans from single images. Specifically, VS first stretches the estimated SMPL-X mesh into a coarse 3D human model using shift fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network embedded with implicit-function-learned features. This \"stretch-refine\" strategy addresses large deformations required for reconstructing loose clothing and delicate deformations for recovering intricate and detailed surfaces, achieving high-fidelity reconstructions that faithfully convey the pose, clothing, and surface details from the input images. The graph convolutional network's ability to exploit neighborhood vertices coupled with the advantages inherited from the deformation methods ensure VS rarely produces artifacts like distortions and non-human shapes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.694 + ], + "angle": 0, + "content": "and never produces artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothing. Experimental results on three benchmarks and two in-the-wild datasets demonstrate that VS significantly outperforms current state-of-the-art methods. The code and models of VS are available for research purposes at https://github.com/starVisionTeam/VS." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.724, + 0.633, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Human avatars will be an essential medium for connecting the physical and digital worlds. Recent advancements in 3D human reconstruction [7, 40] have made it convenient to create human avatars even from single phone-taken images, but the reconstructions are not yet perfect. An ideal reconstruction should be high-fidelity and artifact-free. More specifically, it should faithfully convey the pose, clothing, and surface details of the human in the image while being free from artifacts like holes, broken or missing parts, disembodied limbs, distortions, and non-human shapes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Recently, encouraging reconstructions of humans with diverse clothing and poses have been yielded by implicit-function-based (IF-based) methods [14, 16, 36, 37]. They utilize deep neural networks to estimate occupancy fields [9, 31] or signed distance fields [32], from which detailed 3D human meshes can be reconstructed via MarchingCubes [29]. However, these IF-based methods often suffer from producing severe artifacts, as they do not exploit any knowledge about the human body structure. To address this issue, follow-up methods [3, 26, 48, 52, 54] utilize parametric body models [28, 34] to regularize the free-form implicit functions. While this regularization alleviates artifacts under the scenario of complex poses, it restricts the flexibility to loose clothing. The most recent work, ECON [49], has improved the quality of reconstructions to a new level, but it still produces reconstructions with wrong thickness, inaccurate poses, and missing parts in many cases. This indicates that more efforts are still needed to enhance the fidelity of the reconstructions and to eliminate artifacts. Besides IF-based methods, there is another line of works that use deformation methods, which naturally do not produce artifacts like holes and broken parts, for 3D object reconstruction [12, 13, 45]. Despite several works [22, 39, 47, 55] have applied deformation methods to clothed 3D human reconstruction, none of them has successfully reconstructed humans with loose clothing from single images. On the one hand, large deformations are required to fit loose clothing; on the other hand, recovering surface details needs to shift the mesh's vertices delicately." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.473, + 0.904 + ], + "angle": 0, + "content": "In this paper, we propose VS, which stands for vertex shift, to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS is a two-stage deformation method that uses a \"stretch-refine\" strategy to reconcile the contradiction between large deformations required for reconstructing loose clothing and delicate deformations for recovering details. Specifically, VS first stretches the estimated SMPL-X mesh [34] into a coarse 3D human model using displacement fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network (GCN) embedded with implicit-function-learned features. In the stretching stage, VS shifts all vertices on the SMPL-X mesh to property positions where the coarse model can well align with the human in the input image. This enables VS to deal with loose clothing and correct pose/shape errors in the estimated SMPL-X. In the refining stage, the GCN driven by implicit-function-learned features can delicately deform the coarse model to recover the intricate and detailed surfaces of clothed humans. Moreover, each vertex on the graph model is constrained by its neighborhoods, which takes human body priors into account to suppress artifacts like distortions and non-human shapes. During the whole deforming process, VS only shifts vertices without altering edges," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.182 + ], + "angle": 0, + "content": "guaranteeing that it consistently produces holistic 3D human meshes without artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans on both benchmark datasets and in-the-wild images even under scenarios of challenging poses and loose clothing (See Fig.1)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.183, + 0.822, + 0.196 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.894, + 0.272 + ], + "angle": 0, + "content": "(1) We propose a two-stage deformation method that uses a \"stretch-refine\" strategy for clothed 3D human reconstruction, contributing to reconciling the contradiction between large deformations for reconstructing loose clothing and delicate formations for recovering surface details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.894, + 0.332 + ], + "angle": 0, + "content": "(2) We introduce displacement fields inferred from normal maps for stretching the coarse model to align well with the input image, allowing our deformation method to handle loose clothing and correct inaccurate pose estimates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.379 + ], + "angle": 0, + "content": "(3) We combine implicit-function-learned features with a graph convolutional network, making VS not only recover surface details but also suppress artifacts." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.642, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Body-agnostic methods. Given the remarkable expressive capability of deep implicit functions [8-10, 31, 32] for arbitrary 3D object surfaces, mainstream body-agnostic methods [1, 14, 23, 36, 37] employ them to represent clothed 3D humans. PIFu [36] proposes pixel-aligned image features for training the implicit function. PIFuHD [37] further enhances the geometric details by introducing a multi-level architecture that leverages both low- and high-resolution images. While PIFu and PIFuHD can recover intricate local shapes of 3D humans, they often suffer from artifacts such as disembodied limbs and non-human shapes, especially in the case of complex or novel poses. To reduce artifacts, Geo-PIFu [14] predicts a voxel-based coarse geometry model to regularize the deep implicit function. However, for humans with novel poses that are not included in the training data, Geo-PIFu often predicts poor coarse geometry models. Besides geometry, Stereo-PIFu [16] and Self-portrait [23] utilize depth information to regularize implicit functions, while PHORHUM [1] considers shading and lighting information to infer better geometric details. Since body-agnostic methods do not exploit the human body structure, which is a strong prior for 3D human reconstruction, such methods often struggle with artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Body-aware methods. Body-aware methods [3, 5, 11, 15, 18, 19, 24-27, 42, 44, 48, 49, 52-54] utilize 3D parametric body models such as SMPL [28] and SMPL-X [34] or 2D joint points [6] as additional priors for clothed 3D human reconstruction. DeepHuman [53] fuses a semantic volume generated from the SMPL model with image features. However, due to an over-reliance on SMPL, DeepHuman is only able to reconstruct humans in tight clothing. PAMIR [54], HEI-Human [26], IP-Net [3], and DeepMult" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10499" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Figure 2. Overview of VS. VS employs a \"stretch-refine\" strategy to stepwise deform the SMPL-X \\((\\mathcal{M}_o)\\) into a coarse human model \\((\\mathcal{M}_c)\\) and a detailed human model \\((\\mathcal{M}_d)\\) using StretchVS and RefineVS modules, respectively. (1) First, two displacement fields \\((S = \\{\\mathcal{S}_f,\\mathcal{S}_b\\})\\) are inferred by warping the body normals \\((\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})\\) into clothing normal maps \\((\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})\\) via a CNN (ShiftEstNet, \\(\\mathcal{G}_s\\)). Then, StretchVS shifts vertices of the SMPL-X to from the coarse model using the displacement fields. (2) Taking the coarse model as input, RefineVS employs a GCN (DeformNet, \\(\\mathcal{G}_{\\mathrm{m}}\\)) embedded with implicit-function-learned features \\((\\Psi)\\) to infer vertex locations of the detailed model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.473, + 0.669 + ], + "angle": 0, + "content": "iCap [52] use voxelized SMPL to regularize free-form implicit functions. Nevertheless, this regularization restricts these methods' ability to reconstruct humans in loose clothing. Moreover, these methods are sensitive to the accuracy of SMPL estimates, and inaccurate SMPLs may introduce severe artifacts. ARCH [19], ARCH++ [15], FITE [25], and CAR [24] first learn the 3D shape of a human in the canonical space and then transform the learned shape into the posed space using SMPL. ICON [48] employs the SMPL model to guide both the detailed normals estimator and the visibility-aware implicit surface regressor. Recently, ECON [49] first utilizes BINI [4] to recover 2.5D front and back surfaces based on estimated normal maps, and then employs SMPL-X as a \"canvas\" to stitch together the two surfaces and \"inpaint\" the missing geometry. Although encouraging improvements have been achieved by these body-aware methods, they still inevitably produce artifacts and cannot yield high-fidelity reconstructions in many cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Deformation methods for 3D object reconstruction. Our work is also inspired by the deformation-based 3D object reconstruction methods [13, 21, 22, 39, 45]. Pix2Mesh [45] employs a graph-based neural network to progressively deform ellipsoid meshes into 3D object models. Point2Mesh [13] iteratively shrinks and wraps a convex hull of the input point cloud to a watertight mesh. GCMR [21] uses a Graph-CNN to directly regress the 3D location of the down-sampled SMPL vertices. MGT-Net [22] utilizes a vertex-based deformation representation to model human shapes and employs cascaded multiscale graph transformation networks to generate topology-consistent 3D human models. Although holistic 3D models can be yielded, these methods are only suitable for reconstructing humans wearing tight clothing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.591, + 0.41 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.526 + ], + "angle": 0, + "content": "An overview of VS is depicted in Fig. 2. Given an input image, VS first estimates an SMPL-X mesh and reorganizes it as a watertight 2-manifold graph (Sec. 3.1), then stretches the SMPL-X mesh into a coarse human mesh using displacement fields inferred from visual cues (Sec. 3.2), and finally refines the coarse human mesh by a GCN to form the final detailed human mesh (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.707, + 0.549 + ], + "angle": 0, + "content": "3.1. Model Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.601 + ], + "angle": 0, + "content": "In VS, 3D meshes including the SMPL-X \\((\\mathcal{M}_o)\\), the coarse and detailed clothed human models \\((\\mathcal{M}_c\\) and \\(\\mathcal{M}_d)\\) are represented by graphs:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.607, + 0.891, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {*} = \\left\\langle \\mathcal {V} _ {*}, \\mathcal {E} \\right\\rangle \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.81 + ], + "angle": 0, + "content": "where \\(* = \\{o, c, d\\}\\), \\(\\mathcal{V}_* = \\{\\pmb{v}_*,^1, \\pmb{v}_*,^2, \\dots\\}\\) are 3D vertices, and \\(\\mathcal{E} = \\{e^1, e^2, \\dots\\}\\) are edges. Particularly, we reorganize the vertices and edges to form watertight 2-manifold triangular meshes using [17]. The manifold and watertight properties ensure that our meshes have consistent surfaces without any holes, broken parts, or disembodied limbs, which are often observed in meshes reconstructed by IF-based methods. As mentioned above, VS stepwise deforms the SMPL-X model into coarse and detailed human models. In order to preserve the manifold and watertight properties of meshes, VS only shifts vertices without altering edges during the deforming process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.9 + ], + "angle": 0, + "content": "The original SMPL-X[34] only has 10,475 vertices and 20,908 faces, and meshes deformed from it are challenging to accurately represent clothed 3D humans, especially when loose clothing is involved. To address this issue, we use the method proposed in [17] to increase the vertices and faces of the SMPL-X mesh." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10500" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.43, + 0.108 + ], + "angle": 0, + "content": "3.2. Vertex Shift for Coarse Model Stretching" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.113, + 0.47, + 0.16 + ], + "angle": 0, + "content": "The StretchVS module stretches the SMPL-X mesh \\((\\mathcal{M}_o)\\) into the coarse human mesh \\((\\mathcal{M}_c)\\) using displacement fields \\((S = \\{\\mathcal{S}_f, \\mathcal{S}_b\\}\\) inferred from visual cues:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.167, + 0.469, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S h i f t} \\left(\\mathcal {M} _ {o}, \\mathcal {S}\\right)\\rightarrow \\mathcal {M} _ {c} \\Leftrightarrow \\mathcal {V} _ {\\mathrm {o}} + \\mathcal {V} _ {\\mathrm {s}} \\rightarrow \\mathcal {V} _ {\\mathrm {c}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.189, + 0.47, + 0.237 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}_{\\mathrm{o}}\\) and \\(\\mathcal{V}_{\\mathrm{c}}\\) are respectively the vertices of \\(\\mathcal{M}_o\\) and \\(\\mathcal{M}_c\\), and \\(\\mathcal{V}_{\\mathrm{s}} = \\{\\pmb{v}_{s}^{1},\\pmb{v}_{s}^{2},\\dots \\}\\) is a set of shift vectors derived from \\(\\mathcal{S}\\). For each vertex \\(\\pmb{v}_s^i\\) in \\(\\mathcal{V}_s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.242, + 0.469, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\langle \\mathcal {S} _ {f} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {i f} \\boldsymbol {v} _ {o} ^ {i} \\text {i s v i s i b l e} \\\\ \\left\\langle \\mathcal {S} _ {b} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {e l s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.469, + 0.394 + ], + "angle": 0, + "content": "where and \\(\\pi (\\cdot)\\) represents the orthogonal transformation that projects a 3D vertex onto the 2D displacement field plane. As described in Equ. 3, StretchVS shifts vertices on X-Y planes while maintaining the depth of SMPL-X, since it is challenging to estimate high-precision depth from 2D images using concise algorithms. Vertex shifts in the Z-direction will be handled by the RefineVS module later." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.469, + 0.485 + ], + "angle": 0, + "content": "The key to StretchVS is inferring the displacement fields from visual cues. Pioneering works like PIFuHD [37], ICON [48] and ECON [49] have already demonstrated the effectiveness of normals in clothed 3D human reconstruction. Encouraged by these works, we choose normals as the visual cues for inferring the displacement fields." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.485, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Normals generation. We employ PyMAF-X [51] to estimate the SMPL-X, and then render two body normal maps \\((\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})\\) from the SMPL-X respectively from the front and back views. To estimate clothing normal maps \\((\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})\\) from the input image, the approach presented in ECON [49] is used." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.469, + 0.651 + ], + "angle": 0, + "content": "Displacement field inferring. In our method, the displacement fields are inferred by establishing the pixel-wise correspondences [41] between the body normal maps \\((\\mathcal{R})\\) and the clothing normal maps \\((\\mathcal{N})\\). Once \\(\\mathcal{S}\\) is inferred, \\(\\mathcal{R}\\) can be warped into \\(\\mathcal{N}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.657, + 0.469, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {N}} _ {*} = \\mathcal {W} \\left(\\mathcal {R} _ {*}, \\mathcal {S} _ {*}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.726 + ], + "angle": 0, + "content": "where \\( * = \\{f, b\\} \\), and \\( \\mathcal{W}(\\cdot) \\) is the dense image warp function [2]. For simplicity, we omit the subscripts 'f' and 'b' for the front and back views in the subsequent formulas." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.469, + 0.757 + ], + "angle": 0, + "content": "We design a 5-layer lightweight CNN, named ShiftEst-Net \\((\\mathcal{G}_s)\\), to infer the displacement fields \\((S)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.764, + 0.469, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {s} (\\mathcal {R}, \\mathcal {N}) \\rightarrow \\mathcal {S} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Since the body normals and the clothing normals lay in different domains, there is no direct photometric evidence for establishing the dense correspondences between them. Inspired by [56] and [41], we propose a \"cycle wrapping\" strategy. To this end, ShiftEstNet is designed to have the capacity of inverse warping:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.884, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {s} (\\mathcal {N}, \\mathcal {R}) \\rightarrow \\tilde {\\mathcal {S}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.121 + ], + "angle": 0, + "content": "where \\(\\tilde{S}\\) is the inferred displacement fields for inverse warping:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.121, + 0.892, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {R}} = \\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.144, + 0.892, + 0.189 + ], + "angle": 0, + "content": "In a cycle, we can warp body normal maps into the clothing normal maps and then warp them back to body normal maps, and vice versa:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.198, + 0.892, + 0.247 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {\\mathcal {R}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {R}, \\mathcal {S}), \\tilde {\\mathcal {S}}\\right) \\\\ \\hat {\\mathcal {N}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}), \\mathcal {S}\\right) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.299 + ], + "angle": 0, + "content": "In this way, ShiftEstNet can be optimized in an self-supervised manner. The objective function consists of three items:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.3, + 0.891, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {S}, \\hat {\\mathcal {S}}} \\lambda_ {\\mathrm {w}} \\mathcal {L} _ {\\mathrm {w}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.331, + 0.688, + 0.346 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{w}}\\) is a warping loss:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.355, + 0.892, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {w}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.383, + 0.691, + 0.398 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{c}}\\) is a cycle consistent loss:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.408, + 0.891, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} _ {c} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} _ {c} \\right\\| _ {2} ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.466 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{\\mathrm{s}}\\) is a smoothness loss that encourages the estimated displacement fields to be locally smooth [38]:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.476, + 0.891, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s}} = \\left\\| \\mathcal {N} _ {g} - \\hat {\\mathcal {N}} _ {g} \\right\\| _ {2} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.503, + 0.891, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {N}} _ {g} = \\mathcal {W} \\left(\\mathcal {W} \\left(\\mathcal {N} _ {g}, \\tilde {\\mathcal {S}}\\right), \\mathcal {S}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.532, + 0.875, + 0.548 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_g\\) is an uniform-distributed random normal map." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.846, + 0.572 + ], + "angle": 0, + "content": "3.3. Vertex Shift for Detailed Model Refining" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.579, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Driven by embedded features \\((\\Psi)\\), the RefineVS module deforms the coarse human mesh \\((\\mathcal{M}_c)\\) into the detailed human mesh \\((\\mathcal{M}_d)\\) via a GCN named DeformNet \\((\\mathcal{G}_{\\mathrm{m}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.636, + 0.891, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {\\mathrm {m}} \\left(\\mathcal {M} _ {c}; \\Psi\\right)\\rightarrow \\mathcal {M} _ {d} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.662, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Network structure. DeformNet is designed using the mesh convolution operation proposed in MeshCNN [12]. Recall that \\(\\mathcal{M}_c\\) is a watertight 2-manifold mesh with triangular faces, which guarantees that each edge in the mesh is exactly adjacent to four other edges. Leveraging this invariant, we apply convolution operations to edges rather than vertices. The convolution for an edge \\(e_0\\) and its four adjacent edges \\(\\{e_1,e_2,e_3,e_4\\}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.794, + 0.891, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\sum_ {k = 0} ^ {4} \\boldsymbol {\\mu} _ {k} \\cdot \\boldsymbol {\\psi} _ {e _ {k}} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.845, + 0.892, + 0.875 + ], + "angle": 0, + "content": "where \\([\\pmb{\\mu}_0, \\dots, \\pmb{\\mu}_4]\\) are convolutional kernels, and \\(\\psi_{e_k}\\) is the feature embedded on edge \\(e_k\\)." + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.886, + 0.891, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\psi} _ {e _ {k}} = \\boldsymbol {\\psi} _ {v _ {i}} \\oplus \\boldsymbol {\\psi} _ {v _ {j}} \\tag {16}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10501" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "where \\( v_{i} \\) and \\( v_{j} \\) are the two vertices of edge \\( e_{k} \\), and \\( \\oplus \\) denotes the concatenate operator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.332 + ], + "angle": 0, + "content": "DeformNet is designed as a fully convolutional network, so that intermediate layers and the output layer (i.e., the detailed mesh \\(\\mathcal{M}_d\\)) after mesh convolution operations can maintain the same structure as the input mesh \\(\\mathcal{M}_c\\). In our implementation, DeformNet consists of 11 mesh convolutional layers, each followed by a LeakyReLU activation. Particularly, the number of channels of the convolution kernel before the output layer is set to 6, allowing us to use the 6-dimensional convolutional result to represent the positions of the two vertices of an edge. Note that a vertex is shared by an uncertain number of edges, which means that DeformNet predicts multiple new positions for each vertex. We take the mean of the multiple positions as the final positions of the vertex." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.469, + 0.454 + ], + "angle": 0, + "content": "Features. From the seminal PIFu [36] to the current ECON [49], it has been demonstrated that \"pixel-aligned\" features trained using deep implicit functions achieve encouraging results in clothed 3D human reconstruction. In our method, we employ an encoder (denoted as \\(\\mathcal{G}_e\\)) with a similar structure to PIFuHD [37] to extract visual features from the input image \\((\\mathcal{I})\\) and the estimated normal maps \\((\\{\\mathcal{N}_f,\\mathcal{N}_b\\})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.455, + 0.469, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {e} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right)\\rightarrow \\Psi \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.624 + ], + "angle": 0, + "content": "Although the encoder can be retrained, we have observed that using an off-the-shelf encoder is a more convenient and feasible solution. Thereby, we use the pre-trained encoder from PIFuHD [37]. Experiments show that DeformNet is not sensitive to feature encoders, and replacing the encoder with that of other methods like PIFu [36] and PaMIR [54] does not significantly decrease the quality of reconstructions (see \"ablation study\" in Sec. 4.4). In addition to visual features extracted by the encoder, we also embed the coordinate \\((\\pmb{v})\\) and normal \\((\\mathcal{N}_v)\\) of each vertex onto the vertex:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.631, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\psi} _ {\\boldsymbol {v}} = \\left\\langle \\Psi (\\pi (\\boldsymbol {v})), \\mathcal {N} _ {\\boldsymbol {v}}, \\boldsymbol {v} \\right\\rangle \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.469, + 0.683 + ], + "angle": 0, + "content": "where \\(\\pi (\\pmb {v})\\) denotes the orthogonal projection that maps a 3D vertex \\(\\pmb{v}\\) onto the 2D feature plane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.469, + 0.819 + ], + "angle": 0, + "content": "Losses. The coarse model \\(\\mathcal{M}_c\\) input into DeformNet already has an approximate geometric topology, but it is not accurate enough and lacks local details. In addition, we expect not to introduce \"bulge\" artifacts when deforming \\(\\mathcal{M}_c\\) into \\(\\mathcal{M}_d\\). Thus, DeformNet faces three tasks: (1) improving geometry accuracy, (2) recovering local details, and (3) preventing \"bulge\" artifacts. To these ends, the loss for training DeformNet consists of a geometric error term \\((\\mathcal{L}_{\\mathrm{d}})\\), a local detail error term \\((\\mathcal{L}_{\\mathrm{n}})\\), and a regularization term \\((\\mathcal{L}_{\\mathrm{e}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.826, + 0.469, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.848, + 0.367, + 0.863 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{\\mathrm{d}}\\) is described by Chamfer distance:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.868, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d}} = \\sum_ {\\hat {\\boldsymbol {p}} \\in \\hat {\\mathcal {P}}} \\min _ {\\boldsymbol {p} \\in \\mathcal {P}} \\| \\hat {\\boldsymbol {p}} - \\boldsymbol {p} \\| _ {2} + \\sum_ {\\boldsymbol {q} \\in \\mathcal {P}} \\min _ {\\hat {\\boldsymbol {q}} \\in \\hat {\\mathcal {P}}} \\| \\hat {\\boldsymbol {q}} - \\boldsymbol {q} \\| _ {2} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.198 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{P}}\\) and \\(\\mathcal{P}\\) are 3D point sets respectively sampled from the predicted ground-truth meshes. Note that, we sample uniformly distributed 3D points using the \"triangle point picking\" approach [46] when computing \\(\\mathcal{L}_{\\mathrm{d}}\\), since the original vertices of the predicted and ground-truth meshes are usually not distributed uniformly on their surfaces. \\(\\mathcal{L}_{\\mathrm{n}}\\) is described using the cosin distance of normals:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.202, + 0.892, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n}} = \\sum_ {\\boldsymbol {p}, \\hat {\\boldsymbol {p}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {p}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}} \\| _ {2}} + \\sum_ {\\boldsymbol {q}, \\hat {\\boldsymbol {q}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {q}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {q}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}} \\| _ {2}} \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.291 + ], + "angle": 0, + "content": "where \\((\\pmb{p},\\hat{\\pmb{p}})\\) and \\((\\pmb{q},\\hat{\\pmb{q}})\\) are bi-directional matched closest point pairs, and \\(\\pmb{n}\\) represents the normal vector. The regularization term \\(\\mathcal{L}_{\\mathrm{e}}\\) penalizes long edges:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.297, + 0.891, + 0.329 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e}} = \\sum_ {e \\in \\mathcal {E}} \\| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\| _ {2} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.366 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_i\\) and \\(\\pmb{v}_j\\) are the two vertices of edge \\(e\\). \\(\\mathcal{L}_{\\mathrm{e}}\\) not only prevents \"bulges\" artifacts, but also helps convergence." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.377, + 0.706, + 0.394 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.584 + ], + "angle": 0, + "content": "VS is compared with five state-of-the-art (SOTA) single-image clothed 3D human reconstruction methods with publicly released code: PIFu [36], PIFuHD [37], PaMIR [54], ICON [48], and ECON [49]. We opt for the \\(\\mathrm{ECON_{EX}}\\) version of ECON [49], as it is reported to have superior and stable performance over the \\(\\mathrm{ECON_{IF}}\\) version [49]. All methods employ the post-processing technique used in ICON. For a fair comparison, we use the re-implemented PIFu and PaMIR provided by Xiu et al. [48], which have uniformized network settings and inputs with ICON. As such, all methods involved in our experiments fairly take an RGB image and the estimated front/back normal maps as input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.602, + 0.605 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Training dataset. As with ECON [49], VS is trained on the THuman2.0 dataset [50], which contains 525 3D scans and SMPL-X models of 525 subjects with lab-acted poses and daily clothes. In our experiments, we split THuman2.0 into a training set and a testing set at a ratio of 20:1 by subjects. To augment the training data, we render 18 RGB images for each 3D scan by rotating a virtual camera w.r.t the frontal view at an interval of \\(20^{\\circ}\\). In the end, 9K pairs (RGB image, 3D scan) are obtained for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Testing datasets. Our VS is extensively evaluated on five datasets: the testing set of THuman 2.0 [50], CAPE [30], RenderPeople [35], SHHQ [20], and VcgPeople. Humans in these testing datasets exhibit a variety of poses and clothes, many of which are novel in the training data. For THuman2.0, CAPE, and RenderPeople, which have ground-truth 3D scans, we render 6 testing images from each scan at random viewpoints. Since SHHQ and VcgPeople just contain in-the-wild images, we only conduct qualitative evaluations on them." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10502" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.163, + 0.089, + 0.812, + 0.174 + ], + "angle": 0, + "content": "
MethodsPublicationsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]ICCV'191.7601.9040.05000.24082.9672.7380.04490.24092.7812.8570.05900.2773
†PIFuHD [37]CVPR'203.0883.1130.08910.36634.7143.8230.05550.27963.3113.31180.08460.3541
PaMIR [54]TPAMI'221.0641.1850.04380.19271.7721.4040.03370.16761.5801.6590.04860.2088
ICON [48]CVPR'220.9470.9250.04220.17611.1331.0960.03110.14311.2651.2510.04310.1871
ECON [49]CVPR'230.9060.8450.03790.18910.9370.9210.03350.16441.2851.0790.04170.1973
VSOurs0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.185, + 0.89, + 0.214 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison against other SOTAs on three benchmarks. Ground-truth SMPLs are used for methods that need SMPL(x) guidance. † Official model trained on Renderpeople." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.298, + 0.256 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.413 + ], + "angle": 0, + "content": "Metrics. Chamfer distance \\((\\varepsilon_{cd}\\), unit: cm), point-to-surface distance \\((\\varepsilon_{p2s}\\), unit: cm), cosine distance \\((\\varepsilon_{cos} \\in [0,1])\\), and L2 normal error \\((\\varepsilon_{l2} \\in [0,1])\\) are employed as quantitative metrics. The first two metrics mainly assess large geometric errors, while the latter two tend to measure local detail differences between reconstructions and ground-truth models. Following ECON [49], when reporting cosine distance and L2 normal error, four normal maps are rendered respectively for the reconstructed and ground-truth models at the angles of \\(\\{0^o, 90^o, 180^o, 270^o\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.52 + ], + "angle": 0, + "content": "Quantitative comparison. As illustrated in Table 1, VS demonstrates superior performance on all three benchmark datasets. Especially, our VS exceeds all other methods by a large margin on the two metrics that represent large geometric errors (i.e., \\(\\varepsilon_{cd}\\) and \\(\\varepsilon_{p2s}\\)). In terms of local surface details (i.e., \\(\\varepsilon_{cos}\\) and \\(\\varepsilon_{L2}\\)), VS performs on par with ECON and much better than PIFu, PIFuHD, PaMIR, and ICON." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.288, + 0.545 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Visual results. Fig. 4 shows the reconstructions by applying our VS on in-the-wild images. It can be seen that VS recovers high-fidelity clothed 3D humans even in challenging poses and complex loose clothes. Note that the poses, clothes, and accessories in these images are out of the distribution of the training data. These results reveal that VS not only has high reconstruction accuracy on challenging images but also has strong generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Visual comparison. Given that ECON [49] is the current SOTA method for single-image clothed 3D human reconstruction, we first compare our VS with it solely. Fig. 3 shows the reconstructions by ECON and our VS. It can be seen that VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. Fig. 5 demonstrates visual comparisons with all five SOTA methods on in-the-wild images. ECON and VS demonstrate a substantial advantage over PIFu [36], PIFuHD [37], PaMIR [54], and ICON [48]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Perceptual study. We also conduct a perceptual study on the in-the-wild SHHQ and VcgPeople datasets. To this end, we designed a perceptual evaluation tool (see Fig. 11 in SupMat) where reconstructions produced by different meth" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.236, + 0.865, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Figure 3. Comparison of VS with ECON on in-the-wild images. VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. More comparisons are shown in Video2 in SupMat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.893, + 0.691 + ], + "angle": 0, + "content": "ods are blindly placed on one page. We recruited 25 volunteers to participate in the perceptual study, and randomly sampled 50 images from each of the two datasets for each participant. Participants were requested to select the best reconstruction for each input image after rotating the reconstructions and carefully inspecting them from all viewpoints. The average percentages of preferences from these participants are illustrated in Table 2. The majority of participants perceived that our VS achieves the best reconstructions on both datasets. It's worth noting that ECON and VS exhibit greater superiorities over other methods on the more challenging VcgPeople dataset." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.702, + 0.87, + 0.741 + ], + "angle": 0, + "content": "
DatasetsPIFuPIFuHDPaMIRICONECONVS
SHHQ3.64.311.13.919.058.1
VcgPeople0.20.15.65.024.764.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.751, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Table 2. Average percentages (\\%) of participants' preferences. VS is significantly preferred over other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.654, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the two vertex shift states. To assess the effectiveness of the StretchVS and RefineVS modules in VS, we implemented two variants of VS, each exclusively incorporating either the StretchVS module or the RefineVS module. The quantitative results achieved by the complete" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.086, + 0.873, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.893, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Results produced by VS on in-the-wild images. Each result is shown from three views. VS recovers high-fidelity and artifact-less clothed 3D humans even in challenging poses (P) and loose clothes (C). More results are shown in Fig.12 and Video1 in SupMat." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.343, + 0.808, + 0.407 + ], + "angle": 0, + "content": "
VersionsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
VS0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
StretchVS0.9790.9150.06450.24690.9350.9070.04040.17431.2091.0400.06270.2443
RefineVS0.6690.5980.03910.16790.7210.7190.02710.12931.2630.9970.05290.2138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.299, + 0.418, + 0.67, + 0.432 + ], + "angle": 0, + "content": "Table 3. Quantitative results achieved by three versions of VS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.64 + ], + "angle": 0, + "content": "VS and the two variants are presented in Table 3. It is evident from these results that: (1) StretchVS exhibits a significant performance decrease on all three datasets, especially in terms of local details. (2) RefineVS performs consistently on THuman2.0 and CAPE without significant degradation, while it exhibits a noticeable decline on RenderPeople. This is because RenderPeople contains many out-of-distribution loose clothes, and it is difficult for the RefineVS module to deform an SMPL-X mesh to fit loose clothes without the assistance of the StretchVS module. The visual results shown in Fig. 6 are mutually corroborated with the quantitative results illustrated in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Robustness to SMPL-X estimates. Most body-aware methods are known to be sensitive to SMPL-(X) estimates, while SMPL-(X) estimation remains an unresolved issue. In this experiment, we use both accurate and inaccurate SMPL-X estimates to initiate VS. As shown in Fig. 7, VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates. Since StretchVS treats SMPL-X just as a \"seed\" and stretches it to align with visual cues, it has the \"error-correcting\" ability to shift vertices on inaccurate SMPL-X estimates to proper positions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Flexibility to feature encoders. To validate that VS is insensitive to IF-based feature encoders, we evaluated VS" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.661, + 0.852, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.882, + 0.865, + 0.898 + ], + "angle": 0, + "content": "Figure 5. Visual comparison with other SOTA methods on in-the-wild images. More results are shown in Fig. \\(16\\sim 19\\) in SupMat." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10504" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.83, + 0.155 + ], + "angle": 0, + "content": "
EncodersTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]0.7660.6880.04110.17900.7350.7360.02720.12911.1040.8840.04650.1939
PaMIR [54]0.8380.7330.04020.17680.8610.7990.02660.13101.1820.9330.04280.1879
PIFuHD [37]0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.165, + 0.694, + 0.178 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation of VS with different feature encoders." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.201, + 0.441, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.352, + 0.471, + 0.407 + ], + "angle": 0, + "content": "Figure 6. Results produced by VS variants w/wo StretchVS and RefineVS modules. StretchVS handles large deformations like loose clothing, while RefineVS recovers surface details. For humans who don't wear loose clothing, RefineVS can work alone." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.421, + 0.439, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Figure 7. Robustness to SMPL-X estimates. We use both accurate (odd rows) and inaccurate (even rows) SMPL-X estimates to initiate VS. VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.471, + 0.786 + ], + "angle": 0, + "content": "with pre-trained encoders from PIFu [36], PaMIR [54], and PIFuHD [37]. As shown in Table 4, the employment of different feature encoders does not result in significant discrepancies in the quantitative metrics achieved by VS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.196, + 0.815 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.9 + ], + "angle": 0, + "content": "We propose VS to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS first stretches the estimated SMPL-X model into a coarse 3D human model using displacement fields inferred from visual cues, and then refines the coarse 3D human model into a de" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.203, + 0.88, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.303, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Figure 8. Failure cases. VS may fail to handle extreme clothing, layer confusion, and large deformation in the Z-axis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.569 + ], + "angle": 0, + "content": "tailed 3D human model via a GCN embedded with implicit-function-learned features. This two-stage deformation strategy reconciles the contradiction between large deformations for reconstructing loose clothing and delicate vertex shifts for recovering intricate and detailed surfaces. Extensive experiments on five datasets demonstrate that VS can reconstruct high-fidelity and artifact-less clothed 3D humans and achieves SOTA performance. VS confirms that deformation methods can reconstruct high-quality clothed 3D humans with complex poses and loose clothing, and even have advantages over IF-based methods in eliminating artifacts. We hope that VS can provide a new perspective for 3D human reconstruction and bring us one step closer to creating \"perfect\" human avatars for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.797 + ], + "angle": 0, + "content": "Limitations & Future work. Although VS achieves impressive results, it still has some limitations, as shown in Fig. 8. (A) Extreme clothing. VS may face challenges in deforming vertices on the legs of SMPL-X into extreme dresses. (B) Layer confusion. Currently, VS cannot distinguish the layers of the human body and clothing. (C) Wrong depth. VS fails in a couple of images exhibiting large deformation in the Z-axis (say, belts drifting towards the camera). In future work, we plan to separate a clothed human into the body and clothing parts and reconstruct them separately. As the clothing styles can be identified from the images [43] and the corresponding preset models can be provided [33], we hypothesize that using VS to deform preset clothing and body models separately can reconstruct extreme clothing and mitigate layer confusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.671, + 0.827 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.835, + 0.892, + 0.895 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (62077026), and the Fundamental Research Funds for the Central Universities (CCNU22QN012)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10505" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1506-1515, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] The TensorFlow Authors. Dense image warp. www.tensorflow.org/addons/api_docs/python/tfa/image/dense_image_warp, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Eur. Conf. Comput. Vis. (ECCV), page 311-329, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.286, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Eur. Conf. Comput. Vis. (ECCV), pages 552 - 567, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[5] Yukang Cao, Kai Han, and Kwan-Yee K. Wong. Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4647-4657, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7291-7299, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[7] Lu Chen, Sida Peng, and Xiaowei Zhou. Towards efficient and photorealistic 3d human reconstruction: A brief survey. Visual Informatics, 5(4):11-19, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[8] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 18522-18531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5932-5941, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Int. Conf. Mach. Learn. (ICML), pages 3789-3799, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[11] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12869-12879, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[12] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes-Or, Shachar Fleishman, and Daniel. MeshCNN: A network with an edge. ACM Trans. Graph. (TOG), 38(4):90:1–90:12, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[13] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph. (TOG), 39(4):126:1-126:12, 2020. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[14] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-PIFu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 2304–2314, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. \\(\\mathrm{ARCH}++\\): Animation-ready clothed human re" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "construction revisited. In Int. Conf. Comput. Vis. (ICCV), pages 11046-11056, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. StereoPIFu: Depth aware clothed human digitization via stereo vision. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 119-128, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[17] Jingwei Huang, Hao Su, and Leonidas J. Guibas. Robust watertight manifold surface generation method for shapenet models. ArXiv, 2018. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.225, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[18] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In Int. Conf. 3D Vis.(3DV), 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH:animatable reconstruction of clothed humans. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3093-3102, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[20] Fu Jianglin, Li Shikai, Jiang Yuming, Lin Kwan-Yee, Qian Chen, Loy Chen Change, Wu Wayne, and Liu Ziwei. Stylegan-human: A data-centric odyssey of human generation. In Eur. Conf. Comput. Vis. (ECCV), pages 1-19, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4501-4510, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[22] Kun Li, Wen Hao, Feng Qiao, Yuxiang Zhang, Xiongzheng Li, Jing Huang, Cunkuan Yuan, Yu-Kun Lai, and Yebin Liu. Image-guided human reconstruction via multi-scale graph transformation networks. IEEE Trans. Image Process. (TIP), 30:5239-5251, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[23] Zhe Li, Tao Yu, Chuanyu Pan, Zerong Zheng, and Yebin Liu. Robust 3d self-portraits in seconds. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1344–1353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[24] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8662–8672, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[25] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In Eur. Conf. Comput. Vis. (ECCV), page 210-228, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[26] Leyuan Liu, Jianchi Sun, Yunqi Gao, and Jingying Chen. HEI-Human: A hybrid explicit and implicit method for single-view 3d clothed human reconstruction. In Chinese Conf. Pattern Recog. Comput. Vis. (PRCV), pages 251-262, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Leyuan Liu, Yunqi Gao, Jianchi Sun, and Jingying Chen. Single-image clothed 3d human reconstruction guided by a well-aligned parametric body model. Multimedia Systems, page 1579-1592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 34(6):1-16, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] William Lorensen and Harvey Cline. Marching Cubes: A high resolution 3d surface construction algorithm. ACM Trans. Graph. (TOG), 21(4):163-169, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[30] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3d people in generative clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6469-6478, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d reconstruction in function space. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4455-4465, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.345 + ], + "angle": 0, + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[33] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7365-7375, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10975-10985, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.45, + 0.489 + ], + "angle": 0, + "content": "[35] RenderPeople. www.renderpeople.com, 2018. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[36] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis. (ICCV), pages 2304-2314, 2019. 2, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[37] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 84-93, 2020. 2, 4, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[38] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I-Chao Chang, Yubo Fan, and Yan Xu. Unsupervised end-to-end learning for deformable medical image registrations. ArXiv, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[39] Qingyang Tan, Lin Gao, Yu-Kun Lai, Jie Yang, and Shihong Xia1. Mesh-based autoencoders for localized deformation component analysis. In AAAI, page 2452-2459, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[40] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3D human mesh from monocular images: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), pages 15406 - 15425, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[41] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Int. Conf. Comput. Vis. (ICCV), pages 10346-10356, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[42] Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: Volumetric inference of 3d human body shapes. In Eur. Conf. Comput. Vis. (ECCV), pages 20-36, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[43] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-YOLOv4: Scaling cross stage" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "partial network. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13029-13038, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.893, + 0.188 + ], + "angle": 0, + "content": "[44] Junying Wang, Jae Shin Yoon, Tuanfeng Y. Wang, Krishna Kumar Singh, and Ulrich Neumann. Complete 3d human reconstruction from a single incomplete image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8748-8758, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In Eur. Conf. Comput. Vis. (ECCV), pages 52-67, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[46] Eric W. Weissstein. Triangle point picking. mathworld. wolfram.com/TrianglePointPicking.html, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[47] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Int. Conf. 3D Vis.(3DV), pages 322-332, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[48] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit clothed humans obtained from normals. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13296-13306, 2022. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[49] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit clothed humans optimized via normal integration. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 512-523, 2023. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[50] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5746-5756, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[51] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: towards well-aligned full-body model regression from monocular images. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 45(10):12287 - 12303, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[52] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. DeepMultiCap: Performance capture of multiple characters using sparse multiview cameras. In Int. Conf. Comput. Vis. (ICCV), pages 6219-6229, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[53] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3d human reconstruction from a single image. In Int. Conf. Comput. Vis. (ICCV), pages 7739-7749, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 44(6):3170 - 3184, 2022. 2, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[55] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4491-4500, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Jun-yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis. (ICCV), pages 2223-2232, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10507" + } + ] +] \ No newline at end of file diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_origin.pdf b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d85fc61a764c9b97698d57ebf1522cd3bb2160b0 --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/1d0a8a04-2343-4d6f-871f-2ebf0c5dbdc7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d317d5370c5c6481d3f9c17e39c93f970e073371a42d2b8cb273244f4d5cdc3c +size 5799234 diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/full.md b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/full.md new file mode 100644 index 0000000000000000000000000000000000000000..562145559afbad3e7b92fb9d61b6a95870348f12 --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/full.md @@ -0,0 +1,352 @@ +# VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift + +Leyuan Liu $^{1}$ , Yuhan Li $^{1}$ , Yunqi Gao $^{1}$ , Changxin Gao $^{2}$ , Yuanyuan Liu $^{3}$ , Jingying Chen $^{1,*}$ +$^{1}$ National Engineering Research Center for E-Learning, Central China Normal University $^{2}$ School of Artificial Intelligence and Automation, Huazhong University of Science and Technology $^{3}$ School of Computer Science, China University of Geosciences (Wuhan) {lyliu,chenjy} $@$ mail.ccnu.edu.cn, {lyh985812,gaoyunqi} $@$ mails.ccnu.edu.cn, cgao@hust.edu.cn, liuyy@cug.edu.cn + +![](images/419cbacb89d2b0a9ca89e8ff5af4c7ab12497b193c85adaa0ba1d2e3c5cdfcc3.jpg) +Figure 1. Using two stages of vertex shift, VS first stretches the estimated SMPL-X into a coarse human model and then refines it into a detailed clothed human model. Results on both benchmarks (1st row) and in-the-wild images (2nd row) show that VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothes. To comprehensively inspect results from all perspectives, we rotate the results through a full circle and show them from 8 views. + +# Abstract + +Various applications require high-fidelity and artifact-free 3D human reconstructions. However, current implicit function-based methods inevitably produce artifacts while existing deformation methods are difficult to reconstruct high-fidelity humans wearing loose clothing. In this paper, we propose a two-stage deformation method named Vertex Shift (VS) for reconstructing clothed 3D humans from single images. Specifically, VS first stretches the estimated SMPL-X mesh into a coarse 3D human model using shift fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network embedded with implicit-function-learned features. This "stretch-refine" strategy addresses large deformations required for reconstructing loose clothing and delicate deformations for recovering intricate and detailed surfaces, achieving high-fidelity reconstructions that faithfully convey the pose, clothing, and surface details from the input images. The graph convolutional network's ability to exploit neighborhood vertices coupled with the advantages inherited from the deformation methods ensure VS rarely produces artifacts like distortions and non-human shapes + +and never produces artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothing. Experimental results on three benchmarks and two in-the-wild datasets demonstrate that VS significantly outperforms current state-of-the-art methods. The code and models of VS are available for research purposes at https://github.com/starVisionTeam/VS. + +# 1. Introduction + +Human avatars will be an essential medium for connecting the physical and digital worlds. Recent advancements in 3D human reconstruction [7, 40] have made it convenient to create human avatars even from single phone-taken images, but the reconstructions are not yet perfect. An ideal reconstruction should be high-fidelity and artifact-free. More specifically, it should faithfully convey the pose, clothing, and surface details of the human in the image while being free from artifacts like holes, broken or missing parts, disembodied limbs, distortions, and non-human shapes. + +Recently, encouraging reconstructions of humans with diverse clothing and poses have been yielded by implicit-function-based (IF-based) methods [14, 16, 36, 37]. They utilize deep neural networks to estimate occupancy fields [9, 31] or signed distance fields [32], from which detailed 3D human meshes can be reconstructed via MarchingCubes [29]. However, these IF-based methods often suffer from producing severe artifacts, as they do not exploit any knowledge about the human body structure. To address this issue, follow-up methods [3, 26, 48, 52, 54] utilize parametric body models [28, 34] to regularize the free-form implicit functions. While this regularization alleviates artifacts under the scenario of complex poses, it restricts the flexibility to loose clothing. The most recent work, ECON [49], has improved the quality of reconstructions to a new level, but it still produces reconstructions with wrong thickness, inaccurate poses, and missing parts in many cases. This indicates that more efforts are still needed to enhance the fidelity of the reconstructions and to eliminate artifacts. Besides IF-based methods, there is another line of works that use deformation methods, which naturally do not produce artifacts like holes and broken parts, for 3D object reconstruction [12, 13, 45]. Despite several works [22, 39, 47, 55] have applied deformation methods to clothed 3D human reconstruction, none of them has successfully reconstructed humans with loose clothing from single images. On the one hand, large deformations are required to fit loose clothing; on the other hand, recovering surface details needs to shift the mesh's vertices delicately. + +In this paper, we propose VS, which stands for vertex shift, to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS is a two-stage deformation method that uses a "stretch-refine" strategy to reconcile the contradiction between large deformations required for reconstructing loose clothing and delicate deformations for recovering details. Specifically, VS first stretches the estimated SMPL-X mesh [34] into a coarse 3D human model using displacement fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network (GCN) embedded with implicit-function-learned features. In the stretching stage, VS shifts all vertices on the SMPL-X mesh to property positions where the coarse model can well align with the human in the input image. This enables VS to deal with loose clothing and correct pose/shape errors in the estimated SMPL-X. In the refining stage, the GCN driven by implicit-function-learned features can delicately deform the coarse model to recover the intricate and detailed surfaces of clothed humans. Moreover, each vertex on the graph model is constrained by its neighborhoods, which takes human body priors into account to suppress artifacts like distortions and non-human shapes. During the whole deforming process, VS only shifts vertices without altering edges, + +guaranteeing that it consistently produces holistic 3D human meshes without artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans on both benchmark datasets and in-the-wild images even under scenarios of challenging poses and loose clothing (See Fig.1). + +Our contributions are summarized as follows: + +(1) We propose a two-stage deformation method that uses a "stretch-refine" strategy for clothed 3D human reconstruction, contributing to reconciling the contradiction between large deformations for reconstructing loose clothing and delicate formations for recovering surface details. +(2) We introduce displacement fields inferred from normal maps for stretching the coarse model to align well with the input image, allowing our deformation method to handle loose clothing and correct inaccurate pose estimates. +(3) We combine implicit-function-learned features with a graph convolutional network, making VS not only recover surface details but also suppress artifacts. + +# 2. Related Work + +Body-agnostic methods. Given the remarkable expressive capability of deep implicit functions [8-10, 31, 32] for arbitrary 3D object surfaces, mainstream body-agnostic methods [1, 14, 23, 36, 37] employ them to represent clothed 3D humans. PIFu [36] proposes pixel-aligned image features for training the implicit function. PIFuHD [37] further enhances the geometric details by introducing a multi-level architecture that leverages both low- and high-resolution images. While PIFu and PIFuHD can recover intricate local shapes of 3D humans, they often suffer from artifacts such as disembodied limbs and non-human shapes, especially in the case of complex or novel poses. To reduce artifacts, Geo-PIFu [14] predicts a voxel-based coarse geometry model to regularize the deep implicit function. However, for humans with novel poses that are not included in the training data, Geo-PIFu often predicts poor coarse geometry models. Besides geometry, Stereo-PIFu [16] and Self-portrait [23] utilize depth information to regularize implicit functions, while PHORHUM [1] considers shading and lighting information to infer better geometric details. Since body-agnostic methods do not exploit the human body structure, which is a strong prior for 3D human reconstruction, such methods often struggle with artifacts. + +Body-aware methods. Body-aware methods [3, 5, 11, 15, 18, 19, 24-27, 42, 44, 48, 49, 52-54] utilize 3D parametric body models such as SMPL [28] and SMPL-X [34] or 2D joint points [6] as additional priors for clothed 3D human reconstruction. DeepHuman [53] fuses a semantic volume generated from the SMPL model with image features. However, due to an over-reliance on SMPL, DeepHuman is only able to reconstruct humans in tight clothing. PAMIR [54], HEI-Human [26], IP-Net [3], and DeepMult + +![](images/3be71975415f71a92397ddfddc356489c52a4cf95da5995070bfa1105602f762.jpg) +Figure 2. Overview of VS. VS employs a "stretch-refine" strategy to stepwise deform the SMPL-X $(\mathcal{M}_o)$ into a coarse human model $(\mathcal{M}_c)$ and a detailed human model $(\mathcal{M}_d)$ using StretchVS and RefineVS modules, respectively. (1) First, two displacement fields $(S = \{\mathcal{S}_f,\mathcal{S}_b\})$ are inferred by warping the body normals $(\mathcal{R} = \{\mathcal{R}_f,\mathcal{R}_b\})$ into clothing normal maps $(\mathcal{N} = \{\mathcal{N}_f,\mathcal{N}_b\})$ via a CNN (ShiftEstNet, $\mathcal{G}_s$ ). Then, StretchVS shifts vertices of the SMPL-X to from the coarse model using the displacement fields. (2) Taking the coarse model as input, RefineVS employs a GCN (DeformNet, $\mathcal{G}_{\mathrm{m}}$ ) embedded with implicit-function-learned features $(\Psi)$ to infer vertex locations of the detailed model. + +iCap [52] use voxelized SMPL to regularize free-form implicit functions. Nevertheless, this regularization restricts these methods' ability to reconstruct humans in loose clothing. Moreover, these methods are sensitive to the accuracy of SMPL estimates, and inaccurate SMPLs may introduce severe artifacts. ARCH [19], ARCH++ [15], FITE [25], and CAR [24] first learn the 3D shape of a human in the canonical space and then transform the learned shape into the posed space using SMPL. ICON [48] employs the SMPL model to guide both the detailed normals estimator and the visibility-aware implicit surface regressor. Recently, ECON [49] first utilizes BINI [4] to recover 2.5D front and back surfaces based on estimated normal maps, and then employs SMPL-X as a "canvas" to stitch together the two surfaces and "inpaint" the missing geometry. Although encouraging improvements have been achieved by these body-aware methods, they still inevitably produce artifacts and cannot yield high-fidelity reconstructions in many cases. + +Deformation methods for 3D object reconstruction. Our work is also inspired by the deformation-based 3D object reconstruction methods [13, 21, 22, 39, 45]. Pix2Mesh [45] employs a graph-based neural network to progressively deform ellipsoid meshes into 3D object models. Point2Mesh [13] iteratively shrinks and wraps a convex hull of the input point cloud to a watertight mesh. GCMR [21] uses a Graph-CNN to directly regress the 3D location of the down-sampled SMPL vertices. MGT-Net [22] utilizes a vertex-based deformation representation to model human shapes and employs cascaded multiscale graph transformation networks to generate topology-consistent 3D human models. Although holistic 3D models can be yielded, these methods are only suitable for reconstructing humans wearing tight clothing. + +# 3. Method + +An overview of VS is depicted in Fig. 2. Given an input image, VS first estimates an SMPL-X mesh and reorganizes it as a watertight 2-manifold graph (Sec. 3.1), then stretches the SMPL-X mesh into a coarse human mesh using displacement fields inferred from visual cues (Sec. 3.2), and finally refines the coarse human mesh by a GCN to form the final detailed human mesh (Sec. 3.3). + +# 3.1. Model Representation + +In VS, 3D meshes including the SMPL-X $(\mathcal{M}_o)$ , the coarse and detailed clothed human models $(\mathcal{M}_c$ and $\mathcal{M}_d)$ are represented by graphs: + +$$ +\mathcal {M} _ {*} = \left\langle \mathcal {V} _ {*}, \mathcal {E} \right\rangle \tag {1} +$$ + +where $* = \{o, c, d\}$ , $\mathcal{V}_* = \{\pmb{v}_*,^1, \pmb{v}_*,^2, \dots\}$ are 3D vertices, and $\mathcal{E} = \{e^1, e^2, \dots\}$ are edges. Particularly, we reorganize the vertices and edges to form watertight 2-manifold triangular meshes using [17]. The manifold and watertight properties ensure that our meshes have consistent surfaces without any holes, broken parts, or disembodied limbs, which are often observed in meshes reconstructed by IF-based methods. As mentioned above, VS stepwise deforms the SMPL-X model into coarse and detailed human models. In order to preserve the manifold and watertight properties of meshes, VS only shifts vertices without altering edges during the deforming process. + +The original SMPL-X[34] only has 10,475 vertices and 20,908 faces, and meshes deformed from it are challenging to accurately represent clothed 3D humans, especially when loose clothing is involved. To address this issue, we use the method proposed in [17] to increase the vertices and faces of the SMPL-X mesh. + +# 3.2. Vertex Shift for Coarse Model Stretching + +The StretchVS module stretches the SMPL-X mesh $(\mathcal{M}_o)$ into the coarse human mesh $(\mathcal{M}_c)$ using displacement fields $(S = \{\mathcal{S}_f, \mathcal{S}_b\}$ inferred from visual cues: + +$$ +\operatorname {S h i f t} \left(\mathcal {M} _ {o}, \mathcal {S}\right)\rightarrow \mathcal {M} _ {c} \Leftrightarrow \mathcal {V} _ {\mathrm {o}} + \mathcal {V} _ {\mathrm {s}} \rightarrow \mathcal {V} _ {\mathrm {c}} \tag {2} +$$ + +where $\mathcal{V}_{\mathrm{o}}$ and $\mathcal{V}_{\mathrm{c}}$ are respectively the vertices of $\mathcal{M}_o$ and $\mathcal{M}_c$ , and $\mathcal{V}_{\mathrm{s}} = \{\pmb{v}_{s}^{1},\pmb{v}_{s}^{2},\dots \}$ is a set of shift vectors derived from $\mathcal{S}$ . For each vertex $\pmb{v}_s^i$ in $\mathcal{V}_s$ : + +$$ +\boldsymbol {v} _ {s} ^ {i} = \left\{ \begin{array}{l l} \left\langle \mathcal {S} _ {f} \left(\pi \left(\boldsymbol {v} _ {o} ^ {i}\right)\right), 0 \right\rangle , & \text {i f} \boldsymbol {v} _ {o} ^ {i} \text {i s v i s i b l e} \\ \left\langle \mathcal {S} _ {b} \left(\pi \left(\boldsymbol {v} _ {o} ^ {i}\right)\right), 0 \right\rangle , & \text {e l s e} \end{array} \right. \tag {3} +$$ + +where and $\pi (\cdot)$ represents the orthogonal transformation that projects a 3D vertex onto the 2D displacement field plane. As described in Equ. 3, StretchVS shifts vertices on X-Y planes while maintaining the depth of SMPL-X, since it is challenging to estimate high-precision depth from 2D images using concise algorithms. Vertex shifts in the Z-direction will be handled by the RefineVS module later. + +The key to StretchVS is inferring the displacement fields from visual cues. Pioneering works like PIFuHD [37], ICON [48] and ECON [49] have already demonstrated the effectiveness of normals in clothed 3D human reconstruction. Encouraged by these works, we choose normals as the visual cues for inferring the displacement fields. + +Normals generation. We employ PyMAF-X [51] to estimate the SMPL-X, and then render two body normal maps $(\mathcal{R} = \{\mathcal{R}_f,\mathcal{R}_b\})$ from the SMPL-X respectively from the front and back views. To estimate clothing normal maps $(\mathcal{N} = \{\mathcal{N}_f,\mathcal{N}_b\})$ from the input image, the approach presented in ECON [49] is used. + +Displacement field inferring. In our method, the displacement fields are inferred by establishing the pixel-wise correspondences [41] between the body normal maps $(\mathcal{R})$ and the clothing normal maps $(\mathcal{N})$ . Once $\mathcal{S}$ is inferred, $\mathcal{R}$ can be warped into $\mathcal{N}$ : + +$$ +\hat {\mathcal {N}} _ {*} = \mathcal {W} \left(\mathcal {R} _ {*}, \mathcal {S} _ {*}\right) \tag {4} +$$ + +where $* = \{f, b\}$ , and $\mathcal{W}(\cdot)$ is the dense image warp function [2]. For simplicity, we omit the subscripts 'f' and 'b' for the front and back views in the subsequent formulas. + +We design a 5-layer lightweight CNN, named ShiftEst-Net $(\mathcal{G}_s)$ , to infer the displacement fields $(S)$ : + +$$ +\mathcal {G} _ {s} (\mathcal {R}, \mathcal {N}) \rightarrow \mathcal {S} \tag {5} +$$ + +Since the body normals and the clothing normals lay in different domains, there is no direct photometric evidence for establishing the dense correspondences between them. Inspired by [56] and [41], we propose a "cycle wrapping" strategy. To this end, ShiftEstNet is designed to have the capacity of inverse warping: + +$$ +\mathcal {G} _ {s} (\mathcal {N}, \mathcal {R}) \rightarrow \tilde {\mathcal {S}} \tag {6} +$$ + +where $\tilde{S}$ is the inferred displacement fields for inverse warping: + +$$ +\hat {\mathcal {R}} = \mathcal {W} (\mathcal {N}, \tilde {\mathcal {S}}) \tag {7} +$$ + +In a cycle, we can warp body normal maps into the clothing normal maps and then warp them back to body normal maps, and vice versa: + +$$ +\left\{ \begin{array}{l} \hat {\mathcal {R}} _ {c} = \mathcal {W} \left(\mathcal {W} (\mathcal {R}, \mathcal {S}), \tilde {\mathcal {S}}\right) \\ \hat {\mathcal {N}} _ {c} = \mathcal {W} \left(\mathcal {W} (\mathcal {N}, \tilde {\mathcal {S}}), \mathcal {S}\right) \end{array} \right. \tag {8} +$$ + +In this way, ShiftEstNet can be optimized in an self-supervised manner. The objective function consists of three items: + +$$ +\min _ {\mathcal {S}, \hat {\mathcal {S}}} \lambda_ {\mathrm {w}} \mathcal {L} _ {\mathrm {w}} + \lambda_ {\mathrm {c}} \mathcal {L} _ {\mathrm {c}} + \lambda_ {\mathrm {s}} \mathcal {L} _ {\mathrm {s}} \tag {9} +$$ + +where $\mathcal{L}_{\mathrm{w}}$ is a warping loss: + +$$ +\mathcal {L} _ {\mathrm {w}} = \left\| \mathcal {N} - \hat {\mathcal {N}} \right\| _ {2} ^ {2} + \left\| \mathcal {R} - \hat {\mathcal {R}} \right\| _ {2} ^ {2} \tag {10} +$$ + +$\mathcal{L}_{\mathrm{c}}$ is a cycle consistent loss: + +$$ +\mathcal {L} _ {\mathrm {c}} = \left\| \mathcal {N} - \hat {\mathcal {N}} _ {c} \right\| _ {2} ^ {2} + \left\| \mathcal {R} - \hat {\mathcal {R}} _ {c} \right\| _ {2} ^ {2} \tag {11} +$$ + +and $\mathcal{L}_{\mathrm{s}}$ is a smoothness loss that encourages the estimated displacement fields to be locally smooth [38]: + +$$ +\mathcal {L} _ {\mathrm {s}} = \left\| \mathcal {N} _ {g} - \hat {\mathcal {N}} _ {g} \right\| _ {2} ^ {2} \tag {12} +$$ + +$$ +\hat {\mathcal {N}} _ {g} = \mathcal {W} \left(\mathcal {W} \left(\mathcal {N} _ {g}, \tilde {\mathcal {S}}\right), \mathcal {S}\right) \tag {13} +$$ + +where $\mathcal{N}_g$ is an uniform-distributed random normal map. + +# 3.3. Vertex Shift for Detailed Model Refining + +Driven by embedded features $(\Psi)$ , the RefineVS module deforms the coarse human mesh $(\mathcal{M}_c)$ into the detailed human mesh $(\mathcal{M}_d)$ via a GCN named DeformNet $(\mathcal{G}_{\mathrm{m}})$ : + +$$ +\mathcal {G} _ {\mathrm {m}} \left(\mathcal {M} _ {c}; \Psi\right)\rightarrow \mathcal {M} _ {d} \tag {14} +$$ + +Network structure. DeformNet is designed using the mesh convolution operation proposed in MeshCNN [12]. Recall that $\mathcal{M}_c$ is a watertight 2-manifold mesh with triangular faces, which guarantees that each edge in the mesh is exactly adjacent to four other edges. Leveraging this invariant, we apply convolution operations to edges rather than vertices. The convolution for an edge $e_0$ and its four adjacent edges $\{e_1,e_2,e_3,e_4\}$ is defined as: + +$$ +\sum_ {k = 0} ^ {4} \boldsymbol {\mu} _ {k} \cdot \boldsymbol {\psi} _ {e _ {k}} \tag {15} +$$ + +where $[\pmb{\mu}_0, \dots, \pmb{\mu}_4]$ are convolutional kernels, and $\psi_{e_k}$ is the feature embedded on edge $e_k$ . + +$$ +\boldsymbol {\psi} _ {e _ {k}} = \boldsymbol {\psi} _ {v _ {i}} \oplus \boldsymbol {\psi} _ {v _ {j}} \tag {16} +$$ + +where $v_{i}$ and $v_{j}$ are the two vertices of edge $e_{k}$ , and $\oplus$ denotes the concatenate operator. + +DeformNet is designed as a fully convolutional network, so that intermediate layers and the output layer (i.e., the detailed mesh $\mathcal{M}_d$ ) after mesh convolution operations can maintain the same structure as the input mesh $\mathcal{M}_c$ . In our implementation, DeformNet consists of 11 mesh convolutional layers, each followed by a LeakyReLU activation. Particularly, the number of channels of the convolution kernel before the output layer is set to 6, allowing us to use the 6-dimensional convolutional result to represent the positions of the two vertices of an edge. Note that a vertex is shared by an uncertain number of edges, which means that DeformNet predicts multiple new positions for each vertex. We take the mean of the multiple positions as the final positions of the vertex. + +Features. From the seminal PIFu [36] to the current ECON [49], it has been demonstrated that "pixel-aligned" features trained using deep implicit functions achieve encouraging results in clothed 3D human reconstruction. In our method, we employ an encoder (denoted as $\mathcal{G}_e$ ) with a similar structure to PIFuHD [37] to extract visual features from the input image $(\mathcal{I})$ and the estimated normal maps $(\{\mathcal{N}_f,\mathcal{N}_b\})$ : + +$$ +\mathcal {G} _ {e} \left(\mathcal {I}, \mathcal {N} _ {f}, \mathcal {N} _ {b}\right)\rightarrow \Psi \tag {17} +$$ + +Although the encoder can be retrained, we have observed that using an off-the-shelf encoder is a more convenient and feasible solution. Thereby, we use the pre-trained encoder from PIFuHD [37]. Experiments show that DeformNet is not sensitive to feature encoders, and replacing the encoder with that of other methods like PIFu [36] and PaMIR [54] does not significantly decrease the quality of reconstructions (see "ablation study" in Sec. 4.4). In addition to visual features extracted by the encoder, we also embed the coordinate $(\pmb{v})$ and normal $(\mathcal{N}_v)$ of each vertex onto the vertex: + +$$ +\boldsymbol {\psi} _ {\boldsymbol {v}} = \left\langle \Psi (\pi (\boldsymbol {v})), \mathcal {N} _ {\boldsymbol {v}}, \boldsymbol {v} \right\rangle \tag {18} +$$ + +where $\pi (\pmb {v})$ denotes the orthogonal projection that maps a 3D vertex $\pmb{v}$ onto the 2D feature plane. + +Losses. The coarse model $\mathcal{M}_c$ input into DeformNet already has an approximate geometric topology, but it is not accurate enough and lacks local details. In addition, we expect not to introduce "bulge" artifacts when deforming $\mathcal{M}_c$ into $\mathcal{M}_d$ . Thus, DeformNet faces three tasks: (1) improving geometry accuracy, (2) recovering local details, and (3) preventing "bulge" artifacts. To these ends, the loss for training DeformNet consists of a geometric error term $(\mathcal{L}_{\mathrm{d}})$ , a local detail error term $(\mathcal{L}_{\mathrm{n}})$ , and a regularization term $(\mathcal{L}_{\mathrm{e}})$ : + +$$ +\mathcal {L} = \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\mathrm {n}} + \lambda_ {\mathrm {e}} \mathcal {L} _ {\mathrm {e}} \tag {19} +$$ + +Here, $\mathcal{L}_{\mathrm{d}}$ is described by Chamfer distance: + +$$ +\mathcal {L} _ {\mathrm {d}} = \sum_ {\hat {\boldsymbol {p}} \in \hat {\mathcal {P}}} \min _ {\boldsymbol {p} \in \mathcal {P}} \| \hat {\boldsymbol {p}} - \boldsymbol {p} \| _ {2} + \sum_ {\boldsymbol {q} \in \mathcal {P}} \min _ {\hat {\boldsymbol {q}} \in \hat {\mathcal {P}}} \| \hat {\boldsymbol {q}} - \boldsymbol {q} \| _ {2} \tag {20} +$$ + +where $\hat{\mathcal{P}}$ and $\mathcal{P}$ are 3D point sets respectively sampled from the predicted ground-truth meshes. Note that, we sample uniformly distributed 3D points using the "triangle point picking" approach [46] when computing $\mathcal{L}_{\mathrm{d}}$ , since the original vertices of the predicted and ground-truth meshes are usually not distributed uniformly on their surfaces. $\mathcal{L}_{\mathrm{n}}$ is described using the cosin distance of normals: + +$$ +\mathcal {L} _ {\mathrm {n}} = \sum_ {\boldsymbol {p}, \hat {\boldsymbol {p}}} - \frac {\boldsymbol {n} _ {\boldsymbol {p}} \cdot \boldsymbol {n} _ {\hat {\boldsymbol {p}}}}{\| \boldsymbol {n} _ {\boldsymbol {p}} \| _ {2} \| \boldsymbol {n} _ {\hat {\boldsymbol {p}}} \| _ {2}} + \sum_ {\boldsymbol {q}, \hat {\boldsymbol {q}}} - \frac {\boldsymbol {n} _ {\boldsymbol {q}} \cdot \boldsymbol {n} _ {\hat {\boldsymbol {q}}}}{\| \boldsymbol {n} _ {\boldsymbol {q}} \| _ {2} \| \boldsymbol {n} _ {\hat {\boldsymbol {q}}} \| _ {2}} \tag {21} +$$ + +where $(\pmb{p},\hat{\pmb{p}})$ and $(\pmb{q},\hat{\pmb{q}})$ are bi-directional matched closest point pairs, and $\pmb{n}$ represents the normal vector. The regularization term $\mathcal{L}_{\mathrm{e}}$ penalizes long edges: + +$$ +\mathcal {L} _ {\mathrm {e}} = \sum_ {e \in \mathcal {E}} \| \boldsymbol {v} _ {i} - \boldsymbol {v} _ {j} \| _ {2} \tag {22} +$$ + +where $\pmb{v}_i$ and $\pmb{v}_j$ are the two vertices of edge $e$ . $\mathcal{L}_{\mathrm{e}}$ not only prevents "bulges" artifacts, but also helps convergence. + +# 4. Experimental Results + +VS is compared with five state-of-the-art (SOTA) single-image clothed 3D human reconstruction methods with publicly released code: PIFu [36], PIFuHD [37], PaMIR [54], ICON [48], and ECON [49]. We opt for the $\mathrm{ECON_{EX}}$ version of ECON [49], as it is reported to have superior and stable performance over the $\mathrm{ECON_{IF}}$ version [49]. All methods employ the post-processing technique used in ICON. For a fair comparison, we use the re-implemented PIFu and PaMIR provided by Xiu et al. [48], which have uniformized network settings and inputs with ICON. As such, all methods involved in our experiments fairly take an RGB image and the estimated front/back normal maps as input. + +# 4.1. Datasets + +Training dataset. As with ECON [49], VS is trained on the THuman2.0 dataset [50], which contains 525 3D scans and SMPL-X models of 525 subjects with lab-acted poses and daily clothes. In our experiments, we split THuman2.0 into a training set and a testing set at a ratio of 20:1 by subjects. To augment the training data, we render 18 RGB images for each 3D scan by rotating a virtual camera w.r.t the frontal view at an interval of $20^{\circ}$ . In the end, 9K pairs (RGB image, 3D scan) are obtained for training. + +Testing datasets. Our VS is extensively evaluated on five datasets: the testing set of THuman 2.0 [50], CAPE [30], RenderPeople [35], SHHQ [20], and VcgPeople. Humans in these testing datasets exhibit a variety of poses and clothes, many of which are novel in the training data. For THuman2.0, CAPE, and RenderPeople, which have ground-truth 3D scans, we render 6 testing images from each scan at random viewpoints. Since SHHQ and VcgPeople just contain in-the-wild images, we only conduct qualitative evaluations on them. + +
MethodsPublicationsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]ICCV'191.7601.9040.05000.24082.9672.7380.04490.24092.7812.8570.05900.2773
†PIFuHD [37]CVPR'203.0883.1130.08910.36634.7143.8230.05550.27963.3113.31180.08460.3541
PaMIR [54]TPAMI'221.0641.1850.04380.19271.7721.4040.03370.16761.5801.6590.04860.2088
ICON [48]CVPR'220.9470.9250.04220.17611.1331.0960.03110.14311.2651.2510.04310.1871
ECON [49]CVPR'230.9060.8450.03790.18910.9370.9210.03350.16441.2851.0790.04170.1973
VSOurs0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
+ +# 4.2. Quantitative Evaluation + +Metrics. Chamfer distance $(\varepsilon_{cd}$ , unit: cm), point-to-surface distance $(\varepsilon_{p2s}$ , unit: cm), cosine distance $(\varepsilon_{cos} \in [0,1])$ , and L2 normal error $(\varepsilon_{l2} \in [0,1])$ are employed as quantitative metrics. The first two metrics mainly assess large geometric errors, while the latter two tend to measure local detail differences between reconstructions and ground-truth models. Following ECON [49], when reporting cosine distance and L2 normal error, four normal maps are rendered respectively for the reconstructed and ground-truth models at the angles of $\{0^o, 90^o, 180^o, 270^o\}$ . + +Quantitative comparison. As illustrated in Table 1, VS demonstrates superior performance on all three benchmark datasets. Especially, our VS exceeds all other methods by a large margin on the two metrics that represent large geometric errors (i.e., $\varepsilon_{cd}$ and $\varepsilon_{p2s}$ ). In terms of local surface details (i.e., $\varepsilon_{cos}$ and $\varepsilon_{L2}$ ), VS performs on par with ECON and much better than PIFu, PIFuHD, PaMIR, and ICON. + +# 4.3. Qualitative Evaluation + +Visual results. Fig. 4 shows the reconstructions by applying our VS on in-the-wild images. It can be seen that VS recovers high-fidelity clothed 3D humans even in challenging poses and complex loose clothes. Note that the poses, clothes, and accessories in these images are out of the distribution of the training data. These results reveal that VS not only has high reconstruction accuracy on challenging images but also has strong generalization capabilities. + +Visual comparison. Given that ECON [49] is the current SOTA method for single-image clothed 3D human reconstruction, we first compare our VS with it solely. Fig. 3 shows the reconstructions by ECON and our VS. It can be seen that VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. Fig. 5 demonstrates visual comparisons with all five SOTA methods on in-the-wild images. ECON and VS demonstrate a substantial advantage over PIFu [36], PIFuHD [37], PaMIR [54], and ICON [48]. + +Perceptual study. We also conduct a perceptual study on the in-the-wild SHHQ and VcgPeople datasets. To this end, we designed a perceptual evaluation tool (see Fig. 11 in SupMat) where reconstructions produced by different meth + +![](images/2081b4adff711e99ed34b9e08f7bc5813cea47e4c2d96e1afa27f088b25471fa.jpg) +Figure 3. Comparison of VS with ECON on in-the-wild images. VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. More comparisons are shown in Video2 in SupMat. + +ods are blindly placed on one page. We recruited 25 volunteers to participate in the perceptual study, and randomly sampled 50 images from each of the two datasets for each participant. Participants were requested to select the best reconstruction for each input image after rotating the reconstructions and carefully inspecting them from all viewpoints. The average percentages of preferences from these participants are illustrated in Table 2. The majority of participants perceived that our VS achieves the best reconstructions on both datasets. It's worth noting that ECON and VS exhibit greater superiorities over other methods on the more challenging VcgPeople dataset. + +Table 1. Quantitative comparison against other SOTAs on three benchmarks. Ground-truth SMPLs are used for methods that need SMPL(x) guidance. † Official model trained on Renderpeople. + +
DatasetsPIFuPIFuHDPaMIRICONECONVS
SHHQ3.64.311.13.919.058.1
VcgPeople0.20.15.65.024.764.4
+ +Table 2. Average percentages (\%) of participants' preferences. VS is significantly preferred over other methods. + +# 4.4. Ablation Study + +Effectiveness of the two vertex shift states. To assess the effectiveness of the StretchVS and RefineVS modules in VS, we implemented two variants of VS, each exclusively incorporating either the StretchVS module or the RefineVS module. The quantitative results achieved by the complete + +![](images/ce4d7c47b41e61d6afc3a0589ff67f6503a04b0ad6749bb4ce69fbb28f3e33ed.jpg) +Figure 4. Results produced by VS on in-the-wild images. Each result is shown from three views. VS recovers high-fidelity and artifact-less clothed 3D humans even in challenging poses (P) and loose clothes (C). More results are shown in Fig.12 and Video1 in SupMat. + +
VersionsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
VS0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
StretchVS0.9790.9150.06450.24690.9350.9070.04040.17431.2091.0400.06270.2443
RefineVS0.6690.5980.03910.16790.7210.7190.02710.12931.2630.9970.05290.2138
+ +Table 3. Quantitative results achieved by three versions of VS. + +VS and the two variants are presented in Table 3. It is evident from these results that: (1) StretchVS exhibits a significant performance decrease on all three datasets, especially in terms of local details. (2) RefineVS performs consistently on THuman2.0 and CAPE without significant degradation, while it exhibits a noticeable decline on RenderPeople. This is because RenderPeople contains many out-of-distribution loose clothes, and it is difficult for the RefineVS module to deform an SMPL-X mesh to fit loose clothes without the assistance of the StretchVS module. The visual results shown in Fig. 6 are mutually corroborated with the quantitative results illustrated in Table 3. + +Robustness to SMPL-X estimates. Most body-aware methods are known to be sensitive to SMPL-(X) estimates, while SMPL-(X) estimation remains an unresolved issue. In this experiment, we use both accurate and inaccurate SMPL-X estimates to initiate VS. As shown in Fig. 7, VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates. Since StretchVS treats SMPL-X just as a "seed" and stretches it to align with visual cues, it has the "error-correcting" ability to shift vertices on inaccurate SMPL-X estimates to proper positions. + +Flexibility to feature encoders. To validate that VS is insensitive to IF-based feature encoders, we evaluated VS + +![](images/2c5cbfa23ef8d740a8138bc74716a0eb529558f22197610441183f767423d4bf.jpg) +Figure 5. Visual comparison with other SOTA methods on in-the-wild images. More results are shown in Fig. $16\sim 19$ in SupMat. + +
EncodersTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]0.7660.6880.04110.17900.7350.7360.02720.12911.1040.8840.04650.1939
PaMIR [54]0.8380.7330.04020.17680.8610.7990.02660.13101.1820.9330.04280.1879
PIFuHD [37]0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
+ +Table 4. Quantitative evaluation of VS with different feature encoders. + +![](images/119374a63e0ddbe4089eb4681bf6b6511380d82c00ace9bc1ab260c9f0630a5f.jpg) +Figure 6. Results produced by VS variants w/wo StretchVS and RefineVS modules. StretchVS handles large deformations like loose clothing, while RefineVS recovers surface details. For humans who don't wear loose clothing, RefineVS can work alone. + +![](images/f996322bece70b52b8cfd11cc5d8b73a324e67a14099762e01223caf22f562c5.jpg) +Figure 7. Robustness to SMPL-X estimates. We use both accurate (odd rows) and inaccurate (even rows) SMPL-X estimates to initiate VS. VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates. + +with pre-trained encoders from PIFu [36], PaMIR [54], and PIFuHD [37]. As shown in Table 4, the employment of different feature encoders does not result in significant discrepancies in the quantitative metrics achieved by VS. + +# 5. Conclusion + +We propose VS to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS first stretches the estimated SMPL-X model into a coarse 3D human model using displacement fields inferred from visual cues, and then refines the coarse 3D human model into a de + +![](images/7862f9bd0f67eed98ff37b7663fd7b342715d5e9a5ad0f36874141a57bedb442.jpg) +Figure 8. Failure cases. VS may fail to handle extreme clothing, layer confusion, and large deformation in the Z-axis. + +tailed 3D human model via a GCN embedded with implicit-function-learned features. This two-stage deformation strategy reconciles the contradiction between large deformations for reconstructing loose clothing and delicate vertex shifts for recovering intricate and detailed surfaces. Extensive experiments on five datasets demonstrate that VS can reconstruct high-fidelity and artifact-less clothed 3D humans and achieves SOTA performance. VS confirms that deformation methods can reconstruct high-quality clothed 3D humans with complex poses and loose clothing, and even have advantages over IF-based methods in eliminating artifacts. We hope that VS can provide a new perspective for 3D human reconstruction and bring us one step closer to creating "perfect" human avatars for real-world applications. + +Limitations & Future work. Although VS achieves impressive results, it still has some limitations, as shown in Fig. 8. (A) Extreme clothing. VS may face challenges in deforming vertices on the legs of SMPL-X into extreme dresses. (B) Layer confusion. Currently, VS cannot distinguish the layers of the human body and clothing. (C) Wrong depth. VS fails in a couple of images exhibiting large deformation in the Z-axis (say, belts drifting towards the camera). In future work, we plan to separate a clothed human into the body and clothing parts and reconstruct them separately. As the clothing styles can be identified from the images [43] and the corresponding preset models can be provided [33], we hypothesize that using VS to deform preset clothing and body models separately can reconstruct extreme clothing and mitigate layer confusion. + +# 6. Acknowledgment + +This work was supported by the National Natural Science Foundation of China (62077026), and the Fundamental Research Funds for the Central Universities (CCNU22QN012). + +# References + +[1] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1506-1515, 2022. 2 +[2] The TensorFlow Authors. Dense image warp. www.tensorflow.org/addons/api_docs/python/tfa/image/dense_image_warp, 2023.4 +[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Eur. Conf. Comput. Vis. (ECCV), page 311-329, 2020. 2 +[4] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Eur. Conf. Comput. Vis. (ECCV), pages 552 - 567, 2022. 3 +[5] Yukang Cao, Kai Han, and Kwan-Yee K. Wong. Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4647-4657, 2023. 2 +[6] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7291-7299, 2017. 2 +[7] Lu Chen, Sida Peng, and Xiaowei Zhou. Towards efficient and photorealistic 3d human reconstruction: A brief survey. Visual Informatics, 5(4):11-19, 2021. 1 +[8] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 18522-18531, 2022. 2 +[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5932-5941, 2019. 2 +[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Int. Conf. Mach. Learn. (ICML), pages 3789-3799, 2020. 2 +[11] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12869-12879, 2023. 2 +[12] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes-Or, Shachar Fleishman, and Daniel. MeshCNN: A network with an edge. ACM Trans. Graph. (TOG), 38(4):90:1–90:12, 2019. 2, 4 +[13] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph. (TOG), 39(4):126:1-126:12, 2020. 2, 3, 1 +[14] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-PIFu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 2304–2314, 2020. 2 +[15] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\mathrm{ARCH}++$ : Animation-ready clothed human re + +construction revisited. In Int. Conf. Comput. Vis. (ICCV), pages 11046-11056, 2021. 2, 3 +[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. StereoPIFu: Depth aware clothed human digitization via stereo vision. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 119-128, 2021. 2 +[17] Jingwei Huang, Hao Su, and Leonidas J. Guibas. Robust watertight manifold surface generation method for shapenet models. ArXiv, 2018. 3, 1 +[18] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In Int. Conf. 3D Vis.(3DV), 2024. 2, 1 +[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH:animatable reconstruction of clothed humans. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3093-3102, 2020. 2, 3 +[20] Fu Jianglin, Li Shikai, Jiang Yuming, Lin Kwan-Yee, Qian Chen, Loy Chen Change, Wu Wayne, and Liu Ziwei. Stylegan-human: A data-centric odyssey of human generation. In Eur. Conf. Comput. Vis. (ECCV), pages 1-19, 2022. 5 +[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4501-4510, 2019. 3 +[22] Kun Li, Wen Hao, Feng Qiao, Yuxiang Zhang, Xiongzheng Li, Jing Huang, Cunkuan Yuan, Yu-Kun Lai, and Yebin Liu. Image-guided human reconstruction via multi-scale graph transformation networks. IEEE Trans. Image Process. (TIP), 30:5239-5251, 2021. 2, 3 +[23] Zhe Li, Tao Yu, Chuanyu Pan, Zerong Zheng, and Yebin Liu. Robust 3d self-portraits in seconds. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1344–1353, 2020. 2 +[24] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8662–8672, 2023. 2, 3 +[25] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In Eur. Conf. Comput. Vis. (ECCV), page 210-228, 2022. 3 +[26] Leyuan Liu, Jianchi Sun, Yunqi Gao, and Jingying Chen. HEI-Human: A hybrid explicit and implicit method for single-view 3d clothed human reconstruction. In Chinese Conf. Pattern Recog. Comput. Vis. (PRCV), pages 251-262, 2021. 2 +[27] Leyuan Liu, Yunqi Gao, Jianchi Sun, and Jingying Chen. Single-image clothed 3d human reconstruction guided by a well-aligned parametric body model. Multimedia Systems, page 1579-1592, 2023. 2 +[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 34(6):1-16, 2015. 2 + +[29] William Lorensen and Harvey Cline. Marching Cubes: A high resolution 3d surface construction algorithm. ACM Trans. Graph. (TOG), 21(4):163-169, 1987. 2 +[30] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3d people in generative clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6469-6478, 2020. 5, 1 +[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d reconstruction in function space. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4455-4465, 2019. 2 +[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 165-174, 2019. 2 +[33] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7365-7375, 2020. 8 +[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10975-10985, 2019. 2, 3 +[35] RenderPeople. www.renderpeople.com, 2018. 5, 1 +[36] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis. (ICCV), pages 2304-2314, 2019. 2, 5, 6, 8, 1 +[37] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 84-93, 2020. 2, 4, 5, 6, 8, 1 +[38] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I-Chao Chang, Yubo Fan, and Yan Xu. Unsupervised end-to-end learning for deformable medical image registrations. ArXiv, 2018. 4 +[39] Qingyang Tan, Lin Gao, Yu-Kun Lai, Jie Yang, and Shihong Xia1. Mesh-based autoencoders for localized deformation component analysis. In AAAI, page 2452-2459, 2018. 2, 3 +[40] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3D human mesh from monocular images: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), pages 15406 - 15425, 2023. 1 +[41] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Int. Conf. Comput. Vis. (ICCV), pages 10346-10356, 2021. 4 +[42] Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: Volumetric inference of 3d human body shapes. In Eur. Conf. Comput. Vis. (ECCV), pages 20-36, 2018. 2 +[43] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-YOLOv4: Scaling cross stage + +partial network. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13029-13038, 2021. 8 +[44] Junying Wang, Jae Shin Yoon, Tuanfeng Y. Wang, Krishna Kumar Singh, and Ulrich Neumann. Complete 3d human reconstruction from a single incomplete image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8748-8758, 2023. 2 +[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In Eur. Conf. Comput. Vis. (ECCV), pages 52-67, 2018. 2, 3 +[46] Eric W. Weissstein. Triangle point picking. mathworld. wolfram.com/TrianglePointPicking.html, 2023.5 +[47] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Int. Conf. 3D Vis.(3DV), pages 322-332, 2020. 2 +[48] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit clothed humans obtained from normals. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13296-13306, 2022. 2, 3, 4, 5, 6, 1 +[49] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit clothed humans optimized via normal integration. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 512-523, 2023. 2, 3, 4, 5, 6, 1 +[50] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5746-5756, 2021. 5, 1 +[51] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: towards well-aligned full-body model regression from monocular images. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 45(10):12287 - 12303, 2023. 4 +[52] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. DeepMultiCap: Performance capture of multiple characters using sparse multiview cameras. In Int. Conf. Comput. Vis. (ICCV), pages 6219-6229, 2021. 2, 3 +[53] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3d human reconstruction from a single image. In Int. Conf. Comput. Vis. (ICCV), pages 7739-7749, 2019. 2 +[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 44(6):3170 - 3184, 2022. 2, 5, 6, 8, 1 +[55] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4491-4500, 2019. 2 +[56] Jun-yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis. (ICCV), pages 2223-2232, 2017. 4 \ No newline at end of file diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/images.zip b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e423d47bafb8389f5a90d65affb04eeba4ed0b16 --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2ec4f39e555ed99b40c75448fe2ff88281b02f7c64805bc8cc5f95c8d0bf975 +size 672666 diff --git a/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/layout.json b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d87d241550a073faccf1244a5dafd78b4fd2866d --- /dev/null +++ b/2024/VS_ Reconstructing Clothed 3D Human from Single Image via Vertex Shift/layout.json @@ -0,0 +1,9107 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 529, + 121 + ], + "type": "text", + "content": "VS: Reconstructing Clothed 3D Human from Single Image via Vertex Shift" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 142, + 539, + 213 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": "Leyuan Liu" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": ", Yuhan Li" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": ", Yunqi Gao" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": ", Changxin Gao" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": ", Yuanyuan Liu" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "text", + "content": ", Jingying Chen" + }, + { + "bbox": [ + 82, + 142, + 509, + 157 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "content": " National Engineering Research Center for E-Learning, Central China Normal University " + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "content": " School of Artificial Intelligence and Automation, Huazhong University of Science and Technology " + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "content": " School of Computer Science, China University of Geosciences (Wuhan) {lyliu,chenjy} " + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "content": " mail.ccnu.edu.cn, {lyh985812,gaoyunqi} " + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 54, + 157, + 539, + 213 + ], + "type": "text", + "content": " mails.ccnu.edu.cn, cgao@hust.edu.cn, liuyy@cug.edu.cn" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 225, + 548, + 361 + ], + "blocks": [ + { + "bbox": [ + 48, + 225, + 548, + 361 + ], + "lines": [ + { + "bbox": [ + 48, + 225, + 548, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 548, + 361 + ], + "type": "image", + "image_path": "419cbacb89d2b0a9ca89e8ff5af4c7ab12497b193c85adaa0ba1d2e3c5cdfcc3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 369, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 45, + 369, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 45, + 369, + 547, + 413 + ], + "type": "text", + "content": "Figure 1. Using two stages of vertex shift, VS first stretches the estimated SMPL-X into a coarse human model and then refines it into a detailed clothed human model. Results on both benchmarks (1st row) and in-the-wild images (2nd row) show that VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothes. To comprehensively inspect results from all perspectives, we rotate the results through a full circle and show them from 8 views." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 440, + 192, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 440, + 192, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 440, + 192, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 462, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 462, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 462, + 290, + 715 + ], + "type": "text", + "content": "Various applications require high-fidelity and artifact-free 3D human reconstructions. However, current implicit function-based methods inevitably produce artifacts while existing deformation methods are difficult to reconstruct high-fidelity humans wearing loose clothing. In this paper, we propose a two-stage deformation method named Vertex Shift (VS) for reconstructing clothed 3D humans from single images. Specifically, VS first stretches the estimated SMPL-X mesh into a coarse 3D human model using shift fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network embedded with implicit-function-learned features. This \"stretch-refine\" strategy addresses large deformations required for reconstructing loose clothing and delicate deformations for recovering intricate and detailed surfaces, achieving high-fidelity reconstructions that faithfully convey the pose, clothing, and surface details from the input images. The graph convolutional network's ability to exploit neighborhood vertices coupled with the advantages inherited from the deformation methods ensure VS rarely produces artifacts like distortions and non-human shapes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 547, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 549 + ], + "type": "text", + "content": "and never produces artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothing. Experimental results on three benchmarks and two in-the-wild datasets demonstrate that VS significantly outperforms current state-of-the-art methods. The code and models of VS are available for research purposes at https://github.com/starVisionTeam/VS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 573, + 387, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 387, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 387, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Human avatars will be an essential medium for connecting the physical and digital worlds. Recent advancements in 3D human reconstruction [7, 40] have made it convenient to create human avatars even from single phone-taken images, but the reconstructions are not yet perfect. An ideal reconstruction should be high-fidelity and artifact-free. More specifically, it should faithfully convey the pose, clothing, and surface details of the human in the image while being free from artifacts like holes, broken or missing parts, disembodied limbs, distortions, and non-human shapes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10498" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": "Recently, encouraging reconstructions of humans with diverse clothing and poses have been yielded by implicit-function-based (IF-based) methods [14, 16, 36, 37]. They utilize deep neural networks to estimate occupancy fields [9, 31] or signed distance fields [32], from which detailed 3D human meshes can be reconstructed via MarchingCubes [29]. However, these IF-based methods often suffer from producing severe artifacts, as they do not exploit any knowledge about the human body structure. To address this issue, follow-up methods [3, 26, 48, 52, 54] utilize parametric body models [28, 34] to regularize the free-form implicit functions. While this regularization alleviates artifacts under the scenario of complex poses, it restricts the flexibility to loose clothing. The most recent work, ECON [49], has improved the quality of reconstructions to a new level, but it still produces reconstructions with wrong thickness, inaccurate poses, and missing parts in many cases. This indicates that more efforts are still needed to enhance the fidelity of the reconstructions and to eliminate artifacts. Besides IF-based methods, there is another line of works that use deformation methods, which naturally do not produce artifacts like holes and broken parts, for 3D object reconstruction [12, 13, 45]. Despite several works [22, 39, 47, 55] have applied deformation methods to clothed 3D human reconstruction, none of them has successfully reconstructed humans with loose clothing from single images. On the one hand, large deformations are required to fit loose clothing; on the other hand, recovering surface details needs to shift the mesh's vertices delicately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we propose VS, which stands for vertex shift, to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS is a two-stage deformation method that uses a \"stretch-refine\" strategy to reconcile the contradiction between large deformations required for reconstructing loose clothing and delicate deformations for recovering details. Specifically, VS first stretches the estimated SMPL-X mesh [34] into a coarse 3D human model using displacement fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network (GCN) embedded with implicit-function-learned features. In the stretching stage, VS shifts all vertices on the SMPL-X mesh to property positions where the coarse model can well align with the human in the input image. This enables VS to deal with loose clothing and correct pose/shape errors in the estimated SMPL-X. In the refining stage, the GCN driven by implicit-function-learned features can delicately deform the coarse model to recover the intricate and detailed surfaces of clothed humans. Moreover, each vertex on the graph model is constrained by its neighborhoods, which takes human body priors into account to suppress artifacts like distortions and non-human shapes. During the whole deforming process, VS only shifts vertices without altering edges," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "guaranteeing that it consistently produces holistic 3D human meshes without artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans on both benchmark datasets and in-the-wild images even under scenarios of challenging poses and loose clothing (See Fig.1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 144, + 503, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 144, + 503, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 503, + 155 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 156, + 547, + 300 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 304, + 156, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 215 + ], + "type": "text", + "content": "(1) We propose a two-stage deformation method that uses a \"stretch-refine\" strategy for clothed 3D human reconstruction, contributing to reconciling the contradiction between large deformations for reconstructing loose clothing and delicate formations for recovering surface details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "content": "(2) We introduce displacement fields inferred from normal maps for stretching the coarse model to align well with the input image, allowing our deformation method to handle loose clothing and correct inaccurate pose estimates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 264, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 300 + ], + "type": "text", + "content": "(3) We combine implicit-function-learned features with a graph convolutional network, making VS not only recover surface details but also suppress artifacts." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 311, + 392, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 392, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 392, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 605 + ], + "type": "text", + "content": "Body-agnostic methods. Given the remarkable expressive capability of deep implicit functions [8-10, 31, 32] for arbitrary 3D object surfaces, mainstream body-agnostic methods [1, 14, 23, 36, 37] employ them to represent clothed 3D humans. PIFu [36] proposes pixel-aligned image features for training the implicit function. PIFuHD [37] further enhances the geometric details by introducing a multi-level architecture that leverages both low- and high-resolution images. While PIFu and PIFuHD can recover intricate local shapes of 3D humans, they often suffer from artifacts such as disembodied limbs and non-human shapes, especially in the case of complex or novel poses. To reduce artifacts, Geo-PIFu [14] predicts a voxel-based coarse geometry model to regularize the deep implicit function. However, for humans with novel poses that are not included in the training data, Geo-PIFu often predicts poor coarse geometry models. Besides geometry, Stereo-PIFu [16] and Self-portrait [23] utilize depth information to regularize implicit functions, while PHORHUM [1] considers shading and lighting information to infer better geometric details. Since body-agnostic methods do not exploit the human body structure, which is a strong prior for 3D human reconstruction, such methods often struggle with artifacts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": "Body-aware methods. Body-aware methods [3, 5, 11, 15, 18, 19, 24-27, 42, 44, 48, 49, 52-54] utilize 3D parametric body models such as SMPL [28] and SMPL-X [34] or 2D joint points [6] as additional priors for clothed 3D human reconstruction. DeepHuman [53] fuses a semantic volume generated from the SMPL model with image features. However, due to an over-reliance on SMPL, DeepHuman is only able to reconstruct humans in tight clothing. PAMIR [54], HEI-Human [26], IP-Net [3], and DeepMult" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10499" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 218 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 218 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 218 + ], + "type": "image", + "image_path": "3be71975415f71a92397ddfddc356489c52a4cf95da5995070bfa1105602f762.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of VS. VS employs a \"stretch-refine\" strategy to stepwise deform the SMPL-X " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_o)" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " into a coarse human model " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_c)" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " and a detailed human model " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_d)" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " using StretchVS and RefineVS modules, respectively. (1) First, two displacement fields " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(S = \\{\\mathcal{S}_f,\\mathcal{S}_b\\})" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " are inferred by warping the body normals " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " into clothing normal maps " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " via a CNN (ShiftEstNet, " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_s" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": "). Then, StretchVS shifts vertices of the SMPL-X to from the coarse model using the displacement fields. (2) Taking the coarse model as input, RefineVS employs a GCN (DeformNet, " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": ") embedded with implicit-function-learned features " + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\Psi)" + }, + { + "bbox": [ + 46, + 226, + 547, + 293 + ], + "type": "text", + "content": " to infer vertex locations of the detailed model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 289, + 529 + ], + "type": "text", + "content": "iCap [52] use voxelized SMPL to regularize free-form implicit functions. Nevertheless, this regularization restricts these methods' ability to reconstruct humans in loose clothing. Moreover, these methods are sensitive to the accuracy of SMPL estimates, and inaccurate SMPLs may introduce severe artifacts. ARCH [19], ARCH++ [15], FITE [25], and CAR [24] first learn the 3D shape of a human in the canonical space and then transform the learned shape into the posed space using SMPL. ICON [48] employs the SMPL model to guide both the detailed normals estimator and the visibility-aware implicit surface regressor. Recently, ECON [49] first utilizes BINI [4] to recover 2.5D front and back surfaces based on estimated normal maps, and then employs SMPL-X as a \"canvas\" to stitch together the two surfaces and \"inpaint\" the missing geometry. Although encouraging improvements have been achieved by these body-aware methods, they still inevitably produce artifacts and cannot yield high-fidelity reconstructions in many cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "Deformation methods for 3D object reconstruction. Our work is also inspired by the deformation-based 3D object reconstruction methods [13, 21, 22, 39, 45]. Pix2Mesh [45] employs a graph-based neural network to progressively deform ellipsoid meshes into 3D object models. Point2Mesh [13] iteratively shrinks and wraps a convex hull of the input point cloud to a watertight mesh. GCMR [21] uses a Graph-CNN to directly regress the 3D location of the down-sampled SMPL vertices. MGT-Net [22] utilizes a vertex-based deformation representation to model human shapes and employs cascaded multiscale graph transformation networks to generate topology-consistent 3D human models. Although holistic 3D models can be yielded, these methods are only suitable for reconstructing humans wearing tight clothing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 312, + 361, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 361, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 361, + 324 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 332, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 416 + ], + "type": "text", + "content": "An overview of VS is depicted in Fig. 2. Given an input image, VS first estimates an SMPL-X mesh and reorganizes it as a watertight 2-manifold graph (Sec. 3.1), then stretches the SMPL-X mesh into a coarse human mesh using displacement fields inferred from visual cues (Sec. 3.2), and finally refines the coarse human mesh by a GCN to form the final detailed human mesh (Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 422, + 432, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 432, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 432, + 434 + ], + "type": "text", + "content": "3.1. Model Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": "In VS, 3D meshes including the SMPL-X " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_o)" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": ", the coarse and detailed clothed human models " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_d)" + }, + { + "bbox": [ + 304, + 440, + 545, + 475 + ], + "type": "text", + "content": " are represented by graphs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 394, + 480, + 545, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 480, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 394, + 480, + 545, + 494 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {*} = \\left\\langle \\mathcal {V} _ {*}, \\mathcal {E} \\right\\rangle \\tag {1}", + "image_path": "5b1ff836ff0c8f34eabdd5a22a2451842f3eca643b4a9ce83251f746d87ccbd4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "inline_equation", + "content": "* = \\{o, c, d\\}" + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_* = \\{\\pmb{v}_*,^1, \\pmb{v}_*,^2, \\dots\\}" + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "text", + "content": " are 3D vertices, and " + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e^1, e^2, \\dots\\}" + }, + { + "bbox": [ + 304, + 498, + 546, + 641 + ], + "type": "text", + "content": " are edges. Particularly, we reorganize the vertices and edges to form watertight 2-manifold triangular meshes using [17]. The manifold and watertight properties ensure that our meshes have consistent surfaces without any holes, broken parts, or disembodied limbs, which are often observed in meshes reconstructed by IF-based methods. As mentioned above, VS stepwise deforms the SMPL-X model into coarse and detailed human models. In order to preserve the manifold and watertight properties of meshes, VS only shifts vertices without altering edges during the deforming process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "The original SMPL-X[34] only has 10,475 vertices and 20,908 faces, and meshes deformed from it are challenging to accurately represent clothed 3D humans, especially when loose clothing is involved. To address this issue, we use the method proposed in [17] to increase the vertices and faces of the SMPL-X mesh." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10500" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "type": "text", + "content": "3.2. Vertex Shift for Coarse Model Stretching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "text", + "content": "The StretchVS module stretches the SMPL-X mesh " + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_o)" + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "text", + "content": " into the coarse human mesh " + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_c)" + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "text", + "content": " using displacement fields " + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "inline_equation", + "content": "(S = \\{\\mathcal{S}_f, \\mathcal{S}_b\\}" + }, + { + "bbox": [ + 47, + 89, + 287, + 126 + ], + "type": "text", + "content": " inferred from visual cues:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 85, + 132, + 287, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 132, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 85, + 132, + 287, + 144 + ], + "type": "interline_equation", + "content": "\\operatorname {S h i f t} \\left(\\mathcal {M} _ {o}, \\mathcal {S}\\right)\\rightarrow \\mathcal {M} _ {c} \\Leftrightarrow \\mathcal {V} _ {\\mathrm {o}} + \\mathcal {V} _ {\\mathrm {s}} \\rightarrow \\mathcal {V} _ {\\mathrm {c}} \\tag {2}", + "image_path": "6f102e1776e8a08a8be4cb0050b6447f88bdb4a50a88c12e62974913c9a156bc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{o}}" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{c}}" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": " are respectively the vertices of " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_o" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{s}} = \\{\\pmb{v}_{s}^{1},\\pmb{v}_{s}^{2},\\dots \\}" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": " is a set of shift vectors derived from " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": ". For each vertex " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{v}_s^i" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s" + }, + { + "bbox": [ + 47, + 149, + 287, + 187 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 191, + 287, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 191, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 80, + 191, + 287, + 223 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\langle \\mathcal {S} _ {f} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {i f} \\boldsymbol {v} _ {o} ^ {i} \\text {i s v i s i b l e} \\\\ \\left\\langle \\mathcal {S} _ {b} \\left(\\pi \\left(\\boldsymbol {v} _ {o} ^ {i}\\right)\\right), 0 \\right\\rangle , & \\text {e l s e} \\end{array} \\right. \\tag {3}", + "image_path": "8496716144d59a71cc66d2cb2642d0675d5efbbcf869b681b0fd4282146c73ac.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "text", + "content": "where and " + }, + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 46, + 228, + 287, + 312 + ], + "type": "text", + "content": " represents the orthogonal transformation that projects a 3D vertex onto the 2D displacement field plane. As described in Equ. 3, StretchVS shifts vertices on X-Y planes while maintaining the depth of SMPL-X, since it is challenging to estimate high-precision depth from 2D images using concise algorithms. Vertex shifts in the Z-direction will be handled by the RefineVS module later." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": "The key to StretchVS is inferring the displacement fields from visual cues. Pioneering works like PIFuHD [37], ICON [48] and ECON [49] have already demonstrated the effectiveness of normals in clothed 3D human reconstruction. Encouraged by these works, we choose normals as the visual cues for inferring the displacement fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "text", + "content": "Normals generation. We employ PyMAF-X [51] to estimate the SMPL-X, and then render two body normal maps " + }, + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "inline_equation", + "content": "(\\mathcal{R} = \\{\\mathcal{R}_f,\\mathcal{R}_b\\})" + }, + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "text", + "content": " from the SMPL-X respectively from the front and back views. To estimate clothing normal maps " + }, + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "inline_equation", + "content": "(\\mathcal{N} = \\{\\mathcal{N}_f,\\mathcal{N}_b\\})" + }, + { + "bbox": [ + 47, + 384, + 287, + 455 + ], + "type": "text", + "content": " from the input image, the approach presented in ECON [49] is used." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": "Displacement field inferring. In our method, the displacement fields are inferred by establishing the pixel-wise correspondences [41] between the body normal maps " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "(\\mathcal{R})" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": " and the clothing normal maps " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "(\\mathcal{N})" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": ". Once " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": " is inferred, " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": " can be warped into " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 520, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 520, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 128, + 520, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {N}} _ {*} = \\mathcal {W} \\left(\\mathcal {R} _ {*}, \\mathcal {S} _ {*}\\right) \\tag {4}", + "image_path": "86b0c0f138649ef8e02e6ecf917936032ec9a24aead46d461a5bf6d93dd1b1ba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "inline_equation", + "content": "* = \\{f, b\\}" + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot)" + }, + { + "bbox": [ + 46, + 539, + 287, + 574 + ], + "type": "text", + "content": " is the dense image warp function [2]. For simplicity, we omit the subscripts 'f' and 'b' for the front and back views in the subsequent formulas." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "content": "We design a 5-layer lightweight CNN, named ShiftEst-Net " + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_s)" + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "content": ", to infer the displacement fields " + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 605, + 287, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 287, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {s} (\\mathcal {R}, \\mathcal {N}) \\rightarrow \\mathcal {S} \\tag {5}", + "image_path": "947bcebac1ccabac96e8b940593bab912e997c2aa1a2b06019a3f32b173e578d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "content": "Since the body normals and the clothing normals lay in different domains, there is no direct photometric evidence for establishing the dense correspondences between them. Inspired by [56] and [41], we propose a \"cycle wrapping\" strategy. To this end, ShiftEstNet is designed to have the capacity of inverse warping:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 700, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 700, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 133, + 700, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {s} (\\mathcal {N}, \\mathcal {R}) \\rightarrow \\tilde {\\mathcal {S}} \\tag {6}", + "image_path": "33fe19a9a4a862653b05620226986c31d7a4b195c246a19802038e22a82842c7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " is the inferred displacement fields for inverse warping:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 394, + 95, + 545, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 95, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 394, + 95, + 545, + 108 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {R}} = \\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}) \\tag {7}", + "image_path": "7a71a08c0e794da55788816c712c7de8f92b8b2bd8bfee9a1960b486fac80133.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 114, + 545, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 545, + 149 + ], + "type": "text", + "content": "In a cycle, we can warp body normal maps into the clothing normal maps and then warp them back to body normal maps, and vice versa:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 156, + 545, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 156, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 365, + 156, + 545, + 195 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {\\mathcal {R}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {R}, \\mathcal {S}), \\tilde {\\mathcal {S}}\\right) \\\\ \\hat {\\mathcal {N}} _ {c} = \\mathcal {W} \\left(\\mathcal {W} (\\mathcal {N}, \\tilde {\\mathcal {S}}), \\mathcal {S}\\right) \\end{array} \\right. \\tag {8}", + "image_path": "eba527ffcdd7db8c0a761e5082cf687c638a1eb27e2fddbdeeb69bd27e3bc85b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 201, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 236 + ], + "type": "text", + "content": "In this way, ShiftEstNet can be optimized in an self-supervised manner. The objective function consists of three items:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 237, + 545, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 237, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 369, + 237, + 545, + 257 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {S}, \\hat {\\mathcal {S}}} \\lambda_ {\\mathrm {w}} \\mathcal {L} _ {\\mathrm {w}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} \\tag {9}", + "image_path": "7677d02e5671f8a4583e469cc5fa404ab51b065dab4b113b8b469ef466aecea2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 262, + 421, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 421, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 421, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 262, + 421, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{w}}" + }, + { + "bbox": [ + 306, + 262, + 421, + 274 + ], + "type": "text", + "content": " is a warping loss:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 281, + 545, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 281, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 359, + 281, + 545, + 295 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {w}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\tag {10}", + "image_path": "8866af063dc5c208279e077d83e2005cb539dfdb490752d2e9a7d108460576e4.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 303, + 422, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 422, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 422, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 306, + 303, + 422, + 315 + ], + "type": "text", + "content": " is a cycle consistent loss:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 358, + 323, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 323, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 358, + 323, + 545, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\left\\| \\mathcal {N} - \\hat {\\mathcal {N}} _ {c} \\right\\| _ {2} ^ {2} + \\left\\| \\mathcal {R} - \\hat {\\mathcal {R}} _ {c} \\right\\| _ {2} ^ {2} \\tag {11}", + "image_path": "08c34c6e3f792ee01d104c6064bb3f72e2186fd3c2cf7b1a2550cefda970f1e2.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 345, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 545, + 369 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 305, + 345, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 305, + 345, + 545, + 369 + ], + "type": "text", + "content": " is a smoothness loss that encourages the estimated displacement fields to be locally smooth [38]:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 386, + 376, + 545, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 376, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 386, + 376, + 545, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s}} = \\left\\| \\mathcal {N} _ {g} - \\hat {\\mathcal {N}} _ {g} \\right\\| _ {2} ^ {2} \\tag {12}", + "image_path": "7ed1cdbec7bec2e157937d0a0f3da4ce55b49cef42c208f5e28d20366d3a3bff.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 398, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 398, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 372, + 398, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {N}} _ {g} = \\mathcal {W} \\left(\\mathcal {W} \\left(\\mathcal {N} _ {g}, \\tilde {\\mathcal {S}}\\right), \\mathcal {S}\\right) \\tag {13}", + "image_path": "4e942f6cc5d74124b431f73d439c64534d47aae5d699f86c7b4348bfa89557b9.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 421, + 535, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 535, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 535, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 421, + 535, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_g" + }, + { + "bbox": [ + 306, + 421, + 535, + 434 + ], + "type": "text", + "content": " is an uniform-distributed random normal map." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 440, + 517, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 517, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 517, + 453 + ], + "type": "text", + "content": "3.3. Vertex Shift for Detailed Model Refining" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "content": "Driven by embedded features " + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "inline_equation", + "content": "(\\Psi)" + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "content": ", the RefineVS module deforms the coarse human mesh " + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_c)" + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "content": " into the detailed human mesh " + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_d)" + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "content": " via a GCN named DeformNet " + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_{\\mathrm{m}})" + }, + { + "bbox": [ + 305, + 458, + 545, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 382, + 503, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 503, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 382, + 503, + 545, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {\\mathrm {m}} \\left(\\mathcal {M} _ {c}; \\Psi\\right)\\rightarrow \\mathcal {M} _ {d} \\tag {14}", + "image_path": "9caa40aa12633cd717beb83f75c483bf7ba0a4771bcf3cbd6051ee989bb33170.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": "Network structure. DeformNet is designed using the mesh convolution operation proposed in MeshCNN [12]. Recall that " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": " is a watertight 2-manifold mesh with triangular faces, which guarantees that each edge in the mesh is exactly adjacent to four other edges. Leveraging this invariant, we apply convolution operations to edges rather than vertices. The convolution for an edge " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "e_0" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": " and its four adjacent edges " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\{e_1,e_2,e_3,e_4\\}" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 398, + 628, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 628, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 398, + 628, + 545, + 661 + ], + "type": "interline_equation", + "content": "\\sum_ {k = 0} ^ {4} \\boldsymbol {\\mu} _ {k} \\cdot \\boldsymbol {\\psi} _ {e _ {k}} \\tag {15}", + "image_path": "4a9e33588b71a50645fc66534e350cd5e1c893579ec3afbb0cf572bb7cc2fa06.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "[\\pmb{\\mu}_0, \\dots, \\pmb{\\mu}_4]" + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": " are convolutional kernels, and " + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\psi_{e_k}" + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": " is the feature embedded on edge " + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 388, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 388, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\psi} _ {e _ {k}} = \\boldsymbol {\\psi} _ {v _ {i}} \\oplus \\boldsymbol {\\psi} _ {v _ {j}} \\tag {16}", + "image_path": "55296a639d93b16d3707e3c00b64bf6dca3e7eeeb87f3812ac83fe8a7018e887.jpg" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10501" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " are the two vertices of edge " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "e_{k}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " denotes the concatenate operator." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "text", + "content": "DeformNet is designed as a fully convolutional network, so that intermediate layers and the output layer (i.e., the detailed mesh " + }, + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_d" + }, + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "text", + "content": ") after mesh convolution operations can maintain the same structure as the input mesh " + }, + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 96, + 287, + 262 + ], + "type": "text", + "content": ". In our implementation, DeformNet consists of 11 mesh convolutional layers, each followed by a LeakyReLU activation. Particularly, the number of channels of the convolution kernel before the output layer is set to 6, allowing us to use the 6-dimensional convolutional result to represent the positions of the two vertices of an edge. Note that a vertex is shared by an uncertain number of edges, which means that DeformNet predicts multiple new positions for each vertex. We take the mean of the multiple positions as the final positions of the vertex." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "text", + "content": "Features. From the seminal PIFu [36] to the current ECON [49], it has been demonstrated that \"pixel-aligned\" features trained using deep implicit functions achieve encouraging results in clothed 3D human reconstruction. In our method, we employ an encoder (denoted as " + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e" + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "text", + "content": ") with a similar structure to PIFuHD [37] to extract visual features from the input image " + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "text", + "content": " and the estimated normal maps " + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "inline_equation", + "content": "(\\{\\mathcal{N}_f,\\mathcal{N}_b\\})" + }, + { + "bbox": [ + 47, + 263, + 287, + 359 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 360, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 360, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 123, + 360, + 287, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {e} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right)\\rightarrow \\Psi \\tag {17}", + "image_path": "3a3732381b41f4a664c550201de7099276eff7670b44603d6dfefa5060c10beb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "content": "Although the encoder can be retrained, we have observed that using an off-the-shelf encoder is a more convenient and feasible solution. Thereby, we use the pre-trained encoder from PIFuHD [37]. Experiments show that DeformNet is not sensitive to feature encoders, and replacing the encoder with that of other methods like PIFu [36] and PaMIR [54] does not significantly decrease the quality of reconstructions (see \"ablation study\" in Sec. 4.4). In addition to visual features extracted by the encoder, we also embed the coordinate " + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "inline_equation", + "content": "(\\pmb{v})" + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "inline_equation", + "content": "(\\mathcal{N}_v)" + }, + { + "bbox": [ + 46, + 374, + 287, + 494 + ], + "type": "text", + "content": " of each vertex onto the vertex:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 499, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 499, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 287, + 512 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\psi} _ {\\boldsymbol {v}} = \\left\\langle \\Psi (\\pi (\\boldsymbol {v})), \\mathcal {N} _ {\\boldsymbol {v}}, \\boldsymbol {v} \\right\\rangle \\tag {18}", + "image_path": "358b98b23b7ad5de442af07a72b6e7f2b511085a86ab21fcac83aca9e9a350e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\pi (\\pmb {v})" + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": " denotes the orthogonal projection that maps a 3D vertex " + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 47, + 517, + 287, + 540 + ], + "type": "text", + "content": " onto the 2D feature plane." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": "Losses. The coarse model " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": " input into DeformNet already has an approximate geometric topology, but it is not accurate enough and lacks local details. In addition, we expect not to introduce \"bulge\" artifacts when deforming " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_d" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": ". Thus, DeformNet faces three tasks: (1) improving geometry accuracy, (2) recovering local details, and (3) preventing \"bulge\" artifacts. To these ends, the loss for training DeformNet consists of a geometric error term " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{d}})" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": ", a local detail error term " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{n}})" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": ", and a regularization term " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{e}})" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 654, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 654, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 110, + 654, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} \\tag {19}", + "image_path": "fe673e742ee6fc4cf6c73a50209d6b49d158f966e37a393da94abff15f8d1904.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 224, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 224, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 224, + 683 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 671, + 224, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 47, + 671, + 224, + 683 + ], + "type": "text", + "content": " is described by Chamfer distance:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 687, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 687, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 63, + 687, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d}} = \\sum_ {\\hat {\\boldsymbol {p}} \\in \\hat {\\mathcal {P}}} \\min _ {\\boldsymbol {p} \\in \\mathcal {P}} \\| \\hat {\\boldsymbol {p}} - \\boldsymbol {p} \\| _ {2} + \\sum_ {\\boldsymbol {q} \\in \\mathcal {P}} \\min _ {\\hat {\\boldsymbol {q}} \\in \\hat {\\mathcal {P}}} \\| \\hat {\\boldsymbol {q}} - \\boldsymbol {q} \\| _ {2} \\tag {20}", + "image_path": "c9e710c4406e6259cd5229c46f028620289730a8cf92d3c2984933ea6b27797d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " are 3D point sets respectively sampled from the predicted ground-truth meshes. Note that, we sample uniformly distributed 3D points using the \"triangle point picking\" approach [46] when computing " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", since the original vertices of the predicted and ground-truth meshes are usually not distributed uniformly on their surfaces. " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{n}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is described using the cosin distance of normals:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 159, + 545, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 159, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 315, + 159, + 545, + 189 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n}} = \\sum_ {\\boldsymbol {p}, \\hat {\\boldsymbol {p}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {p}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {p}}} \\| _ {2}} + \\sum_ {\\boldsymbol {q}, \\hat {\\boldsymbol {q}}} - \\frac {\\boldsymbol {n} _ {\\boldsymbol {q}} \\cdot \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}}}{\\| \\boldsymbol {n} _ {\\boldsymbol {q}} \\| _ {2} \\| \\boldsymbol {n} _ {\\hat {\\boldsymbol {q}}} \\| _ {2}} \\tag {21}", + "image_path": "fa39b7a449bbc6a3c13deb38973f6c1e700340954ea63bd77f510799e85932a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "inline_equation", + "content": "(\\pmb{p},\\hat{\\pmb{p}})" + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "inline_equation", + "content": "(\\pmb{q},\\hat{\\pmb{q}})" + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "content": " are bi-directional matched closest point pairs, and " + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "content": " represents the normal vector. The regularization term " + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 194, + 545, + 230 + ], + "type": "text", + "content": " penalizes long edges:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 235, + 545, + 260 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 235, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 381, + 235, + 545, + 260 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e}} = \\sum_ {e \\in \\mathcal {E}} \\| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\| _ {2} \\tag {22}", + "image_path": "30a6a2e3aed2d49ba38a46abbf23ddb1b6f63ffca4f9fe9efddd4d5097a5f65b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{v}_j" + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "content": " are the two vertices of edge " + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 265, + 545, + 289 + ], + "type": "text", + "content": " not only prevents \"bulges\" artifacts, but also helps convergence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 298, + 432, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 432, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 432, + 312 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "text", + "content": "VS is compared with five state-of-the-art (SOTA) single-image clothed 3D human reconstruction methods with publicly released code: PIFu [36], PIFuHD [37], PaMIR [54], ICON [48], and ECON [49]. We opt for the " + }, + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{ECON_{EX}}" + }, + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "text", + "content": " version of ECON [49], as it is reported to have superior and stable performance over the " + }, + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{ECON_{IF}}" + }, + { + "bbox": [ + 304, + 318, + 545, + 462 + ], + "type": "text", + "content": " version [49]. All methods employ the post-processing technique used in ICON. For a fair comparison, we use the re-implemented PIFu and PaMIR provided by Xiu et al. [48], which have uniformized network settings and inputs with ICON. As such, all methods involved in our experiments fairly take an RGB image and the estimated front/back normal maps as input." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 468, + 368, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 368, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 368, + 479 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Training dataset. As with ECON [49], VS is trained on the THuman2.0 dataset [50], which contains 525 3D scans and SMPL-X models of 525 subjects with lab-acted poses and daily clothes. In our experiments, we split THuman2.0 into a training set and a testing set at a ratio of 20:1 by subjects. To augment the training data, we render 18 RGB images for each 3D scan by rotating a virtual camera w.r.t the frontal view at an interval of " + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": ". In the end, 9K pairs (RGB image, 3D scan) are obtained for training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Testing datasets. Our VS is extensively evaluated on five datasets: the testing set of THuman 2.0 [50], CAPE [30], RenderPeople [35], SHHQ [20], and VcgPeople. Humans in these testing datasets exhibit a variety of poses and clothes, many of which are novel in the training data. For THuman2.0, CAPE, and RenderPeople, which have ground-truth 3D scans, we render 6 testing images from each scan at random viewpoints. Since SHHQ and VcgPeople just contain in-the-wild images, we only conduct qualitative evaluations on them." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10502" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 496, + 137 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 496, + 137 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 496, + 137 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 496, + 137 + ], + "type": "table", + "html": "
MethodsPublicationsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]ICCV'191.7601.9040.05000.24082.9672.7380.04490.24092.7812.8570.05900.2773
†PIFuHD [37]CVPR'203.0883.1130.08910.36634.7143.8230.05550.27963.3113.31180.08460.3541
PaMIR [54]TPAMI'221.0641.1850.04380.19271.7721.4040.03370.16761.5801.6590.04860.2088
ICON [48]CVPR'220.9470.9250.04220.17611.1331.0960.03110.14311.2651.2510.04310.1871
ECON [49]CVPR'230.9060.8450.03790.18910.9370.9210.03350.16441.2851.0790.04170.1973
VSOurs0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
", + "image_path": "43d23c8ff2959153c96ecec15f93e0266e7ab9cbb7fc821304ece82a9fee9e41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 182, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 182, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 182, + 202 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "Metrics. Chamfer distance " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "(\\varepsilon_{cd}" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": ", unit: cm), point-to-surface distance " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "(\\varepsilon_{p2s}" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": ", unit: cm), cosine distance " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "(\\varepsilon_{cos} \\in [0,1])" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": ", and L2 normal error " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "(\\varepsilon_{l2} \\in [0,1])" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " are employed as quantitative metrics. The first two metrics mainly assess large geometric errors, while the latter two tend to measure local detail differences between reconstructions and ground-truth models. Following ECON [49], when reporting cosine distance and L2 normal error, four normal maps are rendered respectively for the reconstructed and ground-truth models at the angles of " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\{0^o, 90^o, 180^o, 270^o\\}" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "Quantitative comparison. As illustrated in Table 1, VS demonstrates superior performance on all three benchmark datasets. Especially, our VS exceeds all other methods by a large margin on the two metrics that represent large geometric errors (i.e., " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon_{cd}" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon_{p2s}" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "). In terms of local surface details (i.e., " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon_{cos}" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\varepsilon_{L2}" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "), VS performs on par with ECON and much better than PIFu, PIFuHD, PaMIR, and ICON." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 176, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 176, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 176, + 431 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "Visual results. Fig. 4 shows the reconstructions by applying our VS on in-the-wild images. It can be seen that VS recovers high-fidelity clothed 3D humans even in challenging poses and complex loose clothes. Note that the poses, clothes, and accessories in these images are out of the distribution of the training data. These results reveal that VS not only has high reconstruction accuracy on challenging images but also has strong generalization capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 664 + ], + "type": "text", + "content": "Visual comparison. Given that ECON [49] is the current SOTA method for single-image clothed 3D human reconstruction, we first compare our VS with it solely. Fig. 3 shows the reconstructions by ECON and our VS. It can be seen that VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. Fig. 5 demonstrates visual comparisons with all five SOTA methods on in-the-wild images. ECON and VS demonstrate a substantial advantage over PIFu [36], PIFuHD [37], PaMIR [54], and ICON [48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "Perceptual study. We also conduct a perceptual study on the in-the-wild SHHQ and VcgPeople datasets. To this end, we designed a perceptual evaluation tool (see Fig. 11 in SupMat) where reconstructions produced by different meth" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 186, + 529, + 332 + ], + "blocks": [ + { + "bbox": [ + 324, + 186, + 529, + 332 + ], + "lines": [ + { + "bbox": [ + 324, + 186, + 529, + 332 + ], + "spans": [ + { + "bbox": [ + 324, + 186, + 529, + 332 + ], + "type": "image", + "image_path": "2081b4adff711e99ed34b9e08f7bc5813cea47e4c2d96e1afa27f088b25471fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 340, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 384 + ], + "type": "text", + "content": "Figure 3. Comparison of VS with ECON on in-the-wild images. VS recovers high-fidelity 3D humans while ECON struggles with missing parts, incomplete loose clothing, wrong thickness, and incorrect poses. More comparisons are shown in Video2 in SupMat." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 403, + 546, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 547 + ], + "type": "text", + "content": "ods are blindly placed on one page. We recruited 25 volunteers to participate in the perceptual study, and randomly sampled 50 images from each of the two datasets for each participant. Participants were requested to select the best reconstruction for each input image after rotating the reconstructions and carefully inspecting them from all viewpoints. The average percentages of preferences from these participants are illustrated in Table 2. The majority of participants perceived that our VS achieves the best reconstructions on both datasets. It's worth noting that ECON and VS exhibit greater superiorities over other methods on the more challenging VcgPeople dataset." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 320, + 555, + 532, + 586 + ], + "blocks": [ + { + "bbox": [ + 46, + 146, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 544, + 169 + ], + "type": "text", + "content": "Table 1. Quantitative comparison against other SOTAs on three benchmarks. Ground-truth SMPLs are used for methods that need SMPL(x) guidance. † Official model trained on Renderpeople." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 555, + 532, + 586 + ], + "lines": [ + { + "bbox": [ + 320, + 555, + 532, + 586 + ], + "spans": [ + { + "bbox": [ + 320, + 555, + 532, + 586 + ], + "type": "table", + "html": "
DatasetsPIFuPIFuHDPaMIRICONECONVS
SHHQ3.64.311.13.919.058.1
VcgPeople0.20.15.65.024.764.4
", + "image_path": "4dca442b15a1fa097347e88623de112915bf3353b106bd7be138b70212bb607a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 594, + 545, + 616 + ], + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 616 + ], + "type": "text", + "content": "Table 2. Average percentages (\\%) of participants' preferences. VS is significantly preferred over other methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of the two vertex shift states. To assess the effectiveness of the StretchVS and RefineVS modules in VS, we implemented two variants of VS, each exclusively incorporating either the StretchVS module or the RefineVS module. The quantitative results achieved by the complete" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10503" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 534, + 231 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 534, + 231 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 534, + 231 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 534, + 231 + ], + "type": "image", + "image_path": "ce4d7c47b41e61d6afc3a0589ff67f6503a04b0ad6749bb4ce69fbb28f3e33ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "type": "text", + "content": "Figure 4. Results produced by VS on in-the-wild images. Each result is shown from three views. VS recovers high-fidelity and artifact-less clothed 3D humans even in challenging poses (P) and loose clothes (C). More results are shown in Fig.12 and Video1 in SupMat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 271, + 494, + 322 + ], + "blocks": [ + { + "bbox": [ + 99, + 271, + 494, + 322 + ], + "lines": [ + { + "bbox": [ + 99, + 271, + 494, + 322 + ], + "spans": [ + { + "bbox": [ + 99, + 271, + 494, + 322 + ], + "type": "table", + "html": "
VersionsTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
VS0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
StretchVS0.9790.9150.06450.24690.9350.9070.04040.17431.2091.0400.06270.2443
RefineVS0.6690.5980.03910.16790.7210.7190.02710.12931.2630.9970.05290.2138
", + "image_path": "684c39efe8b82876456e5e195f9dc34fb1ea06acb37b5c9db2103a86a9f95927.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 331, + 410, + 342 + ], + "lines": [ + { + "bbox": [ + 182, + 331, + 410, + 342 + ], + "spans": [ + { + "bbox": [ + 182, + 331, + 410, + 342 + ], + "type": "text", + "content": "Table 3. Quantitative results achieved by three versions of VS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 363, + 288, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 506 + ], + "type": "text", + "content": "VS and the two variants are presented in Table 3. It is evident from these results that: (1) StretchVS exhibits a significant performance decrease on all three datasets, especially in terms of local details. (2) RefineVS performs consistently on THuman2.0 and CAPE without significant degradation, while it exhibits a noticeable decline on RenderPeople. This is because RenderPeople contains many out-of-distribution loose clothes, and it is difficult for the RefineVS module to deform an SMPL-X mesh to fit loose clothes without the assistance of the StretchVS module. The visual results shown in Fig. 6 are mutually corroborated with the quantitative results illustrated in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "text", + "content": "Robustness to SMPL-X estimates. Most body-aware methods are known to be sensitive to SMPL-(X) estimates, while SMPL-(X) estimation remains an unresolved issue. In this experiment, we use both accurate and inaccurate SMPL-X estimates to initiate VS. As shown in Fig. 7, VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates. Since StretchVS treats SMPL-X just as a \"seed\" and stretches it to align with visual cues, it has the \"error-correcting\" ability to shift vertices on inaccurate SMPL-X estimates to proper positions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 485, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 546, + 510 + ], + "type": "text", + "content": "Flexibility to feature encoders. To validate that VS is insensitive to IF-based feature encoders, we evaluated VS" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 72, + 523, + 521, + 690 + ], + "blocks": [ + { + "bbox": [ + 72, + 523, + 521, + 690 + ], + "lines": [ + { + "bbox": [ + 72, + 523, + 521, + 690 + ], + "spans": [ + { + "bbox": [ + 72, + 523, + 521, + 690 + ], + "type": "image", + "image_path": "2c5cbfa23ef8d740a8138bc74716a0eb529558f22197610441183f767423d4bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 698, + 529, + 711 + ], + "lines": [ + { + "bbox": [ + 63, + 698, + 529, + 711 + ], + "spans": [ + { + "bbox": [ + 63, + 698, + 529, + 711 + ], + "type": "text", + "content": "Figure 5. Visual comparison with other SOTA methods on in-the-wild images. More results are shown in Fig. " + }, + { + "bbox": [ + 63, + 698, + 529, + 711 + ], + "type": "inline_equation", + "content": "16\\sim 19" + }, + { + "bbox": [ + 63, + 698, + 529, + 711 + ], + "type": "text", + "content": " in SupMat." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10504" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 507, + 122 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 507, + 122 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 507, + 122 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 507, + 122 + ], + "type": "table", + "html": "
EncodersTHuman 2.0CAPERenderPeople
εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓εcd↓εp2s↓εcos↓εl2↓
PIFu [36]0.7660.6880.04110.17900.7350.7360.02720.12911.1040.8840.04650.1939
PaMIR [54]0.8380.7330.04020.17680.8610.7990.02660.13101.1820.9330.04280.1879
PIFuHD [37]0.6280.5550.03730.15550.6210.6150.02620.11380.9760.7880.04190.1618
", + "image_path": "8cb52039b278bd51f4671b3b97221d6ccb44d5dc69b30d79e009cc5e43ba79b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 130, + 424, + 140 + ], + "lines": [ + { + "bbox": [ + 168, + 130, + 424, + 140 + ], + "spans": [ + { + "bbox": [ + 168, + 130, + 424, + 140 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation of VS with different feature encoders." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 159, + 269, + 270 + ], + "blocks": [ + { + "bbox": [ + 67, + 159, + 269, + 270 + ], + "lines": [ + { + "bbox": [ + 67, + 159, + 269, + 270 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 269, + 270 + ], + "type": "image", + "image_path": "119374a63e0ddbe4089eb4681bf6b6511380d82c00ace9bc1ab260c9f0630a5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 278, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 288, + 322 + ], + "type": "text", + "content": "Figure 6. Results produced by VS variants w/wo StretchVS and RefineVS modules. StretchVS handles large deformations like loose clothing, while RefineVS recovers surface details. For humans who don't wear loose clothing, RefineVS can work alone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 333, + 268, + 499 + ], + "blocks": [ + { + "bbox": [ + 67, + 333, + 268, + 499 + ], + "lines": [ + { + "bbox": [ + 67, + 333, + 268, + 499 + ], + "spans": [ + { + "bbox": [ + 67, + 333, + 268, + 499 + ], + "type": "image", + "image_path": "f996322bece70b52b8cfd11cc5d8b73a324e67a14099762e01223caf22f562c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "type": "text", + "content": "Figure 7. Robustness to SMPL-X estimates. We use both accurate (odd rows) and inaccurate (even rows) SMPL-X estimates to initiate VS. VS can achieve high-quality reconstructions even with inaccurate SMPL-X estimates." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": "with pre-trained encoders from PIFu [36], PaMIR [54], and PIFuHD [37]. As shown in Table 4, the employment of different feature encoders does not result in significant discrepancies in the quantitative metrics achieved by VS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "type": "text", + "content": "We propose VS to reconstruct high-fidelity and artifact-less clothed 3D humans from single images. VS first stretches the estimated SMPL-X model into a coarse 3D human model using displacement fields inferred from visual cues, and then refines the coarse 3D human model into a de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 160, + 538, + 232 + ], + "blocks": [ + { + "bbox": [ + 314, + 160, + 538, + 232 + ], + "lines": [ + { + "bbox": [ + 314, + 160, + 538, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 160, + 538, + 232 + ], + "type": "image", + "image_path": "7862f9bd0f67eed98ff37b7663fd7b342715d5e9a5ad0f36874141a57bedb442.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 239, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 239, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 545, + 262 + ], + "type": "text", + "content": "Figure 8. Failure cases. VS may fail to handle extreme clothing, layer confusion, and large deformation in the Z-axis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 450 + ], + "type": "text", + "content": "tailed 3D human model via a GCN embedded with implicit-function-learned features. This two-stage deformation strategy reconciles the contradiction between large deformations for reconstructing loose clothing and delicate vertex shifts for recovering intricate and detailed surfaces. Extensive experiments on five datasets demonstrate that VS can reconstruct high-fidelity and artifact-less clothed 3D humans and achieves SOTA performance. VS confirms that deformation methods can reconstruct high-quality clothed 3D humans with complex poses and loose clothing, and even have advantages over IF-based methods in eliminating artifacts. We hope that VS can provide a new perspective for 3D human reconstruction and bring us one step closer to creating \"perfect\" human avatars for real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 451, + 546, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 631 + ], + "type": "text", + "content": "Limitations & Future work. Although VS achieves impressive results, it still has some limitations, as shown in Fig. 8. (A) Extreme clothing. VS may face challenges in deforming vertices on the legs of SMPL-X into extreme dresses. (B) Layer confusion. Currently, VS cannot distinguish the layers of the human body and clothing. (C) Wrong depth. VS fails in a couple of images exhibiting large deformation in the Z-axis (say, belts drifting towards the camera). In future work, we plan to separate a clothed human into the body and clothing parts and reconstruct them separately. As the clothing styles can be identified from the images [43] and the corresponding preset models can be provided [33], we hypothesize that using VS to deform preset clothing and body models separately can reconstruct extreme clothing and mitigate layer confusion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 640, + 410, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 410, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 410, + 654 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 545, + 708 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (62077026), and the Fundamental Research Funds for the Central Universities (CCNU22QN012)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10505" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1506-1515, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] The TensorFlow Authors. Dense image warp. www.tensorflow.org/addons/api_docs/python/tfa/image/dense_image_warp, 2023.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "type": "text", + "content": "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Eur. Conf. Comput. Vis. (ECCV), page 311-329, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 226, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 226, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 226, + 287, + 259 + ], + "type": "text", + "content": "[4] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Eur. Conf. Comput. Vis. (ECCV), pages 552 - 567, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "text", + "content": "[5] Yukang Cao, Kai Han, and Kwan-Yee K. Wong. Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4647-4657, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[6] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7291-7299, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "text", + "content": "[7] Lu Chen, Sida Peng, and Xiaowei Zhou. Towards efficient and photorealistic 3d human reconstruction: A brief survey. Visual Informatics, 5(4):11-19, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[8] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 18522-18531, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "text", + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5932-5941, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 508 + ], + "type": "text", + "content": "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Int. Conf. Mach. Learn. (ICML), pages 3789-3799, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[11] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12869-12879, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[12] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes-Or, Shachar Fleishman, and Daniel. MeshCNN: A network with an edge. ACM Trans. Graph. (TOG), 38(4):90:1–90:12, 2019. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "text", + "content": "[13] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph. (TOG), 39(4):126:1-126:12, 2020. 2, 3, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-PIFu: Geometry and pixel aligned implicit functions for single-view human reconstruction. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 2304–2314, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. " + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ARCH}++" + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": ": Animation-ready clothed human re" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "construction revisited. In Int. Conf. Comput. Vis. (ICCV), pages 11046-11056, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 545, + 140 + ], + "type": "text", + "content": "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. StereoPIFu: Depth aware clothed human digitization via stereo vision. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 119-128, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "text", + "content": "[17] Jingwei Huang, Hao Su, and Leonidas J. Guibas. Robust watertight manifold surface generation method for shapenet models. ArXiv, 2018. 3, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 545, + 221 + ], + "type": "text", + "content": "[18] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In Int. Conf. 3D Vis.(3DV), 2024. 2, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "type": "text", + "content": "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH:animatable reconstruction of clothed humans. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3093-3102, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "type": "text", + "content": "[20] Fu Jianglin, Li Shikai, Jiang Yuming, Lin Kwan-Yee, Qian Chen, Loy Chen Change, Wu Wayne, and Liu Ziwei. Stylegan-human: A data-centric odyssey of human generation. In Eur. Conf. Comput. Vis. (ECCV), pages 1-19, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 326, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 370 + ], + "type": "text", + "content": "[21] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4501-4510, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "text", + "content": "[22] Kun Li, Wen Hao, Feng Qiao, Yuxiang Zhang, Xiongzheng Li, Jing Huang, Cunkuan Yuan, Yu-Kun Lai, and Yebin Liu. Image-guided human reconstruction via multi-scale graph transformation networks. IEEE Trans. Image Process. (TIP), 30:5239-5251, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 429, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 462 + ], + "type": "text", + "content": "[23] Zhe Li, Tao Yu, Chuanyu Pan, Zerong Zheng, and Yebin Liu. Robust 3d self-portraits in seconds. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1344–1353, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "text", + "content": "[24] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8662–8672, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "text", + "content": "[25] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. In Eur. Conf. Comput. Vis. (ECCV), page 210-228, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "text", + "content": "[26] Leyuan Liu, Jianchi Sun, Yunqi Gao, and Jingying Chen. HEI-Human: A hybrid explicit and implicit method for single-view 3d clothed human reconstruction. In Chinese Conf. Pattern Recog. Comput. Vis. (PRCV), pages 251-262, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 667 + ], + "type": "text", + "content": "[27] Leyuan Liu, Yunqi Gao, Jianchi Sun, and Jingying Chen. Single-image clothed 3d human reconstruction guided by a well-aligned parametric body model. Multimedia Systems, page 1579-1592, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 34(6):1-16, 2015. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10506" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] William Lorensen and Harvey Cline. Marching Cubes: A high resolution 3d surface construction algorithm. ACM Trans. Graph. (TOG), 21(4):163-169, 1987. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[30] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3d people in generative clothing. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6469-6478, 2020. 5, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d reconstruction in function space. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4455-4465, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 273 + ], + "type": "text", + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 165-174, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 319 + ], + "type": "text", + "content": "[33] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7365-7375, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10975-10985, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 275, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 275, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 275, + 387 + ], + "type": "text", + "content": "[35] RenderPeople. www.renderpeople.com, 2018. 5, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "text", + "content": "[36] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis. (ICCV), pages 2304-2314, 2019. 2, 5, 6, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[37] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 84-93, 2020. 2, 4, 5, 6, 8, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[38] Siyuan Shan, Wen Yan, Xiaqing Guo, Eric I-Chao Chang, Yubo Fan, and Yan Xu. Unsupervised end-to-end learning for deformable medical image registrations. ArXiv, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[39] Qingyang Tan, Lin Gao, Yu-Kun Lai, Jie Yang, and Shihong Xia1. Mesh-based autoencoders for localized deformation component analysis. In AAAI, page 2452-2459, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[40] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3D human mesh from monocular images: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), pages 15406 - 15425, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[41] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Int. Conf. Comput. Vis. (ICCV), pages 10346-10356, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[42] Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: Volumetric inference of 3d human body shapes. In Eur. Conf. Comput. Vis. (ECCV), pages 20-36, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "text", + "content": "[43] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-YOLOv4: Scaling cross stage" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "partial network. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13029-13038, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 546, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 546, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 546, + 148 + ], + "type": "text", + "content": "[44] Junying Wang, Jae Shin Yoon, Tuanfeng Y. Wang, Krishna Kumar Singh, and Ulrich Neumann. Complete 3d human reconstruction from a single incomplete image. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 8748-8758, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3d mesh models from single rgb images. In Eur. Conf. Comput. Vis. (ECCV), pages 52-67, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 226 + ], + "type": "text", + "content": "[46] Eric W. Weissstein. Triangle point picking. mathworld. wolfram.com/TrianglePointPicking.html, 2023.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[47] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Int. Conf. 3D Vis.(3DV), pages 322-332, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 273, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 316 + ], + "type": "text", + "content": "[48] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit clothed humans obtained from normals. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 13296-13306, 2022. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 317, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 545, + 360 + ], + "type": "text", + "content": "[49] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit clothed humans optimized via normal integration. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 512-523, 2023. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 361, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 414 + ], + "type": "text", + "content": "[50] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5746-5756, 2021. 5, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "type": "text", + "content": "[51] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. PyMAF-X: towards well-aligned full-body model regression from monocular images. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 45(10):12287 - 12303, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "text", + "content": "[52] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. DeepMultiCap: Performance capture of multiple characters using sparse multiview cameras. In Int. Conf. Comput. Vis. (ICCV), pages 6219-6229, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "text", + "content": "[53] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3d human reconstruction from a single image. In Int. Conf. Comput. Vis. (ICCV), pages 7739-7749, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 624 + ], + "type": "text", + "content": "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 44(6):3170 - 3184, 2022. 2, 5, 6, 8, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[55] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4491-4500, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[56] Jun-yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis. (ICCV), pages 2223-2232, 2017. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10507" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_content_list.json b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1de700f19b393c280b41af63531c9a9c9a664676 --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_content_list.json @@ -0,0 +1,1946 @@ +[ + { + "type": "text", + "text": "VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning", + "text_level": 1, + "bbox": [ + 84, + 128, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kang Chen Xiangqian Wu", + "bbox": [ + 361, + 210, + 612, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Faculty of Computing, Harbin Institute of Technology (HIT)", + "bbox": [ + 245, + 229, + 723, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suzhou Research Institute of HIT", + "bbox": [ + 352, + 250, + 617, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenkangcs@stu.hit.edu.cn, xqwu@hit.edu.cn", + "bbox": [ + 295, + 268, + 671, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving the optimal form of Visual Question Answering mandates a profound grasp of understanding, grounding, and reasoning within the intersecting domains of vision and language. Traditional VQA benchmarks have predominantly focused on simplistic tasks such as counting, visual attributes, and object detection, which do not necessitate intricate cross-modal information understanding and inference. Motivated by the need for a more comprehensive evaluation, we introduce a novel dataset comprising 23,781 questions derived from 10,124 image-text pairs. Specifically, the task of this dataset requires the model to align multimedia representations of the same entity to implement multi-hop reasoning between image and text and finally use natural language to answer the question. Furthermore, we evaluate this VTQA dataset, comparing the performance of both state-of-the-art VQA models and our proposed baseline model, the Key Entity Cross-Media Reasoning Network (KECMRN). The VTQA task poses formidable challenges for traditional VQA models, underscoring its intrinsic complexity. Conversely, KECMRN exhibits a modest improvement, signifying its potential in multimedia entity alignment and multi-step reasoning. Our analysis underscores the diversity, difficulty, and scale of the VTQA task compared to previous multimodal QA datasets. In conclusion, we anticipate that this dataset will serve as a pivotal resource for advancing and evaluating models proficient in multimedia entity alignment, multi-step reasoning, and open-ended answer generation. Our dataset and code is available at https://visual-text-qa.github.io/.", + "bbox": [ + 75, + 349, + 473, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 818, + 209, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A paramount objective in AI research is to endow systems with the capacity to comprehend the intricacies of the real", + "bbox": [ + 76, + 844, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "world, akin to human understanding. Question Answering (QA) stands out as an effective task for evaluating the cognitive capabilities of AI systems. To answer questions, people need to extract information from multiple modalities, such as text, images and structured data like knowledge bases, graphs and tables. And furthermore, people need to align the information and do multi-step reasoning between different modalities.", + "bbox": [ + 500, + 319, + 892, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although Visual Question Answering (VQA) [5] has been widely researched as a multimedia QA task, VQA models only extract information from image when answering questions and focus mainly on scene recognition, counting, color and other visual detection tasks, which do not require much logical reasoning or assignment between different modalities. Only recently, there are some attempts to introduce more modal information into VQA tasks. For example: (1) FVQA [24] and KBVQA [23] combine knowledge base (KB) with VQA task, which requires the ability of knowledge understanding and multi-step reasoning. But it is difficult to construct a comprehensive KB in real world, thus limiting its ability to address open-ended questions; (2) TextbookQA [10] and ScienceQA [14] use textbook as data sources, involving texts, images, tables and other multimodal information. But most images in these datasets are manually drawn schematic diagrams, and the questions are all in a multiple-choice setting, which are far from the real world.", + "bbox": [ + 496, + 443, + 892, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, more recently, there appear some new datasets such as MultiModalQA [21] and MuMuQA [18], which involve reasoning across texts, images and tables. However, it's noteworthy that all these datasets are conducted with extracted QA. And in the MultiModalQA dataset, each image corresponds to a Wikipedia entity, simplifying image reasoning to ranking images based on these entities, which diminishes the necessity for crossmedia reasoning abilities. As for MuMuQA, although it requires grounding between image and text along with multi-hop reasoning, there are still some problems: (1) the data", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author", + "bbox": [ + 94, + 887, + 217, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27218", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/818423c22feecce763a7f581f2384e4cf7654af630e049baf37a7ae7d29ee0d3.jpg", + "image_caption": [ + "Image:", + "Figure 1. Example in our dataset with the question-answer pairs and their corresponding image and text. Different representations of the same object in text and image are identified with the same color. For example, 'Elena' in the text and the object bounding box corresponding to 'Elena' in the image are marked red." + ], + "image_footnote": [], + "bbox": [ + 106, + 104, + 392, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Chinese Annotation:", + "text_level": 1, + "bbox": [ + 408, + 93, + 509, + 102 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "公司举行了一场饭桌会议，顾名思义就是一边吃饭一边商讨事宜。爱莲娜的金色头发果然是人群中最显眼的一个，乌尔蒂盖拉穿着橘黄色开衫，正在埋头切一块牛排；赛特温穿着深蓝色衬衫，胸前挂着会议牌，他的左手正抬起来准备去拿酒杯。寒特温的手左边，卡维利耶和阿吉姆正严肃地讨论下个月的预算案。卡维利耶一头褐色齐肩短发，干净利落。", + "bbox": [ + 408, + 109, + 578, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q1：爱莲娜穿着什么类型的上衣？", + "bbox": [ + 408, + 176, + 532, + 185 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A1：西装外套", + "bbox": [ + 408, + 189, + 460, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q2：乌尔蒂盖拉的右手边是谁？", + "bbox": [ + 408, + 203, + 519, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A2：卡维利耶", + "bbox": [ + 408, + 215, + 455, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "English Translation:", + "text_level": 1, + "bbox": [ + 607, + 93, + 705, + 103 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The company held a table meeting, which, as the name implies, is to discuss matters while eating. Elena's blond hair is indeed the most conspicuous one in the crowd, and Urigera, wearing an orange cardigan, is burying her head in cutting a steak; Setwin was wearing a dark blue shirt, with a meeting card hanging on his chest, and his left hand was lifting up to get a glass of wine. On the left of Setwin, Kavelier and Agum are seriously discussing the budget for next month. Kavelier has short brown shoulder-length hair, clean and clean.", + "bbox": [ + 604, + 109, + 864, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q1: What type of blouse does Elena wear?", + "bbox": [ + 606, + 176, + 759, + 185 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A1: Suit", + "bbox": [ + 607, + 189, + 635, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q2: Who is on the right-hand side of Urtigera?", + "bbox": [ + 607, + 203, + 776, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A2: Kavelier", + "bbox": [ + 607, + 215, + 650, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is all from news, resulting in most questions related to human beings; (2) questions all follow a specific pattern: first perform image entity grounding and then find the answer in the news body text; (3) there is only 1384 human-curated examples and the training data is automatically generated, which leads to poor quality and difficult to train with. Current multimedia benchmarks are still far from the real-world QA scene and cannot measure the multimedia understanding ability of AI system well.", + "bbox": [ + 75, + 306, + 467, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to address these gaps, we propose a novel task named 'Visual Text Question Answering' (VTQA), accompanied by a dedicated dataset, which includes only questions that require multi-hop reasoning through both images and text. All the annotations in VTQA dataset are first marked in Chinese, and then translated into English.", + "bbox": [ + 75, + 445, + 467, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer VTQA questions, the proposed model needs to: (1) learn to identify entities in image and text referred to the question, (2) align multimedia representations of the same entity, and (3) conduct multi-step reasoning between text and image and output open-ended answer. The VTQA dataset consists of 10,124 image-text pairs and 23,781 questions. The images are real images from MSCOCO dataset [12], containing a variety of entities. The annotators are required to first annotate relevant text according to the image, and then ask questions based on the image-text pair, and finally answer the question open-ended. To ensure textual richness, we enforce a minimum text length requirement of over 100 Chinese words. Additionally, we systematically exclude questions that can be answered solely based on either the image or the text, thereby guaranteeing the complexity of the questions.", + "bbox": [ + 75, + 537, + 467, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we conduct an evaluation on the VTQA dataset, employing both state-of-the-art VQA models and our newly proposed baseline model. The baseline model we propose is called Key Entity Cross-Media Reasoning Network (KECMRN), which follows the general paradigm for answering questions on this dataset by iteratively performing key entity extraction and alignment, and cross-modal multi-step reasoning to answer questions. Evaluation re", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sults on the VTQA dataset show that existing state-of-the-art VQA models struggle to achieve satisfactory performance on this dataset, which illustrates the challenges and potential of this dataset for cross-modal question answering tasks.", + "bbox": [ + 496, + 306, + 890, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work can be summarized as follows: (1) we propose a new cross-modal QA dataset named VTQA. Information diversity, multimedia multi-step reasoning and open-ended answer make our dataset more challenging than the existing datasets; (2) we benchmark the state-of-the-art VQA model on our new dataset and show the performance of these models degrades drastically; (3) we propose a baseline that is capable of multimedia entity alignment and multi-step reasoning.", + "bbox": [ + 496, + 383, + 890, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 539, + 640, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) aims to answer a natural language question based on an image, which requires model to understand and reason in the vision-language joint space. Several datasets have been proposed in the past few years, such as DAQUAR [15], FM-IQA [4], VQA [2, 5], COCOQA [19], Visual7W [28], Visual Genome [11], GQA [8], OKVQA [16], A-OKVQA [20], VizWizQA [6] and so on. However, the natural language questions in these datasets can be regarded as instructions to guide the model to complete visual tasks such as object detection, scene recognition, counting, etc. And these questions often only involve the surface information of the image, without considering the complexities between different modalities. Understanding of the reasoning process. In order to improve the multimodal perception ability of the model, we need to design more challenging problems, requiring the model to combine the features of images and texts for cross-modal knowledge fusion and logical reasoning.", + "bbox": [ + 496, + 566, + 890, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some datasets attempt to introduce more modal information into VQA tasks, such as FVQA [24], KB-VQA [23], TextbookQA [10], and ScienceQA [14]. But the information used in these datasets includes manually col", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27219", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lected databases or manually drawn schematic diagrams in textbooks, which are far from real-world QA scenarios. The VTQA we propose is different from these, using the semantically rich real images of the COCO dataset and human-annotated text passages, which can be better generalized to real-world QA scenarios.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most similar to our dataset is MuMuQA [18]. The MuMuQA dataset is a news-based QA dataset whose images and questions mainly involve people and events. This dataset has only a few human-labeled data, and most of the data is generated by automated methods, so there may be a lot of noise and errors. In contrast, our VTQA dataset covers a variety of genres and topics, and all texts, questions, and answers are annotated by humans, ensuring the high quality and accuracy of the data. Our VTQA dataset has 23781 items, which is richer and more diverse than MuMuQA dataset.", + "bbox": [ + 75, + 181, + 467, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VTQA Dataset", + "text_level": 1, + "bbox": [ + 76, + 364, + 225, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the details of our Visual Text Question Answering (VTQA) dataset. We first introduce the VTQA task of our dataset (Section 3.1) and then detail how the dataset was constructed (Section 3.2), along with the statistical properties of our dataset.", + "bbox": [ + 75, + 390, + 467, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. VTQA Task", + "text_level": 1, + "bbox": [ + 76, + 477, + 205, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, given an image-text pair and a question, a system is required to answer the question by natural language. Importantly, the system needs to: (1) analyze the question and find out the key entities, (2) align the key entities between image and text, and (3) generate the answer according to the question and aligned entities. For example, in Fig. 1, the key entity of Q1 is \"Elena\". By referencing the descriptor \"gold hair\" in the text, it is discerned that the second person from the right in the image is \"Elena\". Subsequently, the answer \"suit\" is generated based on the visual information. As for Q2, which is a more complex question, the previous steps need to be repeated several times to answer it.", + "bbox": [ + 75, + 501, + 467, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dataset Construction", + "text_level": 1, + "bbox": [ + 76, + 709, + 274, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To promote the progress of this open-ended multimedia multi-hop VTQA task, we collect a new dataset. Our dataset consists of 10124 image-text pairs and 23,781 questions. We collect data through newly developed annotation interface.", + "bbox": [ + 75, + 733, + 467, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first round of labeling, we present images from COCO and corresponding image descriptions, as well as the object detection labels, to the annotators. The annotators are required to generate a text that exceeds 100 words, which should involve the object in the image and contain information that is not included in the image description.", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure compliance with annotation requirements, annotators can choose to skip some images (12198 images were ultimately skipped, 9830 images were selected, and some of them were selected more than once). The annotators then come up with questions based on the image-text pair and the annotation process requires that the questions cannot be answered only by image or text. Each image-text pair is labeled with 1-4 questions.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a second round of labeling, different annotators were asked to determine whether the question could be answered solely by relying on either images or text within the corresponding text-question pair or image-question pair. Through this step, we filtered down to 23781 questions from a pool of 28919 questions. The questions that pass this step will be labeled with answers and answer categories. We set three categories for answers: (1) $YN$ means yes-or-no answer, (2) $E$ means that the answer is extracted from the text, and (3) $G$ means that the answer is generated from the text-image pair. And the annotators need to label 'yes' or 'no' in English for the yes-or-no answer, since there are too many words to express 'yes' or 'no' in Chinese, for example, '可以' and '是的' both mean 'yes'.", + "bbox": [ + 496, + 218, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We randomly split the dataset into training, validation, test-dev and test splits and each image will only appear in one split. Each split has 11312, 1245, 2189, 9035 samples, respectively.", + "bbox": [ + 496, + 436, + 890, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Dataset Analysis", + "text_level": 1, + "bbox": [ + 500, + 525, + 663, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 shows a comparison of VTQA and other VQA datasets. As shown in the table, VTQA is much larger than most other datasets, especially with a much larger number of multimodal questions than existing multimodal QA datasets. The text length of the VTQA dataset is shorter than that of MuMuQA, but still much longer than the length of the questions in all datasets, and the texts in VTQA are manually generated, containing more types of entities than existing texts extracted from textbooks, wikis, and news.", + "bbox": [ + 496, + 555, + 890, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Images in VTQA dataset are from MSCOCO dataset, which contains multiple objects and rich contextual information. Fig. 2 presents the top-10 categories distribution of the images used in this dataset. Unlike MuMuQA, in which most images are related to people, images in our dataset contain more kinds of objects.", + "bbox": [ + 496, + 696, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The texts and questions length statistics are shown in Fig. 3. It is evident that most texts range from 300 to 600 and most questions range from 9 to 21. Obviously, compared with the previous VQA datasets (usually less than 20 words), our dataset also puts forward higher requirements for text understanding to extract information from the long text.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27220", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d4878855d4c17c1bfbb74d012ec6ffb840f44537284815f78bfe1c67882e0ed0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#I#Q#MMQContextsAvgQAvgTAnswer TypeImage SourceText Source
VQA200K1.1M-I6.1-Open/MCCOCO-
FVQA58265826-I+KB9.5-OpenCOCO-
KBVQA24022402-I+KB6.8-OpenCOCO-
TextbookQA34552626012567I+T9.2-MCTextbookTextbook
ScienceQA10332212086532I+T12.1-MCTextbookTextbook
MultimodalQA57058299188240I+T18.266.2OpenWikipediaWikipedia
MuMuQA*138413841384I+T11.8633.5OpenNewsNews
VTQA98302378123781I+T10238.4OpenCOCOannotated
", + "bbox": [ + 78, + 87, + 895, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Statistics for VTQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, #MMQ: number of multimodal questions, AvgQ: average question length, AvgT: average text length. *We only count the manually annotated parts of the MuMuQA dataset.", + "bbox": [ + 75, + 238, + 893, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77fda0048ce186dc869078aa41d20b3c5c3614e553d8fad2fa4d7170ea59ad38.jpg", + "image_caption": [ + "Figure 2. Top-10 categories distribution of the images used in our dataset." + ], + "image_footnote": [], + "bbox": [ + 89, + 327, + 434, + 526 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bbfce12d0bc15be359464fba6b5bddd858aeeca43429ac99273c77d99cb48c2.jpg", + "image_caption": [ + "Figure 3. Percentage of questions and texts with different Chinese word lengths." + ], + "image_footnote": [], + "bbox": [ + 89, + 625, + 267, + 806 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f405c200360739f022bd0be75bf43be3eea20ddc55f1782d4c8fcdc2bc17a1f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 626, + 454, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. VTQA Model", + "text_level": 1, + "bbox": [ + 500, + 306, + 638, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe a competitive baseline method for evaluation on our benchmark, which is called Key Entity Cross-Media Reasoning Network (KECMRN). Before presenting the KECMRN, we first introduce its basic component, the KECMR layer. The KECMR layer is a modular composition which consists of one Key Entity Extract (KEE) layer and multiple Cross-Media Reason (CMR) layers. The KEE layer and CMR layer are composed of attention unit and feed-forward unit from [22]. Then we use the KECMR module with other layers to combine our KECMRN. An overview flowchart of KECMRN is shown in Fig. 4.", + "bbox": [ + 496, + 330, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Attention and Feed-Forward Units", + "text_level": 1, + "bbox": [ + 500, + 521, + 803, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in [22], the combination of attention unit and feed-forward unit has strong representational and learning ability. We use the same settings to construct our units.", + "bbox": [ + 496, + 544, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Head Scaled Dot-Product Attention Unit. Given a query $q \\in R^{1 \\times d}$ , n key-value pairs (packed into a key matrix $K \\in R^{n \\times d}$ and a value matrix $V \\in R^{n \\times d}$ ), the attended feature is obtained as:", + "bbox": [ + 496, + 590, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 672, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-head attention divides the input into $h$ parts and makes single attention on each part. The attended feature is given by:", + "bbox": [ + 496, + 715, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM H (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 771, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh e a d _ {i} = \\text {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 797, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{i}^{Q}, W_{i}^{Q}, W_{i}^{Q} \\in R^{d \\times d_{h}}$ and $W^{O} \\in R^{(h \\times d_{h}) \\times d}$ are the projection matrices. $d_{h}$ is the dimension of the output features from each head.", + "bbox": [ + 496, + 821, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feed-Forward Unit. The feedforward unit takes the output features of the multi-head attention layer, and further", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27221", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9bc01fc458078bcbca0521cf05a108a8e1f7658670a217d6675859d8d91f088f.jpg", + "image_caption": [ + "Figure 4. Overall flowchart of the Key Entity Cross-Media Reasoning Network." + ], + "image_footnote": [], + "bbox": [ + 161, + 90, + 812, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b8347d993c440a8dbf60409e029071f10310c96d670bae4b9e55e1595026f52e.jpg", + "image_caption": [ + "Figure 5. Key Entity Extract Layer." + ], + "image_footnote": [], + "bbox": [ + 96, + 325, + 450, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "transforms them through two fully-connected layers with ReLU activation in between.", + "bbox": [ + 76, + 568, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF F N (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 614, + 468, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the linear transformations are the same across different positions, they use different parameters from layer to layer.", + "bbox": [ + 76, + 638, + 468, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Key Entity Extract Layer", + "text_level": 1, + "bbox": [ + 76, + 695, + 308, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As stated in Section 3.1, the first step to answer VTQA questions is to find out the key entity according to the questions. We compose the attention unit and feed-forward unit to integrate question information into the text and the image respectively. Then we apply a fully-connected layer to the question-aware text/image features to get the score for each feature. Finally, we extract the top-k features as key entities.", + "bbox": [ + 75, + 719, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete KKE layer is shown in the Fig. 5. Mathematical, given input text features $T$ , image features $I$ and question features $Q$ , the KEE layer can be formulated by (the processing of feature $I$ is aligned with that of feature", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2cede4c626558eb454f68ee3c4f238e30d538d84b175a51073cc46532c4bdd75.jpg", + "image_caption": [ + "Figure 6. Cross-Media Reason Layer. 'G' means gather and 'S' means scatter." + ], + "image_footnote": [], + "bbox": [ + 506, + 327, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "T):", + "bbox": [ + 500, + 500, + 524, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = F F N (M H (Q, Q, Q)) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 527, + 890, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT = F F N (M H (M H (T, T, T), Q, Q)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 550, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s c o r e} _ {T} = W _ {T} T + b _ {T} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 574, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final score will be used to sort the image/text features and the top-k features are considered as image/text key entities. To unify the expression, we treat all the question features as key entities.", + "bbox": [ + 496, + 595, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Cross-Media Reason Layer", + "text_level": 1, + "bbox": [ + 498, + 666, + 746, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The CMR layer is designed for multi-step reasoning across medias. As show in Fig. 6, we first gather the key entities as $S \\in R^{l_k \\times d}$ from input features, where $l_k$ is the total number of all key entities. Then the key entities pass through self-attention, global cross-attention based on original features and feed-forward unit. Finally, the key entities are scattered to the input features. As the requirement for multi-step reasoning, we inserted multiple CMR layers in a KECMR module. The CMR layer can be formulated by:", + "bbox": [ + 496, + 689, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = \\operatorname {g a t h e r} ([ T, I, Q ]) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 839, + 890, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = F F N (M H (M H (S, S, S), [ T, I, Q ], [ T, I, Q ])) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 862, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ T, I, Q ] = \\operatorname {s c a t t e r} (S, [ T, I, Q ]) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 883, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27222", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Key Entity Cross-Media Reason Network", + "text_level": 1, + "bbox": [ + 76, + 90, + 431, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the same network framework as [26] but replace MCA layer with our KECMR module and add an extra text stream. As show in Figure 4, The input image is represented as a set of regional visual features in a bottom-up manner [1]. The input question and text are transformed to features by passing through a one-layer LSTM network [7]. Then we use our KECMR module several times to extract key entities and conduct multi-steps cross-media reasoning. Finally, we use the attention reduce layer to fuse the multimedia features and project the fused feature into the answer probability distribution.", + "bbox": [ + 75, + 112, + 468, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The attention reduce layer can be formulated by:", + "bbox": [ + 96, + 280, + 418, + 295 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM L P (x) = W _ {2} \\left(\\max \\left(0, x W _ {1} + b _ {1}\\right)\\right) + b _ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 320, + 468, + 338 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t R e} (X) = \\operatorname {s u m} (\\operatorname {s o f t m a x} (M L P (X)) X) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 347, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can obtain the attended image feature $I_{a}$ by passing image feature I through attention reduce layer, and the same applies to text feature T and question feature Q. Then we fuse the multimedia features by:", + "bbox": [ + 76, + 369, + 468, + 431 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz = W _ {i} I _ {a} + W _ {t} T _ {a} + W _ {q} Q _ {a} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 445, + 468, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The fused feature $\\mathbf{z}$ is projected into a vector $s \\in R^{N}$ followed by a sigmoid function, where $\\mathbf{N}$ is the number of the answers in the training set. We use binary cross-entropy (BCE) as the loss function to train an N-way classifier on top of the fused feature $\\mathbf{z}$ .", + "bbox": [ + 76, + 467, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 556, + 209, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the current state-of-the-art VQA model and our KECMRN model, and provide results of them in VTQA dataset.", + "bbox": [ + 75, + 580, + 468, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 635, + 261, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the answers divided into three types, we use different metrics for distinct types of answers.", + "bbox": [ + 75, + 659, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Exact match (EM). This metric measures the percentage of predictions that match the ground truth answer exactly and will be used in all types of answers.", + "bbox": [ + 75, + 689, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Macro-averaged) F1 score (F1). This metric measures the average overlap between the prediction and ground truth answer. We treat the prediction and ground truth as bags of tokens and compute their F1. This metric will be used for the $E$ and $G$ types of answers.", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YN accuracy (YNAcc). This metric is only used for the YN type of answers. The answer will be transformed into 'yes' or 'no' by a pre-defined yes-or-no dictionary, which is manually collected based on the answers in our dataset. Then we calculate the accuracy just based on the yes-or-no answer.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 90, + 718, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KECMRN: the hyper-parameters of our model used in the experiments are as follows. The dimensionality of input image features, input question features, input text features and fused multimodal features are 2,048, 1024, 1024, and 2,048, respectively. The latent dimensionality $d$ in the multi-head attention is 1024, the number of heads $h$ is set to 8. The number of CMR layer in each KECMR module and the number of KECMR module are 2 and 2. The number of key entities $k$ is set to 16. We train our model on the train set for 13 epochs and evaluate our model on test set. The Faster-RCNN used for image feature extraction is pretrained and frozen, while all other modules are trained from scratch.", + "bbox": [ + 496, + 113, + 890, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPT4V [17]/LLaVA [13]/MMICL [27]: popular multimodal large language models (MLLM). These models demonstrated impressive multimodel chat abilities and exceptional performance across numerous intricate visual reasoning datasets, particularly excelling in zero-shot evaluations. Notably, LLaVA achieved an impressive $80.0\\%$ accuracy, while GPT4V attained $77.2\\%$ and MMICL attained $70.3\\%$ on the VQAv2 dataset, underscoring their remarkable capabilities in the realm of multimodal understanding and visual question answering. We follow the same evaluation strategy as on the VQAv2 dataset, with the sole modification of adapting the prompt to 'Search for answers in the context above, do not use additional knowledge beyond the text, and answer the question using a single word or phrase'. This adjustment facilitated enhanced performance on VTQA datasets.", + "bbox": [ + 496, + 295, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BEiT3 [25]/VLMo [3]: visual language models directly pre-trained on image-text pairs, offering versatility in addressing a spectrum of visual-language tasks through finetuning on downstream tasks. Notably, on the VQAv2 dataset, BEiT3-base achieved an accuracy of $77.65\\%$ , while VLMo-base demonstrated a commendable accuracy of $76.6\\%$ . To evaluate the transferability of BEiT3 and VLMo on our VTQA task, we follow the same fine-tuning strategy as on the VQAv2 dataset, but we concatenate the text and questions into a single paragraph of text and increase the maximum number of training epochs from 10 to 50 to obtain better results, since our dataset is relatively small compared to VQAv2.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MCAN-region [26]/MCAN-grid [9]: models based on MCAN [26] but using different image features. They achieve SOTA results on VQAv2 dataset (70.93% for MCAN-region and 72.71% for MCAN-grid) among non-pretrained visual-language models(non-PVLM), which indicates that these models were not pretrained on visual-language data, although individual pretraining on either visual or text data is allowed. We concatenate the text and questions and increase the maximum number of input words to train MCAN in VTQA task.", + "bbox": [ + 496, + 734, + 890, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trans-CA/Trans-SA: models based on Trans", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d9758a3b2fcd16879464829e0748b55fe50360977a2edee7b9aee26b03410586.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsEMYN-ACCE-F1G-F1
English Version
LLaVA39.9065.0560.0626.20
GPT4V48.9375.5970.2448.37
MMICL38.7662.9062.3245.96
BEiT3-base41.0256.5150.1828.14
VLMO-base51.8565.4664.5942.74
MCAN-region56.0972.0667.0340.52
MCAN-grid55.8870.7365.3939.19
Trans-CA57.3873.5769.0244.14
Trans-SA54.3967.8965.9541.11
KECMRN(ours)57.9574.9968.9144.25
Chinese Version
MCAN-region49.7871.2659.9445.74
MCAN-grid50.2274.1759.0245.59
Trans-CA49.8774.2759.6347.45
Trans-SA48.9973.2958.2646.31
KECMRN(ours)51.3277.5960.5251.11
", + "bbox": [ + 78, + 87, + 468, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "former [22]. Trans-CA treat image, text and question features as independent input and perform Self Attention and Question Aware Cross Attention in each layer. The layers in Trans-CA resemble those of the KEE module, excluding the \"score&sort\" step. Conversely, Trans-SA adopts a different approach by concatenating all image, text, and question features and only applying Self Attention in each layer. The layers in Trans-SA resemble those of the CMR module, excluding the \"Global Cross-Attention\" step.", + "bbox": [ + 75, + 430, + 468, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results and Analysis", + "text_level": 1, + "bbox": [ + 76, + 590, + 272, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 provides the results of various baselines on the test set of our VTQA benchmark. VLMs, including MMICL, LLaVA, BEiT3, and VLMo, primarily pretrained on English datasets, are exclusively evaluated in the English version.", + "bbox": [ + 75, + 613, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While models such as GPT4V, MMICL, and LLaVA demonstrate superior performance across a variety of visual-language datasets in zero-shot evaluation, they encounter a significant drop in performance when evaluated on the VTQA dataset. This is partly because the model lacks exposure to our training data during training, unlike the VQA dataset included in the model's VL fine-tuning data. The decline also results from the heightened complexity of our VTQA task compared to VQA, posing new challenges to the model's image-text alignment and inference capabilities.", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BEiT3 and VLMo, pretrained directly on image-text pairs, achieve impressive results on various multimodal datasets through fine-tuning on downstream tasks. Despite", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17d16f672adcd92bed16ae679e2828a351a4743954b41d26ef66d48ffc3d6288.jpg", + "table_caption": [ + "Table 2. Results of our KECMRN on test set compared with other multimodal models." + ], + "table_footnote": [], + "table_body": "
MethodsEMYN-ACCE-F1G-F1
Trans(1)55.4572.4365.3542.27
Trans(3)55.0770.9865.743.58
Trans(6)54.0769.4364.8343.91
LSTM57.9574.9968.9144.25
", + "bbox": [ + 531, + 87, + 862, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of different text encoders based on KECMRN.", + "bbox": [ + 506, + 178, + 883, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine-tuning on VTQA training data, BEiT3 and VLMo still exhibit a substantial performance decrease. This is partly attributed to the limitation that the pretraining data for BEiT3 and VLMo only includes image-text pairs with relatively short texts, while the texts in the VTQA dataset are longer, averaging 284 words in English version. This, once again, underscores the challenges inherent in the VTQA task.", + "bbox": [ + 496, + 218, + 890, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While MCAN models outperform the pre-trained models, a notable decrease is observed. In addition, Trans-CA exceeded the results of all VQA models, while the results of Trans-SA are relatively poor. This highlights the necessity for a more intricate independent processing of text, image, and question features in our VTQA dataset. Our KECMRN achieves the best result with $57.95\\%$ on the English version and $51.32\\%$ in Chinese version.", + "bbox": [ + 496, + 324, + 890, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 illustrates the outcomes of attention visualization across different models. The attention weights from the 'Attention Reduce' module were employed to colorize image regions and organize the Top Words. Notably, our KECMRN model exhibits precise identification of Zhang Yanqi on the left side, engaged in ball play. This accurate recognition is based on key phrases such as sports balls and dribbling skills, leading to a correct inference for the associated question ('Yes'). In contrast, although MCAN also directs attention to the person on the left, the absence of Key-Entity Extract (KEE) causes attention to be excessively dispersed. Consequently, there is an inaccurate localization of the person's upper garment, resulting in an incorrect answer. Furthermore, Trans-SA mistakenly focuses on the person on the right, while Trans-CA struggles to discern whether Zhang Yanqi is positioned on the left or right, both yielding incorrect answers. The attention visualization results underscore the superior precision of our KECMRN model in focusing on the correct visual regions and words.", + "bbox": [ + 496, + 445, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 741, + 653, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a number of ablation experiments to investigate the reasons for the effectiveness of KECMR. All the reported results are evaluated on English version.", + "bbox": [ + 496, + 763, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text Encoders: Table 3 provides an overview of the results obtained using different text encoders in conjunction with KECMRN. In this table, Trans(k) denotes a k-layer transformer encoder model. Contrary to expectations, the performance of deeper transformer encoder models tends to be inferior, with LSTM outperforming all transformer en", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27224", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39c36cb5a72c3e3bf4d9007f568fc99c6fc196eb84974c4e47a160a3afb6edc5.jpg", + "image_caption": [ + "Text:" + ], + "image_footnote": [], + "bbox": [ + 120, + 104, + 352, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Song Shuai and Zhang Jiaqi are two local sports ball enthusiasts whose video about sports ball on the short video platform caught the attention of a club in a nearby city. Today they were invited to the city where the club is located for a technical exchange, and now Song Shuai is talking to the head of the club, and Zhang Jiaqi is too obsessed with sports balls on the road and still does not forget to practice dribbling skills.", + "bbox": [ + 124, + 236, + 346, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question:", + "bbox": [ + 125, + 377, + 176, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is Zhang Jiaqi wearing a blue top?", + "bbox": [ + 125, + 388, + 302, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Answer:", + "bbox": [ + 125, + 412, + 171, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yes", + "bbox": [ + 125, + 424, + 148, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c64ae40c2457a6d4a0a8113215c982f8620aebae8bf75f42fa9d8bc9bca25502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 89, + 594, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Top Words:", + "bbox": [ + 367, + 212, + 431, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "balls, Jiaqi, enthusiasts, Zhang, road, skills, obsessed, dribbling, sports, forget Predict Answer: Yes", + "bbox": [ + 367, + 224, + 573, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0074f83903e06aef247d91cb0e37d8d18c8f0d1f3fb57dbf1b4f28bda46b286.jpg", + "image_caption": [ + "KECMRN" + ], + "image_footnote": [], + "bbox": [ + 362, + 277, + 591, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Top Words:", + "bbox": [ + 367, + 400, + 431, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sports, dribbling, Jiaqi, club, Shuai, Zhang, video, caught, platform, talking Predict Answer: No", + "bbox": [ + 367, + 412, + 568, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trans-SA", + "bbox": [ + 449, + 450, + 504, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea7925dc6cd291f2c8dc6a3a4a014aaac157aae22f482d51db66db1ea0bd8c6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 848, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Top Words:", + "bbox": [ + 624, + 212, + 687, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "skills, dribbling, enthusiasts, nearby, road, balls, Zhang, club, still, sports Predict Answer: No", + "bbox": [ + 624, + 224, + 839, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13a5d0a27bd4e1632a8daf46b6ff066113c94c816376ea668c4eaab3aa39dd1e.jpg", + "image_caption": [ + "MCAN", + "Figure 7. An example of attention visualization for different models." + ], + "image_footnote": [], + "bbox": [ + 617, + 277, + 848, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Top Words:", + "bbox": [ + 624, + 400, + 687, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Jiaqi, Song, exchange, dribbling, Zhang, obsessed, head, forget, short, Shuai Predict Answer: No", + "bbox": [ + 624, + 412, + 831, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trans-CA", + "bbox": [ + 705, + 450, + 761, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/895b81b74cebcaf0f3008f36c3489f093e55c4800e7d322157ed503a3491f3f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsEMYN-ACCE-F1G-F1
Trans-CA57.3873.5769.0244.14
CA+SA57.1773.5768.9642.88
KEE+SA57.6374.4269.143.06
KEE+CMR57.9574.9968.9144.25
", + "bbox": [ + 99, + 515, + 450, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of the KECMRN model with different module combinations.", + "bbox": [ + 76, + 606, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coder models. This phenomenon is attributed to both insufficient training data and the intricate structure of transformer models.", + "bbox": [ + 76, + 679, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "KEE/CA and CMR/SA: Table 4 displays the results of the KECMRN model with different module combinations. In this table, CA+SA denotes the alternate use of CA and SA layers, similar to our KECMRN design. KEE+SA follows a similar pattern. Compared to Trans-CA, replacing some CA layers with SA layers results in a slight performance decrease (↓0.21). Conversely, the inclusion of the KEE layer improves performance (↑0.46) by concentrating information on key entities. Additionally, the CMR layer further optimizes features related to key entities, leading to enhanced results (↑0.32).", + "bbox": [ + 76, + 734, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 518, + 617, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel cross-modal question answering dataset, VTQA, which necessitates models to acquire pertinent information from both text and image sources and perform complex cross-modal reasoning to answer questions. Our experimental findings reveal that even state-of-the-art pretrained models fall short of achieving satisfactory performance on VTQA, underscoring the challenges and potential inherent in the dataset for cross-modal question-answering tasks. Looking ahead, we aim to augment the dataset's size and coverage and provide additional reasoning annotations in future work to more effectively assess and catalyze the development of cross-modal reasoning models. Simultaneously, we aspire to explore more fitting evaluation metrics to surmount the limitations of EM metrics and more equitably gauge the model's question-answering proficiency.", + "bbox": [ + 496, + 545, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 790, + 669, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of China under Grant 62073105, U23A20389 and 62306095, by the Natural Science Foundation of Heilongjiang Province of China under Grant ZD2022F002, and by the Heilongjiang Touyan Innovation Team Program.", + "bbox": [ + 496, + 816, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27225", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086. Computer Vision Foundation / IEEE Computer Society, 2018. 6", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 2425-2433, 2015. 2", + "[3] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems, pages 32897-32912. Curran Associates, Inc., 2022. 6", + "[4] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a machine? dataset and methods for multilingual image question. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2", + "[5] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 6325-6334. IEEE Computer Society, 2017. 1, 2", + "[6] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2", + "[7] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 6", + "[8] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[9] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10264-10273, 2020. 6", + "[10] Aniruddha Kembhavi, Min Joon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 5376-5384. IEEE Computer Society, 2017. 1, 2", + "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalan-" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vision, 123(1):32-73, 2017. 2", + "[12] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 2", + "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6", + "[14] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 1, 2", + "[15] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1682-1690, 2014. 2", + "[16] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[17] OpenAI, Josh Achiam, Steven Adler, et al. Gpt-4 technical report, 2024. 6", + "[18] Revanth Gangi Reddy, Xilin Rui, Manling Li, Xudong Lin, Haoyang Wen, Jaemin Cho, Lifu Huang, Mohit Bansal, Avirup Sil, Shih-Fu Chang, Alexander G. Schwing, and Heng Ji. Mumuqa: Multimedia multi-hop news question answering via cross-media knowledge extraction and grounding. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 11200-11208. AAAI Press, 2022. 1, 3", + "[19] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2", + "[20] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision - ECCV 2022, pages 146-162, Cham, 2022. Springer Nature Switzerland. 2", + "[21] Alon Talmor, Ori Yoran, Amnon Catav, Dan Lahav, Yizhong Wang, Akari Asai, Gabriel Ilharco, Hannaneh Hajishirzi, and Jonathan Berant. Multimodalqa: complex question answering over text, tables and images. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021. 1" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 4, 7", + "[23] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton Van Den Henge. Explicit knowledge-based reasoning for visual question answering. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, page 1290–1296. AAAI Press, 2017. 1, 2", + "[24] Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. FVQA: fact-based visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 40(10): 2413-2427, 2018. 1, 2", + "[25] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 6", + "[26] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6281-6290. Computer Vision Foundation / IEEE, 2019. 6", + "[27] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 6", + "[28] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4995–5004, 2016. 2" + ], + "bbox": [ + 78, + 90, + 468, + 625 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27227", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_model.json b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_model.json new file mode 100644 index 0000000000000000000000000000000000000000..26a6728184cb9d4b2dc60c8d086c54db3ef8f3a1 --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.13, + 0.888, + 0.178 + ], + "angle": 0, + "content": "VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.211, + 0.613, + 0.231 + ], + "angle": 0, + "content": "Kang Chen Xiangqian Wu" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.231, + 0.724, + 0.249 + ], + "angle": 0, + "content": "Faculty of Computing, Harbin Institute of Technology (HIT)" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.25, + 0.619, + 0.265 + ], + "angle": 0, + "content": "Suzhou Research Institute of HIT" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.269, + 0.672, + 0.284 + ], + "angle": 0, + "content": "chenkangcs@stu.hit.edu.cn, xqwu@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.318, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.474, + 0.79 + ], + "angle": 0, + "content": "Achieving the optimal form of Visual Question Answering mandates a profound grasp of understanding, grounding, and reasoning within the intersecting domains of vision and language. Traditional VQA benchmarks have predominantly focused on simplistic tasks such as counting, visual attributes, and object detection, which do not necessitate intricate cross-modal information understanding and inference. Motivated by the need for a more comprehensive evaluation, we introduce a novel dataset comprising 23,781 questions derived from 10,124 image-text pairs. Specifically, the task of this dataset requires the model to align multimedia representations of the same entity to implement multi-hop reasoning between image and text and finally use natural language to answer the question. Furthermore, we evaluate this VTQA dataset, comparing the performance of both state-of-the-art VQA models and our proposed baseline model, the Key Entity Cross-Media Reasoning Network (KECMRN). The VTQA task poses formidable challenges for traditional VQA models, underscoring its intrinsic complexity. Conversely, KECMRN exhibits a modest improvement, signifying its potential in multimedia entity alignment and multi-step reasoning. Our analysis underscores the diversity, difficulty, and scale of the VTQA task compared to previous multimodal QA datasets. In conclusion, we anticipate that this dataset will serve as a pivotal resource for advancing and evaluating models proficient in multimedia entity alignment, multi-step reasoning, and open-ended answer generation. Our dataset and code is available at https://visual-text-qa.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.819, + 0.21, + 0.835 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.845, + 0.47, + 0.876 + ], + "angle": 0, + "content": "A paramount objective in AI research is to endow systems with the capacity to comprehend the intricacies of the real" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.32, + 0.893, + 0.44 + ], + "angle": 0, + "content": "world, akin to human understanding. Question Answering (QA) stands out as an effective task for evaluating the cognitive capabilities of AI systems. To answer questions, people need to extract information from multiple modalities, such as text, images and structured data like knowledge bases, graphs and tables. And furthermore, people need to align the information and do multi-step reasoning between different modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.732 + ], + "angle": 0, + "content": "Although Visual Question Answering (VQA) [5] has been widely researched as a multimedia QA task, VQA models only extract information from image when answering questions and focus mainly on scene recognition, counting, color and other visual detection tasks, which do not require much logical reasoning or assignment between different modalities. Only recently, there are some attempts to introduce more modal information into VQA tasks. For example: (1) FVQA [24] and KBVQA [23] combine knowledge base (KB) with VQA task, which requires the ability of knowledge understanding and multi-step reasoning. But it is difficult to construct a comprehensive KB in real world, thus limiting its ability to address open-ended questions; (2) TextbookQA [10] and ScienceQA [14] use textbook as data sources, involving texts, images, tables and other multimodal information. But most images in these datasets are manually drawn schematic diagrams, and the questions are all in a multiple-choice setting, which are far from the real world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address these issues, more recently, there appear some new datasets such as MultiModalQA [21] and MuMuQA [18], which involve reasoning across texts, images and tables. However, it's noteworthy that all these datasets are conducted with extracted QA. And in the MultiModalQA dataset, each image corresponds to a Wikipedia entity, simplifying image reasoning to ranking images based on these entities, which diminishes the necessity for crossmedia reasoning abilities. As for MuMuQA, although it requires grounding between image and text along with multi-hop reasoning, there are still some problems: (1) the data" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.218, + 0.901 + ], + "angle": 0, + "content": "* corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27218" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.095, + 0.143, + 0.104 + ], + "angle": 0, + "content": "Image:" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.106, + 0.393, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.409, + 0.094, + 0.51, + 0.103 + ], + "angle": 0, + "content": "Chinese Annotation:" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.11, + 0.58, + 0.173 + ], + "angle": 0, + "content": "公司举行了一场饭桌会议，顾名思义就是一边吃饭一边商讨事宜。爱莲娜的金色头发果然是人群中最显眼的一个，乌尔蒂盖拉穿着橘黄色开衫，正在埋头切一块牛排；赛特温穿着深蓝色衬衫，胸前挂着会议牌，他的左手正抬起来准备去拿酒杯。寒特温的手左边，卡维利耶和阿吉姆正严肃地讨论下个月的预算案。卡维利耶一头褐色齐肩短发，干净利落。" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.178, + 0.533, + 0.186 + ], + "angle": 0, + "content": "Q1：爱莲娜穿着什么类型的上衣？" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.19, + 0.461, + 0.199 + ], + "angle": 0, + "content": "A1：西装外套" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.204, + 0.52, + 0.212 + ], + "angle": 0, + "content": "Q2：乌尔蒂盖拉的右手边是谁？" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.216, + 0.457, + 0.224 + ], + "angle": 0, + "content": "A2：卡维利耶" + }, + { + "type": "title", + "bbox": [ + 0.608, + 0.094, + 0.706, + 0.104 + ], + "angle": 0, + "content": "English Translation:" + }, + { + "type": "text", + "bbox": [ + 0.605, + 0.11, + 0.865, + 0.173 + ], + "angle": 0, + "content": "The company held a table meeting, which, as the name implies, is to discuss matters while eating. Elena's blond hair is indeed the most conspicuous one in the crowd, and Urigera, wearing an orange cardigan, is burying her head in cutting a steak; Setwin was wearing a dark blue shirt, with a meeting card hanging on his chest, and his left hand was lifting up to get a glass of wine. On the left of Setwin, Kavelier and Agum are seriously discussing the budget for next month. Kavelier has short brown shoulder-length hair, clean and clean." + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.178, + 0.761, + 0.186 + ], + "angle": 0, + "content": "Q1: What type of blouse does Elena wear?" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.19, + 0.637, + 0.198 + ], + "angle": 0, + "content": "A1: Suit" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.204, + 0.777, + 0.212 + ], + "angle": 0, + "content": "Q2: Who is on the right-hand side of Urtigera?" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.216, + 0.651, + 0.224 + ], + "angle": 0, + "content": "A2: Kavelier" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.894, + 0.281 + ], + "angle": 0, + "content": "Figure 1. Example in our dataset with the question-answer pairs and their corresponding image and text. Different representations of the same object in text and image are identified with the same color. For example, 'Elena' in the text and the object bounding box corresponding to 'Elena' in the image are marked red." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.468, + 0.444 + ], + "angle": 0, + "content": "is all from news, resulting in most questions related to human beings; (2) questions all follow a specific pattern: first perform image entity grounding and then find the answer in the news body text; (3) there is only 1384 human-curated examples and the training data is automatically generated, which leads to poor quality and difficult to train with. Current multimedia benchmarks are still far from the real-world QA scene and cannot measure the multimedia understanding ability of AI system well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.468, + 0.536 + ], + "angle": 0, + "content": "In order to address these gaps, we propose a novel task named 'Visual Text Question Answering' (VTQA), accompanied by a dedicated dataset, which includes only questions that require multi-hop reasoning through both images and text. All the annotations in VTQA dataset are first marked in Chinese, and then translated into English." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.779 + ], + "angle": 0, + "content": "To answer VTQA questions, the proposed model needs to: (1) learn to identify entities in image and text referred to the question, (2) align multimedia representations of the same entity, and (3) conduct multi-step reasoning between text and image and output open-ended answer. The VTQA dataset consists of 10,124 image-text pairs and 23,781 questions. The images are real images from MSCOCO dataset [12], containing a variety of entities. The annotators are required to first annotate relevant text according to the image, and then ask questions based on the image-text pair, and finally answer the question open-ended. To ensure textual richness, we enforce a minimum text length requirement of over 100 Chinese words. Additionally, we systematically exclude questions that can be answered solely based on either the image or the text, thereby guaranteeing the complexity of the questions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Moreover, we conduct an evaluation on the VTQA dataset, employing both state-of-the-art VQA models and our newly proposed baseline model. The baseline model we propose is called Key Entity Cross-Media Reasoning Network (KECMRN), which follows the general paradigm for answering questions on this dataset by iteratively performing key entity extraction and alignment, and cross-modal multi-step reasoning to answer questions. Evaluation re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.383 + ], + "angle": 0, + "content": "sults on the VTQA dataset show that existing state-of-the-art VQA models struggle to achieve satisfactory performance on this dataset, which illustrates the challenges and potential of this dataset for cross-modal question answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.522 + ], + "angle": 0, + "content": "The contributions of this work can be summarized as follows: (1) we propose a new cross-modal QA dataset named VTQA. Information diversity, multimedia multi-step reasoning and open-ended answer make our dataset more challenging than the existing datasets; (2) we benchmark the state-of-the-art VQA model on our new dataset and show the performance of these models degrades drastically; (3) we propose a baseline that is capable of multimedia entity alignment and multi-step reasoning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.642, + 0.556 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) aims to answer a natural language question based on an image, which requires model to understand and reason in the vision-language joint space. Several datasets have been proposed in the past few years, such as DAQUAR [15], FM-IQA [4], VQA [2, 5], COCOQA [19], Visual7W [28], Visual Genome [11], GQA [8], OKVQA [16], A-OKVQA [20], VizWizQA [6] and so on. However, the natural language questions in these datasets can be regarded as instructions to guide the model to complete visual tasks such as object detection, scene recognition, counting, etc. And these questions often only involve the surface information of the image, without considering the complexities between different modalities. Understanding of the reasoning process. In order to improve the multimodal perception ability of the model, we need to design more challenging problems, requiring the model to combine the features of images and texts for cross-modal knowledge fusion and logical reasoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, some datasets attempt to introduce more modal information into VQA tasks, such as FVQA [24], KB-VQA [23], TextbookQA [10], and ScienceQA [14]. But the information used in these datasets includes manually col" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27219" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "lected databases or manually drawn schematic diagrams in textbooks, which are far from real-world QA scenarios. The VTQA we propose is different from these, using the semantically rich real images of the COCO dataset and human-annotated text passages, which can be better generalized to real-world QA scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.349 + ], + "angle": 0, + "content": "The most similar to our dataset is MuMuQA [18]. The MuMuQA dataset is a news-based QA dataset whose images and questions mainly involve people and events. This dataset has only a few human-labeled data, and most of the data is generated by automated methods, so there may be a lot of noise and errors. In contrast, our VTQA dataset covers a variety of genres and topics, and all texts, questions, and answers are annotated by humans, ensuring the high quality and accuracy of the data. Our VTQA dataset has 23781 items, which is richer and more diverse than MuMuQA dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.226, + 0.381 + ], + "angle": 0, + "content": "3. VTQA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.468, + 0.466 + ], + "angle": 0, + "content": "In this section, we present the details of our Visual Text Question Answering (VTQA) dataset. We first introduce the VTQA task of our dataset (Section 3.1) and then detail how the dataset was constructed (Section 3.2), along with the statistical properties of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.206, + 0.494 + ], + "angle": 0, + "content": "3.1. VTQA Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.468, + 0.697 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, given an image-text pair and a question, a system is required to answer the question by natural language. Importantly, the system needs to: (1) analyze the question and find out the key entities, (2) align the key entities between image and text, and (3) generate the answer according to the question and aligned entities. For example, in Fig. 1, the key entity of Q1 is \"Elena\". By referencing the descriptor \"gold hair\" in the text, it is discerned that the second person from the right in the image is \"Elena\". Subsequently, the answer \"suit\" is generated based on the visual information. As for Q2, which is a more complex question, the previous steps need to be repeated several times to answer it." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.276, + 0.725 + ], + "angle": 0, + "content": "3.2. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.809 + ], + "angle": 0, + "content": "To promote the progress of this open-ended multimedia multi-hop VTQA task, we collect a new dataset. Our dataset consists of 10124 image-text pairs and 23,781 questions. We collect data through newly developed annotation interface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In the first round of labeling, we present images from COCO and corresponding image descriptions, as well as the object detection labels, to the annotators. The annotators are required to generate a text that exceeds 100 words, which should involve the object in the image and contain information that is not included in the image description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "To ensure compliance with annotation requirements, annotators can choose to skip some images (12198 images were ultimately skipped, 9830 images were selected, and some of them were selected more than once). The annotators then come up with questions based on the image-text pair and the annotation process requires that the questions cannot be answered only by image or text. Each image-text pair is labeled with 1-4 questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.43 + ], + "angle": 0, + "content": "In a second round of labeling, different annotators were asked to determine whether the question could be answered solely by relying on either images or text within the corresponding text-question pair or image-question pair. Through this step, we filtered down to 23781 questions from a pool of 28919 questions. The questions that pass this step will be labeled with answers and answer categories. We set three categories for answers: (1) \\(YN\\) means yes-or-no answer, (2) \\(E\\) means that the answer is extracted from the text, and (3) \\(G\\) means that the answer is generated from the text-image pair. And the annotators need to label 'yes' or 'no' in English for the yes-or-no answer, since there are too many words to express 'yes' or 'no' in Chinese, for example, '可以' and '是的' both mean 'yes'." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.498 + ], + "angle": 0, + "content": "We randomly split the dataset into training, validation, test-dev and test splits and each image will only appear in one split. Each split has 11312, 1245, 2189, 9035 samples, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.664, + 0.541 + ], + "angle": 0, + "content": "3.3. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Table 1 shows a comparison of VTQA and other VQA datasets. As shown in the table, VTQA is much larger than most other datasets, especially with a much larger number of multimodal questions than existing multimodal QA datasets. The text length of the VTQA dataset is shorter than that of MuMuQA, but still much longer than the length of the questions in all datasets, and the texts in VTQA are manually generated, containing more types of entities than existing texts extracted from textbooks, wikis, and news." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Images in VTQA dataset are from MSCOCO dataset, which contains multiple objects and rich contextual information. Fig. 2 presents the top-10 categories distribution of the images used in this dataset. Unlike MuMuQA, in which most images are related to people, images in our dataset contain more kinds of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The texts and questions length statistics are shown in Fig. 3. It is evident that most texts range from 300 to 600 and most questions range from 9 to 21. Obviously, compared with the previous VQA datasets (usually less than 20 words), our dataset also puts forward higher requirements for text understanding to extract information from the long text." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27220" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.897, + 0.228 + ], + "angle": 0, + "content": "
#I#Q#MMQContextsAvgQAvgTAnswer TypeImage SourceText Source
VQA200K1.1M-I6.1-Open/MCCOCO-
FVQA58265826-I+KB9.5-OpenCOCO-
KBVQA24022402-I+KB6.8-OpenCOCO-
TextbookQA34552626012567I+T9.2-MCTextbookTextbook
ScienceQA10332212086532I+T12.1-MCTextbookTextbook
MultimodalQA57058299188240I+T18.266.2OpenWikipediaWikipedia
MuMuQA*138413841384I+T11.8633.5OpenNewsNews
VTQA98302378123781I+T10238.4OpenCOCOannotated
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 1. Statistics for VTQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, #MMQ: number of multimodal questions, AvgQ: average question length, AvgT: average text length. *We only count the manually annotated parts of the MuMuQA dataset." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.328, + 0.436, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.543, + 0.471, + 0.571 + ], + "angle": 0, + "content": "Figure 2. Top-10 categories distribution of the images used in our dataset." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.626, + 0.268, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.627, + 0.455, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.82, + 0.47, + 0.849 + ], + "angle": 0, + "content": "Figure 3. Percentage of questions and texts with different Chinese word lengths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.64, + 0.324 + ], + "angle": 0, + "content": "4. VTQA Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.514 + ], + "angle": 0, + "content": "In this section, we describe a competitive baseline method for evaluation on our benchmark, which is called Key Entity Cross-Media Reasoning Network (KECMRN). Before presenting the KECMRN, we first introduce its basic component, the KECMR layer. The KECMR layer is a modular composition which consists of one Key Entity Extract (KEE) layer and multiple Cross-Media Reason (CMR) layers. The KEE layer and CMR layer are composed of attention unit and feed-forward unit from [22]. Then we use the KECMR module with other layers to combine our KECMRN. An overview flowchart of KECMRN is shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.804, + 0.537 + ], + "angle": 0, + "content": "4.1. Attention and Feed-Forward Units" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.59 + ], + "angle": 0, + "content": "As shown in [22], the combination of attention unit and feed-forward unit has strong representational and learning ability. We use the same settings to construct our units." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Multi-Head Scaled Dot-Product Attention Unit. Given a query \\( q \\in R^{1 \\times d} \\), n key-value pairs (packed into a key matrix \\( K \\in R^{n \\times d} \\) and a value matrix \\( V \\in R^{n \\times d} \\)), the attended feature is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.673, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Multi-head attention divides the input into \\( h \\) parts and makes single attention on each part. The attended feature is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.772, + 0.891, + 0.79 + ], + "angle": 0, + "content": "\\[\nM H (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.799, + 0.891, + 0.818 + ], + "angle": 0, + "content": "\\[\nh e a d _ {i} = \\text {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.823, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\( W_{i}^{Q}, W_{i}^{Q}, W_{i}^{Q} \\in R^{d \\times d_{h}} \\) and \\( W^{O} \\in R^{(h \\times d_{h}) \\times d} \\) are the projection matrices. \\( d_{h} \\) is the dimension of the output features from each head." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Feed-Forward Unit. The feedforward unit takes the output features of the multi-head attention layer, and further" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.163, + 0.092, + 0.813, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.291, + 0.722, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Overall flowchart of the Key Entity Cross-Media Reasoning Network." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.327, + 0.451, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.527, + 0.38, + 0.541 + ], + "angle": 0, + "content": "Figure 5. Key Entity Extract Layer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.47, + 0.598 + ], + "angle": 0, + "content": "transforms them through two fully-connected layers with ReLU activation in between." + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.615, + 0.469, + 0.632 + ], + "angle": 0, + "content": "\\[\nF F N (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.47, + 0.686 + ], + "angle": 0, + "content": "While the linear transformations are the same across different positions, they use different parameters from layer to layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.31, + 0.712 + ], + "angle": 0, + "content": "4.2. Key Entity Extract Layer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.838 + ], + "angle": 0, + "content": "As stated in Section 3.1, the first step to answer VTQA questions is to find out the key entity according to the questions. We compose the attention unit and feed-forward unit to integrate question information into the text and the image respectively. Then we apply a fully-connected layer to the question-aware text/image features to get the score for each feature. Finally, we extract the top-k features as key entities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The complete KKE layer is shown in the Fig. 5. Mathematical, given input text features \\( T \\), image features \\( I \\) and question features \\( Q \\), the KEE layer can be formulated by (the processing of feature \\( I \\) is aligned with that of feature" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.328, + 0.891, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.444, + 0.89, + 0.472 + ], + "angle": 0, + "content": "Figure 6. Cross-Media Reason Layer. 'G' means gather and 'S' means scatter." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.5, + 0.525, + 0.513 + ], + "angle": 0, + "content": "T):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.529, + 0.891, + 0.546 + ], + "angle": 0, + "content": "\\[\nQ = F F N (M H (Q, Q, Q)) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.551, + 0.891, + 0.568 + ], + "angle": 0, + "content": "\\[\nT = F F N (M H (M H (T, T, T), Q, Q)) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.575, + 0.891, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s c o r e} _ {T} = W _ {T} T + b _ {T} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.658 + ], + "angle": 0, + "content": "The final score will be used to sort the image/text features and the top-k features are considered as image/text key entities. To unify the expression, we treat all the question features as key entities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.667, + 0.747, + 0.683 + ], + "angle": 0, + "content": "4.3. Cross-Media Reason Layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.827 + ], + "angle": 0, + "content": "The CMR layer is designed for multi-step reasoning across medias. As show in Fig. 6, we first gather the key entities as \\( S \\in R^{l_k \\times d} \\) from input features, where \\( l_k \\) is the total number of all key entities. Then the key entities pass through self-attention, global cross-attention based on original features and feed-forward unit. Finally, the key entities are scattered to the input features. As the requirement for multi-step reasoning, we inserted multiple CMR layers in a KECMR module. The CMR layer can be formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.84, + 0.891, + 0.857 + ], + "angle": 0, + "content": "\\[\nS = \\operatorname {g a t h e r} ([ T, I, Q ]) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.863, + 0.891, + 0.88 + ], + "angle": 0, + "content": "\\[\nS = F F N (M H (M H (S, S, S), [ T, I, Q ], [ T, I, Q ])) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.885, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\n[ T, I, Q ] = \\operatorname {s c a t t e r} (S, [ T, I, Q ]) \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27222" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.432, + 0.108 + ], + "angle": 0, + "content": "4.4. Key Entity Cross-Media Reason Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.47, + 0.28 + ], + "angle": 0, + "content": "We use the same network framework as [26] but replace MCA layer with our KECMR module and add an extra text stream. As show in Figure 4, The input image is represented as a set of regional visual features in a bottom-up manner [1]. The input question and text are transformed to features by passing through a one-layer LSTM network [7]. Then we use our KECMR module several times to extract key entities and conduct multi-steps cross-media reasoning. Finally, we use the attention reduce layer to fuse the multimedia features and project the fused feature into the answer probability distribution." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.281, + 0.419, + 0.296 + ], + "angle": 0, + "content": "The attention reduce layer can be formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.321, + 0.469, + 0.339 + ], + "angle": 0, + "content": "\\[\nM L P (x) = W _ {2} \\left(\\max \\left(0, x W _ {1} + b _ {1}\\right)\\right) + b _ {2} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.348, + 0.469, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t R e} (X) = \\operatorname {s u m} (\\operatorname {s o f t m a x} (M L P (X)) X) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.469, + 0.432 + ], + "angle": 0, + "content": "We can obtain the attended image feature \\( I_{a} \\) by passing image feature I through attention reduce layer, and the same applies to text feature T and question feature Q. Then we fuse the multimedia features by:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.446, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\nz = W _ {i} I _ {a} + W _ {t} T _ {a} + W _ {q} Q _ {a} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.543 + ], + "angle": 0, + "content": "The fused feature \\( \\mathbf{z} \\) is projected into a vector \\( s \\in R^{N} \\) followed by a sigmoid function, where \\( \\mathbf{N} \\) is the number of the answers in the training set. We use binary cross-entropy (BCE) as the loss function to train an N-way classifier on top of the fused feature \\( \\mathbf{z} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.21, + 0.574 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.627 + ], + "angle": 0, + "content": "In this section, we evaluate the current state-of-the-art VQA model and our KECMRN model, and provide results of them in VTQA dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.262, + 0.65 + ], + "angle": 0, + "content": "5.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.689 + ], + "angle": 0, + "content": "As the answers divided into three types, we use different metrics for distinct types of answers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Exact match (EM). This metric measures the percentage of predictions that match the ground truth answer exactly and will be used in all types of answers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.81 + ], + "angle": 0, + "content": "(Macro-averaged) F1 score (F1). This metric measures the average overlap between the prediction and ground truth answer. We treat the prediction and ground truth as bags of tokens and compute their F1. This metric will be used for the \\(E\\) and \\(G\\) types of answers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "YN accuracy (YNAcc). This metric is only used for the YN type of answers. The answer will be transformed into 'yes' or 'no' by a pre-defined yes-or-no dictionary, which is manually collected based on the answers in our dataset. Then we calculate the accuracy just based on the yes-or-no answer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.719, + 0.107 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.295 + ], + "angle": 0, + "content": "KECMRN: the hyper-parameters of our model used in the experiments are as follows. The dimensionality of input image features, input question features, input text features and fused multimodal features are 2,048, 1024, 1024, and 2,048, respectively. The latent dimensionality \\(d\\) in the multi-head attention is 1024, the number of heads \\(h\\) is set to 8. The number of CMR layer in each KECMR module and the number of KECMR module are 2 and 2. The number of key entities \\(k\\) is set to 16. We train our model on the train set for 13 epochs and evaluate our model on test set. The Faster-RCNN used for image feature extraction is pretrained and frozen, while all other modules are trained from scratch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.537 + ], + "angle": 0, + "content": "GPT4V [17]/LLaVA [13]/MMICL [27]: popular multimodal large language models (MLLM). These models demonstrated impressive multimodel chat abilities and exceptional performance across numerous intricate visual reasoning datasets, particularly excelling in zero-shot evaluations. Notably, LLaVA achieved an impressive \\(80.0\\%\\) accuracy, while GPT4V attained \\(77.2\\%\\) and MMICL attained \\(70.3\\%\\) on the VQAv2 dataset, underscoring their remarkable capabilities in the realm of multimodal understanding and visual question answering. We follow the same evaluation strategy as on the VQAv2 dataset, with the sole modification of adapting the prompt to 'Search for answers in the context above, do not use additional knowledge beyond the text, and answer the question using a single word or phrase'. This adjustment facilitated enhanced performance on VTQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "BEiT3 [25]/VLMo [3]: visual language models directly pre-trained on image-text pairs, offering versatility in addressing a spectrum of visual-language tasks through finetuning on downstream tasks. Notably, on the VQAv2 dataset, BEiT3-base achieved an accuracy of \\(77.65\\%\\), while VLMo-base demonstrated a commendable accuracy of \\(76.6\\%\\). To evaluate the transferability of BEiT3 and VLMo on our VTQA task, we follow the same fine-tuning strategy as on the VQAv2 dataset, but we concatenate the text and questions into a single paragraph of text and increase the maximum number of training epochs from 10 to 50 to obtain better results, since our dataset is relatively small compared to VQAv2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.884 + ], + "angle": 0, + "content": "MCAN-region [26]/MCAN-grid [9]: models based on MCAN [26] but using different image features. They achieve SOTA results on VQAv2 dataset (70.93% for MCAN-region and 72.71% for MCAN-grid) among non-pretrained visual-language models(non-PVLM), which indicates that these models were not pretrained on visual-language data, although individual pretraining on either visual or text data is allowed. We concatenate the text and questions and increase the maximum number of input words to train MCAN in VTQA task." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Trans-CA/Trans-SA: models based on Trans" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27223" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.47, + 0.367 + ], + "angle": 0, + "content": "
MethodsEMYN-ACCE-F1G-F1
English Version
LLaVA39.9065.0560.0626.20
GPT4V48.9375.5970.2448.37
MMICL38.7662.9062.3245.96
BEiT3-base41.0256.5150.1828.14
VLMO-base51.8565.4664.5942.74
MCAN-region56.0972.0667.0340.52
MCAN-grid55.8870.7365.3939.19
Trans-CA57.3873.5769.0244.14
Trans-SA54.3967.8965.9541.11
KECMRN(ours)57.9574.9968.9144.25
Chinese Version
MCAN-region49.7871.2659.9445.74
MCAN-grid50.2274.1759.0245.59
Trans-CA49.8774.2759.6347.45
Trans-SA48.9973.2958.2646.31
KECMRN(ours)51.3277.5960.5251.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.377, + 0.472, + 0.406 + ], + "angle": 0, + "content": "Table 2. Results of our KECMRN on test set compared with other multimodal models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.583 + ], + "angle": 0, + "content": "former [22]. Trans-CA treat image, text and question features as independent input and perform Self Attention and Question Aware Cross Attention in each layer. The layers in Trans-CA resemble those of the KEE module, excluding the \"score&sort\" step. Conversely, Trans-SA adopts a different approach by concatenating all image, text, and question features and only applying Self Attention in each layer. The layers in Trans-SA resemble those of the CMR module, excluding the \"Global Cross-Attention\" step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.273, + 0.607 + ], + "angle": 0, + "content": "5.3. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Table 2 provides the results of various baselines on the test set of our VTQA benchmark. VLMs, including MMICL, LLaVA, BEiT3, and VLMo, primarily pretrained on English datasets, are exclusively evaluated in the English version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.856 + ], + "angle": 0, + "content": "While models such as GPT4V, MMICL, and LLaVA demonstrate superior performance across a variety of visual-language datasets in zero-shot evaluation, they encounter a significant drop in performance when evaluated on the VTQA dataset. This is partly because the model lacks exposure to our training data during training, unlike the VQA dataset included in the model's VL fine-tuning data. The decline also results from the heightened complexity of our VTQA task compared to VQA, posing new challenges to the model's image-text alignment and inference capabilities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "BEiT3 and VLMo, pretrained directly on image-text pairs, achieve impressive results on various multimodal datasets through fine-tuning on downstream tasks. Despite" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.088, + 0.863, + 0.169 + ], + "angle": 0, + "content": "
MethodsEMYN-ACCE-F1G-F1
Trans(1)55.4572.4365.3542.27
Trans(3)55.0770.9865.743.58
Trans(6)54.0769.4364.8343.91
LSTM57.9574.9968.9144.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.179, + 0.884, + 0.193 + ], + "angle": 0, + "content": "Table 3. Results of different text encoders based on KECMRN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.324 + ], + "angle": 0, + "content": "fine-tuning on VTQA training data, BEiT3 and VLMo still exhibit a substantial performance decrease. This is partly attributed to the limitation that the pretraining data for BEiT3 and VLMo only includes image-text pairs with relatively short texts, while the texts in the VTQA dataset are longer, averaging 284 words in English version. This, once again, underscores the challenges inherent in the VTQA task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.445 + ], + "angle": 0, + "content": "While MCAN models outperform the pre-trained models, a notable decrease is observed. In addition, Trans-CA exceeded the results of all VQA models, while the results of Trans-SA are relatively poor. This highlights the necessity for a more intricate independent processing of text, image, and question features in our VTQA dataset. Our KECMRN achieves the best result with \\(57.95\\%\\) on the English version and \\(51.32\\%\\) in Chinese version." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.733 + ], + "angle": 0, + "content": "Fig. 7 illustrates the outcomes of attention visualization across different models. The attention weights from the 'Attention Reduce' module were employed to colorize image regions and organize the Top Words. Notably, our KECMRN model exhibits precise identification of Zhang Yanqi on the left side, engaged in ball play. This accurate recognition is based on key phrases such as sports balls and dribbling skills, leading to a correct inference for the associated question ('Yes'). In contrast, although MCAN also directs attention to the person on the left, the absence of Key-Entity Extract (KEE) causes attention to be excessively dispersed. Consequently, there is an inaccurate localization of the person's upper garment, resulting in an incorrect answer. Furthermore, Trans-SA mistakenly focuses on the person on the right, while Trans-CA struggles to discern whether Zhang Yanqi is positioned on the left or right, both yielding incorrect answers. The attention visualization results underscore the superior precision of our KECMRN model in focusing on the correct visual regions and words." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.655, + 0.759 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We conducted a number of ablation experiments to investigate the reasons for the effectiveness of KECMR. All the reported results are evaluated on English version." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Text Encoders: Table 3 provides an overview of the results obtained using different text encoders in conjunction with KECMRN. In this table, Trans(k) denotes a k-layer transformer encoder model. Contrary to expectations, the performance of deeper transformer encoder models tends to be inferior, with LSTM outperforming all transformer en" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27224" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.105, + 0.353, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.226, + 0.155, + 0.236 + ], + "angle": 0, + "content": "Text:" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.237, + 0.348, + 0.366 + ], + "angle": 0, + "content": "Song Shuai and Zhang Jiaqi are two local sports ball enthusiasts whose video about sports ball on the short video platform caught the attention of a club in a nearby city. Today they were invited to the city where the club is located for a technical exchange, and now Song Shuai is talking to the head of the club, and Zhang Jiaqi is too obsessed with sports balls on the road and still does not forget to practice dribbling skills." + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.378, + 0.178, + 0.389 + ], + "angle": 0, + "content": "Question:" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.39, + 0.303, + 0.402 + ], + "angle": 0, + "content": "Is Zhang Jiaqi wearing a blue top?" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.414, + 0.172, + 0.424 + ], + "angle": 0, + "content": "Answer:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.425, + 0.149, + 0.435 + ], + "angle": 0, + "content": "Yes" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.09, + 0.596, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.213, + 0.432, + 0.224 + ], + "angle": 0, + "content": "Top Words:" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.225, + 0.575, + 0.258 + ], + "angle": 0, + "content": "balls, Jiaqi, enthusiasts, Zhang, road, skills, obsessed, dribbling, sports, forget Predict Answer: Yes" + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.263, + 0.508, + 0.273 + ], + "angle": 0, + "content": "KECMRN" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.279, + 0.593, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.401, + 0.432, + 0.412 + ], + "angle": 0, + "content": "Top Words:" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.413, + 0.57, + 0.446 + ], + "angle": 0, + "content": "sports, dribbling, Jiaqi, club, Shuai, Zhang, video, caught, platform, talking Predict Answer: No" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.452, + 0.505, + 0.462 + ], + "angle": 0, + "content": "Trans-SA" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.091, + 0.849, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.213, + 0.688, + 0.224 + ], + "angle": 0, + "content": "Top Words:" + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.225, + 0.841, + 0.258 + ], + "angle": 0, + "content": "skills, dribbling, enthusiasts, nearby, road, balls, Zhang, club, still, sports Predict Answer: No" + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.263, + 0.755, + 0.273 + ], + "angle": 0, + "content": "MCAN" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.279, + 0.849, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.401, + 0.688, + 0.413 + ], + "angle": 0, + "content": "Top Words:" + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.413, + 0.832, + 0.447 + ], + "angle": 0, + "content": "Jiaqi, Song, exchange, dribbling, Zhang, obsessed, head, forget, short, Shuai Predict Answer: No" + }, + { + "type": "text", + "bbox": [ + 0.706, + 0.451, + 0.762, + 0.462 + ], + "angle": 0, + "content": "Trans-CA" + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.479, + 0.688, + 0.493 + ], + "angle": 0, + "content": "Figure 7. An example of attention visualization for different models." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.516, + 0.451, + 0.597 + ], + "angle": 0, + "content": "
MethodsEMYN-ACCE-F1G-F1
Trans-CA57.3873.5769.0244.14
CA+SA57.1773.5768.9642.88
KEE+SA57.6374.4269.143.06
KEE+CMR57.9574.9968.9144.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.607, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Table 4. Results of the KECMRN model with different module combinations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.68, + 0.47, + 0.725 + ], + "angle": 0, + "content": "coder models. This phenomenon is attributed to both insufficient training data and the intricate structure of transformer models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "KEE/CA and CMR/SA: Table 4 displays the results of the KECMRN model with different module combinations. In this table, CA+SA denotes the alternate use of CA and SA layers, similar to our KECMRN design. KEE+SA follows a similar pattern. Compared to Trans-CA, replacing some CA layers with SA layers results in a slight performance decrease (↓0.21). Conversely, the inclusion of the KEE layer improves performance (↑0.46) by concentrating information on key entities. Additionally, the CMR layer further optimizes features related to key entities, leading to enhanced results (↑0.32)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.519, + 0.619, + 0.535 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.773 + ], + "angle": 0, + "content": "We propose a novel cross-modal question answering dataset, VTQA, which necessitates models to acquire pertinent information from both text and image sources and perform complex cross-modal reasoning to answer questions. Our experimental findings reveal that even state-of-the-art pretrained models fall short of achieving satisfactory performance on VTQA, underscoring the challenges and potential inherent in the dataset for cross-modal question-answering tasks. Looking ahead, we aim to augment the dataset's size and coverage and provide additional reasoning annotations in future work to more effectively assess and catalyze the development of cross-modal reasoning models. Simultaneously, we aspire to explore more fitting evaluation metrics to surmount the limitations of EM metrics and more equitably gauge the model's question-answering proficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.671, + 0.808 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of China under Grant 62073105, U23A20389 and 62306095, by the Natural Science Foundation of Heilongjiang Province of China under Grant ZD2022F002, and by the Heilongjiang Touyan Innovation Team Program." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27225" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.472, + 0.295 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 2425-2433, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.472, + 0.38 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems, pages 32897-32912. Curran Associates, Inc., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[4] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a machine? dataset and methods for multilingual image question. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[5] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 6325-6334. IEEE Computer Society, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.605 + ], + "angle": 0, + "content": "[6] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.472, + 0.634 + ], + "angle": 0, + "content": "[7] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[8] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[9] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10264-10273, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.775, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[10] Aniruddha Kembhavi, Min Joon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 5376-5384. IEEE Computer Society, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalan-" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "tidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vision, 123(1):32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[12] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[14] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[15] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1682-1690, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.894, + 0.503 + ], + "angle": 0, + "content": "[16] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[17] OpenAI, Josh Achiam, Steven Adler, et al. Gpt-4 technical report, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[18] Revanth Gangi Reddy, Xilin Rui, Manling Li, Xudong Lin, Haoyang Wen, Jaemin Cho, Lifu Huang, Mohit Bansal, Avirup Sil, Shih-Fu Chang, Alexander G. Schwing, and Heng Ji. Mumuqa: Multimedia multi-hop news question answering via cross-media knowledge extraction and grounding. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 11200-11208. AAAI Press, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[19] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[20] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision - ECCV 2022, pages 146-162, Cham, 2022. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Alon Talmor, Ori Yoran, Amnon Catav, Dan Lahav, Yizhong Wang, Akari Asai, Gabriel Ilharco, Hannaneh Hajishirzi, and Jonathan Berant. Multimodalqa: complex question answering over text, tables and images. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27226" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[23] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton Van Den Henge. Explicit knowledge-based reasoning for visual question answering. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, page 1290–1296. AAAI Press, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[24] Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. FVQA: fact-based visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 40(10): 2413-2427, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[25] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.416, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[26] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6281-6290. Computer Vision Foundation / IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.5, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[27] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[28] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4995–5004, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27227" + } + ] +] \ No newline at end of file diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_origin.pdf b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..404b0d5207453ca105d56c34445f2bd1333be8df --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/e4ec05ff-3fa1-4ffb-b826-5f31038b1d51_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e03faa08065fb01c8fb02327233b5e808082b295b8f0cb67edb3ae697d6bd2a6 +size 3477781 diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/full.md b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d995ec947f7220d432d02502392de7781b3242c9 --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/full.md @@ -0,0 +1,368 @@ +# VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning + +Kang Chen Xiangqian Wu + +Faculty of Computing, Harbin Institute of Technology (HIT) + +Suzhou Research Institute of HIT + +chenkangcs@stu.hit.edu.cn, xqwu@hit.edu.cn + +# Abstract + +Achieving the optimal form of Visual Question Answering mandates a profound grasp of understanding, grounding, and reasoning within the intersecting domains of vision and language. Traditional VQA benchmarks have predominantly focused on simplistic tasks such as counting, visual attributes, and object detection, which do not necessitate intricate cross-modal information understanding and inference. Motivated by the need for a more comprehensive evaluation, we introduce a novel dataset comprising 23,781 questions derived from 10,124 image-text pairs. Specifically, the task of this dataset requires the model to align multimedia representations of the same entity to implement multi-hop reasoning between image and text and finally use natural language to answer the question. Furthermore, we evaluate this VTQA dataset, comparing the performance of both state-of-the-art VQA models and our proposed baseline model, the Key Entity Cross-Media Reasoning Network (KECMRN). The VTQA task poses formidable challenges for traditional VQA models, underscoring its intrinsic complexity. Conversely, KECMRN exhibits a modest improvement, signifying its potential in multimedia entity alignment and multi-step reasoning. Our analysis underscores the diversity, difficulty, and scale of the VTQA task compared to previous multimodal QA datasets. In conclusion, we anticipate that this dataset will serve as a pivotal resource for advancing and evaluating models proficient in multimedia entity alignment, multi-step reasoning, and open-ended answer generation. Our dataset and code is available at https://visual-text-qa.github.io/. + +# 1. Introduction + +A paramount objective in AI research is to endow systems with the capacity to comprehend the intricacies of the real + +world, akin to human understanding. Question Answering (QA) stands out as an effective task for evaluating the cognitive capabilities of AI systems. To answer questions, people need to extract information from multiple modalities, such as text, images and structured data like knowledge bases, graphs and tables. And furthermore, people need to align the information and do multi-step reasoning between different modalities. + +Although Visual Question Answering (VQA) [5] has been widely researched as a multimedia QA task, VQA models only extract information from image when answering questions and focus mainly on scene recognition, counting, color and other visual detection tasks, which do not require much logical reasoning or assignment between different modalities. Only recently, there are some attempts to introduce more modal information into VQA tasks. For example: (1) FVQA [24] and KBVQA [23] combine knowledge base (KB) with VQA task, which requires the ability of knowledge understanding and multi-step reasoning. But it is difficult to construct a comprehensive KB in real world, thus limiting its ability to address open-ended questions; (2) TextbookQA [10] and ScienceQA [14] use textbook as data sources, involving texts, images, tables and other multimodal information. But most images in these datasets are manually drawn schematic diagrams, and the questions are all in a multiple-choice setting, which are far from the real world. + +To address these issues, more recently, there appear some new datasets such as MultiModalQA [21] and MuMuQA [18], which involve reasoning across texts, images and tables. However, it's noteworthy that all these datasets are conducted with extracted QA. And in the MultiModalQA dataset, each image corresponds to a Wikipedia entity, simplifying image reasoning to ranking images based on these entities, which diminishes the necessity for crossmedia reasoning abilities. As for MuMuQA, although it requires grounding between image and text along with multi-hop reasoning, there are still some problems: (1) the data + +![](images/818423c22feecce763a7f581f2384e4cf7654af630e049baf37a7ae7d29ee0d3.jpg) +Image: +Figure 1. Example in our dataset with the question-answer pairs and their corresponding image and text. Different representations of the same object in text and image are identified with the same color. For example, 'Elena' in the text and the object bounding box corresponding to 'Elena' in the image are marked red. + +# Chinese Annotation: + +公司举行了一场饭桌会议，顾名思义就是一边吃饭一边商讨事宜。爱莲娜的金色头发果然是人群中最显眼的一个，乌尔蒂盖拉穿着橘黄色开衫，正在埋头切一块牛排；赛特温穿着深蓝色衬衫，胸前挂着会议牌，他的左手正抬起来准备去拿酒杯。寒特温的手左边，卡维利耶和阿吉姆正严肃地讨论下个月的预算案。卡维利耶一头褐色齐肩短发，干净利落。 + +Q1：爱莲娜穿着什么类型的上衣？ + +A1：西装外套 + +Q2：乌尔蒂盖拉的右手边是谁？ + +A2：卡维利耶 + +# English Translation: + +The company held a table meeting, which, as the name implies, is to discuss matters while eating. Elena's blond hair is indeed the most conspicuous one in the crowd, and Urigera, wearing an orange cardigan, is burying her head in cutting a steak; Setwin was wearing a dark blue shirt, with a meeting card hanging on his chest, and his left hand was lifting up to get a glass of wine. On the left of Setwin, Kavelier and Agum are seriously discussing the budget for next month. Kavelier has short brown shoulder-length hair, clean and clean. + +Q1: What type of blouse does Elena wear? + +A1: Suit + +Q2: Who is on the right-hand side of Urtigera? + +A2: Kavelier + +is all from news, resulting in most questions related to human beings; (2) questions all follow a specific pattern: first perform image entity grounding and then find the answer in the news body text; (3) there is only 1384 human-curated examples and the training data is automatically generated, which leads to poor quality and difficult to train with. Current multimedia benchmarks are still far from the real-world QA scene and cannot measure the multimedia understanding ability of AI system well. + +In order to address these gaps, we propose a novel task named 'Visual Text Question Answering' (VTQA), accompanied by a dedicated dataset, which includes only questions that require multi-hop reasoning through both images and text. All the annotations in VTQA dataset are first marked in Chinese, and then translated into English. + +To answer VTQA questions, the proposed model needs to: (1) learn to identify entities in image and text referred to the question, (2) align multimedia representations of the same entity, and (3) conduct multi-step reasoning between text and image and output open-ended answer. The VTQA dataset consists of 10,124 image-text pairs and 23,781 questions. The images are real images from MSCOCO dataset [12], containing a variety of entities. The annotators are required to first annotate relevant text according to the image, and then ask questions based on the image-text pair, and finally answer the question open-ended. To ensure textual richness, we enforce a minimum text length requirement of over 100 Chinese words. Additionally, we systematically exclude questions that can be answered solely based on either the image or the text, thereby guaranteeing the complexity of the questions. + +Moreover, we conduct an evaluation on the VTQA dataset, employing both state-of-the-art VQA models and our newly proposed baseline model. The baseline model we propose is called Key Entity Cross-Media Reasoning Network (KECMRN), which follows the general paradigm for answering questions on this dataset by iteratively performing key entity extraction and alignment, and cross-modal multi-step reasoning to answer questions. Evaluation re + +sults on the VTQA dataset show that existing state-of-the-art VQA models struggle to achieve satisfactory performance on this dataset, which illustrates the challenges and potential of this dataset for cross-modal question answering tasks. + +The contributions of this work can be summarized as follows: (1) we propose a new cross-modal QA dataset named VTQA. Information diversity, multimedia multi-step reasoning and open-ended answer make our dataset more challenging than the existing datasets; (2) we benchmark the state-of-the-art VQA model on our new dataset and show the performance of these models degrades drastically; (3) we propose a baseline that is capable of multimedia entity alignment and multi-step reasoning. + +# 2. Related Work + +Visual Question Answering (VQA) aims to answer a natural language question based on an image, which requires model to understand and reason in the vision-language joint space. Several datasets have been proposed in the past few years, such as DAQUAR [15], FM-IQA [4], VQA [2, 5], COCOQA [19], Visual7W [28], Visual Genome [11], GQA [8], OKVQA [16], A-OKVQA [20], VizWizQA [6] and so on. However, the natural language questions in these datasets can be regarded as instructions to guide the model to complete visual tasks such as object detection, scene recognition, counting, etc. And these questions often only involve the surface information of the image, without considering the complexities between different modalities. Understanding of the reasoning process. In order to improve the multimodal perception ability of the model, we need to design more challenging problems, requiring the model to combine the features of images and texts for cross-modal knowledge fusion and logical reasoning. + +Recently, some datasets attempt to introduce more modal information into VQA tasks, such as FVQA [24], KB-VQA [23], TextbookQA [10], and ScienceQA [14]. But the information used in these datasets includes manually col + +lected databases or manually drawn schematic diagrams in textbooks, which are far from real-world QA scenarios. The VTQA we propose is different from these, using the semantically rich real images of the COCO dataset and human-annotated text passages, which can be better generalized to real-world QA scenarios. + +The most similar to our dataset is MuMuQA [18]. The MuMuQA dataset is a news-based QA dataset whose images and questions mainly involve people and events. This dataset has only a few human-labeled data, and most of the data is generated by automated methods, so there may be a lot of noise and errors. In contrast, our VTQA dataset covers a variety of genres and topics, and all texts, questions, and answers are annotated by humans, ensuring the high quality and accuracy of the data. Our VTQA dataset has 23781 items, which is richer and more diverse than MuMuQA dataset. + +# 3. VTQA Dataset + +In this section, we present the details of our Visual Text Question Answering (VTQA) dataset. We first introduce the VTQA task of our dataset (Section 3.1) and then detail how the dataset was constructed (Section 3.2), along with the statistical properties of our dataset. + +# 3.1. VTQA Task + +As illustrated in Fig. 1, given an image-text pair and a question, a system is required to answer the question by natural language. Importantly, the system needs to: (1) analyze the question and find out the key entities, (2) align the key entities between image and text, and (3) generate the answer according to the question and aligned entities. For example, in Fig. 1, the key entity of Q1 is "Elena". By referencing the descriptor "gold hair" in the text, it is discerned that the second person from the right in the image is "Elena". Subsequently, the answer "suit" is generated based on the visual information. As for Q2, which is a more complex question, the previous steps need to be repeated several times to answer it. + +# 3.2. Dataset Construction + +To promote the progress of this open-ended multimedia multi-hop VTQA task, we collect a new dataset. Our dataset consists of 10124 image-text pairs and 23,781 questions. We collect data through newly developed annotation interface. + +In the first round of labeling, we present images from COCO and corresponding image descriptions, as well as the object detection labels, to the annotators. The annotators are required to generate a text that exceeds 100 words, which should involve the object in the image and contain information that is not included in the image description. + +To ensure compliance with annotation requirements, annotators can choose to skip some images (12198 images were ultimately skipped, 9830 images were selected, and some of them were selected more than once). The annotators then come up with questions based on the image-text pair and the annotation process requires that the questions cannot be answered only by image or text. Each image-text pair is labeled with 1-4 questions. + +In a second round of labeling, different annotators were asked to determine whether the question could be answered solely by relying on either images or text within the corresponding text-question pair or image-question pair. Through this step, we filtered down to 23781 questions from a pool of 28919 questions. The questions that pass this step will be labeled with answers and answer categories. We set three categories for answers: (1) $YN$ means yes-or-no answer, (2) $E$ means that the answer is extracted from the text, and (3) $G$ means that the answer is generated from the text-image pair. And the annotators need to label 'yes' or 'no' in English for the yes-or-no answer, since there are too many words to express 'yes' or 'no' in Chinese, for example, '可以' and '是的' both mean 'yes'. + +We randomly split the dataset into training, validation, test-dev and test splits and each image will only appear in one split. Each split has 11312, 1245, 2189, 9035 samples, respectively. + +# 3.3. Dataset Analysis + +Table 1 shows a comparison of VTQA and other VQA datasets. As shown in the table, VTQA is much larger than most other datasets, especially with a much larger number of multimodal questions than existing multimodal QA datasets. The text length of the VTQA dataset is shorter than that of MuMuQA, but still much longer than the length of the questions in all datasets, and the texts in VTQA are manually generated, containing more types of entities than existing texts extracted from textbooks, wikis, and news. + +Images in VTQA dataset are from MSCOCO dataset, which contains multiple objects and rich contextual information. Fig. 2 presents the top-10 categories distribution of the images used in this dataset. Unlike MuMuQA, in which most images are related to people, images in our dataset contain more kinds of objects. + +The texts and questions length statistics are shown in Fig. 3. It is evident that most texts range from 300 to 600 and most questions range from 9 to 21. Obviously, compared with the previous VQA datasets (usually less than 20 words), our dataset also puts forward higher requirements for text understanding to extract information from the long text. + +
#I#Q#MMQContextsAvgQAvgTAnswer TypeImage SourceText Source
VQA200K1.1M-I6.1-Open/MCCOCO-
FVQA58265826-I+KB9.5-OpenCOCO-
KBVQA24022402-I+KB6.8-OpenCOCO-
TextbookQA34552626012567I+T9.2-MCTextbookTextbook
ScienceQA10332212086532I+T12.1-MCTextbookTextbook
MultimodalQA57058299188240I+T18.266.2OpenWikipediaWikipedia
MuMuQA*138413841384I+T11.8633.5OpenNewsNews
VTQA98302378123781I+T10238.4OpenCOCOannotated
+ +Table 1. Statistics for VTQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, #MMQ: number of multimodal questions, AvgQ: average question length, AvgT: average text length. *We only count the manually annotated parts of the MuMuQA dataset. + +![](images/77fda0048ce186dc869078aa41d20b3c5c3614e553d8fad2fa4d7170ea59ad38.jpg) +Figure 2. Top-10 categories distribution of the images used in our dataset. + +![](images/4bbfce12d0bc15be359464fba6b5bddd858aeeca43429ac99273c77d99cb48c2.jpg) +Figure 3. Percentage of questions and texts with different Chinese word lengths. + +![](images/f405c200360739f022bd0be75bf43be3eea20ddc55f1782d4c8fcdc2bc17a1f4.jpg) + +# 4. VTQA Model + +In this section, we describe a competitive baseline method for evaluation on our benchmark, which is called Key Entity Cross-Media Reasoning Network (KECMRN). Before presenting the KECMRN, we first introduce its basic component, the KECMR layer. The KECMR layer is a modular composition which consists of one Key Entity Extract (KEE) layer and multiple Cross-Media Reason (CMR) layers. The KEE layer and CMR layer are composed of attention unit and feed-forward unit from [22]. Then we use the KECMR module with other layers to combine our KECMRN. An overview flowchart of KECMRN is shown in Fig. 4. + +# 4.1. Attention and Feed-Forward Units + +As shown in [22], the combination of attention unit and feed-forward unit has strong representational and learning ability. We use the same settings to construct our units. + +Multi-Head Scaled Dot-Product Attention Unit. Given a query $q \in R^{1 \times d}$ , n key-value pairs (packed into a key matrix $K \in R^{n \times d}$ and a value matrix $V \in R^{n \times d}$ ), the attended feature is obtained as: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V \tag {1} +$$ + +Multi-head attention divides the input into $h$ parts and makes single attention on each part. The attended feature is given by: + +$$ +M H (Q, K, V) = \operatorname {C o n c a t} \left(\text {h e a d} _ {1}, \dots , \text {h e a d} _ {h}\right) W ^ {O} \tag {2} +$$ + +$$ +h e a d _ {i} = \text {A t t e n t i o n} \left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\right) \tag {3} +$$ + +where $W_{i}^{Q}, W_{i}^{Q}, W_{i}^{Q} \in R^{d \times d_{h}}$ and $W^{O} \in R^{(h \times d_{h}) \times d}$ are the projection matrices. $d_{h}$ is the dimension of the output features from each head. + +Feed-Forward Unit. The feedforward unit takes the output features of the multi-head attention layer, and further + +![](images/9bc01fc458078bcbca0521cf05a108a8e1f7658670a217d6675859d8d91f088f.jpg) +Figure 4. Overall flowchart of the Key Entity Cross-Media Reasoning Network. + +![](images/b8347d993c440a8dbf60409e029071f10310c96d670bae4b9e55e1595026f52e.jpg) +Figure 5. Key Entity Extract Layer. + +transforms them through two fully-connected layers with ReLU activation in between. + +$$ +F F N (x) = \max \left(0, x W _ {1} + b _ {1}\right) W _ {2} + b _ {2} \tag {4} +$$ + +While the linear transformations are the same across different positions, they use different parameters from layer to layer. + +# 4.2. Key Entity Extract Layer + +As stated in Section 3.1, the first step to answer VTQA questions is to find out the key entity according to the questions. We compose the attention unit and feed-forward unit to integrate question information into the text and the image respectively. Then we apply a fully-connected layer to the question-aware text/image features to get the score for each feature. Finally, we extract the top-k features as key entities. + +The complete KKE layer is shown in the Fig. 5. Mathematical, given input text features $T$ , image features $I$ and question features $Q$ , the KEE layer can be formulated by (the processing of feature $I$ is aligned with that of feature + +![](images/2cede4c626558eb454f68ee3c4f238e30d538d84b175a51073cc46532c4bdd75.jpg) +Figure 6. Cross-Media Reason Layer. 'G' means gather and 'S' means scatter. + +T): + +$$ +Q = F F N (M H (Q, Q, Q)) \tag {5} +$$ + +$$ +T = F F N (M H (M H (T, T, T), Q, Q)) \tag {6} +$$ + +$$ +\operatorname {s c o r e} _ {T} = W _ {T} T + b _ {T} \tag {7} +$$ + +The final score will be used to sort the image/text features and the top-k features are considered as image/text key entities. To unify the expression, we treat all the question features as key entities. + +# 4.3. Cross-Media Reason Layer + +The CMR layer is designed for multi-step reasoning across medias. As show in Fig. 6, we first gather the key entities as $S \in R^{l_k \times d}$ from input features, where $l_k$ is the total number of all key entities. Then the key entities pass through self-attention, global cross-attention based on original features and feed-forward unit. Finally, the key entities are scattered to the input features. As the requirement for multi-step reasoning, we inserted multiple CMR layers in a KECMR module. The CMR layer can be formulated by: + +$$ +S = \operatorname {g a t h e r} ([ T, I, Q ]) \tag {8} +$$ + +$$ +S = F F N (M H (M H (S, S, S), [ T, I, Q ], [ T, I, Q ])) \tag {9} +$$ + +$$ +[ T, I, Q ] = \operatorname {s c a t t e r} (S, [ T, I, Q ]) \tag {10} +$$ + +# 4.4. Key Entity Cross-Media Reason Network + +We use the same network framework as [26] but replace MCA layer with our KECMR module and add an extra text stream. As show in Figure 4, The input image is represented as a set of regional visual features in a bottom-up manner [1]. The input question and text are transformed to features by passing through a one-layer LSTM network [7]. Then we use our KECMR module several times to extract key entities and conduct multi-steps cross-media reasoning. Finally, we use the attention reduce layer to fuse the multimedia features and project the fused feature into the answer probability distribution. + +The attention reduce layer can be formulated by: + +$$ +M L P (x) = W _ {2} \left(\max \left(0, x W _ {1} + b _ {1}\right)\right) + b _ {2} \tag {11} +$$ + +$$ +\operatorname {A t t R e} (X) = \operatorname {s u m} (\operatorname {s o f t m a x} (M L P (X)) X) \tag {12} +$$ + +We can obtain the attended image feature $I_{a}$ by passing image feature I through attention reduce layer, and the same applies to text feature T and question feature Q. Then we fuse the multimedia features by: + +$$ +z = W _ {i} I _ {a} + W _ {t} T _ {a} + W _ {q} Q _ {a} \tag {13} +$$ + +The fused feature $\mathbf{z}$ is projected into a vector $s \in R^{N}$ followed by a sigmoid function, where $\mathbf{N}$ is the number of the answers in the training set. We use binary cross-entropy (BCE) as the loss function to train an N-way classifier on top of the fused feature $\mathbf{z}$ . + +# 5. Experiments + +In this section, we evaluate the current state-of-the-art VQA model and our KECMRN model, and provide results of them in VTQA dataset. + +# 5.1. Evaluation Metrics + +As the answers divided into three types, we use different metrics for distinct types of answers. + +Exact match (EM). This metric measures the percentage of predictions that match the ground truth answer exactly and will be used in all types of answers. + +(Macro-averaged) F1 score (F1). This metric measures the average overlap between the prediction and ground truth answer. We treat the prediction and ground truth as bags of tokens and compute their F1. This metric will be used for the $E$ and $G$ types of answers. + +YN accuracy (YNAcc). This metric is only used for the YN type of answers. The answer will be transformed into 'yes' or 'no' by a pre-defined yes-or-no dictionary, which is manually collected based on the answers in our dataset. Then we calculate the accuracy just based on the yes-or-no answer. + +# 5.2. Implementation Details + +KECMRN: the hyper-parameters of our model used in the experiments are as follows. The dimensionality of input image features, input question features, input text features and fused multimodal features are 2,048, 1024, 1024, and 2,048, respectively. The latent dimensionality $d$ in the multi-head attention is 1024, the number of heads $h$ is set to 8. The number of CMR layer in each KECMR module and the number of KECMR module are 2 and 2. The number of key entities $k$ is set to 16. We train our model on the train set for 13 epochs and evaluate our model on test set. The Faster-RCNN used for image feature extraction is pretrained and frozen, while all other modules are trained from scratch. + +GPT4V [17]/LLaVA [13]/MMICL [27]: popular multimodal large language models (MLLM). These models demonstrated impressive multimodel chat abilities and exceptional performance across numerous intricate visual reasoning datasets, particularly excelling in zero-shot evaluations. Notably, LLaVA achieved an impressive $80.0\%$ accuracy, while GPT4V attained $77.2\%$ and MMICL attained $70.3\%$ on the VQAv2 dataset, underscoring their remarkable capabilities in the realm of multimodal understanding and visual question answering. We follow the same evaluation strategy as on the VQAv2 dataset, with the sole modification of adapting the prompt to 'Search for answers in the context above, do not use additional knowledge beyond the text, and answer the question using a single word or phrase'. This adjustment facilitated enhanced performance on VTQA datasets. + +BEiT3 [25]/VLMo [3]: visual language models directly pre-trained on image-text pairs, offering versatility in addressing a spectrum of visual-language tasks through finetuning on downstream tasks. Notably, on the VQAv2 dataset, BEiT3-base achieved an accuracy of $77.65\%$ , while VLMo-base demonstrated a commendable accuracy of $76.6\%$ . To evaluate the transferability of BEiT3 and VLMo on our VTQA task, we follow the same fine-tuning strategy as on the VQAv2 dataset, but we concatenate the text and questions into a single paragraph of text and increase the maximum number of training epochs from 10 to 50 to obtain better results, since our dataset is relatively small compared to VQAv2. + +MCAN-region [26]/MCAN-grid [9]: models based on MCAN [26] but using different image features. They achieve SOTA results on VQAv2 dataset (70.93% for MCAN-region and 72.71% for MCAN-grid) among non-pretrained visual-language models(non-PVLM), which indicates that these models were not pretrained on visual-language data, although individual pretraining on either visual or text data is allowed. We concatenate the text and questions and increase the maximum number of input words to train MCAN in VTQA task. + +Trans-CA/Trans-SA: models based on Trans + +
MethodsEMYN-ACCE-F1G-F1
English Version
LLaVA39.9065.0560.0626.20
GPT4V48.9375.5970.2448.37
MMICL38.7662.9062.3245.96
BEiT3-base41.0256.5150.1828.14
VLMO-base51.8565.4664.5942.74
MCAN-region56.0972.0667.0340.52
MCAN-grid55.8870.7365.3939.19
Trans-CA57.3873.5769.0244.14
Trans-SA54.3967.8965.9541.11
KECMRN(ours)57.9574.9968.9144.25
Chinese Version
MCAN-region49.7871.2659.9445.74
MCAN-grid50.2274.1759.0245.59
Trans-CA49.8774.2759.6347.45
Trans-SA48.9973.2958.2646.31
KECMRN(ours)51.3277.5960.5251.11
+ +former [22]. Trans-CA treat image, text and question features as independent input and perform Self Attention and Question Aware Cross Attention in each layer. The layers in Trans-CA resemble those of the KEE module, excluding the "score&sort" step. Conversely, Trans-SA adopts a different approach by concatenating all image, text, and question features and only applying Self Attention in each layer. The layers in Trans-SA resemble those of the CMR module, excluding the "Global Cross-Attention" step. + +# 5.3. Results and Analysis + +Table 2 provides the results of various baselines on the test set of our VTQA benchmark. VLMs, including MMICL, LLaVA, BEiT3, and VLMo, primarily pretrained on English datasets, are exclusively evaluated in the English version. + +While models such as GPT4V, MMICL, and LLaVA demonstrate superior performance across a variety of visual-language datasets in zero-shot evaluation, they encounter a significant drop in performance when evaluated on the VTQA dataset. This is partly because the model lacks exposure to our training data during training, unlike the VQA dataset included in the model's VL fine-tuning data. The decline also results from the heightened complexity of our VTQA task compared to VQA, posing new challenges to the model's image-text alignment and inference capabilities. + +BEiT3 and VLMo, pretrained directly on image-text pairs, achieve impressive results on various multimodal datasets through fine-tuning on downstream tasks. Despite + +Table 2. Results of our KECMRN on test set compared with other multimodal models. + +
MethodsEMYN-ACCE-F1G-F1
Trans(1)55.4572.4365.3542.27
Trans(3)55.0770.9865.743.58
Trans(6)54.0769.4364.8343.91
LSTM57.9574.9968.9144.25
+ +Table 3. Results of different text encoders based on KECMRN. + +fine-tuning on VTQA training data, BEiT3 and VLMo still exhibit a substantial performance decrease. This is partly attributed to the limitation that the pretraining data for BEiT3 and VLMo only includes image-text pairs with relatively short texts, while the texts in the VTQA dataset are longer, averaging 284 words in English version. This, once again, underscores the challenges inherent in the VTQA task. + +While MCAN models outperform the pre-trained models, a notable decrease is observed. In addition, Trans-CA exceeded the results of all VQA models, while the results of Trans-SA are relatively poor. This highlights the necessity for a more intricate independent processing of text, image, and question features in our VTQA dataset. Our KECMRN achieves the best result with $57.95\%$ on the English version and $51.32\%$ in Chinese version. + +Fig. 7 illustrates the outcomes of attention visualization across different models. The attention weights from the 'Attention Reduce' module were employed to colorize image regions and organize the Top Words. Notably, our KECMRN model exhibits precise identification of Zhang Yanqi on the left side, engaged in ball play. This accurate recognition is based on key phrases such as sports balls and dribbling skills, leading to a correct inference for the associated question ('Yes'). In contrast, although MCAN also directs attention to the person on the left, the absence of Key-Entity Extract (KEE) causes attention to be excessively dispersed. Consequently, there is an inaccurate localization of the person's upper garment, resulting in an incorrect answer. Furthermore, Trans-SA mistakenly focuses on the person on the right, while Trans-CA struggles to discern whether Zhang Yanqi is positioned on the left or right, both yielding incorrect answers. The attention visualization results underscore the superior precision of our KECMRN model in focusing on the correct visual regions and words. + +# 5.4. Ablation Study + +We conducted a number of ablation experiments to investigate the reasons for the effectiveness of KECMR. All the reported results are evaluated on English version. + +Text Encoders: Table 3 provides an overview of the results obtained using different text encoders in conjunction with KECMRN. In this table, Trans(k) denotes a k-layer transformer encoder model. Contrary to expectations, the performance of deeper transformer encoder models tends to be inferior, with LSTM outperforming all transformer en + +![](images/39c36cb5a72c3e3bf4d9007f568fc99c6fc196eb84974c4e47a160a3afb6edc5.jpg) +Text: + +Song Shuai and Zhang Jiaqi are two local sports ball enthusiasts whose video about sports ball on the short video platform caught the attention of a club in a nearby city. Today they were invited to the city where the club is located for a technical exchange, and now Song Shuai is talking to the head of the club, and Zhang Jiaqi is too obsessed with sports balls on the road and still does not forget to practice dribbling skills. + +Question: + +Is Zhang Jiaqi wearing a blue top? + +Answer: + +Yes + +![](images/c64ae40c2457a6d4a0a8113215c982f8620aebae8bf75f42fa9d8bc9bca25502.jpg) + +Top Words: + +balls, Jiaqi, enthusiasts, Zhang, road, skills, obsessed, dribbling, sports, forget Predict Answer: Yes + +![](images/b0074f83903e06aef247d91cb0e37d8d18c8f0d1f3fb57dbf1b4f28bda46b286.jpg) +KECMRN + +Top Words: + +sports, dribbling, Jiaqi, club, Shuai, Zhang, video, caught, platform, talking Predict Answer: No + +Trans-SA + +![](images/ea7925dc6cd291f2c8dc6a3a4a014aaac157aae22f482d51db66db1ea0bd8c6a.jpg) + +Top Words: + +skills, dribbling, enthusiasts, nearby, road, balls, Zhang, club, still, sports Predict Answer: No + +![](images/13a5d0a27bd4e1632a8daf46b6ff066113c94c816376ea668c4eaab3aa39dd1e.jpg) +MCAN +Figure 7. An example of attention visualization for different models. + +Top Words: + +Jiaqi, Song, exchange, dribbling, Zhang, obsessed, head, forget, short, Shuai Predict Answer: No + +Trans-CA + +
MethodsEMYN-ACCE-F1G-F1
Trans-CA57.3873.5769.0244.14
CA+SA57.1773.5768.9642.88
KEE+SA57.6374.4269.143.06
KEE+CMR57.9574.9968.9144.25
+ +Table 4. Results of the KECMRN model with different module combinations. + +coder models. This phenomenon is attributed to both insufficient training data and the intricate structure of transformer models. + +KEE/CA and CMR/SA: Table 4 displays the results of the KECMRN model with different module combinations. In this table, CA+SA denotes the alternate use of CA and SA layers, similar to our KECMRN design. KEE+SA follows a similar pattern. Compared to Trans-CA, replacing some CA layers with SA layers results in a slight performance decrease (↓0.21). Conversely, the inclusion of the KEE layer improves performance (↑0.46) by concentrating information on key entities. Additionally, the CMR layer further optimizes features related to key entities, leading to enhanced results (↑0.32). + +# 6. Conclusion + +We propose a novel cross-modal question answering dataset, VTQA, which necessitates models to acquire pertinent information from both text and image sources and perform complex cross-modal reasoning to answer questions. Our experimental findings reveal that even state-of-the-art pretrained models fall short of achieving satisfactory performance on VTQA, underscoring the challenges and potential inherent in the dataset for cross-modal question-answering tasks. Looking ahead, we aim to augment the dataset's size and coverage and provide additional reasoning annotations in future work to more effectively assess and catalyze the development of cross-modal reasoning models. Simultaneously, we aspire to explore more fitting evaluation metrics to surmount the limitations of EM metrics and more equitably gauge the model's question-answering proficiency. + +# 7. Acknowledgment + +This work was supported in part by the Natural Science Foundation of China under Grant 62073105, U23A20389 and 62306095, by the Natural Science Foundation of Heilongjiang Province of China under Grant ZD2022F002, and by the Heilongjiang Touyan Innovation Team Program. + +# References + +[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086. Computer Vision Foundation / IEEE Computer Society, 2018. 6 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 2425-2433, 2015. 2 +[3] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems, pages 32897-32912. Curran Associates, Inc., 2022. 6 +[4] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a machine? dataset and methods for multilingual image question. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2 +[5] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 6325-6334. IEEE Computer Society, 2017. 1, 2 +[6] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2 +[7] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 6 +[8] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[9] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10264-10273, 2020. 6 +[10] Aniruddha Kembhavi, Min Joon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 5376-5384. IEEE Computer Society, 2017. 1, 2 +[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalan- + +tidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vision, 123(1):32-73, 2017. 2 +[12] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 2 +[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6 +[14] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 1, 2 +[15] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1682-1690, 2014. 2 +[16] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[17] OpenAI, Josh Achiam, Steven Adler, et al. Gpt-4 technical report, 2024. 6 +[18] Revanth Gangi Reddy, Xilin Rui, Manling Li, Xudong Lin, Haoyang Wen, Jaemin Cho, Lifu Huang, Mohit Bansal, Avirup Sil, Shih-Fu Chang, Alexander G. Schwing, and Heng Ji. Mumuqa: Multimedia multi-hop news question answering via cross-media knowledge extraction and grounding. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 11200-11208. AAAI Press, 2022. 1, 3 +[19] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2 +[20] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision - ECCV 2022, pages 146-162, Cham, 2022. Springer Nature Switzerland. 2 +[21] Alon Talmor, Ori Yoran, Amnon Catav, Dan Lahav, Yizhong Wang, Akari Asai, Gabriel Ilharco, Hannaneh Hajishirzi, and Jonathan Berant. Multimodalqa: complex question answering over text, tables and images. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021. 1 + +[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 4, 7 +[23] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton Van Den Henge. Explicit knowledge-based reasoning for visual question answering. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, page 1290–1296. AAAI Press, 2017. 1, 2 +[24] Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. FVQA: fact-based visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 40(10): 2413-2427, 2018. 1, 2 +[25] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 6 +[26] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6281-6290. Computer Vision Foundation / IEEE, 2019. 6 +[27] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 6 +[28] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4995–5004, 2016. 2 \ No newline at end of file diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/images.zip b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1666178c4632a6460f02803144533313370564fd --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c8ff3a07ff96a4293f20862d06b54ddac7669c1c8a9ad160dfd88438fc4c967 +size 526008 diff --git a/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/layout.json b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a21048b2bfe3967beade53fe4717429d2d4c4e0 --- /dev/null +++ b/2024/VTQA_ Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning/layout.json @@ -0,0 +1,7726 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 52, + 102, + 543, + 140 + ], + "type": "text", + "content": "VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 221, + 167, + 375, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 167, + 375, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 167, + 375, + 182 + ], + "type": "text", + "content": "Kang Chen Xiangqian Wu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 182, + 443, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 182, + 443, + 197 + ], + "spans": [ + { + "bbox": [ + 150, + 182, + 443, + 197 + ], + "type": "text", + "content": "Faculty of Computing, Harbin Institute of Technology (HIT)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 198, + 378, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 198, + 378, + 209 + ], + "spans": [ + { + "bbox": [ + 216, + 198, + 378, + 209 + ], + "type": "text", + "content": "Suzhou Research Institute of HIT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "type": "text", + "content": "chenkangcs@stu.hit.edu.cn, xqwu@hit.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 277, + 290, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 290, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 290, + 625 + ], + "type": "text", + "content": "Achieving the optimal form of Visual Question Answering mandates a profound grasp of understanding, grounding, and reasoning within the intersecting domains of vision and language. Traditional VQA benchmarks have predominantly focused on simplistic tasks such as counting, visual attributes, and object detection, which do not necessitate intricate cross-modal information understanding and inference. Motivated by the need for a more comprehensive evaluation, we introduce a novel dataset comprising 23,781 questions derived from 10,124 image-text pairs. Specifically, the task of this dataset requires the model to align multimedia representations of the same entity to implement multi-hop reasoning between image and text and finally use natural language to answer the question. Furthermore, we evaluate this VTQA dataset, comparing the performance of both state-of-the-art VQA models and our proposed baseline model, the Key Entity Cross-Media Reasoning Network (KECMRN). The VTQA task poses formidable challenges for traditional VQA models, underscoring its intrinsic complexity. Conversely, KECMRN exhibits a modest improvement, signifying its potential in multimedia entity alignment and multi-step reasoning. Our analysis underscores the diversity, difficulty, and scale of the VTQA task compared to previous multimodal QA datasets. In conclusion, we anticipate that this dataset will serve as a pivotal resource for advancing and evaluating models proficient in multimedia entity alignment, multi-step reasoning, and open-ended answer generation. Our dataset and code is available at https://visual-text-qa.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 648, + 128, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 128, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 128, + 661 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "content": "A paramount objective in AI research is to endow systems with the capacity to comprehend the intricacies of the real" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 253, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 546, + 348 + ], + "type": "text", + "content": "world, akin to human understanding. Question Answering (QA) stands out as an effective task for evaluating the cognitive capabilities of AI systems. To answer questions, people need to extract information from multiple modalities, such as text, images and structured data like knowledge bases, graphs and tables. And furthermore, people need to align the information and do multi-step reasoning between different modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 351, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 579 + ], + "type": "text", + "content": "Although Visual Question Answering (VQA) [5] has been widely researched as a multimedia QA task, VQA models only extract information from image when answering questions and focus mainly on scene recognition, counting, color and other visual detection tasks, which do not require much logical reasoning or assignment between different modalities. Only recently, there are some attempts to introduce more modal information into VQA tasks. For example: (1) FVQA [24] and KBVQA [23] combine knowledge base (KB) with VQA task, which requires the ability of knowledge understanding and multi-step reasoning. But it is difficult to construct a comprehensive KB in real world, thus limiting its ability to address open-ended questions; (2) TextbookQA [10] and ScienceQA [14] use textbook as data sources, involving texts, images, tables and other multimodal information. But most images in these datasets are manually drawn schematic diagrams, and the questions are all in a multiple-choice setting, which are far from the real world." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "To address these issues, more recently, there appear some new datasets such as MultiModalQA [21] and MuMuQA [18], which involve reasoning across texts, images and tables. However, it's noteworthy that all these datasets are conducted with extracted QA. And in the MultiModalQA dataset, each image corresponds to a Wikipedia entity, simplifying image reasoning to ranking images based on these entities, which diminishes the necessity for crossmedia reasoning abilities. As for MuMuQA, although it requires grounding between image and text along with multi-hop reasoning, there are still some problems: (1) the data" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "text", + "content": "* corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27218" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 83, + 240, + 178 + ], + "blocks": [ + { + "bbox": [ + 65, + 75, + 87, + 82 + ], + "lines": [ + { + "bbox": [ + 65, + 75, + 87, + 82 + ], + "spans": [ + { + "bbox": [ + 65, + 75, + 87, + 82 + ], + "type": "text", + "content": "Image:" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 83, + 240, + 178 + ], + "lines": [ + { + "bbox": [ + 65, + 83, + 240, + 178 + ], + "spans": [ + { + "bbox": [ + 65, + 83, + 240, + 178 + ], + "type": "image", + "image_path": "818423c22feecce763a7f581f2384e4cf7654af630e049baf37a7ae7d29ee0d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "type": "text", + "content": "Figure 1. Example in our dataset with the question-answer pairs and their corresponding image and text. Different representations of the same object in text and image are identified with the same color. For example, 'Elena' in the text and the object bounding box corresponding to 'Elena' in the image are marked red." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 250, + 74, + 312, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 74, + 312, + 81 + ], + "spans": [ + { + "bbox": [ + 250, + 74, + 312, + 81 + ], + "type": "text", + "content": "Chinese Annotation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 87, + 354, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 87, + 354, + 137 + ], + "spans": [ + { + "bbox": [ + 250, + 87, + 354, + 137 + ], + "type": "text", + "content": "公司举行了一场饭桌会议，顾名思义就是一边吃饭一边商讨事宜。爱莲娜的金色头发果然是人群中最显眼的一个，乌尔蒂盖拉穿着橘黄色开衫，正在埋头切一块牛排；赛特温穿着深蓝色衬衫，胸前挂着会议牌，他的左手正抬起来准备去拿酒杯。寒特温的手左边，卡维利耶和阿吉姆正严肃地讨论下个月的预算案。卡维利耶一头褐色齐肩短发，干净利落。" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 140, + 326, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 140, + 326, + 147 + ], + "spans": [ + { + "bbox": [ + 250, + 140, + 326, + 147 + ], + "type": "text", + "content": "Q1：爱莲娜穿着什么类型的上衣？" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 150, + 282, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 150, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 250, + 150, + 282, + 157 + ], + "type": "text", + "content": "A1：西装外套" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 250, + 161, + 318, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 161, + 318, + 167 + ], + "spans": [ + { + "bbox": [ + 250, + 161, + 318, + 167 + ], + "type": "text", + "content": "Q2：乌尔蒂盖拉的右手边是谁？" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 171, + 279, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 171, + 279, + 177 + ], + "spans": [ + { + "bbox": [ + 250, + 171, + 279, + 177 + ], + "type": "text", + "content": "A2：卡维利耶" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 74, + 432, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 74, + 432, + 82 + ], + "spans": [ + { + "bbox": [ + 372, + 74, + 432, + 82 + ], + "type": "text", + "content": "English Translation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 87, + 529, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 87, + 529, + 137 + ], + "spans": [ + { + "bbox": [ + 370, + 87, + 529, + 137 + ], + "type": "text", + "content": "The company held a table meeting, which, as the name implies, is to discuss matters while eating. Elena's blond hair is indeed the most conspicuous one in the crowd, and Urigera, wearing an orange cardigan, is burying her head in cutting a steak; Setwin was wearing a dark blue shirt, with a meeting card hanging on his chest, and his left hand was lifting up to get a glass of wine. On the left of Setwin, Kavelier and Agum are seriously discussing the budget for next month. Kavelier has short brown shoulder-length hair, clean and clean." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 371, + 140, + 465, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 140, + 465, + 147 + ], + "spans": [ + { + "bbox": [ + 371, + 140, + 465, + 147 + ], + "type": "text", + "content": "Q1: What type of blouse does Elena wear?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 150, + 389, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 150, + 389, + 156 + ], + "spans": [ + { + "bbox": [ + 372, + 150, + 389, + 156 + ], + "type": "text", + "content": "A1: Suit" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 161, + 475, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 161, + 475, + 167 + ], + "spans": [ + { + "bbox": [ + 372, + 161, + 475, + 167 + ], + "type": "text", + "content": "Q2: Who is on the right-hand side of Urtigera?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 171, + 398, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 171, + 398, + 177 + ], + "spans": [ + { + "bbox": [ + 372, + 171, + 398, + 177 + ], + "type": "text", + "content": "A2: Kavelier" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 243, + 286, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 351 + ], + "type": "text", + "content": "is all from news, resulting in most questions related to human beings; (2) questions all follow a specific pattern: first perform image entity grounding and then find the answer in the news body text; (3) there is only 1384 human-curated examples and the training data is automatically generated, which leads to poor quality and difficult to train with. Current multimedia benchmarks are still far from the real-world QA scene and cannot measure the multimedia understanding ability of AI system well." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 353, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 286, + 424 + ], + "type": "text", + "content": "In order to address these gaps, we propose a novel task named 'Visual Text Question Answering' (VTQA), accompanied by a dedicated dataset, which includes only questions that require multi-hop reasoning through both images and text. All the annotations in VTQA dataset are first marked in Chinese, and then translated into English." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 426, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 616 + ], + "type": "text", + "content": "To answer VTQA questions, the proposed model needs to: (1) learn to identify entities in image and text referred to the question, (2) align multimedia representations of the same entity, and (3) conduct multi-step reasoning between text and image and output open-ended answer. The VTQA dataset consists of 10,124 image-text pairs and 23,781 questions. The images are real images from MSCOCO dataset [12], containing a variety of entities. The annotators are required to first annotate relevant text according to the image, and then ask questions based on the image-text pair, and finally answer the question open-ended. To ensure textual richness, we enforce a minimum text length requirement of over 100 Chinese words. Additionally, we systematically exclude questions that can be answered solely based on either the image or the text, thereby guaranteeing the complexity of the questions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "Moreover, we conduct an evaluation on the VTQA dataset, employing both state-of-the-art VQA models and our newly proposed baseline model. The baseline model we propose is called Key Entity Cross-Media Reasoning Network (KECMRN), which follows the general paradigm for answering questions on this dataset by iteratively performing key entity extraction and alignment, and cross-modal multi-step reasoning to answer questions. Evaluation re" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "type": "text", + "content": "sults on the VTQA dataset show that existing state-of-the-art VQA models struggle to achieve satisfactory performance on this dataset, which illustrates the challenges and potential of this dataset for cross-modal question answering tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 304, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 413 + ], + "type": "text", + "content": "The contributions of this work can be summarized as follows: (1) we propose a new cross-modal QA dataset named VTQA. Information diversity, multimedia multi-step reasoning and open-ended answer make our dataset more challenging than the existing datasets; (2) we benchmark the state-of-the-art VQA model on our new dataset and show the performance of these models degrades drastically; (3) we propose a baseline that is capable of multimedia entity alignment and multi-step reasoning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 427, + 392, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 392, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 392, + 440 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 449, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 664 + ], + "type": "text", + "content": "Visual Question Answering (VQA) aims to answer a natural language question based on an image, which requires model to understand and reason in the vision-language joint space. Several datasets have been proposed in the past few years, such as DAQUAR [15], FM-IQA [4], VQA [2, 5], COCOQA [19], Visual7W [28], Visual Genome [11], GQA [8], OKVQA [16], A-OKVQA [20], VizWizQA [6] and so on. However, the natural language questions in these datasets can be regarded as instructions to guide the model to complete visual tasks such as object detection, scene recognition, counting, etc. And these questions often only involve the surface information of the image, without considering the complexities between different modalities. Understanding of the reasoning process. In order to improve the multimodal perception ability of the model, we need to design more challenging problems, requiring the model to combine the features of images and texts for cross-modal knowledge fusion and logical reasoning." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Recently, some datasets attempt to introduce more modal information into VQA tasks, such as FVQA [24], KB-VQA [23], TextbookQA [10], and ScienceQA [14]. But the information used in these datasets includes manually col" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27219" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "lected databases or manually drawn schematic diagrams in textbooks, which are far from real-world QA scenarios. The VTQA we propose is different from these, using the semantically rich real images of the COCO dataset and human-annotated text passages, which can be better generalized to real-world QA scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 276 + ], + "type": "text", + "content": "The most similar to our dataset is MuMuQA [18]. The MuMuQA dataset is a news-based QA dataset whose images and questions mainly involve people and events. This dataset has only a few human-labeled data, and most of the data is generated by automated methods, so there may be a lot of noise and errors. In contrast, our VTQA dataset covers a variety of genres and topics, and all texts, questions, and answers are annotated by humans, ensuring the high quality and accuracy of the data. Our VTQA dataset has 23781 items, which is richer and more diverse than MuMuQA dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 138, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 138, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 138, + 301 + ], + "type": "text", + "content": "3. VTQA Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 309, + 286, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 286, + 369 + ], + "type": "text", + "content": "In this section, we present the details of our Visual Text Question Answering (VTQA) dataset. We first introduce the VTQA task of our dataset (Section 3.1) and then detail how the dataset was constructed (Section 3.2), along with the statistical properties of our dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 378, + 126, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 126, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 126, + 391 + ], + "type": "text", + "content": "3.1. VTQA Task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 286, + 552 + ], + "type": "text", + "content": "As illustrated in Fig. 1, given an image-text pair and a question, a system is required to answer the question by natural language. Importantly, the system needs to: (1) analyze the question and find out the key entities, (2) align the key entities between image and text, and (3) generate the answer according to the question and aligned entities. For example, in Fig. 1, the key entity of Q1 is \"Elena\". By referencing the descriptor \"gold hair\" in the text, it is discerned that the second person from the right in the image is \"Elena\". Subsequently, the answer \"suit\" is generated based on the visual information. As for Q2, which is a more complex question, the previous steps need to be repeated several times to answer it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 168, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 168, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 168, + 574 + ], + "type": "text", + "content": "3.2. Dataset Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "type": "text", + "content": "To promote the progress of this open-ended multimedia multi-hop VTQA task, we collect a new dataset. Our dataset consists of 10124 image-text pairs and 23,781 questions. We collect data through newly developed annotation interface." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "In the first round of labeling, we present images from COCO and corresponding image descriptions, as well as the object detection labels, to the annotators. The annotators are required to generate a text that exceeds 100 words, which should involve the object in the image and contain information that is not included in the image description." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "To ensure compliance with annotation requirements, annotators can choose to skip some images (12198 images were ultimately skipped, 9830 images were selected, and some of them were selected more than once). The annotators then come up with questions based on the image-text pair and the annotation process requires that the questions cannot be answered only by image or text. Each image-text pair is labeled with 1-4 questions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "text", + "content": "In a second round of labeling, different annotators were asked to determine whether the question could be answered solely by relying on either images or text within the corresponding text-question pair or image-question pair. Through this step, we filtered down to 23781 questions from a pool of 28919 questions. The questions that pass this step will be labeled with answers and answer categories. We set three categories for answers: (1) " + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "inline_equation", + "content": "YN" + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "text", + "content": " means yes-or-no answer, (2) " + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "text", + "content": " means that the answer is extracted from the text, and (3) " + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 173, + 545, + 340 + ], + "type": "text", + "content": " means that the answer is generated from the text-image pair. And the annotators need to label 'yes' or 'no' in English for the yes-or-no answer, since there are too many words to express 'yes' or 'no' in Chinese, for example, '可以' and '是的' both mean 'yes'." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "type": "text", + "content": "We randomly split the dataset into training, validation, test-dev and test splits and each image will only appear in one split. Each split has 11312, 1245, 2189, 9035 samples, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 416, + 406, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 406, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 406, + 428 + ], + "type": "text", + "content": "3.3. Dataset Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "type": "text", + "content": "Table 1 shows a comparison of VTQA and other VQA datasets. As shown in the table, VTQA is much larger than most other datasets, especially with a much larger number of multimodal questions than existing multimodal QA datasets. The text length of the VTQA dataset is shorter than that of MuMuQA, but still much longer than the length of the questions in all datasets, and the texts in VTQA are manually generated, containing more types of entities than existing texts extracted from textbooks, wikis, and news." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 624 + ], + "type": "text", + "content": "Images in VTQA dataset are from MSCOCO dataset, which contains multiple objects and rich contextual information. Fig. 2 presents the top-10 categories distribution of the images used in this dataset. Unlike MuMuQA, in which most images are related to people, images in our dataset contain more kinds of objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "The texts and questions length statistics are shown in Fig. 3. It is evident that most texts range from 300 to 600 and most questions range from 9 to 21. Obviously, compared with the previous VQA datasets (usually less than 20 words), our dataset also puts forward higher requirements for text understanding to extract information from the long text." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27220" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 548, + 180 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 548, + 180 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 548, + 180 + ], + "type": "table", + "html": "
#I#Q#MMQContextsAvgQAvgTAnswer TypeImage SourceText Source
VQA200K1.1M-I6.1-Open/MCCOCO-
FVQA58265826-I+KB9.5-OpenCOCO-
KBVQA24022402-I+KB6.8-OpenCOCO-
TextbookQA34552626012567I+T9.2-MCTextbookTextbook
ScienceQA10332212086532I+T12.1-MCTextbookTextbook
MultimodalQA57058299188240I+T18.266.2OpenWikipediaWikipedia
MuMuQA*138413841384I+T11.8633.5OpenNewsNews
VTQA98302378123781I+T10238.4OpenCOCOannotated
", + "image_path": "d4878855d4c17c1bfbb74d012ec6ffb840f44537284815f78bfe1c67882e0ed0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 222 + ], + "type": "text", + "content": "Table 1. Statistics for VTQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, #MMQ: number of multimodal questions, AvgQ: average question length, AvgT: average text length. *We only count the manually annotated parts of the MuMuQA dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 55, + 259, + 266, + 417 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 266, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 266, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 266, + 417 + ], + "type": "image", + "image_path": "77fda0048ce186dc869078aa41d20b3c5c3614e553d8fad2fa4d7170ea59ad38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 430, + 288, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 452 + ], + "type": "text", + "content": "Figure 2. Top-10 categories distribution of the images used in our dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 495, + 164, + 639 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 164, + 639 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 164, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 164, + 639 + ], + "type": "image", + "image_path": "4bbfce12d0bc15be359464fba6b5bddd858aeeca43429ac99273c77d99cb48c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 649, + 287, + 672 + ], + "lines": [ + { + "bbox": [ + 46, + 649, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 287, + 672 + ], + "type": "text", + "content": "Figure 3. Percentage of questions and texts with different Chinese word lengths." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 496, + 278, + 639 + ], + "blocks": [ + { + "bbox": [ + 170, + 496, + 278, + 639 + ], + "lines": [ + { + "bbox": [ + 170, + 496, + 278, + 639 + ], + "spans": [ + { + "bbox": [ + 170, + 496, + 278, + 639 + ], + "type": "image", + "image_path": "f405c200360739f022bd0be75bf43be3eea20ddc55f1782d4c8fcdc2bc17a1f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 243, + 391, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 391, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 391, + 256 + ], + "type": "text", + "content": "4. VTQA Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 262, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 407 + ], + "type": "text", + "content": "In this section, we describe a competitive baseline method for evaluation on our benchmark, which is called Key Entity Cross-Media Reasoning Network (KECMRN). Before presenting the KECMRN, we first introduce its basic component, the KECMR layer. The KECMR layer is a modular composition which consists of one Key Entity Extract (KEE) layer and multiple Cross-Media Reason (CMR) layers. The KEE layer and CMR layer are composed of attention unit and feed-forward unit from [22]. Then we use the KECMR module with other layers to combine our KECMRN. An overview flowchart of KECMRN is shown in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 413, + 492, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 492, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 492, + 425 + ], + "type": "text", + "content": "4.1. Attention and Feed-Forward Units" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 431, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 467 + ], + "type": "text", + "content": "As shown in [22], the combination of attention unit and feed-forward unit has strong representational and learning ability. We use the same settings to construct our units." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "text", + "content": "Multi-Head Scaled Dot-Product Attention Unit. Given a query " + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "inline_equation", + "content": "q \\in R^{1 \\times d}" + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "text", + "content": ", n key-value pairs (packed into a key matrix " + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "inline_equation", + "content": "K \\in R^{n \\times d}" + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "text", + "content": " and a value matrix " + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "inline_equation", + "content": "V \\in R^{n \\times d}" + }, + { + "bbox": [ + 304, + 468, + 545, + 515 + ], + "type": "text", + "content": "), the attended feature is obtained as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 533, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 533, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 335, + 533, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {1}", + "image_path": "b16f6eb14965d52e8e60f0a04277a714a5a14b094887a243c459e449989e525c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": "Multi-head attention divides the input into " + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": " parts and makes single attention on each part. The attended feature is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 611, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 611, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 315, + 611, + 545, + 625 + ], + "type": "interline_equation", + "content": "M H (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {2}", + "image_path": "1732bdcae1ec164d3e571790f1f681669350392076d7a1aacff4ff9cd9494fd7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "type": "interline_equation", + "content": "h e a d _ {i} = \\text {A t t e n t i o n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\tag {3}", + "image_path": "3bc7d4b11bbd68a2c384e38c9e505fd9d21f7b582aec86399daadf1e9dd462ca.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}, W_{i}^{Q}, W_{i}^{Q} \\in R^{d \\times d_{h}}" + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "inline_equation", + "content": "W^{O} \\in R^{(h \\times d_{h}) \\times d}" + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "text", + "content": " are the projection matrices. " + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 304, + 651, + 545, + 689 + ], + "type": "text", + "content": " is the dimension of the output features from each head." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Feed-Forward Unit. The feedforward unit takes the output features of the multi-head attention layer, and further" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27221" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 72, + 497, + 222 + ], + "blocks": [ + { + "bbox": [ + 99, + 72, + 497, + 222 + ], + "lines": [ + { + "bbox": [ + 99, + 72, + 497, + 222 + ], + "spans": [ + { + "bbox": [ + 99, + 72, + 497, + 222 + ], + "type": "image", + "image_path": "9bc01fc458078bcbca0521cf05a108a8e1f7658670a217d6675859d8d91f088f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 230, + 441, + 243 + ], + "lines": [ + { + "bbox": [ + 151, + 230, + 441, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 230, + 441, + 243 + ], + "type": "text", + "content": "Figure 4. Overall flowchart of the Key Entity Cross-Media Reasoning Network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 258, + 276, + 406 + ], + "blocks": [ + { + "bbox": [ + 59, + 258, + 276, + 406 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 276, + 406 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 276, + 406 + ], + "type": "image", + "image_path": "b8347d993c440a8dbf60409e029071f10310c96d670bae4b9e55e1595026f52e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 417, + 232, + 428 + ], + "lines": [ + { + "bbox": [ + 102, + 417, + 232, + 428 + ], + "spans": [ + { + "bbox": [ + 102, + 417, + 232, + 428 + ], + "type": "text", + "content": "Figure 5. Key Entity Extract Layer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 450, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 473 + ], + "type": "text", + "content": "transforms them through two fully-connected layers with ReLU activation in between." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 487, + 287, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 487, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 83, + 487, + 287, + 500 + ], + "type": "interline_equation", + "content": "F F N (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2} \\tag {4}", + "image_path": "5632131beae7aad9d08f047352cbe2be40ee258173cf6c00046c61c0985a75e9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 506, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 543 + ], + "type": "text", + "content": "While the linear transformations are the same across different positions, they use different parameters from layer to layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "type": "text", + "content": "4.2. Key Entity Extract Layer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 663 + ], + "type": "text", + "content": "As stated in Section 3.1, the first step to answer VTQA questions is to find out the key entity according to the questions. We compose the attention unit and feed-forward unit to integrate question information into the text and the image respectively. Then we apply a fully-connected layer to the question-aware text/image features to get the score for each feature. Finally, we extract the top-k features as key entities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The complete KKE layer is shown in the Fig. 5. Mathematical, given input text features " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", image features " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " and question features " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", the KEE layer can be formulated by (the processing of feature " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is aligned with that of feature" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 259, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 310, + 259, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 310, + 259, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 310, + 259, + 545, + 344 + ], + "type": "image", + "image_path": "2cede4c626558eb454f68ee3c4f238e30d538d84b175a51073cc46532c4bdd75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 351, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 351, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 544, + 373 + ], + "type": "text", + "content": "Figure 6. Cross-Media Reason Layer. 'G' means gather and 'S' means scatter." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 396, + 321, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 321, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 321, + 406 + ], + "type": "text", + "content": "T):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 418, + 545, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 418, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 367, + 418, + 545, + 432 + ], + "type": "interline_equation", + "content": "Q = F F N (M H (Q, Q, Q)) \\tag {5}", + "image_path": "8d64e4e43bdd06f0335c7d56e2c2efccbd43e92816b9533da910525d42334ff5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 436, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 436, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 342, + 436, + 545, + 449 + ], + "type": "interline_equation", + "content": "T = F F N (M H (M H (T, T, T), Q, Q)) \\tag {6}", + "image_path": "a49dfed42e8002b0ff53feece346ad7527b56173840521cd9eec072893ffef85.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 455, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 455, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 380, + 455, + 545, + 467 + ], + "type": "interline_equation", + "content": "\\operatorname {s c o r e} _ {T} = W _ {T} T + b _ {T} \\tag {7}", + "image_path": "adb093ae5b520e5512a9f2d1a2c50a6b500fcb67b89f5382d82718545a337caf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 521 + ], + "type": "text", + "content": "The final score will be used to sort the image/text features and the top-k features are considered as image/text key entities. To unify the expression, we treat all the question features as key entities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 528, + 457, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 457, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 457, + 540 + ], + "type": "text", + "content": "4.3. Cross-Media Reason Layer" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "text", + "content": "The CMR layer is designed for multi-step reasoning across medias. As show in Fig. 6, we first gather the key entities as " + }, + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "inline_equation", + "content": "S \\in R^{l_k \\times d}" + }, + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "text", + "content": " from input features, where " + }, + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "inline_equation", + "content": "l_k" + }, + { + "bbox": [ + 304, + 546, + 545, + 654 + ], + "type": "text", + "content": " is the total number of all key entities. Then the key entities pass through self-attention, global cross-attention based on original features and feed-forward unit. Finally, the key entities are scattered to the input features. As the requirement for multi-step reasoning, we inserted multiple CMR layers in a KECMR module. The CMR layer can be formulated by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 378, + 665, + 545, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 665, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 378, + 665, + 545, + 678 + ], + "type": "interline_equation", + "content": "S = \\operatorname {g a t h e r} ([ T, I, Q ]) \\tag {8}", + "image_path": "ab3650a8dcbd63fe08f797873751d339b1149116aaed7b5de6dae45dc1a857c7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 683, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 683, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 312, + 683, + 545, + 696 + ], + "type": "interline_equation", + "content": "S = F F N (M H (M H (S, S, S), [ T, I, Q ], [ T, I, Q ])) \\tag {9}", + "image_path": "39de692f5829eebb3d89f6fbbeae605c6bc335aef364fb923fae4db3b47a630a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 359, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 359, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "[ T, I, Q ] = \\operatorname {s c a t t e r} (S, [ T, I, Q ]) \\tag {10}", + "image_path": "265c7759088208b40258cf36e8b2d4772cc03f44a5be5a0ceeee0e6271f6c16c.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27222" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 264, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 264, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 264, + 85 + ], + "type": "text", + "content": "4.4. Key Entity Cross-Media Reason Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 287, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 287, + 221 + ], + "type": "text", + "content": "We use the same network framework as [26] but replace MCA layer with our KECMR module and add an extra text stream. As show in Figure 4, The input image is represented as a set of regional visual features in a bottom-up manner [1]. The input question and text are transformed to features by passing through a one-layer LSTM network [7]. Then we use our KECMR module several times to extract key entities and conduct multi-steps cross-media reasoning. Finally, we use the attention reduce layer to fuse the multimedia features and project the fused feature into the answer probability distribution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 222, + 256, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 222, + 256, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 222, + 256, + 234 + ], + "type": "text", + "content": "The attention reduce layer can be formulated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 254, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 254, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 70, + 254, + 287, + 268 + ], + "type": "interline_equation", + "content": "M L P (x) = W _ {2} \\left(\\max \\left(0, x W _ {1} + b _ {1}\\right)\\right) + b _ {2} \\tag {11}", + "image_path": "03b5bd5a284832df5cfa861fcb53dd748e94e6d247eaf2f0a12f1ddf26c69c2f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 275, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 275, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 67, + 275, + 287, + 289 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t R e} (X) = \\operatorname {s u m} (\\operatorname {s o f t m a x} (M L P (X)) X) \\tag {12}", + "image_path": "1d5bc625f9ef69cf2808f9f3e8da594640f880a2cd202457b701b07cad769b6f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": "We can obtain the attended image feature " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": " by passing image feature I through attention reduce layer, and the same applies to text feature T and question feature Q. Then we fuse the multimedia features by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 353, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 353, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 107, + 353, + 287, + 365 + ], + "type": "interline_equation", + "content": "z = W _ {i} I _ {a} + W _ {t} T _ {a} + W _ {q} Q _ {a} \\tag {13}", + "image_path": "8cda0cac3f9b0e67b6dd957c7a59786773e72e6d6c33e703c0f5671fd60bde1a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "content": "The fused feature " + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "content": " is projected into a vector " + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "s \\in R^{N}" + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "content": " followed by a sigmoid function, where " + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "content": " is the number of the answers in the training set. We use binary cross-entropy (BCE) as the loss function to train an N-way classifier on top of the fused feature " + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 370, + 287, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 128, + 454 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 460, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 496 + ], + "type": "text", + "content": "In this section, we evaluate the current state-of-the-art VQA model and our KECMRN model, and provide results of them in VTQA dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 503, + 160, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 160, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 160, + 514 + ], + "type": "text", + "content": "5.1. Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 545 + ], + "type": "text", + "content": "As the answers divided into three types, we use different metrics for distinct types of answers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 546, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 581 + ], + "type": "text", + "content": "Exact match (EM). This metric measures the percentage of predictions that match the ground truth answer exactly and will be used in all types of answers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "(Macro-averaged) F1 score (F1). This metric measures the average overlap between the prediction and ground truth answer. We treat the prediction and ground truth as bags of tokens and compute their F1. This metric will be used for the " + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": " types of answers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "YN accuracy (YNAcc). This metric is only used for the YN type of answers. The answer will be transformed into 'yes' or 'no' by a pre-defined yes-or-no dictionary, which is manually collected based on the answers in our dataset. Then we calculate the accuracy just based on the yes-or-no answer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 440, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 440, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 440, + 84 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": "KECMRN: the hyper-parameters of our model used in the experiments are as follows. The dimensionality of input image features, input question features, input text features and fused multimodal features are 2,048, 1024, 1024, and 2,048, respectively. The latent dimensionality " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " in the multi-head attention is 1024, the number of heads " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " is set to 8. The number of CMR layer in each KECMR module and the number of KECMR module are 2 and 2. The number of key entities " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " is set to 16. We train our model on the train set for 13 epochs and evaluate our model on test set. The Faster-RCNN used for image feature extraction is pretrained and frozen, while all other modules are trained from scratch." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "text", + "content": "GPT4V [17]/LLaVA [13]/MMICL [27]: popular multimodal large language models (MLLM). These models demonstrated impressive multimodel chat abilities and exceptional performance across numerous intricate visual reasoning datasets, particularly excelling in zero-shot evaluations. Notably, LLaVA achieved an impressive " + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "text", + "content": " accuracy, while GPT4V attained " + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "text", + "content": " and MMICL attained " + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "inline_equation", + "content": "70.3\\%" + }, + { + "bbox": [ + 304, + 234, + 545, + 425 + ], + "type": "text", + "content": " on the VQAv2 dataset, underscoring their remarkable capabilities in the realm of multimodal understanding and visual question answering. We follow the same evaluation strategy as on the VQAv2 dataset, with the sole modification of adapting the prompt to 'Search for answers in the context above, do not use additional knowledge beyond the text, and answer the question using a single word or phrase'. This adjustment facilitated enhanced performance on VTQA datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "BEiT3 [25]/VLMo [3]: visual language models directly pre-trained on image-text pairs, offering versatility in addressing a spectrum of visual-language tasks through finetuning on downstream tasks. Notably, on the VQAv2 dataset, BEiT3-base achieved an accuracy of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "77.65\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ", while VLMo-base demonstrated a commendable accuracy of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "76.6\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ". To evaluate the transferability of BEiT3 and VLMo on our VTQA task, we follow the same fine-tuning strategy as on the VQAv2 dataset, but we concatenate the text and questions into a single paragraph of text and increase the maximum number of training epochs from 10 to 50 to obtain better results, since our dataset is relatively small compared to VQAv2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": "MCAN-region [26]/MCAN-grid [9]: models based on MCAN [26] but using different image features. They achieve SOTA results on VQAv2 dataset (70.93% for MCAN-region and 72.71% for MCAN-grid) among non-pretrained visual-language models(non-PVLM), which indicates that these models were not pretrained on visual-language data, although individual pretraining on either visual or text data is allowed. We concatenate the text and questions and increase the maximum number of input words to train MCAN in VTQA task." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "Trans-CA/Trans-SA: models based on Trans" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27223" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 287, + 290 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 287, + 290 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 287, + 290 + ], + "type": "table", + "html": "
MethodsEMYN-ACCE-F1G-F1
English Version
LLaVA39.9065.0560.0626.20
GPT4V48.9375.5970.2448.37
MMICL38.7662.9062.3245.96
BEiT3-base41.0256.5150.1828.14
VLMO-base51.8565.4664.5942.74
MCAN-region56.0972.0667.0340.52
MCAN-grid55.8870.7365.3939.19
Trans-CA57.3873.5769.0244.14
Trans-SA54.3967.8965.9541.11
KECMRN(ours)57.9574.9968.9144.25
Chinese Version
MCAN-region49.7871.2659.9445.74
MCAN-grid50.2274.1759.0245.59
Trans-CA49.8774.2759.6347.45
Trans-SA48.9973.2958.2646.31
KECMRN(ours)51.3277.5960.5251.11
", + "image_path": "d9758a3b2fcd16879464829e0748b55fe50360977a2edee7b9aee26b03410586.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 341, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 461 + ], + "type": "text", + "content": "former [22]. Trans-CA treat image, text and question features as independent input and perform Self Attention and Question Aware Cross Attention in each layer. The layers in Trans-CA resemble those of the KEE module, excluding the \"score&sort\" step. Conversely, Trans-SA adopts a different approach by concatenating all image, text, and question features and only applying Self Attention in each layer. The layers in Trans-SA resemble those of the CMR module, excluding the \"Global Cross-Attention\" step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 468, + 167, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 167, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 167, + 480 + ], + "type": "text", + "content": "5.3. Results and Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 544 + ], + "type": "text", + "content": "Table 2 provides the results of various baselines on the test set of our VTQA benchmark. VLMs, including MMICL, LLaVA, BEiT3, and VLMo, primarily pretrained on English datasets, are exclusively evaluated in the English version." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "While models such as GPT4V, MMICL, and LLaVA demonstrate superior performance across a variety of visual-language datasets in zero-shot evaluation, they encounter a significant drop in performance when evaluated on the VTQA dataset. This is partly because the model lacks exposure to our training data during training, unlike the VQA dataset included in the model's VL fine-tuning data. The decline also results from the heightened complexity of our VTQA task compared to VQA, posing new challenges to the model's image-text alignment and inference capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "BEiT3 and VLMo, pretrained directly on image-text pairs, achieve impressive results on various multimodal datasets through fine-tuning on downstream tasks. Despite" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 325, + 69, + 528, + 133 + ], + "blocks": [ + { + "bbox": [ + 47, + 298, + 288, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 298, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 288, + 321 + ], + "type": "text", + "content": "Table 2. Results of our KECMRN on test set compared with other multimodal models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 69, + 528, + 133 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 528, + 133 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 528, + 133 + ], + "type": "table", + "html": "
MethodsEMYN-ACCE-F1G-F1
Trans(1)55.4572.4365.3542.27
Trans(3)55.0770.9865.743.58
Trans(6)54.0769.4364.8343.91
LSTM57.9574.9968.9144.25
", + "image_path": "17d16f672adcd92bed16ae679e2828a351a4743954b41d26ef66d48ffc3d6288.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 141, + 541, + 152 + ], + "lines": [ + { + "bbox": [ + 310, + 141, + 541, + 152 + ], + "spans": [ + { + "bbox": [ + 310, + 141, + 541, + 152 + ], + "type": "text", + "content": "Table 3. Results of different text encoders based on KECMRN." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 256 + ], + "type": "text", + "content": "fine-tuning on VTQA training data, BEiT3 and VLMo still exhibit a substantial performance decrease. This is partly attributed to the limitation that the pretraining data for BEiT3 and VLMo only includes image-text pairs with relatively short texts, while the texts in the VTQA dataset are longer, averaging 284 words in English version. This, once again, underscores the challenges inherent in the VTQA task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "text", + "content": "While MCAN models outperform the pre-trained models, a notable decrease is observed. In addition, Trans-CA exceeded the results of all VQA models, while the results of Trans-SA are relatively poor. This highlights the necessity for a more intricate independent processing of text, image, and question features in our VTQA dataset. Our KECMRN achieves the best result with " + }, + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "inline_equation", + "content": "57.95\\%" + }, + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "text", + "content": " on the English version and " + }, + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "inline_equation", + "content": "51.32\\%" + }, + { + "bbox": [ + 304, + 257, + 545, + 352 + ], + "type": "text", + "content": " in Chinese version." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 353, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 580 + ], + "type": "text", + "content": "Fig. 7 illustrates the outcomes of attention visualization across different models. The attention weights from the 'Attention Reduce' module were employed to colorize image regions and organize the Top Words. Notably, our KECMRN model exhibits precise identification of Zhang Yanqi on the left side, engaged in ball play. This accurate recognition is based on key phrases such as sports balls and dribbling skills, leading to a correct inference for the associated question ('Yes'). In contrast, although MCAN also directs attention to the person on the left, the absence of Key-Entity Extract (KEE) causes attention to be excessively dispersed. Consequently, there is an inaccurate localization of the person's upper garment, resulting in an incorrect answer. Furthermore, Trans-SA mistakenly focuses on the person on the right, while Trans-CA struggles to discern whether Zhang Yanqi is positioned on the left or right, both yielding incorrect answers. The attention visualization results underscore the superior precision of our KECMRN model in focusing on the correct visual regions and words." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 587, + 400, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 400, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 400, + 601 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "We conducted a number of ablation experiments to investigate the reasons for the effectiveness of KECMR. All the reported results are evaluated on English version." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Text Encoders: Table 3 provides an overview of the results obtained using different text encoders in conjunction with KECMRN. In this table, Trans(k) denotes a k-layer transformer encoder model. Contrary to expectations, the performance of deeper transformer encoder models tends to be inferior, with LSTM outperforming all transformer en" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27224" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 83, + 216, + 177 + ], + "blocks": [ + { + "bbox": [ + 74, + 83, + 216, + 177 + ], + "lines": [ + { + "bbox": [ + 74, + 83, + 216, + 177 + ], + "spans": [ + { + "bbox": [ + 74, + 83, + 216, + 177 + ], + "type": "image", + "image_path": "39c36cb5a72c3e3bf4d9007f568fc99c6fc196eb84974c4e47a160a3afb6edc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 178, + 94, + 186 + ], + "lines": [ + { + "bbox": [ + 77, + 178, + 94, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 94, + 186 + ], + "type": "text", + "content": "Text:" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 187, + 212, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 187, + 212, + 289 + ], + "spans": [ + { + "bbox": [ + 76, + 187, + 212, + 289 + ], + "type": "text", + "content": "Song Shuai and Zhang Jiaqi are two local sports ball enthusiasts whose video about sports ball on the short video platform caught the attention of a club in a nearby city. Today they were invited to the city where the club is located for a technical exchange, and now Song Shuai is talking to the head of the club, and Zhang Jiaqi is too obsessed with sports balls on the road and still does not forget to practice dribbling skills." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 299, + 108, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 299, + 108, + 308 + ], + "spans": [ + { + "bbox": [ + 77, + 299, + 108, + 308 + ], + "type": "text", + "content": "Question:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 308, + 185, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 308, + 185, + 318 + ], + "spans": [ + { + "bbox": [ + 77, + 308, + 185, + 318 + ], + "type": "text", + "content": "Is Zhang Jiaqi wearing a blue top?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 327, + 105, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 105, + 335 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 105, + 335 + ], + "type": "text", + "content": "Answer:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 336, + 91, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 336, + 91, + 344 + ], + "spans": [ + { + "bbox": [ + 77, + 336, + 91, + 344 + ], + "type": "text", + "content": "Yes" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 222, + 71, + 364, + 166 + ], + "blocks": [ + { + "bbox": [ + 222, + 71, + 364, + 166 + ], + "lines": [ + { + "bbox": [ + 222, + 71, + 364, + 166 + ], + "spans": [ + { + "bbox": [ + 222, + 71, + 364, + 166 + ], + "type": "image", + "image_path": "c64ae40c2457a6d4a0a8113215c982f8620aebae8bf75f42fa9d8bc9bca25502.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 168, + 264, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 168, + 264, + 177 + ], + "spans": [ + { + "bbox": [ + 225, + 168, + 264, + 177 + ], + "type": "text", + "content": "Top Words:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 178, + 351, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 178, + 351, + 204 + ], + "spans": [ + { + "bbox": [ + 225, + 178, + 351, + 204 + ], + "type": "text", + "content": "balls, Jiaqi, enthusiasts, Zhang, road, skills, obsessed, dribbling, sports, forget Predict Answer: Yes" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 222, + 220, + 362, + 315 + ], + "blocks": [ + { + "bbox": [ + 274, + 208, + 310, + 216 + ], + "lines": [ + { + "bbox": [ + 274, + 208, + 310, + 216 + ], + "spans": [ + { + "bbox": [ + 274, + 208, + 310, + 216 + ], + "type": "text", + "content": "KECMRN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 220, + 362, + 315 + ], + "lines": [ + { + "bbox": [ + 222, + 220, + 362, + 315 + ], + "spans": [ + { + "bbox": [ + 222, + 220, + 362, + 315 + ], + "type": "image", + "image_path": "b0074f83903e06aef247d91cb0e37d8d18c8f0d1f3fb57dbf1b4f28bda46b286.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 225, + 317, + 264, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 317, + 264, + 326 + ], + "spans": [ + { + "bbox": [ + 225, + 317, + 264, + 326 + ], + "type": "text", + "content": "Top Words:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 225, + 327, + 348, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 327, + 348, + 353 + ], + "spans": [ + { + "bbox": [ + 225, + 327, + 348, + 353 + ], + "type": "text", + "content": "sports, dribbling, Jiaqi, club, Shuai, Zhang, video, caught, platform, talking Predict Answer: No" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 275, + 357, + 309, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 357, + 309, + 365 + ], + "spans": [ + { + "bbox": [ + 275, + 357, + 309, + 365 + ], + "type": "text", + "content": "Trans-SA" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 519, + 166 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 519, + 166 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 519, + 166 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 519, + 166 + ], + "type": "image", + "image_path": "ea7925dc6cd291f2c8dc6a3a4a014aaac157aae22f482d51db66db1ea0bd8c6a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 168, + 421, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 168, + 421, + 177 + ], + "spans": [ + { + "bbox": [ + 382, + 168, + 421, + 177 + ], + "type": "text", + "content": "Top Words:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 178, + 514, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 178, + 514, + 204 + ], + "spans": [ + { + "bbox": [ + 382, + 178, + 514, + 204 + ], + "type": "text", + "content": "skills, dribbling, enthusiasts, nearby, road, balls, Zhang, club, still, sports Predict Answer: No" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 378, + 220, + 519, + 315 + ], + "blocks": [ + { + "bbox": [ + 436, + 208, + 462, + 216 + ], + "lines": [ + { + "bbox": [ + 436, + 208, + 462, + 216 + ], + "spans": [ + { + "bbox": [ + 436, + 208, + 462, + 216 + ], + "type": "text", + "content": "MCAN" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 220, + 519, + 315 + ], + "lines": [ + { + "bbox": [ + 378, + 220, + 519, + 315 + ], + "spans": [ + { + "bbox": [ + 378, + 220, + 519, + 315 + ], + "type": "image", + "image_path": "13a5d0a27bd4e1632a8daf46b6ff066113c94c816376ea668c4eaab3aa39dd1e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 379, + 421, + 390 + ], + "lines": [ + { + "bbox": [ + 171, + 379, + 421, + 390 + ], + "spans": [ + { + "bbox": [ + 171, + 379, + 421, + 390 + ], + "type": "text", + "content": "Figure 7. An example of attention visualization for different models." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 382, + 317, + 421, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 317, + 421, + 327 + ], + "spans": [ + { + "bbox": [ + 382, + 317, + 421, + 327 + ], + "type": "text", + "content": "Top Words:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 382, + 327, + 509, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 327, + 509, + 354 + ], + "spans": [ + { + "bbox": [ + 382, + 327, + 509, + 354 + ], + "type": "text", + "content": "Jiaqi, Song, exchange, dribbling, Zhang, obsessed, head, forget, short, Shuai Predict Answer: No" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 432, + 357, + 466, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 357, + 466, + 365 + ], + "spans": [ + { + "bbox": [ + 432, + 357, + 466, + 365 + ], + "type": "text", + "content": "Trans-CA" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 61, + 408, + 276, + 472 + ], + "blocks": [ + { + "bbox": [ + 61, + 408, + 276, + 472 + ], + "lines": [ + { + "bbox": [ + 61, + 408, + 276, + 472 + ], + "spans": [ + { + "bbox": [ + 61, + 408, + 276, + 472 + ], + "type": "table", + "html": "
MethodsEMYN-ACCE-F1G-F1
Trans-CA57.3873.5769.0244.14
CA+SA57.1773.5768.9642.88
KEE+SA57.6374.4269.143.06
KEE+CMR57.9574.9968.9144.25
", + "image_path": "895b81b74cebcaf0f3008f36c3489f093e55c4800e7d322157ed503a3491f3f0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 480, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 502 + ], + "type": "text", + "content": "Table 4. Results of the KECMRN model with different module combinations." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 538, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 287, + 574 + ], + "type": "text", + "content": "coder models. This phenomenon is attributed to both insufficient training data and the intricate structure of transformer models." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 288, + 713 + ], + "type": "text", + "content": "KEE/CA and CMR/SA: Table 4 displays the results of the KECMRN model with different module combinations. In this table, CA+SA denotes the alternate use of CA and SA layers, similar to our KECMRN design. KEE+SA follows a similar pattern. Compared to Trans-CA, replacing some CA layers with SA layers results in a slight performance decrease (↓0.21). Conversely, the inclusion of the KEE layer improves performance (↑0.46) by concentrating information on key entities. Additionally, the CMR layer further optimizes features related to key entities, leading to enhanced results (↑0.32)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 411, + 378, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 378, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 378, + 423 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 432, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 612 + ], + "type": "text", + "content": "We propose a novel cross-modal question answering dataset, VTQA, which necessitates models to acquire pertinent information from both text and image sources and perform complex cross-modal reasoning to answer questions. Our experimental findings reveal that even state-of-the-art pretrained models fall short of achieving satisfactory performance on VTQA, underscoring the challenges and potential inherent in the dataset for cross-modal question-answering tasks. Looking ahead, we aim to augment the dataset's size and coverage and provide additional reasoning annotations in future work to more effectively assess and catalyze the development of cross-modal reasoning models. Simultaneously, we aspire to explore more fitting evaluation metrics to surmount the limitations of EM metrics and more equitably gauge the model's question-answering proficiency." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 626, + 410, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 410, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 410, + 639 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of China under Grant 62073105, U23A20389 and 62306095, by the Natural Science Foundation of Heilongjiang Province of China under Grant ZD2022F002, and by the Heilongjiang Touyan Innovation Team Program." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27225" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 178 + ], + "type": "text", + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 179, + 288, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 233 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 2425-2433, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 235, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 288, + 300 + ], + "type": "text", + "content": "[3] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In Advances in Neural Information Processing Systems, pages 32897-32912. Curran Associates, Inc., 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "text", + "content": "[4] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking to a machine? dataset and methods for multilingual image question. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "text", + "content": "[5] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 6325-6334. IEEE Computer Society, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "text", + "content": "[6] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 479, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 502 + ], + "type": "text", + "content": "[7] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "text", + "content": "[8] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "text", + "content": "[9] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10264-10273, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 613, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 288, + 690 + ], + "type": "text", + "content": "[10] Aniruddha Kembhavi, Min Joon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pages 5376-5384. IEEE Computer Society, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 288, + 714 + ], + "type": "text", + "content": "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalan-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "text", + "content": "tidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vision, 123(1):32-73, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "type": "text", + "content": "[12] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "type": "text", + "content": "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[14] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 342 + ], + "type": "text", + "content": "[15] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1682-1690, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 344, + 547, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 547, + 398 + ], + "type": "text", + "content": "[16] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "type": "text", + "content": "[17] OpenAI, Josh Achiam, Steven Adler, et al. Gpt-4 technical report, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 423, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 544 + ], + "type": "text", + "content": "[18] Revanth Gangi Reddy, Xilin Rui, Manling Li, Xudong Lin, Haoyang Wen, Jaemin Cho, Lifu Huang, Mohit Bansal, Avirup Sil, Shih-Fu Chang, Alexander G. Schwing, and Heng Ji. Mumuqa: Multimedia multi-hop news question answering via cross-media knowledge extraction and grounding. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 11200-11208. AAAI Press, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "text", + "content": "[19] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 590, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 646 + ], + "type": "text", + "content": "[20] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision - ECCV 2022, pages 146-162, Cham, 2022. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "text", + "content": "[21] Alon Talmor, Ori Yoran, Amnon Catav, Dan Lahav, Yizhong Wang, Akari Asai, Gabriel Ilharco, Hannaneh Hajishirzi, and Jonathan Berant. Multimodalqa: complex question answering over text, tables and images. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27226" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 495 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "text", + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[23] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton Van Den Henge. Explicit knowledge-based reasoning for visual question answering. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, page 1290–1296. AAAI Press, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "text", + "content": "[24] Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. FVQA: fact-based visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 40(10): 2413-2427, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 327 + ], + "type": "text", + "content": "[25] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 329, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 329, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 329, + 287, + 393 + ], + "type": "text", + "content": "[26] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6281-6290. Computer Vision Foundation / IEEE, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "type": "text", + "content": "[27] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[28] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4995–5004, 2016. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27227" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/VTimeLLM_ Empower LLM to Grasp Video Moments/93551b6a-586f-4f31-83ad-4d86b76784d8_content_list.json b/2024/VTimeLLM_ Empower LLM to Grasp Video Moments/93551b6a-586f-4f31-83ad-4d86b76784d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b95b698d579b4cf8b6ca587f7520241cd3f9a04e --- /dev/null +++ b/2024/VTimeLLM_ Empower LLM to Grasp Video Moments/93551b6a-586f-4f31-83ad-4d86b76784d8_content_list.json @@ -0,0 +1,1526 @@ +[ + { + "type": "text", + "text": "VTimeLLM: Empower LLM to Grasp Video Moments", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Huang, Xin Wang\\*, Hong Chen, Zihan Song, Wenwu Zhu\\* Tsinghua University", + "bbox": [ + 241, + 180, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangb23, h-chen20, songzh23}@mails.tsinghua.edu.cn", + "bbox": [ + 254, + 218, + 712, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xin_wang,wwzhu}@tsinghua.edu.cn", + "bbox": [ + 343, + 237, + 622, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have shown remarkable text understanding capabilities, which have been extended as Video LLMs to handle video data for comprehending visual details. However, existing Video LLMs can only provide a coarse description of the entire video, failing to capture the precise start and end time boundary of specific events. In this paper, we solve this issue via proposing VTimeLLM, a novel Video LLM designed for fine-grained video moment understanding and reasoning with respect to time boundary. Specifically, our VTimeLLM adopts a boundary-aware three-stage training strategy, which respectively utilizes image-text pairs for feature alignment, multiple-event videos to increase temporal-boundary awareness, and high-quality video-instruction tuning to further improve temporal understanding ability as well as align with human intents. Extensive experiments demonstrate that in fine-grained time-related comprehension tasks for videos such as Temporal Video Grounding and Dense Video Captioning, VTimeLLM significantly outperforms existing Video LLMs. Besides, benefits from the fine-grained temporal understanding of the videos further enable VTimeLLM to beat existing Video LLMs in video dialogue benchmark, showing its superior cross-modal understanding and reasoning abilities.", + "bbox": [ + 75, + 316, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have garnered significant attention due to their exceptional capabilities in text understanding and generation [23, 31]. However, harnessing the potential of LLMs for understanding and reasoning over multimodal data, especially videos, still remains a substantial challenge. This is because analyzing videos requires deep understanding of both visual details and temporal dynamics for models. Several preliminary at", + "bbox": [ + 75, + 733, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fa84e9592bfae4b89febfafb4bf1d51005717a182cd5aa1ed4beaf0db2eb080.jpg", + "image_caption": [ + "Figure 1. Existing Video LLMs, such as VideoChat and VideoLLaMA, exhibit a deficiency in boundary awareness, leading to challenges in accurately capturing the precise timestamps of specific events." + ], + "image_footnote": [], + "bbox": [ + 531, + 285, + 875, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tempts [15, 20, 34] utilize LLMs for video understanding. Nevertheless, these works predominantly focus on generation of generic video captions and can merely offer surface-level summaries of the content, thus failing to capture the relationships between specific moment boundaries and the bounded events, as shown in Figure 1.", + "bbox": [ + 496, + 534, + 890, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the problem, in this paper, we investigate improving the boundary-aware ability of Video LLM, which faces the following two challenges.", + "bbox": [ + 498, + 626, + 890, + 671 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. There is a scarcity of large-scale video datasets with accurate boundary annotations to train the Video LLM for temporal alignment.", + "2. It is non-trivial to design effective temporal-related video tasks for training LLM to understand the content of multiple moments within videos." + ], + "bbox": [ + 500, + 672, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose VTimeLLM, a novel Video LLM that can perceive fine-grained segments in videos with better temporal reasoning ability. VTimeLLM consists of i) a visual encoder and a visual adapter to process the input video, and ii) a tailored LLM to understand both text and video content, which is trained via a novel boundary-aware three-stage training strategy. Specifically, visual features are aligned with LLM's semantic space through image-text training in the first stage.", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors.", + "bbox": [ + 94, + 862, + 230, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our project page is at https://github.com/huangb23/", + "bbox": [ + 96, + 875, + 467, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "VTimeLLM", + "bbox": [ + 78, + 888, + 145, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14271", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the second stage, we then design the single-turn and multi-turn related question answering (QA) tasks to endue VTimeLLM with the awareness of time boundaries and the ability to understand the corresponding events bounded within the boundaries. We employ a large-scale video-text dataset containing multiple segments together with their roughly annotated labels for training VTimeLLM with the QA tasks. Finally, in the third stage, we further create a high-quality dialogue dataset for instruction tuning, which simultaneously aligns VTimeLLM with human intention and enables VTimeLLM to conduct temporal understanding for video segments more precisely. Extensive experiments show that VTimeLLM significantly outperforms existing Video LLMs in time-related video understanding tasks, such as Temporal Video Grounding and Dense Video Captioning. In addition, benefiting from the fine-grained temporal understanding of videos, VTimeLLM is able to beat existing Video LLMs in video dialogue benchmark, demonstrating its superiority in cross-modal understanding and reasoning for videos. Our contributions in this paper are listed as follows,", + "bbox": [ + 76, + 90, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose VTimeLLM, the first boundary-aware Video LLM, to the best of our knowledge.", + "- We propose the boundary-aware three-stage training strategy, which consecutively leverages i) large-scale imag-text data for feature alignment, ii) large-scale multi-event video-text data together with the temporal-related single-turn and multi-turn QA to enhance the awareness of time boundary, and iii) instruction tuning on the high-quality dialog dataset for better temporal reasoning ability.", + "- We conduct extensive experiments to demonstrate that the proposed VTimeLLM significantly outperforms existing Video LLMs in various fine-grained temporal-related video tasks, showing its superior ability for video understanding and reasoning." + ], + "bbox": [ + 76, + 411, + 467, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 638, + 225, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Model", + "text_level": 1, + "bbox": [ + 76, + 665, + 387, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image LLMs To enable Large Language Models (LLMs) to comprehend visual information, significant efforts have been made to align visual and linguistic modalities. BLIP-2 [14] introduced the concept of Q-Former, utilizing learnable query vectors to extract visual features from frozen image encoders. MiniGPT-4 [37] demonstrated that further fine-tuning with detailed image descriptions significantly enhances its usability. LLAVA [16] explored diverse multi-modal instruction-following data, aiming to construct a general-purpose visual assistant. Recent endeavors, such as Kosmos-2 [21] and VisionLLM [26], delved into more detailed aspects of image comprehension, including referencing and grounding, significantly enhancing the capability to describe intricate image details.", + "bbox": [ + 76, + 688, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video LLMs Driven by the success of Image LLM, researchers have naturally extended their focus from single-frame images to multi-frame videos, leading to the emergence of Video-compatible LLMs like VideoChat [15], Video-LLaMA [34], and Video-ChatGPT [20]. These models employ a two-stage training strategy. In the first stage, large-scale datasets align video features with the feature space of LLMs. In the second stage, a limited amount of GPT-annotated or human-annotated datasets are used for instruction tuning. While these models exhibit impressive overall video comprehension, their abilities to describe specific video segments and perform temporal reasoning remain limited. The limitation arises mainly due to the nature of datasets used in the first training stage, such as WebVid [2], which usually consist of one-event videos and noisy textual annotations. Moreover, the scarcity of high-quality, temporally annotated data in the second stage poses a challenge for models to conduct temporal reasoning. To bridge this gap, our approach, VTimeLLM, introduces a boundary perception stage between these two stages. This stage enables the model to precisely locate events within videos and describe multiple distinct events accurately, empowering our model to grasp fine-grained details of video moments.", + "bbox": [ + 496, + 90, + 893, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Fine-Grained Video Understanding", + "text_level": 1, + "bbox": [ + 500, + 463, + 808, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-grained video understanding, the ability to precisely locate and comprehend specific events within a video, is a crucial challenge for video analysis. When integrated with natural language, there are two primary tasks: Temporal Video Grounding [1, 8] and Dense Video Captioning [12, 25].", + "bbox": [ + 496, + 486, + 890, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Video Grounding Temporal Video Grounding aims to identify corresponding video segments for given textual inputs. Traditional approaches can be categorized into two types: proposal-based [4, 29, 32] and proposal-free methods [9, 30, 33]. Proposal-based techniques generate candidate proposals before ranking them based on relevance. In contrast, proposal-free methods directly predict the start and end boundaries of the target moment.", + "bbox": [ + 496, + 595, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense Video Captioning Dense Video Captioning is a more intricate task, demanding both temporal localization and captioning for all events within an untrimmed video. Earlier methods [6, 11, 12] employed a two-stage process involving temporal localization followed by event captioning. Recent developments [25, 28, 36] in this field have witnessed a shift towards joint training of captioning and localization modules. For instance, Vid2Seq [28], enhances a language model by incorporating specific time tokens, enabling the model to generate event boundaries and textual descriptions within the unified output sequence.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14272", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both these two tasks share a fundamental requirement: the alignment of video segments with semantic context. Leveraging the power of LLM with the help of our training strategy, our VTimeLLM model unifies these tasks and has demonstrated remarkable effectiveness. Concurrently, VTimeLLM enables natural language interaction with humans, establishing itself as an excellent assistant for comprehending video content.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VTimeLLM: Being Aware of Time Boundaries in Videos", + "text_level": 1, + "bbox": [ + 76, + 224, + 468, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce VTimeLLM, which is designed to grasp precise video moments for LLMs. We first provide a detailed description of the model architecture, and then our innovative boundary-aware three-stage training framework, as shown in Figure 2.", + "bbox": [ + 75, + 268, + 468, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 76, + 353, + 210, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the LLM to comprehend videos, our VTimeLLM model incorporates two additional modules within LLM, i.e., the visual encoder and the visual adapter, which transform the visual information into text space.", + "bbox": [ + 75, + 375, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Encoder Our VTimeLLM model utilizes a frozen CLIP ViT-L/14 [22] as the visual encoder, and for simplicity we denote it as ViT. Given a video $V \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ with $T$ frames, we uniformly sample $N = 100$ frames, represented as $\\tilde{V} \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ , where $\\tilde{V}_1 = V_1$ and $\\tilde{V}_N = V_T$ . Each frame $\\tilde{V}_i$ is independently processed through the visual encoder:", + "bbox": [ + 75, + 454, + 468, + 560 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{v _ {i} ^ {c l s}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {p} \\} = \\operatorname {V i T} (\\tilde {V} _ {i}), i = 1, 2, \\dots , N, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 570, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p$ represents the number of patches in the ViT.", + "bbox": [ + 76, + 599, + 426, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Adapter We utilize the global feature $v_{i}^{cls}$ as the feature for the i-th frame, and apply a linear layer $f(\\cdot)$ to project the features of each frame into the same embedding space as that of LLM:", + "bbox": [ + 75, + 631, + 468, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = f \\left(v _ {i} ^ {\\text {c l s}}\\right), i = 1, 2, \\dots , N. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 702, + 468, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, $Z = \\{z_{i}\\} \\in \\mathbb{R}^{N\\times d}$ becomes the input sequence that LLM can comprehend, and $d$ is the hidden dimension of LLM.", + "bbox": [ + 75, + 729, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that in the visual modules, we do not model the temporal relationships for the frames, inspired by the fact that the LLM itself can receive sequential input embeddings and capture their temporal relations.", + "bbox": [ + 75, + 777, + 468, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input of LLM To enable the simultaneous processing of video and text inputs, we introduce a special token, $\\langle \\text{video} \\rangle$ , to represent the video content. More intuitively,", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "if we want to ask the LLM questions about the video, we can use a mixed description, e.g., \"This is a video